Proceedings of The Eleventh International Conference On Management Science and Engineering Management

Lecture Notes on Multidisciplinary Industrial Engineering
Jiuping Xu · Mitsuo Gen

Asaf Hajiyev · Fang Lee Cooke Editors
Proceedings of the Eleventh

International Conference
on Management Science
and Engineering
Management
Lecture Notes on Multidisciplinary Industrial
Engineering
Series editor
J Paulo Davim, Dept of Mechanical Engineering, University of Aveiro
Aveiro, Portugal
More information about this series at http://www.springer.com/series/15734
Jiuping Xu Mitsuo Gen
•
Asaf Hajiyev Fang Lee Cooke

•
Editors
Proceedings of the Eleventh

International Conference
on Management Science
and Engineering
Management
123
Editors
Jiuping Xu Asaf Hajiyev
Business and Administration Azerbaijan National Academy of Sciences
Sichuan University Baku
Chengdu Azerbaijan
China
Fang Lee Cooke
Mitsuo Gen Monash University
Fuzzy Logic Systems Institute Clayton, VIC
Tokyo University of Science Australia
Tokyo, Tokyo
Japan
Lecture Notes on Multidisciplinary Industrial Engineering

ISBN 978-3-319-59279-4 ISBN 978-3-319-59280-0 (eBook)
DOI 10.1007/978-3-319-59280-0
Library of Congress Control Number: 2017942984
© Springer International Publishing AG 2018

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made. The publisher remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.
Printed on acid-free paper
This Springer imprint is published by Springer Nature

The registered company is Springer International Publishing AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
Welcome to the proceedings of the Eleventh International Conference on

Management Science and Engineering Management (ICMSEM 2017), which will
be held from July 28 to 31, 2017, at Kanazawa, Japan. The International
Conference on Management Science and Engineering Management (ICMSEM) is
an annual conference organized by the International Society of Management
Science and Engineering Management (ISMSEM). The conference is to foster
international research collaborations in Management Science and Engineering
Management as well as to provide a forum for presenting current research work in
the forms of technical sessions, roundtable discussions during the conference period
in a nice relaxing enjoyable atmosphere. The participants can share their academic
achievements in MS and EM, and communicate with others during the conference.
The ICMSEM has been held ten times since 2007 in meeting locations across
Asia, Europe, and the Americas and has had a great influence on MS and EM
research. In the past ten years, the ICMSEM has been successfully held in Chengdu,
Chongqing, Bangkok, Chungli, Macau, Islamabad, Philadelphia, Lisbon,
Karlsruhe, and Baku. All accepted papers were put into the proceedings for each
International Conference on Management Science and Engineering Management
and proceedings for the last five conferences can be retrieved from EI Compendex.
This year, 939 papers from 32 countries were received and 150 papers from 20
countries accepted for presentation or poster display at the conference. These papers
were from Algeria, Australia, Azerbaijan, Bangladesh, Canada, China, Indonesia,
Iran, Iraq, Japan, Korea, the Republic of Moldova, Pakistan, Portugal, Spain,
Thailand, the UK, the USA, Uzbekistan, and Vietnam. Each accepted paper was
reviewed by three reviewers, who gave objective and helpful revision advice to the
authors where necessary, which has made these conference proceedings of very
high quality. The papers in the proceedings have been classified into eight sections:
Computing Methodology, Data Analysis, Enterprise Operation Management,
Decision Support System, Green Supply Chain, Resource Optimization
Management, Risk Control, and Integrated Project Management. The key issues at
the eleventh ICMSEM covered many areas of current popularity in Management
Science and Engineering Management. To further encourage state-of-the-art
v
vi Preface
research in the field of Management Science and Engineering Management, the

ISMSEM Advancement Prize for Management Science and Engineering
Management was awarded for the excellent papers at the conference.
We would like to take this opportunity to thank all participants, all of whom
worked exceptionally hard to ensure this conference was a success. We want to
express our sincere gratitude to the following prestigious academies and institutions
for their high-quality papers and great support for the ICMSEM: the Azerbaijan
Academy of Sciences, the Moldova Academy of Sciences, Academy of Sciences
of the Republic of Uzbekistan, Fuzzy Logic Systems Institute, and Tokyo
University of Science, Brock University. We would also like to acknowledge the
assistance we received from the International Society of Management Science and
Engineering Management, Kanazawa University, and Sichuan University for the
conference organization. We also appreciate Springer-Verlag for the publication
of the proceedings. We are grateful to Professor Yoshihiko Uesugi as the General
Chair, Professor Mistuo Gen and Professor Hidetaka Nambo for their work as the
Organizing Committee Chair, and Professor Takashi Oyabu as Local Arrangement
Committee Chair. We appreciate the great support received from all members of the
Organizing Committee, the Local Arrangement Committee, and the Program
Committee as well as all participants. Finally, we would like to thank all researchers
for their excellent conference papers. Because of the high quality of the research,
the ISMSEM Advancement Prize for MSEM is to be awarded again for papers
which have focused on innovative practical applications for Management Science
and Engineering Management.
As MSEM research is in continuous development, many new MSEM devel-
opment trends have emerged. Our work needs to continue to focus on MSEM
development so as to encourage greater and more innovative development activity.
Next year, we plan to continue the novel and successful ICMSEM and intend to
increase our efforts to improve the quality of the proceedings and to recommend
more excellent papers for the ISMSEM Advancement Prize. The Twelfth
International Conference on Management Science and Engineering Management
will be hosted by Monash University, Melbourne, Australia, in July 2018. Professor
Lee has been nominated as the Organizing Committee Chair for the 2018
ICMSEM. We sincerely hope you can submit your new MSEM findings and share
your ideas in Melbourne, Australia.
May 2017 Jiuping Xu

Mitsuo Gen
Asaf Hajiyev
Fang Lee Cooke
Organization
ICMSEM 2017 was organized by International Society of Management Science and

Engineering Management, Sichuan University (Chengdu, China), and Kanazawa
University (Kanazawa, Japan). It was held in cooperation with Lecture Notes on
Multidisciplinary Industrial Engineering (LNMIE) of Springer.
Executive Committee
General Chairs
Jiuping Xu Sichuan University, China

Yoshihiko Uesugi Kanazawa University, Japan
Program Committee Chairs
Mitsuo Gen Fuzzy Logic Systems Institute and Tokyo

University of Science
Shavkat Ayupov Academy of Sciences of Uzbekistan, Uzbekistan
Le Dung Trang Abdus Salam ICTP, Vietnamese
Benjamin Lev Drexel University, Philadelphia, USA
José Antonio de la Peña Mena Mexican Academy of Sciences, Mexico
Asaf Hajiyev Institute of Systems Control, Azerbaijan National
Academy of Sciences, Baku, Azerbaijan
V. Cruz Machado Universidade Nova de Lisboa, Lisbon, Portugal
Alberto Pignotti University of Buenos Aires, Argentina
Baldev Raj Indian National Academy of Engineering, India
Roshdi Rashed National Center of Scientific Research, France
Haruhiko Kimura Kanazawa University, Japan
vii
viii Organization
Hayato Ohwada Tokyo University of Science, Japan

Takashi Oyabu Kokusai Business Gakuin College, Kanazawa,
Japan
Organizing Committee Chair
Mitsuo Gen Fuzzy Logic Systems Institute and Tokyo

University of Science
Hidetaka Nambo Kanazawa University
Local Arrangement Committee Chair
Takashi Oyabu Kokusai Business Gakuin College, Kanazawa,

Japan
Local Arrangement Committee
Ayako Sawada Hokuriku Gakuin Junior College, Kanazawa,

Japan
Program Committee
Mohammad Z. Abu-Sbeih King Fahd University of Petroleum and Minerals,

Saudi Arabia
Joseph G. Aguayo University of Concepcion, Chile
Basem S. Attili United Arab Emirates University,
United Arab Emirates
Alain Billionnet Ecole National Superieure Informatics
for Industry and Enterprise, France
Borut Buchmeister University of Maribor, Slovenia
Daria Bugajewska Adam Mickiewicz University, Poland
Saibal Chattopadhyay Indian Institute of Management, India
Edwin Cheng Hong Kong Polytechnic University, Hong Kong
Anthony Shun Fung Chiu De La Salle University, Philippines
Jeong-Whan Choi Department of Mathematics, Republic of Korea
Kaharudin Dimyati University of Malaya, Malaysia
Behloul Djilali University of Sciences and Technology Houari
Boumediene, Algeria
Eid Hassan Doha Cairo University, Giza, Egypt
O’Regan Donal National University of Ireland, Ireland
Siham El-Kafafi Manukau Institute of Technology, New Zealand
Organization ix
Christodoulos A. Floudas Princeton University, USA

Masao Fukushima Kyoto University, Japan
Oleg Granichin Sankt-Petersburg State University, Russia
Bernard Han Western Michigan University, USA
Rene Henrion Humboldt University, Germany
Voratas Kachitvichyanukul Asian Institute of Technology, Thailand
Arne Løkketangen Molde University College, Norway
Andres Medaglia University of the Andes, Colombia
Venkat Murali Rhodes University, South Africa
Shmuel S. Oren University of California Berkeley, USA
Turgut Öziş Ege University, Turkey
Panos M. Pardalos University of Florida, USA
Gianni Di Pillo Sapienza University of Rome, Italy
Nasrudin Abd Rahim University of Malaya, Malaysia
Celso Ribeiro Fluminense Federal University, Brazil
Hsin Rau Chung Yuan Christian University, Taiwan
Jan Joachim Ruckmann University of Birmingham, UK
Martin Skitmore Queensland University of Technology, Australia
Frits C.R. Spieksma Katholieke University Leuven, Belgium
Yong Tan University of Washington, USA
Albert P.M. Wagelmans Erasmus University Rotterdam, Netherlands
Desheng Dash Wu University of Toronto, Canada
Hong Yan Hong Kong Polytechnic University, Hong Kong
Secretary-General
Zhineng Hu Sichuan University, China
Under-Secretary
Lin Zhong Sichuan University, China
General
Zongmin Li Sichuan University, China
Secretaries
Ning Ma, Jingqi Dai, Lurong Fan, Yan Wang

Contents
Volume 1
Advances in Management Science of the Eleventh ICMSEM

Advancement of Computing Methodology, Data Analysis,
Enterprise Operation Management and Decision Support
System Based on the Eleventh ICMSEM Proceedings . . . . . . . . . . . . . . . 3
Jiuping Xu
Computing Methodology
A Comparison of Pretest, Stein-Type and Penalty Estimators
in Logistic Regression Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Orawan Reangsephet, Supranee Lisawadi, and Syed Ejaz Ahmed
Multi-objective Job Shop Rescheduling with Estimation
of Distribution Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
Xinchang Hao, Lu Sun, and Mitsuo Gen
Multi-Queue Priority Based Algorithm for CPU Process Scheduling . . . . 47
Usman Rafi, Muhammad Azam Zia, Abdul Razzaq, Sajid Ali,
and Muhammad Asim Saleem
An Order-Based GA for Robot-Based Assembly Line
Balancing Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Lin Lin, Chenglin Yao, and Xinchang Hao
A Hybrid Model of AdaBoost and Back-Propagation Neural
Network for Credit Scoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
Feng Shen, Xingchao Zhao, Dao Lan, and Limei Ou
Effects of Urban and Rural Residents’ Behavior Differences in Sports
and Leisure Activity: Application of the Theory of Planned Behavior
and Structural Equation Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Linling Zhang
xi
xii Contents
Fast Multiobjective Hybrid Evolutionary Algorithm Based

on Mixed Sampling Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
Wenqiang Zhang, Yu Wang, Chunxiao Wang, Le Xiao, and Mitsuo Gen
Literature Mining Based Hydrogen Fuel Cell Research . . . . . . . . . . . . . . 117
Luoji Li, Ying Zhang, and Qiulin Li
GM(1.1) Model Based Leshan’s Industries Shift-Share Analysis
and Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
Hongxing Lan, Caiyang Xu, and Zhiyi Meng
Scheduling Problem for Allocating Worker with Class-Type Skill
in JSP by Hybrid Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
Kenichi Ida, Daiki Takano, and Mitsuo Gen
The Rules Determination of Numerical Association
Rule Mining Optimization by Using Combination of PSO
and Cauchy Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
Imam Tahyudin and Hidetaka Nambo
An Analytical Framework for the Conflict Coordination
Mechanism of Knowledge Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
Qifeng Wei
Flight Arrival Scheduling Optimization on Two Runways Based
on IGEP Algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
Rui Wang, Minglei Qu, and Fuzheng Wang
A Study of Urban Climate Change Vulnerability Assessment Based
on Catastrophe Progression Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
Yang Sun, Yi Lu, and Yinghan Li
Judging Customer Satisfaction by Considering Fuzzy Random
Time Windows in Vehicle Routing Problems . . . . . . . . . . . . . . . . . . . . . . 204
Yanfang Ma, Cuiying Feng, Jing Zhang, and Fang Yan
Hybrid Multiobjective Evolutionary Algorithm with Differential
Evolution for Process Planning and Scheduling Problem . . . . . . . . . . . . 212
Chunxiao Wang, Wenqiang Zhang, Le Xiao, and Mitsuo Gen
A New Approach for Solving Optimal Control Problem
by Using Orthogonal Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
Akram Kheirabadi, Asadollah Mahmoudzadeh Vaziri, and Sohrab Effati
The Sustainable Interaction Analysis of Cause Marketing and Ethical
Consumption in Electric Business Platform: Based
on Game Theory and Simulation Analysis . . . . . . . . . . . . . . . . . . . . . . . . 233
Xu Zu and Weiping Yu
Contents xiii
Data Analysis
How to Predict Financing Efficiency in Public-Private Partnerships–In
an Aspect of Uncertainties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
Yixin Qiu, Umair Akram, Sihan Lin, and Muhammad Nazam
The Moderating Effects of Capacity Utilization on the Relationship
Between Capacity Changes and Asymmetric Labor Costs Behavior . . . . 260
Abdulellah Azeez Karrar, DongPing Han, and Sobakinova Donata
The Empirical Analysis of the Impact of Technical Innovation
on Manufacturing Upgrading-Based on Subdivision
Industry of China . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
Dan Jiang and Yuan Yuan
A Crash Counts by Severity Based Hotspot Identification Method
and Its Application on a Regional Map Based Analytical Platform . . . . 286
Xinxin Xu, Ziqiang Zeng, Yinhai Wang, and John Ash
Comparison Between K-Means and Fuzzy C-Means Clustering
in Network Traffic Activities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
Purnawansyah, Haviluddin, Achmad Fanany Onnilita Gafar,
and Imam Tahyudin
RDEU Evolutionary Game Model and Simulation of the Network
Group Events with Emotional Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
Guoqiang Xiong, Xian Wang, Ying Yang, and Yuxi Liu
Effects of Internet Word-of-Mouth of a Tourism Destination
on Consumer Purchase Intention: Based on Temporal Distance
and Social Distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
Mo Chen and Jingdong Chen
Analysis and Prediction of Population Aging Trend Based
on Population Development Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331
Jiancheng Hu
Evaluation of Progressive Team Intervention on Promoting
Physical Exercise Behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
Xinyan Guo
SEM-Based Value Generation Mechanism from Open Government
Data in Environment/Weather Sector . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
Xiaoling Song, Charles Shen, Lin Zhong, and Feniosky Peña-Mora
Impact of Management Information Systems Techniques
on Quality Enhancement Cell’s Report for Higher Education
Commission of Pakistan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
Faraz Ullah Khan and Asif Kamran
xiv Contents
A Priority-Based Genetic Representations for Bicriteria Network

Design Optimizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382
Lin Lin, Jingsong Zhao, Sun Lu, and Mitsuo Gen
Study of County Level Government’s E-Government Efficiency
Evaluation in Sichuan Province Based on DEA . . . . . . . . . . . . . . . . . . . . 398
Yu Liu, Jiawei Liu, Zhaohui Zhang, and Liangqing Zhang
Comparing Visitors’ Behavior Through Mobile Phone Users’
Location Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411
Masahide Yamamoto
Research on Geo/Geo/1 Retrial Queue with Working Vacation
Interruption and Nonpersistent Customers . . . . . . . . . . . . . . . . . . . . . . . . 421
Mingcong Wu, Yong Huang, Yang Song, Liang Zhao, and Jian Liu
Pan-Tourism Urbanization Model Based on System Dynamics:
A Case Study of Barkam . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 438
Lin Hu, Lu Gan, and Qian Yang
A Study on the Three Different Channel Models in Supply
Chain Under the Network Externalities . . . . . . . . . . . . . . . . . . . . . . . . . . 452
Jinjiang Yan, Jiankai Xing, Kai Zhu, and Ke Liu
Construction of Restoration System for Old Books Written
in Braille . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 469
Yuko Shimomura, Hiroyuki Kawabe, Hidetaka Nambo, and Shuichi Seto
Enterprise Operation Management

Discrimination of Personal Opinions from Commercial Speech
Based on Subjectivity of Text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 481
Yoshio Seino and Takahiro Hayashi
Shift in the Regional Balance of Power from Europe to Asia:
A Case Study of ICT Industry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 490
Zahid Latif, Jianqiu Zeng, Shafaq Salam, Zulfiqar Hussain, Lei Wang,
Nasir Jan, and Muhammad Salman
The Empirical Evidence of the Effect on the Enterprises R&D
from Government Subsidies, Political Connections
and Rent-Seeking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499
Dongliang Cai, Zhen Yang, Wu Jiang, and Qiuhua Xu
Scenario-Based Location Arc Routing Problems:
Introducing Mathematical Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 511
Alireza Amini, Reza Tavakkoli-Moghaddam, and Sadoullah Ebrahimnejad
The Impact of Industrial Structure Change on Economic Growth . . . . . 522
Hongxia Liu, Zongtang Xie, and Xin Song
Contents xv
Disclosure Behavior of Annual Reports of Listed Companies

Under Digital Worship . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 530
Xiaojing Xu
Energy Finance in Promoting Renewable Energy Sustainable
Development in China: The Channels and Means of Financing . . . . . . . 538
Lei Xu, Yanfei Deng, Yuan Yuan, and Karen Mancl
The Study on Factors Affecting the Attraction of Microblog – An
Empirical Study Based on Sina Microblogging Site . . . . . . . . . . . . . . . . . 547
Yue He, Yue Zhang, Min Xiao, and Lingxi Song
A Genetic Algorithm Approach for Loading Cells
with Flow Shop Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 559
Patrick Gannon and Gürsel A. Süer
Research on the Competitive Strategy of Two Sided Platform
Enterprises Based on Hotelling Model. . . . . . . . . . . . . . . . . . . . . . . . . . . . 577
Zeming Wang and Yinpeng Guo
Research on Tourists Experience in Traditional Culture Festival
Activities Based on the Importance-Performance Analysis
Research Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 588
Lu Li
How Cross Listing Effects Corporate Performance: Measurement
by Propensity Score Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 600
Yingkai Tang, Huang Huang, Hui Wang, Yue Liu, and Jinhua Xu
Factors Affecting Employee Motivation Towards Employee
Performance: A Study on Banking Industry of Pakistan . . . . . . . . . . . . . 615
Abdullah Khan, Shariq Ahmed, Sameer Paul,
and Syed Hasnain Alam Kazmi
How Content Marketing Can Help the Bank Industrial:
Experience from Iran . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 626
Sima Zomorodian and Yi Lu
The Establishment and Application of AHP-BP Neural Network
Model for Entrepreneurial Project Selection . . . . . . . . . . . . . . . . . . . . . . . 634
Ke Wu and Xiaofeng Li
Improving the User Experience and Virality of Tourism-Related
Facebook Pages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 644
Ayako Sawada and Taketoshi Yoshida
Meta-analysis of the Factors Influencing the Employees’
Creative Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 658
Yang Xu, Ying Li, Hari Nugroho, John Thomas Delaney, and Ping Luo
xvi Contents
The Impact of Mixed Ownership Reform on Enterprise

Performance–An Empirical Study Based on A-Share
Listing Corporation in China . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 670
Sheng Ma, Wenjie Li, and Sicheng Yan
Online Impulse Buying on “Double Eleven” Shopping Festival:
An Empirical Investigation of Utilitarian and Hedonic Motivations . . . . 680
Umair Akram, Peng Hui, Muhammad Kaleem Khan, Muhammad Hashim,
Yixin Qiu, and Ying Zhang
Decision Support System

Fuzzy Multi-attribute Grey Relational Analysis
Using DEA and AHP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 695
Mohammad Sadegh Pakkar
A Fuzzy Multi-criteria Decision Making Approach
for Supplier Selection Under Fuzzy Environment . . . . . . . . . . . . . . . . . . 708
Adnan Sarwar, Ziqiang Zeng, Richard AduAgyapong,
Nguyen ThiHoaiThuong, and Talat Qadeer
Cement Plant Site Selection Problem with Carbon Emission
Trading Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 721
Lurong Fan, Zongmin Li, and Pingwen Wu
Representation and Analysis of Multi-dimensional Data
to Support Decision Making in the Management . . . . . . . . . . . . . . . . . . . 735
Shavkat Ayupov and Abdulla Arifjanov
Fuzzy Logic Applied to SCADA Systems . . . . . . . . . . . . . . . . . . . . . . . . . 749
Tahar Benmessaoud, Alberto Pliego Marugán, Kamal Mohammedi,
and Fausto Pedro García Márquez
On Causal Analysis of Accident and Design of Risk-Proof Procedure
for Nuclear Materials Operation: The Case of JCO Accident . . . . . . . . . 758
Sachiko Oshima
Multi-stage Logistics Inventory for Automobile Manufacturing
by Random Key-Based GA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 768
Hisaki Inoue, Jung Bok Jo, and Mitsuo Gen
The Impact of Temporary Employment on Employees’
Organizational Citizenship Behavior and Turnover Intention:
The Moderating Effect of Organizational Identification . . . . . . . . . . . . . . 791
Xiaoye Qian, Qian Li, Qiong Wu, and Yilun Wu
Contents xvii
Integration of Sound and Image Data for Detection

of Sleep Apnea . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 804
Takehiro Kasahara, Kota Nomura, Yoshihiro Ueda, Yuji Yonezawa,
Masatoshi Saito, Hirohisa Toga, Yuki Fujimoto, Koji Kojima,
Haruhiko Kimura, and Hidetaka Nambo
Pricing Strategy Study on Product Crowdfunding . . . . . . . . . . . . . . . . . . 814
Min Luo, Yong Zhang, Huanmei Zhang, and Xia Li
The Support System of Innovation-Driven Strategy in Private
Enterprises: A Theoretical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 825
Tao Xie, Guichuan Zhou, and Shuangyi Zheng
The Emission Reduction and Recycling of Coal Mine Water
Under Emission Allowance Allocation of Government . . . . . . . . . . . . . . . 835
Ning Ma and Shuangyi Zheng
Modeling and Solving the Vehicle Routing Problem with Multiple
Fuzzy Time Windows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 847
Fang Yan and Yuanyuan Wang
Optimization of Operating of the Systems with Recurrent
Service by Delays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 858
Asaf Hajiyev and Narmina Abdullayeva
An Epidemic Spreading Model Based on Dynamical Network . . . . . . . . 868
Yi Zhang
Analysis of Enterprise Microblog Marketing in Different Industries
Based on DEA Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 878
Dan Zhang, Yufeng Ma, Aixin Wang, Yue He, and Changzheng He
Accounting for Clustering and Non-ignorable Missingness in Binomial
Responses: Application to the Canadian National Survey on Child
Safety Seat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 891
Syed Ejaz Ahmed, Abdulkadir A. Hussein, Anne Snowdon,
and Yiwen Tang
Equity and Sustainability Based Model for Water
Resources Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 902
Yan Tu, Jun Gang, Benjamin Lev, and Xiaoyang Zhou
SCADA and Artificial Neural Networks
for Maintenance Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 912
Alberto Pliego Marugán and Fausto Pedro García Márquez
xviii Contents
Volume 2
Advances in Engineering Management of the Eleventh ICMSEM

Advances in Green Supply Chain, Resource Optimization
Management, Risk Control and Integrated Project Management
Based on the Eleventh ICMSEM Proceedings . . . . . . . . . . . . . . . . . . . . . 923
Jiuping Xu
Green Supply Chain

Modelling a Supply Chain Network of Processed Seafood
to Meet Diverse Demands by Multi-branch Production System . . . . . . . 937
Koichi Murata, Seiichiro Isobe, and Hiroshi Katayama
Hybrid Global Supply Chain Design Considering Product Families
and Layered Cellular Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 947
Jue Jiang and Gürsel A. Süer
Design of Closed-Loop Supply Chain Model with Efficient
Operation Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 962
Xing Chen, Anudari Chuluunsukh, YoungSu Yun, and Mitsuo Gen
Optimization of Supply Chain Based on Internet Plus . . . . . . . . . . . . . . . 975
Wen Hua
Assessing the Recycling Efficiency of Resource in E-Waste Based
on MFA of Reverse Logistics System . . . . . . . . . . . . . . . . . . . . . . . . . . . . 986
Minxi Wang, Xiaoling You, and Xin Li
Analysis of Supply Chain Collaboration with Big Data Suppliers
Participating in Competition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 998
Shuai Liu and Hongchun Wang
Fuzzy GM(1,1) Model Based per Capital Income Predicted of Farmers
in the World Natural and Cultural Heritage Area: Take Leshan City
for an Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1007
Zhiyi Meng, Caiyang Xu, and Hongxing Lan
Eco-Cycle Comprehensive Operation Performance
Evaluation–A Case Study of Baotou Steel Group . . . . . . . . . . . . . . . . . . . 1019
Yuyan Luo, Zhong Wang, Yao Chen, Yahong Wang, and Jiming Xie
Optimization of Closed-Loop Supply Chain Model with Part
and Module Inventory Centers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1030
YoungSu Yun, Anudari Chuluunsukh, Xing Chen, and Mitsuo Gen
The Innovation Research of College Students’ Academic
Early-Warning Mechanism Under the Background of Big Data . . . . . . . 1043
Yu Li and Ye Zhang
Contents xix
Bottleneck Management of Multi-stage Sorting-Packing Operations

with Large-Scale Warehouse System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1054
Tetsuya Sato and Hiroshi Katayama
A Carbon-Constrained Supply Chain Planning Model . . . . . . . . . . . . . . 1067
Zhimiao Tao and Jing Xu
Supply Chain Coordination by Revenue Sharing Contract Under
Different Carbon Emission Policies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1078
Li Lu
Research on Integration of Livestock Products Supply Chain Based
on the Optimal Match Between Supply and Demand. . . . . . . . . . . . . . . . 1089
Liang Zhao, Yong Huang, Zhusheng Liu, Mingcong Wu, and Lili Jiang
A Novel Multi-Objective Programming Model Based
on Transportation Disruption in Supply Chain
with Insurance Contract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1103
Kai Kang, Jing Zhao, and Yanfang Ma
An Interval-Parameter Based Two-Stage Stochastic Programming
for Regional Electric Power Allocation . . . . . . . . . . . . . . . . . . . . . . . . . . . 1111
Jingqi Dai and Xiaoping Li
Pricing Strategies of Closed Loop Supply Chain with Uncertain
Demand Based on Ecological Cognition . . . . . . . . . . . . . . . . . . . . . . . . . . 1122
Dongjing Yu and Chunxiang Guo
Beds Number Prediction Under Centralized Management Mode
of Day Surgery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1136
Jianchao Yang, Luo Li, Hongsheng Ma, and Yong Luo
A Standardization Methodology for Visual Management
in Lean Supply Chain Environments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1147
André Simas and Virgilio Cruz-Machado
Resource Optimization Management

Online Fault Detection in Solar Plants Using a Wireless Radiometer
in Unmanned Aerial Vehicles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1161
Carlos Quiterio Gómez Muñoz, Alfredo Peinado Gonzalo,
Isaac Segovia Ramirez, and Fausto Pedro García Márquez
Demand Response Mechanism of a Hybrid Energy Trading Market
for Residential Consumers with Distributed Generators . . . . . . . . . . . . . 1175
Nastaran Naseri, Babak Rezaee, and Shahin Kazemzadeh
A Fuzzy Multi-Criteria Evaluation Method of Water Resource
Security Based on Pressure-Status-Response Structure . . . . . . . . . . . . . . 1186
Talat Qadeer and Zongmin Li
xx Contents
Management of Technological Modes of System Distributed

Generation Electric Energy on the Basis of Daily Schedules
of Electric Loadings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1198
Abdulla Arifjanov and Romen Zakhidov
Haze-Related Air Pollution and Impact on the Stock Returns
of the Listed Steel Companies in China. . . . . . . . . . . . . . . . . . . . . . . . . . . 1209
Kai Liu, Ying Li, and John Thomas Delaney
Low Carbon-Oriented Coupling Mechanism and Coordination
Model for Energy Industry Investment and Financing of China . . . . . . . 1220
Yanfei Deng, Lei Xu, and Ning Hou
Space-Time Analysis for Water Utilization Efficiency of Urban
Agglomeration and Its Economic Influence Factors łłin Chengdu
Urban Agglomeration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1230
Yeyun Huang and Yunqiang Liu
Exploring Linkages Between Lean and Green Supply Chain
and the Industry 4.0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1242
Susana Duarte and Virgilio Cruz-Machado
A Model of Maker Education in Chinese Universities:
The Perspective of Innovation Ecosystem . . . . . . . . . . . . . . . . . . . . . . . . . 1253
Qinglong Zhan and Mengjia Yang
Pricing and Greening Policy for Manufacturer
with Low Carbon Strategic Customers . . . . . . . . . . . . . . . . . . . . . . . . . . . 1266
Wen Jiang
Evolutionary Game Analysis of the Reservoir Immigrants’
Joint Venture. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1276
Xiaofeng Liu, Tingting Song, and Yanhua Liu
Solid Waste Management in the Republic of Moldova . . . . . . . . . . . . . . . 1283
Gheorghe Duca and Aliona Mereuţa
How to Popularize Green Residential Buildings in China:
A Survey Study from Sichuan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1296
Jun Gang, Dirong Xu, and Ying Wei
Prognostics of Lithium-Ion Batteries Under Uncertainty
Using Multiple Capacity Degradation Information . . . . . . . . . . . . . . . . . . 1307
Fan Li and Yusheng Wang
Analysis of the Effect of Low-Carbon Traffic Behavior
on Public Bicycles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1318
Jing Ma, Xin Liu, Zhineng Hu, and Yong Cheng
Contents xxi
An Empirical Study on the Impact of Niche Overlap of Tourism

Enterprise on Tourist Satisfaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1330
Shaojiang Lin, Jiaying Chen, and Jing Tang
Long Term Scheduling for Optimal Sizing of Renewable Energy
Sources for Hospitals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1342
Shabnam Mahmoudzadeh Vaziri, Babak Rezaee,
and Masoud Amel Monirian
Life Cycle Assessment of Waste Mobile Phone Recycling–A Case
Study in China . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1351
Tingting Liu, Moudi Mahdi, and Liming Yao
A Malmquist Index-Based Dynamic Industrial Green Efficiency
Evaluation in Sichuan Province . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1361
Xuhong Liu and Xiaowen Jie
Risk Control
Machine Learning and Neural Network
for Maintenance Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1377
Alfredo Arcos Jiménez, Carlos Quiterio Gómez Muñoz,
and Fausto Pedro García Márquez
Volatility Spillover Between Foreign Exchange Market
and Stock Market in Bangladesh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1389
Shibli Rubayat and Mohammad Tareq
Cost/Efficiency Assessment of Alternative Maintenance
Management Policies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1395
Diego Ruiz-Hernández and Jesús María Pinar-Pérez
Heijunka Operation Management of Agri-Products Manufacturing
by Yield Improvement and Cropping Policy . . . . . . . . . . . . . . . . . . . . . . . 1407
Ritsuko Aoki and Hiroshi Katayama
Optimizing Reserve Combination with Uncertain Parameters
(Case Study: Football) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1417
Masoud Amel Monirian, Hamideh Razavi,
and Shabnam Mahmoudzadeh Vaziri
A Bayesian-Based Co-Cooperative Particle Swarm Optimization
for Flexible Manufacturing System Under Stochastic Environment . . . . 1428
Lu Sun, Lin Lin, and Haojie Li
Exchange Rate Movements, Political Environment and Chinese
Outward FDI in Countries Along “One Belt One Road”. . . . . . . . . . . . . 1439
Wenjing Zu and Haiyue Liu
xxii Contents
Post-Traumatic Stress Disorder Among Survivors in Hard-Hit Areas

of the Lushan Earthquake: Prevalence and Risk Factors . . . . . . . . . . . . 1457
Zongtang Xie and Hongxia Xiu
How Corruption Affects Economic Growth: Perception of Religious
Powers for Anti-corruption in Iraq . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1466
Marwah Abdulkareem Mahmood, Yizhuang Tian,
and Karrar Abdulelah Azeez
Procurement Risk Mitigation for Rebar Using Commodity Futures . . . . 1476
Jian Ni, Wei Zhou, and Dan Yang
Ubiquitous Healthcare and Ubiquitousness of Chronic Disease
Prevention and Control: Theory and Design . . . . . . . . . . . . . . . . . . . . . . 1490
Zhihan Liu
Can Media Supervision Improve the Quality of the Firms
Internal Control? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1497
Tian Yang, Qianwei Ying, and Xingyao Li
Fuzzy Chance Constrained Twin Support Vector Machine
for Uncertain Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1508
Qilin Cao, Xiaodan Fu, and Yaomin Guo
A Projection Pursuit Combined Method for PPP Risk Evaluation . . . . . 1522
Xinli Zhang, Tianjin Wang, and Sihan Li
Optimal Ownership Pattern to Control Agency Conflict
in Manufacturing Industry of Pakistan . . . . . . . . . . . . . . . . . . . . . . . . . . . 1535
Muhammad Kaleem Khan, He Ying, Umair Akram, Muhammad Hashim,
Xiaoyue Yuan, and Lv Gaoyu
The Impact of Institutional Investors on Stock Price Synchronicity:
Evidence from the Shanghai Stock Market . . . . . . . . . . . . . . . . . . . . . . . . 1548
Li Kun, Lei Yu, and Xiaoxue Hu
Research on Risk Allocation Model in PPP Projects: Perspectives
from the SOEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1559
Chuan Chen, Tengteng Yan, and Ping Chen
An Exploratory Case Study of the Mature Enterprise’s Corporate
Brand Building Based on Strategic Perspective . . . . . . . . . . . . . . . . . . . . 1573
Yuan Yuan, Jing Xu, Liming Zhang, and Rui Zhou
Case Study: Packing and Distribution Logistics Optimization
of Fashion Goods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1583
Young Jae Jang and Shin Woong Sung
Contents xxiii
Integrated Project Management

Algorithmical and Program Functions of Innovation Project
Management in Technoloji Park . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1595
Elchan Ghuseynov, Javanshir Mammadov, and Shafahat Rahim Rahimov
Two-Stage Fuzzy DEA Models with Undesirable Outputs
for Banking System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1604
Xiaoyang Zhou, Rui Luo, Benjamin Lev, and Yan Tu
Increasing Effect in Lodger Number of Hot Spring Hotel According
to the Started Operation of Hokuriku Shinkansen . . . . . . . . . . . . . . . . . . 1616
Takashi Oyabu and Junko Nakamura
Charging Infrastructure Allocation for Wireless Charging
Transportation System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1630
Min Seok Lee and Young Jae Jang
The Research of Incentive Model Based on Principal-Agent
for R&D Personnel in System-Transformed Institutes . . . . . . . . . . . . . . . 1645
Ying Wei and Jun Gang
Research on the Influencing Factors of Public Sense of Security
and Countermeasures in Major Emergencies–An Empirical Analysis
on the Public Sense of Security in View Of the Explosion
in Binhai New Area in Tianjin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1653
Jing Yang and Wei Xu
Effect Assessment of the Free Tissue Flap Colorimetric Card
in the Postoperative Flap Management for Tumors of Oral
and Maxilla-Facial Region . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1667
Xiongtao Yang, Xueying Wang, and Qing Yang
The Impact of Intellectual Capital on High-Tech Enterprise
Performance: Applied Study in China’s Second-Board Market . . . . . . . 1677
Hongchang Mei and Kunling Wang
Site Selection of Public Fast Electric Vehicle Charging Station
by Using an Intuitionistic Fuzzy Projection-Based Approach . . . . . . . . . 1688
Lin Zhong and Zhibin Wu
If the Medical Expenses Do Effect the Rural Residents’ Consume . . . . . 1697
Songtao Jiang and Yuan Ji
A Discrete Time-Cost-Environment Trade-Off Problem
with Multiple Projects: The Jinping-I Hydroelectric
Station Project . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1709
Huan Zheng
xxiv Contents
Research on Equalization of Public Services Based on Changes

of Urban and Rural Population Structure . . . . . . . . . . . . . . . . . . . . . . . . 1722
Qian Fang and Yi Sheng
Economic Cycle, Accounting Conservatism
and Financial Constraints. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1737
Hong Wang, Pan Liang, and Juqiu Deng
Traffic Lights Dynamic Timing Algorithm Based
on Reinforcement Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1752
Chenqing Lu, Feng Wen, and Mitsuo Gen
Research on the Collaborative Governance of Innovation Network
Based on the Extended JM Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1762
Chiyan Zou, Changyi Zhao, Tao Wang, and Xin Gu
Performance Evaluation of Housing Price Regulation Policy in China:
Based on ARIMA Model and Intervention Analysis . . . . . . . . . . . . . . . . 1773
Chang Liu, Yixiao Zhou, Wei Zhao, Qiang Jiang, Xuedong Liang, Hao Li,
Hua Huang, and Shucen Fan
A Descriptive Analysis of the Impact of Air Pollution
on the Mortality of Urban and Rural Residents in Mianyang . . . . . . . . . 1786
Jianping Xu, John Thomas Delaney, Xudong Chen, and Liming Yao
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1797
Advances in Management Science
of the Eleventh ICMSEM
Advancement of Computing Methodology,
Data Analysis, Enterprise Operation
Management and Decision Support System
Based on the Eleventh ICMSEM Proceedings
Jiuping Xu(B)
Uncertainty Decision-Making Laboratory, Sichuan University,

Chengdu 610065, People’s Republic of China
xujiuping@scu.edu.cn
Abstract. Management is playing an increasingly more important role

in modern society. In particular, the development of efficient innovative
managerial tools has significantly influenced social progress in manage-
ment science and engineering management. In this paper, we first give
a brief introduction to the eleventh ICMSEM proceedings Volume I. An
analysis of the current management science topics reveals that comput-
ing methodology, data analysis, enterprise operation management, and
decision support system have become key foci in the past few years.
We then conduct a literature review to examine the main research in
these four areas, after which the central issues in the eleventh ICMSEM
Proceedings Volume I are examined using NodeXL. Finally, by analyz-
ing the main keywords using CiteSpace, frontier management science is
identified. All in all, the ICMSEM continues to provide a valuable forum
for academic exchange and communication and will continue to play an
important role in promoting MSEM advancements in the future.
Keywords: Computing methodology · Data analysis · Enterprise oper-

ation management · Decision support system
1 Introduction
The Eleventh International Conference on Management Science and Engineering
Management (ICMSEM) in Kanazawa, Japan, has given Management Science
and Engineering Management (MSEM) academics the opportunity to present
their innovative findings in this increasingly popular research field. The papers
in this volume demonstrate the substantial growth in interdisciplinary MSEM
methodologies and practical applications. The ICMSEM aims to be the primary
forum for academic researchers and practitioners to become involved in discus-
sions on state-of-the-art MSEM research and development.
MSEM has shown tremendous growth and expansion since its beginnings. In
particular, Management Science (MS) has had a long history associated with uni-
versal and impersonal management research disciplines. Scientific management,
c Springer International Publishing AG 2018
J. Xu et al. (eds.), Proceedings of the Eleventh International Conference
on Management Science and Engineering Management, Lecture Notes
on Multidisciplinary Industrial Engineering, DOI 10.1007/978-3-319-59280-0 1
4 J. Xu
of course, owes its beginnings to the innovative ideas of Taylor and Fayol and
their associates, who at the turn of the century attempted to solve industry prob-
lems through strict time monitoring, technical production methods, incentive
wage payments, and rational factory organization based on efficiently structured
work assignments [1,6]. In the period following the Second World War, modern
analytical management methods brought the business practice into a new era.
Terms such as Decision Theory, Operations Research, System Engineering, and
Industrial Dynamics, which were practically unknown in the early fifties, are now
as well-known as Accounting or Finance [12]. As time passed, with societal, eco-
nomic, organizational, and cultural, factors playing increasingly more important
roles in management practice, MS used mathematics information science, sys-
tems science, cybernetics, statistics and other theories and methods from natural
science to develop innovative management and control systems. The integration
of these research areas has brought significant improvements to the ICMSEM
proceedings Volume I, which is mainly focused on computing methodology, data
analysis, enterprise operations management, and decision support system this
year.
2 Literature Review
Literature reviews give insights into the focus and new research directions in
a particular research field. The widespread research attention given to MS has
been mainly focused on four main areas; computing methodology, data analysis,
enterprise operation management, and decision support system. Therefore, in
this section, we review the pertinent research that had taken place in each of
these four areas.
2.1 Computing Methodology
By providing the algorithms for management science and engineering manage-

ment problems, computing methodology (CM) could be seen to be the theoretical
foundation to MSEM. Through a comparative study, an appraisal of wind speed
distribution prediction using soft computing methodologies has been examined
for energy applications [20]. This soft computing methodologies combination
meant that a more complete array from the information database could be used
to develop tactical planning forecasts. Obrad Anicic et al. built an effective wind
turbine noise level prediction model with the polynomial and radial basis func-
tion being applied as the kernel function for support vector regression (SVR) to
estimate wind turbine noise levels [4]. Gotmare et al. presented a comprehensive
review on the various swarm and evolutionary computing schemes that could
be employed for system identification and digital filter design, and envisioned a
quick reference for a few popular evolutionary algorithms [11]. A hybrid parallel
programming model was designed for which a novel sparse matrix partitioning
algorithm, a performance analysis, and an SpMV optimization on a CPU−GPU
heterogeneous computing platform was developed to solve the problem [25].
Advancement of EINT, CM, IT and DSS 5
Jardim and Silva presented a simple and computationally efficient methodol-

ogy for computing robust dynamic equivalents that could be easily embedded
in time domain simulation packages [14]. Current scholars are not only focus-
ing on computing methodological theoretical research, but also concentrating on
applications to ensure better results.
2.2 Data Analysis

Data analysis (DA) refers to the statistical analysis of large amounts of col-
lected data to extract useful information and form conclusions. West and Post
conducted research to quantify potential soil C sequestration rates for different
crops with the goal of decreasing tillage intensity and enhancing rotation com-
plexity [24]. For environmental management, carbon sequestration rates have
been used in spatial modeling analyses to more accurately predict regional,
national, and global C sequestration potential, and data envelopment analysis
(DEA) methods have been adapted for eco-efficiency. DEA can provide substi-
tution possibilities for different natural resources and emissions as Kuosmanen
illustrated in the development of an application to assess road transportation
in the three largest towns in eastern Finland [17]. To demonstrate the accuracy
and potential of the methodologies, Muxika et al. used historical data, expert
judgement, and multivariate analysis together with a definition of reference con-
ditions to examine changes in the benthic communities in the Basque Country
over the last decade [18]. With the development of information technology, big
data has become a popular research focus because of its potential to give val-
ued insights for enhanced decision-making. For example, big data and big data
analytics methods have been used to assist decision makers better reach their
targets [23]. DA is also playing an increasingly dominant role in modern society
and as a result, data analysis research has yielded many competing models and
contributed both conceptually and empirically to technology and organizational
resource management.
2.3 Enterprise Operation Management

Enterprise operation management (EOM) generally refers to enterprise teams
with the visibility and automation tools to increase returns on assets, improve
management performance, and reduce enterprise costs. In the enterprise net-
work, computing and efficient resource use can provide enterprise-wide informa-
tion sharing, allowing for convenient communication between employees as well
as external information access. Bi et al. investigated the impact of the emerging
internet of things on enterprise systems in modern manufacturing by discussing
the evolution of manufacturing system paradigms [7]. Camarinha-Matos et al.
presented a discussion on the contribution of the collaborative networked orga-
nizations paradigm to the challenges faced by manufacturing systems [8]. It is
undeniable that big data has attracted significant attention from researchers in
the information sciences as well as policy and decision makers in governments
and enterprises. Therefore, there is no doubt that business productivity and
6 J. Xu
technological innovation in the future will draw on the advantages offered by big
data [9,15]. Because of the promise of more automated information exchanges
in networked enterprise scenarios, enterprise information systems (EIS) have
become increasingly important for interoperability to increase productivity and
efficiency [2]. With improvements in enterprise operations management and the
application of advanced management tools, operating efficiency is being contin-
ually enhanced, thereby benefiting enterprise operations.
2.4 Decision Support System
Decision support system (DSS) are computer-based information systems that

support knowledge management, decision making, and management reporting
to assist managers make decisions in highly uncertain and complex environ-
ments. As is known, when making decisions, people weigh up individual opin-
ions and combine them using a thought process to come to a final informed
decision. Critically analyzing the nature and state of the history of decision sup-
port system (DSS), we can see the evolution of research and practice in areas
such as personal DSS, group support systems, negotiation support systems, and
knowledge management-based DSS [5]. Aiello et al. proposed a methodological
approach to the development of a DSS in the specific context of Integrated Pest
Management for intensive greenhouse production and provided an experimental
validation based on real data [3]. A hybrid decision support system consisting of
a three level humanitarian relief chain: a simulator, a rule-based inference engine,
and a knowledge-based system: was designed for a humanitarian relief chain in
which three main performance measures were considered for an explicit trade-off
analysis between cost efficiency and responsiveness [21]. DSS can be applied to
most decision making processes through the research, design and development
of strategies to ensure adoption and integration.
3 The Central Issues in the Proceedings Volume I

Literature mining identifies the most pertinent scientific research in areas of
particular interest [22] and has proven to be a powerful method for revealing
major trends in published scientific literature across the years to allow for topic
maps to be built [10]. NodeXL was designed to facilitate the learning of the
concepts and methods of social network analysis using visualization as a key
component [13] and is a powerful, easy-to-use interactive network visualization
tool that leverages the widely available Microsoft Excel application as a platform
for representing generic graphical data, performing advanced network analyses,
and providing a visual exploration of the networks. NodeXL generally supports
multiple social network data providers who import graphical data (node and
edge lists) into an Excel sheet, and has also been widely used by researchers
[16,19]. Therefore, it plays a vital role in the analysis of the keyword trends in
our research. Based on the most popular research topics, we called for papers
from around the world. This year, 150 papers were accepted which were divided
into two proceedings volumes of 75 papers each. The significance of the keywords
lies not only in the frequency or ratio but also in the key word connections that
demonstrate how these papers revolve around MS concepts and metrics. Infor-
mation visualization methods have also proven valuable in discovering patterns,
trends, clusters, and outliers, even in complex social networks. We have now
received the submissions and sorted the keywords using the software Keywords
Match to perform keyword matching and to identify similarities. Finally, the
processed keywords were put into the NodeXL software, and the results are
shown in Fig. 1.
Fig. 1. Research topics in MS for the Eleventh ICMSEM
From this figure, it is not easy to identify the most important and most heav-
ily researched areas. Further, from calculating and screening keyword degrees
greater than six, the following graphical results were obtained. It can be seen the
central issues of the eleventh ICMSEM proceedings Volume I have been mainly
about computing methodology, data analysis, enterprise operations management
and decision support system, as shown as Fig. 2. In the following, we focus on
some related papers published in the proceedings to highlight the mainstream
research this year.
There was a great deal of interest in the computing methodology section.
Usman Rafi et al. discusses various scheduling terms and scheduling algorithms
and proposes a new approach for scheduling based on a mixture of MQMS, a
priority scheduling mechanism, and round robin scheduling. Lin et al. proposes
a Bayesian network based particle swarm optimization (BNPSO) for solving
a stochastic scheduling problem. Tahyudin and Nambo determine the rules of
numerical association rule mining and also calculate the multi-objective function
using a combination of PSO and the Cauchy distribution. Shen et al. proposes
an AdaBoost algorithmic model based on a back-propagation neural network
8 J. Xu
Fig. 2. Research topics in MS for the eleventh ICMSEM
for highly accurate and efficient credit scoring. Both classical and intelligence
algorithms are used to develop the model solution.
The data analysis section also presents some innovative work. Yuko et al. pro-
vides a method to convert Braille books into machine-readable electronic data.
Another interesting study is that in which Yamamoto researches the number of
visitors in each period and their characteristics based on mobile phone user loca-
tion data collected by a mobile phone company. Zeng et al. develops a severity
crash count based on a hotspot identification method by extending the tradi-
tional empirical Bayes method to a generalized nonlinear model-based mixed
multinomial logit approach. In addition, Hu determines the future population
trends in Sichuan, China through the establishment of a population development
equation. All these papers solve problems by analyzing data and developing use-
ful tools.
For enterprise operations management, Sawada and Yoshida present a novel
approach to increase the attractiveness of Facebook pages, Guo et al. uses a
Helloing model to analyze the influence of pricing and profit on platform enter-
prises, Lu uses IPA methods to help identify the strengths and weaknesses of
tourist satisfaction towards traditional culture, and Khan et al. provides cog-
nitive support for the development of optimized employee motivation levels
through consciousness raising about applied approaches and unrealistic ideas. A
method that measures propensity score matching is proposed by Tang to assess
how cross listing affects corporate performance. In a word, researchers have stud-
ied many aspects of enterprise operations management to improve efficiency and
performance.
In the decision support system section, some practical applications to solve
decision problems are presented. Pakkak proposes an integrated data envelop-
ment analysis (DEA) and analytic hierarchy process (AHP) approach to obtain
attribute weights, Sarwar et al. uses a fuzzy analytical hierarchical process
(AHP) and an extent analysis method to choose an appropriate supplier under
fuzzy environment, Benmessaoud et al. focuses on the monitoring of a wind farm
in real time based on big data collected by the Supervisory Control and Data
Acquisition (SCADA) system, and shows how the maintenance decision-making
can be assured by the SCADA system, and Zhang et al. provides a corresponding
pricing strategy for product crowdfunding, providing inspiration and reference
for enterprises and entrepreneurs when making pricing strategy decisions. Inte-
grating modern technology into traditional methods has become a tendency in
the study of decision support system.
4 Development Trends for MS and ICMSEM

In this section, we use CiteSpace software to understand the effects of scientific
research on future MS research and practice. CiteSpace is citation visualization
analysis software based on the analysis of latent knowledge in scientific analysis.
CiteSpace, proposed by the Chinese scholar Chen from Drexel University based
on Java language, is visual document analysis software that shows the trends
and tendencies within a domain of knowledge over a certain period and allows
for the identification of the evolution occurring at the research frontier; it has
been widely used for knowledge map visualization. The purpose of this work
was to visualize and analyze the management science scientific knowledge map.
In the “Web of Science” database, “management science” was used as the sub-
ject word to search and retrieve articles from 1990 to 2017, from which 62924
articles were identified. To ensure an adequate correlation for the data clean-
ing, TI = (Management Science) was used as the title word retrieval to exclude
other databases and eliminate redundancies, thereby reducing the total to 2231
articles.
4.1 CiteSpace Analysis Results

There are 402 keyword categories and 633 edges in the map for the co-occurring
keywords, in which the nodes represent each individual keyword and the lines
indicate keyword co-occurrence, with the overall network density being 0.0079.
To simplify the display, interactive modification was used on the visualization. By
setting up “threshold = 30”, we obtained a new keyword co-occurrence network
map, as shown in Fig. 3.
However, on the whole, there is a loose structure in the keyword co-occurrence
network as the density is not high, indicating that outside collaborations are
needed to ensure that sufficient attention is paid to the research subject for
international and practical needs. Deeply expanding research around a theme is
vital as a lack of breadth and depth is not conducive to the sound development
of management science.
In these “Co-occurring keywords”, 402 categories appear. To better present
the high frequency themes, the top thirty were analyzed and ranked from top to
bottom in Table 1.
To investigate the use of different keywords in different time windows, the
management keyword frontier research title terms were clustered into 51 cate-
gories based on the time statistics and then labelled in sequence. The first ten
10 J. Xu
Fig. 3. Results of co-occurring keywords
classes are shown in timezone view in Fig. 4. As can be seen, management science
has been in dynamic development at the points where the key word repetition
is relatively high; however, there are no stable phases. There were also some
indirectly related key words scattered in the small nodes around the center (not
shown), which reflected the immaturity in some of the past management science
research and the combinations with other research areas.
4.2 Tendency in the ICMSEM Evaluation

To further analyze the popular management science research fields, the research
areas were analyzed using “Threshold = 200” in CiteSpace, the results of which
are shown in Fig. 5. It can be seen that safety management, uncertainty and
weighted objectives, and improved water management were the most popular
study areas. From the ICMSEM papers, it can be seen that the research theme
directions are similar to the research themes in journal publications. However,
there are some small discrepancies between the ICMSEM proceedings volume I
themes and the journal publication themes, as the proceedings are focused on
innovations and more popular topics.
In addition, Table 2 indicates the highest frequency for the top thirty from
the 145 categories. Of these, the most studied areas have been computer science,
with the research direction having changed from computer science (2000) to
environmental science & ecology (2007) to materials science & multidisciplinary
research (2012) over the years. In recent years, research has focused on extending
Table 1. The first thirty co-occurring keywords categories
Frequency Centrality Year Cited references

67 0.43 2002 System
55 0.31 2002 Model
46 0.04 2000 Information technology
33 0.23 2003 Performance
27 0.25 2006 Management
21 0.08 2006 Perspective
20 0.1 2001 Innovation
16 0.08 2000 Network
15 0.06 2000 Information system
15 0.04 2004 Internet
13 0.04 2006 Integration
12 0.02 2008 Impact
12 0.02 2013 Cloud computing
12 0.03 2014 Support vector regression
11 0.01 2014 Neural network
11 0.09 2007 Algorithm
11 0.06 2006 Framework
10 0.03 2008 SME
9 0 2007 Decision support system
9 0.03 2008 Organization
9 0.02 2006 Strategy
8 0.01 2014 ICT
8 0.05 2008 Prediction
8 0.02 2006 Design
7 0.03 2012 Knowledge
7 0 2006 Industry
7 0.02 2011 Architecture
7 0.04 2015 Anfi
7 0.02 2006 Technology
7 0.02 2007 Competitive advantage
the range of management science research applications in areas such as resource

management, energy management, and education. Though these keywords are
not in the top thirty, they are attracting increasing research attention.
When comparing Fig. 2 with the Tables 1 and 2, some of the papers in this
volume not only reflect popular research areas but also introduce innovations in
other areas. The articles in computing methodology (computer science theory
12 J. Xu
Fig. 4. Results of clustering keywords
Fig. 5. The category clustering of MS timezone view
and methods, models systems, and frameworks), data analysis (statistical tools,
data mining, data processes and big data), enterprise operation management
(technology, information, innovation governance, business and economics) and
decision support system (systems, decision making and strategies) all reflect the
most pressing areas at the moment. Environmental management (ecosystems,
sustainability, water resources) also has appeared as one of the main foci in
the ICMSEM proceedings. When reviewing the previous years’ proceedings, the
Table 2. Results of keywords cluster

6372 0.07 2000 Computer science
273 0.31 2000 Engineering
160 0.01 2000 Computer science, Information systems
159 0.24 2000 Engineering, Electrical & Electronic
127 0.01 2000 Computer science, Theory & Methods
117 0.03 2000 Business & Economics
114 0.33 2000 Computer science, Interdisciplinary applications
87 0.06 2001 Computer science, Artificial intelligence
60 0.09 2002 Telecommunications
51 0.02 2004 Business
49 0.01 2000 Information science & Library science
48 0.04 2001 Computer science, Software engineering
44 0.03 2001 Operations research & Management science
39 0.16 2004 Education & Educational research
32 0.05 2003 Automation & Control systems
30 0.01 2003 Computer science, Hardware & Architecture
29 0.03 2007 Environmental sciences & Ecology
28 0.13 2007 Economics
25 0.06 2000 Engineering, Industrial
22 0.1 2004 Engineering, Multidisciplinary
20 0.11 2006 Engineering, Mechanical
17 0.05 2012 Materials science
16 0 2012 Social sciences - Other topics
16 0.09 2007 Environmental sciences
16 0.04 2000 Engineering, Manufacturing
14 0.06 2012 Materials science, Multidisciplinary
14 0.05 2001 Geosciences, Multidisciplinary
14 0.07 2003 Imaging science & Photographic technology
14 0.09 2012 Social sciences, Interdisciplinary
articles included in the ICMSEM proceedings this year appear to be consistent

with previous predictions. In the future, the present research vitality will hope-
fully result in a new round of research so that the ICMSEM continues to highlight
the most popular research trends.
14 J. Xu
5 Conclusion
In this paper, we briefly introduced the four main areas covered in proceedings
volume I and summarized previous research in these areas. Using the keyword
analysis function in NodeXL, the most prominent topics in these four areas were
identified. We also itemized the main research foci in the ICMSEM proceedings
Volume I to assist readers better understand the content in this year’s papers.
Finally, we analyzed the MS and ICMSEM development trends using the litera-
ture analysis tool CiteSpace, which found that the research foci in the ICMSEM
proceedings volume I are consistent with but slightly different from mainstream
MS research. Further work is needed to identify all areas from MSEM journals
so as to identify the ICMSEM expectations for the future.
Acknowledgements. The author gratefully acknowledges Jingqi Dai and Lin Zhong’s
efforts on the paper collection and classification, Zongmin Li and Lurong Fan’s efforts
on data collation and analysis, and Ning Ma and Yan Wang’s efforts on the chart
drawing.
References
1. Ackoff RL (1962) Scientific method: optimizing applied research decisions
2. Agostinho C, Ducq Y et al (2015) Towards a sustainable interoperability in net-
worked enterprise information systems: trends of knowledge and model-driven tech-
nology. Comput Ind 79:64–76
3. Aiello G, Giovino I et al (2017) A decision support system based on multisensor
data fusion for sustainable greenhouse management. J Cleaner Prod. doi:10.1016/
j.jclepro.2017.02.197
4. Anicic O, Petkovi D, Cvetkovic S (2016) Evaluation of wind turbine noise by soft
computing methodologies: a comparative study. Renew Sustain Energy Rev 56(1–
2):1122–1128
5. Arnott D, Pervan G (2005) A critical analysis of decision support systems research.
J Inf Technol 20(2):67–87
6. Beged-Dov AG (1967) An overview of management science and information sys-
tems. Manage Sci 13(12):817–831
7. Bi Z, Xu LD, Wang C (2014) Internet of things for enterprise systems of modern
manufacturing. IEEE Trans Ind Inf 10(2):1537–1546
8. Camarinha-Matos LM, Afsarmanesh H et al (2009) Collaborative networked
organizations-concepts and practice in manufacturing enterprises. Comput Ind Eng
57(1):46–60
9. Chen CLP, Zhang CY (2014) Data-intensive applications, challenges, techniques
and technologies: a survey on big data. Inf Sci 275(11):314–347
10. De Bruijn B, Martin J (2002) Getting to the (c) ore of knowledge: mining biomed-
ical literature. Int J Med Inf 67(1):7–18
11. Gotmare A, Bhattacharjee SS et al (2016) Swarm and evolutionary computing algo-
rithms for system identification and filter design: a comprehensive review. Swarm
Evol Comput 32:68–84
12. Hall AD (1962) A methodology for systems engineering. Van Nostrand, Princeton
13. Hansen D, Shneiderman B, Smith MA (2010) Analyzing social media networks

with NodeXL: insights from a connected world. Morgan Kaufmann, San Francisco
14. Jardim JL, Silva AMLD (2017) A methodology for computing robust dynamic
equivalents of large power systems. Electr Power Syst Res 143:513–521
15. Kambatla K, Kollias G et al (2014) Trends in big data analytics. J Parallel Distrib
Comput 74(7):2561–2573
16. Kumar S, Jan JM (2012) Discovering knowledge landscapes: an epistemic analysis
of business and management field in Malaysia. Procedia Soc Behav Sci 65:1027–
1032
17. Kuosmanen T, Kortelainen M (2005) Measuring eco-efficiency of production with
data envelopment analysis. J Ind Ecol 9(4):59–72
18. Muxika I, Angel B, Bald J (2007) Using historical data, expert judgement and
multivariate analysis in assessing reference conditions and benthic ecological status,
according to the European water framework directive. Mar Pollut Bull 55(1–6):16–
29
19. Parr CS, Guralnick R et al (2012) Evolutionary informatics: unifying knowledge
about the diversity of life. Trends Ecol Evol 27(2):94–103
20. Petkovi D, Shamshirband S et al (2014) An appraisal of wind speed distribution
prediction by soft computing methodologies: a comparative study. Energy Convers
Manage 48:133–139
21. Sahebjamnia N, Torabi SA, Mansouri SA (2016) A hybrid decision support system
for managing humanitarian relief chains. Decis Support Syst 95:12–26
22. Scherf M, Epple A, Werner T (2005) The next generation of literature analysis:
integration of genomic analysis into text mining. Briefings Bioinformatics 6(3):287–
297
23. Sivarajah U, Kamal MM et al (2017) Critical analysis of big data challenges and
analytical methods. J Bus Res 70:263–286
24. West TO, Post WM (2002) Soil organic carbon sequestration rates by tillage and
crop rotation. Soil Sci Soc Am J 66(6):1930–1946
25. Yang W, Li K, Li K (2017) A hybrid computing method of SPMV on CPUCGPU
heterogeneous computing systems. J Parallel Distrib Comput 104:49–60
Computing Methodology
A Comparison of Pretest, Stein-Type
and Penalty Estimators in Logistic
Regression Model
Orawan Reangsephet1(B) , Supranee Lisawadi1 , and Syed Ejaz Ahmed2

1
Department of Mathematics and Statistics,
Thammasat University, Bangkok, Thailand
num.stat@gmail.com
2
Faculty of Mathematics and Science, Brock University,
St. Catharines, ON, Canada
Abstract. Various estimators are proposed based on the preliminary

test and Stein-type strategies to estimate the parameters in a logistic
regression model when it is priori suspected that some parameters may
be restricted to a subspace. Two different penalty estimators as LASSO
and ridge regression are also considered. A Monte Carlo simulation exper-
iment was conducted for different combinations, and the performance of
each estimator was evaluated in terms of simulated relative efficiency.
The positive-part Stein-type shrinkage estimator is recommended for use
since its performance is robust regardless of the reliability of the sub-
space information. The proposed estimators are applied to a real dataset
to appraise their performance.
Keywords: Monte Carlo simulation · Logistic regression model · Like-

lihood ratio test · Preliminary test estimator · Shrinkage estimator ·
Penalty estimator
1 Introduction
For the past few decades, simultaneous variable selection and estimation of sub-
model parameters has become popular. Many predictors exist to infer an inter-
esting response in the initial model. Some of these predictors may be inactive and
not influential; these should be excluded from the final model that represents
a sparsity pattern in the predictor space to achieve parsimony, flexibility and
reliability. Several researchers, following this information in statistical modeling,
have used either the full model or a candidate submodel.
The logistic regression model also called the logit model, is the most widely
used for an analysis of the independent binary response data in medical, engi-
neering, and other studies. This model assumes that the logit of the response
variable can be modelled by a linear combination of unknown parameters xi β

20 O. Reangsephet et al.
where xi = (xi1 , xi2 , · · · , xip ) is a p × 1 vector of the p predictors for the ith sub-
ject and β = (β1 , β2 , · · · , βp ) is a p × 1 vector of regression parameters. Detailed
information on logistic regression can be found in the books by Hilbe [8] and
Hosmer and Lameshow [10].
In this article, we consider the problem of estimating the logistic regression
model when the response variable may be related to many predictors, some
of which may be inactive. Prior information about inactive predictors may be
incorporated in the full model to produce the candidate submodel.
The pretest (preliminary test) estimation strategy, is inspired by Bancroft,
and the shrinkage estimation strategy, is inspired by Stein, efficiently combine
both full model and submodel estimators in an optimal way to achieve an
improved estimator. Numerous authors have discussed the pretest, shrinkage,
and penalty estimation strategies in many fields including Ahmed and Amezziane
[2], Ahmed and Yüzbaşı [4], Al-Momaniet et al. [5], Gao, Ahmed, and Feng
[6], Hossain, Ahmed, and Doksum [12], and Yüzbaşı and Ahmed [16,17]. For a
logistic regression model, shrinkage estimators and three penalty estimators as
LASSO, adaptive LASSO and SCAD were considered by Hossain and Ahmed
[11] and Lisawadi, Shah, and Ahmed [13] considered the pretest estimation.
As we know, ridge regression (Hoerl and Kennard [9]) has been widely used
when there are many possible predictors to achieve the precision of an esti-
mate. Ahmed et al. [3] found that the ridge regression is highly efficient and
stable when there are many predictors with small effect. Hence, we suggest the
ridge regression for a logistic regression model. In this article, we propose the
pretest and shrinkage estimators in the logistic regression model when it is priori
suspected that parameters may be restricted to a subspace and compares the
resulting estimators to the classical maximum likelihood estimator as well as
the penalty estimators, i.e. LASSO estimator and ridge regression. Monte Carlo
simulation study is carried out using the simulated relative efficient to appraise
the performance of the proposed estimators.
To further illustrate the proposed estimators in the logistic regression model,
we apply the proposed estimator to the South African heart disease data set and
provide a bootstrap approach to compute simulated relative efficiency (SRE) and
simulated relative prediction error (SPE) of the estimators. The detail of this
data set will be described in the Sect. 4. Hossain, Ahmed, and Doksum [12] also
considered this data in the generalized linear model via the pretest estimator,
positive-part Stein-type shrinkage estimator, and three penalty estimators as
LASSO, adaptive LASSO, and SCAD The performance of these estimators are
evaluated in terms of simulated relative efficient (SRE).
Under the prior information about inactive predictors, the full parameter
vector β can be partitioned as β = (β1 , β2 ) where β1 and β2 represent a p1 × 1
active parameter and a p2 × 1 inactive parameter subvector, respectively, such
that p = p1 + p2 . Therefore, our interest lies in the estimation of the active
parameter subvector β1 when the information on β2 is readily available. In other
words, this information about the inactive parameters may be used to estimate
the active parameter subvector β1 when their values are near to some specified
A Comparison of Pretest, Stein-Type and Penalty Estimators 21
value β20 . Without the loss of generality, it is plausible that β2 may be set to a
zero vector, β2 = 0. Keep in mind that the candidate submodel estimator is more
efficient than the full model estimator when the candidate submodel is correct.
On the other hand, the submodel estimator may not be reliable and become
considerably inefficient when the candidate submodel incorrectly represents the
data at hand.
The remainder of this article is organized as follows; the model and the
efficient estimation strategies are proposed in Sect. 2, the results of a Monte Carlo
simulation study are reported in Sect. 3, real data applications are described in
Sect. 4, and finally, discussions and conclusions are presented in Sect. 5.
2 Model and Estimation Strategies
Let y1 , y2 , · · · , yn be independent binary response variables which contain only

two possible outcomes, and xi = (xi1 , xi2 , · · · , xip ) is a p × 1 predictors vector
for the ith subject and i = 1, 2, · · · , n. The simplest idea would be to let zi be a
linear function of the predictors, suppose
zi = xi β, (1)
where β is a p × 1 vector of regression coefficients. Thus, the logistic regression

model assume that

exp (zi ) exp xi β
P (yi = 1|xi ) = π (zi ) = = . (2)
1 + exp (zi ) 1 + exp xi β
The log-likelihood function of the logistic regression model is given by

n

n
l (β) = yi ln π(xi β) + (1 − yi ) ln 1 − π xi β . (3)
i=1 i=1
The derivative of the log-likelihood function with respect to β is obtained by

solving the score equation:
n
∂l (β)
= yi − π(xi β) xi = 0. (4)
∂β i=1
2.1 The Unrestricted and Restricted Maximum Likelihood

Estimator
The unrestricted maximum likelihood estimator (UE) of the parameter vector β

denoted by β̂ U E is obtained by solving the non-linear score Eq. 4, and this can
be solved by using an iterative method like Newton-Raphson.
Under the certain regularity conditions of maximum likelihood estimator
(MLE), Gourieroux and Monfort [7] showed that β̂ U E is a consistent estimator
of β and asymptotically normally distributed with a variance-covariance matrix

(I(β))−1 , where I(β) is the information matrix which is defined as
n

I (β) = π(xi β) 1 − π(xi β) xi xi . (5)
i=1
The restricted maximum likelihood estimator (RE) of β denoted by β̂ RE can

be obtained by maximizing the log-likelihood function (3) under the subspace
restriction β2 − β20 = 0.
2.2 The Linear Shrinkage Estimator

The linear shrinkage estimator (LS) of β denoted by β̂ LS is a linear combination
of the unrestricted and restricted estimator, that is
β̂ LS = λβ̂ RE + (1 − λ) β̂ U E , λ ∈ [0, 1] , (6)
where λ defines the degree of confidence in the given prior information and is a
fixed constant. The linear shrinkage estimator shrinks β̂ U E toward β̂ RE . If λ = 0,
then LS simplifies to an unrestricted estimator, while it simplifies to a restricted
estimator when λ = 1. The performance of the linear shrinkage estimator is
better than the unrestricted and restricted MLE in some part of the parameter
space.
2.3 The Preliminary Test Estimator

The preliminary test estimator or pretest estimator (P T ) of β denoted by β̂ P T
is defined as

β̂ P T = β̂ U E − β̂ U E − β̂ RE I (Ln ≤ Ln,α ) , (7)
where I(.) is an indicator function, and Ln,α is the α-level critical value of the
exact distribution of a suitable test statistic Ln under H0 : β2 = β20 . For testing
H0 : β2 = β20 , the likelihood ratio statistic Ln is suggested:
⎛ ⎞
L β̂ RE
Ln = −2 log ⎝ ⎠ = 2 l β̂ U E − l β̂ RE , (8)
L β̂ U E
where l(β̂ U E ) and l(β̂ RE ) are values of the log-likelihood at the unrestricted and
restricted estimates, respectively. Under H0 , the distribution of Ln converges to
Chi-square distribution with p2 degree of freedom as n → ∞.
Clearly, the pretest estimator takes the value of the unrestricted estimator
when the test statistic lies in a rejection region, otherwise, it takes the value of
the restricted estimator. This estimator has limits due to the large size of the
pretest.
2.4 The Shrinkage Pretest Estimator
The shrinkage pretest estimator (SP) of β denoted by β̂ SP is defined by replacing

the restricted estimator with the linear shrinkage estimator in Eq. (7), that is

β̂ SP = β̂ U E − β̂ U E − β̂ LS I (Ln ≤ Ln,α ) . (9)
An alternative form of the estimator is

β̂ SP = β̂ U E − λ β̂ U E − β̂ RE I (Ln ≤ Ln,α ) . (10)
Ahmed [1] found that the shrinkage pretest estimator significantly improves
upon the pretest estimator in terms of size α, and it dominates the unrestricted
estimator in a large portion of the parameter space. For λ = 1, the pretest
estimators are used to estimate the parameter, while we use a UE as λ = 0.
Generally, the estimators based on the pretest strategy are biased and inefficient
when the null hypothesis does not hold.
2.5 The Stein-Type Shrinkage Estimator

The Stein-type shrinkage estimator which combines the unrestricted and the
restricted estimator in an optimal way to dominate the unrestricted estimator.
The Stein-type shrinkage estimator (S) of β denoted by β̂ S is given as follows

β̂ S = β̂ RE + 1 − (p2 − 2) Ln −1 β̂ U E − β̂ RE , p2 ≥ 3, (11)
alternatively,

β̂ S = β̂ U E − (p2 − 2) Ln −1 β̂ U E − β̂ RE , p2 ≥ 3. (12)
For some insight to this estimator, we refer to Hossain, Ahmed, and Doksum
[12], Yüzbaşı and Ahmed [17] among others. The Stein-Type shrinkage estimator
will provide uniform improvement over the unrestricted estimator. However, the
Stein-type shrinkage estimator tends to over-shrink the unrestricted estimator
towards the restricted estimator when the test statistic Ln is very small in
comparison with p2 − 2. To avoid the over-shrink behavior of this estimator,
the truncated version is suggested which is called the positive-part Stein-type
shrinkage estimator.
2.6 The Positive-Part Stein-Type Shrinkage Estimator

+
The positive-part Stein-type shrinkage estimator (S + ) of β denoted by β̂ S is a
convex combination of the unrestricted and restricted estimator, that is
+ + U E
β̂ S = β̂ RE + 1 − (p2 − 2) Ln−1 β̂ − β̂ RE , p2 ≥ 3, (13)
where a+ = max{0, a}. Alternatively, it can be written in the following conical

form as
+
β̂ S = β̂ RE + 1 − (p2 −2) Ln −1 β̂ U E − β̂ RE I (Ln < (p2 − 2)) , p2 ≥ 3. (14)
The positive-part Stein-type shrinkage estimator is particularly important to

control the over-shrinking inherent in β̂ S .
2.7 The LASSO Estimator

Tibshirani [15] introduced the LASSO estimator of β which minimizes the neg-
ative log-likelihood in Eq. (3) under the L1 constraint. It can be defined as
p

β̂ LASSO
= argminβ −l (β) + γ |βi | , γ ≥ 0, (15)
i=1
where λ is the tuning parameter which controls the amount of a shrinkage. The
LASSO shrinks some coefficients to exactly zero. Therefore, LASSO procedure
performs variable selection and parameter estimation simultaneously.
2.8 The Ridge Regression Estimator

Hoerl and Kennard [9] proposed the ridge regression estimator of β which min-
imizes the negative log-likelihood in Eq. (3) under the L2 constraint. It can be
defined as
p

RIDGE
β̂ = arg minβ −l (β) + γ βi , γ ≥ 0,
2
(16)
i=1
where λ is a tuning parameter which controls the amount of shrinkage. The

ridge regression estimator always keeps all the predictors in the model; thus,
this estimator cannot produce a parsimonious model.
3 Monte-Carlo Simulation Studies

In this section, we carry out a Monte Carlo simulation to compare the perfor-
mance of the pretest, Stein-type and penalty estimators in terms of the quadratic
risk, namely mean squared error (MSE). Our simulations are based on a logis-
tic regression model with the sample size n = 250. A binary response data is
generated from the following model

pi
ln = xi β = β1 xi1 + β2 xi2 + · · · + βp xip , i = 1, 2, · · · , n, (17)
1 − pi
where pi = P (yi = 1|xi ) and the predictor values xi have been drawn from a
standardized multivariate normal distribution.
We consider the hypothesis H0 : β2 = 0. We partition the parameter vector

β as β = (β1 , β2 ) and where β1 and β2 represent a p1 × 1 and a p2 × 1 vector,

respectively, such that p = p1 + p2 . We set the true value of β = (β1 , β2 ) =

(β1 , 0) with β1 = (1.90, −1.05, 0.25, −0.78).
The value of λ is set to 0.25, 0.50, and 0.75. The value of significance level α
is set to 0.01, 0.05, 0.10 and a higher value 0.35.
We now define the parameter Δ∗ representing the distance between the sim-
ulated model and the candidate submodel estimator by
p 2
Δ∗ = β − β (0)
(0)
β − β (0) = βi − βi , (18)
i=1

where β (0) = β1 , 0 and β is the true parameter in the simulated model.
Samples were generated using Δ∗ between 0 and 4.
The number of replications in the simulation was initially varied and it was
determined that N = 1, 000 iterations were adequate to obtain a stable result
for each combination of parameters.
Based on the simulated data, we estimated the MSE of all the proposed
estimators. The performance of the estimators was evaluated using the notion
of simulated relative efficient (SRE), which is the MSE relative to the MSE of
β̂ U E . For any estimator β̂ ∗ , the SRE of β̂ ∗ with respect to β̂ U E is defined as
p
UE 2
UE ∗ SimulatedMSE(β̂ U E ) Simulated i=1 βi − βi
SRE β̂ , β̂ = = p ∗ 2
.
SimulatedMSE(β̂ ∗ ) Simulated i=1 (βi − βi )
(19)
Keep in mind that an SRE is larger than the one that indicates the degree
of superior of the estimator β̂ ∗ over β̂ U E .
3.1 Model with Correct Candidate Submodel (Δ∗ = 0)

First, the case when the candidate submodel is assumed to be correct, is
Δ∗ = 0. Various choices of active and inactive predictors are provided for
(p1 , p2 ) = (4, 3), (4, 5), (4, 7), (4, 10), and (4, 15), and the SRE results are reported
in Tables 1, 2 and 3. The tuning parameter γ of the two penalty estimators is
estimated using 10 fold-cross validation. The findings from Tables 1, 2 and 3 are
summarized as follows:
We note that SREs of all the estimators increase as the number of inactive
predictors 2 is increased for fixed λ and α. Interestingly, the restricted estimator
is the best, and all estimators are superior to the unrestricted estimator for
all configurations except the ridge regression estimator. The linear shrinkage
estimator depends on the choice of λ. Its SRE decreases sharply to 1 as λ → 0
and approaches to SRE of the restricted estimator for higher value of λ. The
SREs of the pretest estimators depend on the size of the test α. For small α,
Table 1. The SREs of the estimators with respect to the UE for λ = 0.25 at Δ∗ = 0.
Estimator Number of in active (p2 )

3 5 7 10 15
RE 1.693 2.104 2.735 3.544 6.120
LS 1.204 1.280 1.356 1.424 1.543
PT α = 0.01 1.631 2.002 2.542 3.162 4.558
α = 0.05 1.472 1.759 2.147 2.513 3.037
α = 0.10 1.369 1.606 1.804 2.108 2.398
α = 0.35 1.128 1.238 1.312 1.368 1.441
SP α = 0.01 1.190 1.263 1.335 1.396 1.489
α = 0.05 1.152 1.220 1.283 1.333 1.393
α = 0.10 1.125 1.187 1.227 1.278 1.325
α = 0.35 1.049 1.088 1.109 1.125 1.147
S 1.158 1.467 1.818 2.300 3.363
S+ 1.204 1.549 1.924 2.428 3.604
LASSO 1.226 1.293 1.470 1.819 2.383
Ridge 0.561 0.668 0.867 1.029 1.332
Table 2. The SREs of the estimators with respect to the UE for λ = 0.50 at Δ∗ = 0

3 5 7 10 15
RE 1.693 2.104 2.735 3.544 6.120
LS 1.416 1.610 1.839 2.069 2.556
PT α = 0.01 1.631 2.002 2.542 3.162 4.558
α = 0.05 1.472 1.759 2.147 2.513 3.037
α = 0.10 1.369 1.606 1.804 2.108 2.398
α = 0.35 1.128 1.238 1.312 1.368 1.441
SP α = 0.01 1.384 1.566 1.773 1.968 2.315
α = 0.05 1.298 1.454 1.622 1.765 1.954
α = 0.10 1.239 1.376 1.473 1.608 1.737
α = 0.35 1.089 1.162 1.206 1.240 1.286
S 1.158 1.467 1.818 2.300 3.363
S+ 1.204 1.549 1.924 2.428 3.604
LASSO 1.226 1.293 1.470 1.819 2.383
Ridge 0.561 0.668 0.867 1.029 1.332
Table 3. The SREs of the estimators with respect to the UE for λ = 0.75 at Δ∗ = 0

3 5 7 10 15
RE 1.693 2.104 2.735 3.544 6.120
LS 1.597 1.927 2.381 2.913 4.354
PT α = 0.01 1.631 2.002 2.542 3.162 4.558
α = 0.05 1.472 1.759 2.147 2.513 3.037
α = 0.10 1.369 1.606 1.804 2.108 2.398
α = 0.35 1.128 1.238 1.312 1.368 1.441
SP α = 0.01 1.546 1.848 2.245 2.669 3.556
α = 0.05 1.413 1.655 1.953 2.227 2.616
α = 0.10 1.326 1.53 1.689 1.925 2.159
α = 0.35 1.116 1.215 1.278 1.326 1.392
S 1.158 1.467 1.818 2.300 3.363
S+ 1.204 1.549 1.924 2.428 3.604
LASSO 1.226 1.293 1.470 1.819 2.383
Ridge 0.561 0.668 0.867 1.029 1.332
the pretest estimator is comparable to that of the restricted estimator. On the

contrary, its SRE decreases as the size increases.
Moreover, the performance of the shrinkage pretest estimator depends on
the choice of λ and α. Its performance becomes poorer with an increase in and
a decrease in λ. Unsurprisingly, the pretest estimator outperforms the linear
shrinkage estimator at Δ∗ = 0. As we would expect, the positive-part Stein-
type shrinkage estimator outperforms the Stein-type shrinkage estimator in any
situation.
Similar results are observed for the LASSO estimator which performs better
than the estimator based on the Stein-type strategy for small p2 . On the other
hand, the estimator based on the Stein-type strategy is preferable when there
are many inactive predictors. However, the ridge regression estimator does not
perform well, but its performance slowly improves as the number of inactive
predictors increases.
3.2 Model with Correct and Incorrect Candidate Submodel

(Δ∗ ≥ 0)
The penalty estimators are not included in the Δ∗ > 0 case because these
estimators do not take advantage of the fact that the regression parameter lies
in the subspace β2 = 0. In this case, the simulation model has an active coefficient
vector β1 = (0.5, −1.5, 0.2) and inactive coefficient vector β2 = (β4 , a) where β4
is a scalar and assumes various values; thus, Δ∗ = (β4 ) when the candidate
2
submodel is used. We choose β4 = 0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.8, 1.0, 1.5, and 2.0
and a is k × 1 zero vector with different dimensions, k = p2 − 1 to be the

number of inactive predictors in the model and we use p2 = 5, 7, 10, and 15 in
the simulations.
The choice of α was fixed to be 0.01, 0.05, 0.10, λ = 0.75 and n = 250. SREs
of the proposed estimators are are reported in Table 4 are graphically represented
in Figs. 1, 2 and 3. The findings are summarized as follows:
The restricted estimator β̂ RE outperforms all the estimators at and near
Δ = 0. In contrast, as Δ∗ becomes larger than zero, the relative efficiency of
∗
β̂ RE as well as the linear shrinkage estimator decrease and become unbounded.

However, the relative efficiency of all the other estimators remains bounded and
approaches to 1. These show that an incorrect candidate submodel is fatal to
the restricted and linear shrinkage estimators.
The pretest estimator is much better than the shrinkage pretest estimator
when the candidate submodel is correct, i.e. Δ∗ = 0. On the contrary, the shrink-
age pretest estimator does well relative to the pretest estimator in a small part
of the parameter space. However, the SRE of both pretest and shrinkage pretest
estimators approaches to one as Δ∗ moves away from zero, but after becoming
inferior to β̂ U E , and later at some point, they join the SRE of one from below.
In addition, they outshine the estimator based on Stein-type strategy where Δ∗
is near zero and for small and moderate p2 . The estimators based on Stein-type
strategy are superior to the unrestricted estimator in the entire range of Δ∗ espe-
cially their gain in risk reduction is impressive as p2 increases. Lastly, we found
that the estimators based on Stein-type strategy perform better than all other
estimators in the wider range of Δ∗ and these estimators are little impacted by
severe departure from the restriction.
4 Real Data Example: South African Heart Disease Data
In this section, we apply the proposed estimators to the South African heart
disease data set. Rousseauw et al. [14] described a retrospective sample of males
in a heart-disease high-risk region of the Western Cape, South Africa. This study
comprised over 462 samples and the set of variables is described in Table 5.
We notice that the condition index (CI) value is calculated as 392.718 which
implies the existence of multicollinearity in this data set. After applying the
variable selection procedure based on AIC criterion, BIC criterion, and LASSO,
the results are given in Table 6.
Table 6 shows that the candidate submodel based on AIC and BIC criteria
contains 5 active predictors, while LASSO selection procedure contains 7 active
predictors. Hence, we will consider the candidate submodel with 5 active pre-
dictors that is tobacco, famhist, ldl, typea, and age. The restricted subspace is
β2 = (βadiposity , βobesity , βalcohol , βsbp ) = (0, 0, 0, 0), p = 9, p1 = 5, and p2 = 4.
To examine the performance of the proposed estimators for the candidate
submodel, we draw m = 250 bootstrap rows with replacement N = 1, 000 times
from the data. The performance of the proposed estimators with respect to the
Table 4. The SREs of RE, LS, PT, SP, S, and S+ with respect to the UE for λ = 0.75
p2 Δ∗ RE LS PT SP S S+
α = 0.01 α = 0.05 α = 0.10 α = 0.01 α = 0.05 α = 0.10
5 0.0 2.581 2.311 2.428 2.057 1.806 2.195 1.909 1.705 1.597 1.726
0.1 1.328 1.611 1.141 1.046 1.016 1.299 1.139 1.082 1.279 1.309
0.2 0.884 1.227 0.836 0.882 0.907 0.976 0.957 0.960 1.187 1.195
0.3 0.666 0.988 0.773 0.873 0.920 0.882 0.928 0.954 1.150 1.152
0.4 0.54 0.833 0.795 0.910 0.958 0.879 0.947 0.975 1.126 1.127
0.5 0.452 0.716 0.863 0.948 0.968 0.917 0.967 0.980 1.111 1.111
0.8 0.309 0.515 0.958 0.988 0.996 0.973 0.992 0.998 1.088 1.088
1.0 0.255 0.434 0.984 1.000 1.000 0.989 1.000 1.000 1.081 1.081
1.5 0.184 0.322 1.000 1.000 1.000 1.000 1.000 1.000 1.074 1.074
2.0 0.152 0.269 1.000 1.000 1.000 1.000 1.000 1.000 1.070 1.070
7 0.0 3.065 2.66 2.873 2.477 2.079 2.523 2.232 1.924 1.959 2.147
0.1 1.701 1.962 1.406 1.215 1.147 1.550 1.286 1.197 1.499 1.540
0.2 1.183 1.548 1.000 0.973 0.975 1.140 1.042 1.021 1.349 1.359
0.3 0.909 1.29 0.876 0.927 0.955 0.985 0.978 0.984 1.282 1.284
0.4 0.748 1.116 0.857 0.939 0.963 0.938 0.971 0.981 1.240 1.241
0.5 0.643 0.993 0.891 0.962 0.979 0.944 0.979 0.989 1.217 1.217
0.8 0.449 0.735 0.969 0.991 0.996 0.981 0.994 0.997 1.169 1.169
1.0 0.376 0.632 0.989 0.997 1.000 0.993 0.998 1.000 1.155 1.155
1.5 0.278 0.483 1.000 1.000 1.000 1.000 1.000 1.000 1.139 1.139
2.0 0.222 0.394 1.000 1.000 1.000 1.000 1.000 1.000 1.129 1.129
10 0.0 4.693 3.702 3.987 3.087 2.396 3.277 2.680 2.175 2.612 2.942
0.1 2.43 2.661 1.982 1.531 1.366 2.124 1.604 1.414 1.944 2.048
0.2 1.669 2.124 1.301 1.135 1.074 1.499 1.230 1.135 1.686 1.730
0.3 1.263 1.755 1.029 0.991 0.983 1.197 1.079 1.037 1.569 1.584
0.4 1.024 1.510 0.911 0.931 0.949 1.048 0.997 0.99 1.485 1.49
0.5 0.868 1.335 0.857 0.926 0.955 0.974 0.974 0.983 1.433 1.434
0.8 0.610 1.001 0.890 0.959 0.980 0.942 0.977 0.989 1.326 1.327
1.0 0.508 0.858 0.928 0.980 0.992 0.959 0.988 0.995 1.300 1.300
1.5 0.368 0.648 0.996 1.000 1.000 0.998 1.000 1.000 1.262 1.262
2.0 0.295 0.529 1.000 1.000 1.000 1.000 1.000 1.000 1.236 1.236
15 0.0 6.502 4.665 4.986 3.510 2.651 3.879 2.976 2.372 3.58 3.961
0.1 3.570 3.566 2.568 1.735 1.462 2.570 1.742 1.469 2.569 2.662
0.2 2.466 2.926 1.575 1.258 1.164 1.703 1.310 1.197 2.163 2.196
0.3 1.906 2.506 1.219 1.083 1.046 1.330 1.132 1.075 1.949 1.956
0.4 1.549 2.188 1.062 1.018 1.011 1.152 1.053 1.029 1.818 1.819
0.5 1.323 1.951 0.993 0.993 0.992 1.057 1.017 1.005 1.712 1.713
0.8 0.927 1.503 0.971 0.992 0.997 0.991 0.997 0.999 1.560 1.560
1.0 0.793 1.331 0.984 0.995 0.998 0.993 0.998 0.999 1.500 1.500
1.5 0.574 1.015 0.999 0.999 1.000 0.999 0.999 1.000 1.423 1.423
2.0 0.458 0.827 1.000 1.000 1.000 1.000 1.000 1.000 1.380 1.380
Fig. 1. SREs of RE, LS, PT, SP, S, and S + with respect to the UE when the candidate
subspace misspecifies β4 as zero as of Δ∗ = (β4 )2 . Here, p1 = 3, α = 0.01.
Fig. 2. SREs of RE, LS, PT, SP, S, and S+ with respect to the UE when the candidate
Fig. 3. SREs of RE, LS, PT, SP, S, and S+ with respect to the UE when the candidate
Table 5. List of variables
Variable Description
Response variable
Chd Coronary heart disease
Predictor variable
Tobacco Cumulative tobacco (kg)
Famhist Family history of heart disease, a factor with levels absent and present
Ldl Low density lipoprotein cholesterol
Typea Type-A behavior
Age Age at onset
Adiposity Adiposity
Obesity Obesity
Alcohol Current alcohol consumption
Sbp Systolic blood pressure
Table 6. Full and candidate sub-models for South African heart disease data
Selection criterion Response Active predictor

Full model Chd tobacco, famhist, ldl, typea, age, adiposity, obesity,
alcohol, sbp
LASSO Chd tobacco, famhist, ldl, typea, age, obesity, sbp
AIC,BIC Chd tobacco, famhist, ldl, typea, age
unrestricted estimator is evaluated by simulated relative efficiency (SRE) and

simulated relative prediction error (SPE) of the estimators which is defined as:

Simulated pi=i βitrue − β̂iU E
SRE β̂ U E , β̂ ∗ = p
Simulated i=1 βitrue − β̂i∗

Simulated m i=1 Yi − π(x i β̂ U E )
and SPE β̂ U E , β̂ ∗ = m , i = 1, 2, · · · , m.
Simulated i=1 Yi − π(x i β̂ ∗ )
Note that SPE is less than one; this means the unrestricted estimator is doing
better. This study assumed the empirical distribution F̂ based on 462 actual
observations to be the true distribution and the resulting logistic regression
coefficient β̂’s to be the true parameter values. We assumed α = 0.01 and λ =
0.50. The results of the point estimates, standard errors, SREs, and SPEs of the
estimators are shown in Table 7.
Table 7 reveals that the restricted estimator is the best, and all the estima-
tors outperform the unrestricted estimator. The performance of linear shrink-
Table 7. Estimates (first row) and standard errors (second row) of the coefficients for
active predictors. The SRE and SPE columns give the relative efficiency and relative
prediction error of the estimators with respect to UE, respectively.
Estimator βtobacco βfamhist βldl βtypea βage SRE SPE

UE 0.084 0.948 0.186 0.042 0.048 1.000 1.000
0.041 0.340 0.087 0.018 0.018
RE 0.083 0.917 0.167 0.038 0.052 2.156 1.364
0.039 0.324 0.077 0.017 0.015
LS 0.085 0.933 0.176 0.040 0.050 1.653 1.360
0.040 0.330 0.079 0.018 0.015
PT 0.084 0.922 0.170 0.039 0.051 1.836 1.230
0.039 0.329 0.079 0.018 0.016
SP 0.085 0.935 0.178 0.040 0.050 1.507 1.274
0.040 0.333 0.081 0.018 0.016
S 0.084 0.938 0.180 0.041 0.049 1.309 1.156
0.040 0.334 0.082 0.018 0.016
S+ 0.084 0.938 0.180 0.041 0.049 1.361 1.191
0.040 0.334 0.082 0.018 0.016
LASSO 0.071 0.788 0.147 0.029 0.042 1.360 1.051
0.038 0.353 0.083 0.020 0.014
Ridge 0.072 0.761 0.146 0.028 0.035 1.552 1.210
0.030 0.311 0.068 0.018 0.015
age, pretest, and shrinkage pretest estimators is dominated by the restricted

estimator as our data in the resampling scheme are generated from an empirical
distribution where the candidate submodel is correct, that is Δ∗ = 0.
Furthermore, the positive-part Stein-type shrinkage estimator outshines the
LASSO estimator when there are moderate or relatively large numbers of the
inactive predictor in the model. In fact, the true parameter values are not exactly
zero, and there is the multicollinearity problem. Unsurprisingly, the ridge regres-
sion estimator performs well. Lastly, the LASSO estimator shows good perfor-
mance in terms of SRE but not in terms of SPE because of the instability
estimation.
5 Discussion and Conclusions
In this article, we compared various estimators based on pretest and Stein-type

strategy and two penalty estimators to the unrestricted and restricted maximum
likelihood estimators in the context of the logistic regression model under the
restriction of parameter. We established the properties of the proposed estima-
tors via Monte Carlo simulation study.
By using Monte Carlo simulation study, we found that the performance of a
restricted maximum likelihood estimator depends heavily on the quality of the
candidate submodel. The restricted maximum likelihood estimator is the best
estimator when the candidate submodel is correct or nearly correct. For any sce-
nario, the estimators based on Stein-type strategy outperform the unrestricted
maximum likelihood in the entire parameter space, especially the truncated ver-
sion of the Stein-type shrinkage estimator. On the contrary, the performance of
the estimators based on the preliminary test procedure lacks this property. The
LASSO estimator is preferable to the Stein-type estimators when the number
of inactive predictors is small. In contrast, the Stein-type estimators dominate
the LASSO estimator only when the number of inactive predictors is relatively
large. The ridge regression cannot produce a parsimonious model, thus the ridge
regression estimator does not perform well when the model is sparse. However,
the positive-part Stein-type shrinkage estimator is robust. It would be interest-
ing, therefore, to investigate the relative performances of adaptive LASSO and
smoothly clipped absolute deviation (SCAD) estimators. We will leave this for
further consideration.
Acknowledgements. We would like to thank all the referees, the Editor, and the
Associate Editor for their valuable suggestions on the revision of the article.
References
1. Ahmed SE (1992) Shrinkage preliminary test estimation in multivariate normal
distributionsm. J Stat Comput Simul 43(3–4):177–195
2. Ahmed SE, Amezziane M (2016) Shrinkage-based semiparametric density estima-
tion. Statistical Methodology
3. Ahmed SE, Raheem E, Hossain S (2011) Absolute Penalty Estimation. Springer

International Encyclopedia of statistical science, New York
4. Ahmed SE, Yüzbaşı B (2016) Big data analytics: integrating penalty strate-
gies. International Journal of Management Science & Engineering Management
11(2016):105–115
5. Al-Momani M, Hussein AA, Ahmed SE (2016) Penalty and related estimation
strategies in the spatial error model: estimation strategies in sem model
6. Gao X, Ahmed SE, Feng Y (2016) Post selection shrinkage estimation for high-
dimensional data analysis. Appl Stoch Models Bus Ind
7. Gourieroux CS, Monfort A (1981) Asymptotic properties of the maximum likeli-
hood estimator in dichotomous logit models. J Econometrics 17(1):83–97
8. Hilbe JM (2009) Logistic Regression Models. Chapman and Hall, Boca Raton
9. Hoerl AE, Kennard RW (2000) Ridge regression: biased estimation for nonorthog-
onal problems. Technometrics 42(1):80–86
10. Hosmer DW, Lemeshow S (2000) Applied Logistic Regression. Wiley
11. Hossain S, Ahmed S (2014) Shrinkage estimation and selection for a logistic regres-
sion model
12. Hossain S, Ahmed SE, Doksum KA (2014) Shrinkage, pretest, and penalty esti-
mators in generalized linear models. Stat Methodol 24(2):52–68
13. Lisawadi S, Shah MKA, Ahmed SE (2016) Model selection and post estimation
based on a pretest for logistic regression models. J Stat Comput Simul 86(17):1–17
14. Rossouw JE, Du PJ, et al (1983) Coronary risk factor screening in three rural
communities. The CORISs baseline study. S Afr Med J 64(12):430–436
15. Tibshirani R (2011) Regression shrinkage and selection via the lasso. J Roy Stat
Soc 58(3):267–288
16. Yüzbaşı B, Ahmed SE (2015) Shrinkage ridge regression estimators in high-
dimensional linear models. Springer, Heidelberg
17. Yüzbaşı B, Ahmed SE (2016) Shrinkage and penalized estimation in semi-
parametric models with multicollinear data. J Stat Comput Simul
Multi-objective Job Shop Rescheduling
with Estimation of Distribution Algorithm
Xinchang Hao1(B) , Lu Sun2 , and Mitsuo Gen3

1
Waseda University, Tokyo 163-8001, Japan
haoxc@live.com
2
Dalian University of Technology, Dalian 116024, People’s Republic of China
3
Fuzzy Logic Systems Institute, Tokyo University of Science, Tokyo 163-8001, Japan
Abstract. To solve the moJSRP model, with the framework of pro-

posed MoEDA, the probability model of the operation sequence is esti-
mated firstly. For sampling the processing time of each operation with
the Monte Carlo methods, allocation method is used to decide the oper-
ation sequence, and then the expected makespan and total tardiness of
each sampling are evaluated. Subsequently, updating mechanism of the
probability models is proposed according to the best solutions to obtain.
Finally, for comparing with some existing algorithms by numerical exper-
iments on the benchmark problems, we demonstrate the proposed effec-
tive estimation of distribution algorithm can obtain an acceptable solu-
tion in the aspects of schedule quality and computational efficiency.
Keywords: Job shop rescheduling · Multi-objective optimization ·

Estimation of distribution algorithms
1 Introduction
In current manufacturing systems, production processes and management are
involved in many unexpected events and new requirements are emerging con-
stantly. For modern flexible decision making in manufacturing systems it is nec-
essary to adapt the unexpected changes with rescheduling processes. As a result,
production managers in the volatile environment have to generate high quality
schedules [3,8]. Rescheduling is defined as periodically or event-driven operation
to better arranging activities for the next time period based on both the state of
the system and the vested schedule. Many literatures have reported researches
on the rescheduling topic. Typically, Vieira et al. presented a comprehensive
survey for most applications of rescheduling in manufacturing systems [8]. This
methodology is summarized in Table 1.
In this paper, job shop rescheduling problem (JSRP) in dynamic environment
will be studied, more specifically, the study takes into account the effect of ran-
dom job arrivals and machine breakdowns. Since JSRP is ubiquitous in manufac-
turing systems, it has attracted great attention to design optimization methods
36 X. Hao et al.
Table 1. Rescheduling framework
Rescheduling Static (finite set of jobs) Deterministic (all information given)

environments Stochastic (some information uncertain)
Dynamic No arrival variability (cyclic production)
(infinite set of jobs) Arrival variability (flow shop)
Process flow variability (job shop)
Rescheduling Dynamic (no schedule) Dispatching rules
strategies Control-theoretic
Predictive-reactive Rescheduling policies Periodic
(generate and update) Event-driven
Hybrid
Rescheduling Schedule generation Nominal schedules
methods Robust schedules
Schedule repair Right-shift rescheduling
Partial rescheduling
Complete regeneration
for finding the desired effective solutions. It is well-known that Job shop schedul-
ing problem is NP-hard, and various methods have been introduced to obtain
optimum solutions. From the practical point of view, many dispatching rules and
composite dispatching rule have been developed for the JSRP [8]. Sabuncuoglu
and Kizilisik proposed several reactive scheduling policies and tested their per-
formances with various experimental conditions, processing time variations, and
random machine breakdown Table 1. Sha and Liu presented extended data min-
ing tools for extracting knowledge of job scheduling with respect to due date
assignment in a dynamic job shop environment [7]. Vinod and Sridharan pre-
sented the salient aspects of a simulation-based experimental study on scheduling
rules in a dynamic job shop where the setup times are sequence dependent [9].
However, a key challenge is to balance available capacities among jobs at differ-
ent processing stages, especially in the shops with re-entrant flow, like those in
semiconductor wafer fabrication plants [5]. Furthermore, in job shop reschedul-
ing, job arrival of specific product is often unexpected, critical job information
is not available in advance which would eliminate schedulers to anticipate future
workloads effectively. However, all these researches transform the multi-objective
problem to a single objective problem using a weighting method. They do not dis-
cuss the implications of non-dominated solutions for the multi-objective dynamic
job shop scheduling problems.
In this paper, to solve the moJSRP model, with the framework of proposed
MoEDA, the probability model of the operation sequence is estimated firstly.
For sampling the processing time of each operation with the Monte Carlo meth-
ods, allocation method is used to decide the operation sequence, and then the
expected makespan and total tardiness of each sampling are evaluated. Subse-
quently, updating mechanism of the probability models is proposed according to
the best solutions to obtain.
Multi-objective Job Shop Rescheduling with Estimation 37
The rest of the paper is organized as follows: in Sect. 2, the mathematical

model of moJSRP is formulated; Sect. 3 describes the proposed EDA-based app-
roach for solving the moJSRP; in Sect. 4, the computational experiments and
analysis of results are presented; finally, Sect. 5 gives the summary and the con-
clusions of this research.
2 Mathematical Formulation
(1) Assumption
The job shop scheduling problem (JSP) concerns with the determination of
the operation sequences on the machines so that the makespan is minimized. It
may consist of several assumptions as follows:
A1. Each machine processes only one job at a time.

A2. Each job is processed on one machine at a time.
A3. Jobs never visit the same machine twice.
A4. For each operation, the processing time and machine assignment are
known in advance.
A5. There are no precedence constraints among operations of different jobs.
A6. Operations cannot be interrupted.
A7. Neither release times nor due dates are specified.
For a new job arrival, the following assumptions are given:
A8. The 1rst operation of the new job has to be started after the processing
operation is finished on one machine.
A9. For new job, the processing time and machine assignment of each oper-
ation are determinate.
A10. There are no precedence constraints between operations of original jobs
and new job.
A11. For new job, neither release times nor due dates are specified.
The JSP has already been confirmed as one of the NP-hard combinatorial
problems. There are N jobs and M machines to be scheduled; furthermore, each
job is composed of a set of operations and the operation order on machines is
pre-specified. Each operation is characterized by the required machine and the
fixed processing time [3].
(2) Notation
Indices:
i, i , k: job index; where i is original job index, i is any job index, k is the
index of new arrival job;
j, j , l: operation index; where j is operation index of original job, j is any
operation index of any job, and l is the operation index of new arrival job.
38 X. Hao et al.
Parameters:
J: number of original jobs;

N : number of new jobs;
M : number of machines;
Ni : number of operations of original job i;
Nk : number of operations of new job k;
oij : operation j of original job i;
okl : operation l of new job k;
oi j : operation j of any job i ;
mij : machine on which operation is to be assigned;
pij : processing time of operation oij ;
pkl : processing time of operation okl ;
ti j: begin time of original operation oij ;
tb : predefined time at which schedule is renewed, i.e., rescheduling point’.
Decision Variables:
tij : begin time of operation oij in the original schedule;

tkl : begin time of operation okl in the new schedule.
The first objective Eq. (1) of the moJSRP is to minimize the makespan of new
job. The second objective Eq. (2) is to minimize the costs caused by disruption
on original schedule, where the term Δtij and Δt̄ denote the deviation and its
mean of begin time respectively in the original schedule.
min tM S = max(tkNk pkNk ) (1)

k
1
min σ =
2
(Δtij − Δt̄) (2)
i Ni
i j

where Δtij = tij − tij (3)
1
Δt̄ = |Δtij | (4)
i Ni
i j

s. t. tt,l−1 + pk,i−1 ≤ tkl , ∀k, l > 1 (5)
(ti j − (tij + pij )) ∧ (tij − (ti j + pi j )) + ψ |mij − mi j | ≥ 0, ∀oij , oi j
(6)
tij = tij , tij ≤ tb (7)
tij > 0, ∀i, j (8)
tkl > 0, ∀k, l, (9)
where, inequity (5) ensures that the (l − 1)th operation of job k should be
processed before the l − th operation in the same job. Equation (6) expresses the
operation precedence constraints. Equation (7) ensures that the original opera-
tions that are fixed for execution at the rescheduling point cannot be changed.
Inequities (8) and (9) represent the nonnegative restrictions.
3 Effective Multiobjectives Estimation of Distribution

Algorithm
3.1 A Descriptive Analysis of Air Pollution and Death Data
The multiple objective optimization problems have been receiving growing inter-
est from researchers with various backgrounds since early 1960. There are a
number of scholars who have made significant contributions to the problem. In
the paper, we present a scheme of hybrid multiobjective estimation of distrib-
ution algorithm (h-MoEDA). In h-MoEDA, MoEDA takes care of exploration
that tries to identify the most promising search space regions, and modeling the
distribution of the stochastic variables. For NP-hard combinatorial optimization
problems, it became necessary that the global search dynamics of EDAs ought to
be complemented with local search refinement [6]. Variable neighborhood search
(VNS) is proposed to improve the solution quality, and the simple principle for
driving this improvement is based on the systematic change of neighborhood
within a possibly randomized local search.
In Fig. 1, h-MoEDA starts by generating the solutions kept in the population
randomly. The set of solutions of high fitness, which refers to promising solu-
tions, is selected from the population using a selection method, and the promis-
ing solutions are used to learn the probability model. Thereafter, probabilities
defined by the probability model are estimated, and the new candidate solutions
Fig. 1. Pseudo-code for h-MoEDA

40 X. Hao et al.
are sampled according to the given sampling method. For a candidate solution,
problem specific local search algorithm is used to improve each candidate solu-
tion to reach a local optimum. Finally, the new solutions are evaluated, and
these with high fitness are incorporated into a solution pool, which keeps these
individuals contributing to the makeup of promising solutions. The iteration will
continue until the predefined termination criteria are met.
The following subparagraph presents the vital components of the proposed h-
MoEDA in detail. Firstly, the representation of a chromosome for an individual
is described. Then, transition probability model is presented. Finally, fitness
assignment mechanisms are discussed.
3.2 Solution Representation

The representation of the operation sequence uses job-based encoding [1], and the
length of the chromosome equals the total number of operations. The job number
denotes the operation of each job, and the lth occurrence of a job number refers
to the lth operation in the sequence of this job. Figure 1 presents an example
considering a rescheduling with a 4 jobs and 3 machines.
Fig. 2. Illustration of the representation of a solution for moJSRP
Giffler-Thompson based heuristic algorithm [4] based decoding procedure is

applied to generate a feasible schedule. This produces a timetable with the start
and end of the processing period, and the makespan time. The Gantt chart of
this solution is illustrated in Fig. 3.
Fig. 3. Gantt chart of the schedule encoded by the operation sequence shown in Fig. 2
An example of reschedule repaired priority-based chromosome are shown in

Fig. 4.
3.3 Transition Probability

For traditional EAs, the representation of a chromosome for an individual is gen-
erated by mapping the decision space into the search space or directly encod-
ing the decision variables. Each position in a probability vector indicates the
Fig. 4. An illustration of reschedule generated by repaired priority-based representation
distribution of probability regarding each variable. When prior knowledge of

distribution is not assumed, the domain of discrete variable X is a set of pre-
defined values (x). The distribution of random variable X has the same equal
probability; the initialization is as following:
pt=0 (X) = 1/ |X| , (10)
where |X| denotes the number of values in the set of domain X.

After the initialization of EDA, probability sample for new alternative solu-
tions and the new solutions are evaluated according to a specific system objec-
tive. EDA collects all new alternative solutions and replaces the inferior solutions
in the promising data. The probability distribution of X can be estimated as
follows:
B t (X = x) = (N (X = x) + 1/ |X|) /prSize + 1/ |X| , (11)
where N (X = x) denotes the number of instances in promising solutions with

variable X = x, and represents the low bound to the probability of X.
The distribution probability of X in the probability vector is learned toward
the estimated distribution of promising data, as follows:
Pt+1 (X = x) = (1 − α) Pt (X = x) + αBt (X = x) , (12)
where α denotes the learning rate from the current promising solutions; in par-
ticular, for α = 1, the probability distribution is completely reconstructed by
the current promising solutions.
To maintain the diversity of sampling, the distribution probability of X is
updated toward the estimation distribution. The distribution can be tuned with
probability pm of the mutation, and the mutation is performed using the follow-
ing definition:
Pt (X = x) + λm
Pt+1 (X = x) = ,
λm

x ∈X\{x} max P t (x) − |X|−1 , ε + (P t (X = x) + λ m )
(13)
where λm is the mutation shift that controls the amount for mutation operation,
and ε is a small probability value to avoid the negative probability value.
42 X. Hao et al.
3.4 Fitness Assignment Mechanisms
The multiple objective optimization problems have been receiving growing inter-
est from researchers with various backgrounds since early 1960. There are a num-
ber of scholars who have made significant contributions to the problem. Among
them, Pareto is perhaps one of the most recognized pioneers in this field and
recently. Many multi-objective genetic algorithms differ mainly in the fitness
assignment strategy which is known as an important issue in solving multiple
objectives optimization problem [1]. Zhang et al. proposed a hybrid sampling
strategy-based evolutionary algorithm (HSS-EA) in which a Pareto dominating
and dominated relationship-based fitness function (PDDR-FF) is used to evalu-
ate the individuals. PDDR-FF of an individual Si is calculated by the following
fitness assignment function (14):
eval (Si ) = q (Si ) + 1/ (p (Si ) + 1) i = 1, 2, · · · , popSize, (14)
where q(·)is the number of individuals which can dominate the individual S · p(·)
is the number of individuals which can be dominated by the individual Si . The
PDDR-FF can set the obvious difference values between the nondominated and
dominated individuals as shown in Fig. 5.
Fig. 5. The description of HSS-EA
4 Experiments and Discussion
To examine the practical viability and efficiency of the proposed MoEDA, we
designed a numerical study to compare MoEDA with efficient algorithms from
previous studies. The proposed MoEDA was compared with adaptive weight
genetic algorithm (awGA) [1], Non-dominated Sorting Genetic Algorithm II
(NSGA-II) [2] and Strength Pareto Evolutionary Algorithm 2 (SPEA2) [10] for a
set of simulation data of testing standard benchmark problems. All of the above
algorithms were implemented using JAVA under the Eclipse environment, and
the simulation experiments were conducted on Intel Core i5 (2.3 GHz clock)
with 4G memory. Data were collated from 30 test runs for each algorithm. In
order to compare the performance of these algorithms fairly and under the same
environment, the strategies of related algorithms and their respective parameters
are presented in Table 2.
Table 2. The parameters of NSGA-II, SPEA2, AWGA and h-MoEDA
NSGA-II, SPEA2 and awGA MoEDA

Interaction 1000 1000
Population 200 200
Selection Roulette Tournament (k)
Operators Crossover (Pc ) Sampling
Mutation (Pm ) Improvement (Pi )
Parameters Pm = 0.20 PromisingRate = 0.7
Pc = 0.80, w2 = 0.6 α = 0.02, β = 0.02
w1 = 0.4, w = 0.6 Pm = 0.4, k = 2
In order to evaluate the performance of a given algorithm for MoJSRP, the

following two performance measures are considered in this paper. Let Sj be a
solution set for each solution method (j = 1, 2, 3, 4). S∗ is a known set of the
Pareto-optimal set.
Coverage C(S1 , S2 ) is the percent of the individuals in S2 who are weakly
dominated by S1 . The value C(S1 , S2 ) = 1 means that all individuals in S2 are
weakly dominated by S1 .
On the contrary, C(S1 , S2 ) = 0 denotes that none of the individuals in S2 is
weakly dominated by S1 [10]. The larger C(S1 , S2 ) is; the better S1 outperforms
S2 in C.
4.1 Convergence
Table 4 shows a comparison of coverage by MoEDA, awGA, NSGA-II and
SPEA2. Figure 6 presents the distribution range of coverage conducted on LA35.
From Table 3 and Fig. 6, it is easy to see that the h-MoEDA is better than
MSGA-II, SPEA2 and awWA on C measure. Such better convergence should
mainly attribute to the hybrid sampling strategy of VEGA’s preference for the
edge region of the Pareto front and PDDR-FF’s tendency converging toward the
centre area of the Pareto front.
They preserve better performances both in efficacy and efficiency. Especially,
h-MoEDA can also keep diversity evenly without special distribution mecha-
nisms like NSGA-II and SPEA2. h-MoEDA is better than MSGA-II, SPEA2
44 X. Hao et al.
Table 3. Comparison of coverage measure by h-MoEDA, AWGA, NSGA-II, SPEA2

and AWGA
Problem Algorithm Mean Improved SD

C(h-MoEDA,A) C(A, (C(h-MoEDA,A)- C(h-MoEDA,A) C(A,
h-MoEDA) C(A,h-MoEDA))(%) h-MoEDA)
LA29 NSGA-II 0.4669 0.3524 11.45 0.0842 0.0709
SPEA2 0.4965 0.3224 17.41 0.0696 0.0603
awGA 0.5731 0.2541 31.90 0.0451 0.0553
LA35 NSGA-II 0.6846 0.2824 36.63 0.0777 0.0794
SPEA2 0.6606 0.2618 39.89 0.0860 0.0631
awGA 0.6844 0.2127 46.73 0.0672 0.0460
LA38 NSGA-II 0.4543 0.3425 11.19 0.0779 0.0754
SPEA2 0.5063 0.3412 16.52 0.0696 0.0603
awGA 0.5625 0.2767 28.58 0.0572 0.0619
and awWA on C measure. Such better convergence should mainly attribute to

the hybrid sampling strategy of VEGA’s preference for the edge region of the
Pareto front and PDDR-FF’s tendency converging toward the center area of the
Pareto front. They preserve better performances both in efficacy and efficiency.
Especially, h-MoEDA can also keep diversity evenly without special distribution
mechanisms like NSGA-II and SPEA2.
Fig. 6. Box plot of coverage measure conducted on LA35
4.2 Distribution Performance

The experiment results of dispersion performances, SP are shown in Table 4 and
Fig. 7 presents the distribution range of space conducted on LA35. It indicates
that MoEDA is obviously better than awGA, while better than NSGA-II and
SPEA2. Without special mechanism to preserve the diversity evenly, h-MoEDA
can also achieve satisfactory dispersion performance.
4.3 Computation Costs

The computation costs of evolutionary-based algorithms mainly depend on the
number of fitness evaluations. We need to compare h-MoEDA with other meth-
ods, and the termination criterion is to reach the maximum number of iterations
Table 4. Comparison of the spacing measure by h-MoEDA, NSGA-II, and SPEA2 and
AWGA
Problem Algorithm Mean Improved Sd

SP SP(A) (SP(h-MoEDA) SP(h-MoEDA) SP(A)
(h-MoEDA) SP(A))/SP(A)(%)
LA29 NSGA-II 16.3568 20.1712 −18.91 4.4081 6.4981
SPEA2 16.3568 22.2868 −26.62 4.4081 6.7942
awGA 16.3568 25.8421 −36.70 4.4081 7.2521
LA35 NSGA-II 19.0001 24.8423 −20.31 4.9569 6.2297
SPEA2 19.0001 24.9510 −23.85 4.9569 7.0526
awGA 19.0001 33.2410 −42.84 4.9569 7.1686
LA38 NSGA-II 17.1736 21.4569 −19.96 5.2566 6.9774
SPEA2 17.1736 24.0161 −28.49 5.2566 7.1136
awGA 17.1736 27.0155 −36.43 5.2566 7.9576
Fig. 7. Box plot of space measure conducted on LA35
(generations). The mean computation cost of h-MoEDA is 217.78 s, NSGA-II

and SPEA2 are 245.54, 390.24 s respectively. Compared with SPEA2, SPEA2
needs to calculate all the distances of each individual to all the other individuals
first. Then SPEA2 gets the k-th nearest distance for each individual after sorting
its distances and the time complexity of SPEA2 is O(mN 2 log N ) where m is the
number of objectives. It is clear that the NSGA-II achieved better computation
costs. Without special dispersion preservation mechanism, the time complexity
of fitness calculation (PDDR-FF) in h-MoEDA is O(0). It is slightly better than
NSGA-II while h-MoEDA needs to estimate the parameters of probability model
using the promising solutions.
5 Conclusion
This paper presents an effective h-MoEDA, which solves the MoJSRP. It min-
imized the expected average makespan and expected total tardiness within a
46 X. Hao et al.
reasonable amount of calculation time. With the framework of the proposed

EDA, the explicit probability model of the operation sequence is estimated on
the distribution of good solutions found so far and use the constructed model to
guide further search behavior. The sampling operator based on the probability
model achieves better convergence and stability than the conventional operator
such crossover and mutation. In our future work, further experiments will be
conducted to determine accuracy of the proposed EDA in response to variations
among the parameters.
Acknowledgement. This work is partly supported by the National Natural Science

Foundation of China under Grant 61572100, and in part by the Grant-in-Aid for Sci-
entific Research (C) of Japan Society of Promotion of Science (JSPS) No. 15K00357.
References
1. Cauty R (2000) Genetic algorithms and engineering optimization. Wiley, New York
2. Deb K, Pratap A, Agarwal S et al (2002) A fast and elitist multiobjective genetic
algorithm: NSGA-II. IEEE Trans Evol Comput 6(2):182–197
3. Gen M, Cheng R, Lin L (2008) Network models and optimization: multiobjective
genetic algorithm approach. Springer, London
4. Giffler B, Thompson GL (1960) Algorithms for solving production-scheduling prob-
lems. Oper Res 8(4):487–503
5. Kempf K (1994) Intelligently scheduling semiconductor wafer fabrication. In: Intel-
ligent Scheduling, pp 517–544
6. Michiels W, Aarts E, Korst J (2007) Theoretical aspects of local search. Springer,
Berlin
7. Sha DY, Liu CH (2005) Using data mining for due date assignment in a dynamic
job shop environment. Int J Adv Manufact Technol 25(11):1164–1174
8. Vieira GE, Herrmann JW, Lin E (2003) Rescheduling manufacturing systems: a
framework of strategies, policies, and methods. J Sched 6(1):39–62
9. Vinod V, Sridharan R (2008) Scheduling a dynamic job shop production system
with sequence-dependent setups: an experimental study. Rob Comput-Integr Man-
ufact 24(3):435–449
10. Zitzler E, Laumanns M, Thiele L (2001) SPEA 2: improving the strength pareto
evolutionary algorithm. In: Evolutionary methods for design, optimization and
control. CIMNE, Barcelona
Multi-Queue Priority Based Algorithm for CPU
Process Scheduling
Usman Rafi1 , Muhammad Azam Zia2,3(B) , Abdul Razzaq4 , Sajid Ali5 ,

and Muhammad Asim Saleem6
1
Department of CS and IT, UOS Lyallpur Campus, Faisalabad, Pakistan
2
State Key Laboratory of Networking and Switching Technology, BUPT, Beijing,
The People’s Republic of China
3
University of Agriculture, Faisalabad, Pakistan
zia uaf@hotmail.com
4
Department, NFC Institute of Engineering and Technology, Multan, Pakistan
5
Department, D.G. Khan Campus, University of Education,
Lahore, Pakistan
6
Department of IT, Government College University, Faisalabad, Pakistan
Abstract. Operating Systems acts as a base software and acts as a driver

for both application programs and system programs. All the programs
residing in an operating system has to become process for execution. A
modern computer system supports multitasking by single user or multi-
ple users. Different processes have different priorities. The major goal of
an operating system is to reduce waiting time and enhance throughput by
scheduling processes in some way. This paper discusses various scheduling
terms and scheduling algorithms. We have proposed a new approach for
scheduling. This proposed algorithm is based on the mixture of MQMS,
Priority Scheduling mechanism and Round Robin scheduling. The pro-
posed algorithm facilitates operating system by managing separate queue
for separate priority of process and manages queue scheduling in round
robin fashion with dynamic time slicing. Processes are added to appropri-
ate queue and this decision is based on any user defined or system defined
criteria. We have also discussed various case studies regarding this algo-
rithm and compared its results with priority scheduling algorithm. These
case studies are limited to two queuing system up till now. We have also
proposed multiple queue management (more than 2), dynamic time slic-
ing instead of half execution scheme and varying execution times of queues
as future work of this algorithm scheme.
Keywords: Priority scheduling · Multi-queue scheduling · Dynamic

time slice execution · Fair priority scheduling · Single processor multi-
queue scheduling · Improved priority scheduling
1 Introduction
A computer system comprises of software and hardware. Software includes pro-
grams for system and user needs. Hardware is a set of resources including CPU,
48 U. Rafi et al.
GPU, RAM and many more of them. The most important software is Operating
System that acts as an interfacing layer which serves the users in using hardware
effectively. User needs are fulfilled by set of Application Software that requires a
base software to run i.e. operating system. Software may be in running state or
it may not be running. The former is called a Process while the latter is called
Program. A process requires certain resources for execution. CPU and RAM are
major resources that each process requires for execution. CPU is the most costly
and important resource. It must be utilized properly. CPU should never be left
underutilized. Operating System must ensure proper utilization of the CPU.
A computer system may be Single Processor System or it may be Multi-
Processor System. Single processor system allows only one process to acquire
the CPU and gets execute on it. On the other hand multi-processor system
allows more than one processes to be executed with one on each processor at a
time.
A process changes its activity from time to time as it executes. Process activ-
ity is defined in terms of two-state model or five-state model. Two-state model
defines that a process can have two states: running and not running. It gives an
abstract idea. Two state model is shown in Fig. 1. Stallings [21] described a more
elaborative version of two state model called five-state model. Five state model
took into account the reason of process for not being in the running state. So,
five-state model described five process activities: new, ready, running, waiting
and terminated. Figure 2 shows five state model for process execution states.
Fig. 1. Two state model for process states
Fig. 2. Five state model for process states
Section 2 of the paper describes CPU Scheduling, its need, various queue man-
agement systems, queues involved in scheduling and scheduling criteria. Section 3
Multi-Queue Priority Based Algorithm for CPU Process Scheduling 49
is organized to give a brief idea of various scheduling algorithms. Section 4

describes some of related work. Sections 5 and 6 describes our proposed algo-
rithm, its working logic and pseudo code. Section 7 comprises of various case
studies and results of our proposed algorithm. The results of our proposed algo-
rithm are compared with the priority scheduling algorithm in Sect. 8.
2 CPU Scheduling
Operating system may be designed to be Single Programmed Systems or Multi

Programmed Systems.
A single programmed system can only accommodate single process in memory
for execution. CPU has to remain idle in the case of process waiting for some
I/O operation. This results in underutilization and causes increase in waiting
time and turnaround time.
Silberchatz et al. [18] narrated that a multi-programmed system facilitated
the efficient utilization of processor by allowing multiple processes or jobs
ready/waiting (residing in memory) for execution and switching the CPU among
these processes. If an executing process had to wait for some I/O operation, the
CPU no longer remained idle. It switched to the next job waiting for CPU alloca-
tion to get executed. In this manner proper utilization of CPU was achieved. To
achieve this utilization, scheduling was required. Scheduling scheme comprised
of criteria and algorithm to implement this criteria. CPU scheduling was of
immense importance in multi-programmed systems because CPU was expansive
and scarce resource.
2.1 Ready Queue Management
The ready queue is a bulk of processes ready to get executed. Ready queue
may be partitioned as it is done in Multilevel Queue Scheduling and Multilevel
Feedback Scheduling. There are following two ways to manage ready queue:
(1) SQMS
SQMS (Single Queue Management System) involves no partitioning of ready
queue. SQMS consists of single partition ready queue in which all the incoming
processes are stored. It is a straight forward queue management scheme. It is due
to the fact that processes are to be picked for execution from a single queue. So
operating system requires only process scheduling algorithm. Only a single queue
has to be managed and scheduled. No overhead is involved in terms of scheduling
ready queue partitions or multiple ready queues. In addition to this, operating
system doesn’t require any overhead for deciding the addition of processes to
un-partitioned ready queue.
50 U. Rafi et al.
(2) MQMS
MQMS (Multi-Queue Management System) scheme maintains multiple ready
queues of jobs or partitions of ready queue. Processes are added to appropriate
ready queue portion depending upon certain criteria. MQMS scheme is complex
scheme in terms of implementation, management and scheduling. The selection
of process for execution requires two important decisions: (i) appropriate queue
selection and (ii) appropriate process selection from the selected queue. Hence
MQMS involves both queue scheduling algorithm and process scheduling algo-
rithm. MQMS scheme involves overhead in terms of queue selection.
A process is first added to job queue. After this, it is moved to ready queue.
Process can now move itself among ready queue, running state and waiting queue
as and when required. All these state changes depend upon a criteria on the basis
of which the Operating System decides that which process move to ready queue,
which started execution next, which process will be needed to move to waiting
queue and which process pre-empted and moved to ready queue. These decisions
are taken by intelligent algorithms, executed by CPU Scheduler.
3 Scheduling Algorithms
Tanenbaum [22] described that First Come First Served (FCFS), Shortest Job
First (SJF), Shortest Remaining Time First (SRTF), Round Robin (RR) and
Priority Scheduling (PS) algorithms were commonly used Scheduling Algo-
rithms. These algorithms were implemented in the form of SQMS. However,
priority scheduling may had MQMS version. Mishra and Khan [11] narrated
that SJF only permitted process with lowest burst time to acquire CPU for exe-
cution and lowered average waiting time. RR scheduling defined a time quantum
for assigning CPU fairly to each process present in the ready queue for specific
amount of time. We discuss Shortest Job First (SJF), Shortest Remaining Time
First (SRTF), Round Robin (RR) and Priority Scheduling algorithms below.
3.1 Shortest Job First (SJF)
Shortest Job First is based on the Burst Time of processes. SJF is based on
a simple idea derived from FCFS, that is, executing smaller processes before
larger ones lowers the Average Waiting Time (AWT) of process set. If a sub set
of processes have same burst time, then they are executed in FCFS fashion. The
processes are arranged in a queue on the base of their Burst Times in a way that
process with smallest burst time is placed at the Head/Front of queue while the
process with largest burst time is placed at Rear of queue. In this way, CPU
Scheduler always picks the process with smallest burst time from ready queue.
It’s a non-preemptive scheduling algorithm. The main advantage of SJF is to
decrease the Average Waiting Time of process set. However, this algorithm does
not take into account process priorities. SJF may result in starvation of larger
processes.
3.2 Shortest Remaining Time First (SRTF)

Shortest Remaining Time First scheduling algorithm is simply preemptive ver-
sion of SJF. In SRTF, the main idea is simply the same, that is, execute small
process before large one. But in this algorithm, if a process is under execution
and during course of its execution a new process enters the system. Then the
burst time of newly entered process is compared with remaining burst time of
currently executing process. If the comparison results in the fact that enter-
ing process has smaller burst time, then CPU is preempted and allocated to the
newly entering process. However, if the new entering process has same burst time
as that of already executing process, then FCFS manner is used for their schedul-
ing. The major advantage is the exploitation of phenomenon to lower Average
Waiting Time of process set. The demerit of SRTF is that, the processes in real
environment have priorities and SRTF does not support priorities. Moreover, it
is not necessary that smaller processes are important while larger ones are less
important. This algorithm also includes high overheads for queue managements,
queue rearrangements, burst time comparisons and context switches. It may also
give rise to starvation for large processes.
3.3 Round Robin (RR)

Round Robin scheduling algorithm is same as that of FCFS with certain innova-
tion. This innovation includes preemption and Time Quantum. Time Quantum
is a small amount of time. Processes are arranged in a circular queue. CPU is
allocated a process from the queue; CPU executes the allocated process for a
time period equal to that of time quantum. When the time quantum expires, the
CPU is preempted and next process from the queue is selected. Next process is
also executed for a time length equal to time quantum. This algorithm continues
to schedule processes in this fashion until queue is empty. The major advantage
of this scheme is that each of the process gets fair share and equal opportunity
of execution. The major drawback is that, if time quantum is very small, then
too many context switches are involved. On the other hand, if time quantum is
large, then RR may perform same as FCFS. Moreover, RR also does not consider
priorities of processes. Siregar [23] narrated that starvation could be avoided if
CPU was allocated fairly to processes. Time quantum was an important factor
in RR scheduling scheme for lowering waiting time of process set. Time quantum
could take any value. But best time quantum should be discovered for minimizing
average waiting time.
3.4 Priority Scheduling (PS)

Priority Scheduling s involves assigning a Priority to each process. A Prior-
ity is either indicated by an integer number (0, 1, 2, 3, etc.) or a string (low,
medium, high etc.). In the case of numeric priority, two choices exist: higher
number may indicate a higher priority (our proposed algorithm consideration)
or lower priority (other algorithm consideration). This scheduling algorithm is
52 U. Rafi et al.
implemented using a Priority Queue. Processes are arranged in decreasing order

of priority (Front to Rear) with process having highest priority placed at front
while process having second highest priority placed at a place next to front
and so on. The CPU Scheduler picks process with highest priority and allocates
CPU for execution. If a number of processes have same priority then they are
scheduled in FCFS fashion as of in SJF and RR case. PS may be preemptive
or non-preemptive. If non-preemptive version is being used, then no matter how
much higher priority process enters the queue, if CPU is executing a process
then it will not be preempted unless the process is completed. On the other
hand, the preemptive version preempted the CPU if such a situation occurs.
The main attraction in PS is realistic approach about priorities. However, PS
can lead to starvation situation preventing lower priority processes from execu-
tion. Its implementation involves high overheads of queue rearrangements and
priority comparisons in case of preemptive version. PS can be implemented in the
form of either SQMS (Traditional Priority Scheduling) or MQMS (Multi-Level
Queue Scheduling). Rajput and Gupta [14] narrated that in priority scheduling,
processes were arranged in queue according to their fixed priority values. A high
priority incoming process interrupted low priority process. Waiting time of low
priority processes were thus increased and hence starvation occurred.
Adekunle et al. [2] analysed the performance of FCFS, SJF, RR, SRTF and
Lottery scheduling algorithms. SJF showed higher while RR showed medium
CPU utilization. SJF also showed lower average waiting time and good response
time as compared to RR. However, RR algorithm was found to be fair and
starvation free.
Almakdi et al. [4] simulated FCFS, SJF, RR and SRT algorithms for
analysing scheduling algorithm performance. The results showed that the lowest
average waiting time and the lowest average turnaround time were showed by
SJF. However, high average core utilization and high average throughput were
resulted by RR.
4 Related Work
Rao et al. [15] elaborated that Multilevel Feedback Queue Scheduling involved
use of multiple queues. Process could be migrated (promoted/demoted) from
one queue to another. Each queue had its own scheduling algorithm. Similarly,
criteria were also defined for promotion and demotion of processes. PMLFQ
resulted in starvation because of execution of high priority process queues before
low priority process queues.
Goel and Garg [5] explained fairness was an important scheduling criteria.
Fairness ensured that each process got fair share of CPU for execution. Further,
higher priority processes exhibited smaller waiting time and starvation could
take place for lower priority process.
Shrivastav et al. [16] proposed Fair Priority Round Robin with Dynamic
Time Quantum (FPRRDQ). FPRRDQ calculated time quantum on the basis of
priority and burst time for each individual process. Experimental results when
compared with Priority Based Round Robin (PBRR) and Shortest Execution
First Dynamic Round Robin (SEFDRR) scheduling algorithms, showed that
FPRRDQ had improved performance considerably.
Patel and Patel [13] proposed a Shortest Job Round Robin (SJRR) schedul-
ing algorithm. SJRR suggested to arrange processes in SJF fashion. The shortest
process’s burst time was designated as time quantum. Case study showed reduc-
tion in average waiting time of process set.
Abdulrahim et al. [1] proposed New Improved Round Robin (NIRR) algo-
rithm for process scheduling. The algorithm involved two queues ARRIVE and
REQUEST. In this algorithm, the time quantum was calculated by finding the
ceiling of average burst time of process set. First process in the REQUEST queue
was allocated to CPU for one time quantum. In any case, the executing process
had burst time less than or equal to half time quantum then CPU reallocated for
remaining CPU burst time. Otherwise, process was removed from REQUEST
and added to ARRIVE. NIRR algorithm improved scheduling performance and
also lowered number of context switches as the case of RR algorithm.
Mishra and Rashid [12] introduced Improved Round Robin with Varying
Time Quantum (IRRVQ) scheduling algorithm. IRRVQ proposed arrangement
of processes (in ready queue) in SJF fashion and setting time quantum equal
to burst time of first process. After a complete execution of first process, next
processes were selected and assigned one time quantum. If finished, processes
were removed from ready queue otherwise, processes were preempted and added
to tail of ready queue.
Sirohi et al. [20] described an improvised round robin scheduling algorithm
for CPU scheduling. This algorithm calculated the time quantum by finding the
average of the burst times of processes. The processes were allocated to ready
queue in SJF order. CPU was allocated to first process for one time quantum.
The reallocation of CPU was based on remaining burst time of process under
execution.
Akhtar et al. [3] proposed a preemptive hybrid scheduling algorithm based
on SJF. Initially, it showed similar steps as that of SJF. The algorithm reduced
number of context switches considerably but only showed minor change in aver-
age waiting time. The algorithm could exhibit starvation.
Joshi and Tyagi [7] explained Smart Optimized Round Robin (SORR)
scheduling algorithm for enhancing performance. At first step, the SORR algo-
rithm arranged processes in increasing order of burst times. In the second step,
the mean burst time was calculated. Third step involved calculation of Smart
Time Quantum (STQ) after which the CPU was allocated to first process in
the ready queue for a time quantum equal to calculated STQ. If the allocated
process was remained with burst time less or equal to 1 time quantum, CPU was
reallocated for the burst time. Otherwise process was preempted, next process
was selected from ready queue and CPU was allocated to the newly selected
process for STQ.
Lulla et al. [10] devised a new algorithm for CPU scheduling. The algo-
rithm introduced mathematical model for calculating initial time quantum and
54 U. Rafi et al.
dynamic time quantum. Dynamic time quantum assisted in lowering number

of context switches. The algorithm proposed execution of top two high priority
processes by assigning calculated initial time quantum followed by the calcula-
tion and assignment of dynamic time quantum for execution of top two high
priority processes.
Goel and Garg [6] proposed Optimized Multilevel Dynamic Round Robin
Scheduling (OMDRRS) algorithm. This algorithm suggested the allocation of
dynamic time slices. Factor analysis based on arrival time, burst time and prior-
ity was calculated for sorting processes in ready queue. Burst values of first and
last process were used for time slice calculation. Simulated results, statistically
analysed using ANOVA and t-test, showed considerable performance improve-
ment.
Kathuria et al. [8] devised Revamped Mean Round Robin (RMRR) schedul-
ing algorithm. RMRR involved two queues ReadyQueue (RQ) and Pre-
ReadyQueue (PRQ). Time quantum was calculated by taking mean of the burst
time of processes. The proposed algorithm lowered average waiting time, average
turnaround time and number of context switches as compared to RR scheduling.
RMRR showed the best response time when compared with FCFS, SJF and RR
scheduling algorithms.
Singh and Singh [19] proposed a hybrid scheduling algorithm based on Multi
Level Feedback Queue (MLFQ) and Improved Round Robin (IRR) scheduling
principles. The proposed algorithm suggested the sub-division of memory in
three queues as in MLFQ. Fonseca and Fleming’s Genetic Algorithm (FFGA)
was implemented for optimization. Results showed optimized performance in
comparison to various algorithms (Fig. 3).
Fig. 3. Schedulers and scheduling queues
5 Proposed Algorithm
In this paper, we proposed a new scheduling algorithm as an extension of pri-
ority scheduling. Here, we proposed it only for single processor system (neither
multiple processors nor multi-core processors). It is based on hybrid concepts of

multiple queue management, priority scheduling and round robin scheme. In our
algorithm, multiple queue management scheme facilitates subdivision of ready
queue in multiple portions. Shukla et al. [17] described the use of multi-level
queuing approach. Multi-level queuing involved partitioning of ready queue into
various sub queues. In such scheduling, processes were assigned permanently to
specific queue. Memory size, process type and priority were some of parameters
used for such assignments.
In our proposed algorithm, each job is allocated to specific queue according
to criteria predefined criteria. The criteria consists of two factors: (1) Priority
of individual process and (2) Burst Time of process. Firstly, a priority criteria
is defined in a set of processes. The processes are allocated after comparing
process priorities to priority criteria. Our proposed algorithm assumes that the
largest number represented highest priority and smallest number represented
lowest priority. Shortest job first scheme is applied while arranging processes
when they are transferred from the job pool to the ready queue partitions. Jobs
inside a specific queue are scheduled as defined by priority scheduling scheme.
Khan and Kakhani [9] suggested SJF based PS was proposed for performance
improvement. Processes were placed in ready queue in higher to lower order
of priority. SJF was employed for arranging process in case of more than one
processes having similar priority. Results of SJF based PS were compared to
FCFS based PS and results showed considerable decrease in average waiting
and average turnaround times. Moreover, the queues are scheduled in a round
robin fashion from the highest priority queue to the lowest priority queue. Round
robin scheme is based on dynamic time slicing. Whenever a queue is given time
slice it is according to the process burst time and some variations of it. Our
algorithm has following main steps:
Step 1. Create Multiple Queues (one for high priority processes and other for
low priority processes) based on Priority Scheme. These queues are
designated as High Priority and Low Priority Queues.
Step 2. Create Processes in Job Pool.
Step 3. Define Priority Criteria.
Step 4. Allocate processes to queues based on priority criteria and burst time.
Processes are arranged in decreasing order of priority from front to
rear. Additionally following are considered:
Step 4.1.1 If process priority is less than or equal to priority criteria
then add it to the Low Priority Queue. Otherwise add it
to High Priority Queue.
Step 4.1.2 If multiple processes have same priority then arranges
them in a way that the process with shortest burst time
comes first in the queue.
Step 5. Select first process from High Priority Queue and execute it fully (for
entire burst time by assigning time slice equal to the burst time) and
remove it from queue head.
56 U. Rafi et al.
Step 6. Select first process from Low Priority Queue and execute for a time
equal to half that of burst time (assign time slice equal to half the
burst time). Important thing to consider is that, in this case, each low
priority process completed in two halves.
Step 7. Repeat Steps 5–6 until both queues are empty.
One of the interesting things to note is that, CPU is alternated between

high priority and low priority processes in order to increase the fairness. Our
approach lowers the starvation faced by low priority processes in the presence of
high priority processes. Another point may arise that why low priority processes
are executed in two halves? The answer of this question is that, it is done so
that if a low priority process is executed fully, then the purpose of two queues
is unclear. Similarly, if high priority process is executed for half time, then the
use of priorities becomes meaningless.
6 Pseudo Code
Proposed Algorithm (P[1, · · · , n], Priority Criteria) (Fig. 4)
begin
Create Queue Portions QH[1....nh], QL[1....nl]
For each P[i] in P[1.......n]
If (P[i].Priority <= PriorityCriteria) Then
Add P[i] to QL in such a way that processes are arranged in
decreasing order of priority and increasing order of burst time
Else
Add P[i] to QH in such a way that processes are arranged in
decreasing order of priority and increasing order of burst time
End If
Repeat following steps while QH and QL are not empty
Select a process from QH head, allocate CPU, execute completely and
remove it from QH
Select a process from QL head, allocate CPU, execute for the time half to
its original burst time
If (P[i].RBT) = 0 Then
Removes it from QL
Else
place it back to QL head
End If
Goto step 4
End while
Completion of process set
End Algorithm
*RBT = Remaining Burst Time
Fig. 4. The proposed algorithm
7 Case Studies
7.1 Case Study 1
Gantt Chart of above process data when executed by Priority Scheduling is as
Fig. 5 (Table 1).
Table 1. Process set with five processes
Process Burst time Priority

P1 10 3
P2 1 1
P3 2 3
P4 1 4
P5 5 2
Fig. 5. Gantt chart for priority scheduling algorithm
Average Waiting Time = 8.6 ms.

Average Turnaround Time = 12.4 ms.
Now we define certain parameters related to proposed algorithm.
(1) Priority Criteria: Process with Priority > 2 is placed in High Priority
Queue while process with Priority <= 2 is placed in Low Priority Queue.
(2) Ready Queue Partitions: As Fig. 6.
The proposed algorithm on given data results in Fig. 7.
Fig. 6. Scheduling queues Fig. 7. Gantt chart for proposed algorithm
Average Waiting Time = 6.5 ms

Average Turnaround Time = 10.3 ms
7.2 Case Study 2

Fig. 8 (Table 2).

58 U. Rafi et al.

P1 10 3
P2 1 1
P3 2 4
P4 1 5
P5 5 2
Queue while process with Priority ≤ 2 is placed in Low Priority Queue.
(2) Ready Queue Partitions: As Fig. 9.

7.3 Case Study 3

Fig. 11 (Table 3).
(2) Ready Queue Partitions: As Fig. 12

Fig. 9. Scheduling queues Fig. 10. Gantt chart for proposed algo-
rithm

P1 12 3
P2 19 3
P3 21 5
P4 13 2
P5 15 3
7.4 Case Study 4–Special Case

This case deals with a situation involving same priority processes (high priority).
Similarly, inverse case of this can also be assumed. In such cases all the processes
added to a single queue either high or low priority (Table 4).
Gantt Chart of above process data when executed by Priority Scheduling is
as Fig. 14.

(2) Ready Queue Partitions: As Fig. 15

In the case with all the processes getting place in a single queue, the proposed
algorithm gives an average waiting time and average turnaround time exactly
equal to priority scheduling algorithm.
rithm
60 U. Rafi et al.
Table 4. Process set with three processes
Process Burst Time Priority

P1 10 5
P2 3 6
P3 4 7
Fig. 14. Gantt chart for proposed scheduling algorithm
8 Results and Discussion

8.1 Comparison of Average Waiting Time
In this section we compare average waiting time of our proposed algorithm and
traditional priority scheduling algorithm. This comparison is based on case stud-
ies discussed in the previous sections.
The average waiting time is minimized by our proposed algorithm as com-
pared to the baseline scheduling algorithm as shown in Fig. 17. However, the last
case showed same average waiting time. This was because, the last case com-
prised of special dataset involving all the processes with same priority (high). In
this case, only one priority queue was formed rather two queues. So, in this case
it exhibited same behaviour as that of traditional priority scheduling algorithm.
8.2 Comparison of Average Turnaround Time

In this section we presented comparison of average turnaround time of our pro-
posed algorithm with traditional priority scheduling algorithm. This comparison
based on case studies discussed in the previous sections.
Figure 18 clearly showed that in first three case studies the average turn-
around time of our proposed scheduling algorithm was less than traditional pri-
ority scheduling algorithm. However, the last case showed same average turn-
around time. This was because, the last case comprised of special dataset having
rithm
Fig. 17. Comparison of average waiting time
Fig. 18. Comparison of average turnaround time
all the processes having high priority. In this case only one priority queue was
formed rather two queues.
9 Conclusion and Future Work

It is concluded that our proposed algorithm used minimum waiting and turn-
around time as compared to baseline Priority Scheduling algorithms. Our pro-
posed algorithm also supported multiple queue supporting operating systems
using different priorities. Each type of priority can be managed with a separate
queue. The proposed algorithm provided fairness to both high priority and low
priority jobs and thus saved low priority jobs from starvation. The importance is
given to high priority process within both high priority and low priority list by
our proposed algorithm. Our algorithm also enhanced throughput by lowering
average waiting and average turnaround time. As compared to Round Robin
scheduling with small time slice, proposed algorithm used a scheme requiring
only two context switches per process. In this way, our proposed algorithm lim-
ited the number of context switches required for a job. This algorithm is a good
mixture of priority scheduling, shortest job first and round robin scheduling.
As a future work, we will try out different execution lengths or time slices for
high priority processes and low priority processes to find optimized solution pro-
viding optimized results of scheduling criteria. We will also extend this algorithm
in a way that for each priority type a single queue may be used.
62 U. Rafi et al.
References
1. Abdulrahim A, Abdullahi SE, Sahalu JB (2014) A new improved round robin (nirr)
cpu scheduling algorithm. Int J Comput Appl 90(4):27–33
2. Adekunle O (2014) A comparative study of scheduling algorithms for multipro-
gramming in real-time systems. Int J Innov Sci Res 12:180–185
3. Akhtar M, Hamid B et al (2015) An optimized shortest job first scheduling algo-
rithm for cpu scheduling. J Appl Environ Biol Sci 5:42–46
4. Almakdi S (2015) Simulation and Performance Evaluation of CPU Scheduling
Algorithms. LAP LAMBERT Academic Publishing, Saarbrücken
5. Goel N, Garg RB (2013) A comparative study of cpu scheduling algorithms. Int J
Graph Image Process 2(4):245–251
6. Goel N, Garg RB (2016) Performance analysis of cpu scheduling algorithms with
novel omdrrs algorithm. Int J Adv Comput Sci Appl 7(1):216–221
7. Joshi R, Tyagi SB (2015) Smart optimized round robin (sorr) cpu scheduling algo-
rithm. Int J Adv Res Comput Sci Softw Eng 5:568–574
8. Kathuria S, Singh PP et al (2016) A revamped mean round robin (rmrr) cpu
scheduling algorithm. Int J Innov Res Comput Commun Eng 4:6684–6691
9. Khan R, Kakhani G (2015) Analysis of priority scheduling algorithm on the basis
of fcfs and sjf for similar priority jobs. Int J Comput Sci Mob Comput 4:324–331
10. Lulla D, Tayade J, Mankar V (2015) Priority based round robin cpu scheduling
using dynamic time quantum. Int J Emerg Trends Technol 2:358–363
11. Mishra MK (2012) An improved round robin cpu scheduling algorithm. J Glob Res
Comput Sci 3(6):64–69
12. Mishra MK, Rashid F (2014) An improved round robin cpu scheduling algorithm
with varying time quantum. Int J Comput Sci Eng Appl 4(4):1–8
13. Patel R, Patel M (2013) Sjrr cpu scheduling algorithm. Int J Eng Comput Sci
2:3396–3399
14. Rajput G (2012) A priority based round robin cpu scheduling algorithm for real
time systems. Int J Innov Eng Technol 1:1–10
15. Rao MVP, Shet KC, Roopa K (2009) A simplified study of scheduler for real time
and embedded system domain. Comput Sci Telecommun 12(5):1–6
16. Shrivastav MK, Pandey S et al (2012) Fair priority round robin with dynamic time
quantum: Fprrdq. Int J Mod Eng Res 2:876–881
17. Shukla D, Ojha S, Jain S (2010) Data model approach and markov chain based
analysis of multi-level queue scheduling. J Appl Comput Sci Math 8(4):50–56
18. Silberschatz A, Gagne G, Galvin PB (1983) Operating System Concepts, 8th edn.
Addison-Wesley Pub. Co, Boston Binder Ready Version
19. Singh N, Singh Y (2016) A practical approach on mlq-fuzzy logic in cpu scheduling.
Int J Res Educ Sci Methods 4:50–60
20. Sirohi A, Pratap A, Aggarwal M (2014) Improvised round robin (cpu) scheduling
algorithm. Int J Comput Appl 99(18):40–43
21. Stallings W (2011) Operating Systems–Internals and Design Principles, 7th edn.
DBLP
22. Tanenbaum AS, Tanenbaum AS (2001) Modern Operating Systems, 2nd edn.
Prentice-Hall, Upper Saddle River
23. Ulfahsiregar M (2012) A new approach to cpu scheduling algorithm: Genetic round
robin. Int J Comput Appl 47(19):18–25
An Order-Based GA for Robot-Based Assembly
Line Balancing Problem
Lin Lin1,2,3(B) , Chenglin Yao1 , and Xinchang Hao4

1
lin@dlut.edu.cn
2
Fuzzy Logic Systems Institute, Tokyo, Japan
3
Key Laboratory for Ubiquitous Network and Service Software of Liaoning Province,
Dalian 116024, People’s Republic of China
4
Waseda University, Tokyo 169-0051, Japan
Abstract. In the real world, there are a lot of scenes from which the
product is made by using the robot, which needs different assembly times
to perform a given task, because of its capabilities and specialization. For
a robotic assembly line balancing (rALB) problem, a set of tasks have
to be assigned to stations, and each station needs to select one robot to
process the assigned tasks. In this paper, we propose a hybrid genetic
algorithm (hGA) based on an order encoding method for solving rALB
problem. In the hGA, we use new representation method. Advanced
genetic operators adapted to the specific chromosome structure and the
characteristics of the rALB problem are used. In order to strengthen the
search ability, a local search procedure is integrated under the frame-
work the genetic algorithm. Some practical test instances demonstrate
the effectiveness and efficiency of the proposed algorithm.
Keywords: Balancing · Improvement · Genetic algorithms · Local

search · Neighborhood structure · Robotic assembly line
1 Introduction
An assembly line system is a manufacturing process in which interchangeable
parts are added to a product in a sequential manner to produce a finished prod-
uct. It is important that an assembly line is designed and balanced so that it
works as efficiently as possible. Most of the work related to the assembly lines
concentrate on the assembly line balancing (ALB). The ALB model deals with
the allocation of the tasks among stations so that the precedence relations are
not violated and a given objective function is optimized.
Since the ALB model was first formulated by Helgeson et al. [13], many ver-
sions of ALB arise by varying the objective function [26]: Type-F is an objective
independent problem which is to establish whether or not a feasible line balance
exists. Type-1 and Type-2 have a dual relationship; the first one tries to min-
imize the number of stations for a given cycle time, and the second one tries
64 L. Lin et al.
to minimize the cycle time for a given number of stations. Type-E is the most
general problem version, which tries to maximize the line efficiency by simul-
taneously minimizing the cycle time and a number of stations. Finally, Type-3,
4 and 5 correspond to maximization of workload smoothness, maximization of
work relatedness and multiple objectives with Type-3 and Type-4, respectively
[15]. Furthermore, most versions of above models are NP-hard [12].
Recently, genetic algorithm (GA) and other evolutionary algorithms (EAs)
have been successfully applied in a wide variety of ALB problems. Falkenauer
and Delchambre [6] were the first to solve ALB with GAs. Following application
of GAs for solving ALB model was studied by many researchers, e.g., Anderson
and Ferris [2], Rubinovitz and Levitin [22], Gen et al. [10], Bautista et al. [3],
Sabuncuoglu et al. [24], Goncalves and Almeida [11], Brown and Sumichrast [4],
and Nearchou [21]. However, most of the researchers focused on the simplest
version of the problem, with single objective and ignored the recent trends, i.e.,
mixed-model production, u-shaped lines, robotic lines and etc.
Sigeru focused on implementing particle swarm optimization (PSO) to opti-
mize the robotic assembly line balancing (RALB) problems with an objective of
maximizing line efficiency. By maximizing the line efficiency, industries tend to
utilize their resources in an efficient manner [14]. Slim studied a robotic assem-
bly line balancing problem which goal is to maximize the efficiency of the line
and proposed some resolution methods which define the suitable component
and point positions in order to define the strategy of pick and place for each
robot [5]. In Nilakantan’s research, bio-inspired search algorithms, viz. particle
swarm optimization (PSO) algorithm and a hybrid cuckoo search and particle
swarm optimization (CS-PSO) were proposed to balance the robotic assembly
line with the objective of minimizing the cycle time [20]. Armin et al. extended
the existing algorithms to solve the robotic assembly line problem [27]. Yoose-
felahi et al. [29] considered a different type II robotic assembly line balancing
problem (RALB-II). One of the two main differences with the existing literature
is objective function which is a multi-objective one. The aim is to minimize the
cycle time, robot setup costs and robot costs. The second difference is on the
procedure proposed to solve the problem. In addition, a new mixed-integer lin-
ear programming model is developed. Snke [16] introduced a new robot concept
that aims at closing the gap between a manual assembly and a fully automatic
assembly. It is intended to be used for handling and assembly of small parts in a
highly agile production scenario, which employs both human workers and robots
in the same line, with a frequent need for reconfiguration. The objective is to
minimize the cost of the stations. This formulation is very similar to the rALB
problem.
Gao and Gen [7] proposed a hybrid GA for solving rALB-2, in which different
robots may be assigned to the assembly line tasks, and each robot needs different
assembly times to perform a given task due to its capabilities and specialization.
They focused on the genetic operators adapted to the specific chromosome struc-
ture and the characteristics of the rALB problems.
An Order-Based GA for Robot-Based Assembly Line Balancing Problem 65
In this paper, we also consider a hybrid GA approach for solving rALB-2

problems. Different with prophase works [7], we not only focus on the design of
GA to minimize the cycle time of the assembly systems, but also hammer at
the key point of improvement of assembly system depended on the final result
analysis. We propose a hybrid genetic algorithm (hGA) to solve the rALB-2.
The genetic algorithm uses partial representation technique, that is, the coding
space only contains a part of the feasible solutions, in which the optimal solution
is included. New crossover and mutation operators are also developed to adapt
to the specific chromosome structure and the nature of the problem. In order to
strengthen the search ability, a local search procedure works under the framework
of GA, this only investigates the neighbor solutions that have possibilities to
improve the solutions of current population.
The rest of the paper is organized as follows. In Sect. 2, we formulate the
mathematical model of rALB-2. The proposed GA for the problem is described
in Sect. 3. In Sect. 4, introduce a local search procedure. In Sect. 5, we demon-
strate effectiveness comparing with different GA approaches. This paper gives
the conclusion in Sect. 6.
2 Robotic Assembly Line Balancing Problem
The assembly of each product unit requires the execution of n tasks (indivisible
elements of work). Precedence constraints partially specify the order in which the
tasks have to be performed. They can be represented by an acyclic precedence
graph which contains nodes for all tasks. As mentioned above, the type II robotic
assembly line balancing problem usually occurs when changes in the production
process of a product take place. In this case, the robotic assembly line has to
be reconfigured using the present resources (such as robots) so as to improve it
efficiency for the new production process. The problem concerns how to assign
the tasks to stations and how to allocate the available robots for each station in
order to minimize the cycle time under the constraint of precedence relationships.
Let assembly tasks (i = 1, 2, · · · , 10) assigned on 4 stations. Robots (R1, R2,
R3, R4) are to be equipped on the 4 stations. The balancing chart of the solution
was drawn for analyzing the solution. Figure 1 shows the idle time of the station
1, 2, 3 is very big, and it also means this line did not get the balancing for
producing. In the real world, the assembly line is not just for producing one unit
of the product. It should produce several units. We give the Gantt chart about
3 units for analyzing the solution just like Fig. 2.
We can see the waiting time occurs in the Fig. 2, and the waiting time means
the idle time in the line. For example, the part waiting is occurred when the
maximum processing time of the station, which is before the current station,
is larger than the processing time of the current station. It means the current
station will wait for the parts to process, which come from the anterior station.
The processing waiting is occurred when the processing time of the current
station is smaller than the processing time of the next station. It means the
parts which were produced by the current station will wait for being processed
66 L. Lin et al.
stn/rbn
4/4 8 9 10
3/3 1 4 6
2/1 3 5
1/2 2 7
50 85 time
Fig. 1. Balancing chart of the feasible solution
stn / rbn (case of 3 products) : Part waiting

1/2 2 7 : Processing waiting
2/1 3 5
3/3 1 4 6
4/4 8 9 10
50 100 150 200 250 300 350 400
Fig. 2. Gantt chart for producing 3 units
by the next station. Both of the waiting time is the idle time of the assembly
line. We want to reduce it by getting the balance of the line. The notation used
in this section can be summarized as follows:
Indices:
i, j: index of assembly tasks, i, j = 1, 2, · · · , n;

k: index of stations, k = 1, 2, · · · , m;
l: index of robots, l = 1, 2, · · · , m.
Parameters:
n: total number of assembly tasks;
m: total number of stations (robots);
til : processing time of the i − th task by robot l;
pre(i): the set of predecessor of task i in the precedence diagram.
Decision variables:

1, if task j is assigned to work station k
xjk =
0, otherwise;
1, if robot j is assigned to work station k
yjk =
0, otherwise.
Problem Formulation:
n
m
min CT = min til xik ykl (1)
1≤k≤m
i=1 l=1
m
m

s. t. kxj k − kxik ≥ 0, ∀j; i ∈ pre(i) (2)
k=1 k=1
m
xik = 1, ∀i (3)
k=1
m
xkl = 1, ∀k (4)
l=1
m
xkl = 1, ∀l (5)
k=1
xik ∈ 0, 1, ∀k, i (6)
xik ∈ 0, 1, ∀l, k. (7)
The objective (1) is to minimize the cycle time (CT ). Inequity (2) represents
the precedence constraints. It ensures that for each pair of assembly activities,
the precedent cannot be assigned to a station behind the station of the successor,
if there is precedence between the two activities. Equation (3) ensures that each
task has to be assigned to one station. Equation (4) ensures that each station
is equipped with one robot. Equation (5) ensures that each robot can only be
assigned to one station.
3 Order-Based GA Design
Genetic algorithms (GAs) are powerful and broadly applicable stochastic search
and optimization techniques based on principles from evolutionary theory [9].
GA’s have been applied to solve various assembly line balancing problems
[8,23,28]. We divide this algorithm into 3 phases. It is as follows:
S1 S2 S3 S4
S1 S2 S3 S4
2 1 3 4 9 6 5 7 8 10
Fig. 3. Structure of example problem
Phase 1. Deciding Task Sequence
Phase 1.1 Order encoding for task sequence: A GA’s structure and parameter
settings affect its performance. However, the primary determinants
of a GA’s success or failure are the coding by which its genotypes
represent candidate solutions and the interaction of the coding with
the GA’s recombination and mutation operators.
68 L. Lin et al.
Locus 1 2 3 4 5 6 7 8 9 10
Phase 1: Task Sequence (v1) 2 1 3 4 9 6 5 7 8 10
Locus: station 1 2 3 4
Phase 2:
Robot assignment (v2) 1 2 3 4
Fig. 4. Solution representation of a sample problem
A solution of the rALB problem can be represented by two integer

vectors. Task sequence vector (v1 ), which contains a permutation of
assembly tasks, ordered according to their technological precedence
sequence, and robot assignment vector (v2 ). The solution represen-
tation method can be visually illustrated as in Figs. 3 and 4.
Phase 1.2 Reordering procedure: Considering precedence constraints, a task
sequence may be infeasible since a precedent may appear before its
successors. The reordering procedure will repair it into a feasible one
before it is divided m parts to form a solution in the phase 1.
Phase 1.3 Breakpoint decoding: So, we use just one decoding procedure for
phase 1, and the set of decoding procedure is to generate a fea-
sible solution based on the task sequence and robot assignment
schemes which are contained in the chromosomes. Then, the break-
point decoding procedure inserts m points along the reordered task
sequence vector to divide it into m parts, each of which corresponds
to a station. The breakpoint decoding procedure consists of four
main steps:
Step 1. Calculate the lower bound of the cycle time (CLB ) for the
solution represented by the task sequence vector and robot
assignment vector;
n
1
cLB = min {til }; (8)
m i=1 1≤l≤m
Step 2. Find out a feasible cycle time as the upper bound of the cycle
time (CU B );
Step 3. Find out the optimal cycle time via bisection method;
Step 4. Partition the task sequence into m parts with the optimal
cycle time based on the robot assignment vector.
Here, a cycle time is said to be feasible if all the tasks can be allocated
to the stations by allowing as many tasks as possible for each station
under the constraint of the cycle time. The procedure to calculate
the upper bound of cycle time is illustrated in procedure 2, and
the bisection method to find out the optimal cycle time is shown in
procedure 3. After calculating the optimal cycle time, it is easy to
generate the breakpoints on the task sequence to divide it into m
parts, each of which will correspond to a station, based on the robot
assignment vector. Therefore, the tasks assigned for each station are
determined according to a chromosome, and the cycle time can be

calculated. An example of breakpoint decoding is showed in Fig. 5.
Phase 2. Assigning Robots to Station
Phase 2.1 Order encoding for robot assignment: A solution of the rALB problem
can be represented by two integer vectors. In this phase, we code
the robot assignment vector (v2 ), which indicates the robot assigned
for each station. The solution representation method can be visually
illustrated as in Fig. 6. The coding space takes an exponential growth
with the length of the chromosome; therefore, even one allele saved
in the chromosome can decrease the coding space significantly. An
obvious advantage to intentionally omit breakpoint vector, which is
used in recently research, in the chromosome is that the coding space
is dramatically decreased. As a result, the speed to find the global
optimal is accelerated.
Phase 2.2 Breakpoint decoding: In this phase, this robot assignment vector (v2 )
will decode into a solution with the task sequence vector (v1 ), simul-
taneously by the breakpoint decoding procedure. It means when the
tasks are assigned into each station, the robots will be assigned into
each station at the same time. The illustration of the solution is shown
in Phase 1.3.
Phase 3. Making a Schedule
Phase 3.1 Creating a schedule for the assembly line: We can make a schedule
based on the chromosome (in Fig. 6) as fellow:
S = {(t1 , R1 : 0 − 17), (t2 , R1 : 17 − 38), (t3 , R1 : 38 − 50), (t4 , R2 : 50 − 71),
(t5 , R3 : 102 − 128), (t6 , R2 : 71 − 89), (t7 , R3 : 128 − 151),
(t8 , R4 : 151 − 176), (t9 , R2 : 89 − 102), (t10 , R4 : 176 − 202)}
Phase 3.2 Drawing a Gantt chart: Firstly, we draw a balancing chart for analyz-
ing the solution (in Fig. 8). We can see the solution got the balancing
for the line by comparing with the feasible solution from Fig. 2.
Breakpoint 1 Breakpoint 3 Breakpoint 7 Breakpoint 9

1 2 3 4 5 6 7 8 9 10
V1: 2 1 3 4 9 6 5 7 8 10
R1 R2 R3 R4
Fig. 5. Breakpoint decoding
In the real world, the assembly line is not just for producing one unit of the
product. It should produce several units. So, we give the Gantt chart with 3 units
for analyzing the solution just like Fig. 7, which is shown at the beginning. We
can see the solution reduce the waiting time for the line by comparing with the
feasible solution from Fig. 6. It also means the better solution got the balancing
for the assembly line.
70 L. Lin et al.
stn/rbn (Balancing chart of the best solution)
4/4 8 10
3/3 5 7
2/2 4 6 9
1/1 1 2 3
52 time
Fig. 6. The balancing chart of the best solution
stn / rbn (case of 3 products)
1/1 1 2 3
: Processing waiting
2/2 4 6 9 : Part waiting
3/3 5 7
4/4 8 10
50 100 150 200 250 300 time
Fig. 7. Gantt chart for producing 3 units
3.1 Crossover Operators
In genetic algorithms, crossover recombines two chromosomes to generate a num-

ber of children. Offspring of crossover should represent solutions that combine
substructures of their parental solutions [8]. In this paper, we use to kind of
crossover methods: exchange crossover and mixed crossover. The detailed expla-
nation of OX and PMX is introduced in [8].
(1) Exchange Crossover: A good schedule could be expected by exchanging task

sequence and robot assignment schemes between a pair of parents. This
kind of crossover is accomplished by selecting two parents and exchanging
the task sequence vectors of the two parents to generate offspring.
(2) Mixed Crossover: The mixed order crossover consists of two crossover meth-
ods: Order Crossover (OX) and Partial-Mapped Crossover (PMX). For the
rALB chromosome, the task sequence vector and the robot assignment vector
are both of the style of permutation representation. Yet, the useful features
of the task sequence vector and the robot assignment vector are different.
• For the task sequence vector, the useful feature of permutation representation
to represent a rALB solution is the order information among the tasks.
• For the robot assignment vector, the acting feature is the number at each
allele which indicates the robot no. assigned for a specific station.
3.2 Mutation Operators
In this study, two kinds of mutation operators are implemented: allele-based

mutation and immigration mutation. For both the robot task sequence and
robot assignment vectors, allele-based mutation randomly decides whether an
allele should be mutated in a certain probability. Then, another position will
be generated at random to perform exchange values with the mutated allele.
In contrary to the canonical gene-by-gene mutation with very small probability
at each generation, immigration mutation randomly generates one or more new
members of the population from the same distribution as the initial population.
This process prevents premature convergence of the population, and leads to a
simple statement of convergence.
3.3 Fitness Evaluation

How to evaluate the fitness of chromosome is also key issue for GAs. In the most
researches, objective function is used for fitness evaluation. However, the effective
fitness function has to be different function and suitable to different problems.
A central problem of fitness definition is how to guide to the most promising areas
from the individuals of current generation. The characteristic of ALB problems
are: 1 We can calculate several different schedules with same cycle time. That
means other conditions, such as idle time and total processing time are different,
though the cycle time is same. For improvement of ALB, we also have to improve
other conditions, not only cycle time. 2 General, the solutions with better fitness
of idle time and total processing time also can create better fitness of cycle time.
We define the fitness function with considering not only the objective function
(cycle time), but also idle time and total processing time. It is useful to guide to
the most promising areas depending on the wide conditions from the individuals
of current generation. We first define the fitness value f1 (v) as the reciprocal
of the cycle time CT (v) of the solution candidate v, the fitness values f2 (v) as
number of critical stations nc and the fitness value f3 (v) as the reciprocal of the
standard deviation by the idle time of each station for balancing the workload
of each station. tW
k is the total processing time of each station. Then we give the
adaptive weights for each fitness, where w1 + w2 + w3 = 1. The special definition
of fitness function is showed as follows:
f (v) = w(f (v)z min ) + w(f (v)z min ) + w(f (v)z min ), (9)
where
1
f1 (v) = , (10)
CT (v)
f2 (v) = nc (v), (11)
m m
1 1
f3 (v) = ((CT (v) − tW
k (v)) − (CT (v) − tW
j (v))). (12)
m−1 m j=1
k=1
72 L. Lin et al.
4 Local Search Procedure

Two individuals with high fitness values are likely to have dissimilar machine
assignment and operation sequences, and the recombination may result in
offspring of poor performance. This means that the genetic operations by
themselves have limited ability in finding the global optimal solution. Many
researchers have found that the convergence speed of simple GAs is relatively
slow [7]. One promising approach for improving the convergence speed to the
global optimum is the use of local search in GAs [17]. In this study, two local
search methods are proposed to enhance the search ability of GA. Both the two
local search methods are based on critical stations in order to improve their
effectiveness and efficiency.
4.1 Task Sequence Search
Let s(i) be the station to which task i is assigned. First we introduce the logical
function Wi (k) which return false if task i cannot be transferred from station
s(i) to station k, and true otherwise:
Wi (k) = false if
s(i) > k and j ∈ suc(i), s(j) > k, or
s(i) < k and j; i ∈ suc(j), s(j) < k
Wi (k) = true, otherwise.
For a pair of tasks i, j, if
Wi (s(j)) = Wj (s(i)) = true ∧ i ∈

/ pre(i) j ∈
/ pre(i),
then, the two tasks are said to be exchangeable. Let R(i) be the robot which is
allocated for station s(i), and T l be the total assembly time of station l. For a
critical task i and task j which is not in s(i), if
tj,s(i) < ti,s(i) and Ts (j) − tj,s(j) + ti,s(j) < CT ,
then the exchange between tasks i and j is called worthwhile. The task sequence
search is as follows:
Step 1. Decode the chromosome and identify a critical station l∗ ;

Step 2. For a task i that assigned to the critical station l∗ , look for a task j that
is not assigned to station l∗ , and is exchangeable with task i.
Step 3. Repeat Step 2 until the exchange between tasks i and j is worthwhile or
all the tasks in the critical station are tried.
Step 4. Repeatedly perform the above search steps until the chromosome cannot
be improved any more.
4.2 Robot Assignment Search

Let N (i) denote the set of machine assignment neighborhood of solution i. The
enlarged two-pace machine assignment neighborhood is defined as the union of
the neighborhood of each robot assignment neighbor of solution i
N 2 (i) = ∪j∈N (i) N (j). (13)
During the robot assignment search, the local search will implement over two-
pace neighborhood when it reaches the local optima of one-pace neighborhood,
and is called two-pace robot assignment search. During the robot assignment
search, when the local optima of two-pace robot assignment neighborhood is
reached, the neighbors of the two-pace local optima are improved by the task
sequence search to help the robot assignment search escape from the local
optima. Since the size of the two-pace neighborhood is very large and the task
sequence search is very computationally complex, to improve each neighbor solu-
tion by task sequence search needs a great computation time. Therefore, only
those neighbor solutions that are local optimums of the two-pace robot assign-
ment neighborhood are improved by task sequence search so as to save compu-
tation time.
5 Numerical Experiments
In the literature, no benchmark data sets are available for rALB. We collect 8
representative precedence graphs from [1], which are widely used in the sALB-1
literature [25]. These precedence graphs contain with 25-297 tasks. From each
precedence graph, 4 different rALB-2 problems are generated by using different
WEST ratios: 3, 5, 7, 10, 15. WEST ratio, as defined by Dar-EI [19], measures
the average number of activities per station. For each problem, the number of
station is equal to the number of robots, and each task can be processed on
any robot. The task time data are generated at random, while two statistical
dependence are maintained: (1) statistical dependence of task times on the task
type, (2) statistical dependence of task times on the robot on which the task is
processed.
To validate the effectiveness of our hybrid GA, we compare our approach with
Levitin et al.’s two algorithms. Recently, Levitin et al. [18] develop an efficient
approach for rALB problems. It is similar research with our purpose, aims to
achieve a balanced distribution of work between different stations (balance the
line) while assigning to each station the robot best fit for the activities assigned
to it. Levitin et al. proposed two algorithms named as recursive assignment
method and consecutive assignment method. If we look each robot as one type
and deleting this type from the total types of robots when the robot is allocated
for a station, Levitin et al.’s two algorithms can be adapted to solve the 32 rALB-
2 problems here. Hence, the two algorithms proposed by Levitin et al. are also
used to solve the 32 problems. The adopted parameters of the hGA are listed as
following: maximal generations maxGen = 1000; population size popSize = 100;
74 L. Lin et al.
Table 1. Performance of the proposed algorithm
Test problems Cycle time (CT ) Test problems Cycle time (CT )
of of West Levitin Levitin Proposed of of West Levitin Levitin Proposed
task station ratio et al.’s et al.’s approach task station ratio et al.’s et al.’s approach
recursive consecutive recursive consecutive
25 3 8.33 518 503 503 89 8 11.13 638 505 494
4 6.25 351 330 327 12 7.42 455 371 370
6 4.17 343 234 213 16 5.56 292 246 236
9 2.78 138 125 123 21 4.24 277 209 205
35 4 8.75 551 450 449 111 9 12.33 695 586 557
5 7.00 385 352 344 13 8.54 401 339 319
7 5.00 250 222 222 17 6.53 322 257 257
12 2.92 178 120 113 22 5.05 265 209 192
53 5 10.60 903 565 554 148 10 14.80 708 638 600
7 7.57 390 342 320 14 10.57 537 441 427
10 5.30 35 251 230 21 7.05 404 325 300
14 3.79 243 166 162 29 5.10 249 210 202
70 7 10.00 546 490 449 297 19 15.63 1129 674 646
10 7.00 313 287 272 29 10.24 571 444 430
14 5.00 231 213 204 38 7.82 442 348 344
19 3.68 198 167 154 50 5.94 363 275 256
exchange crossover probability pC = 0.20; mixed crossover probability p1M =

0.80; task sequence mutation probability p2M = 0.05; robot assign. mutation
probability p3M = 0.10. Levitin et al.’s two algorithms used the same population
size and maximal generations as the hGA.
Table 1 shows the experiment results of Levitin et al.’s two algorithms and
proposed approach for 32 different scale test problems. As depicted in Table 1, all
of results by proposed approach are obvious better than Levitin et al.’s recursive
assignment method. And most results of proposed approach are better than and
Levitin et al.’s consecutive assignment method, except to the test 25-3, test 35-7
and test 111-17. The differences of 32 test problems between proposed approach
with Levitin et al.’s approaches are showed in Fig. 8.
To clearly see the convergence performance of proposed approach, we give
the evolutionary process of proposed approach and Levitin et al.’s two methods
by combining 3 different scale test problems: small-scale test 35-12, middle-scale
test 111-17 and large-scale test 148-21. The evolutionary processes are showed in
Fig. 9. As depicted in Fig. 9, the convergence of proposed approach is faster than
each of other approaches. The difference of convergence increased with increasing
the test problems size. Furthermore, though Levitin et al.’s consecutive found the
same result with proposed approach for test 111-17, but the convergence speed
of proposed approach was faster. The simulations show that propose hybrid GA
with combining local search approach is computationally efficient and effective
to find the best solution and convergence speed.
Fig. 8. The difference between different approaches
Recursive assign.
Recursive assign.
Consecutive assign.
Consecutive assign. Proposed approach
Proposed approach
0 250 500 750 1000

0 250 500 750 1000
(a) test 35-12 (b) test 111-17
Recursive assign.
Proposed approach Consecutive assign.
0 250 500 750 1000
(c) test 148-21
Fig. 9. Evolutionary process of 3 different approaches
6 Conclusion
Robotic assembly line has been an important manufacturing system in the mod-
ern era. The objective of this work is to develop an efficient solution for the
robotic assembly line balancing problem. This solution aims to achieve a bal-
anced distribution of work between different stations and assign to each station
the robot best fit for the activities assigned to it. The result of such solution
would be an increased production rate of the line.
A new representation method adapting the GA to the rALB-II problem
is proposed. Advanced genetic operators adapted to the specific chromosome
76 L. Lin et al.
structure and the characteristics of the rALB problem are used. In order to
strengthen the search ability, two kinds of local search are integrated under the
framework the genetic algorithm. The coordination among the three kinds of
local search is well considered. The neighborhood structure of the local search can
be adjusted dynamically. The balance between genetic search and local search
is also investigated. The performance of proposed method is validated through
simulation experiments. The simulation shows that our algorithm is computa-
tionally efficient and effective to find the best solution. The solution obtained by
our algorithm outperforms the results from previous works.
Acknowledgements. This work is partly supported by the National Natural Science

References
1. ALB: Assembly line balancing: Data sets & research topics (2016). http://www.
assembly-line-balancing.de/
2. Anderson EJ, Ferris MC (1994) Genetic algorithms for combinatorial optimization:
the assemble line balancing problem. Informs J Comput 6(2):161–173
3. Bautista J, Suarez R et al (2000) Local search heuristics for the assembly line
balancing problem with incompatibilities between tasks. In: IEEE International
Conference on Robotics and Automation, 2000, Proceedings, ICRA, pp 2404–2409
4. Brown EC, Sumichrast RT (2005) Evaluating performance advantages of grouping
genetic algorithms. Eng Appl Artif Intell 18(1):1–12
5. Daoud S, Chehade H et al (2014) Solving a robotic assembly line balancing problem
using efficient hybrid methods. J Heuristics 20(3):235–259
6. Falkenauer E, Delchambre A (1997) A genetic algorithm for bin packing and line
balancing. In: IEEE International Conference on Robotics and Automation, 1992,
Proceedings, vol 2, pp 1186–1192
7. Gao J, Sun L et al (2009) An efficient approach for type II robotic assembly line
balancing problems. Comput Ind Eng 56(3):1065–1080
8. Gen M, Cheng R (1997) Genetic Algorithms and Engineering Design. Wiley, New
York
9. Gen M, Cheng R (2000) Genetic Algorithms and Engineering Optimization. Wiley,
New York
10. Gen M, Tsujimura Y, Li Y (1996) Fuzzy assembly line balancing using genetic
algorithms. Comput Ind Eng 31(31):631–634
11. Gonçalves JF, Almeida JRD (2002) A hybrid genetic algorithm for assembly line
balancing. J Heuristics 8(6):629–642
12. Gutjahr AL, Nemhauser GL (1964) An algorithm for the line balancing problem.
Manage Sci 11(2):308–315
13. Helgeson WB, Salveson ME, Smith WW (1954) How to Balance an Assembly Line.
Technical Report Carr Press, New Caraan
14. Janardhanan MN, Nielsen P, Ponnambalam SG (2016) Application of particle
swarm optimization to maximize efficiency of straight and U-shaped robotic assem-
bly lines. Springer
15. Kim YK, Kim YJ, Kim Y (1996) Genetic algorithms for assembly line balancing
with various objectives. Comput Ind Eng 30(3):397–409
16. Kock S, Vittor T et al (2011) Robot concept for scalable, flexible assembly automa-
tion: a technology study on a harmless dual-armed robot. In: IEEE International
Symposium on Assembly and Manufacturing, pp 1–5
17. Krasnogor N, Smith J (2000) A memetic algorithm with self-adaptive local search:
TSP as a case study. In: Proceedings of 2000 Genetic and Evolutionary Computa-
tion Conference, pp 987–994
18. Levitin G, Rubinovitz J, Shnits B (2006) A genetic algorithm for robotic assembly
line balancing. Eur J Oper Res 168(3):811–825
19. (Mansoor) EMDE (1973) Malbla heuristic technique for balancing large single-
model assembly lines. IIE Trans 5(4):343–356
20. Mukund Nilakantan J, Ponnambalam SG et al (2015) Bio-inspired search algo-
rithms to solve robotic assembly line balancing problems. Neural Comput Appl
26(6):1379–1393
21. Nearchou AC (2007) Balancing large assembly lines by a new heuristic based on
differential evolution method. Int J Adv Manufact Technol 34(9):1016–1029
22. Rubinovitz J, Levitin G (1995) Genetic algorithm for assembly line balancing. Int
J Prod Econ 41(41):343–354
23. Rubinovitz J, Levitin G (1995) Genetic algorithm for line balancing. Int J Prod
Econ 41:343–354
24. Sabuncuoglu I, Erel E, Tanyer M (2000) Assembly line balancing using genetic
algorithms. J Intell Manufact 11(3):295–310
25. Scholl A (1993) Data of assembly line balancing problems. Schriften zur Quanti-
tativen Betriebswirtschaftslehre
26. Scholl A (1999) Balancing and Sequencing of Assembly Lines. Physica-Verlag, Hei-
delberg
27. Scholl A, Fliedner M, Boysen N (2010) Absalom: balancing assembly lines with
assignment restrictions. Eur J Oper Res 200(3):688–701
28. Tsujimura Y, Gen M, Kubota E (1995) Solving fuzzy assembly-line balancing
problem with genetic algorithms. In: International Conference on Computers and
Industrial Engineering, pp 543–547
29. Yoosefelahi A, Aminnayeri M et al (2012) Type II robotic assembly line balancing
problem: an evolution strategies algorithm for a multi-objective model. J Manufact
Syst 31(2):139–151
A Hybrid Model of AdaBoost and
Back-Propagation Neural Network
for Credit Scoring
Feng Shen1(B) , Xingchao Zhao1 , Dao Lan2 , and Limei Ou2

1
School of Finance, Southwestern University of Finance and Economics,
shenfeng1213@gmail.com
2
Southwestern University of Finance and Economics,
Abstract. Owing to the development of internet finance in China, credit

scoring is growing into one of the most important issues in the field of
financial risk management. Quantitative credit scoring models are widely
used tools for credit risk assessment in financial institutions. In this study,
an AdaBoost algorithm model based on back-propagation neural net-
work for credit scoring with high accuracy and efficiency is proposed.
We first illustrate the basic concepts of back-propagation neural network
and AdaBoost algorithm and propose a hybrid model of AdaBoost and
back-propagation neural network, then two real-world credit data sets
are selected to demonstrate the effectiveness and feasibility of the pro-
posed model. The results show that the proposed model can get higher
accuracy compared to other classifiers listed in this study.
Keywords: Credit scoring · AdaBoost model · Back-propagation neural

network
1 Introduction
Credit scoring has grown an increasingly important issue of financial risk man-
agement in financial institutions since the 2008 financial crisis. It often calculates
by following a set of decision models and other underlying technologies, does
a great favor for lenders’ judging whether an application of credit should be
approved or rejected [27]. When some applicants fail to repay their debt, it leads
to a direct economic loss for the lenders. In addition, the sub-prime mortgage
crisis occurred in the USA has caused some financial institutions loss billions
of dollars due to customers’ default. However, if a credit-granting-institution
rejects all applicants of loans even with good credit scores, it will suffer the
potential revenues it can earn from the applicants in the future. Therefore, an
efficient decision support with high accuracy becomes a clear need for financial
institution.
AdaBoost and Back-Propagation Neural Network 79
Due to the great importance of credit risk assessment, an increasing research

stream is focusing upon credit risk assessment and credit scoring using different
methods and models, approaches such as linear discriminant analysis (LDA)
[1], logistic regression analysis LRA [26], k-nearest neighbor (KNN) [12,13] and
decision tree [5] which uses the statistics disciplines. With the development of
artificial intelligence(AI) techniques, artificial neural network (ANN) [3,8,14,21,
30,31], support vector machines (SVM) [2,11,25,34], genetic algorithm (GA)
[7,22,28], rough set [4] and case-based reasoning [15–19] are used for credit
scoring. Also some combined and ensemble approaches perform quite well with
high accuracy and efficiency, including fuzzy system and artificial neural network
[20,23], rough set and support vector machine [33], fuzzy system and support
vector machines [29], case-based reasoning and support vector machines [19],
neural network ensemble [32] etc.
Inspired by the combined and ensemble theories, this study attempts to pur-
pose an ensemble AdaBoost model based on BP neural network for credit scoring.
The main idea of AdaBoost algorithm is to maintain a distribution of weights
over the training samples and adjust them after each basic classifier sorting cycle
adaptively. In our study, we use ten single back propagation neural networks as
the weak learners contends of a three-layer feed-forward network each, and the
final strong ensemble classifier is constructed by the method of weighted voting
associated with the performances of the weak learners. Based on the experi-
mental results achieved from two public available credit data sets, our proposed
AdaBoost model based on BP neural network obtains a good performance with
higher accuracy efficiency compared to other models used in this study, which
indicates a wide prospect usage in financial risk management.
The rest of this paper is organized as follows. Section 2 gives a brief formula-
tion concerns about back propagation neural network and AdaBoost algorithm.
Section 3 presents a hybrid model of AdaBoost and BP neural network. To ver-
ify the accuracy and effectiveness of the purposed model, empirical validation
results of the model using the German credit data set and Australian credit
data set are analyzed in Sect. 4. Finally, a short conclusion and discussion are
presented in Sect. 5.
2 Methodology Formulation
2.1 Back-Propagation Neural Network Theory
Back-propagation neural network model was proposed by Rumelhart and Mccel-

land in 1985 [24], it is a kind of multi-layered forward feed type error counter-
biography neural network. This model usually consists of input layer, out-
put layer and hidden layer. Each layer includes certain nodes, and each node
expresses a neuron. There exists interconnection among the nodes of different
layers, but there is no connection among nodes in the same layer. Among them,
the single hidden layer BP network’s application is the most common. The typ-
ical structure of BP neural network is as shown in Fig. 1.
80 F. Shen et al.
Fig. 1. Structure of single hidden layer BP neural network
The core thought of BP neural network can be described as following: training

signal forward-propagating and error signal reverse dissemination. During the
process of forward-propagating, the signal spreads from input layer, dealt with
connection weights between input layer and hidden layer, then is transformed
by the active function of hidden layer and transmitted to the output layer. If
expected output can’t be achieved from the output layer, then another process of
error back propagation begins, which would continuously adjust the weights and
bias of each layer in order to ultimately minimize the error of the network. This
kind of signal forward-propagating and error back-propagating is to go iteration
and iteration till the system output error is reduced to an acceptance degree or
obtains the network constraints. Its concrete process is shown in Fig. 2.
Fig. 2. Process of BP neural network model

2.2 Basic Theory of AdaBoost Algorithm

AdaBoost is a very popular boosting algorithm for binary classification, and
it solved many practical difficulties of the earlier boosting algorithms. It was
initially proposed by Yoav Freund and Robert Schapire in 1997 [10]. This algo-
rithm constructs a high-quality compositive classifier by combining trained weak
classifiers sequentially while putting more attention on those weak classifiers
accompanied with good performance. There are several methods for combining
results from base weak learners into one stronger predictor. Uniform voting, dis-
tribution summation, Bayesian combination, etc. are widely used for ensemble
algorithms. Adaptive boosting algorithm has been applied in many fields such as
speech recognition [9], moving vehicle classification based on images [6] and etc.
The main ideas of AdaBoost algorithm is to maintain a distribution of weights
over the training samples and adjust them after each basic classifier sorting cycle
adaptively. The weights of training data which are wrongly classified by current
weak learner will be increased, otherwise, decreased if the samples are correctly
classified. During the training process, the prediction error of the weak classi-
fiers should be less than 0.5. And the voting weights of base classifiers will be
increased while the decreasing of the prediction error, which means a larger vot-
ing weight this weak learner will take in the final ensemble output. Pseudocode
for AdaBoost algorithm is shown in Fig. 3. As a series of base learners have
been achieved, AdaBoost calculates a value αt that is assigned to ht , and the
final hypothesis H is constructed via T weak classifiers using a weighted voting
ensemble method.
Fig. 3. Pseudocode for AdaBoost algorithm

82 F. Shen et al.
3 Hybrid Model of AdaBoost and BP Neural Network

In this section, a hybrid model is constructed based on AdaBoost and BP neural
network for credit scoring. According to the AdaBoost algorithm, the efficiency
and accuracy of the ensemble predictor based on a series of weak classifiers
largely depends on the accuracy of those base learners, the correlation coeffi-
cient of the ensemble predictor and the base classifiers is directly proportional.
As BP neural network model is very fledged both in theory and practical appli-
cations with high accuracy and efficiency compared to other widely used models.
So, in this study, BP neural network is selected as the weak classifier, and as
the final ensemble classifier requires a certain method to combine those weak
learners into an efficient ensemble system, we adopt the weighted voting method
to combine the outputs of each weak classifier for the final outputs. The frame
and pseudocode of hybrid model based on AdaBoost and BP neural network are
shown in Figs. 4 and 5. The main steps used to construct ensemble predictor are
shown as follows.
Step 1. Select the training samples randomly from the chosen databases with
the method of 10-fold cross-validation, and each group training sample
is assigned with the same weights:
T = {(x1, y1), · · · , (xi, yi), · · · , (xn, yn)} while xi ∈ X, yi ∈ Y = {−1, +1}.
The weights of this distribution on training samples on round t is

denoted D1(i). Initially, all weights are assigned with equal value D1(i):
1
D1(i) = , i = 1, 2, · · · , n, (1)
n
where n represents the total number of samples in the training data set;
Step 2. Parameters of the BP neural network, which include the structure of the
network, neurons of each layer, expected goal, are set up according to
the attributes of the input and output training samples. Then initialize
the weights and biases of BP neural network.
Step 3. Training the BP neural network with the processed databases:
Step 3.1 Calculate input and output values of hidden layer: netj , bj
n

netj = wij × xi − θj , (2)
i=1
bj = f (netj ), (3)
where wij is the connection weights between input layer and

hidden layer and θj represents the Threshold values of hidden
layer, f (x) represents the transfer function Sigmoid. Namely:
1
f (net) = . (4)
1 + e−net
Fig. 4. The frame of hybrid model of AdaBoost and BP neural network
Fig. 5. Pseudocode of hybrid model of AdaBoost and BP neural network
Step 3.2 Calculate input and output values of output layer:

m

neti = vjt × bj − γi , (5)
j=1
oi = f (neti ), (6)
84 F. Shen et al.
where vjt is the connection weights between hidden layer and

output layer and γi represents the Threshold values of output
layer. In this study, the hidden layer contains m neurons.
Step 3.3 Calculate the error di of each neuron of output using network’s
goal vector and its response output:
di = yi − oi , (i = 1, 2, · · · , n). (7)
Step 3.4 Calculate the total error of the network e(k):

n
1 2
e(k) = (di ) . (8)
2 i=1
Step 3.5 If the total error above is acceptable, the process stops. Oth-
erwise, revise and. There are many ways of weight changes, we
use the most common gradient descent method:
k+1 k k k−1
wij = wij + Δwij + ∂Δwij , (9)
k+1 k k k−1
wjt = wjt + Δwjt + ∂Δwjt . (10)
This process continuous till the system output error is reduced

to an acceptance degree or obtains the network constraints.
Step 4. When training epochs go to round t, a base classifier ht (x) could
be achieved with the weights distribution Dt and then calculate the
weighted error εt from model:
n

εt = Dt (i) × I(ht (xi )), (11)
i=1
where I(ht (xi )) is the indicator function, and its mathematical expres-
sion shows as below:

yi ,
1, if ht (xi ) =
I(ht) = (12)
0, if ht (xi ) = yi .
Step 5. Calculate the weight of base classifier ht (x)

1 1 − εt
at = ln . (13)
2 εt
As is shown above, if the prediction error of the samples is less than 0.5,
the weight is bigger than zero, and with the decreasing of the prediction
error, will be increased meanwhile, which means a larger voting weight
of this weak learner in the final ensemble output.
Step 6. Update the weights of each samples in the training databases for the
next iteration:
Dt (i)
Dt+1 (i) = exp(−at × yi × ht (xi )), (14)
Zt
where Zt is normalization constant for Dt+1 (i) become a proper distri-

bution:
n

Zt = Dt (i) × exp(−at × yi × ht (xi )). (15)
i=1
In this way, the weights of training data which are wrongly classified by
current weak learner will be increased or decreased if the samples are
correctly classified.
Step 7. As the training iteration goes on, a series of base classifiers will be
obtained, then combine these classifiers during the processes:
T

f (x) = at × ht (x). (16)
t=1
A series of base classifiers with diversity are collected with loops from
step 2 to step 5. Finally, an AdaBoost model based on the weak learners
trained above is constructed:
T
H(x) = sign( at × ht (x)) (17)
t=1
4 Empirical Analysis
In order to test the performance of the proposed AdaBoost algorithm based

on BP neural network model, two available real world credit data sets with
detailed input attributes description (German and Australian credit data sets)
are used in this study. These two credit databases are available from open access
UCI Machine Learning Repository. The German credit data set consists of 1000
instances with 700 samples labeled as creditworthy and 300 samples classified
as poor credit. For each instance, there are 24 input variables described 19
attributes with 4 attributes changed to dummy variables. Also, the 25th variable
is the label of the instance with two numerical descriptions, and label 1 stands
for a worthy credit assessment of current instance, label 2, on the contrary, rep-
resents a bad credit evaluation. The meaning of original attributes are described
in Table 1; The Australian credit data set is interesting because there is a good
mix of attributes – continuous, nominal with small numbers of values, and nomi-
nal with larger numbers of values, which includes 690 instances in total with 383
samples classified as trustworthy and 307 samples labeled as poor credit. For
each instance, 15 variables described 15 features of personal information and
financial history of applicants, the last feature is labeled as approved (marked as
0) or rejected (marked as 1). All attribute names and values have been changed
to meaningless symbols to protect confidentiality of the data.
86 F. Shen et al.
Table 1. Original attributes in the German credit data set
Number Description Class

Attribute 1 Status of existing checking account Qualitative
Attribute 2 Duration in month Numerical
Attribute 3 Credit history Qualitative
Attribute 4 Purpose Qualitative
Attribute 5 Credit amount Numerical
Attribute 6 Savings account/bonds Qualitative
Attribute 7 Present employment since Qualitative
Attribute 8 Instalment rate in percentage of disposable income Numerical
Attribute 9 Personal status and sex Qualitative
Attribute 10 Other guarantors Qualitative
Attribute 11 Present residence since Numerical
Attribute 12 Property Qualitative
Attribute 13 Age in years Numerical
Attribute 14 Other instalment plans Qualitative
Attribute 15 Housing Qualitative
Attribute 16 Number of existing credits at this bank Numerical
Attribute 17 Job Qualitative
Attribute 18 Number of people being liable to provide maintenance for Numerical
Attribute 19 Telephone Qualitative
Attribute 20 Foreign worker Qualitative
4.1 Experiment Design
In the experiment, the AdaBoost algorithm model based on BP neural network

is used. The final classifier contents ten weak learners constructed with a three-
layer feed-forward BP neural network model. As to the neurons in hidden of each
weak learner, 9 nodes in hidden layer dealt with German data set and 20 nodes in
hidden layer in Australian data set are set. For comparison purpose, some com-
monly used models, such as linear discriminant analysis (LDA), quadratic dis-
criminant analysis (QDA), logistic analysis (Log A), K-Nearest neighbor (KNN),
artificial neural network (ANN), back propagation neural network (BPNN), and
least squares support vector machine (LSSVM) are also applied. In this study,
training samples are selected by the method of 10-fold cross-validation to deter-
mine the model parameters. In BP neural network model, a common three-layer
feed forward net trained by Levenberg-Marquartdt algorithm is employed, the
transfer function in hidden layer is applied with Sigmoid function, and with
Purelin function in output layer. In addition, three evaluation criteria, Type I
Accuracy, Type II Accuracy and Total Accuracy are used, which are defined as
follows [34]:
number of both observed bad and classified as bad

Type I Accuracy = Specificity = , (18)
number of observed bad
number of both observed good and classified as good
Type II Accuracy = Sensitivity = ,
number of observed good
(19)
number of correct classification
Total Accuracy = . (20)
number of evaluation sample
4.2 Experiment Result

As is shown in Tables 2 and 3, validation results for the German credit data set
and Australian data set are achieved using the different algorithms and settings
described above. The classification results of the first six classifiers in Tables 2
and 3 are from another article [35]. From Tables 2 and 3, several interesting
findings can be drawn.
First of all, the total accuracy of the proposed AdaBoost model based on
BP neural network is much better than other classifiers in both German and
Australian data sets. Followed by the least squares support vector machine.
Probable explaining reasons of the high accuracy achieved by AdaBoost model
includes two aspects. For one thing, the accuracy of AdaBoost model depends
much on the accuracy of the weak learners which it ensembles, and as we can see
from Tables 2 and 3, BP neural network do a good credit classification compared
to other commonly used models. For another reason, the main idea of AdaBoost
system gives larger weights to those base classifiers with small prediction error,
and as a result, the total accuracy of AdaBoost model get the best performance
among all the classifiers compared in this study.
Second, referring to Type I Accuracy, the presented AdaBoost model based
on BP neural network performs the best, compared to other classifiers employed
in this study both in German credit data set and Australian credit data set.
Followed by the linear discriminant analysis model in German data set and
logistic analysis model in Australian data set.
Third, in terms of Type I Accuracy and Type II Accuracy, all models used
in this study perform better with higher accuracy in Type II Accuracy than
Type I Accuracy in general, which indicates that it’s more difficult to classify
customers with bad credit conditions from those of good credit evaluation cause
of the complexity of credit risk.
Fourth, according to the Type II Accuracy shown in Table 2, our proposed
model only ranks the fourth compared to other models for comparison. Supposed
reason of this phenomenon may be the default of the AdaBoost algorithm, there
is no exotic existence of noisy data in German data set, which will attract more
attention of our AdaBoost model based on BP neural network on these noisy
signals, as a result, the accuracy of our proposed model may not be satisfied as
we wish and thus performs ordinarily in German credit data set.
Finally, compared with Tables 2 and 3, it’s easy to find out that the perfor-
mance of the German data set is worse than that of the Australian data set.
88 F. Shen et al.
Table 2. Credit evaluation results comparison of different models for German credit
dataset
Models Type I Accuracy Type II Accuracy Total Accuracy

% Rank % Rank % Rank
LDA 72 2 74.57 6 73.8 5
QDA 66.57 3 69.33 8 67.4 8
Log A 50.33 5 88.14 3 76.8 3
KNN 27 8 90.57 1 71.5 6
ANN 46.89 7 73.46 7 69.43 7
BPNN 63.83 4 79.36 5 75.8 4
LSSVM 49.67 6 88.86 2 77.1 2
BPNN AdaBoost 78.7 1 84.19 4 83 1
Table 3. Credit evaluation results comparison of different models for Australian credit
dataset
Models Type I Accuracy Type II Accuracy Total Accuracy

% Rank % Rank % Rank
LDA 80.94 6 92.18 2 85.94 5
QDA 66.12 8 91.38 3 80.14 6
Log A 85.9 2 86.32 6 86.09 4
KNN 81.72 5 54.4 8 69.57 8
ANN 72.56 7 83.61 7 78.94 7
BPNN 83.95 4 87.65 5 86.23 3
LSSVM 85.12 3 89.25 4 86.96 2
BPNN AdaBoost 89.06 1 94.59 1 92.03 1
There are two possible reasons. On one hand, the credit market in Germany is
more complex than that in Australia. On the other hand, there is more non-
linearity in the German data set than in the Australian data set. Overviewing
the above performance, one interesting phenomenon shows that there is a great
capability raising of all three evaluation criteria both in the German data set
and Australian data set compared BP neural network with the AdaBoost model
based on BP neural network, from the viewpoint of Total Accuracy, there is
a 9.5% raising in the German data set and a 6.7% raising in the Australian
data set.
5 Conclusion
In this paper, a hybrid model of AdaBoost and BP neural network is proposed
for credit risk evaluation. According to the empirical results, we find that our
proposed hybrid model is the best one compared with other seven models for
two publicly available credit data sets, which indicates that our proposed hybrid
model of AdaBoost and BP neural network has a good practicability for credit
scoring. In the future, we will use other method as base classifiers, for example,
support vector machine and decision tree, and research other ensemble algorithm
like bagging algorithm for credit risk assessment.
Acknowledgements. This research was supported by the project of Research Center

for System Sciences and Enterprise Development (Grant No.Xq16C03).
References
1. Altman EI (1968) Financial ratios, discriminant analysis and the prediction of
corporate bankruptcy. J Finan 23(4):589–609
2. Baesens B, van Gestel T et al (2003) Benchmarking state-of-art classification algo-
rithms for credit scoring. Oper Res Soc 54(6):627–635
3. Baesens B, Setiono R et al (2003) Using neural network rule extraction and decision
tables for credit-risk evaluation. Manag Sci 49(3):312–329
4. Beynon MJ, Peel MJ (2001) Variable precision rough set theory and data discreti-
sation: an application to corporate failure prediction. Omega 29(6):561–576
5. Carter C, Catlett J (1987) Assessing credit card applications using machine learn-
ing. IEEE Expert 2(3):71–79
6. Chen M, Ma L, Gao Y (2010) Vehicle detection segmentation based on adaboost
and grabcut. In: IEEE International Conference on Progress in Informatics and
Computing, pp 896–900
7. Chen MC, Huang SH (2003) Credit scoring and rejected instances reassigning
through evolutionary computation techniques. Expert Syst Appl 24(4):433–441
8. Desai VS, Crook JN, Overstreet J (1996) A comparison of neural networks and
linear scoring models in the credit union environment. Oper Res 95(2):24–37
9. Essa EM, Tolba AS, Elmougy S (2008) A comparison of combined classifier archi-
tectures for arabic speech recognition. In: International Conference on Computer
Engineering & Systems, pp 149–153
10. Freund Y, Schapire RE (1995) A decision-theoretic generalization of on-line learn-
ing and an application to boosting. In: European Conference on Computational
Learning Theory, pp 119–139
11. Gestel TV, Baesens B et al (2003) A support vector machine approach to credit
scoring. Banken Financiewezen 2:73–82
12. Henley WE, Hand DJ (1996) A k-nearest-neighbour classifier for assessing con-
sumer credit risk. J Roy Stat Soc 45(1):77–95
13. Henley WE, Hand DJ (1997) Construction of a k-nearest-neighbour credit-scoring
system. IMA J Manag Math 8(4):305–321
14. Khashman A (2009) A neural network model for credit risk evaluation. Int J Neural
Syst 19(4):285–294
15. Li H, Sun J (2008) Ranking-order case-based reasoning for financial distress pre-
diction. Knowl Based Syst 21(8):868–878
16. Li H, Sun J (2009) Gaussian case-based reasoning for business failure prediction
with empirical data in China. Inf Sci 179(1):89–108
17. Li H, Sun J (2009) Predicting business failure using multiple case-based reasoning
combined with support vector machine. Expert Syst Appl 36(6):10085–10096
90 F. Shen et al.
18. Li H, Sun J (2010) Business failure prediction using hybrid2 case-based reasoning
(h2cbr). Comput Oper Res 37(1):137–151
19. Li H, Sun J, Sun BL (2009) Financial distress prediction based on or-cbr in the
principle of k-nearest neighbors. Expert Syst Appl 36(1):643–659
20. Malhotra R, Malhotra DK (2002) Differentiating between good credits and bad
credits using neuro-fuzzy systems. Eur J Oper Res 136(1):190–211
21. Malhotra R, Malhotra DK (2003) Evaluating consumer loans using neural net-
works. Soc Sci Electron Publishing 31(2):83–96
22. Ong CS, Huang JJ, Tzeng GH (2005) Building credit scoring models using genetic
programming. Expert Syst Appl 29(1):41–47
23. Piramuthu S, Piramuthu S (1999) Financial credit-risk evaluation with neural and
neurofuzzy systems. Eur J Oper Res 112(2):310–321
24. Rumelhart DE, Hinton GE, Williams RJ (1986) Learning representations by back-
propagating errors. Nature 323(6088):533–536
25. Schebesch KB, Stecking R (2005) Support vector machines for classifying and
describing credit applicants: detecting typical and critical regions. J Oper Res Soc
56(9):1082–1088
26. Steenackers A, Goovaerts MJ (1989) A credit scoring model for personal loans.
Insur Math Econ 8(1):31–34
27. Thomas LC, Edelman DB, Crook JN (2002) Credit scoring and its applications.
SIAM, Philadelphia
28. Varetto F (1951) Genetic algorithms in the analysis of insolvency risk. University
of Illinois Press, Champaign
29. Wang Y, Wang S, Lai KK (2006) A new fuzzy support vector machine to evaluate
credit risk. IEEE Trans Fuzzy Syst 13(6):820–831
30. West D (2000) Neural network credit scoring models. Comput Oper Res
27(11):1131–1152
31. Yobas MB, Crook JN, Ross P (2000) Credit scoring using neural and evolutionary
techniques. IMA J Manag Math 11(2):111–125
32. Yu L, Wang S, Lai KK (2008) Credit risk assessment with a multistage neural
network ensemble learning approach. Expert Syst Appl 34(2):1434–1444
33. Yu L, Wang SY et al (2008) Designing a hybrid intelligent mining system for credit
risk evaluation. Syst Sci Complex 21(5):527–539
34. Yu L, Wang S, Cao J (2009) A modified least squares support vector machine
classifier with application to credit risk analysis. Int J Inf Technol Decis Making
08(4):697–710
35. Zhou L, Lai KK, Yu L (2009) Credit scoring using support vector machines with
direct search for parameters selection. Soft Comput 13(2):149–155
Effects of Urban and Rural Residents’ Behavior
Differences in Sports and Leisure Activity:
Application of the Theory of Planned Behavior
and Structural Equation Modeling
Linling Zhang1,2(B)
1
School of Business, Sichuan University,
1010519067@qq.com
2
Leisure Sports Department, Chengdu Sport Institute,
Abstract. Because the urban and rural residents’ sports and leisure
behaviors were different, to find out the factors influencing these differ-
ences was the purpose of this study. Therefore we selected urban and
rural residents in Sichuan province as the research object to start our
study. This study which was based on the Theory of Planned Behav-
ior (TPB) used Structural Equation Modeling (SEM) as the research
method. Firstly we constructed the conceptual model and put forward a
series of assumptions. Then we used questionnaire survey to obtain first-
hand information, at the same time we used ARMOS17.0 to work out
the model. At last we verified our original assumption with the model.
Research showed that: (1) The influence on the urban and rural resi-
dents’ sports and leisure behavior from external objective factors (social
atmosphere, venues, facilities, etc.) was much larger than that from the
intrinsic objective factors (leisure time, income, physical condition, etc.);
(2) The richness of sports and leisure activities, venues, facilities and the
social atmosphere impacted the behavior of urban residents much more
than that of rural residents; (3) The attitude to affirm and recognize
sports and leisure had much more influence on urban residents; (4) While
rural residents more cared about the views from family, friends even gov-
ernment.
Keywords: Sports and leisure activity · Urban and rural residents ·

Effects · Theory of planned behavior · Structural equation modeling
supplier selection
1 Introduction
Under the Chinese specific dual social structure between urban and rural
areas, social and economic activities, material conditions, living environment,
92 L. Zhang
life style, cultural consciousness between urban and rural areas appear a huge
differentiation [5,24]. In terms of sports and leisure activity, social system differ-
ences between the urban and rural result in gaps which come from capital invest-
ment, facilities construction, social organization development and concepts. And
the leisure sports behavior of urban and rural residents also shows difference.
What’s more, with the accelerating of urbanization process, urban and rural
demographic differences will also expand the difference of leisure sports behav-
ior [11]. Therefore, the purpose of this study is trying to find out the main factors
which cause huge difference and gap between urban and rural residents, and to
think about how to close the gap.
Sports and leisure behavior is a dynamic process that people try to meet
their leisure needs to participate in sports activities spontaneously and get sat-
isfaction from it. The purpose that people take part in leisure sports activities
mainly includes physical aspect (exercise, beauty fitness, self-defense, etc.) and
emotional aspect (entertainment, relieve pressure, thrill-seeking, social communi-
cation, etc.) [19]. Western researches about sports and leisure behavior is mainly
focused on the influence of behavior by using various theories. Trans-theoretical
Model (TTM), Self-efficacy, Self-determination Theory (SDT), Leisure Con-
strains Theory, Theory of Planned Behaviors (TPB) are the representative the-
ories [3,4,6–9,15,16,18]. In 2006, Xu [21] who published a paper “The Study
on the characteristics of the Residents’ Leisure Sports Behaviors” in the Jour-
nal of Guangzhou Sport University opened the prelude of the domestic sports
and leisure behavior research. Qiu [10,12–14] who published many articles in the
Chinese authoritative core journals is very outstanding in the study of sports
and leisure behavior. It is worth mentioning Yu [23], Zhao [25], Ye [22] and oth-
ers who introduced the theory of planned behavior (TPB) into the sports and
leisure behavior study. Overall, the domestic scholars’ research mainly focused
on cities and towns and little focused on rural areas and it caused the lacking of
empirical studies in such areas.
It is very hard to find out a clear answer from existing researches in the face
of the question what is the deep reason that causes sports and leisure behav-
ior differences between the urban and rural residents. In order to answer this
question better, we introduce the Theory of Planned Behavior (TPB) and Struc-
tural Equation Modeling (SEM) as our theoretical and method guide and choose
Sichuan province as the sample to try to reveal the main influencing factors of
sports and leisure behavior of urban and rural residents and to analyze the
relationship between these factors. We try to work out a theoretical analysis
framework to explain the differences between urban and rural residents’ sports
and leisure behavior and this framework should play an important role to enrich
and expand the leisure science research. According to the results, we put forward
corresponding countermeasures and suggestions to improve the policy environ-
ment effectively and promote the development of urban and rural integration
better.
Effects of Urban and Rural Residents’ Behavior Differences 93
2 Theory and Hypothesis

2.1 Introduction to the Theory of Planned Behavior
Theory of Planned Behaviors (TPB) is one of the most famous theories about
the relationship between attitude and behavior in the field of social psychology.
Based on the Theory of Reasoned Action (TRA), Ajzen [1] published “The The-
ory of Planned Behavior” in 1991, it marked the born of the formal formation of
Planned Behavior Theory. In the TPB, the Behavior is affected by the Behavior
Intention or the Perceived Behavioral Control. And the Behavior Intention is
affected by the Attitude toward the Behaviors, Subjective Norm and Perceived
Behavioral Control, at the same time Behavioral beliefs, Normative beliefs and
Control beliefs respectively affect the Behavioral Attitude, Subjective Norm and
Perceived Behavioral Control.
The effects of sports and leisure behavior involve economic factors (income,
consumption idea), cultural factors (leisure concept, leisure atmosphere, leisure
time, etc.), environmental factors (Sports venues, policy guarantee, product pro-
motion, service level, etc.), etc. In this study we analyze problems from the per-
spective of the TPB and give out the main factors which affect the urban and
rural residents’ behavior. Below are the main factors.
(1) Attitude
If people keep positive attitudes on the sports and leisure activity, they will
actively participate in the activity. It is obvious that the attitude can affect the
behavior.
(2) Subjective Norm
It refers to whether the views and practices of people around will affect residents’
participation in sports and leisure activities.
(3) Perceived Behavioral Control
It refers to whether the past experience and current conditions of residents have
influence on their sports and leisure behaviors.
(4) Behavior Intention
It refers to if the residents are more willing to participate in sports and leisure
activities subjectively, the probability that they actually participate in sports
and leisure activities is greater. And the significant positive correlation between
behavior and intention is also confirmed by many relative researches.
(5) Other Objective Factors
According to interview results from Sichuan residents, the main reasons that
people do not attend sports and leisure activities include lack of enough time,
lack of appropriate sports ground, lack of specific skills for their favorite sports,
lack of information and lack of partner, etc.
94 L. Zhang
2.2 Conceptual Model and Assumptions
According to the TPB and the previous investigation results, we build our con-
ceptual model (Fig. 1). According to the key points in the model, we make the
following theoretical assumptions:
Hypothesis 1 (H1): The attitude of sports and leisure has a significant positive
influence on the behavior intention.
Hypothesis 2 (H2): The attitude of sports and leisure plays an intermediary role
between subjective norm and behavior intention.
Hypothesis 3 (H3): The attitude of sports and leisure behavior plays an inter-
mediary role between perceived behavioral control and behavior intention.
Hypothesis 4 (H4): Subjective norm has a significant positive influence on the
behavior intention.
Hypothesis 5 (H5): The perceived behavioral control has a significant positive
influence on the behavior intention.
Hypothesis 6 (H6): The peripheral objective factors have a significant positive
influence on the TPB system.
Hypothesis 6a (H6a): The peripheral objective factor has a significant positive
influence on the subjective norm.
Hypothesis 6b (H6b): The peripheral objective factors have a significant positive
influence on the attitude.
Hypothesis 6c (H6c): The peripheral objective factors have a significant positive
influence on the perceived behavioral control.
Subjective norm Demographic

variable
Peripheral
objective Attitude Intention
factors Behavior
Perceived behavior
control
Fig. 1. The conceptual model of factors influencing sports and leisure behavior based
on the TPB
3 Data Collection
3.1 Questionnaire Design
The influence factors of sports and leisure behavior mainly include the attitude,
subjective norm, perceived behavioral control, peripheral objective factors and
intention. The relevant information is collected through the questionnaire survey.
This questionnaire chooses the Likert Scale, each question uses a measurement
from strongly disagree (1 point) to strongly agree (5 points). All questions in
Questionnaire come from the discussion of our team and the experts’ suggestion.
Below are the questions designed according to the conceptual model and the
hypothesis (Table 1).
3.2 Data Collection

Empirical investigation in this study mainly involved Chengdu and other seven-
teen cities of Sichuan. We issued questionnaires two times. One is in spring of
2016 and the other is in summer of 2016.
For the first time, we mainly authorized students from Chengdu Sport Insti-
tute to issue the questionnaires in spring. During the vacation we issued 200
questionnaires through network and field investigation, and we received 62 effec-
tive questionnaires (43 ones from urban residents while only 19 ones from rural
residents). Unfortunately, the effective recycling rate was very low and the rea-
son was that: firstly, in the questionnaire we designed three filtering questions
to filter bad ones; Secondly, the types of sports and leisure activities in rural
areas are much less than these in urban areas, participation rate was also lower
than that in town. In addition, in recent year rural migrant workers continued
to increase in young adults, people left were only older and children. The older
and children left were not able to finish questionnaire, so the amount of effective
questionnaires was very small.
For the second time, we still authorized students from Chengdu Sport Insti-
tute to issue questionnaires in summer vacation. During the whole vacation we
issued 400 questionnaires. This time we did some adjustments to issue more ques-
tionnaires to the rural residents. At last we recycled 168 effective questionnaires
(83 ones from urban residents and 85 ones from rural residents).
To sum up, we issued 600 questionnaires out and recycled 230 effective ones
at last. Among the effective questionnaires, 126 ones were from urban residents
and the left 104 ones came from rural residents. Though sample size was not big
enough but it could provide supports for this study.
3.3 Validity and Reliability Analysis
(1) Validity Analysis

Validity analysis of the questionnaire mainly uses the Factor Analysis method.
We handle 20 questions in Table 1 as 20 variables which have influence on sports
96 L. Zhang
Table 1. The questionnaire items assessing the TPB variables
Variables Items (Question)

Attitude a1. It is a pleasure to take part in sports and leisure
activities
a2. Sports and leisure activities are the best choice
for entertainment and relaxation
a3. Sports and leisure can help me to make more
friends
a4. Sports and leisure is an important part of my life
a5. To participate in sports and leisure activities
needs to insist on a long time, for which you are
willing to give up some of the other things you like
Subjective norms a6. The family supports me to participate in sports
and leisure activities
a7. Friends / colleagues / classmates support and
appreciate me to participate in sports and leisure
activities
a8. When people see I take part in sports and leisure
activities, they think I’m a man of life taste
a9. The government encourages us to participate in
sports and leisure activities
Perceived behavior control a10. Enough time
a11. Good income
a12. Good physical condition
a13. Enough information
a14. It is very convenient to participate in sports and
leisure activities
Peripheral objective factors a15. The sports and leisure activities are abundant
a16. Enough ground and facilities for sports and
leisure activities
a17. Good atmosphere of sports and leisure
Intention a18. I am willing to participate in sports and leisure
activities under the existing conditions
a19. I will take part in sports and leisure activities in
the future
a20. I will persuade my family and friends together
to participate in sports and leisure activities
and leisure behavior. By using SPSS19.0 software to calculate, Kaiser-Meyer-

Olkin (KMO) is 0.871 and the significant probability of Bartlett Test of Spheric-
ity statistic is 0.000 (p < 0.01), both of them indicate the data is suitable for
factor analysis. Then we extract 5 common factors from the 20 variables, they
can account for 63.876% of the total variation. Through rotary, we can see the
factor-1 has a great influence on variables-a1, a2, a3, a4, a5, a18 and a19 (load
value > 0.5); Factor-2 has a great influence on variables-a6, a7, a9 and a20;
Factor-3 has a great influence on variables-a12, a15, a16 and a17; Factor-4 has a
greater influence on variables-a10 and a13; and factor-5 has a great influence on
variable-a11. While for variables a8 and a13, each of their load value on every
common factor is less than 0.5, so we have reason to delete them. In the next
step, we find that the item variable-a11 also should be removed through the tests
for every common factor respectively.
(2) Reliability Analysis
We do the internal consistency test for the variables of every factor with Cron-
bach’s alpha reliability analysis. And we find that 4 factors (attitude, subjective
norm, periphery objective factors and intention) pass the test. If we delete the
variables a11 and a12 in the factor-perceived behavioral control, the a-coefficient
increases significantly to 0.698. In order to pass the internal consistency test,
variables a11 and a12 should be deleted.
4 Modeling and Modification

4.1 Structural Equation Modeling
Structural Equation Modeling (SEM) is a common method in social science
research. In the field of social science, the causal relationships among variables
are complex sometimes, or these variables can’t be observed directly, therefore
traditional statistical methods can’t solve this problem well. Since the 1980s,
SEM has been developed quickly, it makes up for the deficiency of the tradi-
tional statistical method and becomes an important tool of multivariate data
analysis [2,20]. SEM is a kind of multivariate analysis technology based on the
regression. For it is of theory apriority, SEM is used to check the hypothesis
between observed variables and latent variables, generally 5 samples are enough
for every observed variable [17]. This research built the SEM of urban and rural
residents’ sports and leisure behavior influence factors with ARMOS17.0 soft-
ware (Fig. 2).
4.2 The Urban Model

(1) Calculation and Correction
According to the SEM shown in Fig. 2, we take the 126 samples of the urban
residents into the model. There are 5 paths that don’t pass the significance
test (p < 0.01) in the results and TLI index also fails in the inspection, so the
model needs to be modified and optimized. After repeating the path correction,
we delete the factors of perceived behavioral control, the results are obviously
optimized (Tables 2 and 3). All indexes pass the test. Every variance estimation
also passes the significance test (p < 0.01). In order to make the original model
optimized better, we adopt Modification Index to fix the model again. We try
98 L. Zhang
z1
1
1
e6 a6
Subjective
1
e7 a7 norm
1 z4 1
1 a15 e15
e9 a9 1
1
Peripheral 1
a16 e16
objective
1 z2
e1 a1 factors 1
1 a17 e17
1
e2 a2
1
e3 a3 Attitude
1
1
e4 a4 z5
1 1 1
e5 a5 1 a18 e18
1
Intention a19 e19
1
Perceived a20 e20
1
z3 behavior
control
1
a10 a14
1 1
e10 e14
Fig. 2. Structural equation modeling
to modify the error terms of some observed variables by calculating MI value

and find that it can optimize the index, but it can’t improve the P value. On
balance, we give up Modification Index and just keep the first adjusted result.
(2) Output
For the urban residents’ sports and leisure behavior influence factor, we sum-
marize the verification results of hypothesis 1 to 6 (Table 4). Then we can build
out the SEM for urban residents.
4.3 The Rural Model
(1) Calculation and correction

According to the SEM shown in Fig. 2, we take the 104 samples of the rural
residents into the model. 4 paths in the results don’t pass the significance test
(p < 0.01), and the AIC and EVCI indexes also fail in the test. It needs to correct
the model. After repeating path correction, we delete the factor of perceived
behavioral control and use Modification Index to revise the error terms of a4, a5
and a18 in the model. The results are optimized, all indexes pass the test, and
every variance estimation also passed the significance test (p < 0.01). (Tables 5
and 6)
(2) Output
In view of the sports and leisure behavior influence factors of rural residents, we
summarize the verification results of hypothesis 1 to 6 (Table 7). Then we can
build out the SEM for rural residents.
Table 2. The adjusted coefficient estimates
Estimate S.E C.R P Label Estimate

∗∗∗
Attitude ← Peripheral 0.884 0.166 5.34 par 12 0.764
objective
factors
∗∗∗
Subjective norm ← Peripheral 0.327 0.099 3.294 par 13 0.433
objective
factors
Intention ← Subjective 0.167 0.095 1.756 0.079 par 10 0.135
norm
∗∗∗
Intention ← Attitude 0.648 0.078 8.258 par 11 0.807
a9 ← Subjective 1 0.564
norm
∗∗∗
a7 ← Subjective 1.548 0.248 6.251 par 1 0.867
norm
∗∗∗
a6 ← Subjective 1.581 0.253 6.259 par 2 0.86
norm
a4 ← Attitude 1 0.831
∗∗∗
a3 ← Attitude 0.836 0.128 6.528 par 3 0.57
∗∗∗
a2 ← Attitude 0.81 0.089 9.069 par 4 0.743
∗∗∗
a1 ← Attitude 0.672 0.079 8.549 par 5 0.71
a15 ← Peripheral 1 0.686
objective
factors
∗∗∗
a16 ← Peripheral 0.859 0.187 4.581 par 6 0.514
objective
factors
∗∗∗
a17 ← Peripheral 0.841 0.166 5.072 par 7 0.586
objective
factors
a18 ← Intention 1 0.843
∗∗∗
a19 ← Intention 1.109 0.101 11.029 par 8 0.856
∗∗∗
a20 ← Intention 1.113 0.124 8.976 par 9 0.73
∗∗∗
a5 ← Attitude 0.91 0.107 8.485 par 14 0.706
∗∗∗
p < 0.01
Table 3. The calculation results of the revised fitting index-2
Fit index χ2 (df) CMIN/DF P TLI IFI CFI RMSEA AIC EVCI
Results 130.356 (73) 1.786 0 0.909 0.929 0.927 0.079 194.356 1.555
Conclusion Suitable Pass Pass Pass Pass Pass Suitable Pass Pass
100 L. Zhang
Table 4. Hypothesis testing results of the effects of urban residents’ behavior in sports
and leisure
Hypotheses Validation results

H1: The Attitude of sports and leisure has a significant positive Support
influence on the behavior intention
H2: The attitude of sports and leisure plays an intermediary role Nonsupport
between subjective norm and behavior intention
H3: The attitude of sports and leisure behavior plays an Nonsupport
intermediary role between perceived behavioral control and
behavior intention
H4: Subjective norm has a significant positive influence on the Support
behavior intention
H5: The perceived behavioral control has a significant positive Nonsupport
H6: The peripheral objective factors have a significant positive Partial support
influence on the TPB system
H6a: The peripheral objective factor has a significant positive Support
influence on the subjective norm
H6b: The peripheral objective factors have a significant positive Support
influence on the attitude
H6c: The peripheral objective factors have a significant positive Nonsupport
influence on the perceived behavioral control
Table 5. The adjusted coefficient estimates

∗∗∗
Subjective ← Peripheral 0.263 0.072 3.633 par 13 0.382
norm objective factors
∗∗∗
Attitude ← Peripheral 0.69 0.132 5.245 par 12 0.582
objective factors
∗∗∗
Attitude ← Subjective norm 0.444 0.141 3.14 par 15 0.257
∗∗∗
Intention ← Subjective norm 0.401 0.098 4.094 par 10 0.289
∗∗∗
Intention ← Attitude 0.504 0.058 8.642 par 11 0.628
a9 ← Subjective norm 1 0.49
∗∗∗
a7 ← Subjective norm 1.834 0.243 7.563 par 1 0.911
∗∗∗
a6 ← Subjective norm 1.76 0.232 7.572 par 2 0.876
a4 ← Attitude 1 0.799
∗∗∗
a3 ← Attitude 0.85 0.093 9.171 par 3 0.604
∗∗∗
a2 ← Attitude 0.892 0.071 12.55 par 4 0.791
∗∗∗
a1 ← Attitude 0.778 0.061 12.86 par 5 0.809
a15 ← Peripheral 1 0.698
objective factors
Table 5. (Continued)

∗∗∗
a16 ← Peripheral objective factors 0.769 0.144 5.341 par 6 0.469
∗∗∗
a17 ← Peripheral objective factors 0.839 0.137 6.129 par 7 0.575
a18 ← Intention 1 0.868
∗∗∗
a19 ← Intention 1.14 0.067 17.094 par 8 0.892
∗∗∗
a20 ← Intention 1.116 0.083 13.469 par 9 0.755
∗∗∗
a5 ← Attitude 0.777 0.069 11.333 par 14 0.61
∗∗∗
p < 0.01
Table 6. The calculation results of the revised fitting index-2
Fit index χ2 (df) CMIN/DF P TLI IFI CFI RMSEA AIC EVCI
Results 133.622(70) 1.909 0 0.947 0.935 0.96 0.063 203.622 0.889
Conclusion Suitable Pass Pass Pass Pass Pass Suitable Pass Pass
Table 7. Hypothesis testing results of the effects of rural residents’ behavior in sports
and leisure
Hypotheses Validation results

H1: The Attitude of sports and leisure has a significant positive Support
H2: The attitude of sports and leisure plays an intermediary role Support
between subjective norm and behavior intention
H3: The attitude of sports and leisure behavior plays an Nonsupport
intermediary role between perceived behavioral control and
behavior intention
H4: Subjective norm has a significant positive influence on the Support
behavior intention
H5: The perceived behavioral control has a significant positive Nonsupport
H6: The peripheral objective factors have a significant positive Partial support
influence on the TPB system
H6a: The peripheral objective factor has a significant positive Support
influence on the subjective norm
H6b: The peripheral objective factors have a significant positive Support
influence on the attitude
H6c: The peripheral objective factors have a significant positive Nonsupport
influence on the perceived behavioral control
102 L. Zhang
5 Results
In view of the above empirical analysis, there are some differences between urban
residents and rural residents in sports and leisure behavior influence factor.
From the perspective of direct effects (Table 8), the influence of peripheral
objective factor that affected urban residents’ sports and leisure attitude was
larger 0.182 units than what affected rural residents, the influence of subjective
norm on urban residents was larger 0.051 units than that on rural residents. For
the influence of sports and leisure attitude on the intention, urban residents were
higher 0.179 units than rural residents. The subjective norm only affected the
intention of urban residents but did not affect their attitude, however subjective
norm not only affected the intention of rural residents but also affected their
attitude. The subjective norm had more influences on the intention of rural
residents.
From the perspective of indirect effects (Table 8), the influence of peripheral
objective factors on the intention of urban residents was 0.675 units, it was higher
than that on rural residents. At the same time, the peripheral objective factors
also indirectly affected the attitude of the rural residents. The subjective norm
could indirectly affect the intention of rural residents through attitude, but it
could not affect urban residents.
Table 8. Direct and indirect effect (Standardized results)
Urban Rural
Direct effect/Indirect effect Direct effect/Indirect effect
Peripheral Attitude Subjective Intention Peripheral Subjective Attitude Intention
objective norm factors objective norm
factors
Attitude .764/ - - - .582/.098 .257/ - -
Subjective .433/ - - - .382/ - - -
norm
Intention /.675 .807/ .135/ - /.537 .289/.162 .628/ -
a5 /.540 .706/ - - /.415 /.157 .610/
a20 /.493 /.588 /.099 .730/ /.406 /.341 /.474 .755/
a19 /.578 /.690 /.116 .856/ /.480 /.402 /.560 .892/
a18 /.569 /.680 /.114 .843/ /.467 /.392 /.545 .868/
a17 .586/ - - - .575/ - - -
a16 .514/ - - - .469/ - - -
a15 .686/ - - - .698/ - - -
a1 /.543 .710/ - - /.550 /.208 .809/ -
a2 /.568 .743/ - - /.538 /.204 .791/ -
a3 /.436 .570/ - - /.411 /.156 .604/ -
a4 /.635 .831/ - - /.544 /.206 .799/ -
a6 /.373 - .860/ - /.335 .876/ - -
a7 /.376 - .867/ - /.348 .911/ - -
a9 /.245 - .564/ - /.187 .490/ - -
From the perspective of total effects (Table 8), the influence of peripheral
objective factors on the attitude, subjective norm and intention of urban resi-
dents was greater than that of rural residents. For the influence of attitude on the
intention, it was larger for urban residents than for rural residents. For the influ-
ence of subjective norm on the intention, it was larger for rural residents than
for urban residents. In addition, the subjective norm also affected the attitude
of rural residents, but it did not affect the attitude of urban residents.
6 Discussion and Conclusion
The sports and leisure behavior of urban residents was affected by the inten-
tion, attitude, subjective norm and peripheral objective factors. The Subjective
norm, attitude and intention were the intermediary variables. The subjective
norm and the behavior attitude were two independent variables, besides, they
were both affected by the peripheral objective factors, at the same time they indi-
rectly affected the sports and leisure behavior by impacting behavioral intention.
Sports leisure and behavior of rural residents was also affected by the intention,
attitude, subjective norm and peripheral objective factors. The subjective norm,
behavior, attitude and intention were the intermediary variables. The relation-
ships among them were more complicated than the relationships of urban resi-
dents. The peripheral objective factors affected the subjective norm and attitude,
at the same time they affected the sports and leisure behavior through behav-
ioral intention. In addition, the subjective norm affected the attitude of rural
residents, but it did not affect the attitude of urban residents.
After we calculated the path coefficient, we got a point of view about the total
effect. The influence of peripheral objective factors on the attitude, subjective
norm and intention of urban residents was higher than that of rural residents. For
the influence of attitude on the intention, it was higher for urban residents than
for rural residents. For the influence of subjective norm on the intention, it was
higher for rural residents than for urban residents. That is to say, the influence of
external objective factors (social atmosphere, sports ground, facilities, etc.) on
the sports and leisure behavior of both urban and rural residents was much larger
than the influence of intrinsic objective factors (personal leisure time, personal
income, physical condition, etc.). The diversity and richness of sports and leisure
activities, sports ground, facilities, and social atmosphere impacted urban resi-
dents more than rural residents. Admissive and positive attitude toward sports
and leisure activities influenced urban residents more. While rural residents more
cared about the view of their family, friends and even the government about their
participation in sports and leisure activities.
Acknowledgement. We thank Guo Xinyan and Liu Ying for technical assistance;
Zhang Xu for guidance. This work was supported by Sichuan Social Science Fund
(Grant # SC16XK025).
104 L. Zhang
References
1. Ajzen I (1991) The theory of planned behavior. Organ Behav Hum Decis Process
50:179–211
2. Ajzen I (2016) Ownership structure, diversification, and corporate performance
based on structural equation modeling. J Real Estate Portfolio Manage 22:63–73
3. Beville JM, Meyer M et al (2014) Gender differences in college leisure time physical
activity: application of the theory of planned behavior and integrated behavioral
model. J Am Coll Health 62(3):173–184
4. Boudreau F, Godin G (2014) Participation in regular leisure-time physical activ-
ity among individuals with type 2 diabetes not meeting canadian guidelines: the
influence of intention, perceived behavioral control, and moral norm. Int J Behav
Med 21:918–926 (in Chinese)
5. Caldwell JT, Ford CL et al (2016) Intersection of living in a rural versus urban
area and race/ethnicity in explaining access to health care in the United States.
Am J Public Health 106(8):1463–1469
6. Cerin E, Vandelanotte C et al (2008) Recreational facilities and leisure-time phys-
ical activity: an analysis of moderators and self-efficacy as a mediator. Health
Psychol 27(2):126–135
7. Heesch KC, Masse LC (2004) Do women have enough time for leisure-time physical
activity. Res Q Exerc Sport 75(1):A101–A101
8. Lloyd K, Little DE (2010) Self-determination theory as a framework for under-
standing women’s psychological well-being outcomes from leisure-time physical
activity. Leisure Sci 32(4):369–385
9. Nurmi J, Hagger MS et al (2016) Relations between autonomous motivation and
leisure-time physical activity participation: the mediating role of self-regulation
techniques. J Sport Exerc Psychol 38(2):128–137
10. Qiu Y (2008) Research on restricted factors of leisure sport behavior developing
stagelan assumption theory framework. China Sport Sci 28:71–75 (in Chinese)
11. Qiu Y (2009) Research on the Motivations and Constraints of the Stages of Leisure
Physical Activity. Zhejiang University Press (in Chinese)
12. Qiu Y (2011) Research on the negotiation strategies of the leisure sports activities.
China Sport Sci 31:8–16 (in Chinese)
13. Qiu YJ, Jiao XU (2014) Research on the characteristics of constraints to women’s
leisure sports activities and the relationship with behaviors. China Sport Sci 32:25–
33 (in Chinese)
14. Qiu YJ, Liang MY (2012) Qualitative research on constraints of Chinese women
leisure sports behavior-based on the perspective of the theory of social gender.
China Sport Sci 34:75–82 (in Chinese)
15. Sabiston CM, Crocker PRE (2008) Exploring self-perceptions and social influences
as correlates of adolescent leisure-time physical activity. J Sport Exerc Psychol
30(1):3
16. Shields M, Ellingson L et al (2008) Determinants of leisure time physical activity
participation among Latina women. Leisure Sci 30(5):429–447
17. Simpson VL, Hyner GC, Anderson JG (2013) Lifestyle behavior change and repeat
health risk appraisal participation: a structural equation modeling approach. Am
J Health Promot 28(2):128–135
18. Stanis SAW, Schneider IE, Russell KC (2009) Leisure time physical activity of park
visitors: retesting constraint models in adoption and maintenance stages. Leisure
Sci 31(3):287–304
19. VanSickle J, Schaumleffel NA (2016) Developing recreation, leisure, and sport pro-
fessional competencies through practitioner/academic service engagement partner-
ships. Schole A J Leisure Stud Recreation Educ 31(2):37–55
20. Vanwesenbeeck I, Walrave M, Ponnet K (2016) Young adolescents and advertising
on social network games: a structural equation model of perceived parental media
mediation, advertising literacy, and behavioral intention. J Advertising 45(2):1–15
21. Xu J (2006) The study on the characteristics of the residents’ leisure sports behav-
iors. J Guangzhou Sport Univ 26:97–100 (in Chinese)
22. Ye W (2014) Theory and empirical research on the intention of leisure sports
behavior. Sport Sci Technol 35:102–104
23. Xu Y (2013) A study on the citizens’ behavior intention of sports leisure based on
the theory of planned behavior. Yunnan Geogr Environ Res 25:71–76 (in Chinese)
24. Zhao B (2011) Consumer ethical beliefs associated with its birthplace: empirical
research of China’s urban and rural dual society background. Manage World 1:92–
100 (in Chinese)
25. Zhao Q, Xu N et al (2014) The study of the relation between leisure sports con-
sumption culture and consumption behaviors. Sichuan Sports Sci 33:101–106 (in
Chinese)
Fast Multiobjective Hybrid Evolutionary
Algorithm Based on Mixed Sampling Strategy
Wenqiang Zhang1,2(B) , Yu Wang1,2 , Chunxiao Wang1,2 , Le Xiao1,2 ,

and Mitsuo Gen3
1
Henan University of Technology, Zhengzhou, China
2
Key Laboratory of Grain Information Processing and Control,
Ministry of Education, Zhengzhou, People’s Republic of China
zhangwq@haut.edu.cn
3
Fuzzy Logic Systems Institute, Tokyo University of Science, Tokyo, Japan
Abstract. In this paper, a fast multiobjective hybrid evolutionary

algorithm (MOHEA) is proposed to solve the multiobjective optimiza-
tion problem (MOOP) in achieving a balance between convergence and
distribution with computational complexity. The proposed algorithm,
MOHEA, improves the vector evaluated genetic algorithm (VEGA) by
combing a new sampling strategy according to the Pareto dominating
and dominated relationship-based fitness function. VEGA is good at
searching the edge region of the Pareto front, but it has neglected the cen-
tral area of the Pareto front, and the new sampling strategy prefers the
center region of the Pareto front. The mixed sampling strategy improves
the convergence performance and the distribution performance while
reducing the computational time. Simulation experiments on multiob-
jective test problems show that, compared with NSGA-II and SPEA2,
the fast multiobjective hybrid evolutionary algorithm is better in the two
aspects of convergence and distribution, and has obvious advantages in
the efficiency.
Keywords: Evolutionary algorithm · Mixed sampling · Multi-objective

optimization · Styling · Benchmark
1 Introduction
Multiobjective evolutionary algorithms (MOEAs) have been identified as very
suitable for solving multiobjective optimization problems (MOP) [4]. MOEA is
a heuristic method, and is useful for searching the Pareto optimal solutions of a
MOP by global and local search between generations. From the results, MOEA
can provide decision maker (DM) with a number of more practical solutions. So
the DM can make decisions based on actual needs. In the optimization process
of a single objective, the collaboration of other objectives should be considered
at the same time. Meanwhile, maintaining the trend of simultaneous evolution

Fast Multiobjective Hybrid Evolutionary Algorithm 107
make the Pareto front solutions as close as possible. Simultaneously, the real
Pareto optimal set is evenly distributed.
Moreover, a number of evolutionary algorithms have their own characteris-
tics. Each algorithm is effective to solve the problem of limited areas of appli-
cation. No algorithm can solve all the problems. How to mix these evolutionary
algorithms to play their respective strengths has become one of the challenging
research fields in evolutionary algorithms.
In order to compensate for the shortcomings of single MOEA, more
researchers began to combine different algorithm to generate hybrid algorithm.
In 2012, Zhang and Fujimura proposed an improved vector evaluation of genetic
algorithm with archive (IVEGA-A) [6]. The IVEGA-A combines the strong con-
vergence ability of the VEGA [7] method to the Pareto frontier boundary region
[11] and the elite population updating mechanism based on the new fitness func-
tion to ensure the overall performance of the algorithm in the central region of
the Pareto front. Zhang and Li introduced differential evolution strategy into
MOEA/D [5].
In this paper, we propose a multi-objective hybrid evolutionary algorithm
(MOHEA) to solve multi-objective optimization problem for improving the con-
vergence, the distribution and reducing the computational time. The rest of
the paper is organized as follows: An overview of sampling strategy is given in
Sect. 2. MOHEA is summarized in Sect. 3. Section 4 presents experimental results
to illustrate the efficiency of the algorithm.
2 Mixed Sampling Strategy

2.1 Pareto Dominating and Dominated Relationship-Based Fitness
Function
In this paper, a new fitness function based on Pareto dominating and dominated
relationship-based fitness function (PDDR-FF) is used to evaluate the individual.
The fitness function value of an individual Si is calculated by the following
function:
1
eval(Si ) = q(Si ) + , i = 1, 2, · · · , popSize, (1)
p(Si ) + 1
where q (Si ) is the number of individuals that dominate Si , p (Si ) is the number
of individuals that dominated by Si . popSize is the size of the population. If the
individual Si is non-dominated solution, q(Si ) = 0. Therefore, the fitness func-
tion of the non-dominated individual Si is related to p(Si ). The more individuals
are dominated by Si , the smaller the value of 1/(p(Si ) + 1), and the smaller the
value of PDDR-FF. If q(Si ) = 0, p(Si ) = 0, the PDDR-FF value of Si is 1. It
means there is no individual dominates Si and no individual is dominated by
Si . Therefore, the PDDR-FF value of the non-dominated solution will no more
than 1. Si is dominated individual, q(Si ) ≥ 1; the less individuals that dominate
Si , the smaller q(Si ). In addition, the more individuals that are dominated by
108 W. Zhang et al.
Si , the bigger p(Si ) and the smaller 1/(p(Si ) + 1). Then, the value of PDDR-FF
is smaller.
The smaller value of PDDR-FF means that the more individuals dominated
by Si and the less individuals dominating Si . Therefore, the smaller the PDDR-
FF value, the better. And besides, PDDR-FF can clearly identify the dominant
and non-dominated solutions. If an individual is a non-dominant, the fitness
function value is not more than 1. Even if the individuals are all non-dominated,
the individuals with different number of dominant are endowed with different
fitness function values. It is obvious that the fitness function value (close to 0)
of the non-dominated individuals locating near the central area of the Pareto
front having big domination area is smaller than the value (close to 1) of the
individual in the edge region.
2.2 Mixed Sample Strategy of VEGA and PDDR-FF

The sampling strategy of PDDR-FF has the advantage of being able to converge
to the central region of Pareto frontier. But the convergence of the edge region
is not so good. Only using the PDDR-FF can cause the uneven distribution
performance. Compared to the central area of Pareto frontier, the sampling
strategy of VEGA is more preferred than the edge region, and the single use of
VEGA sampling can also lead to the problem of the distribution performance.
Therefore, the two sampling strategy can be combined to improve the overall
performance of the algorithm and reduce the computation time.
3 Fast Multi-Objective Hybrid Evolutionary Algorithm

3.1 Detailed Process of MOHEA
The strong convergence capability of mixed sample strategy ensures that the
MOHEA has ability to converge to the center and the edge region of true Pareto
frontier. Meanwhile, different convergence regions ensure the distribution.
The specific algorithm flow of MOHEA is shown in Fig. 1.
The specific process of MOHEA is as follows:
Step 1. Initialization of algorithm parameters: The size of the population:
popSize, the size of archive:archiveSize, the maximum number of calcula-
tion:maxEvaluations.
Step 2. Creating initial population P (t) and initial archive A(t).
Step 3. Loop d) − g) step until the termination condition is satisfied: The num-
ber of the calculation achieves the maximum.
Step 4. Selection (VEGA sampling strategy) Using the VEGA sampling strat-
egy to select good individuals as a part of the mating pool. For the two
objectives of problem, at first all individuals in the population
are ordered
P (t)
according to the value of object function 1 and select 2 individuals with
replacement into the sub-population
1 in order. In a similar way, consider-
P (t)
ing objective 2 and select 2 individuals with replacement into the sub-
population 2.
Fig. 1. The specific algorithm flow of MOHEA
Step 5. Generation of the mating pool: mixing the sub-populations and A(t) to
form the mating pool. In the mating pool, sub-population 1 stores the excel-
lent individuals for the objective 1 and sub-population 2 keeps the excellent
individuals for the another objective. The individuals with good PDDR-FF
value are stored in A(t). For the two objectives optimization problem, the
size of these two sub populations and archive are set to the half size of the
population. As a result, there are three parts in the mating pool, 1/3 of the
individuals serve the objective 1, 1/3 of the individuals tend to the objective
2 and the remaining individuals obeys both the two objectives.
Step 6. Genetic operations: according to the selection operator, two individuals
are selected from the mating pool as the parent and the new generation
P (t + 1) is produced after genetic operations (crossing and mutation).
Step 7. The updating of the archive (PDDR-FF sampling strategy): mixing
the new generation P(t+1) and the archive to form the temp population
A (t). Calculate the fitness function value of all individuals in the population,
according to the ascending order and individuals with the smallest of the
| A(t) | values are chosen as A(t + 1).
The overall procedure of MOHEA is described as the follows Fig. 2.

110 W. Zhang et al.
Fig. 2. The overall procedure of MOHEA
3.2 Genetic Encoding
The genetic code operating is real-coded in this paper. Each gene of the chro-
mosome is consists of a random, evenly distributed, floating-point number in
[0, 1], or the true value of the decision variable. Each chromosome is a vector
of a floating point type and the length of the chromosome is determined by the
number of decision variables. Real coding makes the complex genetic operation
easy to implement. It enhances the accuracy of the operation and strengthens
the search ability of the algorithm to the large space.
For example, the number of benchmark test problem ZDT6 [10] decision
variables is 10. Suppose vector V = (0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1) is
a set of ZDT6 solution, the vector V can be directly used as the chromosome
corresponding gene coding sequence. Or the vector V may be used as the gene
sequence of the chromosome according to the mapping relationship in the Fig. 3.
3.3 Crossover Operator
The crossover operator used in this section is the Simulated Binary Crossover
(SBX) proposed by Deb [1]. When the real-coded chromosomes are crossed by
the SBX, the distribution parameter β of the parents need to be calculated by
the equation:
1
(2u) η+1 , u ≤ 0.5
β= 1
1
η+1
(2)
2(1−u) , othersize,
Fig. 3. Real-Coded mapping
where u ∈ U (0, 1), η represents the cross-distribution index and be determined

by the decision maker. The larger the value, the smaller the distance between
the progeny individual and the parents. Conversely, the greater the distance.
Parent S1 and S2 can produce offspring s1 and s2 by the following equation:

S1 = 0.5 [(1 + β) S1 + (1 − β) S2 ]
(3)
S2 = 0.5 [(1 + β) S1 + (1 − β) S2 ].
3.4 Crossover Operator
The mutation operator used in this section is a polynomial variation proposed

by Deb and Goyal [2]. The main forms of variation are:
S = S + δωmax , (4)
where S is a normal individual and S is a variant individual. ωmax is the max-

imum variable value of decision variables. δ is calculated as:
1
(2v) γ+1 − 1, v < 0.5
δ= 1 (5)
1 − [2(1 − v)] γ+1
, v ≥ 0.5,
where v ∈ V (0, 1). γ > 0 represents the variance distribution index and be
determined by the decision maker.
3.5 Selection Operator
In this study, NSGA-II [3] and SPEA2 [13] use binary contention selection as
the selection operator. SPEA2 need to consider the dominance of the individual
relationship. In this context, NSGA-II also needs to note the crowding distance
of individuals. The mating pool of MOHEA consists of two parts: one is the
individuals that are excellent for a single target. In the selection process, the
individuals in the current population are first sorted according to the number of
targets and the sub-population with excellent single-target value is selected in
112 W. Zhang et al.
order. All sub-populations are included as part of the mating pool. The other
is the non-dominated individuals in the elite population. In order to ensure
sufficient information exchange between individuals in the two populations, it
is necessary to randomly select two individuals from the mating pool during
genetic manipulation.
The benchmark problem is used to test the MOHEA proposed in this paper. At
the same time, two classical multi-objective evolutionary algorithms, NSGA-II
and SPEA2, are tested and compared under the same conditions. The test func-
tion uses the ZDT and DTLZ problems that are currently used in the MOEAs
domain. The algorithm is implemented in Java language. The hardware condi-
tion of the PC computer is: CPU Core 2, clock speed 2.8G, memory 2G, and
operating system Windows XP.
4.1 Evaluation Index
Let Sj be the solutions of each algorithm (j = 1, 2, 3). In this study, the test
considers two kinds of common convergence and distribution performance mea-
sures.
Coverage C(S1 , S2 ) is the percent of individuals dominated by S1 in S2 [12].
The bigger the C(S1 , S2 ) value, the better the convergence performance of S1 .
Spacing SP is the standard deviation of the closest distance of individuals
in Sj [8]. The smaller value of SP means the better distribution performance.
The mathematically defined expressions for the 12 benchmark tests used
in the experiment are shown in [10]. Among them, ZDT1, ZDT2 and ZDT3
have 30 decision variables, ZDT4 and ZDT6 have 10 decision variables. The
decision variables and the number of targets of DTLZ problem can be expanded
arbitrarily. For k and | xk | values, if the target number is 2, then k = 2.
The parameters used in the three algorithms are as same as the original. The
detailed values are shown in the following.
• Population size: 100

• Elite population size: 100
• Crossover rate: 0.4
• Mutation rate: 0.8
• Cross distribution index: 15
• Variation distribution index: 20
• Max generation: 500
The three algorithms have the same crossover and mutation operator. Fit-
ness evaluation mechanism and the elite strategy is not the same. It should be
pointed out that NSGA-II and SPEA2 need to select all non-dominated solu-
tions from the elite as Pareto sets. MOHEA updates the Pareto set by choosing
non-dominated solutions among the elite and the population. For the optimal
stopping criterion of the algorithm, this paper uses the most widely used function
evaluation number as the termination condition of the algorithm.
The performance evaluation index of the algorithm adopts the convergence
index C and the distributed index SP. The results of the three algorithms running
30 times are compared. The time of each algorithm running is recorded. Thus the
algorithm efficiency index CPU is obtained. A boxplot was used to evaluate the
alignment results of the algorithm during 30 calculation results [9]. Box diagram
is an important tool for statistical analysis. The image can clearly reflect the
distribution of data. It is the most effective way to display EAs graphically.
The upper line of the box represents the quartile line of the sample. The lower
line indicates the lower quartile line of the sample. The middle line represents
the bisector (mean). The top of the box represents the maximum value of the
sample and the bottom is the minimum value. The gap in the middle of the
box represents the confidence interval, and the other scatter points represent
the outliers.
4.2 Discussion of the Results

The evaluation indexes C, SP and CPU obtained from MOHEA, NSGA-ii, and
SPEA2 are shown in Fig. 3, and the average and variance of each index are
shown in Tables 1 and 3. It can be concluded that MOHEA performs better on
most problems than NSGA-II and SPEA2 in terms of convergence. On questions
ZDT1, ZDT2, ZDT6 and DTLZ7, NSGA-II and SPEA2 performed well with
MOHEA as shown in Table 1. For the distribution, it can be seen from Tables 2
and 3 that the distribution of MOHEA is better than the other two algorithms
except for the problems ZDT1 and DTLZ6. Since the number of variables and
the range of variables make the focus of the 12 test functions different, it is
difficult for the same algorithm to perform optimally on all test functions.
The efficiency of the algorithm can be discussed from two aspects. M is the
number of targets and N mean the population number. In terms of distance com-
Table 1. C index
Questions C(NSGA-II, MOHEA) C(MOHEA, NSGA-II) C(SPEA2, MOHEA) C(MOHEA, SPEA2)

ZDT1 7.91e-011.5e-01 1.16e-011.4e-01 4.20e-011.8e-01 4.77e-012.0e-01
ZDT2 6.26e-027.1e-02 8.31e-016.8e-02 4.52e-012.6e-01 4.22e-012.2e-01
ZDT3 1.54e-011.5e-01 6.39e-012.2e-01 2.95e-012.2e-01 4.29e-011.9e-01
ZDT4 1.79e-013.2e-01 4.56e-013.1e-01 2.74e-013.2e-01 2.82e-012.6e-01
ZDT6 8.18e-011.9e-01 9.66e-022.1e-01 5.06e-012.6e-01 3.85e-012.6e-01
DTLZ1 1.98e-012.2e-01 6.04e-012.9e-01 3.06e-012.5e-01 4.47e-012.7e-01
DTLZ2 1.60e-022.9e-02 1.01e-012.1e-02 6.00e-031.0e-02 9.40e-022.2e-02
DTLZ3 1.79e-011.7e-01 4.24e-012.5e-01 3.25e-012.3e-01 4.43e-012.3e-01
DTLZ4 2.20e-021.9e-02 9.70e-022.1e-02 2.40e-022.1e-02 9.70e-022.5e-02
DTLZ5 2.50e-022.7e-02 7.90e-021.6e-02 3.00e-036.4e-03 7.40e-021.5e-02
DTLZ6 1.88e-011.3e-01 6.98e-011.8e-01 3.34e-012.9e-01 5.15e-012.7e-01
DTLZ7 7.64e-011.4e-01 1.51e-011.3e-01 5.15e-012.3e-01 3.99e-012.2e-01
114 W. Zhang et al.
Table 2. SP index
Questions NSGA-II SPEA2 MOHEA

ZDT1 1.12e-021.8e-03 1.12e-021.8e-03 1.85e-026.2e-03
ZDT2 2.03e-023.4e-03 2.03e-023.4e-03 1.18e-022.8e-03
ZDT3 1.06e-021.1e-03 1.06e-021.1e-03 9.33e-031.3e-03
ZDT4 1.27e+006.2e-01 1.27e+006.2e-01 3.20e-014.0e-01
ZDT6 5.16e-023.2e-02 5.16e-023.2e-02 2.11e-024.3e-03
DTLZ1 1.31e+004.1e-01 1.31e+004.1e-01 8.39e-011.7e+00
DTLZ2 4.46e-036.2e-04 4.46e-036.2e-04 4.01e-034.1e-03
DTLZ3 2.06e+013.0e+01 2.06e+013.0e+01 4.73e+001.7e+00
DTLZ4 6.54e-036.7e-04 6.54e-036.7e-04 4.45e-034.7e-04
DTLZ5 4.23e-033.6e-04 4.23e-033.6e-04 2.47e-031.7e-03
DTLZ6 9.05e-021.8e-02 9.05e-021.8e-02 1.01e-011.3e-02
DTLZ7 1.83e-027.1e-03 1.83e-027.1e-03 1.79e-022.9e-03
Table 3. CPU index
Questions NSGA-II SPEA2 MOHEA

ZDT1 7.92e+046.8e+03 1.29e+042.6e+03 8.29e+039.9e+02
ZDT2 1.03e+058.1e+03 1.20e+041.9e+03 7.80e+039.3e+02
ZDT3 7.28e+046.1e+03 1.19e+042.0e+03 9.11e+037.2e+02
ZDT4 6.80e+045.3e+03 1.19e+041.9e+03 9.06e+038.0e+02
ZDT6 8.92e+047.6e+03 1.25e+042.1e+03 8.73e+036.9e+02
DTLZ1 8.04e+048.8e+03 1.21e+041.9e+03 8.95e+037.6e+02
DTLZ2 1.06e+051.1e+04 1.16e+041.6e+03 6.88e+035.5e+02
DTLZ3 9.99e+041.8e+04 1.24e+041.9e+03 8.76e+037.9e+02
DTLZ4 1.25e+051.9e+04 1.32e+042.0e+03 8.77e+038.2e+02
DTLZ5 6.57e+043.2e+03 1.38e+041.3e+03 9.23e+039.8e+02
DTLZ6 7.99e+046.3e+03 1.22e+041.7e+03 8.73e+039.5e+02
DTLZ7 1.12e+052.0e+04 1.34e+041.9e+03 8.78e+031.1e+03
putation, the time complexity of NSGA-ii is O (mN log2 N ). The computational

complexity of SPEA2 is O (mN log2 N). Since there is no distance mechanism in
the MOHEA, the computational complexity of MOHEA is 0. In the calculation
of the fitness value, the MOHEA needs mN (N − 1)/2 comparisons to calculate
PDDR-FF value, so the computational complexity of MOHEA is O(mN log2 N ).
The computational complexity of NSGA-II is the same as SPEA2. The value is
O(mN 2 ). In summary based on the quantitative analysis of the complexity, the
MOHEA is clarified faster than NSGA-II and SPEA2, and the experimental
result are shown in Table 1.
5 Conclusions
In this study, a multi-objective hybrid evolutionary algorithm (MOHEA) based
on mixed sampling strategy was proposed to solve the multi-objective optimiza-
tion problem while improving the convergence and distribution performances,
and reducing the computational time at the same time. The VEGA sampling
strategy prefers the edge region of the Pareto front, but deviates from the search
of the central region. The PDDR-FF sampling strategy makes up the shortcom-
ing of VEGA. The mixed sample strategy could converge the multiply regions
of Pareto front to improve the efficacy. Since no computation requirement for
distance, the MOHEA could also improve the efficiency. Numerical comparisons
indicated that the proposed MOHEA outperforms the classical MOEAs such as
SPEA 2 and NSGA ii on convergence and distribution performance as well as
computational time too.
Acknowledgements. This research work is supported by the National Natural Sci-

ence Foundation of China (U1304609), Foundation for Science & Technology Research
Project of Henan Province (162102210044, 152102210068, 152102110076), Program
for Innovative Research Team (in Science and Technology) in University of Henan
Province (17IRTSTHN011), Program for Key Project of Science and Technology in
University of Education Department of Henan Province (17A520030), Fundamental
Research Funds for the Henan Provincial Colleges and Universities (2014YWQQ12,
2015XTCX03, 2015XTCX04), Research Funds for Key Laboratory of Grain Informa-
tion Processing and Control (Henan University of Technology) (KFJJ-2015-106), Min-
istry of Education and the Grant-in-Aid for Scientific Research (C) of Japan Society
of Promotion of Science (JSPS): No. 15K00357.
References
1. Deb K, Beyer HG (2001) Self-adaptive genetic algorithms with simulated binary
crossover. Evol Comput 9(2):197–221
2. Deb K, Goyal M (1999) A combined genetic adaptive search (geneas) for engineer-
ing design. Comput Sci Inform 26:30–45
3. Deb K, Pratap A et al (2002) A fast and elitist multiobjective genetic algorithm:
NSGA-II. IEEE Trans Evol Comput 6(2):182–197
4. Gaspar-Cunha A, Covas JA (2001) Robustness in multi-objective optimization
using evolutionary algorithms. John Wiley & Sons, Inc
5. Li H, Zhang Q (2009) Multiobjective optimization problems with complicated
pareto sets, MOEA/D and NSGA-II. IEEE Trans Evol Comput 13(2):284–302
6. Non-Member WZ, Member SFS (2012) Multiobjective process planning and
scheduling using improved vector evaluated genetic algorithm with archive. IEEJ
Trans Electr Electron Eng 7(3):258–267
7. Schaffer JD (1985) Multiple objective optimization with vector evaluated genetic
algorithms. In: International Conference on Genetic Algorithms, pp 93–100
8. Schott JR (1995) Fault tolerant design using single and multicriteria genetic algo-
rithm optimization. Cell Immunol 37(1):1–13
9. Tukey JW (1978) Variations of box plots. Am Stat 32(1):12–16
116 W. Zhang et al.
10. Yao X, Liu Y, Lin G (1999) Evolutionary programming made faster. IEEE Trans
Evol Comput 3(2):82–102
11. Yu X, Gen M (2010) Introduction to Evolutionary Algorithms. Springer, London
12. Zitzler E, Thiele L (1999) Multiobjective evolutionary algorithms: a comparative
case study and the strength pareto approach. IEEE Trans Evol Comput 3(4):257–
271
13. Zitzler E, Laumanns M, Thiele L (2001) Spea2: Improving the strength pareto
evolutionary algorithm. vol 3242, pp 95–100
Literature Mining Based Hydrogen
Fuel Cell Research
Luoji Li1 , Ying Zhang2(B) , and Qiulin Li3

1
Shenyang Sport University, Shenyang 110102, People’s Republic of China
2
Environment School, Liaoning University,
Shenyang 110036, People’s Republic of China
yingying0809@126.com
3
Business School, Sichuan University, Chengdu 610065, People’s Republic of China
Abstract. In this paper, we reviewed the current technology for the

hydrogen fuel cells and got the light spots when 1271 papers related to
hydrogen fuel cells were analyzed, which were focused on the hydrogen
fuel cell technologies, hydrogen fuel cell application and solar hydro-
gen fuel cell. This paper had outlined these several types of the spe-
cific technology and confirmed that hydrogen fuel cells was an inevitable
choice when faced to serious environmental pollution and energy struc-
ture’s imbalance. And the solar-hydrogen fuel cell system that involved
the hydrogen production, hydrogen storage and hydrogen application is
of significant importance addressing the problem of the production of
renewable energy, power generation and transportation pollution.
Keywords: Hydrogen fuel cell · Solar-hydrogen · FC vehicles ·

Zero-pollution
1 Introduction
Many environmental issues have been caused by or relate to the production,
transformation and use of fossil energies, for example, acid rain, stratospheric
ozone depletion and global climate change [5]. Energy is the material basis for
the country’s social development and progress of science and technology, but due
to the constant use of the three large fossil energies, energy conservation and
environmental problem became more and more serious. As a result of the badly
exceeds bid of PM2.5, human beings got a warning of continuous haze weather.
And the air pollution problem can be settled down when it is mainly closely
related to fossil fuel combustion and transport industries. The increasing trend
in world’s energy need is expected to continue in the future. As a result, a growth
in energy generation capacity will be needed [1]. Cars are important transport
means in people’s lives. However, the automobile exhaust is the important cause
of increasingly serious environmental pollution. Therefore, we need to find an
alternative fuel for the traditional fossil fuels.
118 L. Li et al.
Hydrogen has high energy density, and the energy released is enough to
make the vehicle’s engine running. Hydrogen energy systems appear to be the
one of the most effective solutions and can play a significant role in providing
better environment and sustainability [4,14]. Hydrogen fuel cell vehicle itself
works with no noise, no vibration, no loss, and long service life. The electrode
is only working as the place of chemical reaction and conductive channel and
it does not participate in chemical reactions. The fuel of hydrogen fuel cell is
hydrogen and oxygen, and the product is clean water, with the prefect results of
no CO and CO2 , and no sulfur and particulate exhaust. We could have a clear
look at the advantages when compared to the thermal power as Table 1 shows.
Therefore, hydrogen fuel cell vehicle can run with zero emissions, zero pollution
in real sense and hydrogen is regarded as a perfect energy for vehicle. Fuel cell
technology with the hydrogen energy is the core technology of the auto industry
in the 21st century, and it takes revolutionary significance for the car industry.
Table 1. Comparison of atmospheric pollution between fuel cells and thermal power
Pollution Gas thermal Heavy oil Coal thermal Fuel

components power thermal power power cells
SO2 2.5–230 4550 8200 0–0.12
NO2 1800 3200 3200 63–107
Hydrocarbon 20–1270 135–5000 30–10000 14–102
Dust 0–90 45–320 365–680 0–0.14
(unit: kg·10−6 (KWh)−1 )
In the face of oil price’s rising, increasingly severe atmospheric environment,

increasingly strict emission regulations and the pressing needs of light energy,
fuel cell is the power generation device which can turn the chemical energy
stored in the fuel and oxidant directly into electrical energy. It is considered
to be the best choice as clean and efficient power generation technology in the
21st century because of its energy conversion is not limited by Carnot cycle
[8], energy conversion efficiency can reach 90%, and the actual service efficiency
is 2–3 times of internal combustion engine. And hydrogen fuel cell shows the
unique superiority in such aspects as in the comprehensive energy efficiency,
environment friendly and high reliability, so the hydrogen fuel cell will have
broad market prospects as specific tools for the transition from fossil fuel based
era to hydrogen economy-oriented society.
2 Literature Mining
Hydrogen fuel cell is a cutting-edge research problem in current society, but
researches about this issue has certain basis and achievements. Based on these
Literature Mining Based Hydrogen Fuel Cell Research 119
researches, literature mining has been shown to be a powerful method for elu-
cidating major trends across time in published scientific literature so that topic
maps can be built [2]. Garfield believed that the academic literature citation
indexing was crucial when researching similar topic areas [7]. A citation index is
a synthesized result based on journal articles, keywords, publication dates and
abstracts and is able to separate and highlight the various influences in a specific
field, allowing for the research with the greatest impact to be easily identified.
2.1 Data Analysis
Literature mining from published scientific literature allows for the discovery of
key areas and trends [2]. CiteSpace is one of the latest development of a generic
approaches to detecting and visualizing emerging trends and transient patterns
in scientific literature. The work makes substantial theoretical and methodolog-
ical contributions to progressive knowledge domain visualization. A specialty is
conceptualized and visualized as a time-variant duality between two fundamen-
tal concepts in information science: research fronts and intellectual bases [3].
1271 papers had been retrieved as “TI = HFC” been limited, and all the data
was imported into the CiteSpace with a data process. Finally, 65 categories were
formed after cluster, as Fig. 1 shows.
Fig. 1. Clusters of the hydrogen fuel cells
2.2 Results
Keywords are important index terms used to provide access to articles that
have been published or presented in journals and conference databases and are
120 L. Li et al.
essential for clearly identifying arange within a specified boundary. As Fig. 2

illustrates, keywords about hydrogen fuel cells have been shown from 1960 to
2016, which could be divide into three phases as timezone developed. And the
first stage could be described as hydrogen fuel cell technologies in Sect. 3, the
second could be extracted as hydrogen fuel cell application in Sect. 4, and the
latest stage could be know as solar hydrogen fuel cell in Sect. 5, which is on
behalf of the most frontier and hot issue on hydrogen fuel cells.
Solar
Hydrogen
Fuel Cell
Hydrogen
fuel cell
application
Operational solar-hydrogen
characterist
ics renewable energy
vehicles efficiency
hfc network costs
two-phase mixture
thermoanalytical study transport
temperature-dependence stationary
storage
emissions portable
reaction
heat transfer MCFC
physical layer design PEMFC
two-way hfc environment
hfc refrigerant
presencerefrigerant equipment
1990 2000 2010 2016
Fig. 2. Timezone of the hydrogen fuel cells
3 Hydrogen Fuel Cell Technologies

Hydrogen production has laid the basis for the development of fuel cells (FC) and
the diffusion of associated technologies. The fuel cell technology mechanisms were
discovered in 1838 by the German-Swiss chemist Christian Friedrich Schonbein
and the British lawyer and natural scientist Sir William Grover [17]. The FC
technologies have the potential to be the key technologies that contribute to the
transition to a low carbon energy economy. The fuel cell and its application has
become the most attractive and promising technology in the hydrogen energy
utilization field. Because of its’ promising clean and efficient power generation,
as oil and the internal combustion engine displaced coal and the steam engine,
FCs will make the transition in a similar way in the future.
3.1 Operational Characteristics

A fuel cell is similar to conventional battery but different, being composed of an
anode, cathode and electrolyte. These components generate electricity through
an electrochemical reaction in which the external oxygen supply source (usu-

ally air) and a hydrogen-rich fuel combine to form water as Eq. (2) show. The
difference with a conventional battery is that no combustion occurs in FCs, as
the energy to run the FCs is released electro-catalytically. This electrochemical
reaction at the two electrodes is shown from Eqs. (1) to (2) respectively.
Atanode: H2 → 2H+ +2e− ., (1)

−
Atcathode: 2H +2e +O2 → H2O.+
(2)
As the schematic diagram in Fig. 3 shows, the fuel (hydrogen) and oxidizer
(oxygen) dissociate to evolve electrons and ions, the ions separating from the
pole move towards another under electrolyte conditions while the dissociated
electrons are guided to an external electric circuit. The fuel (hydrogen) is sup-
plied for the cathode, and the oxidizer for the anode. The hydrogen dissociates
the H+ and e− , and the H+ blends into the electrolyte while the e− , tends
towards the positive pole through an external circuit. The H+ and oxygen in the
air react at the positive pole to form water when absorbing e− ,. As long as the
fuel and oxygen are supplied, there is a constant power output in the FCs.
Lo ad
e−
Dep leted fu el an d Dep len ted o xid an t an d
p ro d u ct g as ed o u t p ro d u ct g as es o u t
−
H2 OH O2
AFC
HO
2
HO
2
H2 H
−
O2
P EMFC
HO P AFC
2
H2 CO 3
2-
O2
CO MCFC
2
CO
HO
2
2
2−
H2 O2 O2
S O FC
HO
2
Fu el in Oxid an t in
A n o d e Electro ly te Cath o d e
Fig. 3. The schematic diagram of FCs and different types FCs’ operation principle
3.2 Outlook for Fuel Cell Technologies

FCs are classified by the kinds of electrolytes they use, which determines the kind
of chemical reactions that take place in the cell, the kind of catalysts required,
122 L. Li et al.
the temperature range in which the cell operates, the fuel required, and other
factors. In turn, these chemical reaction condition characteristics affect the appli-
cations for which these cells are most suitable. There are six major types of fuel
cells presently under development, each with its own advantages, limitations,
and potential applications, as shown in Table 2. Respectively, they are proton
exchange membrane fuel cells (PEMFC), alkaline fuel cells (AFC), phosphoric
acid fuel cells (PAFC), molten carbonate fuel cells (MCFC), solid oxide fuel cells
(SOFC) and direct methanol fuel cells (DMFC). The operating characteristics
of these systems are highlighted in Table 2.
Table 2. Operating and applicable properties of FCs technologies
Styles PAFC MCFC SOFC PEMFC AFC

Fuel Natural Natural gas Natural gas Hydrogen, Hydrogen,
gas, methane gas, methane gas, oxygen oxygen
methane air, coal gas, coal gas, air,
hydrogen hydrogen hydrogen
peroxide, peroxide peroxide,
air
Electrolyte phosphate Alkali metal Oxygen ion Proton Potassium
carbonate conductive exchange hydroxide
fused ceramics membrane solution
mixture
Operating 160–220 620–660 800–1000 80 90
temperature
Electrochemical 55% 65% 60–65% 40–60% 60–70%
efficiency
Power output 200KW 2MW– 100KW 1KW 300W–5KW
10MW
AFCs use an aqueous solution of potassium hydroxide (KOH) as an elec-

trolyte with a concentration ranging from 35% to 45%, a fuel of pure hydrogen,
and typically operate at a low temperature of around 70◦ C. It is a one of the most
developed technologies so far and has been used by NASA spacecraft. Because
of the low operating temperature, instead of a platinum catalyst in the system,
a variety of non-precious metals such as the most commonly used catalyst of
nickel can be employed to speed up the reactions taking place at the anode and
cathode in AFC units. Owing to the rate at which the chemical reactions take
place, AFCs offer relatively high fuel to electricity conversion efficiencies, as high
as 60% in some applications. The major advantage of AFCs is the quick start,
but the disadvantage is that they are very sensitive to CO2 which causes a slower
reaction time and consumes the alkaline in the electrolyte, thereby limiting its
commercial applications.
PEMFCs are considered to be a promising technology for clean and efficient
power generation in the 21st century and are known for their high-power density
delivery and short start up time compared with other fuel cells, characteristics
which guarantee that the PEMFCs technology is competitive for transportation
and commercial applications such as stationary, and portable power generation
[13]. PEMFCs use a solid polymer electrolyte (Teflon-like membrane) which is
an excellent conductor of protons and an insulator for electrons to exchange
ions between two porous electrodes, and operate under a temperature of 100◦ C.
The operating schematics are shown in Fig. 3. Therefore, the membrane is the
key point in the PEMFCs technology and current research on PEMFCs also
focuses on the development of a proton exchange membrane with high proton
conductivity, low electronic conductivity, good thermal stability, and low cost.
Figure 3 shows that the working principles of the PAFCs and PEMFCs are
similar even though the structures are different. PAFCs use liquid phosphoric
acid as an electrolyte which is contained in a Teflon-bonded silicon carbide
matrix, and works at about 150 to 200◦ C. Different from PEMFCs and AFCs,
PAFCs are more tolerant and insensitive to any impurities in the hydrogen fuel.
The advantages of PAFCs are that the metal catalysts required are much less
than that of the AFCs and the reducing agent purity requirements are signifi-
cantly lower allowing for up to 5% carbon monoxide. PAFCs are viewed as the
“first generation” of modern FCs. According to the state development, these are
the first mature cell types to be used commercially. The 100200 and 500 kW size
plants are typically available for stationary power generation. A 1.3 MW system
has already been tested in Milan and PAFCs have been installed at 70 sites in
Europe, USA and Japan [16].
MCFCs are high-temperature (650◦ C) fuel cells which commonly use a
molten carbonate salt, such as lithium carbonate, potassium carbonate or sodium
carbonate, suspended in a porous ceramic matrix as the electrolyte. To date, this
type has been developed in natural gas and coal based power plants for indus-
trial and military use. The SOFCs’ electrolyte is a hard, non-porous ceramic
compound such as yttria stabilized zirconia which has a strong conductivity and
operates at a much higher temperature of 800–1000◦ C [12]. SOFCs are consid-
ered to be around 50%–60% efficient at converting fuel to electricity and have
rapidly increased in popularity in stationary applications in recent years.
Different from the five types fuel cells which fueled by hydrogen, DMFCs
are powered by pure methanol or alcohol which can be generated hydrogen via
reforming reaction as presented in Sect. 3. Direct methanol fuel cell technology is
relatively new compared with that of fuel cells powered by pure hydrogen, and
DMFCs research and development is roughly 3–4 years behind that for other
fuel cell types [10]. The working principle and reaction equations are shown in
Eqs. (3)–(4).
Atanode: CH3 OH+H2 O → CO2 +6e− +6H+ , (3)

3
Atcathode: O2 +6e− +6H+ → 3H2 O. (4)
2
124 L. Li et al.
4 Hydrogen Fuel Cell Application

Generally, today, the application of fuel cell technologies can be categorized in
three directions. Respectively, there are portable fuel cells designed to be moved,
including auxiliary power units (APU), stationary power fuel cells designed to
provide power to a fixed location, and transport fuel cells providing either pri-
mary propulsion or range-extending capacity for vehicles as shown in Table 3.
Table 3. A brief summary of hydrogen energy application technical directions [11]
Application Portable Stationary Transport

Definition Units that are built Units that provide Units that
into, or charge up, electricity provide
products that are sometimes, but are propulsive
designed ti be not designed to be power or ranger
moved, including moved extension to
auxiliary power vehicle
units (APU)
Typical power range 5W–20KW 0.5KW–400KW 1KW–100KW
Typical technology PEMFE, DMFC MCFC, PAFC, PEMFC,
PEMFC, SOFC DEMFC
Examples Non-motive APU Large stationary Materials
(campervants, combined heat and handing
boats, lighting); power; small vehicles; fuel
military stationary cell electric
applications; micro-CHP; vehicles
portable products; uninterruptible (FCEV); Trucks
small personal power and buses
electronics supplies(UPS)
4.1 Portable Technological Direction

Using portable fuel cells can solve the low energy power capability and long
charging time problems of current battery technology. Consequently, the global
production of portable fuel cells has grown continuously on the basis of FC
technology. In the 21st century, commercial portable units are being adopted
rapidly in such areas as educational fuel cells and toys, auxiliary power units
(APU), and consumer electronics [11]. With the many specific industrial niches
that can benefit from fuel cells, the technology’s use in APU has been an area
of interest for fuel cell manufacturers and industry players alike. DMFCs are a
relatively expensive cell technology and is most cost-effective in small, low-power
stack configurations under 100 W. There are a number of products emerging for
applications that require a higher APU output. The Danish company Serenergy
uses methanol as a fuel, but processes it ahead of a 350 W high-temperature
PEMFC (HT PEMFC), which it is marketing fora variety of applications.
The long-awaited launch of a number of fuel cell consumer electronics charg-

ers in 2011 and 2012 has marked a turning point in this sector. For many years
it proved too difficult to miniaturize fuel cell technology to a level where it could
realistically compete with external battery supplies. Continued development is
expected to result in units small enough to be integrated inside consumer devices
and large multinational companies, such as Apple and Research In Motion, are
patenting products in this field.
4.2 Stationary Technological Direction

Stationary units that provide electricity/heat are not designed to be moved. They
are used for backup power, in remote locations, as stand-alone power plants for
towns and cities, for distributed generation in buildings, and for co-generation [9].
Compared with traditional power stations, stationary fuel cells use is inevitable
because of today’s subtle change in energy structures and the breakthroughs in
fuel cell technology.
In recent years, both small and large systems in residential, industrial and
prime power applications have experienced a fast growth such as in the highly-
advanced Japanese Ene-Farm residential micro-CHP plan and the emergence
of similar schemes in Europe. Present fuel cell technology allows for a contin-
ued cost reduction and optimization of the technology. At the same time, the
commercialization of backup power systems for the telecommunications industry
has accelerated as the communication field expands with a significant interest
emerging in China and elsewhere for both independent and grid-connected sys-
tems. The introduction of methanol-fuelled telecommunication backup locations,
where delivering hydrogen is difficult, to take advantage of fuel cell technology.
Large stationary fuel cell installations have gained increasing interest from multi-
national companies who are looking for ways to reduce their carbon footprint.
4.3 Transport Technological Direction

The global automotive industry sells in excess of 80 million vehicles per year, and
for fuel cells a portion of this market remains the long-term prize in this sector.
FCs are capable of replacing ICEs (internal-combustion engines) as they have
the potential to achieve higher efficiency and lower GHG emissions. A typical
application in vehicles varies from passenger cars, utility vehicles, to buses, with
power ranges from 20 kW to 250 kW [15].
The production of regular automobiles increased steadily in the early 2000 s.
In the past few years, the fuel cell light-weight vehicle market has been led
by Honda, General Motors, and others. Honda has started shipping its FCX
Clarity to Southern and Northern California. In 2007 General Motors, through
its “Project Driveway” program, delivered over 100 units of its Chevrolet
Equinox fuel-cell vehicles to California, Washington DC, and New York. Other
fuel cell application technologies are in marine and unmanned aerial vehicles
(UAVs), trucks, trains, and bicycles. Siemens has developed a fuel-cell-based
air-independent propulsion system intended for use in submarines, such as
126 L. Li et al.
the U212/214 type hybrid powered submarine manufactured by the German

Howaldtswerke Deutsche Werft GmbH (HDW).
5 Solar Hydrogen Fuel Cell
Solar energy is a ubiquitous and rich source, also a kind of renewable energy
sources with no environmental pollution. China has abundant solar energy
resource, annual total solar radiation area is greater than 1050 kWh/m2 , which
is more than 96% of the total land. There are several methods for producing
hydrogen from solar energy. Currently, the most widely used solar hydrogen
production method is to obtain hydrogen by electrolyzing the water at low tem-
perature. Solar-hydrogen fuel cell power generation system was new-type energy
which was based on the solar power for water decomposition, and it could gen-
erate electricity by electrochemical reaction using the hydrogen and oxygen pro-
duced. Countries are committed to carry on the development and application
of solar-hydrogen fuel cell because of rich energy source, high energy conversion
efficiency, and no pollution in the whole transformation process. In the future
energy system, solar energy will become the main primary energy to replace the
current energy such as the coal, oil and gas, and hydrogen fuel cell will become
the clean energy to replace petrol, diesel and chemical battery.
r
to
c
l le
co
r
la
hot
so
ls
water
elc
tank
ar
ol
cs
lta
MPPT battery
Diode
vo
to
o
Ph
Hydrogen storage
H2
electrolyzer
PV Panel
Fuel ce ll
Fig. 4. Scheme of the solar-hydrogen system
As a reversible fuel cell, it gives you the freedom to invent your own clean
energy applications using fuel cells and renewable hydrogen formed using sun and
water. Figure 4 shows a schematic diagram of hydrogen production system that
uses photovoltaic technology. The electrolysis of distilled water using electricity

produced by the PV panel takes place in electrolyser unit and produces hydrogen
and oxygen as the end product [6]. Through the study of solar-hydrogen fuel
cell system’s characteristics, we known that hydrogen fuel cell power generation
efficiency could be as high as 36.4%, and solar cell power generation efficiency
generally is in 15% to 18% range. Therefore, the hydrogen fuel cells not only
greatly improve the efficiency of the power generation, also produce only water
in the whole power generation process, which implement the clean, pollution-
free and efficient, and will become a leader in the field of new energy in the 21st
century.
6 Conclusion
Through the researches on the existing hydrogen fuel cell technologies and the
application technologies, we learned that the real value of the hydrogen fuel cells
is does not poor. Especially this particular and reversible solar-hydrogen fuel cell
system is bright star prior to other styles with a steady stream of resources and
the cycle of zero pollution. The direct hydrogen fuel cell vehicle is preferred, since
it would be less complex, have better fuel economy, lower greenhouse gas emis-
sions, greater oil import reductions and would lead to a sustainable transporta-
tion system once renewable energy was used to produce hydrogen. All countries
need to increase comprehensive and multi-functional analysis and research on
hydrogen fuel cell’s characteristics if we want to keep in active status in the field
of energy development in the 21st century.
Acknowledgements. This research is supported by the Liaoning economic and social

development project (No. 2017slktjd-063); The authors are also grateful to the Liaoning
doctoral research start-up funding (No. 201501131).
References
1. Acar C, Dincer I (2014) Comparative assessment of hydrogen production methods
from renewable and non-renewable sourcesn. Int J Hydrogen Energy 39(1):1–12
2. Bruijn BD, Martin J (2002) Getting to the (c) ore of knowledge: mining biomedical
literature. Int J Med Inform 67(1):7–18
3. Chen C (2006) CiteSpace II: detecting and visualizing emerging trends and tran-
sient patterns in scientific literature. J Am Soc Inform Sci Technol 57(3):359–377
4. Dincer I (2007) Environmental and sustainability aspects ofhydrogen and fuel cell
systems. Int J Hydrogen Energy 31(1):29–55
5. Dincer I (2012) Green methods for hydrogen production. Int J Hydrogen Energy
37(2):1954–1971
6. Joshi AS, Dincer I, Reddy BV (2010) Exergetic assessment of solar hydrogen pro-
duction methods. Int J Hydrogen Energy 35(10):4901–4908
7. Leydesdorff L (2015) Bibliometrics/citation networks. arXiv preprint
arXiv 1502.06378
8. Li J, Chen L et al (2015) Finite time optimizations of a newton’s law carnot cycle.
Int J Energy Environ 5(6):517
128 L. Li et al.
9. Miller EL, Randolph K et al (2016) Innovative approaches to addressing the funda-

mental materials challenges in hydrogen and fuel cell technologies. MRS Advances
1:1–13
10. Murray EP, Tsai T, Barnett SA (1999) A direct-methane fuel cell with a ceria-
based anode. Nature 400(6745):649–651
11. Narayanan S, Valdez T (2002) Portable direct methanol fuel cell systems.
Handbook of fuel cells
12. Park SD, Vohs JM, Gorte RJ (2000) Direct oxidation of hydrocarbons in a solid-
oxide fuel cell. Nature 404(6775):265–267
13. Stumper J, Stone C (2008) Recent advances in fuel cell technology at ballard. J
Power Sources 176(2):468–476
14. Turner J, Sverdrup G et al (2008) Renewable hydrogen production. Int J Energy
Res 32(5):379–407
15. Wilberforce T, Alaswad A et al (2016) Advances in stationary and portable fuel
cell applications. Int J Hydrogen Energy 41(37):16,509–16,522
16. Winter CJ (2009) Hydrogen energy-abundant, efficient, clean: a debate over the
energy-system-of- change. Int J Hydrogen Energy 34(14):1–52
17. Young S (2001) Tomorrow’s energy: hydrogen, fuel cells and the prospects for a
cleaner planet. Nature 414(6863):487–488
GM(1.1) Model Based Leshan’s Industries
Shift-Share Analysis and Prediction
Hongxing Lan1,2 , Caiyang Xu1 , and Zhiyi Meng1,2(B)

1
Business College, Sichuan Agricultural University,
Dujiangyan 611800, People’s Republic of China
zhiyimengscu@sina.com
2
Southwest Poverty Reduction Center, Sichuan Agricultural University,
Abstract. The technique of shift-share method [6] and GM(1.1) model

[4] are applied for a dynamic assessment and predictive study of indus-
tries between the year of 2010 and 2015 in Leshan city, choosing Sichuan
province as a background. With the actuality, some advice was put for-
ward on how to adjust and optimize the development of industries of
Leshan. The survey shows that the secondary industry is the key factor
to the whole economic development of Leshan City. Leshan’s tertiary
industries had a weak competitiveness and will develop more slowly,
department structure contributed a little to the rapid growth of economic
total; Tourist industry and trade are lack of competitiveness, while the
transportation industry industries ran on a small scale and contribute
little to the growth of economic total. Hence, it’s important and neces-
sary to formulate the industry development strategy based on objective
scientific analysis of Leshan’s industries.
Keywords: GM(1.1) model · Shift-share · Leshan city · Industrial

policy
1 Introduction
Shift-Share Method (SSM) was put forward by America scholars Dunn, Perloff,
Lampard and Muth in the 60s, and it was summarized as the form which is
widely acknowledged now by Dunn in the early 80s [2]. Shift-Share Method is
widely applied in the assessment of regional economy and industrial analysis
with high comprehensiveness and dynamic which is used mainly in analyzing
regional economy efficiency. Due to the particularity of object, every practical
application has its own referring value.
SSM was introduced to China in the 80s. Nowdays lots of domestic scholars
study in inter-region with SSM widespread and some surveys already found that
it is evident and suitable for the area of the industrial development assessment
[1,5,9]. However, the utilizing of SSM by existing research is based on the analysis
130 H. Lan et al.
of economic development, depending on the experience to predict the growing

trend of the future whose forecast of the trend still has a high level of subjectivity.
As a result, this text applies GM(1.1) model to predict the chief data of industrial
development. To employ this predicative study to estimate the development of
regional industry in the future on the basis of dynamically assessing industrial
development with SSM.
Following a longitudinal study, we can explore the rule of industry develop-
ment to provide the adjustment of regional industry organization and industry
policy formulation with more scientific decision source.
Leshan city is a significant connection point and port of the economic zones
of the southern Sichuan and Chengdu areas, also the west of Panzhihua which
plays a vital role in economic development structure. On one hand, Leshan city
has an outstanding advantage in agriculture, industry, tourism, transportation
and business. On the other hand, the problems of inadequate use of ascendant
resource still exist. From the view of whole Sichuan province, the overall economic
development of Leshan city remains an upper medium level; in the perspective
of industrial department, the superior industries are not obvious enough which
cannot be judged by experience to find out the departments with advantages
and growth potential. Energy-extensive consumption, high-emission and high-
pollution still have an influence on the economic development of Leshan city. In
this situation, if efforts are made in pursuit of long-term development, it will
solve the problem of increasing load of resource and environment. Therefore,
getting a clear understanding of industrial development trend will offer reliable
evidence to the transformation of low-carbon industry. As a result, it is of great
realistic significance to apply the technique of SSM, dynamic assessment and
predictive study to industrial development of Leshan city.
2 Research Methods and Data Sources

2.1 Shift-Share Method and GM(1.1) Model
SSM regards regional economy as a dynamic process and views its locating region
or whole country as a reference system. Then this method divides the varia-
tion of the regions’ own economic aggregate into three subprojects, which are
the national growth effect, the industry mix effect and the competitive effect
to approve the reason of this region’s economy development and recession. At
the same time, it also can evaluate the strength of regional economic struc-
ture’s competitiveness. Nij = bij × Rj is called as national growth effect which
means Department J’s total quantity is distributed pro rata. The variation of
Department J’s scale in area “i” is the variation caused by regional industrial
department for standardization in average growth rate of locating area or whole
country. Pij = (bij,0 − bij ) × Rj is marked as the industry mix effect (industry
structure effect) which reflects the influence and contribution from the structure
of department to the increase: the rate is greater and that contribution is greater.
Dij = bij,0 × (rij − Rj ) is named as the competitive effect (the local share effect)
which shows the relative competitiveness of Department J, so that a greater rate
Leshan’s Industries Shift-Share Analysis and Prediction 131
means a greater effect to economic increase made by Department J’s compet-

b −b B −B b ×B
itiveness, including rij = ij,tbij,0ij,0 , Rj = j,tBj,0 j,0 , bij = ij,0B0 j,0 . Introduce
Kj,0 = bij,0 /Bj,0 , Kj,t = bij,t /Bj,t which means the relative department’s share
of the whole area or country at the beginning and at last. The relative increasing
rate of area i, compared with its locating area or country, is L = W × U ,
n n n
j=1 Kj,0 × Bj,t j=1 Bj,t j=1 Kj,t × Bj,t
W = n : n , U = n ,
j=1 Kj,0 × Bj,0 j=i Bj,0 j=1 Kj,0 × Bj,t
where W and U separately represent result the effect index and effect of regional
competition index. If Gi is a little bit big while L is greater than 1, the regional
increase is faster than another which is of locating area and whole country. If Pi
is a little bit large while W is greater than 1, it proves the industry department
with high-speed growth has major share. In this situation, the regional overall
economic structure is a little bit good and the structure makes a great contribu-
tion to economic growth. If Di is great and U is greater than 1, the increasing
trend of every industry department is obvious with a strong competitiveness.
GM(1.1) model has part of the information known and partial information
unknown. It can accurately describe the state and behavior of the social economic
system. On one hand, this method can help avoid the Achilles’ heel of the lack of
relevant data. On the other hand, it also can avoid the subjective assume due to
personal preference, experience, knowledge and macro policy which can better
grasp the self-evolution law of the system.
GM(1.1) was modeled as follows:
(1) Set the original sequence X (0) and figure up the original sequence according
i
to x(1) (i) = m=1 x(0) (m), i = 1, 2, · · · , n to create a new sequence X (1) ;
(2) Create a winterization differential equation [12] GM(1.1) as predictive model
(l)
to sequence X (1) , that is dXt +aX (l) = b, a, b are undetermined parameter,
separately as evolution parameter and gray action;
(3) Suppose â as undetermined parameter vector,
with the help of least square
a
method, to figure out â and get â = = (B B)−1 B T X;
T
b
⎡ ⎤ ⎡ (0) ⎤
− 12 [x(1) (1) + x(1) (2)] 1 x (2)
⎢ − 1 [x(1) (2) + x(1) (3)] 1⎥ ⎢ x(0) (3) ⎥
⎢ 2 ⎥ ⎢ ⎥
B = ⎢. .. ⎥ , X = ⎢ .. ⎥
⎣ .. .⎦ ⎣. ⎦
− 12 [x(1) (n − 1) + x(1) (n)] 1 x(0) (n)
(4) Figure out GM(1.1) and get X̂ (1) (i + 1) = (x(0) (1) − ab )e−ai + ab ;
(5) According to Inverse Accumulated Generating Operation to calculate
x(0) (i + 1); Make a counter current to the predictive value X̂ (0) (i + 1) =
X̂ (1) (i + 1) − X̂ (1) (i);
(6) Make an error checking to prediction model, including residual test, rela-
tional coefficient test and posterior-variance-test [11].
132 H. Lan et al.
2.2 Definition of Indicator

In recent years, Leshan city’s GDP has been maintaining a double-digit growth
in 10 years whose comprehensive strength significantly strengthened. In 2009, the
domestic GDP is 61.902 billion yuan–15.3% more than the number of the pre-
vious year. Three industries separately stimulate GDP growth by 0.5%, 11.1%,
3.7% whose contribution rate of economic growth separately are 3.1%, 72.6%,
24.3%. It is necessary to analyze the three industries and they are marked as
GDP1, GDP2, GDP3 in the sections below. During The Eleventh Five-Year-
Plan, by building “the transportation project with one hundred billion yuan”,
developing the economy of Culture & Tourisim, business, mountain area and
port-surrounding, it promoted the position as a south-west hub and accelerated
the development of economy and society. Leshan city is a vital city of industry
and tourism of Sichuan province. At the same time, it belongs to the core area
in Chengdu economic zone as an important city of business, so that its tourism,
transportation industry and business are regarded as leading industries for eco-
nomic development. Therefore, in the current study, based on the analysis of
Leshan city’s three industries, for further study, choose tourism, transportation
and storage postal service, business industry from thrice industry to analyze. In
consideration of data’s access, the current study chooses Total Retail Sales of
Consumer Goods as an alternative to reflect the prosperity degree of the busi-
ness industry. Tourism is marked as GDP31 while transportation storage and
transport the postal service is marked as GDP31. SXT shows Total Retail Sales
of Consumer Goods.
2.3 Data Sources
The data come from Leshan City Statistical Yearbook (1999–2009) [3] and the
documents from relevant departments; Sichuan Province Statistical Yearbook
(1999–2009) [8]. The data are calculated in current prices.
3 The Analysis and Prediction of Industry Development

Share-Shift
(1) The Prediction of Industry Development
According to GM(1.1) model, every index will be predicted and calculated to
get Index Prediction as follow Table 1:
ε(i)
Get relative error sequence of every index Δi = x(0) (i)
and make a residual
test to model. If average relative error is not more than 5%, its predicted value
meets the precision requirement. Specific contents are as follows Table 2.
As shown in Table 2, the average relative error of the tourism industry in
Sichuan province, the tourism, transportation storage and transport the postal
service of Leshan are more than 5% which cannot pass the residual test. On
the basis of the analysis of the relative error every year, the above three indica-
tors in 2002–2004 relative error is a little bit obvious while the error of 2003 is
Table 1. 2010–2015 The prediction of industrial data of Leshan city (L) and Sichuan
province (S)
2010 2011 2012 2013 2014 2015

S GDP1 2730.12 3073.91 3461 3896.83 4387.55 4940.06
S GDP2 8298.7 10117.14 12334.04 15036.72 18331.62 22348.51
S GDP3 5941.44 6896.16 8004.29 9290.49 10783.36 12516.12
L GDP1 5941.386 6896.162 8004.37 9290.67 10783.67 12516.6
L GDP2 462.0822 567.4739 696.9 855.85 1051.06 1290.78
L GDP3 194 226.4765 264.38 308.63 360.28 420.58
L GDP31 1805.09 2157.87 2579.59 3083.73 3686.39 4406.84
S GDP32 595.91 635.097 676.85 721.36 768.79 819.34
S SXT 6704.55 7893.34 9292.92 10940.66 12880.56 15164.42
L GDP31 139.28 164.66 194.66 230.16 272.04 321.6
L GDP32 20.95 24.69 29.1 34.29 40.4 47.61
L SXT 269.35 316.56 372.04 437.24 513.88 603.94
Table 2. GM(1,1)Residual test
Index Average Index Average relative error

relative error
S GDP1 5.00% S GDP31 11.22%
S GDP2 1.68% S GDP32 3.36%
S GDP3 1.28% S SXT 2.30%
L GDP1 5.71% L GDP31 11.13%
L GDP2 4.57% L GDP32 10.09%
L GDP3 2.87% L SXT 1.63%
Relative error Average 2002 2003 2004 2005 2006 2007 2008 2009
S GDP31 11.22% 13.84% 23.00% 9.24% 2.51% 9.77% 13.20% 15.63% 2.55%
L GDP31 11.13% 14.82% 21.18% 9.41% 1.17% 10.99% 16.50% 9.02% 5.97%
L GDP32 10.09% 14.75% 11.35% 15.50% 8.89% 10.20% 6.06% 5.81% 8.16%
the largest–the prediction relative error between Leshan and Sichuan province
in tourism is more than 20% due to the outbreak of SARS which makes a pro-
found impact [7]. Above conclusion is also mentioned in previous studies and
GM(1.1) model still cannot eliminate the phenomenon effectively. Therefore, on
the basis of acknowledging prediction error and considering the reality, the three
sets of predictive data meets the precision requirement.
Calculate the correlation coefficient, correlation degree (while ρ = 0.5, ζ > 0.6
is satisfying), variance ratio (if C < 0.35, and P ≥ 0.95, the forecasting precision
is regarded as superior). The result shows as follow Table 3: the correlation degree
of every index is greater than 0.6; the variance ratio of every index is less than
0.35; every small error possibility is 1. According to the inspection standard, this
prediction accuracy is higher and predicted values have high credibility.
134 H. Lan et al.
Table 3. GM(1,1) Correlation degree and Posterior-variance-test
Project Correlation degree C P Project Correlation degree C P

S GDP1 0.648 0.148 1 S GDP31 0.663 0.142 1
S GDP2 0.762 0.035 1 S GDP32 0.615 0.174 1
S GDP3 0.727 0.019 1 S SXT 0.671 0.026 1
L GDP1 0.623 0.113 1 L GDP31 0.627 0.152 1
L GDP2 0.762 0.087 1 L GDP32 0.875 0.066 1
L GDP3 0.663 0.056 1 L SXT 0.689 0.025 1
(2)The Analysis of Industry Development Share-Shift

According to the SSM model and using Sichuan province as a reference area,
analyze and predict the share-shift of Leshan industrial data from 2002 to 2015
and the results are showed as follow Tables 4, 5 and 6.
Table 4. 2002–2009 The SSM analysis of three industries of Leshan city
Index Leshan 2002 2003 2004 2005 2006 2007 2008 2009
Pij GDP1 2.07 2.641 9.606 3.293 4.302 17.861 12.464 −4.083
GDP2 5.159 10.399 19.066 21.633 26.034 32.233 43.85 29.615
GDP3 3.449 4.776 9.516 8.173 9.469 13.221 12.258 21.453
W 1.0031 1.0028 1.0081 1.0086 1.004 1.0105 1.0011 1.0069
Dij GDP1 −1.15 1.224 1.969 −3.07 3.6 2.341 −3.404 2.807
GDP2 2.678 14.409 3.758 −2.488 3.218 6.75 20.826 −30.634
GDP3 3.035 0.494 −2.644 −0.083 −1.286 −1.924 5.429 9.524
U 1.0622 1.0086 0.9835 1.0127 1.0123 1.0321 0.9716 1.027
L 1.0655 1.0114 0.9915 1.0214 1.0164 1.0429 0.9727 1.0341
Table 4 indicates that during 2002 to 2003, 2005 to 2007, 2009, the regional
economic growth is faster than the overall level of Sichuan province; the growth
rate is lower than the overall level of Sichuan province in 2004 and 2008. The
result from the Chart tells that the main reason is a decline in the industrial
sector competitiveness (U 04 = 0.9835 < 1, U 08 = 0.9716 < 1). During 2002 to
2009, the effective index of industrial result W is always greater than 1 which
indicates the industries with potential and high-speed increase are the majority
in Leshan economy. Moreover, regional overall economic structure is fine and
the structure makes a great contribution to economic growth while secondary
and tertiary industries are in the proportion of the optimization process step
by step. During 2002 to 2003, 2005 to 2007, 2009, the effect of regional com-
petition index U is greater than 1 which indicates the overall increasing trend
of every industrial department in Leshan is great. What’s more, the competi-

tiveness of secondary industry is more obvious while the primary industry and
tertiary industry cannot provide support for the regional overall competitive-
ness, so that the overall competitive advantages are not clear enough. When
the relative competitiveness of secondary industry is on decrease, the tertiary
industry is unable to provide effective support. In this situation, the secondary
industry makes a decisive effect on the overall competitiveness which means this
region depends on the secondary industry too much.
Table 5. 2010–2015 The SSM prediction of three industries of Leshan city
Index Leshan 2010 2011 2012 2013 2014 2015

Pij GDP1 20.446 13.502 15.463 17.703 20.26 23.178
GDP2 55.822 61.718 73.567 87.624 104.291 124.038
GDP3 18.011 23.898 28.195 33.268 39.259 46.332
W 1.006 1.006 1.006 1.006 1.0059
Dij GDP1 0.222 1.159 1.316 1.493 1.695 1.924
GDP2 32.094 5.082 6.242 7.665 9.414 11.561
GDP3 −5.145 1.512 1.765 2.06 2.405 2.807
U 1.0071 1.0071 1.0071 1.0071 1.0071
L 1.0131 1.0131 1.0131 1.0131 1.013
Choose the structure deviation component P [10] as abscissa and the com-
petitiveness deviation component D as ordinate to draw the SSM Picture of
Leshan’s three industries during 2010 to 2015.
Fig. 1. 2010–2015 The SSM analysis Picture of three industries of Leshan city
Table 5 and Fig. 1 indicates: during 2010–2015, the relative growth rate (L) is
greater than 1, meaning that the regional economic growth will be faster than the
overall level of Sichuan province during the period of the Twelfth Five-year; W is
136 H. Lan et al.
always greater than 1 which indicates the regional overall economic structure will
continue to the good development trend; further study of the structure deviation
component of every industrial department find that the structure deviation com-
ponent is the majority of economic part of every department. It proves that the
structure of department make a great contribution to the economy growth and
Leshan city already has some scale. At the same time, it is clear that compared
with the secondary industry, the scale of tertiary industry has a growing differ-
ence. The effectiveness index of competitiveness shows every sector has a strong
growth trend in total during the period of the Twelfth Five-year with a strong
competitiveness. With the further study of the competitiveness deviation com-
ponent, in addition to the obvious advantage of secondary industry, the primary
industry and tertiary industry are in the average level of entire province without
superiority. Still, this region depends on the secondary industry too much.
On the analysis of three industries of Leshan, we stress the further study
of tertiary industry and get the SSM appraisal result of the tertiary industry
during 2002–2009: the relative growth rate fluctuated, that is to say, there was
no obvious advantage in its economic growth compared with the level of Sichuan
province. The effect coefficient and the effect of regional competition index of
every department are fluctuated too which proves that the advantages in the eco-
nomic structure and competitiveness of tertiary industry are not obvious enough.
With further study, we find that the competitiveness deviation component of
tourism are negative generally which indicates the relative competitiveness of
Leshan’s tourism has not formed yet; there is a trend that the structure com-
ponent of transportation industry is optimized step by step although the scale
need to be further improved. Only in that way, an obvious economic effect will
exists. From Table 6 and Fig. 2, the speed of growth of Leshan’s tertiary industry
will be slower than the level of Sichuan province in the next 5 years. We find out
the result from the analysis of effect coefficient and the effect of regional com-
petition index. The reasons are two: first, the departments in tertiary industry
with high growth are minority; then, there is a non-performing growth trend of
Table 6. 2010–2015 The SSM prediction of tertiary industry of Leshan city
Index Leshan 2010 2011 2012 2013 2014 2015

Pij GDP31 28.113 28.723 33.919 40.059 47.315 55.892
GDP32 2.92 1.572 1.859 2.197 2.596 3.066
SXT 26.762 34.073 40.209 47.475 56.085 66.289
W 0.995 0.9941 0.9931 0.9922 0.9913
Dij GDP31 3.747 −2.183 −2.581 −3.051 −3.607 −4.264
GDP32 −1.302 2.782 3.278 3.863 4.552 5.364
SXT −2.206 −0.648 −0.761 −0.895 −1.051 −1.236
U 0.99993 0.99992 0.99991 0.9999 0.99989
L 0.9949 0.994 0.9931 0.9921 0.9912
Fig. 2. 2010–2015 The SSM analysis Picture of three industries of Leshan city
every industrial department. With further analysis, the competitiveness devia-

tion component of tourism and business are negative mainly which don’t reach
the average level of Sichuan province with a tendency of decrease year by year. By
the study of structure deviation component, transportation and storage postal
service make less contribution to economic growth with low level of structure
and small scale.
4 Conclusion and Suggestion
Based on shift-share model to evaluate the industry development, we can predict

the industrial development trend effectively with the help of GM(1.1) model. In
this way, it can reflect the possible problems in the industrial development in the
future more objectively and clearly which is convenient for related departments
to plan ahead and formulate the specific targeted policies or measures. On the
above analysis result, we suggest related departments to make some plans as
follows:
(1) Make an accurate development orientation. It should further strengthen the

development orientation of Leshan to be an important growth pole of eco-
nomic development in Sichuan, a secondary transportation hub in the west-
ern, a southern core of Chengdu economic zone and an international tourist
destination. The Twelfth Five-year Plan should be viewed as a period with
breakthrough. Guided by the theory of low-carbon economy development,
we can create a pattern of “low energy consumption, low emission, high
benefit”. If we speed up the pace of the adjustment of industrial structure,
the industrial structure of low carbon, cycle, ecological and efficiency will
formed quickly.
(2) Consolidate the leading status of secondary industry. Following with the
theory that “following a new industrialization approach, developing modern
industry system, promoting the combination of informatization and industri-
alization”, we should set a core strategy of economic development in Leshan
138 H. Lan et al.
which will promote the industrialization. Develop high-tech industry such as

polysilicon, electronic information and biomedicine who have high scientific
and technological content, low energy resources consumption and less pollu-
tion; Promote traditional superior industry and increase technology content;
Develop low carbon economy and promote clean production which stick to
the path to a new industrial city with high content of science and tech-
nology, good economic benefit, low resource consumption, less environment
pollution and human capital advantage.
(3) Strengthen the competitiveness of tertiary industry. For the great increase
in the proportion of the tertiary industry in the regional GDP, promote and
optimize all kinds of service industries; Give full play to the unique and rich
ecological tourism resources by developing ecological tourism; blossom the
producer services such as information industry, green Trade and financial
industry to build a new service system. Accelerate the implementation of
“facilitate thousand billion transportation project and the construction of
two flights four railways eight highways and one transportation hub” to
build a fully functional integrated transportation system. Only in this way,
the transport strength of Leshan will be promoted and the modern logistics
services will be developed.
(4) Make full use of the resources rationally and promote tourism development
of Leshan. Protect and make good use of the world cultural, natural her-
itage and bring this city’s superiority into full play with famous mountain,
famous Buddha, famous city and famous person. Prominent the three keys,
like Buddhist culture, natural ecological tourism development and the con-
struction management of scenic area. Improve the quality of ecological envi-
ronment and the cultural taste of the city which improve the function of
urban tourism. It’s obvious that above actions will help speed up the process
of the construction as a tourism city for Leshan.
Acknowledgements. This work was supported by the Humanities and Social Sci-
ences Foundation ot the Ministry of Education (Grant No. 16YJC30089), and the
Sichuan Center for science and Enterprise Development Research (Grant No. Xq16C04)
References
1. Artige L, Neuss L (2014) A new shift-share method. Growth Change 45(4):667–683
2. Fang L (2008) The shift-share analysis of jilin deviate from the industry’s compet-
itiveness. Master’s thesis, Northeast Normal University
3. Statistics Bureau of Leshan City (2015) Leshan statistical yearbook
4. Lin CT, Yang SY (2004) Forecast of the output value of taiwan’s ic industry using
the grey forecasting model. Int J Comput Appl Technol 19(1):23–27
5. Liu YP (2006) Regional economics shift-share analysis of Jiangxi Province. Sci
Mosaic 3:19–20
6. Otsuka A (2017) Determinants of regional energy demand: dynamic shift-share

analysis. In: Regional Energy Demand and Energy Efficiency in Japan. Springer,
pp 23–40
7. Shi J (2004) An analysis of the impact of sars on china. Digest Manage Sci 115:380–
383
8. Statistics Bureau of Leshan City (2016) http://www.sc.stats.gov.cn
9. Wu JA, Zhao X (2009) Application of the spatial shift-share analysis. Stat Res
21:64–68
10. Zhang X, Jiao B, Jun LI (2012) Effect of component deviation on structure and
infrared radiation properties of cordierite. Kuei Suan Jen Hsueh Pao/J Chin Ceram
Soc 40(3):357–361
11. Zhao JH, Chen JQ, Guo LZ (2007) Application of excel in prediction and model
checking of dynamic gm(1,1). J Qingdao Technol Univ 18:26–30
12. Zhou R, Li J (2015) The optimizing whitening differential equation of gm(1,1)
based non-homogenous exponential series. J Math Comput Sci 5(1):50–55
Scheduling Problem for Allocating Worker
with Class-Type Skill in JSP by Hybrid
Genetic Algorithm
Kenichi Ida1(B) , Daiki Takano1 , and Mitsuo Gen2

1
Graduate School of Engineering,
Maebashi Institute of Technology, Maebashi, Japan
ida@maebashi-it.ac.jp
2
Fuzzy Logic Systems Institute and Tokyo University of Science, Tokyo, Japan
Abstract. Scheduling in manufacturing systems is one of the most

important and complex combinatorial optimization problems, where it
can have a major impact on the productivity of a production process.
Moreover, most of manufacturing scheduling models fall into the class of
NP-hard combinatorial problems. In a real world manufacturing system,
a plurality of worker who operates the machine exists, depending on the
skill level by the workers for each machine and working time is different
even if same work on the same machine in job-shop scheduling problem
(JSP). Therefore, it is taking to account for differences in working time
by the worker is scheduling problem with worker allocation. In this paper,
in order to approach the more realistic model by dividing into several
class workers and to determine the skill level for each machine for each
class worker, we propose a new model that introduced the concept of
class-type skill and demonstrate the effectiveness of the computational
result by Hybrid Genetic Algorithm.
Keywords: Job-shop scheduling problem (JSP) · Scheduling prob-

lem for allocating worker (SPAW) · Class-type skill · Hybrid genetic
algorithm
1 Introduction
Machine scheduling problem (MSP) arises in diverse areas such as flexible man-
ufacturing system, production planning, computer design, logistics, communi-
cation, etc. A common feature of many of these problems is that no efficient
solution algorithm is known yet for solving it to optimality in polynomial time.
The classical job-shop scheduling problem (JSP) is one of the most well-known
MSP models. Informally, the JSP model can be described as follows: there are a
set of jobs and a set of machines. Each job consists of a chain of operations, each
of which needs to be processed during an uninterrupted time period of a given
length on a given machine. Each machine can process at most one operation at
Scheduling Problem for Allocating Worker with Class-Type Skill 141
a time. A schedule is an allocation of the operations to time intervals on the

machines. The problem is to find a schedule of minimum length [14].
In real world manufacturing systems there are many combinatorial opti-
mization problems (COP) imposing on more complex issues, such as complex
structure, non-linear constraints, and multiple objectives to be handled simul-
taneously. Manufacturing scheduling is one of the important and complex COP
models, where it can have a major impact on the productivity of a production
process. Moreover, the COP models make the problem intractable to the tra-
ditional optimization techniques be-cause most of scheduling problems fall into
the class of NP-hard combinatorial problems. In order to develop effective and
efficient solution algorithms that are in a sense “good”, i.e., whose computational
time is small as within 3 min, or at least reasonable for NP-hard combinatorial
problems met in practice, we have to consider three important issues: quality of
solution, computational time and effectiveness of the nondominated solutions for
multiobjective optimization problem (MOP) are very important issues [2,5,6].
In a typical job-shop scheduling problem (JSP), hitherto worker’s ability has
not been basically considered. Scheduling problem for allocating worker (SPAW)
by introducing the concept of skill in JSP as a parameter, thereby taking into
account the ability of the workers. And, it is defined as an objective function to
minimize the sum of the delay time for the delivery time of the completion time
of each job.
Generally JSP is known as NP-hard problem which cannot be calculated in
actual SPAW [3,4]. When all the operator’s skill level was 1.00, it can be con-
sidered equivalent to the JSP. In other words, SPAW can be said that the gen-
eralized problem of the JSP. So, the SPAW is also NP-hard. Recently Gen [8,9]
surveyed multiobjective hybrid genetic algorithms for manufacturing scheduling.
Therefore, genetic algorithm has been noted that a metaheuristic approach is as
a solution to the problem of the SPAW in JSP. Then the solution methods have
been proposed to solve the SPAW models [11,13,15].
In this paper, we propose a new model introduced the concept of skill value
by Kitada [12] which used for the nurse scheduling problem (NSP) considering
SPAW based on Iima [11]. And they used the GA algorithm as a solution of the
model.
The paper is organized as follows: Sect. 2 presents Scheduling Problem for
Allocating Worker (SPAW); Sect. 3 describes the detailed of SPWA with the
class-type skill by hybrid genetic algorithm(HGA); Sect. 4 gives a numerical
experiments results; finally, the conclusion are drawn in Sect. 5.
2 Scheduling Problem for Allocating Worker (SPAW)

2.1 Workers Skill Level
In the actual production site, there is a worker to operate the machine, and the
difference in processing time of work is produced in accordance with skill level
on the mechanical of each worker. We introduce a skill level as a parameter to
account for the difference in processing time of the operation by the worker.
142 K. Ida et al.
By using this parameter, each worker was placed in each machine, and it is pos-
sible to determine the processing time in consideration of the worker’s abilities.
In addition, we can create a more realistic schedule.
(1) Processing time in consideration of skill level
When assigning N workers Wn (n = 1, 2, · · · , N ) to the I products Ai (i =
1, 2, · · · , I) using K machines Mk (k = 1, 2, · · · , K), skill level Sn (k) for each
worker Wn is set as the ability to operate the machine Mk .
In addition, for each product Ai (i = 1, 2, · · · , I), there are Ji operations writ-
ten by each one Oij (j = 1, 2, · · · , Ji ) to be processed by the machineMRij (Rij ∈
1, 2, · · · , K) which is determined in advance. The processing time given to each
operation Oij (processing time when the skill level is 1.00) is defined as the
standard processing time P Tij . The processing time ptnij of operation Oij by the
worker Wn is represented by the following equation, and the actual processing
time expands and contracts depending on the skill level [11]:
P Tij
ptnij = . (1)
Sn (Rij )
The proposed worker skill level in SPAW by Iima and Sannomiya [11] is
determined in a range from 0.00 to 1.00 and are given in advance as skill level
table. In addition, the value of the skill level for the worker of the machine is 0.00
cannot operate the machine. Processing time on each machine by each worker
will be determined using the skill level table.
(2) Setting of the Worker Skill Level
Concept of worker’s skill level is twofold. One is uniform proficiency type of
skill level that has been proposed by Iima [11]. Another is variety proficiency
type of skill level that has been proposed by Osawa [15].
Uniform proficiency type of skill level is different for each machine. On the
other hand, variety proficiency type of skill level has a skill level that each worker
is different for each machine. An example of the skill level table as shown in Fig. 1.
Fig. 1. Setting an example of the skill level of worker

2.2 Placement of Workers
The overall schedule is divided into P periods SPp (p = 1, 2, · · · , P ) and it sets

the length of one period T time. At this time, each worker working at some
periods of the P number of periods. For placement of workers, the following
constraints exist [11].
(1) Each worker can operate the machine only with placement period.
(2) Each worker can operate one machine only at the same time.
(3) The number of workers are placed equal to the total number of machine in
each period.
(4) If the work is straddling the two periods, the rest of the work will be taken
over in succession to the worker after the replacement. At this time, the
interruption of work at the time of change is not taken into consideration.
In SPAW proposed by Iima [11], the entire schedule is dividing to 10 periods

and the length of one period is set to 8 hours. If the worker is placed skill
level of 0.00 against the machine, the worker cannot operate the machine. Thus,
the scheduling becomes infeasible. In order to avoid this, Osawa and Ida [15]
proposed an infeasible solution correction algorithm.
2.3 Set of Delivery Time
In this paper, it is defined as an objective function to minimize the total of

the late delivery time of each job. In fact, delivery time of each job is set by
consultation with the customer. However, when performing experiments with an
existing instance, the customer is not present. Therefore, it is necessary to set
the delivery time convenience. In this paper, we carried out the delivery time
settings for each job by using a delivery time coefficient [1,16].
Delivery setting using the delivery time coefficient is a technique that is used
in delivery time with JSP. The value obtained by multiplying the delivery time
coefficient in total processing time of each job is set as the delivery time.
In Asano [1] and Singer and Pinedo [16] of the experiment is set to delivery
time coefficient F = 1.5 and 1.6. In addition, it has given a late delivery penalty
in the case of delay in delivery occurred in each job.
3 SPAW with the Class-Type Skill by HGA
3.1 Introduction of the Class-Type Skill

In SPAW, skill level has been set in the range of 0.00 to 1.00. Skill level setting
can be considered only if processing time than standard processing time is longer.
However, in actual production site, worker who can handle the work in a short
period of time than the standard processing time is present. In this paper, we
propose a new model, which was introduced the concept of skill value Kitada
[12] was used for the NSP in SPAW that Iima [11] proposed.
144 K. Ida et al.
Divided into several class workers, to determine the skill level for each
machine in each worker for each class, considering the case that can handle
jobs in a shorter time than the standard processing time.
In Kitada [12] skills value table used in the NSP, the nurse of the leader,
veteran, mid-level, second year, newcomers has been divided into five classes.
For this reason, divided workers into five classes in this paper, to set the range
of skill level in each class. Range of skill level of each class, the leader from 1.00
to 1.40, a veteran from 0.90 to 1.30, mid-level from 0.80 to 1.20, second year
from 0.60 to 1.00, newcomer from 0.30 to 0.60. To determine the skill level for
each machine of the workers in each of the ranges for each class. In addition, the
number of workers of each class was determined based on the number ratio of
each class in the skill value table of Kitada [12].
3.2 Creating Worker Group in Consideration of Average Skill Level

In SPAW of Iima [11], workers who work in each period have been prepared in
advance as the worker group, and the worker group is divided into three. How-
ever, the group who work in two period (16 hours) consecutive are present in the
worker group created by the Iima [11]. This is not desirable from the viewpoint
of the Labor Standards Law. If you create a worker group in advance, reduced
diversity of individuals, and the possibility that initial convergence occurs is
high. In this paper, we set the total number of workers is three times of the total
number of machines, and three of the worker group without creating in advance,
we set to be created for each individual in each trial.
The creation of the worker group to use the concept of uniform proficiency
type of skill level proposed by Iima [11]. Therefore, the worker has a skill level
for the worker group in each class. Skill level for the worker group of each class
was set the leader 1.20, veteran 1.10, mid-level 1.00, second year 0.80, and rookie
0.50. In addition, in order to prevent bias in the class of workers for each worker
group, the average skill level of the worker group became when less than 0.9 will
be re-create a worker group. Showing the procedure of a worker group created
in consideration of the average skill level of each period of the total number of
periods P as follows:
Step 1. Set i = 1.
Step 2. Select one worker at random from the operator that is not arranged, and
to place the selected workers in the period i.
Step 3. Exclude worker placed in Step 2 from the arrangement of the choices
Step 4. If the worker of the same number as the total machine numbers are
arranged in period i, the process proceeds to Step 5. Otherwise, return to
Step 2.
Step 5. If it is i = 3, the process advances to Step 6. Otherwise, as i = i + 1,
and return to Step 2.
Step 6. If the periods 1 to 3 skill level average of all of the worker group was 0.9
or more, the process proceeds to Step 7. Otherwise, initialize the unplaced all
workers which are disposed in the periods 1–3, and returns to Step 1
Step 7. Assign the worker group of the period from 1 to 3 in order to each period
of the period from 4 to P , and exit when assigned a worker group to period P .
3.3 Decision of Machine Skill Level 0 in Consideration of the Class

In the actual production site, there is a worker who cannot operate the specific
machine (skill level 0.00). To account for this, the Iima [11] and Osawa [15]
was set percentage of a machine skilled level 0 in advance, and based on it,
when performing experiments determines the machine each worker’s skill level at
random becomes 0.However, in this paper, there is more than one class to worker
by introducing the Class-type skill. Therefore, when using a similar method as
Iima [11] and Osawa [15], for example, leader may become greater the number
of machines that cannot operate than newcomers, and it is not realistic setting.
In this paper, to determine machine skill level 0 in consideration of the class of
workers.
An example of skill level table of the class-type skill, including a machine of
skill level 0 in consideration of the class shows in Table 1.
Table 1. Skill level table using a class type skill level
Class M1 M2 M3
Leader 1.01 1.28 1.15
Veteran 1.13 0.93 1.29
Mid-level 0 0.95 1.08
2nd year 0.66 0 0.98
Newcomer 0 0 0.5
In this paper, the leader and the veteran can operate all of the machine, and
mid-level, second year, newcomer is set as the operator cannot operate the part
of the machine (skill level 0). In the experiment, the proportion of the worker,
mid-level 10–30%, second year 20–40%, and newcomer 40–60%, is determined at
random in advance.
3.4 SPAW with the Class-Type Skill by GA

Recently, many researchers tried to adapt different metaheuristic approaches
such as Genetic Algorithm (GA), Tabu Search (TS), Simulated Annealing
(SA) for solving job-shop scheduling problem (JSP). Because of its inherent
intractability, metaheuristic procedures are an attractive alternative (Gen [7]).
To improve the efficiency for finding better solutions in searching space, some
special technical local searches have been adapted in JSP. Osawa [15] reformed
the traditional left shift to short the idle time, and formulated an algorithm
called Eshift. Goncalves et al. proposed another technique based on the critical
146 K. Ida et al.
path to confirm the search space, and swapped the operations in critical block,
and this approach can also improve the efficiency of algorithm for finding active
schedule (Gen [7]).
In GA for SPAW, it represents one solution using two types of chromosomes.
One is a job permutation chromosome representing a processing order for each
task of a job, the other is a worker placement chromosome representing the
arrangement of the worker to operate the machine at each period. Therefore,
genetic operations such as crossover and mutation is carried out for each of the
chromosome.
In this paper, we developed hybrid genetic algorithm (HGA) for solving
scheduling problem to allocate worker with class-type skill in JSP as shown
in a pseudo code in Fig. 2.
Fig. 2. A pseudo code of HGA for SPAW with class-type skill
3.5 Initial Population
In Osawa [15] of the technique, the initial population of job permutation chro-
mosome of each individual are randomly generated. In GA, superior solution can
be obtained by repeating the genetic operations such as crossover and mutation.
Therefore, it is required that to have diversity in the chromosome in the initial
population stage, and eliminate the diversity operation is cause to cause initial
convergence.
However, when starting the operation of GA from the chromosome of the
initial population randomly generated, there is a possibility that the improve-
ment of solution stagnates. Further, since the randomly generated, it is difficult
to obtain stable and excellent solution.
In this paper, we propose the initial population generation method, including

the two types of chromosomes. Type one is a chromosome to be stored in the
random and priority from the early work of the delivery time, and type two is a
chromosome to be stored in the random and priority from the slow work of the
delivery time. Initial population is generated by these chromosomes by encoding
routine is the proportion of each half. This initial population generation method,
while maintaining the diversity of the initial population, it is possible to have
the characteristics and trends for each chromosome. This technique, because
improved based on the delivery time, which is a method that can be applied to
general delivery time with JSP.
3.6 Genetic Operation

(1) Job Permutation Chromosome
Osawa [15], for the job permutation chromosome, is using the proposed 2-
point crossover of Hirano [10] as a crossover method, and it uses a critical block
mutation as a mutation. In this paper, we introduce the concept of delivery time
to 2-point crossover of the Hirano [10], and use a crossover technique to store
the large work of delivery time delay time on the left side of the chromosome.
This technique, because improved based on the delivery time, which is a method
that can be applied to general delivery time with JSP. In addition, mutations
uses Osawa [15] same critical block mutation.
(2) Worker Placement Chromosome
Osawa [15] is using only the crossover method to worker placement chromo-
some. That approach uses a 2-point crossover of Hirano [10] like the job permu-
tation chromosome. In this paper, it improved based on the 2-point crossover
of Hirano [10] to worker placement chromosome. Its crossover technique has a
method considering skill level and total processing time for each machine in each
period.
4 Numerical Experiment
4.1 Experimental Data
Make a comparison experiments with the method (oGA) by Osawa [15] with
using the proposed method (pGA) in this paper. Because the Osawa [15] app-
roach had incorporated worker group as data, the creation of worker group is
using the technique proposed in this paper.
Scheduling of instances considering variety of worker skill level, such as deal-
ing with in this paper, does not exist as far as the authors know. Therefore, as an
instance in this experiment, to use instances for JSP (10job-10machine problem
la16-la20, ft10). Total machine number is 10, so the total number of workers is
set 30. Upon the experiment, the worker’s skill level is used a randomly gener-
ated skill level in the range specified by the class-type skill level. Worker ratio of
each class will be set based on the number ratio of each class in the skill value
table of Kitada [5].
148 K. Ida et al.
If the skill level for the operator of the machine is zero, the method of Osawa
[15] shown in Subsect. 3.3 is used. In addition, delivery time of each job is set
using the delivery time coefficient F, F was defined as F = 1.3 by preliminary
experiment.
Various parameters for the GA are set, the number of attempts 50 times,
the population size 100, crossover probability 0.8, and mutation probability 0.2.
Termination condition is when the number of evaluation individual has reached
one million individuals, or delivery delay time total was set to when it becomes
0. Machine specifications are, Intel Core i5-4430 3.00 GHz, memory 8.00 GB.
Experiments carried out in Microsoft Visual C++ 2010 on PC.
Performing a comparison by the minimum total delivery delay time (Best) in
all the attempts and the average value of the minimum total delivery delay time
(Average) obtained in each trial for each instance. The results were as shown in
Table 2 (number in parentheses represents the number of times when the best
solution is obtained more than once).
Table 2. Experimental result
Instance Best Average

oGA pGA oGA pGA
la16 60 52 135 90
la17 5 1 36 17
la18 30 14 85 55
la19 26 7 77 39
la20 0(17) 0(47) 6 0
4.2 Consideration
From Table 2, significantly better solution than the method of Osawa [15] is
obtained in all instances. Figure 3 shows a Gantt chart of the best solution for
the most improved instance.
In particular, the instances, la18 and la19 has become delivery delay time
following conventional half. In addition, it can be seen that from the experimental
results of the Average, the solution accuracy of each attempt in all instances has
been greatly improved. However, it is also true that still delivery delay time
occurs in these instances. In particular ft10, the width of the delay in delivery
compared with the other instance has become a big thing. This is thought to be
due to features of ft10. ft10 is biased machine to be used in the processing of
the first few of the work of the job. Thus, even if most job can protect delivery
time, jobs that delivery delay time projecting to occur are present. Therefore, it
is required improvement inconsideration of the characteristics of the instances.
Fig. 3. ft10 Gantt chart b oGA and pGA
5 Conclusion
In this paper, scheduling problem for allocating worker (SPAW) with class-type
skill in JSP expanded to the more realistic model by using Iima [11] and improve-
ments were made of various types of algorithm using Osawa [15]. As a result, it
was able to reduce the delivery delay time in all the instances that have occurred
of delivery delay, and solving accuracy was able to present a schedule improved.
Further, by introducing the class-type skill, divided into several class workers, to
determine the skill level for each machine in each worker for each class, consider-
ing the case that can handle jobs in a shorter time than the standard processing
time, and proposed SPAW of Iima [11] could be bought close to realistic models.
However, if the product had been completed earlier than the delivery time,
in fact, there is a need to store the product until the delivery time. This leads to
an extra cost, so it is necessary to consider in order to create a realistic schedule.
The same applies to the total processing time. Therefore, after strictly to defend
delivery time, it is necessary to introduce the inventory control as the second
target and the total processing time as the third target.
Acknowledgements. This research work is supported by the Grant-in-Aid for Sci-

entific Research (C) of Japan Society of Promotion of Science (JSPS): No. 15K00357.
References
1. Mand AK, Hatanaka OH (2003) A heuristic algorithm for job-shop scheduling
to minimize total weighted tardiness. J Jpn Ind Manage Assoc 51(3):245–253 (In
Japanese)
2. Cheng R, Gen M, Tsujimura Y (1996) A tutorial survey of job-shop schedul-
ing problems using genetic algorithmsli: representation. Comput Ind Eng 30(4):
983–997
150 K. Ida et al.
3. Cheng R, Gen M, Tsujimura Y (1999) A tutorial survey of job-shop scheduling

problems using genetic algorithms: Part II. hybrid genetic search strategies. Com-
put Ind Eng 37(1C2):51–55
4. Gen M, Cheng R (1997) Genetic Algorithms and Engineering Design. Wiley, New
York
5. Gen M, Lin L (2014) Multiobjective evolutionary algorithm for manufacturing
scheduling problems: state-of-the-art survey. J Intell Manufact 25(5):849–866
6. Gen M, Tsujimura Y, Kubota E (1994) Solving job-shop scheduling problems by
genetic algorithm. In: IEEE International Conference on Systems, Man, and Cyber-
netics, 1994, Humans, Information and Technology, vol 2, pp 1577–1582
7. Gen M, Cheng R, Lin L (2008) Network models and optimization: Multiobjective
genetic algorithm approach
8. Gen M, Lin L, Zhang W (2015) Multiobjective Hybrid Genetic Algorithms for
Manufacturing Scheduling: Part I Models and Algorithms. Advances in Intelligent
Systems and Computing. Springer, Heidelberg
9. Gen M, Zhang W, Lin L (2015) Multiobjective Hybrid Genetic Algorithms for
Manufacturing Scheduling: Part II Case Studies of HDD and TFT-LCD. Advances
in Intelligent Systems and Computing. Springer, Heidelberg
10. Hirano H (1995) Genetic algorithm with cluster averaging method for solving job-
shop scheduling sroblems. J Jpn Soc Artif Intell 27(5):769–777 (In Japanese)
11. Iima H, Sannomiya N (2002) Proposition of module type genetic algorithm and its
application to modified scheduling problems with worker allocation. IEEJ Trans
Electron Inf Syst 122C(3):409–416 (In Japanese)
12. Kitada M (2014) A study on a nurse scheduling system for unexpected absence of
nurses. PhD thesis, Graduate School of Osaka Prefecture University, (In Japanese)
13. Kurashima Y (2013) Scheduling problem with worker allocation for FSP. Master’s
thesis, Maebashi Institute of Technology, (In Japanese)
14. Lawler EL, Lenstra JK (1993) Sequencing and scheduling: algorithms and com-
plexity. Handb Ope Res Manage Sci 4(2):445–522
15. Osawa A, Ida K (2007) A solution method of scheduling problem with worker
allocation by a genetic algorithm. IEEJ Trans Electron Inf Syst 127(5):755–761
(In Japanese)
16. Singer M, Pinedo M (1998) A computational study of branch and bound techniques
for minimizing the total weighted tardiness in job shops. IIE Trans 30(2):109–118
The Rules Determination of Numerical
Association Rule Mining Optimization by Using
Combination of PSO and Cauchy Distribution
Imam Tahyudin1,2(B) and Hidetaka Nambo1

1
Artificial Intelligence Laboratory, Division of Electrical Engineering
and Computer Science, Graduate School of Natural Science and Technology,
Kanazawa University, Kanazawa, Japan
imam@blitz.ec.t.kanazawa-u.ac.jp
2
Department of Information System, STMIK AMIKOM Purwokerto,
Purwokerto, Indonesia
Abstract. One of the optimization methods to solve the numerical asso-

ciation rule mining problem is particle swarm optimization (PSO). This
method is popularly used in various fields such as in the job scheduling
problem, evaluating stock market, inferring gen regulatory networks and
numerical association rule mining optimization. The weakness of the PSO
is often premature for searching the optimal solution because it traps in
local optima when the best particle is being searched in every iteration.
Combining the PSO with Cauchy distribution for numerical association
rule mining problem (PARCD) is a solution because it is robust for find-
ing the optimal solution in a large neighborhood. The important point
in this proposed method is particle representation which to know the
association between one attribute to another. Therefore, this study has
the aim to determinate rules of numerical association rule mining and
also to calculate the multi-objective function using combination of PSO
and Cauchy distribution. The results show that all of them explain every
attribute to formulate the rule well. In addition, the multi-objective func-
tion value of PARCD method generally produces results which better
than the previous method, MOPAR.
Keywords: PSO · Cauchy distribution · Numerical association rule

mining · Particle representation · Multi-objective functions · PARCD ·
MOPAR
1 Introduction
The issue of numerical association rule mining is interesting to discuss. This prob-
lem already solved by many researchers with many approaches. For instance dis-
cretization [14], distribution and optimization [1,2,6,13,18]. The last approach is
the interesting one because it is said as the recent solution to overcome this problem.

152 I. Tahyudin and H. Nambo
According to [6,18] revealed that this solution can find the multi-objectives func-
tion such as support, confidence, comprehensibility, interestingness, and amplitude
without transformation of numerical dataset to categorical type.
One of the methods in optimization approach is MOPAR method [6]. This
method used the PSO (particle swarm optimization) for solving the numerical
ARM problem. This method already developed by Tahyudin and Nambo [18] by
using combination of PSO and Cauchy distribution because of the weakness of
PSO which trapped in local optima when the iteration goes to infinite then the
velocity particle lead to 0. This condition make the PSO does not have capability
to find the optimal solution [12]. Therefore, the combination of PSO with Cauchy
distribution (PARCD) is the answer for the previous method.
The important step for this method is particle representation to determine
the rules. This step as the initialization process before calculating the multi-
objective function. There are two approach methods, firstly is Pittsburgh method
which is determine the particle as the rule set and secondly, the Michigan app-
roach method which explains that one particle refer to one rule [6]. This research
uses the second method because of being appropriate with the kind of case.
Moreover, the particle representation contains three main parts, they are ACN
(antecedent, consequent and none of them), the lower bound and the upper
bound. Hence, the aim of this research is to determine the rules and to calculate
the multi-objective function values which are support, confidence, comprehensi-
bility and interestingness using PARCD method.
The research is organized as follows: Sect. 2 reviews the literature of recent
research; Sect. 3 presents the proposed method of combining PSO and Cauchy dis-
tribution for numerical association rule mining optimization (PARCD) and the
procedure to determine the particle representation; Sect. 4 gives an experiment and
discussion for determining of particle representation and also the result of multi
objective function; finally, the conclusion and future work are given in Sect. 5.
2 Literature Review
Minaei-Bidgoli et al. [13] demonstrated that the numerical association rule prob-
lem can be solved by discretization, distribution and optimization. The dis-
cretization is performed by partitioning and combining, clustering and fuzzy
logic routines [3,5]. Then, the optimization is approached by optimized ARM
[20], differential evolution [2], Genetic Algorithm (GA) [10,13,15] and Particle
swarm optimization (PSO) [1,6,9]. Using PSO techniques, the numerical data
can be solved to attain the important information without the discretization
process [4,6] and some methods can automatically determine the minimum sup-
port and minimum confidence based on the optimal threshold [15,17,20].
On the other hand, the PSO method has the weakness that the user has to
specify the number of “best rules” and the time of complexity [17] and also it
is not robust when used on large data sets [7,12]. One of the ways to diminish
these weakness is revealed in [12], that the combination of PSO with Cauchy
has approved can rise the leverage of result because the mutation process can
Numerical Association Rule Mining Optimization 153
reach wider and appropriate to a large database. In another research work [7]
the combination of approaches has the ability to optimize two-stage reentrant
flexible flow shop with blocking constraints. In addition, this combination can
improve the interval solution by the average of 15.60% and then the perfor-
mance of this combination higher than Hybrid Genetic Algorithm (HGA) [16].
Then, this combination was used to optimize the Integration of Process Planning
and Scheduling (IPPS) and the result shows the effectiveness of the proposed
IPPS method and the reactive scheduling method [21]. This hybrid method was
developed by [7,16] to increase the wide search space in mutation process by
using Cauchy distribution. His result shows that the method can enhance the
evolutionary process with the wide search space.
3 Proposed Method
3.1 The Particle Representation
The rules of numerical association rule mining by PARCD will be obtained by
the particle representation procedure. This study used Michigan method which
determine for every particle referring to one rule [6], for which the data set will
be extracted into ACN category, based on the value of lower bound and upper
bound. Antecedent is pre condition and consequent is conclusion for describing
a rule. The PARCD method can classify automatically the ACN based on the
optimal threshold in every rules. This concept can be showed clearly by Fig. 1.
Attribute 1 Attribute 2 …… Attribute n

ACNi LBi UBi ACNi LBi UBi ACNn LBn UBn
Fig. 1. The particle representation
If the optimal procedure for one rule are 0 ≤ ACNi ≤ 0.33 for antecedent,
0.34 ≤ ACNi ≤ 0.66 for consequent and 0.67 ≤ ACNi ≤ 1.00 for none of them.
For instance, see Table 1.
Table 1. Example of particle representation
Attribute ACNi LBii U Bii

A 0.32 2.4 6.7
B 0.14 1.7 27.8
C 0.71 0.23 135.1
D 0.66 0.11 78.9
According to the Table 1 the attribute A and B are the antecedent and
the attribute D is consequent. The attribute C is not appearing because it not
includes both of them. Therefore, the rule is AB → D.
3.2 Objective Design

This study uses multi-objective function which are support, confidence, com-
prehensibility and interestingness. The support function measures the ratio of
transactions in D containing X, or Support(x) = |x(t)|/|D|. Witten et al. [19]
explained that support is refer to coverage utility. It means that the support
function can calculate the percentage of number of attributes or their combi-
nation which have the association in every data set. The support of the rule
X → Y is computed using the following equation.
|X ∪ Y |
Support(X ∪ Y ) = . (1)
|D|
The confidence measures the quality of rule based on the number of transac-
tion of an AR in the whole dataset. In other words, the confidence describes the
accuracy [19]. The rule which often emerge in every transaction is considered to
have a better quality because it leads to good accuracy [6].
Support(X ∪ Y )
Confidence(X ∪ Y ) = . (2)
Support(X)
However, this criterion does not guarantee to obtain the appropriate AR,
individually. In order to gain the suitable coverage and reliability the resultant
rule must also be comprehensible and interesting. According to the previous
research, the smaller number of conditions in the antecedent part of a rule, the
more comprehensible or understandable is that rule [8]. Hence, the comprehen-
sibility can be measured as below
log(1 + |Y |)
Comprehensibility = , (3)
log(1 + |X ∪ Y |)
where |Y | is the number of rules in the consequent the and |X ∪ Y | is the total
number of rules in the consequent and antecedent rules.
The interestingness criterion is used for obtaining hidden information by
extracting of such surprising rules. This criterion is based on the support count
of both antecedent and consequent parts [8]. The equation is shown in Eq. (4).

Support(X ∪ Y ) Support(X ∪ Y ) Support(X ∪ Y )
Interestingness = × × 1− . (4)
Support(X) Support(Y ) |D|
This formula consists of three parts. In the first part, [Support(X ∪ Y )/

Support(X)], the generation probability of the rule is computed in terms of the
antecedent part of the rule. In the second part, [Support(X ∪ Y )/Support(Y )],
the generation probability of the rule is computed in terms of the consequence
part of the rule. Finally, in the third part, [1 − Support(X ∪ Y )/|D|], the section
Support(X ∪ Y )/|D| shows the probability of generating the rule according to
the total number of records in the dataset (|D|). So, its complement, [1−Support
(X ∪ Y )/|D|], means the probability of not generating the rule. Therefore, a rule
with a high value of support count will be considered as a less interesting rule [6].
3.3 PSO
The PSO method discovered by Kennedy and Eberhart in 1995. They are animal
psychologist and electrical engineer respectively which observed the swarming
behaviors in flocks of birds, schools of fish, or swarms of bees, and even human
social behavior [11].
The main concept of PSO is to initialize with a group of random particles
(solutions) and then to search for optima by updating generations. During all
iterations, each particle is updated by following two “best” values. The first one
is the best solution (fitness) has achieved so far. This value is called “pBest”.
The other “best” value that is tracked by the particle swarm optimizer is the
best value obtained so far by any particle in the population. This best value is a
global best and is called “gBest”. After finding the two best values; each particle
updates its corresponding velocity and position [9]. Each particle p, at some iter-
ation t, has a position x(t), and a displacement velocity v(t). The personal best
(pBest) and global best (gBest) positions are stored in the associated memory.
The velocity and position are updated using Eqs. (5) and (6) respectively [9,12].
vinew = ωviold + c1 rand()(pBest − xi ) + c2 rand()(gBest − xi ), (5)
xnew
i = xold
i + vinew , (6)
here ω is the inertia weight; vi old is the particle velocity of ith particle before
updating; vi new is the particle velocity of ith particle after updating; xi is the
ith , or current particle; i is the particle’s number; rand() is a random number in
the range (0, 1); c1 is the individual factor; c2 is the societal factor; pBest is the
particle best; gBest is the global best Particle velocities on every dimension are
clamped to a maximum velocity Vmax [9,12].
3.4 PSO for Numerical Association Rule Mining with Cauchy

Distribution (PARCD)
PARCD is the proposed method to implement for solving numerical association
rule mining problem. This combination generates the best result because the
weakness of PSO has solved by Cauchy distribution so it can prevent the trap
of local optima in best particle to gain the global best by long jump searching
using Cauchy mutation.
vi (t + 1) = ω(t)vi (t) + c1 rand()(pBest − xi (t)) + c2 rand()(gBest − xi (t)), (7)
vi (t + 1)
ui (t + 1) = , (8)
(vi1 (t + 1)2 + vi2 (t + 1)2 + · · · + viK (t + 1))2
si (t + 1) = ui (t + 1) × tan(π/2 · rand[0, 1)), (9)
xi (t + 1) = xi (t) + si (t + 1). (10)
3.5 Pseudocode and Flowchart of PARCD

This pseudocode and flowchart depict that the algorithm begins by initializing
the velocity vector and position randomly then it calculates the multi-objective
functions as the current fitness. After that it runs by looping iteration to looking
for pBest until it finds gBest value as the optimal solution.
Pseudocode of PSO [7] is shown as following, and flowchart is as Fig. 2:
Fig. 2. Flowchart of PSO


4.1 Experimental Setup
This experiment uses benchmark datasets from Bilkent University Function
Approximation Repository. There are five datasets which are used, Quake, Bas-
ketball, Body fat, pollution and Bolt (Table 2). This experiment is conducted on
Intel Core i5, 8 GB main memory, running by Windows 7 and the algorithms
are processed by using Matlab software.
We set up the variable of some values in the proposed algorithm such as
parameter, population size, external repository size, number of iteration, the
value of C1 and C2 , W , velocity limit and xRank. They are average, 40, 100,
2000, 2, 0.63, 3.83, and 13.33 respectively. This parameters are referred according
to the previous research of [6] reported in Table 3.
Table 2. Properties of the datasets
Dataset No. of records No. of attributes

Quake 2178 4
Basketball 96 5
Body fat 252 19
Bolt 40 8
Pollution 60 16
Table 3. The variable setup
Parameter Population External Number of C1 and C2 W Velocity xRank

size repository size iteration limit
Average 40 100 2000 2 0.63 3.83 13.33
4.2 Experiments
(1) The output of particle representation
This output is generated by every running times. And then in this experiment
shows the 20th running time which in every running contains 2000 rules. The 20th
running time is the last time so it produces the optimal solution. The results of
ACN of quake data set which are the rule number, ACN and LB and UB in head
title of table is shown in Table 4. According to the result in rule 1, the attributes
as the antecedent are focal depth and latitude. This first rule result is same
like the second rule then the last rule, rule 2000, the antecedent attributes are
latitude, longitude and Richter. However, the consequent attributes are almost
all none.
Table 4. ACN of quake dataset
Rule number (particle) Antecedent, consequent, Lower bound (LB) < attribute
none of them (ACN) < Upper bound (UB)
Rule 1 Antecedent 290.028451 < Att1 < 316.467965
9.068816 < Att2 < 21.329102
Consequence None
9.068816 < Att2 < 21.329102
Consequence None
··· ··· ···
··· ··· ···
33.959986 < Att3 < 71.573894
6.418029 < Att4 < 6.606453
Consequent None
Note: Att1 = focal depth, Att2 = latitude, Att3 = longitude, Att4 = Richter (target).
The result phenomena in Basket Ball Data set (Table 5) is almost similar
to Quake data set, in which almost the rules have no consequent attributes.
The first rule shows that the antecedent attributes are the number of assist per
minute, height and time played. The result of second rule also shows the same
as first rule, then the last rule shows that antecedent attributes are height and
time played. The result of ACN in Quake and Basketball data set are interesting
to study deeply, because generally the optimal solution has the antecedent and
consequent attribute.
In Table 6 the result of ACN of Body Fat data set depicts the complete para-
meter either antecedent or consequent. In the rule 1 we see that there are eight
attributes as the antecedent and three attributes as the consequents. In the rule
2 the number of antecedent and consequent attributes are the same as rule 1and
in the last rule the number of antecedent and consequent attributes are six and
two attributes respectively. The antecedent attributes in rule 1 are case num-
ber, Percent body fat using Siri’s equation, density, age, adiposity index, chest
circumference, Abdomen circumference and thigh circumference. Next, the con-
sequent attributes are Percent body fat using Brozek’s equation, height and hip
circumference. Then, in rule 2 the antecedent and the consequent attributes
are the same as rule 1. So that the rules 1 and 2 are if (att1, att3, att4, att5,
att8, att11, att12, att14) then (att2, att7, att13). And then, in the rule 2000, the
antecedent attributes are Percent body fat using Brozek’s equation, Percent body
fat using Siri’s equation, density, height, neck circumference and knee circum-
ference. Next, the consequent attributes are case number and weight. Therefore,
the rule 2000 is if (att2, att3, att4, att7, att10, att15) then (att1, att6).
Table 7 explains the experiment result from Bolt data set which has eight
attributes; run, speed, total, speed2, number2, Sens, time and T20 Bolt. Based on
Table 5. ACN of basket ball dataset
none of them (ACN) < upper bound (UB)
186.522680 < Att2 < 195.865376
14.763501 < Att2 < 20.155824
Consequence None
186.522680 < Att2 < 195.865376
14.763501 < Att2 < 20.155824
Consequence None
··· ··· ···
··· ··· ···
24.076966 < Att3 < 39.254520
Consequent None
Note: Att1 = Assists per minute, Att2 = height, Att3 = time played, Att4 = age,
Att5 = points per minute (target).
the table, two first rules show the same result both of antecedent and consequent.
The antecedent attributes are total and time while the consequent attributes are
run and speed1. Therefore, the rule is if (total, time) then (run, speed1).
The other rules do not report in this paper because they are so large, only
three rules; the first, the second and the last one, rule 2000. So that, the rule
2000 shows that the antecedent attributes are run and speed 2. However, the
consequent attribute is similar like quake and basket ball datasets, which is
unknown. Hence. This rule can not be declared clearly because it does not have
conclusion.
Table 8 depicts the result of rules from pollution dataset by particle repre-
sentation PARCD method. The result from the first and the second rules are
the same which the antecedent attributes are JANT, EDUC, NONW, WWDRK
while the consequent attributes are PREC, JULT, OVR65, DENS and HUMID.
So that, the rule is if (JANT, EDUC, NONW, WWDRK) then (PREC, JULT,
OVR65, DENS and HUMID).
The rule 2000 has the ACN result which is different from the first and the
second attributes. The antecedent attributes of rule 2000 are JANT, OVR65,
HOUS, POOR, HC and HUMID while its consequent attributes are POPN,
EDUC, DENS, NOX, SO@. The final rule is if (JANT, OVR65, HOUS, POOR,
HC) then (POPN, EDUC, DENS, NOX, SO@).
(2) The output of multi-objective function
Table 9 reveals a comparison for the output of Multi-objective function among
the proposed method, PARCD and MOPAR method. This table compares two
objective functions which are support and confidence from both of methods in
Table 6. ACN of body fat dataset
Rule number (particle) Antecedent, consequent, Lower bound (LB) < attribute <
none of them (ACN) upper bound (UB)
57.988435 < Att3 < 69.574945
309.987803 < Att4 < 314.218245
55.294719 < Att5 < 66.896106
136.234441 < Att8 < 138.744999
40.927433 < Att11 < 41.562953
20.266071 < Att12 < 20.586850
22.220988 < Att14 < 23.180185
Consequence 35.426088 < Att2 < 42.169776

113.825926 < Att7 < 122.261793
32.375620 < Att13 < 33.596051

57.988435 < Att3 < 69.574945
309.987803 < Att4 < 314.218245
55.294719 < Att5 < 66.896106
136.234441 < Att8 < 138.744999
40.927433 < Att11 < 41.562953
20.266071 < Att12 < 20.586850
22.220988 < Att14 < 23.180185
Consequence 35.426088 < Att2 < 42.169776

113.825926 < Att7 < 122.261793
32.375620 < Att13 < 33.596051
··· ··· ···

··· ··· ···
56.221481 < Att3 < 65.667791
139.024098 < Att4 < 289.982951
94.156397 < Att7 < 136.200000
57.669974 < Att10 < 87.300000
18.798957 < Att15 < 19.060978
Consequent 1.054478 < Att1 < 1.108900
31.100000 < Att6 < 40.883823
Note: Att1 = Case Number, Att2 = Percent body fat using Brozek’s equation,
457/Density − 414.2, Att3 = Percent body fat using Siri’s equation, 495/Density −
450, Att4 = Density (gm/cm3 ), Att5 = Age (years), Att6 = Weight (lbs), Att7 = Height
(inches)(target), Att8 = Adiposity index = W eight/Height2 (kg/m2 ), Att9 = Fat Free
Weight = (1 − f raction of body f at) × W eight, using Brozek’s formula (lbs),
Att10 = Neck circumference (cm), Att11 = Chest circumference (cm), Att12 = Abdomen
circumference (cm) “at the umbilicus and level with the iliac crest”, Att13 = Hip
circumference (cm), Att14 = Thigh circumference (cm), Att15 = Knee circumference
(cm), Att16 = Ankle circumference (cm), Att17 = Extended biceps circumference (cm),
Att18 = Forearm circumference (cm), Att19 = Wrist circumference (cm) “distal to the
styloid processes”.
Table 7. ACN of bolt dataset
none of them (ACN) < upper bound (UB)
62.782669 < Att7 < 65.562550
Consequence 23.688468 < Att1 < 31.295955
5.928943 < Att2 < 6.000000
62.782669 < Att7 < 65.562550
Consequence 23.688468 < Att1 < 31.295955
5.928943 < Att2 < 6.000000
··· ··· ···
··· ··· ···
1.761097 < Att4 < 2.325029
Consequent None
Note: Att1 = RUN, Att2 = SPEED1, Att3 = TOTAL, Att4 = SPEED2,
Att5 = NUMBER2, Att6 = SENS, Att7 = TIME, Att8 = T20BOLT (target).
column side, on the row side shows the five datasets which are quake, basket
ball, body fat, bolt and pollution. According to the table the global result of
PARCD method is better than MOPAR method except for the support value of
Quake dataset.
The highest result of support value by PARCD method is at 250.84% (Bolt
dataset) which is double value than the opposite method in the same dataset. On
the other hand, the lowest one is 22.97% by PARCD method which is almost less
twice from the MOPAR method also in the same dataset. Next, the remaining
datasets; Basket ball, Body fat and pollution, show that the PARCD has better
value than MOPAR method in which the gaps are almost 30%, 60% and 10%
respectively.
Then, according to the Table 9, the confidence result there is the additional
value of standard deviation because it shows the stability of confidence value.
Based on the the result, the confidence value generally shows that the PARCD is
better than MOPAR method. It can be seen from the whole datasets from Quake
to pollution data set in which the gaps are by 4%, 0.02%, 40%, 8% and 10%
respectively. Interestingly, the confidence and standard deviation values from
basket ball dataset have the similar values as of both of methods.
The output of comprehensibility and interestingness function can be seen
in Table 10. The comprehensibility function results that the PARCD is better
than MOPAR method, because there are three datasets which the values are
higher. They are basket ball, body fat and pollution which the gaps are almost
125%, 130% and 40% respectively. The comprehensibility means that the results
from the method are easy to understand. So, the percentage of the result can
be explained that both of the methods are better than the traditional method
Table 8. ACN of pollution dataset
Rule number (particle) Antecedent, consequent, Lower bound (LB) < attribute <
non of them (ACN) upper bound (UB)
9.675301 < Att6 < 10.303791
24.171326 < Att9 < 27.345700
42.882070 < Att10 < 44.054696
Consequence 21.695266 < Att1 < 22.757671
77.760994 < Att3 < 80.221960
6.698662 < Att4 < 7.071898
7436.549761 < Att8 < 7801.004046
58.816363 < Att15 < 63.240005
9.675301 < Att6 < 10.303791
24.171326 < Att9 < 27.345700
42.882070 < Att10 < 44.054696
Consequence 21.695266 < Att1 < 22.757671
77.760994 < Att3 < 80.221960
6.698662 < Att4 < 7.071898
7436.549761 < Att8 < 7801.004046
58.816363 < Att15 < 63.240005
··· ··· ···
··· ··· ···
8.721294 < Att4 < 9.206407
89.212389 < Att7 < 90.700000
21.796671 < Att11 < 23.231486
606.938956 < Att12 < 648.000000
67.768113 < Att15 < 73.000000
Consequent 2.956662 < Att5 < 3.005372

9.450171 < Att6 < 10.068287
9345.537477 < Att8 < 9699.000000
225.061313 < Att13 < 288.274133
242.720468 < Att14 < 250.733264
Note: Att1 = PREC Average annual precipitation in inches, Att2 = JANT Average Jan-
uary temperature in degrees F, Att3 = JULT Average July temperature in degrees F,
Att4 = OVR65 SMSA population aged 65 or older, Att5 = POPN Average household size,
Att6 = EDUC Median school years completed by those over 22, Att7 = HOUS of hous-
ing units which are sound & with all facilities, Att8 = DENS Population per sq. mile
in urbanized areas, 1960, Att9 = NONW non-white population in urbanized areas, 1960,
Att10 = WWDRK employed in white collar occupations, Att11 = POOR poor of families
with income < $3000, Att12 = HC Relative hydrocarbon pollution potential, Att13 = NOX
Same as nitric oxides, Att14 = SO@ Same as Sulphur dioxide, Att15 = HUMID Annual
average, relative humidity at 1pm, Att16 = MORT Total age-adjusted mortality rate per
100,000.
Table 9. Output of support and confidence function
Dataset Support (%) Confidence (%)

PARCD MOPAR PARCD MOPAR
Quake 22.97 46.26 86.73 ± 25.88 82.31 ± 28.91
Basket ball 61.04 32.13 92.69 ± 17.87 92.67 ± 16.65
Body fat 73.94 10.1 81.26 ± 30.67 43.59 ± 61.15
Bolt 250.84 107.29 96.88 ± 9.49 88.91 ± 9.49
Pollution 60.45 52.14 34.96 ± 43.91 23.02 ± 40.04
Table 10. Output of comprehensibility and interestingness function
Dataset Comprehensibility (%) Interestingness (%)

PARCD MOPAR PARCD MOPAR
Quake 785.2 ± 37.72 786.14 ± 419.67 2.34 ± 9.30 4.67 ± 11.40
Basket ball 545.80 ± 167.74 424.65 ± 192.63 6.56 ± 21.16 4.99 ± 5.18
Body fat 333.49 ± 218.95 204.87 ± 235.46 10.61 ± 21.03 21.71 ± 9.30
Bolt 231.08 ± 168.35 271.25 ± 168.35 43.43 ± 39.68 23.70 ± 39.68
Pollution 110.63 ± 165.76 65.82 ± 130.49 9.51 ± 18.61 10.23 ± 27.88
(discretization approach). Conversely, the Interestingness result of the MOPAR

is better than PARCD method which the datasets higher are quake, body fat
and pollution. The gaps are about 2%, 10% and 1% respectively. Then the other
data set shows that PARCD is better, using basket ball and bolt data sets by
the gaps around 2% and 20% respectively.
5 Conclusions
The rules of PARCD from five data sets show the good result but there are
two data sets which do not obtain the consequent, they are quake and basket
ball dataset. This output is interesting to be studied continually because actu-
ally the optimal solution has the complete rule. In addition, by referring to the
experiment, the output of multi objective functions which are support, confi-
dence, comprehensibility, Interestingness, it is discovered generally that PARCD
gives a better result than MOPAR method. For the future, since the problem of
numerical association rule mining is still open to be improved, it will be nice to
follow the research for instance by combining with other methods such as genetic
algorithm or fuzzy algorithm.
Acknowledgements. This research supported by various parties. We would like

to thank for scholarship program from Kanazawa University, Japan and Ministry
of Research and Technology and Directorate of Higher Education (KEMENRIS-
TEKDIKTI) and also STMIK AMIKOM Purwokerto, Indonesia. In addition, we thank
for anonymous reviewers who gave input and correction for improving this research.
References
1. Alatas B, Akin E (2008) Rough particle swarm optimization and its applications
in data mining. Soft Comput 12(12):1205–1218
2. Alatas B, Akin E, Karci A (2008) Modenar: multi-objective differential evolution
algorithm for mining numeric association rules. Appl Soft Comput 8(1):646–656
3. Alhajj R, Kaya M (2008) Multi-objective genetic algorithms based automated clus-
tering for fuzzy association rules mining. J Intell Inf Syst 31(3):243–264
4. Álvarez VP, Vázquez JM (2012) An evolutionary algorithm to discover quantitative
association rules from huge databases without the need for an a priori discretiza-
tion. Expert Syst Appl 39(1):585–593
5. Arotaritei D, Negoita MG (2003) An optimization of data mining algorithms used
in fuzzy association rules. In: International Conference on Knowledge-Based and
Intelligent Information and Engineering Systems. Springer, Heidelberg, pp 980–985
6. Beiranvand V, Mobasher-Kashani M, Bakar AA (2014) Multi-objective pso algo-
rithm for mining numerical association rules without a priori discretization. Expert
Syst Appl 41(9):4259–4273
7. Gen M, Lin L, Howada, (2015) Hybrid evolutionary algorithms and data mining:
case studies of clustering. Proc Soc Plant Eng 2015:184–196
8. Ghosh A, Nath B (2004) Multi-objective rule mining using genetic algorithms. Inf
Sci 163(1–3):123–133
9. Indira K, Kanmani S (2015) Association rule mining through adaptive parameter
control in particle swarm optimization. Comput Stat 30(1):251–277
10. Kaya M (2006) Multi-objective genetic algorithm based approaches for mining
optimized fuzzy association rules. Soft Comput 10(7):578–586
11. Kennedy J, Eberhart R (1995) Particle swarm optimization. In: Proceedings of the
IEEE International Conference on Neural Networks, vol 4, pp 1942–1948
12. Li C, Liu Y et al (2007) A fast particle swarm optimization algorithm with cauchy
mutation and natural selection strategy. In: International Symposium on Intelli-
gence Computation and Applications. Springer, Heidelberg, pp 334–343
13. Minaei-Bidgoli B, Barmaki R, Nasiri M (2013) Mining numerical association rules
via multi-objective genetic algorithms. Inf Sci 233:15–24
14. Narita M, Haraguchi M, Okubo Y (2002) Data abstractions for numerical
attributes in data mining. Lecture Notes in Computer Science, vol 2412, pp 35–42
15. Qodmanan HR, Nasiri M, Minaei-Bidgoli B (2011) Multi objective association rule
mining with genetic algorithm without specifying minimum support and minimum
confidence. Expert Syst Appl 38(1):288–298
16. Sangsawang C, Sethanan K et al (2015) Metaheuristics optimization approaches
for two-stage reentrant flexible flow shop with blocking constraint. Expert Syst
Appl 42(5):2395–2410
17. Sarath K, Ravi V (2013) Association rule mining using binary particle swarm
optimization. Eng Appl Artif Intell 26(8):1832–1840
18. Tahyudin I, Nambo H (2017) The combination of evolutionary algorithm method

for numerical association rule mining optimization. In: Proceedings of the Tenth
International Conference on Management Science and Engineering Management.
Springer, Heidelberg, pp 13–23
19. Witten IH, Frank E (2005) Data Mining: Practical machine learning tools and
techniques. Morgan Kaufmann, San Francisco
20. Yan X, Zhang C, Zhang S (2009) Genetic algorithm-based strategy for identifying
association rules without specifying actual minimum support. Expert Syst Appl
36(2):3066–3076
21. Yu M, Zhang Y et al (2015) Integration of process planning and scheduling using
a hybrid ga/pso algorithm. Int J Adv Manuf Technol 78(1–4):583–592
An Analytical Framework for the Conflict
Coordination Mechanism of Knowledge Network
Qifeng Wei(B)
Business School, Chengdu University of Technology,

weiqifeng15@cdut.edu.cn
Abstract. Whether the conflict coordination mechanism works effec-

tively directly affects the achievement of the goal of the knowledge net-
work, and is related to both the existence of the knowledge subject itself
and the knowledge network. Based on principal-agent theory and game
theory, a theoretical framework of conflict coordination mechanism for
knowledge network is constructed in this paper, which divides the con-
flict coordination mechanism into three levels: contract mechanism, self-
implementation mechanism and third-party conflict coordination mech-
anism, then this paper makes a mathematical analysis on the theoretical
framework. Results indicate that the adaptive behavior of the knowledge
subjects within knowledge network evolves along with the conflict coor-
dination mechanism framework; the conflict coordination of knowledge
network contract is mainly embodied in the two aspects of benefit dis-
tribution and knowledge division; the implementation of conflict coor-
dination relies mainly on relational contract mechanism and the trust
mechanism; when the first two fail, the third party helps to relieve the
relationship and make value judgments with a neutral position, in order
to solve the fierce conflict.
Keywords: Knowledge network · Conflict · Coordination mechanism ·

Behaviors
1 Introduction
The conflicts of knowledge network refer to the state of disharmony caused by
the accumulation of contradictions to a certain extent, among two or more than
two knowledge subjects, due to their incompatible behaviors or goals. According
to Pondy [19], the American behavioral scientist, the generation and develop-
ment of conflict could be divided into five stages: latent conflict, perceived con-
flict, felt conflict, manifest conflict and conflict aftermath, so as to construct
the embryonic of organizational conflict theory. There are many reasons for
the conflict of knowledge network, that’s because there are many differences
between different knowledge subjects in organizational goals, management modes
and organizational cultures, which will cause disagreement and lead to conflict;
An Analytical Framework for the Conflict Coordination Mechanism 167
The nature and scope of interdependence among knowledge subjects, are often
in a dynamic reset conditions, which enables conflict inevitable; And because
of the complexity of the knowledge cooperation among organizations, the dis-
location of objectives will also lead to the formation of conflict; In addition,
due to a changing external environment of knowledge network, inconsistencies
between subjects are long-standing. Moreover, the knowledge level, knowledge
structure [20], organizational routines [1], differences of values [10,11], among
the knowledge subjects, also lead to various conflicts inevitably.
Whether knowledge subjects within knowledge network could learn and
improve the knowledge and skills of conflict management, and accurately or effec-
tively implement conflict management, to control the conflict without harm, is
directly affecting the goal, and also it is related to subject itself and the existence
of knowledge network. Therefore, in the evolution process of conflict coordination
mechanism, we should objectively face various conflicts, and adopt correspond-
ing managerial strategies to conflicts with different natures. Obviously, in the
process of the generation, development and operation of the conflict mechanism
of knowledge network, as the carrier of management main body and objects,
knowledge subjects not only have the characteristics of cooperate with the whole
team to acquire a better implementation of conflict management, but also have
the initiative to actively adjust and innovate the managerial behaviors. From
the perspective of long-term development practices, the behavior routines and
routine variation of knowledge subjects in the conflict management, is character-
ized by adaptation, which is specifically the behaviors adaptation under conflicts
management mechanism of knowledge network.
2 Conflict Motivation and Behavioral Framework of

Knowledge Subjects
The knowledge network is composed of many knowledge subjects, and the benefit
distribution scheme among the subjects has both individual rationality and col-
lective rationality. For any subject, cooperation is better than non-cooperation,
and the higher benefit can be expected. Therefore, the cooperation among sub-
jects is a cooperative game process. Take the benefit distribution conflict as an
example, Gu and Li [7] proposed the n-person cooperative game model among
the members of the knowledge chain. The knowledge network consists of sev-
eral knowledge chains, and the cooperation among the knowledge subjects in
the knowledge network inherits the inter-organization cooperative logic of the
knowledge chain, so the model is also applicable here to explain. Quan [12] argued
that the formation of conflict is mainly caused by three factors: intra-member,
inter-member and environment. A three-dimensional conflict model of “benefit-
structure-knowledge” is an effective means to explain the conflict of knowledge
network (Fig. 1).
The rational organization has the expectation of greatest interests for itself,
and conflicts are generated when the benefits of each other can’t be coordinated.
Based on the structuration theory, the “structure” is considered to be those
168 Q. Wei
Y-axis: Z-axis:
knowledge structure
strong explicit
conducive to
communication
X-axis:
uncoordination coordination benefit
not
conducive to weak explicit
communication
Fig. 1. “Benefit- structure - knowledge” Conflict Motivation Model
resources and institutions owned by the knowledge main body, which involves
the organizational structure, organizational culture, social capital, human capi-
tal, capital, core competitive advantage and so on. The overlapping of the sub-
jects is an important source of conflicts. Knowledge, as the basis of the opera-
tion and survival of knowledge network, determines the essence of value pursued
by the knowledge subject. The development of the knowledge subject through
knowledge network is to promote the knowledge creation, so as to form the knowl-
edge advantage. Obviously, knowledge plays the central role in the conflict issue.
Therefore, the conflict coordination of knowledge network also focuses on these
three dimensions to operate management intervention. Wu et al. [22] argued that
the problem of cooperation and conflict between knowledge subjects is actually
a problem of creating and distributing income, including opportunism behav-
ior control and benefit coordination, value creation mechanisms. The former
includes a contract mechanism, a self-implementation mechanism and a third-
party conflict coordination mechanism, while the latter includes a relationship
adjustment mechanism and a knowledge collaboration mechanism [8]. And the
framework of knowledge subject adaptive behaviors is also embedded in these
mechanisms. The contract mechanism and the self-implement mechanism can
effectively distribute the benefits of the knowledge network to avoid the oppor-
tunistic behavior of the knowledge subject, which is based on the coordination
and interaction among the knowledge subjects. When the interaction can’t solve
the conflicts, then the third-party coordination mechanism is introduced to fur-
ther resolve. In the mechanism of value creation, the regulation of relationship
helps to improve the efficiency of cooperation among the knowledge subjects
and increase the value of innovation. The mechanism of knowledge collabora-
tion enhances the collaborative benefits of the knowledge subjects and further
promotes the creation of knowledge to form the knowledge advantage.
3 The Contract Mechanism of Conflict Coordination

Within Knowledge Network
According to the modern contract theory, the cooperative contract is designed
to avoid opportunistic behavior and reduce transaction costs. In addition to
this function, because of the characteristics of knowledge network cooperation
model, the function of knowledge network contract is to promote the knowl-
edge sharing among so as to realize the optimization of the cooperation value,
which is significantly different from other types of cross-organizational consor-
tia. For the contract design of knowledge network, the distribution of benefit
and the division of knowledge are its core contents, and the conflict coordina-
tion is the important objective of those contents, that is, the rules of benefit
distribution and knowledge division should effectively coordinate the possible
conflicts in the process of network operation, if not, the existing benefit distrib-
ution and knowledge division rule is unscientific and unreasonable. According to
the evolutionary theory, the knowledge subject has bounded rationality. Under
the constraint of objective conditions, the knowledge subject tends to choose sat-
isfaction strategy rather than the optimal strategy, which is also shown in the
contract. During the operative process of knowledge network, the information
asymmetry exists in the cooperation among the members, which contributes to
the possibility that the information superiority side obtains the extra benefits
through opportunism behaviors, thus aggravate the conflict of the knowledge
network. Therefore, the design of knowledge network contract is also aimed to
avoid this “rip off” problem. For the traditional contract overemphasizes part-
ner opportunism prevention and behavior motivation, the contract of knowledge
network emphasizes both cognitive and organizational coordination functions,
which have a direct impact on the efficiency of knowledge network cooperation
and conflict coordination. The function of knowledge-based inter-organizational
cooperation contract is embodied in three aspects: motivation, coordination and
cognition. The conflict coordination of knowledge network contract is mainly
embodied in the two aspects of benefit distribution and knowledge division.
3.1 The Contract Mechanism of Conflict Coordination of Benefit

Distribution
The design of the benefit distribution coefficient is the core issue of the conflict
coordination of benefits, and the design of the contract also revolves around the
core. As Wu and Gu et al. [7] stated, there are different approaches to con-
tract optimization when considering knowledge spillovers. While the conflict of
knowledge spillover is not considered, the input level of the knowledge subject
is independent of the input level of the partner. In this type of contract design,
the higher contribution to the overall network, the higher profit the knowledge
subject should gain, which makes the incentive efficiency achieving the best, and
effectively coordinate the conflict, so as to ensure the stable operation of knowl-
edge network. Under the condition of knowledge spillover conflict, the knowl-
edge spillover value is the important reason for the knowledge subject to choose
170 Q. Wei
whether to cooperate and adjust the level of knowledge input, yet the knowledge
spillover loss is also the reason of the conflict formation. At this point, the input
level of the knowledge subject depends on the input level of the partner, showing
a positive correlation.
The study [23] revealed that, the effect of knowledge spillovers on the coop-
eration among knowledge subjects is a “double-edged sword”, which effectively
promotes cooperation but also brings conflict. On the one hand, the innovation
value brought by knowledge spillovers is the driving force for deep cooperation.
The stronger the knowledge absorptive capacity of knowledge subjects, the more
benefits they could get from knowledge spillovers of partners, and the higher the
level of knowledge input. On the other hand, because of the influence of knowl-
edge spillover, the input level of knowledge subject is affected by the level of
partner input, and the drop of knowledge input will affect the benefit from
knowledge spillover. In addition, knowledge spillover may lead to the leakage of
core knowledge assets, which will seriously affect the stability of knowledge net-
work cooperation, because the huge potential value of knowledge spillover will
stimulate the opportunistic behavior of the cooperators, resulting in conflict.
By enhancing the level of cooperation transparency and enhancing the level of
mutual trust among organizations, the observability of knowledge investment in
the cooperation can be improved, which will effectively restrain opportunistic
behaviors and enhance the interaction.
3.2 The Contract Mechanism of Conflict Coordination of

Knowledge Division
What the knowledge division in knowledge network involves, is R&D of knowl-
edge subject belongs to the issue of self-completion or principal-agent prob-
lem. That is, knowledge subject can do scientific research work by themselves,
and make the technology development work outsourced to other subjects in
the knowledge network; or both by other agency to complete. The operation
mechanism of contractual mechanism of knowledge division conflict coordina-
tion includes two stages: knowledge division decision-making and multi-task
principal-agent [8].
In the decision-making of knowledge division, the knowledge subject can
choose independent R&D or principal-agent, and different decision-making
behavior corresponds to different forms of contractual payment. Under the con-
dition of principal-agent, in addition to receiving the entrusted payment, the
entrusted entity can obtain the additional income from the technical capability
growth. On the one hand, the higher the degree of R&D investment, the higher
the R&D capacity gains. On the other hand, the fixed price paid by the entrusted
agent represents the bargaining power in the contract conclusion, which reflects
the technical capacity of the entrusted subject. From this point of view, it’s clear
that the higher the technical base is, the higher the additional benefits of growth
will be. And because the principal does not participate in R&D, so the income
is just the established distribution after the innovation success, but no addi-
tional revenue from the increase of technical capacity. According to the model
of knowledge division decision-making constructed by Gu et al., the higher the

coefficient of technical ability growth income, the higher the input level of the
entrusted subject; the higher the payment price in contract, the higher the input
level of entrusted subject. And finally the income level of principal is constructed
as follows:
ω 1−ω
(1 + k1 )V 2 ω 1−ω
R1F = . (1)
4 c21 c22
In formula (1), k1 is the principal coefficient of technological capability

growth; V is the value of the product after the success of innovation; ω is
the importance level for the scientific research stage; 1 − ω is the level of
importance for the technology development stage; the cost of the scientific
research and technological development of the entrusted subject is respectively
c21 e2r c22 e2d
2 and 2 . and er mean ed the level of knowledge input in two phases,
0 ≤ er ≤ 1,0 ≤ ed ≤ 1,c21 > 0, c22 > 0.
As the input level of the entrusted subject is positively related to its income,
it can be seen that the higher the coefficient of technical capability growth,
the more benefits the entrusted entity will gain. Therefore, in the decision of
entrustment price, the principal will consider the technical capacity growth in
the process of cooperation.
If the knowledge division uses partial delegation, that is, the knowledge sub-
ject only makes the technological development commissioned by the agent. The
success of R&D depends on the degree of knowledge input and the efficiency of
knowledge transfer from the stage of scientific research to technological devel-
opment. Based on the model, the payment level of the principal is positively
related to the importance of the development activities, that is, the more impor-
tant R&D activities are, the higher the payment price will be. In addition, the
higher the technical efficiency growth factor and knowledge transfer efficiency,
the higher the input level of the principal. The principal’s income level is as
follows:
ω
2 1−ω
2ω (1 + k2 ) 2 ω (1 − ω)
R1 = μ
P
V . (2)
2 η 2 c21 2c22
In formula (2), k2 is the coefficient of technological capability growth of

subject; V is the value of the product after the success of innovation; ω is the
importance level for the scientific research stage; 1 − ω is the level of importance
for the technology development stage; μ is the knowledge transfer efficiency,
0 < μ < 1; η is the constant associated with R&D costs, 0 < η < 1.
The choice on the form of knowledge division depends on the level of income.
By comparing formulas (1) and (2), it can be seen that whether the subject
of knowledge conducts the principal-agent of scientific research depends on the
increase of technological capacity in R&D activities [8], the efficiency of knowl-
edge transfer and the scientific research the importance relative to technological
development.
172 Q. Wei
When the knowledge subject entrust scientific research and technological

development activities to other knowledge subjects, the conflict coordination
becomes a multi-task principal-agent problem [9]. Based on the model of knowl-
edge division, the model [8] considers the cost independence of R&D and the risk
aversion problem of agent. In this case, the maximized deterministic equivalent
income of the principal is confronted with two constraints: one is the participa-
tion constraint, the expected utility from the accepting contract is greater than
or equal to the maximum expected utility when the contract is not accepted;
another one is incentive compatibility constraint, when the principal is unable
to follow the behaviors of the entrusted subject and the state of development,
the entrusted subject always chooses a vector that maximizes its expected util-
ity function. It can be seen that the utility behavior of any entrusted subject
can only be realized through the maximization of the utility of the entrusted
subject. Gu argued that the conflict coordination among knowledge subjects is
affected by factors such as the profitability of entrusted subjects, the relevance
of scientific research and technological development tasks, and the traceability
of R&D activities.
In general, when the ability that the principal transferring the R&D results
of the entrusted subject to the profit is stronger, then the degree of incentive
on the entrusted subject to implement innovative behaviors is greater, which
is derived from the larger profits will have more possibility for payment. When
the cost of R&D is complementary, the entrusted subject will give higher moti-
vation to the R&D task; and when the costs of R&D are substitutes for each
other, if the degree of substitution is high, the degree of incentive on a single
R&D task will not be too high. Similarly, if a higher incentive is given to tech-
nology development, the entrusted subject may neglect the input of scientific
research activities, thus affecting the basis of product innovation; When the cost
of R&D are completely independent, the incentive is related to the entrusted
profit transformation ability, the marginal cost change rate, the degree of risk
aversion, and the marginal cost variance and stochastic variance of another task.
4 The Self-implementation Mechanism of Conflict

Coordination Within Knowledge Network
The self-implementation of conflict coordination within knowledge network is a

kind of internal conflict resolution mechanism as opposed to third-party imple-
mentation. As the knowledge subject with the dominant information in the
knowledge network performing “hitchhikes”, “rip off” and some other acts, a
series of conflicts among the organizations are triggered, although the contract
effectively provides for the distribution details of innovation benefits, and to
a certain extent, solve the conflict of knowledge division. However, due to the
incompleteness of the contract, the prior arrangement of rights and the distribu-
tion of benefits cannot cover the content of “residual control” [25], which lead to
the existence of controversy for the contract. The involvement of public relations
departments, arbitration institutions, courts and other third-party conflict medi-

ator is an optional way, but compared to the main subject of the conflict, the
third-party management lacks information superiority, which leads to high man-
agement costs and diseconomy of the contractual mechanism. Therefore, it is the
rational expectation of the knowledge subject to seek the maintenance and con-
flict coordination in the knowledge network. Dyer and Singh [4] have proposed
to use self-imposed conventions to restrain opportunistic behavior. Focusing on
long-term cooperative relationships in knowledge network, relational contracts
among knowledge subjects can effectively reconcile possible conflicts. The rela-
tionship contract helps to make up for the short-sighted and incomplete defects
of the formal contract, for instance, the conclusion of the contract can only
be observed after the completion of the cooperation results, which ensures the
smooth progress of knowledge cooperation. In addition, the trust among organi-
zations is also a “dissolving agent” of the conflict. The self-implementation of the
trust mechanism enhances the transparency of the knowledge subjects’ cooper-
ation, enhances the special property investment and guarantees the openness of
the communication among organizations, which is conducive to the coordination
and settlement for conflict in the knowledge network.
4.1 The Relational Contract Mechanism of Cooperation and

Conflict Coordination
The study of relational contract begins with the relational contract theory pro-
posed by Macneil [15]. The theory holds that social transactions are embedded
in complex relationships, thus, to study market transactions, we must make
clear the various factors in social relations first, such as the specific personal
relationships between contractors, the duration of the contract, the number of
participants, the degree of difficulty in performing the contract, and so on. Baker
and Gibbons et al. [3] argued that relational contracts are closely related to the
relationship of ownership of assets, which is based on the value of future rela-
tions, it is the future relationship value based on the informal relationship. From
this perspective, the purpose of the relational contract is to make the conflict
subjects clear the loss of the future relationship and the loss of the direct bene-
fit, so to change their behaviors, to promote the relaxation of the relationship to
maintain the sustainable cooperation. From the perspective of contract theory,
relational theory is based on the formation of asset specificity [21], whose main
characteristic is “self-executing mechanism” [2]. Liu [14] even believed that the
relational contract is the micro-foundation of formation and operation of the
enterprise network, and the formation of the relational assets is derived from
the common role of specific assets and individual social capital. Gu and Wu
et al. [8] introduced the concept of relational contract to define the rent, the
core enterprise and knowledge specific degree of agent organization of the rela-
tionship among the knowledge chain organizations, and analyzed the influences
from discount factor and knowledge specific degree on the self-implementation of
relational contract. The mechanism by which a relational contract can maintain
cooperation among the contracting parties is that any breach by the other party
174 Q. Wei
will undermine long-term cooperative relations and thereby penalize the loss of
future earnings. In order to maintain the self-implementation of the contract
mechanism in the knowledge network and avoid the intervention of the third-
party managers to increase the management cost, the rational knowledge subject
in cooperation will adjust its behaviors from the factors closely related to the
self-implementation, including relationship rent, discount factor and knowledge-
specific level.
Usually, the principal and the entrusted subject cooperate to create the rela-
tionship. The more the rent is, the more the entrusted subject puts in the specific
knowledge assets, the stronger the self-implementation of the relational contract
is. The higher the discount factor set by the principal, the higher the input
of the entrusted subject, and the stronger the self-implementation of the con-
tract. Knowledge-specific degree refers to the relationship rent created by the
unit-specific knowledge asset invested by the knowledge subject. The higher the
degree of knowledge-specificity of the principal, the higher the degree of knowl-
edge assets locked in the specific cooperative relationship with the entrusted
subject, the replacement cost will be higher, so the self-implementation of rela-
tional contract is stronger; and when the principal commissioned into higher
special assets, the principal must have a very high degree of knowledge in order
to ensure the self-implementation of the relational contract, otherwise the sub-
ject may be subject to the implementation of opportunistic behavior to damage
the benefits of the principal, thus the principal will spend more energy in the
observation of the cooperative behavior of the entrusted subject.
4.2 The Trust Mechanism of Cooperation and Conflict Coordination

Network trust maintains high cooperation intensity among knowledge subjects
and enhances the efficiency of knowledge transfer [18]. Trust also plays a very
important role in cooperation and conflict coordination, that’s because trust
is the basis for establishing and maintaining intimate, cooperative, and efficient
relationships and the devastation of trust due to deception and breaches can have
devastating consequences for relations [24]. Trust is also a self-implementation
mechanism for cooperation and conflict, based on informal rules of society, using
mutual trust among organizations to make people willing to cooperate to elim-
inate conflicts [16,17]. The higher the degree of trust, the more significant the
conflict resolution will be, on the contrary, the contradictions and conflicts will
gradually increase. The role of trust in cooperation and conflict coordination
among knowledge subjects in knowledge networks mainly lies in the following
aspects: First, trust enhances the transparency of knowledge flow. Transparency
between knowledge subjects is not high, mainly because of the vigilance on
opportunistic behavior between each other, which usually lead to increased dif-
ficulty of knowledge sharing. Mutual trust can help reduce the fear of oppor-
tunistic behavior, and enhance the transparency of cooperation between orga-
nizations, thereby reducing the complexity of knowledge flow and uncertainty.
Secondly, trust increases the investment in the specific assets of knowledge sub-
jects. Specific asset investment refers to the existence of specific assets only for
specific users, the knowledge subject once invest the specific asset, the bargaining
power in cooperation will be reduced, which may face “rip off” behaviors. There-
fore, under the condition of lack of confidence, the subject will generally choose
universal asset investment rather than specific asset investment, but specific
asset investment can create relational rent and help to improve performance [4].
In addition, trust helps to ensure the openness of communication among organi-
zations. In-depth communication between organizations means that knowledge
subjects have more comprehensive understanding the strategic vision, business
objectives, resource allocation, financial status and other information, but deep
communication between organizations must be based on a high degree of mutual
trust as the prerequisite.
Trust mechanism for the cooperation between knowledge subjects and coor-
dination of conflict, its essence is the trust game among organizations [8]. Under
completely unconstrained conditions, suppose that there are two knowledge sub-
jects, A and B, are cooperating. When they trust each other, the knowledge is
fully shared. At this time, the two parties get the value of R1 and R2 ; when
A trusts in B, while B is for the individual interests, bad faith and takes the
opportunity to implement opportunistic behavior, so that the interests of A is
damaged, ΔR > 0, then the two sides get the benefits R1 − ΔR and R2 + ΔR; If
two sides do not trust each other, resulting in no knowledge sharing or knowledge
sharing is in low efficiency, assuming that compared to mutual trust, the loss of
benefits this time are both 2ΔR, so two sides respectively get R1 − 2ΔR and
R2 − 2ΔR. In the unconstrained condition, the game matrix of trust behaviors
between knowledge subjects is shown in Table 1.
Table 1. Gaming matrix of trust behaviors of knowledge subjects without constraints
Knowledge subject B
Trust Distrust
Knowledge Subject A Trust R1 , R 2 R1 − ΔR, R2 + ΔR
Distrust R1 + ΔR, R2 − ΔR R1 − 2ΔR, R2 − 2ΔR
As can be seen from Table 1, if knowledge subject A trusts knowledge sub-

ject B at the outset, he must bear the risk of opportunistic behavior that B is
not trustworthy, and he cannot improve his condition as long as he does not
change his behavioral strategy. Therefore, from the beginning, A doesn’t trust
B; vice versa, making the two sides into the Nash equilibrium distrust - distrust
“prisoner’s dilemma”. The two sides of the knowledge subjects even starts from
the rational assumption of self-interest purposes, the result is still detrimental
to others, which is just the reason to strengthen communication to build mutual
trust (Table 2).
When there are constraints, suppose knowledge subject A and knowledge
subject B separately put specific assets r1 and r2 in cooperation. At the same
176 Q. Wei
Table 2. Gaming matrix of trust behaviors of knowledge subjects with constraints
Knowledge subject B
Trust Distrust
Knowledge Subject A Trust R1 , R2 R1 − ΔR, R2 + ΔR − r1 − p
Distrust R1 + ΔR − r1 − p, R2 − ΔR R1 − 2ΔR, R2 − 2ΔR
time, the cooperation is influenced by informal social relations, including rep-

utation, industry regulation and code of ethics. If someone abuses the trust to
implement opportunistic behaviors, he will be punished with the loss of value p,
therefore, in the presence of constraints, the payoff matrix of the trust game of
knowledge subject becomes:
Compared with the income matrix in unconstrained condition, the difference
exists in the losses from investment of specific assets and penalty, from a rational
point of view, if r1 + p > ΔR, knowledge subject B will not choose to implement
opportunistic behaviour, that is, when the extra income gained from opportunis-
tic behaviour is less than the sum total of informal social relations punishments
and specific assets input, knowledge subject B chooses to trust each other with
no suspense, so that the game proceeds to the Nash equilibrium of trust-trust.
Therefore, whether to knowledge subject A or B, in the consideration of the
evolution of trust mechanism, negotiating the level of specific asset investment
making clear social relations punishment are the main adaptive behaviours.
5 The Third-Party Conflict Management Mechanism

Within Knowledge Network
The implementation of the third-party management mechanism of conflict in

knowledge network is derived from the failure of the contract mechanism and
the self-implementation mechanism of cooperation and conflict coordination.
There is a serious disagreement among the knowledge subjects on the target and
interests; it’s unable to negotiate friendly. Only the third party independence
institutions could help to relieve the relationship and make value judgments, to
solve the fierce conflict in the knowledge network. The concept of “third party”
comes from public economics, which can convey information to the conflicting
parties in an impartial manner and attitude, and fairly adjudicates the conflict
[13], the party is generally acted by the organization designated by government
and industry associations, also, arbitration institutions and courts are often used
as potential conflict resolution organizations, but its service cost is much higher.
The role of third party conflict management organization in coordinating the
conflict among organizations is mainly reflected in [6,8]: to protect the bene-
fits of the weaker subjects, to avoid the unreasonable distribution of interests,
to appraise the performance of the knowledge subjects, to avoid the perfor-
mance evaluation conflict of interest arising from subjective differences, and to
investigate relevant conflict facts to determine the specific circumstances of the
statement, so as to consider the dispute resolution. Specific functions include

[13]: conflict evaluation, conflict level control, conflict resolution and conflict pre-
vention. According to Elangovan [5], the strategy of third-party conflict control
has five aspects, namely, means-control strategy (MCS), ends-control strategy
(ECS), low control strategy (LCS), full control strategy (FCS), and part control
strategy (PCS).
MCS mainly affect the formation process of the conflict solution, but not
directly form the solution result. The final conflict resolution strategy is decided
by the conflicting subject, so it is a strategy focusing on the process control,
whose real form of expression is “mediation”. At this time, the adaptive behav-
ior of knowledge subject tends to assist the management of the third party,
strengthen the communication, consciously maintain the order of negotiation,
discuss the solution with other conflict subjects under the witness of the third
party and finally realize the reconciliation. The results of ECS is just the oppo-
site, the strategy influences the outcome of the conflict resolution rather than
the process to intervene in the dispute, directly deciding the conflict resolution,
rather than trying to influence the negotiation process, the subject may posi-
tively and fully demonstrate the conflict content to a third party, the actual form
of such strategies include “arbitration”, “judgment” and so on. LCS advocates
using “negative” way to resolve the conflict, the third party organization only
requests the conflict subject to shelve the dispute or provides the impetus to
urge two sides to realize conflict self-reconciliation, rather than confront conflict
itself, therefore the third-party doesn’t have the substantive involvement in con-
flict coordination process and the result either, For the knowledge subject, the
adaptation behavior at this time is far away from conflict, shelving conflict, and
achieving common development. FCS advocates the comprehensive intervention
in the conflict coordination process, which resolves the conflict by influencing
the way of expounding the cause of the conflict, the way of presenting the facts,
and the final solution to resolve the conflict. In which the third party obtains
information by asking the conflict subject about the conflict, that is, according
to its own conflict resolution routines to actively search for and compile infor-
mation about the conflict, and announce a conflict resolution plan to achieve the
purpose, which as Elangovan stated is referred to as inquisitorial intervention or
autocratic intervention, so as the parties to conflict, the knowledge subject can-
not interfere in any aspect of conflict coordination, what they can only is just co-
ordination. The last strategy PCS is a hybrid management approach, the third
party implements the strategy of sharing the process of conflict-coordination
and control of the strategy outcome, which is presented as: the third-party join
the conflict negotiation to achieve consensus; and by facilitating the interaction
and communication of the conflict subjects and arguing with the conflict sub-
jects on the conflict event, to help think and evaluate the various options to
select the best plan, recommend to convince the conflict subject to accept the
preferred program to resolve conflicts, such as “arbitration after conciliation”,
“collective discussion” and so on. In addition to actively cooperating with the
third party to investigate and participate in collective discussion, the knowledge
178 Q. Wei
subject may try to find a solution to the conflict and obey the decision of the
third party to eliminate conflicts. The third-party conflict intervention strategy
chosen by knowledge network is mainly affected by fairness perception, satisfac-
tion perception, effective perception and strategy implementation efficiency of
knowledge subject.
6 Conclusions
The issue of conflict coordination for knowledge network, which is essential to

knowledge network management, is directly influencing its effectiveness and per-
formance. Due to this, the conflict coordination mechanism of knowledge network
is studied in this paper, and an analytical framework is proposed afterwards.
What’s presented in this paper is: the conflict coordination of knowledge network
is focusing on knowledge, benefits and structure. The contents of the cooperation
and conflict coordination mechanism among knowledge subjects include contract
mechanism, self-implementation mechanism and third-party conflict coordina-
tion mechanism. The framework of knowledge subject’s adaptive behaviors is
embedded in these mechanisms. The contract mechanism and the self-implement
mechanism can effectively distribute the benefits of the knowledge network, to
avoid the opportunistic behaviors of the knowledge subject, which is based on
the coordination and interaction among knowledge subjects. When the interac-
tion among the subjects cannot solve the conflict, the third-party coordination
mechanism is introduced to further resolve.
Acknowledgements. This work is supported by the National Natural Science Foun-

dation of China (71602012), Social Science Planning Project of Sichuan Province
(SC16C002), Soft Science Project of Sichuan Province (2016ZR0194), and Key
Research Base Project of Philosophy and Social Sciences of Sichuan Province (CJY16-
03 and Xq16C06).
References
1. Argote L, Guo JM (2016) Routines and transactive memory systems: creating,
coordinating, retaining, and transferring knowledge in organizations. Res Organ
Behav 36:65–84
2. Baker G, Murphy KJ (2002) Relational contracts and the theory of the firm. Q J
Econ 117(1):39–84
3. Baker G, Gibbons R, Murphy KJ (1999) Informal authority in organizations. J
Law Econ Organ 15(1):56–73
4. Dyer JH, Singh H (1998) The relational view: cooperative strategy and sources of
interorganizational competitive advantage. Acad Manage Rev 23(4):660–679
5. Elangovan AR (1995) Managerial third-party dispute intervention: a prescriptive
model of strategy selection. Acad Manage Rev 20(4):800–830
6. Fisher RJ (2016) Ronald J. Fisher: Advancing the understanding and effectiveness
of third party interventions in destructive intergroup conflict. Springer
7. Gu X, Li J (2005) An n-person cooperative game analysis on the allocation of

payoff among the members of knowledge chain. In: Proceedings of the 8th national
conference on youth management science and system science, pp 55–62
8. Gu X, Wu S (2011) Research on conflict and conflict management among knowledge
chain organizations. Sichuan University Press, Chengdu
9. Holmstrom B, Milgrom P (1991) Multi-task principal-agent analysis: Incentive
contract, asset ownership and job design. J Law Econ Organ 7:24–52
10. Ibert O, Müller FC (2015) Network dynamics in constellations of cultural differ-
ences: relational distance in innovation processes in legal services and biotechnol-
ogy. Res Policy 44(1):181–194
11. Kazadi K, Lievens A, Mahr D (2016) Stakeholder co-creation during the innovation
process: identifying capabilities for knowledge creation among multiple stakehold-
ers. J Bus Res 69(2):590–598
12. Li Q, Xin G (2008) The three-dimensional model of the causations of conflict
among organizations in knowledge chain. Sci Sci Manage S&T 29(12):92–96 (in
Chinese)
13. Liu D, Gao Y, Wu S (2010) Research on the inter-firm conflicts coordination in
knowledge chain based on the third party management. J Intell 29(8):65–68
14. Liu RJ (2006) Relational contract and business networks’ transition. China Ind
Econ 6:91–98
15. Macneil I (1974) The many futures of contracts. South Calif Law Rev 47:691–816
16. Mayer RC, Davis JH, Schoorman FD (1995) An integrative model of organizational
trust. Acad Manage Rev 20(3):709–734
17. Mcallister DJ (1995) Affect-and cognition-based trust as foundations for interper-
sonal cooperation in organizations. Acad Manage J 38(1):24–59
18. Miller K, Mcadam R et al (2016) Knowledge transfer in university quadruple helix
ecosystems: an absorptive capacity perspective. R&D Manage 46(2):383–399
19. Purdy LR (1967) Organizational conflict: concepts and models. Adm Sci Q
13(3):296–320
20. Ranucci RA, Souder D (2015) Facilitating tacit knowledge transfer: routine com-
patibility, trustworthiness, and integration in m & as. J Knowl Manage 19(2):257–
276
21. Williamson OE (1985) The economic institutions of capitalism: firms, markets,
relational contracting. Am Polit Sci Assoc 32(2):61–75
22. Wu S (2009) Research on the stability structuring of inter-firm cooperation and
conflicts in the knowledge chain. Nankai Bus Rev 12(3):54–58
23. Wu S, Gu X, Peng S (2009) Research on the knowledge spillover’s influence on the
stability of inter-organizational cooperation in knowledge chain. In: FSKD 2009,
Sixth international conference on fuzzy systems and knowledge discovery, vol 6, pp
14–16
24. Yao Q (2011) On the trust violation and repair in the field of organizational behav-
ioral. Nankai J 5:133–140 (in Chinese)
25. Zhang Z, Jia M, Wan D (2007) On the allocation of PPP optimal control rights
from the perspective of incomplete contracts and relational contracts. Foreign Econ
Manage 29(8):24–29 (in Chinese)
Flight Arrival Scheduling Optimization on Two
Runways Based on IGEP Algorithm
Rui Wang, Minglei Qu(B) , and Fuzheng Wang

939774881@qq.com
Abstract. With the accelerated pace of life, more and more tourists’
travel mode change from the traditional land travel into air travel. It
has contributed to the rapid development of the aviation industry. But
also produced some problems that troubled the airline. Among them, the
flight delays problem has not been effectively addressed. The cost of flight
delays is still high. This paper launches the research, analyze the cause of
the cost of flight delays. For the factor of terminal area flight scheduling
unreasonable to improve. Combined with the two-runway actual situa-
tion of Chengdu Shuangliu International Airport, minimize the cost of
flight arrival delays, construct the model of flight arrival on two runways.
At the same time, the coding method, selection strategy and fitness func-
tion of GEP are improved combined with the specific problem. Finally,
IGEP and simulation are utilized to solve the practical problem. Com-
pared with the traditional F CF S rules, the cost of flight arrival delays
is significantly reduced, the efficiency of flight arrival and runway uti-
lization is improved, and the interests of airlines are guaranteed. It also
shows the superiority of IGEP in addressing the issue of two-runway
flight arrival.
Keywords: Flight arrival · Improved gene expression programming ·

Cost of flight delays · Two runways
1 Introduction
With the speeding up of life rhythm, people on the quality of travel requirements
are getting higher and higher. Compared with ordinary land travel, people have
a tendency to more efficient and better services’ air travel. China’s civil aviation
passenger traffic for several years to maintain the growth rate of more than 8%
[3]. In order to satisfy the needs of more passengers, major domestic airlines
actively purchased aircraft, the number of aircraft China’s civil aviation has
been increasing for many years. At the same time, several domestic airports
have built new runways for use by airlines. Two-runway airport is becoming
more and more.
But the development of the aviation industry has also brought some problems
cannot be ignored. Air traffic flow increases, airport traffic congestion and flight
Flight Arrival Scheduling Optimization on Two Runways 181
delays are very common. According to statistics, the punctuality rate of flight
in China has been lower than 80% for several years, and has a tendency to
decline [4]. It is contrary to the pursuit of efficient travel. Flight delays led to
some people confidence shook on the air travel, and then give up air travel, the
airlines’ revenue will reduce. In general, flight delays reach two hours, this flight
will be a loss, the longer the delay, the higher the cost of delays. So how to
control flight delays and decrease the cost of flight delays has become an urgent
problem to be solved.
There are numerous reasons for flight delays, such as bad weather and aircraft
faults, these accidental factors are unable to change. However, in many cases,
unreasonable flight scheduling can also lead to flight delays. The data showed
that nearly 40% of flight delays are related to flight scheduling. Unreasonable
scheduling will aggravate flight delays, increase the cost of flight delays. On the
contrary, reasonable flight scheduling can alleviate the problem of flight delays,
reduce the cost of flight delays, protect the interests of airlines and improve
the utilization of the runway. So how to make reasonable flight scheduling has
become more and more attention by the airlines. Scholars in related fields have
studied and made some achievements after noticing the importance of flight
delays:
Hu and Xu [11] pointed out that flight delays and air traffic congestion due
to the restriction of airport capacity, they used the ground maintenance strat-
egy to solve the problem. Xu and Yao [19] established optimization model of
multi-runway in the terminal area, and the genetic algorithm (GA) was used
to solve the problem. Zhang et al. [20] constructed a dynamic multi-objective
optimization model for arriving and departure flights on multiple runways. Then
designed GA to solve the proposed model with the dynamic characteristics of
receding horizon control strategy. Kafle and Zou [12] proposed a novel analytical-
econometric approach, and built an analytical model to reveal the effects of var-
ious influencing factors on flight delays. Liang and Li [14] take flight departure
time as the goal, and used improved gene expression algorithm (IGEA) to study
the departure problem of single runway airport.
But the problem of two-runway flight arrival based on the cost of flight delays
has not been solved. The traditional mode of flight arrival is first come first served
(FCFS), without considering the impact on other flights, the advantage of this
approach is completely fair between the various flights, and the scheduling is sim-
ple. It is more suitable in the early stage of China’s aviation industry, the number
of flights is less. With the increase in the number of flights, the shortcomings are
also obvious leakage. Lack of scheduling lead to high cost of flight arrival delays,
low runway utilization, cannot guarantee the interests of airlines, also cannot
meet the requirements of travel time for passengers. So the traditional FCFS
cannot adapt to the current flight arrival scheduling. There is an urgent need
for new ways to improve the problem. Gene expression programming (GEP) is
a new adaptive evolutionary algorithm based on biological structure and func-
tion [2]. It is evolved from GA and genetic programming (GP). It learns GA
to encode chromosome into a fixed-length linear symbol string, and inherits the
182 R. Wang et al.
characteristics of GP to convert chromosomes into undefined expression trees.

The genotype and phenotype were separated. Therefore, the use of genetic oper-
ators in terms of number and variety is not constrained. Meanwhile, mutation,
insertion, and reorganization expands the search space of GEP. Using simple
coding to address complex problems, make GEP greatly enhance the ability in
solving specific problems. After the algorithm is proposed by Candida Ferreira, it
has aroused great repercussions and became the focus of scholars in some fields.
But the research of GEP is primarily focused on the prediction model and data
mining [1,6,10,13]. Research on aviation management is still very limited. The
application of GEP in flight arrival scheduling is more scarce. In view of GEP’s
powerful function, the problem of two-runway flight arrival scheduling based on
IGEP is proposed in this paper. By using IGEP to solve the problem of two-
runway flight arrival scheduling. The goal is to reduce the cost of flight delays,
the corresponding model and IGEP were designed to optimize the problem. To
ensure that the interests of airlines.
This paper is organized as follows. Section 1 introduce the reasons and influ-
ence of flight delays, discusses the study of flight scheduling, introduce the birth
process and advantages of GEP, and IGEP was proposed to optimize the prob-
lem of two-runway flight arrival. Section 2 introduce the development of ZUUU
and the specific problems of the flight arrival. Section 3 constructs the model
based on the practical problem. In Sect. 4, some improvements of GEP are pro-
posed, and describe the processes of IGEP. Section 5 application IGEP to solve
practical case, and analyze the results of the problem; Sect. 6 is the conclusion
and prospect.
2 Problem Description
Chengdu Shuangliu International Airport (ZUUU) is located in Sichuan

Province, southwest of China. It was founded in 1938. After that, experienced
several important expansion and development, and in November 1995 changed
its name to “Chengdu Shuangliu International Airport”. The second runway
of the ZUUU was successfully completed and put into operation in 2009. It is
located in the eastern part of the airport and connected to the first runway in
the west. The aircraft can land independently on two runways. ZUUU plane
diagram shown in Fig. 1:
Fig. 1. ZUUU plane diagram

There are two runways in Fig. 1. Runway 1 is connected to m taxiways.

Runway 2 is connected to n taxiways. For flight arrival, the runway serves as a
buffer for the speed landing aircraft. Decelerated airplane slides into the taxiway,
and the airplane enters the apron through the taxiway.
After the second runway put into use, the transportation capacity of ZUUU
has been significantly improved, but in some cases the cost of flight arrival delays
is still high. Among them, flight scheduling is the most important factor in the
cost of flight arrival delays, and flight scheduling optimization mainly for the
terminal area. The terminal area has two main functions. First, ensure orderly
flight arrival and departure. Second, ensure safe separation and sequence between
arrival flights. On the premise of flight safety, scheduling flight in the terminal
area can reduce the time and cost of flight delays, improve runway utilization
and ensure benefit of airlines.
Arrivaling is required after the flight reaches the terminal area. But at some
point, due to the effects of bad weather and airport capacity constraints, the
flight to reach the terminal area in a certain period of time will be restricted to
arrival, these flights were forced to circle in the terminal area waiting to arrival.
When the situation of the airport improved, these waiting flights were arranged
on the runway in turn to arrival. But which flights arrival on runway 1 and which
flights arrival on runway 2, as well as the sequence of flight arrival, related to the
cost of flight delays. In order to minimize the total cost of flight arrival delays,
these flights need to be sorted. So the problem can be transformed into the
sequencing optimization problem of flight arrival. We can establish the schedul-
ing model and design the appropriate algorithm to solve it. Flight arrival needs
to consider safe wake vortex separation of the flight. Wake vortex separation is
related to the aircraft type, and the cost of flight delays per unit time is not
same for different aircraft types [5,7,17].
About wake vortex separation, according to the standards of international
civil aviation (ICAO), all aircraft according to the maximum allowable takeoff
weight is divided into three types. Heavy (H), medium (M) and light (L), and
as a basis for the aircraft to set a safe wake vortex separation, ensure the safety
of aircraft [8,15]. The safe wake vortex separation of ICAO is shown in Table 1.
Table 1. The ICAO safe wake vortex separation (Unit/min)
Aircraft types Heavy (H) Medium (M) Light (L)

Heavy(H) 2 2 3
Medium(M) 2 2 3
Light (L) 2 2 2
In Table 1, the rows indicate the arrival flight in the front of a runway, the
columns indicate the flight that is at the back on the same runway, the cross cells
of row and column represent before and after the two adjacent arrival flights safe
wake vortex separation on the same runway.
184 R. Wang et al.
Table 2. The cost of flight delays per unit time for the three aircraft types
Aircraft type The maximum allowable The cost of delays per

takeoff weight (t) unit time (yuan/min)
Heavy (H) > 136 498
Medium (M) 7–136 236
Light (L) ≤7 92
About the cost of flight delays, according to the literature on the analysis of
the cost of flight delays [18]. Table 2 shows the cost of flight delays per unit time
for the three aircraft types, and the classification standards of aircraft type.
3 Model Construction
The total cost of flight arrival delays is minimized as the optimization goal,
establish the scheduling model of flight arrival. In particular, to minimize the cost
of delays in arrival flight on two runways, involved two runways, it is necessary
to allocate runways to these flights, taking into account the sequence of flight
arrival on the runway. The sequence of flight arrival is related to the safe wake
vortex separation. The type of aircraft and the cost of flight delays per unit time
is also required to consider. These factors affect the cost of flight arrival delays
together.
In order to make the model more accurately describe the process of flight
arrival, the following assumptions are made:
Assumption 1. The number of flights required to arrival during a certain period
of time is N times, and these flights are now required to circle in the terminal
area due to environmental impact and airport capacity constraints.
Assumption 2. From a certain moment, the waiting flight in the terminal area
can in turn to arrival, this moment is remarked as 0.
Assumption 3. From 0 moment, the arrival flight can smoothly through the
runway, taxiway into the apron. This is a continuous process. The allocation of
flight on the taxiway and access to the apron does not have to be considered.
i : There are N flights circling in the terminal area waiting for arrival in
a certain period of time, according to the sequence of arrival terminal area
labeled as i = 1,2, · · · ,N ;
j: Indicates a total of two runways j = 1,2;
k: According to the sequence of flight arrival on runway j, the flight is
remarked as k = 1,2, · · · ,M , and M ≤ N ;
j
Xik : If the i − th flight in the terminal area is k − th arrival on runway j,
j j
then Xik = 1; else Xik = 0;
u: Distinguish the aircraft types of flight. Heavy aircraft remarked as u = 1;
medium aircraft remarked as u = 2; light aircraft remarked as u = 3;
Wu : Represents the cost of delays per unit time for the aircraft type u;
j j
Pku : If the k − th flight arrival is the aircraft type on runway j, then Pku = 1;
j
else Pku = 0;
Qjk,k−1 : The safe wake vortex separation of adjacent flight on runway j;
Tkj : The arrival time of the k − th flight on runway j, it is also the time of
flight delays;
j
Tk,k−1 : The difference in arrival time between adjacent flights on runway j;
C: The total cost of flight arrival delays.
The objective function and constraints are as follows:

N
2
M
3
j
Cmin = Pku W uTkj Xik
j
, i = 1, · · · N ; j = 1, 2; k = 1, 2 · · · , M ; u = 1, 2, 3
i=1 j=1 k=1 u=1
(1)

2
M
j
Xik = 1,i = 1, 2, · · · , N ; j = 1, 2; k = 1, 2, · · · , M (2)
j=1 k=1

N
2
M
j
Xik = N, i = 1, 2, · · · N ; j = 1, 2; k = 1, 2, · · · , M (3)
i=1 j=1 k=1

3
j
Pku = 1, j = 1, 2; k = 1, 2, · · · M ; u = 1, 2, 3 (4)
u=1

3
j j j j
Pku W u = Pk1 W 1 + Pk2 W 2 + Pk3 W 3, j = 1, 2; k = 1, 2, · · · M ; u = 1, 2, 3
u=1
(5)
Tkj − j
Tk−1 = j
ΔTk,k−1 ,j = 1, 2; k = 2, · · · , M (6)
j
ΔTk,k−1 ≥ Qjk,k−1 , j = 1, 2; k = 2, · · · , M (7)
T1j = 0, j = 1, 2 (8)
j j
Xik ∈ {0,1} , Pku ∈ {0, 1} . (9)
Among them, Eq. (1) is the objective function of the model, which represents
the total cost of flight arrival delays; Eq. (2) shows that each flight can only
arrival on one runway; Eq. (3) shows that there are N flights need to arrival at 0
moment; Eq. (4) indicates that each flight can only be one of three aircraft types;
Eq. (5) shows the cost of delays per unit time of the k th flight on runway j; Eq. (6)
shows that the difference in arrival time between the adjacent flights on runway j
j
is ΔTk,k−1 ; Eq. (7) indicates that the difference in arrival time between adjacent
flights on the same runway is not shorter than the safe wake vortex separation;
Eq. (8) shows that the first flight arrival on the runway is at 0 moment and the
time of delays is 0; Eq. (9) are the position and aircraft types selection variable.
And then the specific algorithm is designed to solve the model.
186 R. Wang et al.
4 Algorithm Design
FCFS has not been able to adapt to the flight arrival scheduling problem, and
GEP as a new evolutionary algorithm, has a powerful function, compared with
the traditional evolutionary algorithm, the ability to solve problems has been
greatly enhanced. This chapter focusses on the design and process of IGEP.
4.1 IGEP
The IGEP are used to improve the problem of flight arrival on two runways
and multiple taxiways. It is necessary to improve GEP according to the specific
problem and model. It mainly adjusts the coding method, genetic operators
design and so on of GEP, so that the algorithm is more simple and efficient in
solving problems.
(1) Coding Method
Coding method around the problem, there are two runways mentioned on the
model. In order to simplify the problem, it is possible to code the flight arrival
on runway 1 as “1” and the flight arrival on runway 2 as “2”, regardless of flight
specific allocation on the taxiway. There are a total of m “1” and n “2”, and N
flights arrival on two runways. In accordance with the sequence of arrival, the
flight arrival on runway 1 is encoded as “A1 , A2 , · · · , Am ”, the flight arrival on
runway 2 is encoded as “B1 , B2 , · · · , Bn ”.
(2) Mutation Operator
Mutation is the most efficient operator among all operators with the ability to
modify. It can occur anywhere in the chromosome, and the chromosome struc-
ture was complete after mutation [16]. The root insertion, gene insertion, gene
recombination not used in IGEP, excluding recombination and transposition.
Utilising the mutation operator, the mutation probability Pm determines the
point mutation in the chromosome. Since the coding used only contains the
genotypes “1” and “2”, the most effective mutation is to switch “1” and “2” of
some genes. The gene pair represents the arrival flight on two runways, then it
is easy to get the individual and population after mutation, simplify the process
of GEP, increase the efficiency of solving the problem.
(3) Selection Strategy
In GEP, one of the most common selection strategies is roulette selection, in
which the individual chooses according to their fitness through the roulette sam-
pling strategy. Each individual with a round table of the disc to represent the
fitness, the higher the individual fitness, the greater possibility it becomes the
offspring. Instead of choosing this method, a strategy that is more suitable to
solve the problem of flight scheduling is used in this paper. It preserves the indi-
viduals other than the optimal individual, and ensures that the population size
remains constant.
(4) Fitness Function

The degree of fitness of the individual to the environment is expressed by the
fitness, the individual fitness can be obtained by using the fitness function. The
fitness function is the driving force of GEP population evolution, which directly
affects the convergence speed of the algorithm and whether it can find the opti-
mal solution [9]. In designing the fitness function, the analysis must be based on
specific issues, clear objectives and accurate definition. Where the fitness func-
tion does not use the most common fitness function based on relative error. The
improved fitness function is shown in Eq. (10):
m+n
f (c) = . (10)
Cmin
In Eq. (10), Cmin represents the minimum total cost of flight arrival delays on
two runways, m and n represents the number of arrival flights on two runways.
(5) Flight Retention Policy
The retention of the optimal flight is to retain the optimal individual and the
optimal feasible individual. The optimal individual and the optimal feasible indi-
vidual of the initial population are retained. Then the optimal individual and
the optimal feasible individual of the first generation are compared with the
individuals initially retained. If the individuals are better than the individuals
originally retained, then the first generation of individuals will be retained. Oth-
erwise individuals initially retained become the current generation. Then, in this
manner successively selective substitution and retention.
4.2 IGEP Process

The IGEP steps are as follows:
Step 1. Randomly generate 50 chromosomes as the initial population;
Step 2. These chromosomes are translated into expression trees according to
certain rules;
Step 3. According to the fitness function Eq. (10) to calculate the fitness of
each individual;
Step 4. According to fitness value to determine whether the population to meet
the requirements, if meet the requirements, the process is over; other-
wise, proceed to step 5;
Step 5. Selecting the individuals according to the fitness, substitution and
retention individual with selective;
Step 6. Genetic operations, the selected individuals are subjected to modified
replication using genetic operators. Mutation probability is 0.05, so
that the next generation to retain the new features, and then return to
step 2.
The new population is subjected to the same development process described
above, and the process is repeated until a satisfactory solution occurs, or iterates
to a prescribed algebra, which is set to 500. According to the constructed model
and the improved algorithm, the flowchart of IGEP is obtained, as shown in
Fig. 2.
188 R. Wang et al.
Fig. 2. Flowchart of IGEP
5 Case Simulation and Analysis

The case was developed using C + + and tested with multiple sets of data. The
overall parameter design of IGEP is shown in Table 3.
Case taken from the ZUUU suffered a short period of time in a thunderstorm
weather, resulting in flight delays, this period of time a total of 20 arrival flights
were affected, leading to delay in arrival. Flights sequentially numbered 1–20
Table 3. The Parameter Design of IGEP
Parameter Value
Prescribed algebra 500
Population size 50
Chromosome length 10
Number of genes 10
Mutation probability 0.05
Select range 100%
according to the sequence of arrival terminal area. In accordance with the FCFS
rules, the odd-numbered flights arrival on runway 1 in sequence, and the even-
numbered flights arrival on runway 2 in sequence, and two initial queues l1 and
l2 are obtained according to the arrival sequence of the flight on two runways.
Both queues contain 10 arrival flights, representing flights arrival sequence on
runway 1 and runway 2 respectively, which is flight arrival sequence based on
the FCFS rules and is also regarded as the initial flight arrival sequence L0 .
According to L0 , obtained the total cost of flight arrival delays on two runways
is C0 . Then the initial flight arrival sequence L0 is optimized by IGEP according
to the set parameters. When the algorithm reaches the termination condition,
the optimized flight arrival sequence is Ln , and the total cost of flight arrival
delays is Cn . After simulation, the flight arrival delays of the two schemes are
obtained. As shown in Table 4.
Compare flight arrival delays of the two schemes in Table 4. It can be seen
through optimization of IGEP. The total cost of flight arrival delays C has been
significantly decreased. Compare with the traditional FCFS rules, the total cost
of flight arrival delays decreased from 54,862 yuan to 30,716 yuan through the
IGEP optimization. The total cost of flight arrival delays reduced 24,146 yuan,
a decrease of 44.01%. The average delay cost of each flight from 2743.1 yuan
down to 1535.8 yuan. The total time of flight arrival delays dropped from 43 min
to 38 min, the total time decreased by 5 min, the average arrival time of each
flight from 2.15 min down to 1.9 min, the flight arrival efficiency increased by
11.63%. The same flights arrival, reduced arrival time has proved that runway
utilization has improved. According to the data in Table 4, the line chart of the
cost of flight arrival delays on two runways are obtained, the abscissa represents
the arrival sequence on the runway and the ordinate represents the total cost of
flight arrival delays, as shown in Figs. 3 and 4.
As can be seen from Figs. 3 and 4, the cost of flight arrival delays on two
runways has declined. The cost of flight delays on runway 1 decreased from 27406
yuan to 14590 yuan. The cost of flight delays on runway 2 decreased from 27456
yuan to 16176 yuan. Delay cost fell 12,816 yuan and 11,280 yuan respectively.
The total cost of delays in airlines has significantly decreased, the income has
been safeguarded, these have a positive impact on the development of airlines. At
the same time, according to the trend of the line chart, we can see that when the
190
Table 4. The parameter design of IGEP
Flight
R. Wang et al.
The initial flight arrival sequence L0 The optimized flight arrival sequence Ln
arrival Runway 1 T1j Delay cost Runway 2 T2j Delay cost Runway 1 T1j Delay cost Runway 2 T2j Delay cost
order (min) (yuan) (min) (yuan) (min) (yuan) (min) (yuan)
1 H(1) 0 0 H(2) 0 0 H(2) 0 0 H(2) 0 0
2 L(1) 3 276 M(2) 2 472 H(1) 2 996 H(1) 2 996
3 M(1) 5 1456 L(2) 5 932 H(1) 4 2988 H(2) 4 2988
4 L(1) 8 2192 H(2) 7 4418 M(2) 6 4404 M(1) 6 4404
5 M(1) 10 4552 L(2) 10 5338 M(1) 8 6292 M(2) 8 6292
6 H(1) 12 10528 M(2) 12 8170 M(2) 10 8652 M(1) 10 8652
7 L(1) 15 11908 L(2) 15 9550 L(2) 13 9848 M(2) 12 11484
8 H(1) 17 20374 H(2) 17 18016 L(1) 15 11228 L(2) 15 12864
9 L(1) 20 22214 M(2) 19 22500 L(2) 17 12792 L(1) 17 14428
10 M(1) 22 27406 M(2) 21 27456 L(1) 19 14540 L(1) 19 16176
Delay cost 54862
Delay time 43
Fig. 3. The cost of flight arrival delays on runway 1
Fig. 4. The cost of flight arrival delays on runway 2
number of arrival flights reaches a certain scale, with the increasing number of
arrival flights, the difference between the cost of flight arrival delays before and
after optimization is greater, and the economic significance of the optimization
of the flight arrival problem becomes more apparent.
6 Conclusion
The rapid development of social economy has promoted the change of passen-
ger’s travel mode. The demand for aviation industry has been growing. The
development of aviation industry has also resulted in flight delays and high cost
of flight delays, which impose higher demands on flight arrival scheduling. This
paper studies the problem of flight arrival delays based on the results of domestic
and international flight scheduling research. The total cost of flight arrival delays
is minimized as the optimization goal, construct the scheduling model of flight
arrival, and IGEP is designed according to the specific problem and model.
192 R. Wang et al.
Through practical case studied of ZUUU, it was found that IGEP-optimized

flight arrival scheme, compared with the traditional FCFS scheme, the cost of
flight delays is significantly reduced, also reducing the arrival time of the flight,
the efficiency of flight arrival and runway utilization are improved. It also showed
the effectiveness of IGEP in solving the scheduling problem of flight arrival. For
airlines, this approach is worth promoting. It can reduce the loss of the airlines
in solving the scheduling problem of two-runway flight arrival, and ensure the
interests of airlines. It also has a positive impact on the development of the
aviation industry. The research work in the future can be associated with flight
scheduling on the taxiway, improve the process of flight arrival. In the aircraft
type, the three types can be more detailed division, accurate to each type, in
order to get the conclusion more convincing.
References
1. Alavi A (2011) A robust data mining approach for formulation of geotechnical
engineering systems. Eng Comput 28:242–274 (in Chinese)
2. Azamathulla H, Ahmad Z, Aminuddin A (2013) Computation of discharge through
side sluice gate using gene-expression programming. Irrig Drain 62:115–119 (in
Chinese)
3. Civil Aviation Administration of China (2016) The civil aviation industry in 2015
statistical bulletin [eb/ol]
4. Development Planning Department of Civil Aviation Administration of China
(2016) From the statistical view of Civil Aviation. China Civil Aviation, Beijing
5. Deshpande V, Kan M (2012) Impact of airline flight schedules on flight delays.
Manufact Serv Oper Manage 14:423–440 (in Chinese)
6. Divsalar M (2012) A robust data-mining approach to bankruptcy prediction. J
Forecast 31:504–523 (in Chinese)
7. Dorndor U (2007) Disruption management in flight gate scheduling. Statistica
Neerlandica 61:92–114 (in Chinese)
8. Drüe C (2008) Aircraft type-specific errors in amdar weather reports from com-
mercial aircraft. J R Meteorol Soc 134:229–239 (in Chinese)
9. Fernández-Ares A (2016) Analyzing the influence of the fitness function on genet-
ically programmed bots for a real-time strategy game. Entertainment Comput
18:15–29 (in Chinese)
10. Gandomi A (2011) A new prediction model for the load capacity of castellated
steel beams. J Constr Steel Res 67:1096–1105 (in Chinese)
11. Hu M, Xu X (1994) Ground holding strategy for air traffic flow control. J Nanjing
Univ Aeronaut Astronaut 26:26–30 (in Chinese)
12. Kafle N, Zou B (2016) Modeling flight delay propagation: a new analytical-
econometric approach. Transp Res Part B Methodol 93:20–542 (in Chinese)
13. Karbasi M, Azamathulla H (2016) Gep to predict characteristics of a hydraulic
jump over a rough bed. KSCE J Civil Eng 20:1–6 (in Chinese)
14. Liang W, Li Y (2014) Research on optimization of flight scheduling problem based
on improved gene expression algorithm. Comput Technol Dev 7:5–8 (in Chinese)
15. Marques J (2016) On an analytical model of wake vortex separation of aircraft.
Aeronaut J 120:1534–1565 (in Chinese)
16. Peng J (2015) A new evolutionary algorithm based on chromosome hierarchy net-
work. Int J Comput Appl 30:183–191 (in Chinese)
17. Vasilyev I, Avella P, Boccia M (2016) A branch and cut heuristic for a runway
scheduling problem. Autom Remote Control 77:1985–1993 (in Chinese)
18. Xie T (2009) Study of arrival flight scheduling optimizing based on delay cost. PhD
thesis, Beijing Jiaotong University, Beijing (in Chinese)
19. Xu X, Yao Y (2004) Application of genetic algorithm to aircraft sequencing in
terminal area. J Traffic Transp Eng 4:121–126 (in Chinese)
20. Zhang Q, Hu M, Zhang H (2015) Dynamic multi-objective optimization model of
arrival and departure flights on multiple runways based on RHC-GA. J Traffic
Transp Eng 2:012 (in Chinese)
A Study of Urban Climate Change Vulnerability
Assessment Based on Catastrophe Progression
Method
Yang Sun1 , Yi Lu1(B) , and Yinghan Li2

1
Business School, Sichuan University, Chengdu 610065,
People’s Republic of China
luyiscu@163.com
2
Software School, Harbin University of Science and Technology,
Harerbin 150022, People’s Republic of China
Abstract. Climate change can make a profound impact on the entire

ecosystem, and severely affects the stability of human society. The urban
is the economic, political and cultural center of human society. The
assessment of urban disaster vulnerability is of great significance to the
disaster prevention system. Through the review of the relevant litera-
tures on disasters vulnerability and the situations of Sichuan Province,
the study establishes urban climate change actual vulnerability assess-
ment systems. The vulnerability degrees of the five economic zones of
Sichuan Province are evaluated by the improved mutation progression
method based on the entropy method. The study shows that the urban
disaster vulnerability of Southern Sichuan Economic Region is the most
serious, and the degree of disaster vulnerability of Chengdu Economic
Region is the lowest. Chengdu Economic Region should concern about
the floating population. And Southern Sichuan Economic Region needs
to create a good environment for enterprises to increase the number of
enterprises within its jurisdiction.
Keywords: Climate change · Vulnerability · Sichuan province · Muta-

tion progression method · Entropy method
1 Introduction
Climate change has already made a serious impact on human beings [1]. It not
only makes social economy worse [3], increases the flood exposure of cities [6]
and the difficulty of reservoir management [14], severely threaten the human
health [11], but also destroys the diversity and the stability of ecosystem [2,9].
Therefore, climate change has been becoming an important issue related to the
survival and sustainable development of human beings. To minimize the threat
caused by climate change, effective measures need to be formulated.

A Study of Urban Climate Change Vulnerability Assessment 195
The vulnerability can comprehensively assess the impact of natural disasters

[7], and provide a fundamental basis for effective adaptation to climate change
[12]. And the vulnerability assessment of urban climate change disasters is an
important link to enhance the ability of urban disaster response, and has impor-
tant practical significance [8].
After years of development, foreign vulnerability research has acquired
remarkable achievements. The Intergovernmental Panel on Climate Change
(IPCC) has established a system of climate change vulnerability method, which
has been applied to 24 countries [5]. The Tyndall Center constructed a national
climate vulnerability assessment system [15]. China’s relevant researches on the
vulnerability of climate change disasters began in 1990s, which mainly were
related to the impact of climate change on water resources and agricultural dis-
aster, rarely including the regional composite system [17].
The influences caused by climate change are widespread and multi-level [11].
However, the current research focuses on the impact of climate change on the
coastal cities [6], few of those studies of climate change in inland cities. In our
study, Sichuan Province, an inland region of China, is regarded as the research
object. The urban climate change vulnerability study of Sichuan five economical
areas is based on catastrophe progression method. The result of our survey can
provide objective statement of economical regions’ urban vulnerability, and offer
effective references to urban climate change response.
2 Method
The catastrophe progression method combines catastrophe theory and fuzzy
mathematics [10]. And in order to make the sort of indicators more objective
and reduce the influence of subjective factors, the evaluation indexes are ranked
by the weight. The catastrophe progression method can be divided into 6 steps.
Step 1. Establish assessment systems.
The overall goal is divided into multi-level indicators. And the total number
of the corresponding indicators is not more than four, which can be expressed
by the parameter n.
Step 2. Identify the model of system.
⎧ 3
⎪
⎪ x + ax, n=1
⎨ 4
x + ax2 + bx, n=2
f (x) = 1 5 1 3 1 2 (1)
⎪
⎪ x + ax + bx + cx, n=3
⎩ 51 6 31 4 21 3 1 2
6 x + 4 ax + 3 bx + 2 cx + dx, n = 4.
The f (x) represents the potential function of the state variable x. And a, b, c
represent the control variable of the state variable x.
Step 3. Data standardization.
⎧
⎨ xij / max(xij ), xij ∈ x+ ij
∗ j
xij = (2)
⎩ min(xij )/xij , xij ∈ x− ij .
j
196 Y. Sun et al.
The xij means the evaluation index j of object i. The x+

ij is a benefit attribute.
−
The xij is a cost type attribute. The maxj (xij ) is maximum value of evalua-
tion index j. The minj (xij ) is minimum value of evaluation index j.
Step 4. Calculate the weight of each index by the entropy method, and rank the
evaluation indexes based on the weights of evaluation indexes.

m
pij = x∗ij x∗ij , j = 1, 2, · · · , n, (3)
i=1
1
m
ej = − pij ln pij , j = 1, 2, · · · , n, (4)
ln m i=1

n
wj = (1 − ej ) (1 − ej ), j = 1, 2, · · · , n. (5)
j=1
The m represents the number of evaluated objects. And n indicates the num-
ber of indicators. The value of wj is the weight of each index.
Step 5. Confirm the normalized equation of each model.
⎧
⎪
⎪ xa = a1/2 , n=1
⎨
xa = a1/2 , xb = b1/3 , n=2
x= (6)
⎪
⎪ x = a1/2 , xb = b1/3 , xc = c1/4 , n=3
⎩ a
xa = a1/2 , xb = b1/3 , xc = c1/4 , xd = d1/4 , n = 4.
Step 6. Comprehensive evaluation.

According to the multi objective fuzzy decision theory, T1 , T2 , · · · , Tn can be
regarded as a fuzzy goal for the same plan in the multi-objective situation.
Then the ideal membership function is u(x) = uT1 (x) ∩ uT2 (x) ∩ · · · ∩ uTn (x),
of which uTi (x) is the function of the fuzzy decision Ti . By unitary formula,
the corresponding value of the state variable can be confirmed.
3 Case Study
Sichuan province is located in the hinterland of the China’s southwest, and on the
eastern of Qinghai-Tibetan plateau. Due to the unique geographical environment
of the plateau, the climate change in Sichuan is very obvious. Meanwhile, Sichuan
is located in the Longmen Shan Fault, in which geological disasters easily happen
[13]. The sudden heavy rainfall caused by climate change led to the city water
logging, and even geological disasters (e.g., landslide, debris flow). The basic
situation of cities in each economic zone is shown in Fig. 1.
3.1 Construct the Index System

The vulnerability of urban disaster prevention system consists of response, sen-
sitivity, resilience [4]. The related researches of vulnerability don’t form a uni-
fied vulnerability assessment system, and the selection of vulnerability index
GDP
PXER,
NWSER, 7.6%
1.6%
SSER,
17.1%
NESER, CDER,
16.3% 57.5%
NESER
NWSER
CDER
SSER CDER: Chengdu Economic Region

NESER: Northeast Sichuan Economic Region
PXER NWSER: Northwest Sichuan Economic Region
PXER: Pan-Xi Economic Region
SSER: South Sichuan Economic Region
Fig. 1. Overview diagram of Sichuan Province
varies according to the real problems [16]. Sichuan Province is divided into five
major economic zones, including Chengdu Economic Region (CDER), North-
east Sichuan Economic Region (NESER), Northwest Sichuan Economic Region
(NWSER), Pan-Xi Economic Region (PXER), and South Sichuan Economic
Region (SSER). In order to make the economic zone of the social vulnerability
index objective and reasonable, the vulnerability assessment of urban climate
change disaster is established based on actual situation of Sichuan (Table 1).
3.2 Assess the Urban Vulnerability of Climate Catastrophe
Our study regards Sichuan Statistical Yearbook (2015) as the source of the
original data. According to Eqs. (2)–(6), the standard data and the importance
of each indicator can be calculated. Table 1 shows the sorted result of each layer
index.
In the assessment system, the number of each layer’s indexes is not more
than 4. According to the number of each level’s indicators, the vulnerability
of all levels’ indicators can be calculated. The urban vulnerability of Sichuan
economic zone is shown in Table 2. The larger the number of values in the table,
the more serious the degree of vulnerability. As shown in Table 2, the disaster
vulnerability of the CDER is the lowest, and the ability of disaster prevention
and mitigation is the highest. The vulnerability of the SSER is the highest, whose
ability of disaster prevention and reduction is the worst.
198 Y. Sun et al.
Table 1. Vulnerability assessment system and index importance of urban climate

change disasters prevention system
The target layer Criterion layer Index layer Measure (Importance)

(Importance) (Importance)
Urban Capacity of Primary The number of Enterprises A11
vulnerability of coping A organization A1 (0.3190)
climate (0.6285) (0.4811)
catastrophe
The number of neighborhood
committees A12 (0.1620)
Traffic situation A2 Per capita road area A21
(0.0873) (0.0579)
Per capita car number A22
(0.0294)
Awareness of Per capita publicity
disaster A3 (0.0404) expenditure A31 (0.0263)
Per capita educational
expenditure A32 (0.0141)
Medical condition Per capita number of medical
A4 (0.0197) institutions A41 (0.0121)
Per capita number of medical
personnel A42 (0.0039)
Per capita number of sickbeds
A43 (0.0036)
Sensitivity B Urban status B1 Drain pipe length B11 (0.1237)
(0.2125) (0.1240)
Per capita housing area B12
(0.0002)
Greening rate B13 (0.0001)
The structure of Population ratio of Higher
population B2 Education B21 (0.0499)
(0.0885)
Ratio of old people B22
(0.0223)
Population density B23
(0.0062)
Resilience C Self rescue ability Institutional deposit amount
(0.1590) C1 (0.1304) C11 (0.1299)
Per Capita disposable income
C12 (0.0006)
Social security C2 Coverage rate of unemployment
(0.0285) insurance C21 (0.0223)
Coverage rate of medical
insurance C22 (0.0062)
Table 2. Assessment and sorting of urban disaster vulnerability
Region A1 A2 A3 A4 B1 B2 C1 C2 F Rank
CDER 0.1656 0.4301 0.553 0.5751 0.4656 0.4836 0.3493 0.4725 0.6379 5
NESER 0.1917 0.5166 0.5923 0.6026 0.5081 0.6781 0.442 0.6049 0.6616 4
NWSER 0.5643 0.4379 0.3448 0.5118 0.5209 0.5456 0.6892 0.5062 0.8667 2
PXER 0.3514 0.4093 0.4762 0.5398 0.5687 0.4881 0.509 0.4058 0.7699 3
SSER 0.6636 0.6883 0.5596 0.5888 0.5209 0.5456 0.4379 0.5566 0.897 1
3.3 Evaluation Results and Comparative Analysis
The vulnerability distribution map, which is plotted according to the economic

zones’ vulnerability indexes, can visually help decision makers get regional vul-
nerability status and significant factors. Urban planning development and pre-
vention of climate disasters can effectively reduce the city disaster vulnerability
level. According to the calculation results of Table 2, our study draws the status
of the disaster vulnerability, disaster response, sensitivity and resilience of each
economic zone, as shown in Fig. 2.
NESER NESER
NWSER
NWSER
CDER CDER
SSER SSER
PXER PXER
(a) Urbun vulnerability (b) Capacity of coping
NESER NESER
NWSER NWSER
CDER CDER
SSER SSER
PXER PXER
(c) Sensitivity (d) Resilience

Severe High Moderate Low Mild
Fig. 2. Urban vulnerability of change-induced disasters
(1) Chengdu Economic Region (CDER)

CDER consists of 8 cities, including Chengdu, Deyang, Mianyang, Leshan,
Meishan, Ziyang, Suining and Yaan. As shown in Fig. 1, the total GDP
amount and urban built-up areas are more than half of the all urban cor-
responding indicators, the city’s population of which is above one-third of
200 Y. Sun et al.
the total urban population. Therefore, the urban disaster vulnerability of

CDER plays an important role in whole Sichuan urban climate disasters
prevention systems.
Among all economic zones, the disaster vulnerability, coping ability, vul-
nerability, sensitivity and vulnerability resilience vulnerability indicators of the
CDER all place in the lowliest place. The CDER’s overall situation of climate
disasters prevention is relatively benign. The value of the vulnerability index
A3 and A4 is higher than other vulnerability indicators of CDER, which means
disaster awareness and hospital conditions need to be improved. As the CDER
is the core area of the entire Sichuan region, the overall level of medical care
and economy is better than other economic regions. In the context of popula-
tion mobility, the actual population of CDER is much larger than the data of
the Yearbook, so the actual vulnerability is more serious than that in Table 2.
And to minimize the urban vulnerability of the CDER, appropriate measures
must be taken, especially on the disaster awareness and medical conditions. The
urban management of CDER should offer much more attention to the floating
population, and effectively improve the urban climate disaster vulnerability.
(2) Northeast Sichuan Economic Region (NESER)

The NESER includes 5 cities i.e., Guangyuan, Nanchong, Guangan, Dazhou
and Bazhong. As shown in Fig. 2, the rank of the NESER’s resilience vul-
nerability is more serious than that of coping vulnerability and sensitivity
vulnerability. Disaster sensitivity of the NESER needs to be promoted, so
that more energy should be put into the economic construction and the
social security expenditure. Among them, the urban economic development
is the most important link.
NESER is located in the northeast of Sichuan, which is an important

channel to link Chengdu-Chongqing Economic Zone. NESER has rich mineral
resources, especially natural gas reserves. The overall economic development level
of ERSER is relatively backward; the communication and cooperation between
cities in the region is poor, and at present there are no strong core cities to
promote the economic development.
Relying on its geographical advantage and the construction of Chengdu-
Chongqing Economic Zone, NESER can build wide-ranging and multi-level pat-
tern of opening to the outside regions. The establishment of regional economic
center will give priority to develop the core city, and injects impetus for economic
development of the urban. These measures can not only promote the economic
development of cities, but also can improve the disaster recovery, reduce the
vulnerability of urban disasters, and increase the ability to withstand climate
disasters.
(3) Northwest Sichuan Economic Region (NWSER)

NWSER is located in the northwest of Sichuan Province, of which the cli-
mate is plateau-climate. And the tourism resources and the natural resources
are very rich. As a national autonomous region, the Northwest Economic
Zone covers a larger area than other economic zones. However, its urbaniza-
tion process is slow. The ratio of the GDP, urban built-up area and urban
population is much less than other areas (Fig. 1). As shown in Fig. 2, the
sensitivity and vulnerability of the urban economic zone is very serious. The
urban situation, the structure of population, self rescue ability and social
security is poor.
In the vulnerability system, the population structure consists of higher educa-

tion ratio, elderly population ratio and population density. And these indicators
have a great relationship with local talent attractiveness, and the ratio of the
young. In order to reduce the vulnerability, urban management departments
need to increase investment in municipal construction, and improve urban facil-
ities. On another hand, effective measures should be taken to make the urban
appeal to the highly educated young and adults.
NWSER is an important tourist area in China, and its economy is mainly
from agriculture sector. In view of these regional characteristics, the commercial
organizations can build or upgrade related tourist facilities (e.g., the road, the gas
stations and other infrastructure or public services), and make full use of these
resources, so as to retain high-quality personnel and construct the ecological or
tourist city.
(4) Pan-Xi Economic Region (PXER)

PXER is composed of Panzhihua and Liangshan. This economic region has
a huge potential for economic development. As shown in Fig. 2, the overall
urban vulnerability level of PXER is moderate in the five major economic
zones. The root reasons of PXER’s high vulnerability are similar to the
causes of NWSER i.e., the less attractiveness and the shortage of relevant
financial expenditure. And relative measures must be based on the improve-
ment of the urban regional economic level.
In the background of that the central government of China strengthens the

research of the vanadium and hydropower resources, this economic zone can take
full of the mood of the time and transform the mainly strategic shift from the
steel to the vanadium titanium new material with its abundant natural resources.
To provide funding for municipal construction and enhance the attractiveness,
the management of PXER should innovate the development model and increase
revenue.
(5) South Sichuan Economic Region (SSER)

SSER is located in the upper reaches of the upper reaches of the Yangtze
River, adjacent to Yunnan-Guizhou Plateau. The economic region is an
important transportation hub, and has always been the most important
materials trading center.
As shown in Fig. 2, the coping capacity vulnerability of SSER is the most

serious among all economic zones, and need to be paid much more effort. The
urban disaster response consists of grass-roots organizations, traffic conditions,
202 Y. Sun et al.
disaster awareness and medical conditions, in which the importance of grass-

roots organizations is higher than other indicators. In this assessment system,
the primary organization is composed of enterprises and community committees.
According to the data of Table 2, we can know that in order to reduce the urban
vulnerability, the number of enterprises needs to be increased.
Although SSRS has a comprehensive transportation system, the level of
development integration is low. And the overall advantages of this economic
region can’t be fully utilized in the long term. In addition, the transformation
pace of the old industrial base is very slow.
By actively exploring new model of integration development, SSRS can slow
the disaster vulnerability. The process of urban integration can be promoted by
establishing the development fund of Southern Sichuan urban group. SSRS has a
long wine culture. By supporting wine enterprises diversified development, intel-
ligent manufacturing, upgrading of traditional manufacturing machinery and
service oriented manufacturing transformation can be accelerated. Meanwhile,
effective measures should be taken to develop high growth industry, and accel-
erate the exploration or development of shale gas. By various ways, enterprises
can provide a good environment.
4 Conclusion
As an important part of disaster theory, vulnerability assessment is an important

way to estimate the urban coping ability to disasters. To evaluate the urban
vulnerability of climate change disasters, our study regards the five economic
zones of Sichuan Province as objects and establishes the assessment systems.
Through data analysis and discussion, our study found that the economic
zones should take appropriate measures based on its own characteristics to
reduce the urban vulnerability. CDER needs to pay more attention to the float-
ing population. By giving priority to the development of core cities, NESER can
boost the whole urban economic development and enhance the disaster resilience.
NWSER can make full use of its agriculture and tourism advantages, and provide
much more support for enterprises engaged in deep processing of agricultural
products or developing tourism resources. By the establishment of ecological
cities, financial expenditure of NWSER can be guaranteed.
To slow the urban vulnerability, PXER needs to innovate the resource
exploitation model, carry out the strategic transformation and increase the fiscal
revenue. Relying on its rich history of wine culture, SSER can achieve the diver-
sified development of wine enterprises. In addition, through creating a favorable
environment for the development of enterprises, the disaster response capacity
can be increased. And SSER can reduce the level of disaster vulnerability.
Acknowledgments. This research is supported by the Program of Social Science

Foundation of Sichuan (Grant No. SC16C010). The authors are also grateful to Sichuan
University for providing research funding (No. 2016SCU11036; No. skqy201640).
References
1. Bachelet D, Ferschweiler K et al (2016) Climate change effects on southern cali-
fornia deserts. J Arid Environ 127:17–29
2. Bellard C, Bertelsmeier C et al (2012) Impacts of climate change on the future of
biodiversity. Ecol Lett 15(4):365–377
3. Bowen A, Cochrane S, Fankhauser S (2011) Climate change, adaptation and eco-
nomic growth. Clim Change 113(2):95–106
4. Cutter SL (1996) Vulnerability to environmental hazards. Prog Hum Geogr
20(4):529–539
5. Füssel HM, Klein RJT (2006) Climate change vulnerability assessments: an evo-
lution of conceptual thinking. Clim Change 75(3):301–329
6. Hallegatte S, Green C et al (2013) Future flood losses in major coastal cities.
Nature Clim Change 3(9):802–806
7. Mazumdar J, Paul SK (2016) Socioeconomic and infrastructural vulnerability
indices for cyclones in the Eastern Coastal States of India. Nat Hazards 82(3):1621–
1643
8. She L, Wang G, Xu J (2013) Research of influential factors of urban vulnerability
based on interpretative structural modeling: a case study of rescue and recovery in
Japan earthquake. Areal Res Develop (in Chinese) 12:18
9. Shen G, Pimm SL et al (2015) Climate change challenges the current conservation
strategy for the giant panda. Biol Conserv 190:43–50
10. Songhe X, Chuanfeng H, Lingpeng M (2015) Assessment and the key influencing
factors analysis of urban disaster prevention system vulnerability. Soft Science
11. Thompson TM, Rausch S et al (2014) A systems approach to evaluating the air
quality co-benefits of us carbon policies. Nature Clim Change 4(10):917–923
12. Wild M (2009) Global environmental change. J Geophys Res Atmos 114(11):733–
743
13. Xu J, Xie H et al (2015) Post-seismic allocation of medical staff in the longmen
shan fault area: case study of the Lushan earthquake. Environ Hazards 14(4):1–23
14. Yang G, Guo S et al (2016) Multi-objective operating rules for danjiangkou reser-
voir under climate change. Water Resour Manage 30(3):1183–1202
15. Yi L, Xi Z et al (2014) Analysis of social vulnerability to hazards in china. Environ
Earth Sci 71(7):3109–3117
16. Yoon DK (2012) Assessment of social vulnerability to natural disasters: a compar-
ative study. Nat Hazards 63(2):823–843
17. Zhang Q, Meng H (2014) Social vulnerability and poverty to climate change: a
summary on foreign research. J China Agric Univ (in Chinese) 31:29–33
Judging Customer Satisfaction by Considering
Fuzzy Random Time Windows in Vehicle
Routing Problems
Yanfang Ma1,2 , Cuiying Feng2 , Jing Zhang1,2 , and Fang Yan3(B)

1
School of Economics and Management, Hebei University of Technology,
Tianjing 300401, People’s Republic of China
2
College of Economics and Management, Zhejiang University of Technology,
Hangzhou 310014, People’s Republic of China
3
School of Economic and Management, Chongqing Jiaotong University,
Chongqing 400074, People’s Republic of China
yanfang@cqjtu.edu.cn
Abstract. This article puts forward a membership function of the cus-

tomer satisfaction based on fuzzy random time windows in vehicle rout-
ing problems. The objective is to confirm that all the customers are sat-
isfied in an acceptable degree by judging the vehicle arriving time. More
specifically, time windows given by customers are taken as fuzzy random
variables in this paper. And then, fuzzy random theory has been used to
describe customers’ time windows. Finally, a measure function has been
given to calculate customers’ satisfaction based on fuzzy random time
windows.
Keywords: Vehicle routing optimization · Customer satisfaction · Time

windows · Fuzzy random variable
1 Introduction
Generally, in the classical VRP, there is a set of customers, each of whom have
their own demands. Vehicles in the same condition at the depot deliver goods
to these customers, and they are required to start and end at the depot. The
objective of the classical VRP is to minimize the total cost by designing an opti-
mal delivery route for each vehicle. Nowadays, the VRP is a common problem in
almost every industry such as supply chain management and transport planning.
With the development of modern technology, customer satisfaction has become
a hot issue in the VRP.
In the classical VRP, delivery vehicles usually need to meet the following con-
ditions: (1) Serve all customers using a minimum of vehicles; (2) Each customer
is served by only one vehicle once; (3) Each vehicle starts and ends at the depot;
(4)Total customer demand on each route cannot exceed the load capacity of the
vehicle. There are several VRP variants, such as the multi-depot VRP [1,9], the
Considering Fuzzy Random Time Windows in Vehicle Routing Problems 205
Pickup and Delivery VRP [10], and the VRP with Backhauls [16] and so on. In
addition, some problems have pre-set time constraints on the period of the day in
which deliveries should take place. These are known as Vehicle Routing Problems
with Time Windows (VRPTW) which is a well known variant of the VRP.
The VRPTW is classified into two types, namely a VRP with a hard time
window (VRPHTW) in which the time windows are hard constraints and where
a route is not feasible if the service for any customer does not start within the
limits established by the time window [3], and a VRP with a soft time window
(VRPSTW), which is a relaxation of the VRPHTW, where the delivery of goods
is allowed outside the time windows if a penalty is paid [4,5,12]. There are many
uncertainties in real-life applications where hard time windows can be violated,
while the “soft time window” can deal with these situations [13,14]. The travel
time on each road section cannot be determined because of many uncertainties,
such as traffic incidents, vehicle breakdowns, work zones, special events, and
driver skills and experience [6].
In the past, the research used to consider these uncertainties random variables
[6,15], but recently, researchers have been increasingly applying fuzzy member-
ship functions to characterize the service level issues associated with time win-
dow violation in a vehicle routing problem, called the VRP with fuzzy time
windows (VRPFTW) [2,7,14]. This research has highlighted the fact that fuzzy
factors and stochastic factors can exist at the same time in the VRPTM. In this
paper, we propose to use fuzzy random theory to describe the VRPTW, namely
VRPFRTM.
This paper is organized as follows. In Sect. 2, we describe the problem which
includes fuzzy random time windows. And then, in Sect. 3, we discuss the meth-
ods for dealing with these uncertainties, and propose a procedure to handle fuzzy
random time windows. Next, we proposed the membership function of the cus-
tomer satisfaction based on the fuzzy random time windows in Sect. 4. Finally,
some concluding remarks are outlined in Sect. 5.
2 Fuzzy Random Time Windows

The time windows in the traditional VRPTW can be described as follows. Each
customer needs to be served in a certain time window, and no delay is permitted.
This certain time window can be described as [e, l], where e and l are the earliest
and latest time which the customer can accept, respectively. As to vehicle routing
problem with soft time windows (VRPSTW), the vehicle can provide service
before e and also after l. Some researchers in this field believe that the violation
adds some costs and these costs are related to the degree of the violation [14].
Adopting soft time windows often means loss in customer satisfaction. Soft
time windows certainly can bring some economic benefits, while dissatisfaction
causes customer defection in a long term, which leads to economic loss. It is
necessary to maintain the customer satisfaction on a certain level and minimize
the cost in the mean time. Some researchers have use fuzzy theory to describe
the customer satisfaction level, such as [14]. In their research, it is believed that
206 Y. Ma et al.
the customer satisfaction level is related to the start time. If a customer requires
that the service should be started in [e, l] in traditional VRP with hard time
windows, earlier than e or later than l is unacceptable, and the satisfaction level
is 0; otherwise, the customer is satisfied and the satisfaction level is 1. While
considering the soft time windows, the service starts earlier than e or later than
l is permitted in some extent, but no earlier than EET or later than ELT
[14]. Hence, if service starts between e and l, customer satisfaction level is 1; if
service starts between EET and e or between l and ELT customer satisfaction
level is valued in [0, 1]; otherwise, customer satisfaction level is 0. The customer
satisfaction level from hard to soft time windows can be seen in Fig. 1.
Sa Sa
1 1
0 e l t 0 EET e l ELT t
Hard time windows Soft time windows
Fig. 1. The customer satisfaction level from hard to soft time windows
From previous studies, it can be seen that the EET and ELT are usually be
treated as known and deterministic, while in practice, they can not usually be
obtained as deterministic data. There are two common way to get the data, one
is reasoning and the other is consulting the client. For reasoning, for example, the
concreting in a construction project should be started at 9:00, and the process of
unloading the concrete requires 10 min. By reasoning, the ELT is suppose to be
8:50, while actually the acceptable ELT to the project manager may be 8:30. As
seen, reasoning often leads to false EET or ELT due to the lack of flexibility. As
to consulting the client, deterministic data can not be obtained easily because the
client usually gives a response including ambiguous information. The customer
may make a statement, such as “not too early” or “no late than 10:00”. After
getting these information, the specific data can not be acquired obviously, and if
it is dealt as deterministic, the part of information which the customer provides
will be lost. In sum, a response often not only includes some fuzzy information
but also some random information. In this paper, the vehicle routing problem
with fuzzy random time windows (VRPFRTM) is proposed and fuzzy random
theory is chosen to describe the fuzzy random time windows. EET and ELT
are considered to be fuzzy random, namely EET and ELT .
3 Dealing with Fuzzy Random Variables

Uncertainty is absolute and certainty is relative. In practice, the fuzzy random
variable has been applied in many areas. However, it is difficult to deal with
uncertain variables, especially fuzzy random variables. Based on previous studies

about fuzzy random theories, a method to transform a fuzzy random variable
into a deterministic one is proposed in this paper. Take fuzzy random variable
as an example. The procedure of the transform operation can be described
EET
as follows:
Step 1. Consider the endurable earliness time of customer, EET , is fuzzy ran-
dom variable. Through previous data and professional experience using
statistical methods, estimate the parameters [m]L , [m]R , μ0 and σ02 .
Step 2. Obtain the intermediate parameters by using group decision making
approach, namely the decision-makers’ degree of optimism. From Puri
and Ralescu’s definition in [11], a fuzzy random variable is a measur-
able function from a probability space to a collection of fuzzy variables.
Roughly speaking, a fuzzy random variable is a random variable taking
is denoted
fuzzy values. In this paper, the fuzzy random variable EET
= ([m] , ρ(ω), [m] ), where ρ(ω) ∼ N (μ , σ 2 ) whose proba-
as EET L R 0 0
bility density function is ϕρ (x). Thus the expression of ϕρ (x) should
(x−μ0 )2
1 − 2
2σ0
be ϕρ (x) = √2πσ e . Suppose that σ is a probability level
0x
and σ ∈ [0, sup ϕρ (x)], r is a possibility variable and r ∈ [rl , 1] where
R −[m]L
rl = [m]R[m]
−[m]L +ρR L , both of them reflect the decision-maker’s degree
σ −ρσ
of optimism.
Step 3. Let ρσ be the σ-cut of the random variable ρ(ω). According to Xu and
Liu’s lemma in [17], ρσ = [ρL R
σ , ρσ ] = {x ∈ R|ϕρ (x) ≥ σ}, and the value
L R
of ρσ and ρσ can be expressed as
√
ρL
σ = inf{x ∈ R|ϕρ (x) ≥ σ} = inf ϕ−1
ρ (σ) = μ0 − −2σ02 ln( 2πσ0 σ),
√
ρR −1
σ = sup{x ∈ R|ϕρ (x) ≥ σ} = sup ϕρ (σ) = μ0 + −2σ02 ln( 2πσ0 σ).
Step 4. Transform the fuzzy random variables EET = ([m] , ρ(ω), [m] ) into
L R

the (r, σ)-level trapezoidal fuzzy variable ω by the following
EET (r,σ)
equation:
→ω
EET = ([m]L , m, m, [m]R ).
EET (r,σ)
here, we have
√
m = [m]R − r([m]R − ρL
σ ) = [m]R − r([m]R − μ0 + −2σ02 ln( 2πσ0 σ)),
√
m = [m]R − r([m]L + ρR
σ) = [m]L + r(μ0 − [m]L + −2σ02 ln( 2πσ0 σ)).
208 Y. Ma et al.
can be specified by ω
EET = ([m]L , m, m, [m]R ) with the mem-
EET (r,σ)
bership function:
⎧
⎪
⎪ 0[m] −x for x > [m]R
⎪
⎪
⎪ R
⎨ [m]R −m for m ≤ x ≤ [m]R
μω̃ = 1 for m≤x≤m
EET (x) ⎪
⎪ x−[m]L
⎪
⎪ for [m]L ≤ x ≤ m
⎪
⎩ m−[m]L
0 for x < [m]L .
to the (r, σ)-
The process of transforming fuzzy random variable EET

level trapezoidal fuzzy variable ω is illustrated in Fig. 2.
EET (r,σ)
Fig. 2. The process of transforming fuzzy random variable
Step 5. Take expected value operator to convert (r, σ)-level trapezoidal fuzzy
variable into a deterministic one. Based on the definition of fuzzy inter-
val and expected value by Heilpern [8], suppose that there is a fuzzy
number Ñ = (a, b, c, d), its membership function is:
⎧
⎪
⎪ 0 for x < a
⎪
⎪
⎨ fÑ (x) for a ≤ x ≤ b
μÑ (x) = 1 for b ≤ x ≤ c
⎪
⎪
⎪
⎪ gÑ (x) for c ≤ x ≤ d
⎩
0 for x > d.
where fÑ (x) and gÑ (x) are the upper and lower ends of the fuzzy num-
ber respectively. Then, the expected value of fuzzy variables ω̃
EET (r,σ)
is as follows:
m [m]R
1
EV [
ω ]= [(m − fω (x)dx) + (m + gω (x)dx)].
EET (r,σ) 2 EET (r,σ) EET (r,σ)
[m]L m
Based on the what has been discussed above, the fuzzy random time windows
can be achieved, see Fig. 3.
Fig. 3. The customer satisfaction level from soft to fuzzy random time windows
4 Customer Satisfaction by Fuzzy Random Time

Windows
Finally, it comes how to define the customer satisfaction based on the fuzzy
random time windows. In this paper, we suppose that there exists a least average
customer satisfaction level Sa which is accepted by the customers [14].
n
1
Li (ti ) ≥ Sa, (1)
n i=1
where n is the number of customers. Li (ti ) is the satisfaction level serving cus-
tomer i. Sa is the least average customer satisfaction level. The membership
function of the satisfaction level, Li (ti ), is as below [14]:
⎧
⎪ ,
⎪
⎪ 0, ti < EET i
⎪
⎪
⎪
⎪ ti − EET i
, EET i ≤ ti < ei ,
⎪
⎨
ei −EET i
Li (ti ) = 1, e ≤ ti < t,
⎪
⎪
⎪
⎪
−t
ELT ,
, li ≤ ti < ELT
⎪
⎪ −l
i i
i
⎪
⎪ ELT i i
⎩
0, ti ≥ ELT i .
As discussed above, the service may start outside the time window [e, l],
and
and the bounds of acceptable earliness and lateness are described by EET
, respectively. Obviously, the earliness and lateness are closely related to
ELT
the quality of service of the supplier. The response of a customer satisfaction
level to a given service time may not be simply “good” or “bad”; instead, it may
210 Y. Ma et al.
be between “good” and “bad”. For example, the customer might say, “it’s all
, e] or [l, ELT
right” to be served within [EET ]. In either case, the service level
cannot be described by only two states (0 or 1).
5 Conclusion
In this paper, we focus on a special type of the VRPTW, the vehicle rout-
ing problem with fuzzy random time windows and multiple decision makers
(VRPFRTW-MDM) which is seldom considered before. We present a member-
ship function of the customer satisfaction based on fuzzy random time windows
in vehicle routing problems. Since customer satisfaction is becoming more and
more important for suppliers, the objective of this paper is to confirm that all the
customers are satisfied in an acceptable degree by judging the vehicle arriving
time. We also proposed a method to deal with the fuzzy random time windows
based on the fuzzy random theory. In the end, we have given a measure func-
tion on how to obtain the customers’ satisfaction based on fuzzy random time
windows.
Acknowledgement. This research was supported by National Natural Science Foun-

dation of China (Grant No. 71640013, and Grant No. 71401020), and System Science
and Enterprise Development Research Center (Grant No. Xq16C10), National philos-
ophy and Social Science Foundation of Tianjin (TJGL15-007); Provincial Science and
Technology Foundation of Hebei Province (16457643D).
References
1. Aras N, Aksen D, Tuğrul Tekin M (2011) Selective multi-depot vehicle routing
problem with pricing. Transp Res Part C Emerg Technol 19(5):866–884
2. Brito J, Martı́nez F et al (2012) Aco-grasp-vns metaheuristic for vrp with fuzzy
windows time constraints. In: Computer aided systems theory–EUROCAST 2011,
Springer, pp 440–447
3. Cordeau JF, Desaulniers G et al (2002) Vrp with time windows. Veh Routi Prob
9:157–193
4. Errico F, Desaulniers, et al (2016) The vehicle routing problem with hard time
windows and stochastic service times. Euro J Trans Logist:1–29
5. Figliozzi MA (2010) An iterative route construction and improvement algorithm
for the vehicle routing problem with soft time windows. Trans Res Part C Emerg
Technol 18(5):668–679
6. Gao S, Chabini I (2006) Optimal routing policy problems in stochastic time-
dependent networks. Trans Res Part B Methodol 40(2):93–122
7. Ghoseiri K, Ghannadpour SF, Seifi A (2010) Locomotive routing and scheduling
problem with fuzzy time windows. In: Transportation research board 89th Annual
Meeting (No: 10-2592)
8. Heilpern S (1992) The expected value of a fuzzy number. Fuzzy Sets Syst 47:81–86
9. Lightnerlaws C, Agrawal V et al (2016) An evolutionary algorithm approach for
the constrained multi-depot vehicle routing problem. Int J Intell Comput Cybern
9(1)
10. Nair DJ, Grzybowska H, et al (2016) Food rescue and delivery a heuristic algorithm
for periodic unpaired pickup and delivery vehicle routing problem. Trans Res Board
11. Puri ML, Ralescu DA (1986) Fuzzy random variables. J Math Anal Appl
114(2):409–422
12. Qureshi A, Taniguchi E, Yamada T (2009) An exact solution approach for vehicle
routing and scheduling problems with soft time windows. Transp Res Part E Logist
Transp Rev 45(6):960–977
13. Shi C, Li T et al (2016) A heuristics-based parthenogenetic algorithm for the vrp
with potential demands and time windows. Sci Program 1:1–12
14. Tang J, Pan Z et al (2009) Vehicle routing problem with fuzzy time windows. Fuzzy
Sets Syst 160(5):683–695
15. Taş D, Dellaert N et al (2012) Vehicle routing problem with stochastic travel times
including soft time windows and service costs. Comput Oper Res
16. Wassan N, Wassan N et al (2016) The multiple trip vehicle routing problem with
backhauls: formulation and a two-level variable neighbourhood search. Comput
Oper Res
17. Xu J, Liu Y (2008) Multi-objective decision making model under fuzzy random
environment and its application to inventory problems. Inform Sci 178(14):2899–
2914
Hybrid Multiobjective Evolutionary Algorithm
with Differential Evolution for Process Planning
and Scheduling Problem
Chunxiao Wang1,2 , Wenqiang Zhang1,2(B) , Le Xiao1,2 , and Mitsuo Gen3

1
Henan University of Technology, Zhengzhou, People’s Republic of China
2
Key Laboratory of Grain Information Processing and Control,
Ministry of Education, Zhengzhou, People’s Republic of China
zhangwq@haut.edu.cn
3
Fuzzy Logic Systems Institute and Tokyo University of Science, Tokyo, Japan
Abstract. In an intelligent manufacturing environment, process plan-

ning and scheduling (PPS) plays a very important role as a most com-
plex and practical scheduling problem, which processes a set of pris-
matic parts into completed products by determining the optimal process
plans and moments to execute each operation with competitive manu-
facturing resources. Many research works use multiobjective evolution-
ary algorithm (MOEA) to solve PPS problems with consideration of
multiple complicated objectives to be optimized. This paper proposes
a hybrid multiobjective evolutionary algorithm with differential evolution
(HMOEA-DE) to solve the PPS problem. HMOEA-DE uses a special
designed fitness function to evaluate the dominated and nondominated
individuals and divides the population and elitism into two parts, which
close to the center and edge areas of Pareto frontier. Moreover, dif-
ferential evolution applied on elitism tries to improve the convergence
and distribution performances much more by guiding the search direc-
tions through different individuals with different fitness function. Numer-
ical comparisons indicate that the efficacy of HMOEA-DE outperforms
the traditional HMOEA without DE in convergence and distribution
performances.
Keywords: Process planning and scheduling · Multiobjective evolution-

ary algorithm · Differential evolution
1 Introduction
In an intelligent manufacturing system, a set of prismatic parts need to be
processed into a products effectively and economically according to various
resources constraints. The parts have operations with different features, which
to be related with the machines, tools and tool access directions (TADs). More-
over, the precedence relationship constraints among operations must be satis-
fied with the technological and the geometrical considerations. Process planning
Hybrid Multiobjective Evolutionary Algorithm 213
generates the optimal process plans, i.e. optimizes the operations (machine, tool,
tool access direction) and their sequences. The scheduling assignments the most
appropriate moment to execute each operation with competitive resources. The
process planning and scheduling (PPS) optimizes of the optimal process plan
and schedule simultaneously within the precedence relationship constraints and
manufacturing resources. However, it is very difficult to efficiently find an opti-
mal solutions from all of the combinations of all of the operations, manufacturing
resources satisfying one and more specified objectives.
Since the complex and multi-resource constraints as well as multi-objective
requirements of PPS problem, many researchers have addressed the evolutionary
algorithms (EAs) method to deal with it while satisfying the single and multiple
objectives [3,10,11,17,19]. Especially, multi-objective evolutionary algorithms
(MOEAs) are well suitable for solving multi-objective optimizations (MOPs)
[6,18]. In MOEAs, Vector evaluated genetic algorithm (VEGA) divides the pop-
ulation into several sub-populations according to the number of objectives, each
of which evolves toward a single objective [14]. As two classical MOEAs, NSGA-
II [5] and SPEA2 [20] have been proven to get better convergence and distri-
bution performances in solving MOPs. NSGA-II uses the Pareto ranking and
crowding distance mechanism to obtain better performances. SPEA2 proposes
raw fitness assignment mechanism and density mechanism to guarantee the con-
vergence and distribution performances. For solving the multi-objective PPS
problem, Zhang et al., propose a hybrid sampling strategy-based multi-objective
evolutionary algorithm (HSS-MOEA), which combines the sampling strategy of
VEGA and the sampling strategy based on a new Pareto dominating and domi-
nated relationship-based fitness function (PDDR-FF) [19]. The hybrid sampling
strategies preserve both the convergence and the distribution performance while
reduce the computational time.
As a real-parameter optimization method, Differential Evolution (DE) [16]
has been also treated as a population-based approach, in which mutation and
crossover are the variation operators used to generate new solutions. In DE,
the mutation operator has been executed based on the differences of individu-
als to guide the search direction in the current population and the replacement
mechanism uses to maintain the population. The effectiveness but simplicity of
DE has attracted many research interests in single objective and multi-objective
optimization [1,2,4,7–9,12,13,15]. Especially, Iorio and Li [8] proposed three DE
variants incorporating the directional information to improve the optimization
performance. The direction information among different individuals with same
or different rank as well as different crowding distances could guide the search
converge towards the Pareto frontier and/or prefers along the Pareto frontier
to improve the convergence and distribution performances. Iorio and Li [8] pro-
vided an interesting topic to combine the DE with other MOEAs to improve
the performance more especially for most complicated PPS problem, since only
applies the DE on PPS problem could increase the computation time for prac-
tical optimization problem.
214 C. Wang et al.
This paper proposes a hybrid multi-objective evolutionary algorithm with

differential evolution (HMOEA-DE) to solve the multi-objective PPS problem
that combines the DE with HSS-MOEA. The HSS-MOEA could generate the
new individuals with traditional crossover and mutation operators, and update
the elitism with special fitness function and strategy. The DE could enhance the
performance on elitist population after applying the HSS-MOEA.
The paper is organized as follows: Sect. 2 presents the PPS problem and
formulates the mathematical model; Sect. 3 describes the detailed HMOEA-DE
approach; Sect. 4 gives a discussion and analysis of numerical experiments results;
finally, the conclusion are drawn in Sect. 5.
2 Process Planning and Scheduling Problem

In PPS, there are many parts to be processed by a number of available machines
with different tools and different TADs. Each part has several operations, and
each operation can be processed on machines with different processing times.
The PPS problem finally generates a proper process plan and schedule, which
tells decision maker how, when and in which sequence to allocate suitable man-
ufacturing resources to operations effectively.
The operation sequence of PPS should be decided firstly, then, the manu-
facturing resources should be assigned to each operation in operation sequence.
After generating the process plan, the PPS determine the most appropriate
moment to process each operation with competitive resources. The precedence
relationship and resource constraints must be satisfied in deciding the process
plan and schedule. Moreover, the process plan and schedule should also satisfy
the multiple objectives requirements to maintain the feasibility.
Two important and frequently-used objectives are minimizing makespan and
variation of workload for each machine.
The mathematical model of the problem is expressed in the following nota-
tions.
Indices:
i, k: indices of part, (i, k = 1, 2, · · · , I).

j, h: indices of operation for part, (j, h = 1, 2, ..., Ji ).
m: index of machine, (m = 1, 2, · · · , M );
l: index of tool, (l = 1, 2, · · · , L).
d: index of TAD, (d = 1, 2, · · · , D).
Parameters:
I: number of parts;
Ji : number of operations for part i;
M : number of machines;
L: number of tools;
D: number of TADs;
O: set of operations for part i, Oi = {oij |j = 1, 2, · · · , Ji };
oij : the jth operation of part;

mm : the mth machine;
tl : the lth tool;
ad : the dth TAD;
Mij : set of machines that can process oij ;
Am : set of operations that can be processed on machine m; I
S: operation sequence. S = (s1 , s2 , · · · , sw , · · · , sW ), sw = oij , W = i=1 Ji ;
rijh : precedence constraints. If oij is predecessor of oih , rijh = 1; otherwise, 0;
tM i
mij : machining time of o j by machine m;
tMCI : machine change time index. It is the same for each machine change.
tMC
mij : machine change time of oij by machine m. A machine change is described
in article by Zhang et al. [19]. Once machine change is occurred, tMC mij = t
MCI
;
MC
otherwise, tmij = 0;
tTCI : tool change time index. It is the same for each tool change;
tTC
mij : tool change time of oij by machine m. A tool change is described in arti-
cle by Zhang et al. [19] If tool change is occurred, tTC mij = t
TCI
; otherwise,
TC
tmij = 0;
tSCI : set-up change time index. It is the same for each set-up change;
tSC
mij : set-up change time of oij by machine m. A set-up change is described in
article by Zhang et al. [19] If set-up change is occurred, tSC mij = t
SCI
; otherwise,
SC
tmij = 0;
tPRE
mij : preparation time of operation oij by machine m. The preparation time for
an operation consists of machine change time, tool change time and set-up
time for the operation;
tP
mij : processing time of operation oij by machine m.
tPRE MC TC SC
mij = tmij + tmij + tmij (1)
The processing time for an operation consists of the preparation time and
the machining time for the operation.
tP PRE M
mij = tmij + tmij , (2)
where um is workload of machine m.

I Ji
um = tM M
mij xmij , (3)
i=1 j=1
where ū: average workload of machine.

M
ū = 1/M um , (4)
m=1
where tC mij : completion time of oi j by machine m, it should satisfy the inequality

tm i(j−1) + tPmij ≤ tmij that means for every operation, its direct predecessor’s
C C
completion time plus its processing time might be smaller that its completion
time.
216 C. Wang et al.
Decisionvariables:
1, if oij is performed by machinem,
xMmij = 0, otherwise;

1, if oij is performed by tool l,
xTlij =
otherwise;
0,
1, if oij is performed by TAD d,
xDdij = 0, otherwise;

1, if oij is performed directly before okh ,
yijkh =
0, otherwise;
1, if X = Y,
Ω (X, Y ) =
0, otherwise.
The mathematical model can be formulated as the following bi-criteria non-
linear mixed integer programming (NMIP) model:
min tM = max{tC mij } (5)

m,i,j

M 2
min wP = 1/M (um − ū) (6)
m=1
M M
s. t. tCmkh − tmkh − tmij xmij xmkh yijkh ≥ 0, ∀ (i, j) , (k, h) , m
P C
(7)
rijh yihij = 0, ∀(i, j), h (8)
yijij = 0, ∀(i, j) (9)
M
mij = 1, ∀(i, j)
xM (10)
m=1
mij = 0, ∀ (i, j) ∈
xM / Am , ∀m (11)
yijkh ∈ {0, 1} , ∀ (i, j) , (k, h) (12)
mij ∈ {0, 1} , ∀m, (i, j)
xM (13)
tmij ≥ 0, ∀m, (i, j).
C
(14)
Equation illustrates the minimization of makespan. Makespan is the maxi-

mization of completion time among all the operations. Minimization of workload
variance is defined as Equation. Workload variance is defined as standard devi-
ation of workload of all the machines. The equations to impose the precedence
relationship constraints, resource constraints and nonnegative condition.
3 Hybrid Multi-Objective Evolutionary Algorithm

with Differential Evolution
The solution procedure of HMOEA-DE includes 5 phases.

Phase 1: Selection of Better Individuals with Objective Values
The sampling strategy of VEGA is used to select the better individuals from
population P (t) into part of mating pool (sub populations) according to their
objectives value.
Phase 2: Generation of Mating Pool with Hybrid Sampling
The sub populations and elitist population A(t) are combined to form a
mating pool. In the mating pool, sub-pop-1 saves the good individuals for one
objective, and sub-population 2 stores the good individuals for the other objec-
tive. The elitist population (archive) holds the individuals with good PDDR-FF
values. Therefore, one-third of the individuals serve one objective, one-third the
other objective, and the left one-third both the two objectives in the mating
pool.
Phase 3: Reproduction with Traditional Crossover and Mutation Operators
Problem-dependent crossover and mutation operators are used to reproduce
new individuals to form new population P(t+1). Moreover, a local search mech-
anism is proposed to improve the quality of individuals after the reproduction
process. The details of reproduction and local search is described in article by
Zhang et al. [19].
Phase 4: Archive Maintenance by PDDR-FF based Elitist Sampling Strategy
The individuals in A(t) and P (t) are combined to form a temporary archive
A (t). Thereafter, the PDDR-FF values of all individuals in A (t) are calculated
and sorted. If the individual is nondominated one, its fitness value will not exceed
one. The fitness value of dominated individual will exceed one. Moreover, even
the all nondominated individuals have also different PDDR-FF values. The non-
dominated individuals will be smaller values (near to 0) than the edge points
(near to 1). After calculating the fitness function value, the smallest |A(t)| indi-
viduals in A (t) are copied to form A (t + 1).
Phase 5: Archive Enhancement by Combining DE
In this paper, the DE operators such as mutation, crossover and selection
operators only apply on the temporary archive A’(t+1), which like a local search
process to improve the performance of archive. After DE operator, the obtained
solution set (temporary archive) A (t+1) and A (t+1) are combined to generate
the last A(t + 1) according to the PDDR-FF values.
The DE/current/2 is used to generate new mutants and each individuals
xi in A (t + 1) and three individuals r1 , r2 , r3 are randomly picked up to do
reproduction operators according to the following mutation operator.
vi = xi + K(r3 − xi ) + F (r1 − r2 ), (15)
where xi = r1 = r2 = r3 , eval(r1 )< eval(r2 ) and eval(r3 )<eval(xi ), K and F are

scale factors and 0 ≤ K, F ≤ 1. eval(xi ) means the fitness function value of xi .
It is important to note that in PDDR-FF, the smaller value mean better and the
fitness function of nondominated individual will not exceed one. The fitness value
of dominated individual will exceed one. Moreover, the nondominated individuals
locating around the central region of Pareto frontier have smaller values (near
to 0) than the edge points (near to 1).
The mutation operator of DE will improve the performances as follows
schemes according to the different fitness function of current and three random
individuals.
218 C. Wang et al.
Scheme 1: the current individual xi is nondominated solution (eval(si ) ≤ 1).

(1) If the eval(r1 ) ≤ 1 and eval(r2 ) ≤ 1, r1 , r2 are nondominated solutions. This
mutation means the search direction starts from the nondominated solution
(xi ) and firstly follows the direction from edge region to center region(xi
locates in edge region and r3 locates in center region) along the Pareto
frontier, next, the search direction will guide from edge region to center
region (r2 locates in edge region and r1 locates in center region) again. It
will improve the distribution performance.
(2) If the eval(r1 ) > 1 and eval(r2 ) > 1, r1 , r2 are dominated solutions. The
search direction starts from the nondominated solution (xi ) and firstly fol-
lows the direction from edge region to center region(xi locates in edge region
and r3 locates in center region) along the Pareto frontier, next, the search
direction will guide from edge region to center region (r2 locates in edge
region and r1 locates in center region) again. It will improve the distribu-
tion performance.
(3) If the eval(r1 ) ≤ 1 and eval(r2 ) > 1, r1 is nondominated solutions and r2
is dominated solutions. The search direction starts from the nondominated
solution (xi ) and firstly follows the direction from edge region to center
region(xi locates in edge region and r3 locates in center region) along the
Pareto frontier to improve the distribution performance, next, the search
direction will guide from dominated region to nondominated region (r2
locates in the dominated region and r1 locates in the nondominated region)
towards the Pareto frontier to improve the convergence performance.
Scheme 2: the current individual xi is dominated solution (eval(si ) > 1) and
r3 is nondominated solution (eval(r3 ) < 1).
(1) If the eval(r1 ) ≤ 1 and eval(r2 ) ≤ 1, r1 , r2 are nondominated solutions.
This mutation means the search direction starts from the dominated solu-
tion (xi ) and firstly follows the direction from dominated region to non-
dominated region (xi locates in the dominated region and r3 locates in the
nondominated region) along the Pareto frontier to improve the convergence
performance, next, the search direction will guide from edge region to center
region (r2 locates in edge region and r1 locates in center region) to improve
the distribution performance.
search direction starts from the dominated solution (xi ) and firstly follows
the direction from dominated region to nondominated region (xi locates in
the dominated region and r3 locates in the nondominated region) along the
Pareto frontier to improve the convergence performance, next, the search
region and r1 locates in center region) to improve the distribution perfor-
mance.
(3) If the eval(r1 ) ≤ 1 and eval(r2 ) > 1, r1 is nondominated solutions and
r2 is dominated solutions. The search direction starts from the dominated
solution (xi ) and firstly follows the direction from dominated region to non-
dominated region (xi locates in the dominated region and r3 locates in the
nondominated region) along the Pareto frontier to improve the convergence

performance, next, the search direction will guide from dominated region
to nondominated region (r2 locates in the dominated region and r1 locates
in the nondominated region) towards the Pareto frontier to improve the
convergence performance more.
Scheme 3: the current individual xi is dominated solution (eval(si ) > 1) and
r3 also is dominated solution (eval(r3 ) > 1).
(1) If the eval(r1 ) ≤ 1 and eval(r2 ) ≤ 1, r2 are nondominated solutions. This
mutation means the search direction starts from the dominated solution (xi )
and firstly follows the direction from edge region to center region(xi locates
in edge region and r3 locates in center region) along the Pareto frontier,
next, the search direction will guide from edge region to center region (r2
locates in edge region and r1 locates in center region) again. It will improve
the distribution performance.
search direction starts from the dominated solution (xi ) and firstly follows
the direction from edge region to center region(xi locates in edge region
and r3 locates in center region) along the Pareto frontier, next, the search
region and r1 locates in center region) again. It will improve the distribution
performance.
(3) If the eval(r1 ) ≤ 1 and eval(r2 ) > 1, r1 is nondominated solutions and r2 is
dominated solutions. The search direction starts from the dominated solu-
tion (xi ) and firstly follows the direction from edge region to center region(xi
locates in edge region and r3 locates in center region) along the Pareto fron-
tier to improve the distribution performance, next, the search direction will
guide from dominated region to nondominated region (r2 locates in the dom-
inated region and r1 locates in the nondominated region) towards the Pareto
frontier to improve the convergence performance.
From above analysis, the HSS-MOEA could improve the convergence and dis-
tribution performance, the combination of HMOEA and DE could much improve
the performances more.

In this paper, 4 parts (each part having 20, 16, 14 and 7 operations, respectively)
problem are used. All the simulations are run on Core i5 processor (2.6 GHz
clock). The adopted parameters are listed as follows: population size = 100;
maximum generation = 500; archive size = 50; crossover probability = 0.70;
mutation probability = 0.30 and the scale factor K and F are assigned randomly
within 0 to 1. HMOEA-DE and HSS-MOEA are run 30 times to compare the
results.
Let Sj be a solution set for each method (j = 1, 2). P F ∗ is a known refer-
ence Pareto set which comes from combining all of the obtained Pareto set by
220 C. Wang et al.
2 methods with 30 runs. In this study, Coverage C(S1 , S2 ), Average distance

AD(Sj ), Hyper volume HV (Sj ) and Spacing SP (Sj ) are selected to compare
the performances of algorithms.
Figure 1 shows the numerical comparison of the box-and-whisker plots for C,
AD, HV , and SP by HMOEA-DE, and HSS-MOEA. From Fig. 1, it is easy to
see that the HMOEA-DE is better than HSS-MOEA in terms of C measure. The
AD measure also indicates that HMOEA-DE can get smaller value than HSS-
MOEA. For HV measure, HMOEA-DE is better than HSS-MOEA. The distrib-
ution performances, SP , indicates that HMOEA-DE is better than HSS-MOEA.
The PDDR-FF could give different values among nondominated and dominated
solution in elitism, and the combination of DE for individuals in elitism could
significantly improve the convergence and distribution performances.
Coverage for HMOEA-DE and HSS-MOEA Average Distance

1
150
0.8
Average Distance
0.6 100
Coverage
0.4
50
0.2
0
0
C(HMOEA-DE,HSS-MOEA) C(HSS-MOEA,HMOEA-DE) HMOEA-DE HSS-MOEA
4
x 10 Hypervolume Spacing
4
3.5 150
Hypervolume
3
Spacing
100
2.5
2
50
1.5
1
0
HMOEA-DE HSS-MOEA HMOEA-DE HSS-MOEA
Fig. 1. C, AD, HV, and SP by HMOEA-DE, and HSS-MOEA
5 Conclusions
In this study, a HMOEA-DE approach was proposed to solve the multi-objective

PPS problem with consideration of minimization of makespan and minimization
of variation of machine workload simultaneously. The VEGA take a preference
for the edge region of the Pareto front and the PDDR-FF-based elitism maintain
mechanism has the tendency converging toward the center area of the Pareto
front. Moreover, the hybrid DE operators for the elitism has the capabilities of
improving the convergence and distribution performance much more. Numeri-
cal comparisons result demonstrated that HMOEA-DE outperformed the HSS-
MOEA on convergence and distribution performance.
Acknowledgements. This research work is supported by the National Natural Sci-

ence Foundation of China (U1304609), Foundation for Science & Technology Research
Project of Henan Province (162102210044, 152102210068 152102110076), Program for
Innovative Research Team (in Science and Technology) in University of Henan Province
(17IRTSTHN011), Program for Key Project of Science and Technology in University of
Education Department of Henan Province (17A520030), Fundamental Research Funds
for the Henan Provincial Colleges and Universities (2014YWQQ12, 2015XTCX03,
2015XTCX04), Research Funds for Key Laboratory of Grain Information Processing
and Control (Henan University of Technology) (KFJJ-2015-106), Ministry of Educa-
tion and the Grant-in-Aid for Scientific Research (C) of Japan Society of Promotion
of Science (JSPS): No. 15K00357.
References
1. Abbass HA, Sarker R (2011) The pareto differential evolution algorithm. Int J
Artif Intell Tools 11(04):531–552
2. Ali M, Siarry P, Pant M (2012) An efficient differential evolution based algorithm
for solving multi-objective optimization problems. Eur J Oper Res 217(2):404–416
3. Ausaf MF, Gao L, Li X (2015) Optimization of multi-objective integrated process
planning and scheduling problem using a priority based optimization algorithm.
Front Mech Eng 10(4):1–13
4. Babu BV, Jehan MML (2003) Differential evolution for multi-objective optimiza-
tion. In: The 2003 Congress on Evolutionary Computation, 2003, CEC 2003, vol
4, pp 2696–2703
5. Deb K, Pratap A et al (2002) A fast and elitist multiobjective genetic algorithm:
NSGA-II. IEEE Trans Evol Comput 6(2):182–197
6. Gen M, Cheng R, Lin L (2008) Network models and optimization: Multiobjective
7. Guo W, Yu X (2014) Non-dominated sorting differential evolution with improved
directional convergence and spread for multiobjective optimization. In: Proceedings
of the Companion Publication of the 2014 Annual Conference on Genetic and
Evolutionary Computation, pp 87–88
8. Iorio AW, Li X (2006) Incorporating directional information within a differential
evolution algorithm for multi-objective optimization. In: Conference on Genetic
and Evolutionary Computation, pp 691–698
9. Li H, Zhang Q (2006) A multiobjective differential evolution based on decompo-
sition for multiobjective optimization with variable linkages. In: Parallel Problem
Solving From Nature - PPSN Ix, International Conference, Reykjavik, Iceland,
9–13 September 2006, Proceedings, pp 583–592
10. Nayak A (2015) Multi-objective process planning and scheduling using controlled
elitist non-dominated sorting genetic algorithm. Int J Prod Res 53(6):1712–1735
11. Phanden RK, Jain A, Verma R (2013) An approach for integration of process
planning and scheduling. Int J Comput Integr Manufact 26(4):284–302
12. Qu BY, Suganthan PN, Liang JJ (2012) Differential evolution with neighborhood
mutation for multimodal optimization. IEEE Trans Evol Comput 16(5):601–614
13. Santana-Quintero LV, Coello Coello CA (2005) An algorithm based on differential
evolution for multi-objective problems. Int J Comput Intell Res 1(2):151–169
algorithms. In: International Conference on Genetic Algorithms, pp 93–100
222 C. Wang et al.
15. Sindhya K, Ruuska S et al (2011) A new hybrid mutation operator for multiobjec-
tive optimization with differential evolution. Soft Comput 15(10):2041–2055
16. Storn R, Price K (1997) Differential evolution – a simple and efficient heuristic for
global optimization over continuous spaces. J Glob Optim 11(4):341–359
17. Xia H, Li X, Gao L (2016) A hybrid genetic algorithm with variable neighborhood
search for dynamic integrated process planning and scheduling. Comput Ind Eng
102:99–112
18. Yu X (2010) Introduction to evolutionary algorithms. In: International Conference
on Computers and Industrial Engineering, p 1
19. Zhang W, Gen M, Jo J (2014) Hybrid sampling strategy-based multiobjective evo-
lutionary algorithm for process planning and scheduling problem. J Intell Manufact
25(5):881–897
20. Zitzler E, Laumanns M, Thiele L (2001) SPEA2: Improving the strength pareto
evolutionary algorithm
A New Approach for Solving Optimal Control
Problem by Using Orthogonal Function
Akram Kheirabadi1(B) , Asadollah Mahmoudzadeh Vaziri1 , and Sohrab Effati2

1
Department of Mathematics, Faculty of Mathematical Science and Statistics,
University of Birjand, Birjand, Iran
ak kheirabadi@yahoo.com
2
Department of Applied Mathematics, Faculty of Mathematical Science and
Statistics, Ferdowsi University of Mashhad, Mashhad, Iran
Abstract. In the present paper we introduce a numerical technique for

solving fractional optimal control problems (FOCP) based on an ortho-
normal wavelet. First we approximate the involved functions by Sine-
Cosine wavelet basis; then, an operational matrix is used to transfer
the given problem in to a linear system of algebraic equations. In fact
operational matrix of the Riemann-Liouville fractional integration and
derivative of Sine-Cosine wavelet are employed to achieve a linear alge-
braic equation, in place of the dynamical system in terms of the unknown
coefficients. The solution of this system, gives us the solution of original
problem. A numerical example is also given.
Keywords: Fractional optimal control problem · Sine-Cosine wavelet ·

Operational matrix · Caputo derivative · Riemann-Liouville fractional
integration
1 Introduction
Many application of the fractional calculus is in basic sciences and engineering.

Many realistic model of physical [8] phenomenon which has dependence at both
the time instance and on the previous time history, can be utter with fractional
calculus. For example it can be applied in nonlinear oscillations of earthquakes,
fluid-dynamic traffic [9], frequency dependent damping behavior of various vis-
coelastic materials [2], solid mechanics [18], economics [3], signal processing [17],
and control theory [4].
One of the main difficulties is how to solve the fractional differential equa-
tions. The most commonly techniques proposed to solve them are Adomian
decomposition method (ADM) [22], Variational Iteration Method (VIM) [20],
Operational Matrix Method [19], Homotopy Analysis Method [6,7], Fractional
Difference Method (FDM) [15] and Power Series Method [16].
A fractional optimal control problem is an optimal control problem in which
the performance index or the differential equations governing the dynamic of the
224 A. Kheirabadi et al.
system or both contains at least one fractional order derivative term [25]. Inte-
ger order optimal controls have already been well established and a significant
amount of works have been done in the field of optimal control of integer order
systems. Agrawal formulated and developed a numerical scheme for the solution
of FOCP [1] in the Caputo sense. Biswas proposed a pseudo-state space represen-
tation of a fractional dynamical system, which is exploited to solve a fractional
optimal control problem using a direct numerical method [21]. Sweilam et al.
solved some types of fractional optimal control problem with a Hamiltonian for-
mula using a spectral method based on Chebyshev polynomials [24]. Bernstein
polynomials have been used for finding the numerical solution of FOCP by using
Lagrange multipliers [10].
Approximation by orthogonal families of basis functions is widely used in
science and engineering. The main idea behind applying an orthogonal basis is
reduction of the problem under consideration into a system of algebraic equa-
tions. This is possible by truncating series of orthogonal basis functions for the
solution of the problem and applying operational matrices. The orthogonal func-
tions are classified into three main category [23]: the first one is sets of piecewise
constant orthogonal functions such as the Walsh functions and block pulse func-
tions. The second one is orthogonal polynomials such as the Laguerre, Legendre
and Chebyshev functions, and the last one is sine-cosine functions. In one hand
approximating a continuous function with piecewise constant basis functions
results in a piecewise constant approximation, on the other hand, if a discon-
tinuous function is approximated with continuous basis functions, the resulting
approximation is continuous which cannot properly model the discontinuities.
So, neither continuous basis functions nor piecewise constant basis functions, if
used alone, can efficiently model both continuity and discontinuity of phenom-
ena at the same time. In the case that the function under approximation is not
analytic, wavelet functions will be more effective.
In this paper, we propose a computational method based on Sine-Cosine
wavelet with their fractional integration and derivative operational matrix to
solve the FOCP. The main idea is reduction the problem under consideration into
a system of algebraic equations. To this end, we expand the fractional derivative
of the state variable and the control variable using the Sine-Cosine wavelet with
unknown coefficients.
The paper is organized as follows. In first section we will give the definitions
of fractional calculus, then express a brief review of block pulse function and
the related fractional operational matrices. In Sect. 4, we describe Sine-Cosine
wavelets and its application in function approximation. In Sect. 5, operational
matrices of fractional integration and derivative for considered wavelet is given.
In Sect. 6, the proposed method is described for solving the underlying FOCP. In
the last section the proposed method is applied for solving numerical example.
A New Approach for Solving Optimal Control Problem 225
2 Preliminaries of Fractional Calculus

The Riemann-Liouville fractional integration and Caputo differential operator
of a function f of order α ≥ 0 is defined in [13] as:
t
α Γ
1
(α) 0
(t − τ )α−1 f (τ )dτ α > 0
(I f )(t) = (1)
f (t) α = 0,
1 t
Dα f (t) = (t − τ )n−α−1 f (n) (τ )dτ
Γ (n − α) 0 (2)
= I n−α f n (t) n − 1 < α ≤ n.
3 Review of Block Pulse Functions and the Related

Fractional Operational Matrix
In this section first we introduce block pulse function (BPF), then it’s operational
matrix of fractional integration.
3.1 Definition of BPF

A set of BPFs Bm (t) containing m component functions in the interval [0, T)
is given by
Bm (t) [b0 (t)b1 (t) · · · bi (t) · · · bm −1 ]T . (3)
The ith component of the BPF vector Bm (t) is defined as
iT (i+1)T
bi (t) =
1 m ≤ t < m i = 0, 1, 2, · · · , m − 1. (4)
0 O.W.
A square integrable function f can be expanded into a BPF series as
f (t) = [c0 c1 · · · ci · · · cm −1 ]Bm (t) = C T Bm (t), (5)

1 (i+1)h T
ci = f (t)dt h = . (6)
h ih m
3.2 Operational Matrix for Fractional Integration of BPF

Suppose that F α be the block pulse operational matrix of fractional integration
[12]. It is defined as follows,
⎡ ⎤
1 ξ1 ξ2 · · · ξm −1
⎢ 0 1 ξ1 · · · ξm −2 ⎥
⎢ ⎥
⎢ 0 0 1 · · · ξm −3 ⎥
α 1 ⎢ ⎥
Fα = h ⎢. . . . ⎥, (7)
Γ (α + 2) ⎢ .. .. .. . . . .. ⎥
⎢ ⎥
⎣ 0 0 0 0 ξ1 ⎦
00 0 0 1
ξk = (k + 1)α+1 − 2k α+1 + (k − 1)α+1 k = 1, 2, · · · , m − 1. (8)
4 Description of Sine-Cosine Wavelets and Its

Application in Function Approximation
4.1 The Sine-Cosine Wavelet
Sine-cosine wavelets ψn,m (t) are defined as follows [11],

k+1
n n+1
2 2 f (2k t − n) 2k
≤t≤ 2k
ψn,m (t) = (9)
0 o.w
with
⎧ 1
⎨ √2 m=0
fm (t) = cos(2mπt) m = 1, 2, · · · , l (10)
⎩
sin(2(m − l)πt) m = l + 1, · · · , 2l
n = 0, 1, · · · , 2k − 1, k = 0, 1, · · · , where l is any positive integer.
4.2 Function Approximation
A function f (t) ∈ L2 [0, 1) can be approximated as:

k
2 −1
2l
f (t) = cn,m ψn,m = C T Ψ (t) = Ψ T (t)C, (11)
n=0 m=0
where cn,m = f (t), ψn,m and ., . denotes the inner product as:
+∞
cn,m = f (t)ψn,m (t)dt. (12)
−∞
where Ψ (t) represent considered wavelet. C and Ψ (t) are 2k (2l + 1) × 1 matrices
which are given by:
C T = [c00 c01 · · · c0,2l , c10 , · · · , c1,2l , · · · , c2k −1,0 , · · · , c2k −1,2l ], (13)
T
Ψ = [ψ00 ψ01 · · · ψ0,2l , ψ10 , · · · , ψ1,2l , · · · , ψ2k −1,0 , · · · , ψ2k −1,2l ]. (14)
5 Operational Matrix of Fractional Calculus

for Sine-Cosine Wavelet
In this section we find the operational matrix of fractional derivative for the con-
sidered wavelet using the operational matrix of fractional integration for BPF.
5.1 Express Ψ (t) in Terms of BPF

ψn,m (t) as a function can be express in terms of blockpulse function

m
−1
ψn,m fi bi m = 2k (2l + 1), (15)
i=0
i+1 i+1
m m k+1
fi = m ψn,m (x)dx = m 2 2 fm (2k x − n)dx. (16)
i i
m m
Now we calculate fi for different value of i = 0, 1, · · · , m − 1

i+1
m k+1 1 k

m = 0, fi = m 2 2 × √ dx = 2 2
i
m
2 (17)
i = n(2l + 1), · · · , (n + 1)(2l + 1) − 1,
i+1
m k+1
m = 1, 2, · · · , l, fi = m 2 2 cos(2mπ(2k x − n))dx
i
m (18)

m i+1 i
= ψn,m+l − ψn,m+l ,
2
k+1
2 mπ m m
i+1
m k+1

m = l + 1, · · · , 2l, fi = m 2 2 sin(2(m − l)π(2k x − n))dx
i
m

−m i+1 i
= ψn,m−l − ψn,m−l ( ) .
2
k+1
2 (m − l)π m m
(19)
For m = 0 we have
ψn,m = [0, · · · , 0, 2k/2 , 2k/2 , · · · , 2k/2 , 0, · · · , 0] × Bm . (20)

n(2l+1) 2l+1
For m = 1, 2, · · · , l
⎡
m ⎢ n(2l + 1) + 1 n(2l + 1)
ψn,m = ⎣0, 0, · · · , 0, ψn,m+l −ψn,m+l , (21)
2
k+1
2 mπ m m
n(2l+1)
(n + 1)(2l + 1) n(2l + 1) + 2l
· · · , ψn,m+l − ψn,m+l , 0, 0, · · · , 0 × Bm .
m m
And for m = l + 1, · · · , 2l we get
⎡

−m ⎢ n(2l + 1) + 1 n(2l + 1)
ψn,m = ⎣0, 0, · · · , 0, ψn,m−l − ψn,m−l , (22)
(m−l)π m m
k+1
2 2
n(2l+1)
(n + 1)(2l + 1) n(2l + 1) + 2l
· · · , ψn,m−l − ψn,m−l , 0, 0, · · · , 0 × Bm .
m m
Therefore we have Ψ (x) = Φm ×m Bm (x) where Φm ×m = diag(Φ0 , Φ1 , · · · ,
Φ2k −1 ), Φn is defined as follows, in the following matrix, i = n(2l + 1)
⎡ k k
⎤
22 ··· 22
⎢ m i+2l ⎥
⎢ k+1 (ψn,1+l i+1 i i+2l+1
m − ψn,1+l ( m ) · · · ψn,1+l ( m ) − ψn,1+l ( m )) ⎥
⎢ 2 2 1π ⎥
⎢. . . .. ⎥
⎢ .. .. ⎥
⎢ i+1 i ⎥
⎢ m ⎥
Φn = ⎢ k+1 (ψn,2l m − ψn,21 m · · · ψn,2l ( i+2l+1
m ) − ψn,21 ( i+2l
m ))) ⎥.
⎢ 2 2 lπ ⎥
⎢ −m (ψ i+1 − ψ ( i ) i+2l+1 i+2l
· · · ψn,1 ( m ) − ψn,1 ( m )) ⎥
⎢ k+1 n,1 m n,1 m ⎥
⎢ 2 2 1π ⎥
⎢ .. .. . . ⎥
⎢. . . ⎥
⎣ i+1 ⎦
−m i i+2l+1 i+2l
k+1 (ψn,l m − ψn,l ( m ) · · · ψn,l ( m ) − ψn,l ( m ))
2 2 lπ
(23)
5.2 Operational Matrix of Fractional Integration and Derivative for

Sine-Cosine Wavelet
For finding operational matrix of fractional derivative of vector Ψ (t), first of all
we try to find the operational matrix of fractional integration.
(I α Ψ )(x) P α Ψ (x), (24)
where P α is the operational matrix of fractional integration, which calculate as

follows
I α Ψ (x) = I α Φm ×m .Bm (x) = Φm ×m I α Bm (t) = Φm ×m F α Bm (x) (25)
α α α
⇒ P Ψ (x) = P Φ m ×m B
m (x) = Φm ×m F B m (x)
⇒ P = Φm ×m F Φ−1
α
m ×m
α
⇒ I Ψ (x) Φm ×m F α Φ−1

α
m ×m Ψ (x). (26)
Now we calculate operational matrix of derivative using P α
Dα f (x) = I n−α f n (x)n − 1 < α ≤ nn ∈ N, (27)

α α T T α T n−α (n)
D x(t) = D X Ψ (t) = X D Ψ (t) = X I Ψ (t). (28)
For α ∈(0,1) we have n = 1 thus
Dα x(t) X T I 1−α DΨ (t) = X T DI 1−α Ψ (t)

(29)
= X T DΦm ×m F 1−α Φ−1
m ×m Ψ (t),
where D is operational matrix of derivative for Ψ (t) which defined as D =

diag(w, w, · · · , w), which is 2k (2l + 1) × 2k (2l + 1) matrix and w is of size
(2l + 1) × (2l + 1)
⎡ ⎤
0 0 0 ··· 0 0 0 ··· 0
⎢ 0 0 0 · · · 0 −1 0 · · · 0 ⎥
⎢ ⎥
⎢ 0 0 0 · · · 0 0 −2 · · · 0 ⎥
⎢ ⎥
⎢ .. .. .. . . .. .. .. . . .. ⎥
⎢. . . . . . . . . ⎥
k+1 ⎢
⎢ ⎥
w=2 π ⎢ 0 0 0 · · · 0 0 0 · · · −l ⎥ ⎥ . (30)
⎢0 1 0 ··· 0 0 0 ··· 0 ⎥
⎢ ⎥
⎢0 0 2 ··· 0 0 0 ··· 0 ⎥
⎢ ⎥
⎢. . . . . . . . . ⎥
⎣. . . . . . .
. . . . . . . . .. ⎦
.
0 0 0 ··· l 0 0 ··· 0 (2l+1)×(2l+1)
6 Solution of the Fractional Optimal Control Problem

by Sine-Cosine Operational Matrix
Consider the fractional optimal control problem with quadratic performance
index

1 T 1 1 T
min J = X (1)SX(1) + (X (t)QX(t) + U T (t)RU (t))dt (31)
2 2 0
s. t. Dα X(t) = AX(t) + BU (t) (32)
X(0) = X0 0 < α ≤ 1, (33)
where A and B are constant matrices with the appropriate dimensions, also in
cost functional S and Q are symmetric positive semi-definite matrices and R
is a symmetric positive definite matrix. In this section, the Sine-Cosine wavelet
is used for solving the above problem. We approximate each xi (t) and ui (t), in
terms of Sine-Cosine wavelets as
X(t) = [x1 (t), x2 (t), · · · , xs (t)]T xi (t) = Ψ T (t)Xi or XiT Ψ (t), (34)
X(t) = Ψ̂sT (t)X X = [X1T , X2T , · · · , XsT ] Ψ̂s (t) = Is ⊗ Ψ (t), (35)
U (t) = [u1 (t), u2 (t), · · · , uq (t)]T ui (t) = T
Ψ (t)Ui or UiT Ψ (t), (36)
U (t) = Ψ̂qT (t)U U= [U1T , U2T , · · · , UqT ] Ψ̂s (t) = Is ⊗ Ψ (t), (37)
where Xi , Ui are vectors of order 2k (2l + 1) × 1, X and U are vectors of order

s2k (2l + 1) × 1 and q2k (2l + 1) × 1 respectively. ⊗ denotes the kronecker product.
By substituting the above mentioned relation into objective function

1 1 1 T
J = X T Ψ̂s (1)S Ψ̂sT (1)X + [X Ψ̂s QΨ̂sT X + U T Ψ̂q RΨ̂qT U ]dt. (38)
2 2 0
1
Since considered wavelet is orthonormal, it means 0 Ψ T (t)Ψ (t)dt = I, we
can rewrite Eq. (38) as follows
1 T 1
J(X, U ) = X [S ⊗ Ψ̂ (1)Ψ̂ T (1)]X + [X T (Q ⊗ I)X + U T (R ⊗ I)U ]. (39)
2 2
Similarly, we do the same method for Eq. (32)
X(t) = X T Is ⊗ Ψ (t) or (Is ⊗ Ψ T (t))X, (40)

α T T 1−α
D X(t) = I 1−α
X (t) = I 1−α
(X (Is ⊗ Ψ (t))) = X I (Is ⊗ (DΨ (t)))
T T
= X Is ⊗ (I 1−α
DΨ (t)) = X Is ⊗ [DI 1−α
(Ψ (t))] (41)
T
= X Is ⊗ (DΦm ×m F α
Φ−1
m ×m Ψ (x))
R(t) = X T Is ⊗(DΦm ×m F α Φ−1 T T

m ×m )Ψ (t)−AX Is ⊗Ψ (t)−BU Iq ⊗ Ψ (t)R(t)
= [X T Is ⊗ (DΦm ×m F α Φ−1 T
m ×m ) − AX Is ⊗ I2k (2l+1) (42)
T
− BU Iq ⊗ I2k (2l+1) ] ⊗ Ψ (t).
As in a typical tau method [5] we generate 2k (2l + 1) − 1 linear equations by

applying
1
R(t), ψn,m (t) = R(t).ψn,m (t)dt = 0. (43)
0
Also, by substituting Eq. (35) in (33) we get
X(0) = X T Ψ̂ (0) = X0 . (44)
Equations (43) and (44) generate 2k (2l + 1) set of linear equations. These
linear equations can be solved for unknown coefficients of the vectors X T and
U T . Consequently, X(t) and U (t) can be calculated.
7 Illustrative Example
We applied the method presented in this paper and solved the undergoing
example.
Example 1. Consider the following time invariant FOCP [14],

1 1 2
min J = (x (t) + u2 (t))dt
2 0
s. t. Dα x(t) = −x(t) + u(t)
x(0) = 1.
We want to find a control variable u(t) which minimizes the quadratic perfor-
mance index J. This problem is solved by proposed method with α = 1, m = 5
and n = 7, the numerical value obtained for J is 0.1979, which is close to the
exact solutions in the case α = 1(0.1929).
8 Conclusion
In this paper, we derive a numerical method for fractional optimal control based
on the operational matrix for the fractional integration and differentiation. The
procedure of constructing these matrices is summarized. An example is given
to show the efficiency of method. The obtained matrices can also be used to
solve problems such as fractional optimal control with delay. Moreover we could
find these matrices using another set of orthogonal functions instead of BPFs,
it seems if we use a set of continuous orthogonal function the numerical result
will improve.
References
1. Agrawal OP (2008) A formulation and numerical scheme for fractional optimal
control problems. J Vibr Control 14(9–10):1291–1299
2. Bagley RL, Torvik P (1983) A theoretical basis for the application of fractional
calculus to viscoelasticity. J Rheol 27(3):201–210
3. Baillie RT (1996) Long memory processes and fractional integration in economet-
rics. J Econometrics 73(1):5–59
4. Bohannan GW (2008) Analog fractional order controller in temperature and motor
control applications. J Vibr Control 14(9–10):1487–1498
5. Canuto C, Hussaini MY, Quarteroni A, Zang TA Jr (1988) Spectral methods in
fluid dynamics. Annu Rev Fluid Mech 57(196):339–367
6. Dehghan M, Manafian J, Saadatmandi A (2010) The solution of the linear frac-
tional partial differential equations using the homotopy analysis method. Z Natur-
forsch A 65(Z. Naturforsch):935–949
7. Dehghan M, Manafian J, Saadatmandi A (2010) Solving nonlinear fractional partial
differential equations using the homotopy analysis method. Numer Methods Partial
Differ Equ 26(2):448–479
8. He J (1998) Nonlinear oscillation with fractional derivative and its applications.
In: International conference on vibrating engineering, Dalian, China, vol 98, pp
288–291
9. He J (1999) Some applications of nonlinear fractional differential equations and
their approximations. Bull Sci Technol 15(2):86–90
10. Jafari H, Tajadodi H (2014) Fractional order optimal control problems via the
operational matrices of bernstein polynomials. Upb Sci Bull 76(3):115–128
11. Kajani MT, Ghasemi M, Babolian E (2006) Numerical solution of linear integro-
differential equation by using sine-cosine wavelets. Appl Math Comput 180(2):569–
574
12. Li Y, Sun N (2011) Numerical solution of fractional differential equations using the
generalized block pulse operational matrix. Comput Math Appl 62(3):1046–1054
13. Liu F, Agrawal OP et al (1999) Fractional differential equations: an introduction to
fractional derivatives, fractional differential equations, to methods of their solution
and some of their applications. Academic Press
14. Lotfi A, Dehghan M, Yousefi SA (2011) A numerical technique for solving fractional
optimal control problems. Comput Math Appl 62(3):1055–1067
15. Momani S, Odibat Z (2007) Numerical comparison of methods for solving linear
differential equations of fractional order. Chaos Solitons Fractals 31(5):1248–1255
16. Odibat ZM, Shawagfeh NT (2007) Generalized taylor’s formula. Appl Math Com-
put 186(1):286–293
17. Panda R, Dash M (2006) Fractional generalized splines and signal processing. Sig
Process 86(9):2340–2350
18. Rossikhin YA, Shitikova MV (1997) Applications of fractional calculus to dynamic
problems of linear and nonlinear hereditary mechanics of solids. Appl Mech Rev
50:15–67
19. Saadatmandi A, Dehghan M (2010) A new operational matrix for solving
fractional-order differential equations. Comput Math Appl 59(3):1326–1336
20. Saez D (2009) Analytical solution of a fractional diffusion equation by variational
iteration method. Comput Math Appl 57(3):483–487
21. Sen S (2011) Fractional optimal control problems: A pseudo-state space approach.
J Vibr Control 17(17):1034–1041
22. Shawagfeh NT (2002) Analytical approximate solutions for nonlinear fractional
differential equations. Appl Math Comput 131(2–3):517–529
23. Sohrabi S (2011) Comparison chebyshev wavelets method with bpfs method for
solving abel’s integral equation. Ain Shams Eng J 2(3–4):249–254
24. Sweilam NH, Alajami TM, Hoppe RHW (2013) Numerical solution of some types
of fractional optimal control problems. Sci World J 2013(2):306237
25. Tangpong XW, Agrawal OP (2009) Fractional optimal control of continuum sys-
tems. J Vibr Acoust 131(2):557–557
The Sustainable Interaction Analysis of Cause
Marketing and Ethical Consumption in Electric
Business Platform: Based on Game Theory and
Simulation Analysis
Xu Zu and Weiping Yu(B)

331382107@qq.com
Abstract. Based on evolutionary game theory and simulation analysis,

the paper studies the sustainable interaction of cause marketing in elec-
tronic business platform, which constructs the evolutionary game model
of cause marketing and ethical consumption, and studies the existence
and stability of the equilibrium in the four situation interaction system,
and also verifies results by Netlogo. The main contributions are that
ethical consumption is the precondition of sustainable cause marketing
development, reducing costs is the key to promote sustainable interaction
development, electric business platform should vigorously promote social
responsibility consciousness, and promote the policy support. Finally the
paper constructs the theory frame for the realization of benign sustain-
able interaction.
Keywords: Cause marketing · Ethical consumption · The evolutionary

game model · Simulated analysis
1 Introduction
In recent years, the development rate of China’s e-commerce is 2–3 times of GDP
(7% 9%). China’s B2C online shopping market deals reached 609.67 billion yuan
in the second quarter of 2016. With the rise of e-commerce, the businessman
tried to introduce a cause marketing methods to get more customers focus on
their homogeneous products, for example, some businesses join the public wel-
fare plan of Taobao platform. Cause marketing dated from the public welfare
activity of “Renovation of the statue of liberty” by the cooperation of American
Express Company and Alice Island Foundation, and then it obtained the wide-
spread attention and development. Nowadays, more and more traditional cor-
porations have been successfully performed the corporate social responsibility
by the form of cause marketing. The cause marketing is not only contributing
to society, but also obtaining the business interests [1,8], for example, the cause
marketing activity of “buy a bottle of water, donate a penny” improved Nongfu
234 X. Zu and W. Yu
spring’s product sales, and Nongfu Spring raised more than 5 million yuan,
361◦ “ONE CARES ONE” also promoted sales of especially-made shoes and
expanded brand influence. Cause marketing, also known as charity Marketing or
cause-related Marketing, is a special form of donation. It depends on the pur-
chasing behavior of consumers, and donates the amount of a certain percentage
of business to the special public welfare project. Cause marketing has become
one of the important way of charity enterprises to perform social responsibility in
many countries. Cause marketing can promote sales, social responsibility image
and brand awareness [9,13]. In E-commerce platform, corporations, which suc-
cessfully continue to push the cause marketing, not only can achieve the aim of
marketing businesses, but also can effectively promote the healthy development
of Chinese social and economic. However, focusing on E-commerce situation,
businesses joined the cause marketing activities in E-commerce platform, and
didn’t get a good response of consumers.
From the practice experience of western countries, the benign interaction
between the enterprise and the consumer is the key to successfully propel the
corporate social responsibility [12]. The concept of ethical consumption is that
consumers not only consider the commercial value of product, but also consider
the beneficial impact of their purchasing behavior on society, environment, and
so on [5], that effectively promote the corporate cause marketing to fulfill the
social responsibility better. The sustained interaction of cause marketing and
ethical consumption can ensure that businessmen get more consumer approval
ratings in order to cover cost and promote sales and corporate image. According
to the survey in U.S. market, 78% of consumers are willing to buy products with
cause marketing, 66% of consumers willing to switch brands to support cause
marketing [3], and 96.6% consumers also incline to choose homogeneous products
with good corporate social image [6] in China market. However, consumers are
also “rational economic man”. Because of the higher perceived risk in the E-
commerce platform, consumers are not willing to fulfill the ethical consumption
with much cost in order to balance the benefit and the cost [11], so there is often
the inconsistent phenomenon about the consumer ethical attitude and behavior.
The relationship researches of cause marketing and ethical consumption
mainly focus on two aspects: the first is the relationship of cause marketing and
ethical consumption based on the investigation method and experiment method
[1,9], the second is the cause of the inconsistent about consumers’ ethical atti-
tude and behavior based on the interview and other qualitative methods [2].
However, few scholars’ researches focus on the interaction mechanism, especially
the research of systemic balance and dynamic evolution in the E-commerce plat-
form. It belongs to a kind of gambling behavior that cause marketing and ethical
consumption. This research constructs the evolutionary game model based on
evolutionary game theory and systematically explores the evolution rules. The
research result not only enriches the existing theory research of cause market-
ing and ethical consumption, but also provides valuable advice on sustainable
promoting the cause marketing and ethical consumption in the E-commerce
platform.
Analysis of Cause Marketing and Ethical Consumption 235
2 Evolutionary Game Model Building

Evolutionary game theory emphasizes the dynamic equilibrium. It can real reac-
tion behavior diversity and complexity of the body combined the game the-
ory analysis and the dynamic evolution process analysis [7,10]. Based on evolu-
tionary game theory, this research adopts the method that two kinds of group
repeated game and replication dynamic evolutionary game, and puts forward
the evolutionary game model to research like the assumptions:
Hypothesis 1: The two types of players are businesses and consumers in the
E-commerce platform, and they are the rational “economic individuals”. The
businesses behavior sets are cause marketing and non cause marketing and the
consumers behavior sets are ethical consumption and unethical consumption.
Hypothesis 2: The proportion of businesses with cause marketing is x, and the
other is 1 − x; the proportion of consumers with ethical consumption is y, and
the other is 1 − y.
Hypothesis 3: The product is homogeneous, the product cost for cause marketing
is C1 and product sales price is P1 , The endogenous utility, the brand image and
perceived social contribution and so on, and the E− commerce platform support
utility are F ; the product cost for non cause marketing is C2 and product sales
price is P2 , and C1 is higher than C2 .
Hypothesis 4: The social responsibility consciousness of consumers is strength-
ened with the development of society, the ethical consumption not only meets
the basic demand for products, but also makes consumers additional utility of
social contribution, the utility is V , and K is the additional cost that consumers
need to pay higher for ethical consumption than unethical consumption.
Based on the above assumptions, businesses’ revenue is as follows:
(1) When businesses adopt cause marketing and consumers adopt ethical con-
sumption, businesses’ revenue is P1 − C1 + F ;
(2) When businesses adopt cause marketing and consumers adopt unethical
consumption, businesses’ revenue is −C1 + F ;
(3) When businesses adopt non cause marketing and consumers adopt ethical
consumption, businesses’ revenue is −C2 ;
(4) When businesses adopt non cause marketing and consumers adopt unethical
consumption, businesses’ revenue is P2 − C2 .
Consumers’ revenue is as follows:
(1) When consumers adopt ethical consumption and businesses adopt cause
marketing, consumers’ revenue is V − K;
(2) When consumers adopt ethical consumption and businesses adopt non cause
marketing, consumers’ revenue is −V ;
(3) When consumers adopt unethical consumption and businesses adopt cause
marketing or non cause marketing, consumers’ revenue is 0. Payoff matrix
for gaming revenue are shown in Table 1.
236 X. Zu and W. Yu
Table 1. Payoff matrix for gaming revenue about businesses and consumers.
Type Consumers
Ethical Unethical
consumption (y) consumption
(1 − y)
Businesses Cause marketing (x) P1 − C1 + F, V − K −C1 + F, 0
Non cause marketing (1 − x) −C2 , −V P2 − C2 , 0
Expected returns of cause marketing and non cause marketing respectively

are U01 , U02 , the average return of businesses is U 0 :
U01 = y(P1 − C1 + F ) + (1 − y)(−C1 + F ) = yP1 − C1 + F, (1)

U02 = y(−C2 ) + (1 − y)(P2 − C2 ) = (1 − y)P2 − C2 , (2)
U 0 = xU01 + (1 − x)U02 . (3)
Expected returns of ethical consumption and unethical consumption respec-

tively are U11 , U12 , the average return of consumers is U 1 :
U11 = x(V − K) + (1 − x)(−V ) = x(2V − K) − V, (4)

U12 = x · 0 + (1−x) · 0 = 0, (5)
U1 = yU11 + (1 − y)U12 . (6)
According to the above expressions, we can know the replicated dynamic

equations respectively are:
dx
ẋ = = x(U01 − U 0 ) = x(1 − x) [y(P1 + P2 ) + F − C1 − P2 + C2 ] , (7)
dt
dy
ẏ = = y(U11 − U 1 ) = y(1 − y) [x(2V − K) − V ] . (8)
dt
3 Evolutionary Game Model Analysis
3.1 Equilibrium Point and Stability Analysis

dy
Make dxdt = 0 and dt = 0, based on results of replicated dynamic equation of
the interaction between the businesses and the consumers, we can get five local
equilibrium points: (0, 0), (1, 0), (0, 1), (1, 1) and (p∗ , q ∗ ), and
V
p∗ = , (9)
2V − K
C1 − F + P2 − C2
q∗ = . (10)
P1 + P2
Friedman proposed that the stability of the equilibrium point of the evolution
system can be obtained from the local stability analysis of the jacobian matrix
(denoted by J) [4]. The jacobian matrix by the combination of Eqs. (7) and (8):

∂ ẋ ∂ ẋ
a11 a12
J = ∂ ẏ ∂ ẏ =
∂x ∂y
, (11)
∂x ∂y
a21 a22
and
a11 = (1 − 2x) [y(P1 + P2 ) + F − C1 − P2 + C2 ] , (12)
a12 = x(1 − x)(P1 + P2 ), (13)
a21 = y(1 − y)(2V − K), (14)
a22 = (1 − 2y) [x(2V − K) − V ] . (15)
Thus, we can get the numerical results about a11 , a12 , a21 and a22 under the
five local equilibrium points, the results are shown in Table 2.
Table 2. The numerical results about a11 , a12 , a21 and a22 under the five local equi-
librium points.
Equilibrium point a11 a12 a21 a22

(0, 0) F − C1 − P2 + C2 0 0 −V
(1, 0) −F + C1 + P2 − C2 0 0 V −K
(0, 1) P1 + F − C1 + C2 0 0 V
(1, 1) −P1 − F + C1 − C2 0 0 K −V
(p∗ , q ∗ ) 0 A B 0
And the expressions of A and B respectively are:

V V
A= (1 − )(P1 + P2 ), (16)
2V − K 2V − K
C1 − F + P2 − C2 C1 − F + P2 − C2
B= (1 − )(2V − K). (17)
P1 + P2 P1 + P2
When the jacobian matrix satisfies the following conditions of the determi-
nant value (denoted by det J) and trace value (denoted by trJ), the local equi-
librium of system is the stability and can be the evolutionary stable strategy
(denoted by ESS).
Condition 1:

a11 a12

det J = = a11 a22 − a12 a21 > 0, (18)
a21 a22
Condition 2:
trJ = a11 + a22 < 0. (19)
238 X. Zu and W. Yu
Obviously, when the local equilibrium of system is (p∗ , q ∗ ), a11 + a22 = 0 do

not meet condition 2, so it is not as ESS. The study needs to discuss the possi-
bility that four local equilibrium points become ESS from different situations in
the plane of N = {x, y |0 ≤ x, y ≤ 1 }.
3.2 Result Discussion

Case 1: When F > C1 + P2 − C2 and V > K, the stability of four local equi-
librium points is as the Table 3 shown. We can get the ESS when (1, 1) satisfies
the conditions of det J > 0 and trJ < 0.
Table 3. The stability analysis of local equilibrium point when F > C1 + P2 − C2 and
V > K.
Equilibrium point detJ trJ Stability

(0, 0) − Indeterminacy Saddle point
(0, 1) + + Instability point
(1, 1) + − ESS
Case 2: When F < C1 + P2 − C2 and V > K, the stability of four local

equilibrium points is as the Table 4 shown, We can get the ESS when (0,0) and
(1,1) satisfy the conditions of det J > 0 and trJ < 0.
Table 4. The stability analysis of local equilibrium point when F < C1 + P2 − C2 and
V > K.

(0, 0) + − ESS
(1, 1) + − ESS
Case 3: When F > C1 + P2 − C2 and V < K, the stability of four local

equilibrium points is as the Table 5 shown, We can get the ESS when (1, 0)
satisfy the conditions of det J > 0 and trJ < 0.
Case 4: When F < C1 + P2 − C2 and V < K, the stability of four local
equilibrium points is as the Table 6 shown, We can get the ESS when (1,0)
satisfy the conditions of det J > 0 and trJ < 0.
Table 5. The stability analysis of local equilibrium point when F > C1 + P2 − C2 and
V < K.

(1, 0) + − ESS
Table 6. The stability analysis of local equilibrium point when F < C1 + P2 − C2 and
V < K.

(0, 0) + − ESS
4 Netlogo Simulation Modeling and Analysis

The study validates and analyzes the stability of equilibrium points with the aid
of Netlogo modeling and simulation in the four situations. Netlogo is developed
by the center of connectionist learning and computer modeling in northwestern
university. It is based on multi-agent modeling, multi-agent parallel and asyn-
chronous updating, and the whole system is dynamic with time. It can well study
the macroscopic patterns of micro-individual interaction, and is especially suit-
able for modeling and simulating complex systems with time evolution. Based
on the analysis of the stability of the equilibrium point in the evolutionary game
model, this study sets the payment matrix from four cases and performs sim-
ulation verification and analysis. The horizontal axis represents the simulation
time and the vertical axis represents the ratio that the number of subjects to
the total number of game strategy in the simulation results chart. X1 Rate and
Y1 Rate respectively are the cause marketing rate and ethical consumption rate,
and the initial ratio is 1/2.
Case 1: From the stability analysis result of the replicated dynamic equation,
(1,1) is ESS when F > C1 + P2 − C2 and V > K. The study substitutes F = 8,
C1 = 4, P2 = 5, C2 = 2, V = 3 and K = 1 into the simulation model, and the
simulation result is as shown in Fig. 1.
The simulation results validate the deduction of the replicated dynamic equa-
tion. The evolution result is an ideal phenomenon that businesses, consumers and
electric business platform are involved in social responsibility behavior. Con-
sumers need to pay additional costs, but the additional utility of social contri-
bution is always greater than the additional costs paid, consumers still choose
ethical consumption strategy; cause marketing strategy will increase the cost
240 X. Zu and W. Yu
of income, but a strong sense of social responsibility and good support and
encouragement of electric business platform will increase the business additional
benefits, so businesses still prefer to cause marketing strategy. Responsibility
behavior of businesses and consumers will enter a good time of continuous inter-
active development.
(1,1) and (0,0) are ESS when F < C1 + P2 − C2 and V > K. The study respec-
tively substitutes F = 4, C1 = 4, P2 = 5, C2 = 2, V = 3 and K = 1 and F = 1,
C1 = 4, P2 = 5, C2 = 2, V = 3 and K = 0.5 into the simulation model, and the
simulation result is as shown in Figs. 2(a) and 3(a).
The simulation results validate that there are two kinds of behavior patterns
in the evolution system, namely [cause marketing, ethical consumption] and [non
cause marketing, unethical consumption]. When the benefits from a good cor-
porate brand image and self-perception of social contributions and commercial
policy benefits make up for the loss of profits due to cause marketing, businesses
will still choose cause marketing. Consumers, compensated for the additional
cost, will still choose the ethical consumption strategy, so businesses and con-
sumers are in a benign behavioral interaction process. When the additional ben-
efits of cause marketing can not make up for the loss of profits, the system will
quickly move toward a vicious direction: businesses will choose non cause mar-
keting strategy, consumers will choose unethical consumption strategy. Social
responsibility awareness of businesses and consumers is low and the system fails
to achieve the sustainable social interaction due to lack of awareness and support
of electric business platform.
(1,0) is ESS when F > C1 + P2 − C2 and V < K. The study substitutes F = 8,
The simulation results validate that businesses will still keen on cause mar-
keting although they can not sell products when the additional effect is greater
than the sum of non cause marketing profits and the cause marketing cost. Con-
sumers, as “economic man”, can not pay too much extra expense for ethical
consumption, and finally give up ethical consumption. The stability of system
is based on policy support of the electric business platform and businesses own
economic interests, if the electronic business platform reduce support and the
economic benefits can not make up for cause marketing cost, then the system
will be rapidly unbalanced.
(0,0) is ESS when F < C1 + P2 − C2 and V < K. The study substitutes F = 1,
The simulation results validate that consumers can not use the extra utility
of ethical consumption to make up for the extra cost, so they abandon the
ethical consumption, and gradually choose unethical consumption through game
learning. Although businesses get a certain income because of cause marketing,
Fig. 1. The simulation result (F = 8, Fig. 2. The simulation result (F = 4,

C1 = 4, P2 = 5, C2 = 2, V = 3, K = 1) C1 = 4, P2 = 5, C2 = 2, V = 3, K = 1)
Fig. 3. The simulation result (F = 1, Fig. 4. The simulation result (F = 8,

C1 = 4, P2 = 5, C2 = 2, V = 3, K = C1 = 4, P2 = 5, C2 = 2, V = 3, K = 4)
0.5)
Fig. 5. The simulation result (F = 1, C1 = 4, P2 = 5, C2 = 2, V = 3, K = 4)
businesses finally take non cause marketing strategy in order to sell products
for profit without ethical consumption support, so the system will evolve to a
vicious direction, and will be rapidly unbalanced.

From the perspective of evolutionary game, this paper uses replication dynamic
evolutionary game method and Netlogo numerical simulation to analyze the sus-
tainable interaction mechanism between cause marketing and ethical consump-
tion, and obtains the positive and sustainable interaction conditions of cause
marketing and ethical consumption.
242 X. Zu and W. Yu
(1) Ethical consumption is a prerequisite for the sustainable development of

cause marketing. It’s the safeguard for the sustainable development of cause
marketing that consumers are willing to ethical consumption. Consumer
cost affects choices of consumer behavior, low level of social responsibility
awareness will directly increase the support burden of electric business plat-
form to cause marketing, can not guarantee the product sales and achieve
the good interaction of cause marketing and ethical consumption. Therefore,
electronic business platform and businesses should implement the cause mar-
keting, increase the cause marketing publicity to enhance consumer aware-
ness of social responsibility, while reducing the time cost that consumer
understand the specific content of cause marketing.
(2) Reducing cost is the key to promote sustainable development of cause mar-
keting. Cause marketing cost is directly related to the sustainable interactive
development of cause marketing and ethical consumption. Businesses should
work from two aspects: First, reduce the cause marketing cost. The price of
cause marketing product is not necessarily the same as the price of non cause
marketing product, but it can not exceed the maximum afforded range of
ethical consumption. The price should be combined with consumer’s sense
of social responsibility, at the same time, businesses should use other ways
to reduce cause marketing cost. Second, businesses should indirectly reduce
consumer ethical consumption cost. Businesses should accurately analyze
consumer behavior, so as to accurately put cause marketing for advertising,
really more convenient for consumers to understand the specific content of
cause marketing, thereby reducing consumer spending on the time and other
additional costs, better to achieve sustainable interaction of cause marketing
and ethical consumption.
(3) Electronic business platform should vigorously promote the sense of social
responsibility and promote the platform policy support. Businesses, as “eco-
nomic man”, always want their own additional endogenous utility and elec-
tric platform support for cause marketing to make up for opportunity cost
and accounting cost. The platform should guide businesses to enhance the
sense of social responsibility, so that the maximum expectation of policy
support defined in the reasonable range of C1–C2. The specific work can
be implement from two aspects: First, increase social responsibility propa-
ganda. Only relying on the platform support policy is unable to maintain the
sustainable development of cause marketing. Its important to increase the
effectiveness of business and consumer social responsibility that improve
the social responsibility sense in the electric business platform. Electric
business platform should actively promote social responsibility concept to
improve social responsibility awareness of businesses and consumers in order
to increase the effectiveness of social responsibility. Second, improve the
social responsibility incentive mechanism. Cause marketing will increase the
cost of investment, electric business platform should speed up the develop-
ment of policy incentives and other business incentives, to a certain extent
make up the cost burden of business social responsibility, to better promote
the sustainable development of cause marketing.
Fig. 6. The theoretical framework for the realization of benign and sustainable inter-
action.
Based on the above conclusions, this paper constructs a theoretical framework

for the realization of benign and sustainable interaction between cause marketing
and ethical consumption, as shown in Fig. 6:
Acknowledgements. This paper is supported by the General Project of the National

Natural Science Foundation of China “The Mechanism of Effect of Corporate Social
Responsibility Behavior on the Image of Scandal Brand and Strategy Research” (No.
71372189).
References
1. Al-Dmour H, Al-Madani S et al (2016) Factors affecting the effectiveness of cause-
related marketing campaign: moderating effect of sponsor-cause congruence. Int J
Mark Stud 8(5):114–127
2. Carrigan M, Attalla A (2001) The myth of the ethical consumer - do ethics matter
in purchase behaviour? J Consum Mark 18(7):560–578
3. Chang CT (2008) To donate or not to donate? Product characteristics and framing
effects of cause-related marketing on consumer purchase behavior. Psychol Mark
25(12):1089–1110
4. Friedman D (1991) Evolutionary games in economics. Econometrica 59(3):637–666
5. Hoffmann S, Hutter K (2012) Carrotmob as a new form of ethical consumption. The
nature of the concept and avenues for future research. J Consum Policy 35(2):215–
236
6. Huo B, Zhou Y (2014) Research on relationship of corporate social responsibility,
corporate reputation and corporate performance. J Ind. Technol Economics 1:59–
65 (in Chinese)
7. Khalifa NB, El-Azouzi R et al (2016) Evolutionary games in interacting commu-
nities. Dynamic Games Appl 1:1–26
8. Kim YJ, Lee WN (2009) Overcoming consumer skepticism in cause-related market-
ing: the effects of corporate social responsibility and donation size claim objectivity.
J Promot Manage 15(4):465–483
9. Plewa C, Conduit J, Quester PG (2015) The impact of corporate volunteering on
CSR image: a consumer perspective. J Bus Ethics 3:643–659
244 X. Zu and W. Yu
10. Selten R (1986) Evolutionary stability in extensive two-person games - correction

and further development. Math Soc Sci 3:223–266
11. Shaw D, Mcmaster R, Newholm T (2016) Care and commitment in ethical con-
sumption: an exploration of the ‘attitude-behaviour gap’. J Bus Ethics 136(2):251–
265
12. Shin Y, Thai VV (2015) The impact of corporate social responsibility on customer
satisfaction, relationship maintenance and loyalty in the shipping industry. Corp
Soc Responsib Environ Manage 6:381–392
13. Vanhamme J, Lindgreen A et al (2012) To do well by doing good: improving
corporate image through cause-related marketing. J Bus Ethics 109(3):259–274
Data Analysis
How to Predict Financing Efficiency in
Public-Private Partnerships–In an Aspect
of Uncertainties
Yixin Qiu1 , Umair Akram1(B) , Sihan Lin1 , and Muhammad Nazam2

1
School of Economics and Management,
Beijing University of Posts and Telecommunications,
Beijing 100876, People’s Republic of China
leo umairinfo@yahoo.com
2
Institute of Business and Management Science,
University of Agriculture Faisalabad, Faisalabad, Pakistan
Abstract. Financing efficiency in Public-Private Partnership (PPP)

projects is a vital but impressionable aspect, whose likely consequence
always influenced by inevitable uncertainties and confused the stakehold-
ers. This study aims to investigate the methods for evaluating and pre-
dicting the relative efficiency of current financing scheme in PPP project,
which surrounding with various uncertainties. The current study pro-
posed a new algorithm for assessing the inancing efficiency and simu-
lating the flexibilities in PPP projects simultaneously. Furthermore, it
integrates DEA and Monte Carlo to provide the possible probability dis-
tribution of financing efficiency which takes the uncertainties in PPP.
A case study is given to show its feasibility and practicability. It shows
uncertainties in PPP projects considerably count for its deviation from
the optimum value. This innovative model provides an avenue to pre-
dict the effects and possible outcome of an optimal scheme, which is
important for financing scheme decision making.
Keywords: PPP · Monte Carlo · DEA · Financing efficiency
1 Introduction
Infrastructure constructions are essential to support the world economic develop-
ment. Both in developed and developing countries, governments have dedicated
a significant share of the public budget to develop and refurbish infrastruc-
tures [10]. While, with the global economy slows, the lack of economic resource
can prevent further economic growth and the development of already planned
infrastructure projects [5]. Infrastructure builders (always public sectors) are
driven to attract private investment to take part in infrastructure constructions.
This idea has elicited the generation of Public-Private Partnerships (PPP) mode.
248 Y. Qiu et al.
The development and popularization of PPP well bridge the gap between a need
of infrastructures and a lack of budget. Worldwide, PPP grows into one of the
most important procurement mechanisms of project development.
PPP mode holds numerable advantages, among which the financial efficiency
is one of the main superiority. The British government directly name PPP
projects as “Private Finance Initiative”, therefore its effect on financing is evi-
dently reflected. The Asian Development Bank strong introduced PPP for con-
structing infrastructures at the beginning of its establishment, in order to reach
the value of money and promote the efficiency of financing [1]. Many scholars
studied the financing effects of PPP mode either from theoretical aspect or relat-
ing with practical application. Sastoque [12] concluded that it is one of the key
characteristics of PPP to count private sector’s contribution on financial issues
of public projects. Gatti indicated PPP is an efficient financing instrument acts
through designing, structuring, and executing, as an important role in whole
life cycle [6]. The high financing efficiency of PPP mode allows infrastructure
projects obtain the best value for money.
However, stakeholders of PPP are in need of a benchmark to learn about
the financing efficiency quantitatively; it is a necessary prerequisite for making a
sensible decision before it simplement. Uncertainties implicated in PPP projects,
like lengthy durations, high transaction costs and a lack of competition and
transparency, always lead to inefficiencies and ineffectiveness [2], and further
drag the financing efficiency down. Experts are trying to build a systematic
model for the assessment of financing efficiency in PPP. While, PPP projects
are multi-participation activities and multi-indicator processes which makes this
work more complicated than traditional ones, let alone making predictions and
selecting the optimal scheme.
Various methods have been proposed by scholars to settle this issue. The first
family of methods universally used in the assessment of PPP projects is Fuzzy
Synthetic Evaluation (FSE). Verweij [14] employed FSE to analysis satisfactory
outcome of financing in 27 PPP road constructions in the Netherlands. FSE
performs well in comparing existing options, however, fails to propose an opti-
mum scheme and make prediction or simulation. The second family of methods
for addressing the above problem is Analytic Hierarchy Process (AHP) analysis.
Zhuang [21] applied AHP to research the fiscal control in PPP of city infrastruc-
ture. AHP suffered from the same limitation as FSE. Besides, it is easily influ-
enced by the subjective rating of experts. Recently, an improved method had
been proposed for achieving a comparison of the existing options, optimization
of the inefficient ones, as well as eliminating subjective influence from human.
It is Data Envelopment Analysis (DEA). In YuYuanchun’s study, DEA plays
an important role in evaluating the technology transfer efficiency in industry-
university-research institution [17]. Wang Hong investigated the efficiency of
debt risk and fiscal expenditure of local government with an application of DEA
[15]. Another strong point of DEA is that it supports target data of each indi-
cator in relatively inefficient projects, which provides a reference for optimizing
How to Predict Financing Efficiency in Public-Private Partnerships 249
the original scheme. But it ignores the uncertainties and external influences in
projects implement, which can be conducted by Monte Carlo (MC) approach.
The objective of this work is to propose a mathematical model to evaluate
and predict the financing efficiency in PPP Projects. We apply DEA method
to handle difficulties caused by multiple participators and factors, and integrate
MC method to simulate uncertainties embraced in PPP projects. It is may offer
a new insight and academic reference for related research.
2 Problem Statement
The motivation of the proposed method lies in the fact that the crude FSE and
AHP method are limited in giving general evaluation and selecting the optimal
scheme; moreover, they are vulnerable to anthropic rating [9]. These restrictions
result in a series of dilemma in PPP projects, like hard to assess the potential
influence of uncertainties, and lack of quantitative support for decision making
of financing scheme. So, stakeholders in PPP projects are in an urgent need of a
systematic model that offers data support for the decision of financing strategies,
which should be uncertainties considering, optimum proposal providing, and free
from artificial factors.
DEA is able to achieve all the mentioned goals except that it fails to take
uncertainties in the implement and operation of PPP projects into account. The
principle of Monte Carlo analysis is to run stochastic values of independent vari-
able sand then simulate and identifies the possible distribution of outcomes [20].
A massive repetition of simulation processes counts the flexibility of changing
environment, which supplement the flaw of DEA.
In this work, we try to propound an approach by exploiting Monte Carlo
approach and integrating it with DEA. Monte Carlo works for simulating the
uncertainties in PPP projects which facilitate the new method to systematically
fulfill evaluation and optimization of financing efficiency in PPP Projects.
3 Monte Carlo and Its Application

Monte Carlo (MC) simulation is a computerized mathematical technique that
allows people to simulate an approximate solution in quantitative analysis and
decision making. The technique was first exploited by scientists who work in
the nuclear technology field when they were facing with a 6-dimensional integral
equation [20]. Calculating a result by repeated simulation was the only way to
solve this complex multiple problem. Since then, MC was widely applied in the
nuclear area, however, be confined to this area also because of the restriction of
the capability of computing. As in projects, the related influence factor can be
even more complex compared with in nuclear technology, a strong ability in com-
puting and storage is needed [4]. With the booming development of computer
in recent years, MC is widespread to various fields, for handling the simulation
of multidimensional problems, especially in project decision making which con-
stantly face with uncertainty, ambiguity, and variability. By using this method,
250 Y. Qiu et al.
the distribution of all possible outcome of an event is generated by running a

model a large scale of times, and then prompts and assists the decision-making
[11]. The main principle of MC is simulating the input of multiple independent
variables, in accordance with their expected distributions, and drawing the prox-
imate distribution and statistic data of dependent index. In the procedures of
MC, it always involves determining the change rules of independent factors and
exploring their interaction with dependent outcome. Then, the running of numer-
ous simulations could identify the range and distribution of possible outcomes
according to a number of scenarios. While managing a project, administrators
always face up with difficulties when there is a list of alternatives for the project
but lack of clues of their possible impact on the project. MC technique helps
in forecasting the likely outcome of an event and thereby promotes making an
informed decision. The method algorithm is shown in its succession interactive
five steps [4,11,20]:
Step 1. Clearly defining the target problem y, and creating a parametric model
composed by its influence factors x1 , x2 , · · · , xq , the model is noted as
Eq. (1):
y = f (x1 , x2 , · · · , xq ). (1)
Step 2. On the basis of fitted or hypothetic distributions of influence factors,
generating a random input set of data, x1 , x2 , · · · , xq ;
Step 3. Adding the random set which is generated from step 2 into the model
in step 1, obtaining an effective calculation and recording results as yi ;
Step 4. According to the accuracy requirement of specific research, determining
the simulation times n, and repeating steps 2 and 3 for i = 1 to i =
n (n = 1000 in this work);
Step 5. Obtaining a distribution curve and analyzing the results by using sta-
tistic indicators generated from the simulation.
In the execution of PPP projects, forecasting the final financing efficiency is
a complex and repeated issue. It always bewilders stakeholders because of its
high degree of complexity and uncertainty. MC simulation assists in simulating
likely scenarios and giving a probable distribution of objectives with taking the
influence of risk factors and uncertainties into consider. In the model, we put
forward in this study, DEA is applied to take the place of parametric model
included in step 1. The key benefits and applicability of using the MC analysis
in this research are listed below:
• MC simulation settles the uncertainties caused by multiple participators and
complex influence factors in PPP projects well. Differing from the crude
traditional model, MC exempts the complex mathematical calculation, and
directly obtains likely outcomes by simulating a series of possible scenarios,
which make it more operable and practical.
• MC overcomes the irreversibility involved in decision making, especially in
significant determination as those in PPP projects. Simulation is a way to
imitate results by running a large number of potential situations. MC simu-
lation meets the need of foresight the possible outcome in advance.
• With the advanced computing technology, Monte Carlo simulation is easier

to operate. It gains the availability and feasibility in practical application of
PPP projects.
4 Model Establishment
Scholars have explored a number of avenues to calculate and optimize the pro-
ductivity and operational efficiency issues. However, it is noteworthy that DEA is
one of the most popular approaches in this area. It is more recent in applications
among studies and it also draws more effective conclusions in comparison with
some other methodologies like Stochastic Frontier Analysis (SFA) [13]. With a
popularization of the PPP mode, employing DEA as modeling method is becom-
ing more and more popular in latest decade [19]. While in the specific financing
efficiency field of PPP projects, only a few scholars have gone into and DEA is
scarcely used.
In this research, we first adopt DEA to evaluate the financing efficiency of
specific PPP projects and get the initially optimized data. Due to the limita-
tion of theoretical mathematical models, DEA ignores the uncertainties and all
possible scenarios in PPP projects. Therefore, Monte Carlo is designed to play
a role for filling in the vacancy. The integrated algorithm can be used to both
give an assessment and practically optimized scheme and its efficiency.
4.1 Evaluation and Optimization of Financing Efficiency in PPP

Projects
In the projects area, the application of DEA always embodies three main parts:
selecting input and output indexes, forming the decision-making unit set, and
finally analyzing from calculated outcome. For the efficiency of financing in PPP
projects, it involves two major areas: raising funds in low cost and using funds
efficiently. In addition, franchise and public service of public infrastructure are
two characteristics in PPP project. Accordingly, five indicators are selected to
measure the financing efficiency of PPP projects in this study: the ratio of capital
funds, the duration, and the franchise as input indexes; social influence and the
turnover ratio of total capital as output indexes.
Five indicators refer to five aspects of financing efficiency in PPP projects.
Their names, definitions, notations and formulas for indicators are listed in
Table 1.
For making up an effective decision-making unit set, the number of decision-
making units should be no less than twice the number of a sum of the input and
output indicators; otherwise its ability to distinguish efficiency will decline [7].
So, when measures the relative efficiency of a PPP project, the data of the other
nine projects is needed to compose the decision-making unit set. Then input all
indexes into DEA model and results can be generated. The calculating model is
as Eq. (2) [8].
252 Y. Qiu et al.
Table 1. The definition, notation, and calculation formula of five indexes in DEA
model
Index Notation Meaning Formulation

capital invested by partners
Ratio of capital X1 Percentage of capital amount X1 = total construction cost
funds contributed by partners in the
total investment
Duration X2 Construction time takes in PPP X2 = construction time (month)
projects, which related to the
value of money

m
Bt
Franchise rights X3 Privilege the franchise of public x3 =
t=1 (1+it )
infrastructure to co-enterprises
µ1j µ2j
Social influence Y1 Analyze from the number of Y1j = max µ1j + max µ2j
service users (µ1 ) and related 1≤j≤n 1≤j≤n
webpage (µ2 )
net operating income
Turnover ratio of Y2 Evaluate management quality Y2 = average total asset
total capital and utilization efficiency of total
capital
Bt : excess earnings in tth y; m : duration of franchise; i: discount rates.
K L

min θ−ε s− + s+
k=1 l=1
⎧ n
⎪
⎪ Xkj λj + s− = θxko , k = 1, 2, 3
⎪
⎪
⎪
⎪ j=1
⎪
⎪ n (2)
⎨ Y λ − s+ = y , l = 1, 2
lj j ko
s. t. j=1
⎪
⎪ n
⎪
⎪ λj = 1
⎪
⎪
⎪
⎪ j=1
⎩ − +
s , s , λj ≥ 0, j = 1, 2, · · · , n.
Three indicators about efficiency can be obtained: technical efficiency θ, scale

efficiency θSE and pure technical efficiency θP E . θSE and θP E separately show
the financing efficiency of PPP projects in aspects of scale and pure technique.
Only when θP E and θSE equal 1 at the same time, it means that the targeted
project is in a relatively efficient level among other projects in this decision-
making units set. A particular relationship among the three indictors is shown
in Eq. (3).
θ = θP E × θSE . (3)
In addition, the input redundancy s+ and the output deficiency s− stem
from DEA model indicate targets of every independent indicator for reaching
a relative efficiency state. They fail to count uncertainties embodied in PPP
projects, while are important raw data for MC simulation.
4.2 Monte Carlo for Simulating Uncertainties

The Integration of MC and DEA is able to build an avenue for simultaneously
fulfilling the assessment and prediction of financing efficiency in PPP. Although
DEA model have generated an optimal target for each parameter, it is presented
Fig. 1. The flowchart of simulation algorithm integrates Monte Carlo with DEA
as theoretical and fixed data. During the practical development and construction
of a PPP project, all the changes in policy, economics, natural environment
and project itself generate flexibilities and fluctuations. They may result in a
deviation from the expected optimal scheme. MC analysis is strong at simulating
realistic uncertainties by randomly sampling the possible parameters set based
on their fitted distribution.
The simulation process is the vital part in Monte Carlo analysis. It con-
tains defining distributions of independent parameter, inputting random sam-
pling data set, and gathering results to generate a fitted distribution of dependent
indicators. When integrates with DEA model, the sampled random date set need
to be put in DEA model for running and giving the results of efficiency indica-
tors, provided that the simulation times does not reach 1000 yet. An aggregation
of all the results obtained from every simulation forms a possible financing effi-
ciency distribution. Consequently, uncertainty is considered in the majorization.
The flowchart of simulation process of the new model is shown in Fig. 1.
The innovative algorithm combines data processes of MC and DEA, which
contribute to a complement of both methods. DEA creates raw data and algo-
rithm for simulation in MC, and MC simulation bridge the gap between reality
and theoretical values given by DEA. With the integrated algorithm, a prob-
ability distribution of financing efficiency with potential changes in implement
will be drawn. It offers a quantitative reference for managers of PPP projects to
analyze the feasibility and risk of the optimum proposal, and then make more
evidence-based and financing efficient decision.
5 Case Study
5.1 Project Introduction
Australia Adelaide Water Utility project was developed to mitigate the water
threat. PPP mode was applied to ally the public and private sector for supporting
254 Y. Qiu et al.
more efficient and quality service. The United Water Corporation (private sec-
tor) cooperated with the Water Company in South Australia (public sector)
to construct the PPP project. The overall budget was 4.3 billion included 1.6
billion capital investments, which was raised from both main stakeholders and
the others investment from banks and other social capital. During the negotia-
tion stage, the government committed a 27 year franchise right. The completed
water utility was expected to serve 5 million citizens and get 521 million’s annual
operating incomes.
5.2 Simulation and Optimization Analysis

In order to composing an effective decision-making unit, set for DEA analysis,
we collected the related data of nine other PPP projects include Bird Nest and
the Sydney Olympic Games Stadium. In order to guarantee their comparability,
these projects shared similar scale and age of building; further, all the data is
transform into the form of ratio of capital funds, duration, franchise rights, social
influence and turnover ratio of total capital. The results of relative financing effi-
ciency: pure technical efficiency θP E , scale efficiency θSE and technical efficiency
θ are calculated with the application of DEA as listed in Table 2.
From the Table 2, a conclusion is drawn that, the financing efficiency is far
away from relatively efficient state compared with the other projects in this
decision-making unit set. When other projects almost reached 1, a relative effi-
cient state, the pure technical efficiency, scale efficiency and technical efficiency
of financing for Australia Adelaide Water Utility are separate to be 0.667, 0.454
and 0.303. Thus, an optimal scheme is in an urge need to improve the situation
of relative inefficiency in Australia Adelaide Water Utility PPP project.
DEA model delivers an improvement strategy for relative inefficient units
to reach an ideal level, namely a relative efficient financing efficiency state. For
Australia Adelaide Water Utility, the initial input and final target given by DEA
model are shown in Table 3.
Table 2. Financing efficiency of each project stem from DEA model
Project name θP E θSE θ

Australia Adelaide Water Utility 0.667 0.454 irs 0.303
Bird Nest 1 1 1
Sidney Olympic Stadium 1 0.969 irs 0.969
Beijing no. 4 subway 1 1 1
South Africa Ekurhuleni high speed railway 1 0.894 irs 0.894
Perpignan-Figueras rail link 1 1 1
Sidney Harbour Tunnel 1 1 1
Kamchay Hydropower Station 0.725 0.993 irs 0.721
QingDao Sewage Treatment Works 1 1 1
Hong Kong Disneyland 1 1 1
Table 3. Comparison of initial and target data of Australia Adelaide Water Utility
Item Ratioof Duration Franchise Social Turnover ratio

capital funds rights influence of total capital
Initial input 0.37 36 77.28 0.18 0.12
Target value 0.21 24 11.5 0.27 0.3
The target values are more theoretical, rather than feasible and practical.
For in the actual construction, it is surrounded by a changing environment all
the time, any uncertainty can result in a deviation from the optimal target.
If we predict and simulate all the possibility value of independent parameters,
the possible value of dependent parameters can be obtained. MC is created to
imitate this process. Ratio of capital funds, duration, franchise rights, social
influence and turnover ratio of total capital are five independent parameters in
this algorithm, we collect data from over hundred PPP projects so as to generate
a fitting probability distribution for each parameter, and they are hypothesized
as show in Fig. 2.
After sampling 1000 times in Monte Carlo simulation and running extracted
data in DEA model, the crossover process finally generates the probability dis-
tribution of pure technical efficiency θP E , scale efficiency θSE and technical effi-
ciency θ (Fig. 3). It shows the fitted probability curve for scale efficiency and
technical efficiency are lognormal distribution; while for pure technical efficiency
256 Y. Qiu et al.
is beta distribution. With setting the optimal scheme as a target, the expected
value for the pure technical efficiency reaches 0.8, has a promotion over 20%
than the initial value; for the scale efficiency reaches 0.68, get a considerable
50% improvement; for the technical efficiency meets 0.54, improves 80%. The
blue area highlighted in graphs reveals the most probable range of financing
efficiency value with a boundary set at the half of the maximum value.
Finally, we can make a conclusion that the practical implementation of an
optimal scheme hardly fulfills the relative efficient state in the end, because
of all the uncertainties and change of environment related with PPP projects.
While the innovative algorithm proposed in this research offers an approach
to predict possible outcomes and distribution of financing efficiency. It enables
stakeholders make a better decision among financing strategies and prepare for
probable consequences.
6 Conclusions
In this work, we propose an efficient and practical model for estimating and pre-
dicting financing efficiency in PPP projects. It integrates DEA and Monte Carlo
approach to measure financing efficiency scientifically and objectively, as well as
take the uncertainties embodied in PPP project into consider. The innovative
model designs a crossover running between DEA and MC, which enable them to
compensate for each other’s weaknesses. DEA is an effective approach to mea-
sure the relative efficiency and delivered an optimal target to reach the relative
efficiency state [19]. Wanke, Peter F also applied DEA to analyze scale efficiency
of PPP projects in Brazilian ports [16]. While PPP projects are full of risks
and uncertainties which will result in a fluctuation from the optimal target, the
intrinsic restrictions of DEA fail to take it into consider. A number of scholars,
like DijunTan and Yuzba Bahadr, use the features of Monte Carlo to quantify
and visualize volatility and flexibility [3,18]. In this manuscript, the MC method
is combined with DEA to handle uncertainties through a series of simulation.
The proposed algorithm offers a comprehensive method to assess and forecast
the financing efficiency in PPP projects.
In the case study of Australia Adelaide Water Utility project, one- thousand-
time simulation forms the possible probability distribution of final pure technical
efficiency, scale efficiency and technical efficiency of financing. It reveals the
possible outcome of financing efficiency in this PPP project when set the optimal
value derived from DEA as the target. It shows that, although uncertainties make
the relative efficiency hardly to meet a perfectly financing efficient state, while
the three efficiency parameters promoted 20%–80% than the initial financing
scheme. This result enables the stakeholders to have a full preparation for the
probable consequence.
The application of the new method which integrates MC and DEA is not
limited in the financing efficiency of PPP projects. It can be extended to various
fields whose assessment and simulation process involve multiple indicators and
uncertainties. After making corresponding changes in input and output indexes
and fit their possibility distribution, the innovative algorithm can be widely used
in different areas.
7 Summary of Research Results and Future Study
This study proposed a new model to predict the financing efficiency with fluctu-
ation caused by uncertainties in PPP projects, based on integration of DEA and
MC. The case study of Australia Adelaide Water Utility project shows that this
innovative method is capable to evaluate and predict flexibilities quantitatively
and visually. It also indicates that the uncertainties in PPP projects make the
target values in optimal scheme hard to reach, while in this specific project, it
also realizes a considerable promotion in all three efficiency indicators.
Further research, such as a more comprehensive indexes system of DEA and
subdivide uncertainties in to different categories, is necessary to delve more
in-depth and practical in this topic. A more comprehensive and thorough inves-
tigation of related factors in DEA method will contribute to a more accurate
outcome. Considering that the implementation of PPPs would be affected by eco-
nomic, social and environmental conditions, sorted uncertainties can facilitate a
more accurate calculation. Alternatively, a series of diversified case study need
to be carried out in the future to verify and enrich the reliability and feasibility
of this new model.
Acknowledgement. The authors would like to thank the National Natural Science
Foundation of China for financially supporting this research (Grant No.: 71502011).
It is also supported by the Fundamental Funds for Humanities and Social Sciences of
Beijing Jiaotong University (Grant No.: 2015jbwj013).
258 Y. Qiu et al.
References
1. Bank AD (2013) Public-private partnership operational plan 2012-2020: realizing
the vision for strategy 2020: the transformational role of public-private partnerships
in Asian development bank operations
2. Broadbent J, Gill J, Laughlin R (2008) Identifying and controlling risk: the problem
of uncertainty in the private finance initiative in the UK’s national health service.
Crit Perspect Account 19(1):40–78
3. Dijun T, Yixiang T (2007) Volatility modeling with conditional volatility and real-
ized volatility. In: International Conference on Management Science and Engineer-
ing Management, pp 264–272
4. Dubi A (2000) Monte carlo applications in systems engineering. In: IEEE R&M
symposium on a Dubi stochastic modeling of realistic systems topics in reliability
& maintainability & statistics tutorial notes
5. Franke M, John F (2011) What comes next after recession? - airline industry
scenarios and potential end games. J Air Transp Manage 17(1):19–26
6. Gatti S (2013) Project finance in theory and practice: designing, structuring, and
financing private and public projects. Academic Press, Cambridge
7. Li G, Lei Q, Yang Y (2014) The efficiency evaluation of Chinese film and televi-
sion industry listed companies based on DEA method. In: 2014 Proceedings of the
eighth international conference on management science and engineering manage-
ment. Springer, Heidelberg, pp 83–95
8. Lin Q (2013) Introduction of decision analysis. Tsinghua University Press, Beijing,
pp 45–47
9. Liu J, Yang Y (2014) Efficiency evaluation of Chinese press and publication listed
companies based on DEA model
10. Martins J, Rui CM, Cruz CO (2014) Maximizing the value for money of PPP
arrangements through flexibility: an application to airports. J Air Transp Manage
39:72–80
11. Platon V, Constantinescu A (2014) Monte carlo method in risk analysis for invest-
ment projects. Procedia Econ Finance 15:393–400
12. Sastoque LM, Arboleda CA, Ponz JL (2016) A proposal for risk allocation in social
infrastructure projects applying PPP in Colombia. Procedia Eng 145:1354–1361
13. Schøyen H, Odeck J (2013) The technical efficiency of Norwegian container ports:
a comparison to some Nordic and UK container ports using data envelopment
analysis (DEA). Marit Econ Logist 15(2):197–221
14. Verweij S (2015) Producing satisfactory outcomes in the implementation phase of
PPP infrastructure projects: a fuzzy set qualitative comparative analysis of 27 road
constructions in the Netherlands. Int J Project Manage 33(8):1877–1887
15. Wang H, Huang J, Li H (2017) Local government debt risk, fiscal expenditure
efficiency and economic growth. Springer, Singapore
16. Wanke PF, Barros CP (2015) Public-private partnerships and scale efficiency in
Brazilian ports: evidence from two-stage dea analysis. Socio-Econ Plann Sci 51:13–
22
17. Yu Y, Gu X, Chen Y (2017) Research on the technology transfer efficiency eval-
uation in industry-university-research institution collaborative innovation and its
affecting factors based on the two-stage DEA model. In: 2017 Proceedings of the
tenth international conference on management science and engineering manage-
ment. Springer, Singapore, pp 237–249
18. Yüzbaşı B, Ahmed SE (2015) Shrinkage ridge regression estimators in high-

dimensional linear models. Springer, Heidelberg
19. Zhang S, Chan APC et al (2016) Critical review on PPP research - a search from
the Chinese and international journals. Int J Project Manage 34(4):597–612
20. Zhongji X (1985) Monte Carlo Approach. Shanghai Science and Technology Press,
Shanghai
21. Zhuang M (2012) Research on fiscal control in public-private partnerships of city
infrastructure. In: Electronic system-integration technology conference, pp 1–4
The Moderating Effects of Capacity Utilization
on the Relationship Between Capacity Changes
and Asymmetric Labor Costs Behavior
Abdulellah Azeez Karrar1,2(B) , DongPing Han1 , and Sobakinova Donata1

1
School of Management, Harbin Institute of Technology,
Harbin 150001, People’s Republic of China
2
Collage of Economics and Administrative, University of Kufa, Najaf 54001, Iraq
karara.alkhaldy@uokufa.edu.iq
Abstract. The study presents an interactive and theoretical research

model based on the economic theory of asymmetry behavior, which means
labor costs are more likely to increase when activity increases greater than
activity decreases. This paper aims to analyze the effect of capacity utiliza-
tion on prior capacity changes and labor costs behavior. Study method-
ology tests the hypotheses of nonlinear relationships among variables in
the proposed model using multiple regression analysis. Data have been
collected from five factories of Iraqi cement firms by physical output data
from 2006 to 2015 monthly. Findings reveal that degree of cost asymmetry
of manufacturing, selling and administrative, and total labor is associated
with capacity changes and that a moderating effect of capacity utilization
strongly reveals the opposite pattern supported. Our results from Iraqi
manufacturing provide a new evidence on labor costs behavior and contain
important implications for both accounting and economic theories. The
value of the paper is one of the earliest studies that empirically examine
and explain the moderating effects of capacity utilization on employment
capacity and labor costs behavior.
Keywords: Capacity utilization · Employment capacity · Labor costs

behavior · Asymmetry theory
1 Introduction
Alternative behavior model is an important issue based on deliberate managerial
decisions for cost accounting [1]. Underlying the traditional behavior model is
based on the activity level that associates changes linearly and proportionately
[30]. Many studies have documented an asymmetry behavior between costs and
resources in various perspectives [15,22–25,36]. The variances between the level
of costs to activity rise changes and the level of costs to activity fall changes called
“asymmetry” as sticky, anti-stick costs [7]. Conversely, The asymmetric cost
behavior is investigating the consequence cost stickiness on actual and forecast
earnings [28].
The Moderating Effects of Capacity Utilization 261
Managers tend to save capacity when capacity is strained and demand falls
then, add capacity when demand grows [7,13,16]. As optimistic expectations
lead to saving idle capacity in the sales response to necessary declines but these
predictions will make a sticky cost behavior [17]. The asymmetric slack mainte-
nance does not automatically lead to cost stickiness, but if managers retained in
response to a previous sales decrease, the significant is stickiness on a prior sales
increase and anti-stickiness on a prior sales decrease [11]. These motivations lead
us to question is “how capacity utilization moderate the relationship between
labor costs and capacity changes?” Capacity utilization is a measurement of
the percentage of firm productivity capacity that determine unused resources to
total resources [31].
Unemployment fluctuations may make stickiness labor costs in structural
parameters because the statistics of unemployment are replicated as well in
sticky labor behavior [14]. Empirical implications created a counter cyclical
wedge between the real labor and marginal rate [19]. Moreover, the significance
of labor is not flexible or inflexible because it is necessary to manage cost struc-
ture [6]. Moderating model has certain guiding significance for the application
of big data to management decision [37].
Prior studies identified that capacity changes asymmetrically impact on man-
ufacturing labor, selling and administrative, and total costs [5,13]. However,
Banker et al. [9] examined the moderating effect of prior periods on the rela-
tionship between sales changes and costs behavior. Yet, no research available in
literature presents the impact of capacity utilization on asymmetric labor costs
behavior.
This paper aims to analyze the impact of capacity utilization on the asym-
metry behavior between labor costs and employment capacity, and to explain
scientifically the differences without and within the interaction effects.
2 Related Studies and Hypothesis Development
Resources adjustment costs create asymmetry behavior in marginal costs because

managerial expectations depend on optimism and pessimism about sales future
[21]. Sometimes, uncompleted cost information generates asymmetric results [34].
It depends on how managers use resources and understand their mechanism [33].
Managerial optimism weakness on previous changes called cost sticky, and man-
agerial pessimism weakness on previous changes called anti-sticky [2]. Asymme-
try behavior is managerial practical because managers are expected reluctant to
lower committed resources in periods of cost-effective growth [8,12]. Limitation
of labor costs associated with operating’s during sales changes [38]. In addition,
high product competition has affected costs behavior that generates Stickiness
costs or likely cost responsive [17]. This behavior of labor cost will generate a
good control in future because the cost should match with managerial expecta-
tions that mean little variations [11]. Overall, The main reason of asymmetry
behavior is cost response cannot be a rise or decline fast enough to combine
changes in resources [4,30]. Few studies have considered the effect of asymmetric
262 A.A. Karrar et al.
information on reporting decisions and analyze after demand realization under

a decentralized setting [27].
The capacity problems and asymmetry behavior positively are associated
with control on economics variables such as employees’ intensity, production
unit, workforce level [15,26,35,39], and machine level. There are behavioral dif-
ferences between costs categories [24]. Dalla Via and Perego [18] indicated sig-
nificant labor cost stickiness conditional on a prior sales changes for Italian
companies listed. Chen et al. [15] studied the interaction effects among capacity,
economics variables, sales changes, and selling and administrative costs. Their
results indicate asymmetry behavior and find significant. Azeez et al. [5] exam-
ined the relationship of labor costs and employment capacity in which they
activate, and discuss how they share to the growth and improvement rate of cat-
egories of labor cost with capacity. Banker et al. [11] used employment protection
legislation to examine adjustment costs and employment resources in different
countries. Their finding showed that costs respond to sales increase greater than
sales decrease.
We summarize these predictions by three hypotheses as follows:
Hypothesis 1a (H1a): Conditional on a prior capacity utilization increases,

change in manufacturing labor costs to capacity increase is greater than
capacity decrease.
Hypothesis 1b (H1b): Moderating effects on a prior capacity utilization increases,
change in manufacturing labor costs to capacity increase is less than capacity
decrease.
change in selling and administrative labor costs to capacity increase is greater
than capacity decrease.
change in selling and administrative labor costs to capacity increase is less
than capacity decrease.
change in total labor costs to capacity increase is greater than capacity
decrease.
change in total labor costs to capacity increase is less than capacity decrease.
3 Research Methodology
3.1 Sample Description
We have empirically investigated 600 practical observation from five factories

across hundred twenty months each one from 2006 to 2015 of cement in national
Iraqi company. We calculated all log-change using same months in years as
indexes of study variables using OLS estimation for regression and the second
stage we applied hierarchal moderation regression as function of results.
3.2 Model Development

Our research design is based on the established asymmetric labor costs model
of Anderson et al. [1], and Azeez et al. [5], Banker et al. [9] as in equations:

M LCi,t −M LCi,t−1 MEqi,t −MEqi,t−1
ln M LCi,t = ϕ0 + ϕ1 ln
MEqi,t
MEqi,t −MEqi,t−1
+ϕ2 DECi,t ln

MEqi,t
qui,t −qui,t−1
+ϕ3 ln qu (1)
i,t
+ϕ4 DECi,t ln
MEqi,t
qui,t −qui,t−1
× ln qui,t + εi,t ,
where MLCi,t is a manufacturing labor cost for firm i at time t and is a non-
linear function of independent variables (ϕ1 , ϕ2 , ϕ3 , and ϕ4 ). M E is a manu-
facturing Capacity of employees by hours. DECi,t is an indicator set to 1 if
MEqi,t < MEqi,t−1 , and set to 0 otherwise. ϕ0 is a parameter that estimates the
asymmetric labor cost unassociated with employment capacity changes. ϕ1 is a
parameter that estimates the association between manufacturing labor cost and
employment capacity changes increase. ϕ2 is a Parameter of asymmetry mea-
surement that estimates the association between manufacturing labor cost and
employment capacity changes during increasing and decreasing. ϕ3 is a moder-
ating variable that estimates the association between labor cost and capacity
utilization changes. ϕ4 is a critical value that estimates how capacity utiliza-
tion moderates the association between manufacturing employment capacity and
labor costs behavior. εi,t is an error term for variability in labor cost change esti-
mation for firm i at time t.
The capacity utilization was computed for labor hour’s measures of the
unused and total capacity for each of the periods as in Eq. (2) below:
unused capacity
qu = 1 − , (2)
total capacity
where qu is employees Capacity Utilization for each activities. Unused capacity
is that hours cannot exceed effective capacity. Total capacity is the maximum
hours of output designed for the operation and facility other.

S&ALCi,t −S&ALCi,t−1 SEqi,t −SEqi,t−1
ln = θ0 + θ1 ln
S&ALCi,t SEqi,t
SEqi,t −SEqi,t−1
+θ2 DECi,t ln
i,t
SEq
qui,t −qui,t−1
+θ3 ln qu (3)
i,t
+θ4 DECi,t ln
SEqi,t
qui,t −qui,t−1
× ln qui,t + ωi,t ,
where: S&ALCi , is a selling and administrative labor cost for firm i at time t.
SEqi,t is a selling and administrative Capacity of employees by hours, and all
other variables were defined previously.

TLCi,t −TLCi,t−1 TEqi,t −TEqi,t−1
ln = γ0 + γ1 ln
TLCi,t TEqi,t
TEqi,t −TEqi,t−1
+γ2 DECi,t ln
i,t
TEq
qui,t −qui,t−1
+γ3 ln qu (4)
i,t
+γ4 DECi,t ln
TEqi,t
qui,t −qui,t−1
× ln qui,t + αi,t ,
where TLCi , is a total labor cost for firm i at time t. MEqi,t is a total Capacity
of employees by hours include all of the activities, and remainder of variables
are defined previously.
To develop the concept of asymmetry labor costs about employment capacity
structure (high utilization). The coefficients φ1 , and measure the average per-
centage increase in labor costs for one present increase in employment capacity,
whereas the sum of coefficients (φ1 + φ2 ), (θ1 + θ2 ) and (γ1 + γ2 ) in three mod-
els to measure the average percentage of decrease in labor costs for one present
Table 1. Definition of variables in empirical models
Variable/Calculation Description

MLCi,t −MLCi,t−1
ln MLCi,t
Log-change in manufacturing labor costs. All labor costs
of manufacturing process and its supporting, payments
of industrial activity

S&ALCi,t −S&ALCi,t−1
ln S&ALCi,t
Log-change in Selling & administrative labor costs. All
labor costs of selling and administrative process,
payments of selling, and administrative activity

TLCi,t −TLCi,t−1
ln TLCi,t
Log-change in total labor costs. All labor costs of
manufacturing, supporting, selling, and administrative
activities

ln MEqi,t
Log-change in manufacturing employment capacity.
Total hours of employees in manufacturing and
supporting activity

ln SEqi,t
Log-change in selling employment capacity. Total hours
of employees in selling, general, and administrative
activity

ln TEqi,t
Log-change in total employment capacity. Total hours of
employees in the all activities

qui,t −qui,t−1
ln qui,t
Log-change in capacity utilization. The percentage rate
of each kind of actual capacity to design capacity
Manufacturing labor cost is collected from manufacturing, engineering & services, and
quality control activities. Selling and administrative labor cost is collected from mar-
keting and administration activities. Practical employees’ capacity is collected from a
dataset of the planning department. Unused employees’ capacity is calculated based on
differences between actual and practical employees capacity. Actual employees’ capac-
ity is calculated by divided labor costs on payment rates.
decrease in the employment capacity. The labor costs behavior is asymmetry

when the average percentage increase in labor costs not equal the average per-
centage decrease in labor costs, that is, if labor costs decrease less when employ-
ment capacity fall than they increase when employment capacity rise [1,36].
Means the labor costs have sticky behavior (φ1 , θ1 and γ1 , and will be positive,
φ2 , θ2 and γ2 , will be negative). The empirical hypothesis for sticky behavior
means that φ2 < 0. This finding an empirical test of Hypotheses 1a, 2a, and 3a.
In addition, the current study is exciting literature using the interactive effect of
capacity utilization on the asymmetric relationship between employment capac-
ity and labor costs behavior. Hypotheses 1b, 2b, and 3b imply φ2 , are negative
and φ4 , are positive reflecting stickiness and anti-stickiness respectively [9]. The
moderation role of capacity utilization implies that labor costs asymmetry will
be less than in the case without the interactive effects. The variables definitions
are defined in Table 1.
4 Findings and Discussion
In the current study, we estimate three models for each hypothesis of manufac-
turing labor costs, selling and administrative labor costs, and total labor costs
(Tables 2, 3 and 4). For all of those categories, the estimates indicate significance
levels and support our hypotheses. The evidence finds labor costs behavior is
sticky on a prior capacity decrease and indicate significant anti-sticky behavior
on a prior capacity decrease with interactive effects of capacity utilization (<0
and >0 continues to all models).
4.1 Asymmetric Labor Costs Response to Capacity Changes
In the regression analysis, we find a sticky behavior between labor costs and
capacity changes for all of the labor costs categories but in different degrees,
the estimates indicate that ϕ1 is positive, ϕ2 is negative and significant (see
model (1) to each Tables 1, 2 and 3). The results found manufacturing, selling
and administrative labor, total labor costs are sticky on average by magnitudes
of prior employment capacity increase and decrease.
Table 2-panel A. presents a model (1), manufacturing labor costs response to
capacity increase is statistically greater than capacity decrease (ϕ1 ϕ1 +ϕ2 ).
The coefficient is negative (−0.56%, SE. = 0.18) and significantly different from
zero at the 1 % (t-statistics −3.176). On average, the manufacturing labor costs
increase 0.74% per 1% increase in employment capacity (ϕ1 ) and costs decrease
by 0.18% per 1% decrease in employment capacity (ϕ1 + ϕ2 ). The results show
that the adjusted R2 is 0.71, the model is significant (p < 0.001) level. Means
the behavior between employment capacity change and manufacturing labor
cost is sticky (H1a). However, Selling and administrative labor costs response
to capacity increase statistically is greater than capacity decrease (θ1 > + 2).
The coefficient is negative and significant (−0.40%, SE. = 0.12, t-statistics
−3.25). On average, the Selling and administrative labor costs increase 0.71%
Table 2. Validation test of the sticky behavior: Nonlinear analysis of moderation

test among employment capacity, capacity utilization, and manufacturing labor costs
change
Panel A: Regression analysis: direct effects model 1 Dependent variable = manufacturing

labor cost (MLC)
Variable Parameter Parameter Standard Parameter
estimate error significance
(t-statistics)
Intercept ϕ0 0.15 0.08 0.077
(?) −1.77
MEq
i,t −MEqi,t−1
ln MEqi,t
ϕ1 0.74 0.14 0
Asymmetric (+) −5.22
measure
MEq
i,t −MEqi,t−1
ln DECi,t MEqi,t
ϕ2 −0.56 0.18 0.002
(−) (−3.176)

qui,t −qui,t−1
ln qui,t
ϕ3 0.31 0.07 0.015
(+) −4.432
Adjusted R2 0.706
F-value 481.03
Significant level 0
Panel B: Moderation analysis: interactive effects of capacity utilization - model 2
Intercept ϕ0 0.19 0.35 0.593
(?) −0.53
MEq
i,t −MEqi,t−1
ln MEqi,t
ϕ1 0.67 0.09 0
measure
MEq
i,t −MEqi,t−1
ln DECi,t MEqi,t
ϕ2 −0.22 0.34 0.516
(−) (−0.65)

qui,t −qui,t−1
ln qui,t
ϕ3 0.71 0.13 0.001
(+) −5.359
MEq −MEq
i,t i,t−1
ln DECi,t MEqi,t
ϕ4 0.11 0.02 0

qui,t −qui,t−1
× ln qu i,t
(+) −6.047
Adjusted R2 0.723
F-value 391.448
The results present a first examination that explain how the interaction affects of capacity uti-
lization impact on the relation between costs behavior and employment capacity changes. The
sample consists of 600 factor-month observations between 2006 and 2015 from Iraqi industry
on cement produce. All t-statistics calculate by using significant indicate at the 1%, 5%, 10%
level respectively.
Table 3. Validation test of the sticky behavior: Nonlinear analysis of moderation test
among employment capacity, capacity utilization, and selling and administrative labor
costs change
Panel A: Regression analysis: direct effects model 1 Dependent variable = selling and
administrative labor cost (S&ALC)
(t-statistics)
Intercept θ0 0.13 0.08 0.09
(?) −1.7
SEq
i,t −SEqi,t−1
ln SEqi,t
θ1 0.71 0.14 0
measure
SEq
i,t −SEqi,t−1
ln DECi,t SEqi,t
θ2 −0.4 0.12 0.001
(−) (−3.25)

qui,t −qui,t−1
ln qui,t
θ3 0.16 0.04 0.093
(+) −4.3
Adjusted R2 0.704
F-value 475.992
Panel B: Moderation analysis: interactive effects of capacity utilization C model 2
Intercept θ0 0.4 0.33 0.224
(?) −1.22
SEq
i,t −SEqi,t−1
ln SEqi,t
θ1 0.74 0.1 0
behavior
SEq
i,t −SEqi,t−1
ln DECi,t SEqi,t
θ2 −0.19 0.05 0.015
(−) (−3.67)

qui,t −qui,t−1
ln qui,t
θ3 0.7 0.25 0.042
(+) −2.79
SEq
i,t −SEqi,t−1
ln DECi,t SEqi,t
θ4 0.28 0.04 0

qui,t −qui,t−1
× ln qu i,t
(+) −6.81
Adjusted R2 0.725
F-value 395.823
All t-statistics calculate by using significant indicate at the 1%, 5%, 10% level respectively.
per 1% increase in employment capacity (θ1 ) and costs decrease by 0.31% per 1%
decrease in employment capacity (θ1 + θ2 ), for which Selling and administrative
labor costs expected to be sticky (H2a). The adjusted R2 is 0.70, where F-value is
475.99, the model is significant (p < 0.001) level. (Presented in Table 3-panel A.).
Finally, Table 4-panel A. in model (1), presented the empirical examination

of the relationship between total labor costs and employment capacity changes.
Total labor costs asymmetrically respond to capacity changes. The coefficient is
negative and significant (−0.58%, SE. = 0.08, t-statistics −7.68). On average,
the total labor costs increase by 0.67% per 1% increase in employment capacity
(γ1 ) and they decrease by 0.09% per 1% decrease in capacity (γ1 + γ2 ). The
results show that the adjusted R2 is 0.09, where F-value is 21.09, the model is
significant (p < 0.001) level. The result is consistent with (H3a).
4.2 Moderation Analysis: Interactive Effects of Capacity Utilization
In extension analysis, we estimate interactive models for each of the main com-
ponents of labor costs (manufacturing labor, selling and administrative labor,
and total labor), for extending literature of asymmetric cost behavior. The esti-
mates indicate significant stickiness conditional on a prior capacity decrease and
significant anti-stickiness conditional on a prior capacity decrease with moder-
ating effects of capacity utilization change (ϕ2 < 0 and ϕ4 > 0 respectively)
suggesting that capacity utilization changes are related to the effects of employ-
ment capacity changes on labor costs behavior. These findings support H1b, H2b,
and H3b.
Moderation test was conducted to examine the interacting effect of prior
sales changes between sales changes and asymmetric cost behavior [9]. The
labor costs exhibit significant stickiness without the interactive effects of capac-
ity utilization (ϕ2 = −0.22%, SE. = 0.34, t-statistics −0.65), (θ2 = −0.19%,
SE. = 0.05, t-statistics −3.67), and (γ2 = −0.48%, SE. = 0.06, t-statistics
−7.51) respectively, but the labor costs reveal the opposite pattern of significant
anti-stickiness within the interactive effects of capacity utilization (ϕ4 = 0.11%,
SE. = 0.02, t-statistics 6.05), (θ4 = 0.28%, SE. = 0.04, t-statistics 6.81), and
(γ4 = −0.13%, SE. = 0.06, t-statistics 2.16). These results document that labor
costs are description of a broader pattern of asymmetric cost behavior, which
extends to all the major components of labor costs for physical input quantity
(employment capacity) for labor costs behavior.
In additional, Banker et al. [9] indicated that costs respond to activity
increase is greater than activity increase within the moderation effects, which
estimated parameter of total employees costs increase equal 0.62 stronger than
0.42 within the interactive effects. Whereas our finding indicate that coefficients
of asymmetry measure less than within the interactive effects (ϕ2 = 0.56, θ2 =
−0.40, γ2 = −0.58) > (ϕ2 = −0.22, θ2 = −0.19, γ2 = −0.48) respectively
and capacity utilization change was observed to strengthen the relationship
between capacity changes and labor costs behavior (ΔR2 = 0.017, ΔR2 = 0.021,
ΔR2 = 0.009, p < 0.001). The results also underscore the value-added of interac-
tive analysis of capacity utilization, which in the moderation models of categories
of labor costs are important for drawing accurate implications about the nat-
ural of costs behavior. While Banker et al. [9] considered the moderation of
two-period analysis is value- added for theory of asymmetric cost behavior.
Table 4. Validation test of the sticky behavior: Nonlinear analysis of moderation test
among employment capacity, capacity utilization, and total labor costs change
Panel A: Regression analysis: direct effects C model 1 Dependent variable = total

labor cost (TLC)
(t-statistics)
Intercept γ0 0.66 0.29 0
(?) −2.24

ln TEqi,t
γ1 0.67 0.09 0
behavior

ln DECi,t TEqi,t
γ2 −0.58 0.08 0
(−) (−7.68)

qui,t −qui,t−1
ln qui,t
γ3 0.74 0.45 0.1
(+) −1.62
Adjusted R2 0.091
F-value 21.09
Panel B : Moderation analysis: interactive effects of capacity utilization - model 2
Intercept γ0 0.9 0.32 0
(?) −2.76

ln TEqi,t
γ1 0.86 0.11 0
behavior

ln DECi,t TEqi,t
γ2 −0.48 0.06 0
(−) (−7.51)

qui,t −qui,t−1
ln qui,t
γ3 0.41 0.2 0.049
(+) −2.007

ln DECi,t TEqi,t
γ4 0.13 0.06 0.031

qui,t −qui,t−1
× ln qui,t
(+) −2.16
Adjusted R2 0.1
F-value 17.084
All t-statistics calculate by using significant indicate at the 1%, 5%, 10% level
respectively.
4.3 Discussion
This research was conducted to analyze the moderation effects of capacity uti-
lization changes as they related to asymmetric costs behavior at manufacturing
sector firms in Iraq. Until recently, most studies have been conducted to investi-
gate why change in labor costs behavior to activity is asymmetry [3,8,18,20,29],
specially Banker et al. [9] conducted on moderating effects but they ignored the
important dimension of capacity utilization and also focused on multiple periods
analysis for same variables to overcome these issues, this research has presented
an interactive model based on data collection from multiple and different vari-
ables. Maintenance management is necessary an effective system to display the
surface temperature to reduce the capacity losses [32].
From the variable is used in the theoretical research model, three nonlin-
ear relationships were proposed and tested. The results found capacity utiliza-
tion changes can affect asymmetric cost behavior. Conversely, labor costs and
their categories have associated with employment capacity changes in different
degrees such as significant stickiness conditional on a prior capacity decrease
and significant anti-stickiness conditional on an interaction of capacity decreases
with capacity utilization increases. Finally, capacity utilization can decrease the
degree of sticky measure. The results from the nonlinear relationships tested to
confirm and align with the existing literature [1,3,5,9].
Anderson et al. [1] Documented that behavior between costs and resources
changes is asymmetry because costs respond to resources increase greater/less
than resources decrease that rejected the traditional models about fixed and
variable changes. However, costs behavior is sticky when prior sales increase,
whereas costs behavior is anti-sticky when prior sales change decrease. The costs
are anti-sticky when capacity utilization is unusually low [9]. On the other hand,
Azeez et al. [5] examined the labor costs behavior when employment capacity
change is increasing and decreasing for existing asymmetric cost behavior by
using physical output data. Finally, the results found capacity utilization is able
to moderate the relationship between capacity change and labor costs behavior,
and suggest new avenues of exploration for future studies.
5 Conclusion
Significance of the relationship between costs behavior and resources changes has
been an extension of the asymmetric cost behavior. Asymmetry phenomenon is
a new thinking created by [1] but it still under discussion because they did not
use many drivers, their findings used sales revenues change only and ignored a
physical output data for two reasons: physical data is not available, and sales
revenues typically is a more an appropriate empirical measure for activity than
physical data [10]. The current study provides an empirical examining that added
a new dimension for explaining how capacity utilization change moderates the
relationship between capacity changes and labor costs behavior. In addition,
our findings showed a complex fundamental design of asymmetric cost behav-
ior that combines two roles: labor costs behavior is sticky on a prior capacity
increase without the interaction of capacity utilization changes. Second, labor

costs behavior is anti-sticky on a prior capacity increase within the interaction of
capacity utilization changes, means the moderation effects of capacity utilization
made the opposite of the standard predictions.
The findings of the current study have revealed that capacity utilization
changes significantly have affected labor costs behavior. Furthermore, these pre-
dictions explain the influence of managerial expectations for future on firm
resources such as the general structure of optimal decisions with resources adjust-
ment costs, which generate asymmetry behavior. Finally, using capacity utiliza-
tion changes with asymmetric relations can decrease the degree of sticky behavior
between labor costs and capacity change, interaction capacity utilization changes
with a prior capacity decrease can converse the asymmetry degree from sticky to
anti-sticky. The results of moderation are unique and would provide productive
avenues for future studies.
Acknowledgements. The authors would like to thank Iraqi Cultural Attaché in

China and national company for supporting a research about data collecting and their
corresponding.
References
1. Anderson MC, Banker RD, Janakiraman SN (2003) Are selling, general, and
administrative costs “sticky”? J Account Res 41(1):47–63
2. Anderson MC, Lee JH, Mashruwala R (2015) Cost stickiness and cost inertia:
a two-driver model of cost behavior. Social Science Electronic Publishing SSRN
2599108
3. Anderson SW, Lanen WN (2007) Understanding cost management: what can we
learn from the evidence on ‘sticky costs’ ? SSRN Electron J SSRN 975135
4. Argilés JMA, Garcı́ablandón J (2007) Cost stickiness revisited: empirical applica-
tion for farms. Revista Espanola De Financiacion Y Contabilidad 38(187):579–605
5. Azeez KA, DongPing H, Mabula JB et al (2016) Using capacity utilization to
measure asymmetric labor cost behavior and capacity expansion. In: Management
science and engineering (ICMSE)
6. Babecky J, Caju PD et al (2012) How do European firms adjust their labour costs
when nominal wages are rigid? Labour Econ 19(5):792–801
7. Balakrishnan R, Labro E, Soderstrom NS (2014) Cost structure and sticky costs.
J Manage Account Res 26(2):91–116
8. Banker RD (2006) Labor market characteristics and cross-country differences in
cost stickiness. SSRN Electron J
9. Banker RD (2014) The moderating effect of prior sales changes on asymmetric cost
behavior. J Manage Account Res 26:221–242
10. Banker RD, Byzalov D (2014) Asymmetric cost behavior. J Manage Account Res
26(2):43–79
11. Banker RD, Basu S et al (2016) The confounding effect of cost stickiness on con-
servatism estimates. J Account Econ 61(1):203–220
12. Bugeja M, Lu M, Shan Y (2015) Cost stickiness in australia: characteristics and
determinants. Aust Account Rev 25(3):248C261
13. Cannon JN (2014) Determinants of ‘sticky costs’: an analysis of cost behavior using
united states air transportation industry data. Account Rev 89(5):1645
14. Casares M, Moreno A, Vázquez J (2009) Wage stickiness and unemployment fluc-
tuations: an alternative approach. Series 3(3):395–422
15. Chen CX, Hai LU, Sougiannis T (2012) The agency problem, corporate gover-
nance, and the asymmetrical behavior of selling, general, and administrative costs.
Contemp Account Res 29(1):252C282
16. Chen JV, Kama I, Lehavy R (2015) Management expectations and asymmetric
cost behavior. Technical report, Working Paper, University of Illinois at Chicago,
Tel Aviv University, and University of Michigan
17. Cohen S (2015) The sticky cost phenomenon at the local government level: empir-
ical evidence from Greece. Social Science Electronic Publishing SSRN 2575530
18. Dalla Via N, Perego P (2015) Sticky cost behaviour: evidence from small and
medium sized companies. Account Finance 54(3):753–778
19. Dicecio R (2009) Sticky wages and sectoral labor comovement. J Econ Dyn Control
33(3):538–553
20. Gu Z (2016) The political economy of labor cost behavior: evidence from china.
Social Science Electronic Publishing SSRN 2786533
21. Hamermesh DS, Pfann GA (1996) Adjustment costs in factor demand. J Econ Lit
34(3):1264–1292
22. Ibrahim AEA (2015) Economic growth and cost stickiness: evidence from Egypt.
J Financ Report Account 13(1):119–140
23. Jang Y (2016) Asymmetric cost behavior and value creation in m&a deals. Social
Science Electronic Publishing SSRN 2824132
24. Kama I, Dan W (2010) Do managers’ deliberate decisions induce sticky costs?
SSRN Electron J 112(1):39–47
25. Kama I, Dan W (2013) Do earnings targets and managerial incentives affect sticky
costs? J Account Res 51(1):201C224
26. Kee RC (2003) Operational planning and control with an activity-based costing
system. Adv Manage Account 11:59–84
27. Liao Y, Shen W, Lev B (2017) Asymmetric information effect on transshipment
reporting strategy
28. Malik M (2012) A review and synthesis of ‘cost stickiness’ literature. SSRN Elec-
tron J SSRN 2276760
29. Mohammad MN, Masoumeh N (2011) Evidence of cost behavior in Iranian firms.
In: 2011 international conference on advancements in information technology, vol
20, pp 254–258
30. Nematollahi M (2013) Investigation the relation between costs and revenues in
Iranian firms. N Y Sci J 6(11):52–57
31. Nyaoga RB, Wang M, Magutu PO (2015) Does capacity utilization mediate the
relationship between operations constraint management and value chain perfor-
mance of tea processing firms? evidence from Kenya. Int Strateg Manage Rev
3(1–2):81–95
32. Ramirez IS, Munoz CQG, Marquez FPG (2017) A condition monitoring system
for blades of wind turbine maintenance management. In: Proceedings of the tenth
international conference on management science and engineering management, pp
3–11
33. Teece DJ, Pisano G, Shuen A (2009) Dynamic capabilities and strategic manage-
ment. Oxford University Press, New York
34. Von Hippel E (1994) “Sticky Information” and the locus of problem solving: impli-
cations for innovation. Manage Sci 40(4):429–439
35. Watts T, Mcnair CJ et al (2009) Structural limits of capacity and implications for
visibility. J Account Organ Change 5(2):294–312
36. Weidenmier ML, Subramaniam C (2003) Additional evidence on the sticky behav-
ior of costs. Social Science Electronic Publishing
37. Yang A, Lan X et al (2017) An empirical study on the prisoners’ dilemma of
management decision using big data. Springer, Singapore
38. Zanella F, Oyelere P, Hossain S (2015) Are costs really sticky? evidence from
publicly listed companies in the uae. Appl Econ 47(60):6519–6528
39. Zhuang ZY, Chang SC (2015) Deciding product mix based on time-driven activity-
based costing by mixed integer programming. J Intell Manuf 4:1–16
The Empirical Analysis
of the Impact of Technical Innovation
on Manufacturing Upgrading-Based
on Subdivision Industry of China
Dan Jiang(B) and Yuan Yuan
Chongqing Technology and Business School,

719221214@qq.com
Abstract. At this stage, changing the mode of economic growth and

optimizing the industrial structure is the important content for China’s
economy restructuring and steady growth, so the adjustment of industrial
structure is imminent, especially manufacturing. This paper describes the
present situation of China’s technology innovation and the development
of manufacturing, and analyzes the problems of manufacturing upgrad-
ing, and then uses the data of China’ 27 manufacturing sub-sectors from
2005 to 2014 and establishes the regression model. The results showed
that technical innovation has a positive role in promoting the upgrading
of China’s manufacturing, but its effect is still affected by certain fac-
tors, and the differences on technical innovation are obvious. Finally, in
view of the problems found in the research, the paper puts forward some
policy suggestions for the upgrading of China’s manufacturing industry,
such as increasing R&D investment, building system of financial support,
and establishing a technical alliance about industrial-academic-research.
Keywords: Technical innovation · Manufacturing upgrading · Labor-

intensive industry · Capital-intensive industry · Technology-intensive
industry
1 Introduction
The manufacturing industry reflect a country’s productivity level, as well as
present the national competitiveness. It is the cornerstone of the realization
of social progress and national prosperity. With the advantage of reform and
opening policy, the labor force, some other natural resources and other factors,
China’ manufacturing industry has created a miracle and continued growth of
economy. In 2010, China’s manufacture value added exceeded the United States,
becoming a manufacturing country, is worthy of the name. As shown in Fig. 1.
Since the financial crisis in 2008, the development of the real economy has
been paid more and more attention by the countries all over the world. Therefore,
The Empirical Analysis of the Impact of Technical Innovation 275
Fig. 1. Comparison of manufacturing value added (dollar)
the development of manufacturing industry and the expansion of exports are the
key to the transformation of the world economic structure. The developed coun-
tries have taken measures to revitalize the manufacturing sector. For instance,
the USA wishes “re industrialization” once again to opened the gap with the
developing countries, to further promote its economic and social development, to
enhance its competitiveness in the international manufacture. Germany issued
“industry 4.0” which was proposed to establish an information physics system
network in 2013, making itself the creator and supplier of advanced intelligent
manufacturing technology. It provides the opportunities of technological inno-
vation for China’s implementation of “made in China 2025”. At the same time,
some developing countries take advantage of opportunities of the new round of
international industrial transfer, promote the industrialization process, thereby
exacerbating China’s manufacturing competitive environment.
With China’s rising labor costs and resource depletion, the advantage of
traditional manufacturing relying on cheap labor and resources is gradually lost.
In a word, we must rely on technological innovation to speed up the restructuring
of China’s manufacturing and the transformation of development mode. But how
does technological innovation affect the manufacturing upgrading of different
factors intensive industrial structure? This is an urgent problem in the process
of China’s industrialization. In view of this, based on the status of manufacturing
upgrading and the thinking of upgrading power, this paper focuses on the effect
of technical innovation on manufacturing upgrading.
2 Literature Review
Foreign scholars have carried on the thorough discussion on the technology
innovation and the industrial structure evolution, but it mostly based on the
industrial level. Guan [4] analyzed the relationship between technological inno-
vation and industrial competitiveness in the ability of R&D personnel’s learning
and knowledge acquisition, research and development ability, and the research
276 D. Jiang and Y. Yuan
showed that technology innovation is the key way to promote the competitive-
ness of the industry. Castellacci [2] studied the relationship between the external
innovation environment of technology innovation and industrial competitiveness,
the results show that the economic policies based on market oriented and insti-
tutional arrangements will directly affect the innovation mode, and play the role
in industry competitiveness. Cohen [3] considered the Schumpeterian hypothe-
ses relating innovation to market structure and firm size, and considers in more
depth the role of firm characteristics and industry-level variableslbroadly char-
acterized as reflecting demand, technological opportunity, and demand condi-
tionslin affecting firms’ innovative activity and performance. Adak [1] focused
on the influence of technological progress and innovation on the Turkish econ-
omy, and got the result that a significant effect of technological progress and
innovation on economic growth. Hong et al. [5] found that government grants
exert a negative influence on innovation efficiency of high-tech industries. How-
ever, the impact of private R&D funding is significant and positive. In view of
the problem of upgrading the structure of manufacturing industry, this paper
will not discuss it in detail.
Domestic scholars focus on the research on the mechanism and path of tech-
nological innovation to upgrade the manufacturing industry. Sun and Ye [6]
analyzed the role and mechanism of innovation for transformation and upgrad-
ing of manufacturing industry based on the connotation of the innovation drive,
and considered innovation driven to give impetus to manufacturing from three
aspects of power dimension, elements dimensions and competition dimension.
Zhao and Qiao [8] studied the upgrading mode of Shanghai’ manufacturing
industry, and pointed out that Shanghai needed to promote technology improve-
ment and innovation of the traditional manufacturing, and created growth indus-
try through the industrialization of high and new technology industry, and
pushed on the dynamic integration of producer services and advanced manufac-
turing technology, all of which strengthen policy guidance so that it can realize
comprehensively manufacturing upgrading.
From what has been discussed above, the existing literature on the analy-
sis of the technology innovation’s influence on manufacturing upgrading, which
was mostly based on regional classification, and there is less research based on
manufacturing industry so that it failed to reveal the impact of technological
innovation on Internal structure changes in manufacturing industry. Based on
the panel data model, the paper made an empirical analysis of the differences
in the impact of scientific and technical personnel, R&D internal expenditure,
sales revenue of new products, the number of effective invention patent, cost
of technical renovation, main business profit on the manufacturing upgrading
of labor-intensive industry, capital-intensive industry and technology-intensive
industry. Through the analysis of the technology innovation’s influence on dif-
ferent kinds of different factors intensive industries, that can provide guidance for
our country to make the corresponding different incentive policy on technological
innovation, and avoid invalid policy, and make the technological innovation to
better play its role so to realize the upgrading of the manufacturing structure.
It has certain practical significance.
3 Theoretical Basis
According to the classification standards of the National Bureau of Statistics,
the manufacturing industry is divided into 31 categories, due to the lack of data
during some years, this paper chooses 27 Subdivision manufacturing industry of
national above-scale industrial as the research object. Referring to the classifica-
tion method of the factor intensity and Wang and Dong’ classification method
[7] based on the structure of the manufacturing, the 27 Subdivision industry of
China’s manufacturing industry is divided into labor-intensive industry, capital
intensive industry and technology intensive industry. The specific classification
is shown in Table 1:
Table 1. Classification of different factors intensive industries
Manufacturing industry type Industrial distribution

Labor-intensive industry Agricultural food processing industry, food
manufacturing, beverage and refined tea
industry, tobacco industry, textile industry,
textile and apparel and footwear products
industry, leather and feather industry, Wood
processing industry
Capital-intensive industry Furniture manufacturing, paper making and
paper products, printing and recording
media copying, Stationery industry, Oil
processing, coking & nuclear fuel processing,
Chemical Fiber Industry, Rubber and
plastic products industry, non-mental
mineral product industry, Ferrous metal
smelting and rolling processing industry,
Non-ferrous metal smelting and rolling
processing industry, Metal products
manufacturing
Technology-intensive industry Pharmaceutical Industry, the manufactures
of chemical materials and products, General
Equipment Manufacturing sector, Special
Equipment Manufacturing Sector,
Electronic and Communication Equipment
Manufacturing Sector, Electrical machine
equipment manufacturing, Transportation
Manufacturing, instruments and cultural
office machinery
4.1 Index Selection
(1) Manufacturing Upgrading Index Selection

In existing research, the value added rate of manufacturing industry is generally
considered as a comprehensive reflection of the economic benefits and growth
quality of the manufacturing industry. There are also some scholars having used
industrial profit and tax amount, but the main business profit of the Subdivi-
sion manufacturing industry is chosen as an index to measure the manufacturing
upgrading. The reason why the author choose main business profit is that man-
ufacturing upgrading means improving the competitiveness and profitability of
enterprises. On the other hand, choosing the main business profit instead of net
profit ratio, that can eliminate government subsidies which enterprises receive
from government and other non-recurring profit, at the same time, taking into
account the availability of data, this paper chooses main business profit as the
variable of manufacturing upgrading.
(2) The Index Selection of Technological Innovation
In the measurement of technological innovation by domestic scholars, scientific
and technological personnel, R&D investment intensity and other indicators are
generally accepted evaluation indicators. However, considering the transforma-
tion and application of technological innovation achievements, this paper chooses
scientific and technical personnel, R&D internal expenditure, sales revenue of
new products, the number of effective invention patent, cost of technical renova-
tion to measure the level of technology innovation according to the availability
of data. The author tend to specifically study the impact of technological inno-
vation on the manufacturing upgrading, and don’t just use a single indicator of
technological innovation.
(3) The Selection of Control Variable
In this paper, adding the input of labor and capital which his regarded as control
variables into the model, that can reflect the impact of technical innovation and
factor input on the changes of manufacturing added value. The net value of fixed
assets of industrial enterprises above designated size approximately represent
material capital investment, and annual average number of employed persons in
all sectors of the manufacturing industry means the number of workers in the
manufacturing.
4.2 Data Description
The data which is used in the paper come from the “Statistical yearbook in
China”, “Industrial Statistics Yearbook in China”, “Labor Statistics Yearbook
in China”, “Statistical yearbook of scientific and technological activities of indus-
trial enterprises”, Some of the data are compiled according to the original data
of the statistical yearbook.
(1) Comparison of Scientific and Technical Personnel of Various Types of Man-

ufacturing Industry in China
As shown in Fig. 2, all kinds of scientific and technical personnel of manufacturing
industry of our country from 2005 to 2008 has maintained steady growth, due to
the financial crisis in 2008, from 2008 to 2009 it showed a downward trend, various
types of science and technology personnel began to increase slowly after 2009, at
the same time, China’s scientific and technical personnel of labor-intensive indus-
try and capital -intensive industry has been lagging behind in technology-intensive
industry and the total is about half.
(2) Comparison of R&D Internal Expenditure of Various Types of Manufactur-
ing Industry in China
As shown in Fig. 3, General speaking, among 2005–2014, China’s various types
of R&D internal expenditure of manufacturing industry is up forwards except
among 2008–2009. However, R&D internal expenditure of labor-intensive indus-
try and capital-intensive industry is slower than the growth of technology-
intensive industry.
(3) Comparison of Sales Revenue of New Products of Various Types of Manu-
facturing Industry in China
As shown in Fig. 4, China’s sales revenue of new products of technology-
intensive industry grew so slowly and it almost unchanged. It shows that China’s
technology-intensive industry has poor ability of technology transformation, and
the scientific and technological achievements can not be effectively translated
into practical productive forces. But the capital intensive manufacturing and
labor-intensive manufacturing industry has maintained a steady growth. Though
the number of effective invention patent of capital and labor intensive manu-
facturing is far less than the technology intensive industry, the latter continu-
ing innovation investment will make the sales revenue of new products greatly
improved.
(4) Comparison of Effective Invention Patent Number of Various Types of Man-
ufacturing Industry in China
As shown in Fig. 5, the number of effective invention patent in China’s
technology-intensive industry went through a growth spurt, and those of labor-
intensive industry and capital-intensive industry increased much slower. Perhaps
China’s technology-intensive industry more and more pay attention to innova-
tion output and the other two industries funding for innovation is not enough.
(5) Comparison of Cost of Technical Renovation of Various Types of Manufac-
turing Industry in China
As shown in Fig. 6, cost of technical renovation in China’s labor-intensive indus-
try grew slowly, until the peak reached early 2013, after that it was a downward
trend. While cost of technical renovation in China’s technology-intensive indus-
try basically is on the rise from 2005 to 2010, and exceeded the cost of technical
renovation capital- intensive industry in 2010, then it began to decline in 2011,
finally it reached at the same level as capital-intensive industry in 2014.
Fig. 2. Comparison of scientific and technical personnel
Fig. 3. Comparison of R&D internal expenditure (billion)
(6) Comparison of Main Business Profit of Various Types of Manufacturing

Industry in China
As shown in Fig. 7, main business profit in the three kinds of China’s manufactur-
ing overall were rising, and those of technology-intensive manufacturing is much
more profitable than capital-intensive industry and labor-intensive industry, but
capital and labor intensive manufacturing industry’s main business profits are
similar.
4.3 The Construction of the Model

According to the selected model variables, based on the augmented Cobb
Douglas production function, this paper establishes econometric model as
follows:
α β
Yit = ec+δ1 ln RDit +δ2 ln STit +δ3 ln EPit +δ4 ln P Sit +δ5 ln T Rit +uit Kit Lit . (1)
Among them, Yit represents the main business profit of manufacturing during
t period i industry; Kit , Lit represents fixed capital investment and the amount
of labor during t period i industry respectively; RDit , STit , EPit , P Sit and
Fig. 4. Comparison of sales revenue of new products (billion)
Fig. 5. The number of comparison of effective invention patent
Fig. 6. Comparison of cost of technical renovation (billion)
T Rit represents R&D internal expenditure, scientific and technical personnel,

the number of effective invention patent, sales revenue of new products, cost of
technical renovation during t period i industry, uit represents the random error.
Fig. 7. Comparison of main business profit (billion)
Before testing the data of model, the Eq. (1) on both sides are processed into
logarithmic for linear regression model, the model as follows:
ln Yit = c + α ln Kit + β ln Lit + δ1 ln RDit + δ2 ln STit + δ3 ln EPit + δ4 ln P Sit

+ δ5 ln T Rit + uit .
4.4 Variable Inspection
(1) Correlation Analysis of Variables

The following is the correlation analysis of selected variables, and from Table 2
we can see that independent variables were significantly correlated with the
dependent variable. Specifically, the stronger technique innovation capability
is, the higher the main business profit of the industry is. Besides, the capital
and labor input is in positive correlation with the operating profits. At the
same time, the respective variables are also highly relevant. The explanatory
variables including scientific and technical personnel, R&D internal expenditure,
sales revenue of new products, the numbers of effective invention patents and
cost of technical renovation was significantly correlated in 1% level, and the
correlation coefficient is greater than 0.8. It reveals that these indicators selection
of technical innovation is over-repeated, and the correlation is very strong.
(2) Empirical Results
After multicollinearity inspection of several variables, we found that VIF is 20.72
in average, more than 10, which shows that there is serious multicollinearity
between variables. Via detecting x3 and x4 linear with other explanatory vari-
ables exist serious, so they should be removed in the model. The value of VIF is
less than 10 after deletion. So the final determination of the independent variable
is x1 , x2 , x5 and x6 in this paper.
Because different manufacturing industries with individual differences and
mixed OLS cannot reflect individual differences, it is imperative to use fixed
effects and random effects model for regression. Considering the upgrading of
industrial structure with inertia index, this paper uses dynamic panel model
Table 2. Correlation analysis table
y x1 x2 x3 x4 x5 x6 x7
y 1
x1 0.671∗∗∗ 1
x2 0.713∗∗∗ 0.644∗∗∗ 1
x3 0.707∗∗∗ 0.685∗∗∗ 0.822∗∗∗ 1
x4 0.739∗∗∗ 0.747∗∗∗ 0.758∗∗∗ 0.973∗∗∗ 1
x5 0.751∗∗∗ 0.724∗∗∗ 0.692∗∗∗ 0.883∗∗∗ 0.939∗∗∗ 1
x6 0.716∗∗∗ 0.585∗∗∗ 0.651∗∗∗ 0.809∗∗∗ 0.814∗∗∗ 0.841∗∗∗ 1
x7 0.600∗∗∗ 0.760∗∗∗ 0.523∗∗∗ 0.764∗∗∗ 0.845∗∗∗ 0.813∗∗∗ 0.575∗∗∗ 1
and apply the system of GMM estimation to eliminate endogenous for further
analysis. In this paper, the regression model is as follows, model 1 is on behalf
of the panel data model, and model 2 is on behalf of the dynamic panel model.
Model 1 yit = αi + β1 x1 + β2 x2 + β3 x5 + β4 x6 + μit ,

Model 2 yit = αi + γyit−1 +β1 x1 + β2 x2 + β3 x5 + β4 x6 + μit .
Model 1 is the result of fixed effect, model 2 is he result of random effect.

The result of Houseman test showed that Chi2(4) = 7.22, the value of P is
0.1249, that can’t refuse null hypothesis of Houseman test, so random effects
model should be used. It can be seen from model 2, the technical innovation
ability can really promote the upgrading of industrial structure for the manu-
facturing industry. In particular, sales revenue of new products increased by 1%,
the industry’s main business profit increased 0.222%; cost of technical renovation
increased by 1%, the industry’s main business profit increased 0.17%. Because
of the diversification of products and competition, the invention of new products
and the input of technological transformation can increase brand influence and
competitiveness in a highly competitive market environment, thus increasing
corporate profits. From the point of control variables, net value of fixed assets
increased by 1%, the industry’s main business profit increased 0.173%; average
number of employees in manufacturing industry increased by 1%, the industry’s
main business profit increased 0.323%.
After the testing, the best model for dynamic panel is model 3 after the
introduction of the second order. It is found that the equation is over identi-
fied and the value of P is 0.4130, which shows that the instrumental variable
is effective, So the model is tenable. The lagged items of the explanatory vari-
ables entered into the model and the author found that the lagged items had
a significant impact. Specifically, a lag of the main business profit increased by
1%, the current main business profit will be reduced by 0.342%; two lag of main
business profit increased by 1%, the current main business profit will be reduced
by 0.301%; it can be seen that three lag of main business profits will increase
the current main business profit from the model 4. Perhaps it was related to the
Table 3. Regression result table
Variables (1) fe (2) re (3) sysgmm5 (4) sysgmm6

y y y y
L.y −0.342∗∗∗ −0.684∗∗∗
(0.021) (0.104)
L2.y −0.301∗∗∗ −0.440∗∗∗
(0.0245) (0.0631)
L3.y 1.862∗∗∗
(0.482)
x1 0.100 0.173∗∗∗ 0.00873 −1.061∗∗
(0.0857) (0.0643) (0.0667) (0.539)
x2 −0.0677 0.323∗∗∗ 1.752∗∗∗ 2.011∗∗∗
(0.357) (−0.0981) (0.194) (0.531)
x5 0.503∗∗∗ 0.222∗∗∗ 0.289∗∗∗ 0.174
(0.151) (0.0813) (0.0488) (0.151)
x6 0.148* 0.170∗∗∗ 0.447∗∗∗ −0.0328
(0.079) (0.0543) (0.0292) (0.0508)
Constant 10.37∗∗ 9.734∗∗∗ 5.999∗∗∗ 1.966
(5.026) (1.694) (2.058) (4.946)
Observations 270 270 166 114
R-squared 0.312
Number of ind1 52 52 52 44
Standard errors in parentheses (∗∗∗ p < 0.01, ∗∗
p < 0.05, ∗ p < 0.1)
manufacturing industry in our country which is mainly processing enterprises,

and many of which is low value-added industries. Companies tend to increase
plant equipment investment in the current or in the next period to increase the
profits in the future when the current operating profit increased (Table 3).
5 Conclusion
In this paper, it can be seen from the empirical results that the invention of new
products and technological innovation can increase the main business profits and
promote the upgrading of manufacturing industrial structure. China is a big man-
ufacturing country, whose economic growth is mainly driven by the second indus-
try at present, and China’s manufacturing industry accounted for the largest pro-
portion of the second industry. At the same time, the manufacturing industry can
absorb a lot of labor force. The development of the manufacturing industry is of
great significance to the increase of the employment rate, reduction of the gap
between the rich and the poor and the stability of the society. Therefore, the gov-
ernment should support the technological innovation of manufacturing industry,
the enterprise itself also should actively seek innovation to make themselves invin-
cible in the tide of marketization and internationalization.
References
1. Adak M (2015) Technological progress, innovation and economic growth; the case
of turkey. Procedia Soc Behav Sci 195:776–782
2. Castellacci F (2008) Innovation and the competitiveness of industries: comparing
the mainstream and the evolutionary approaches. Technol Forecast Soc Change
75(7):984–1006
3. Cohen WM (2015) Innovation and technological change, economics of. Int Encycl
Soc Behav Sci 2:160–168
4. Guan JC, Yam RCM et al (2006) A study of the relationship between competitive-
ness and technological innovation capability based on dea models. Eur J Oper Res
170(3):971–986
5. Hong J, Bea F (2016) Do government grants promote innovation efficiency in china’s
high-tech industries? Technovation 57:4–13
6. Sun S, Ye Q (2015) Mechanism and strategic choice of innovation driven manufac-
turing transformation. Ind Technol Forum 2:15–18
7. Wang ZH, Dong CT (2012) An analysis on the consistency between manufacturing
structure and quality structure of labor force in china:and the discussion of shortage
of migrant workers, skilled personnel, and difficulty of graduates employment. Popul
Econ 16(3):363–368
8. Zhao X, Qiao M (2012) Research on the mode and path of shanghai promoting
manufacturing upgrading with high technology. Econ Res Shanghai 2:63–69
A Crash Counts by Severity Based Hotspot
Identification Method and Its Application
on a Regional Map Based Analytical Platform
Xinxin Xu1 , Ziqiang Zeng2,3(B) , Yinhai Wang3 , and John Ash3

1
School of Tourism and Economic Management, Chengdu University,
xuxinxin@cdu.edu.cn
2
zengziqiang@scu.edu.cn
3
Smart Transportation Applications and Research Laboratory,
Department of Civil and Environmental Engineering,
University of Washington, Seattle, WA 98195, USA
Abstract. This paper aims to develop a crash counts by severity based

hotspot identification method by extending the traditional empirical
Bayes method to a generalized nonlinear model-based mixed multino-
mial logit approach. A new safety performance index and a new potential
safety improvement index are developed by introducing the risk weight
factor and compared with traditional indexes by employing four hotspot
identification evaluating methods. The comparison results reveal that
the new safety performance index derived by the generalized nonlinear
model-based mixed multinomial logit approach is the most consistent
and reliable method for identifying hotspots. Finally, a regional map
based analytical platform is developed by expanding the safety perfor-
mance module with the new safety performance index and potential
safety improvement functions.
Keywords: Hotspot identification · Crash severity · Safety performance

index · Potential safety improvement index · Regional map based ana-
lytical platform
1 Introduction
The identification of crash hotspots, also referred to as hazardous road locations,

high-risk locations, accident-prone locations, black spots, sites with promise, or
priority investigation locations, is the first step of the highway safety manage-
ment process [10]. Hotspot identification (HSID) is of great importance to land
transport authorities in their efforts to improve the safety of highways. There
is a fairly extensive body of literature focused on methods for HSID [1,2,11].
A Crash Counts by Severity-based Hotspot Identification Method 287
There are papers that discuss methods based on accident count or frequency [3],
papers that employ both accident rate (AR) and rate quality control [13], and
others that adopt the joint use of accident frequency and rate to flag sites with
promise [7]. To correct for the regression-to-the-mean bias associated with typi-
cal HSID methods [4], some researchers have suggested using the empirical Bayes
(EB) techniques [5]. This method combines clues from both the accident history
of a specific site and expected safety of similar sites, and has the advantage of
revealing underlying safety problems which otherwise would not be detected.
Some recent studies indicated that the severity of crashes should not be
neglected in the hotspot identification (HSID) process [12]. The hotspots cor-
responding to high crash risk locations can be quite different when considering
the crash frequency by different levels of crash severity. It is particularly impor-
tant to take into account crash severities in site ranking, because the cost of
crashes could be hugely different at different severity levels. This means that,
for instance, a road segment with higher frequency of fatal accidents may be
considered more hazardous than a road segment with fewer fatal accidents, but
more severe or minor injury accidents. Therefore, it is necessary to consider crash
severity when identifying hotspots.
While the equivalent property damage only (EPDO) method [15] is a way of
comparing severity types among each other, inconsistencies can occur when eval-
uating the HSID methods in different time periods, since the traditional EPDO
method overemphasizes sites with a low frequency of fatal or severe crashes [10].
As a result, a risk weight factor is developed in this research by combining the
average crash cost with the corresponding probability for each type of crash
severity. In order to develop a new safety performance index (SPI) and a new
potential safety improvement index (PSII), a generalized nonlinear model-based
mixed multinomial logit approach is introduced to extend the traditional empir-
ical Bayes method for estimating the probability of crashes in different severity
levels. The new method developed in this paper is compared with other hotspot
identification methods by employing four hotspot identification evaluating meth-
ods and applied on a regional map based analytical platform.
2 Crash Counts by Severity Prediction Model
The EB method requires the use of pertinent crash prediction models. For the
purpose of the study, in order to account for unobserved heterogeneity and non-
linear effects of variables to extract more complex relationship, a refinement of
the generalized nonlinear model-based (GNM-based) mixed multinomial logit
(MNL) approach developed by Zeng et al. [17] was used.
2.1 GNM-based Mixed MNL Model
In this research, three categories are considered for the crash severity (i.e., prop-
erty damage only (PDO) (k = 1), injury (k = 2), and fatal (k = 3)). The crash
severity type, denoted by Y , was the response variable, whereas contributing
288 X. Xu et al.
factors for roadway geometric characteristics, traffic characteristics, and weather

conditions were the independent variables denoted by xij (j = 1, 2, · · · , J), where
i denotes the observation and J denotes the number of independent variables.
Y is then defined as follows:
⎧
⎨ 1, if crash is PDO,
Y = 2, if crash is injury, (1)
⎩
3, if crash is fatal,
Ye and Lord [16] demonstrated that fatal crashes should be set as the baseline
severity for the mixed MNL model. To minimize the bias and reduce the vari-
ability of a model, in this paper, fatal crashes were used as the baseline severity
category for comparison with the other categories.
In order to account for the unobserved heterogeneity, let Ω = (ω1 , ω2 , ω3 ),
as discussed previously, and note that the Ω vector has a continuous density
function f (Ω|Γ ), where Γ is a vector of parameters charactering the density
function, ωk = [ωk1 , ωk2 , · · · , ωkJ ]T is the coefficient vector for the kth category
of the predictor vector. According to [6,9,14], the resulting mixed MNL crash
severity probabilities are as follows:

1
P r(Yi = K) = K−1 U ω +β f (Ω|Γ )dΩ, i = 1, 2, · · · , n, (2)
1 + k=1 e ki k k0

eUki ωk +βk0
P r(Yi = k) = K−1 U ω +β f (Ω|Γ )dΩ, i = 1, 2, · · · , n; k = 1.2, · · · , K − 1,
1 + k=1 e ki k k0
(3)
where P r(Yi = k) is the probability of crash severity type k; Uki =
[Uki1 (xi1 ), Uki2 (xi2 ), · · · , UkiJ (xiJ )] is the nonlinear predictor vector of obser-
vation i for contributing factors (i.e., roadway geometric characteristics, traffic
characteristics, weather conditions); βk0 is an intercept term specific to crash
severity type k. Since Uki is considered as a nonlinear predictor vector of observa-
tion i for contributing factors, Eqs. (2) and (3) are called the prediction functions
of the GNM-based mixed MNL approach.
2.2 Predicting Crash Counts in Different Severity Levels
Based on the statistical approaches available to address the unobserved hetero-

geneity [8], the expected crash density in a roadway segment i can be estimated
as below:
di = eUi ω+β0 f (ω|ϕ)dω, i = 1, 2, · · · , n, (4)
where di = μi /Li yi is the crash density of roadway segment i during a certain

time period; μi is the expected crash frequency (or rate) along segment i during a
certain time period; Li is the segment length in miles; yi is the time period length
(years) of crash frequency of roadway segment i; ω = [ω(1), ω(2), · · · , ω(J)]T is
the coefficient vector for Ui = [U1 (xi1 ), U2 (xi2 ), · · · , UJ (xiJ )] when estimating
the expected crash density.
According to Eqs. (2)–(4), the expected crash density for different severity
levels can be estimated as follows.
(1) Expected PDO crash density:
di1 = di · P r(Yi = 1)
U1i ω1 +β10
= eUi ω+β0 f (ω|ϕ)dω · e
2 i = 1, 2, · · · , n,
f (Ω|Γ )dΩ,
1+ k=1 eUki ωk +βk0
(5)
where di1 is the expected PDO crash density along segment i during a certain
time period.
(2) Expected injury crash density:
di2 = di · P r(Yi = 2)
U2i ω2 +β20
= eUi ω+β0 f (ω|ϕ)dω · e
2 i = 1, 2, · · · , n,
f (Ω|Γ )dΩ,
1+ k=1 eUki ωk +βk0
(6)
where di2 is the expected injury crash density along segment i during a certain
time period.
(3) Expected fatal crash density:
di3 = di · P r(Yi = 3)

= eUi ω+β0 f (ω|ϕ)dω · 1+
2 1
eUki ωk +βk0
i = 1, 2, · · · , n,
f (Ω|Γ )dΩ,
k=1
(7)
where di3 is the expected fatal crash density along segment i during a certain
time period. Based on Eqs. (5)–(7), a crash counts by severity based hotspot
identification (CCS-based HSID) method is developed in the following section.
3 Crash Counts by Severity Based Hotspot Identification

Method
In this section, a CCS-based HSID method is developed by employing the GNM-
based MNL approach discussed above. A new SPI and a new PSII are developed
by introducing the risk weight factor.
3.1 Safety Performance Index
Based on Eqs. (5)–(7), the EPDO crash frequency measure is modified and
employed to weight crashes according to severity (fatal, injury, and PDO) to
develop a combined crash density and severity score (CCDSS) for each site [15].
The weight factors are based on PDO crash costs. An EPDO value summarizes
the crash costs and severity.
In the calculations, weight factors were assessed from the crash cost estimates
developed by WSDOT in the Annual Collision Data Summary Reports (2011–
2014). Using average crash costs for motorways, fatal crashes ($2,227,851) have
a weight factor equal to 981, injury crashes ($20,439) have a weight factor equal
to 9, and PDO crashes ($2,271) have a weight factor equal to 1. However, if we
290 X. Xu et al.
only consider the average crash costs to be the weight factor, inconsistencies can
occur when evaluating the HSID methods in different time periods, since the
traditional EPDO method overemphasizes sites with a low frequency of fatal or
severe crashes [10]. As a result, a risk weight factor is developed in this research
by combining the average crash cost with the corresponding probability for each
type of crash severity. Let Fw denote the fatality risk weight factor, Iw , the
injury risk weight factor, and Pw , the PDO risk weight factor, then they are
defined by using the following equations:
cF · ηF cI · ηI
Fw = , Iw = , Pw = 1, (8)
cP ηP cP ηP
where cF = $2, 227, 851, cI = $20, 439, and cp = $2, 271 are the average costs for
fatal, injury, and PDO crashes; ηF , ηI , and ηP are the probabilities of occurrence
for fatal, injury, and PDO crashes.
Based on the preceding analysis, the expected CCDSS (ECCDSS) for road-
way segment i can be defined as:
ECCDSSi = di1 Fw + di2 Iw + di3 Pw , i = 1, 2, · · · , n. (9)
Equation (9) is regarded as a safety performance function (SPF). In fact, the

ECCDSS is an extension of the expected crash density based on the GNM.
When the fatality weight factor and the injury weight factor are both equal to
1 (i.e., Fw = Iw = 1), the ECCDSS will degrade to an expected crash density
based on the GNM. The EB method is a statistical method that combines the
observed crash frequency with the predicted crash frequency using the SPF to
calculate the expected crash frequency for a site of interest. The EB method
pulls the crash count towards the mean, accounting for the regression to the
mean (RTM) bias. A lot of studies have proved that the EB approach is the
most consistent and reliable method for identifying sites with promise [2,10]. In
this research, the EB method is employed to develop the new SPI as shown in
the following:
SPIi = λi ECCDSSi + (1 − λi )OCCDSSi , i = 1, 2, · · · , n, (10)
where OCCDSSi is the observed combined crash density and severity score
(OCCDSS) for roadway segment i and is defined as below:
OCCDSSi = σi1 Fw + σi2 Iw + σi3 Pw , i = 1, 2, · · · , n, (11)
where σi1 , σi2 , σi3 are the observed fatal, injury, and PDO crash density along
segment i during a certain time period respectively; λi is a weighting factor that
is calculated through the following equation:
1
λi = , (12)
1 + αi ECCDSSi
where αi is the over dispersion parameter, which is a constant for a given model
and is derived during the regression calibration process.
3.2 Potential Safety Improvement Index
The PSII was developed as the difference between the SPI and the ECCDSS, as
follows:
PSIIi = λi ECCDSSi + (1 − λi )OCCDSSi − ECCDSSi
(13)
= SPIi − ECCDSSi , i = 1, 2, · · · , n,
when the PSII value is greater than zero, a site experiences a higher combined
frequency and severity score than expected; when the PSII value is less than zero,
a site experiences a lower combined frequency and severity score than expected.
4 Case Study Data
This study was performed based on crash data records collected in Washington
State from January 2011 to December 2014 (i.e., a four-year period). The
data were obtained from the Washington State Department of Transportation
(WSDOT), Highway Safety Information System (HSIS), and the Digital Road-
way Interactive Visualization and Evaluation Network (DRIVE Net) platform
at the University of Washington (UW). Four major datasets are included in this
study: crash data, roadway geometric characteristics, traffic characteristics, and
weather conditions. These datasets detail all of the information regarding crash
frequency, locations, severities, roadway segment length, average number of lanes
(NOL), horizontal curve type (HCT), curvature of the segment (COS), average
width of outer shoulder (WOS), average width of inner shoulder (WIS), average
width of median (WM), dominant lane surface type (DLST), dominant outer
shoulder type (DOST), dominant inner shoulder type (DIST), dominant median
type (DMT), average speed limit (ASL), AADT, AADT per lane, road surface
conditions (RSC, i.e., dry, wet, snow/ice/slush), and visibility (good, bad).
In this research, we consider using the proportion of the crash frequency for
each type of severity based on the collected crash data of 21,396 road segments
along I-5, I-90, I-82, I-182, I-205, I-405 and I-705 in Washington to represent
the probability of crash occurrence for each severity level in this area. Based on
the crash counts after data quality control, the total number of crashes recorded
during the data collection period was 47,657, including 134 fatal crashes, 13,824
injury crashes, and 33,699 PDO crashes. Thus, we can calculate that ηF =
0.0028, ηI = 0.29, ηP = 0.7072; then, the values of the risk weight factors are
obtained by employing Eq. (8) as Fw = 3.884, Iw = 3.691, Pw = 1.
5 Test Results and Discussion
In order to demonstrate the effectiveness of the SPI and PSII as developed in

this research, they are compared with six reference performance indexes, which
include the expected crash density based on the conventional safety performance
function from the Highway Safety Manual (HSM) (i.e., NB GLM), expected crash
density based on the GNM, EB estimated crash density based on the NB GLM,
292 X. Xu et al.
Table 1. Results of site consistency test of various HSID methods
HSID method HSID method name Parameter Settings

index (h)
γ = 0.01 γ = 0.05
SCTh,t Period 1 SCTh,t+1 SCTh,t Period SCTh,t+1
(2011–2012) Period 2 1 (2011–2012) Period 2
(2013–2014) (2013–2014)
1 SPI 22943.06 21521.79 46310.12 44251.68
2 PSII 22198.11 20879.23 45389.79 43921.56
3 CD (NB GLM) 21036.33 20105.38 43897.54 42868.83
4 CD (GNM) 21901.86 20993.89 44890.67 43571.45
5 EB CD (NB GLM) 21896.77 20981.31 45303.46 43784.09
6 EB CD (GNM) 22856.34 21349.04 45987.98 44012.49
7 ARP (NB GLM) 20131.67 19034.25 43015.82 42106.73
8 ARP (GNM) 20358.78 19734.55 43823.17 42871.29
EB estimated crash density based on the GNM, ARP based on the NB GLM,
and ARP based on the GNM.
Cheng and Washington [2] have developed four new evaluation tests for HSID.
In this research, the site consistency test, method consistency test, total rank
differences test, and the total score test are employed to evaluate the effectiveness
of the developed safety performance indexes and reference performance indexes.
The evaluation experiment uses the following procedure, which closely mimics
how reactive safety management programs are conducted in practice:
(1) For the purpose of comparing alternate HSID approaches, the 4-year acci-
dent data were separated into two periods, Period 1 (Year 2011–2012) and
Period 2 (Years 2013–2014).
(2) Road sections (intersections, ramps, two-lane rural roads, etc.) are segre-
gated so that the safety of similar sites can be fairly compared. In this eval-
uation, the analysis is based on the analysis of nine functional classifications
of road sections.
(3) For each HSID method, similar road sections are sorted in descending order
of estimated safety (noting that the four HSID methods rank sites according
to different criteria).
(4) Sections with the highest rankings are flagged as hotspots (in practice these
sites will be further scrutinized). Typically, a threshold is assigned according
to safety funds available for improvement, such as the top 1% of sites. In this
evaluation, both the top 1% and 5% of the locations are used as experimental
values.
5.1 Site Consistency Test
The site consistency test (SCT) measures the ability of an HSID method to
consistently identify a high-risk site over repeated observation periods. The test
rests on the premise that a site identified as high risk during time period i should
also reveal an inferior safety performance in a subsequent time period t+1, given
that the site is in fact high risk and no significant changes have occurred at the
site. The method that identifies sites in a future period with the highest crash
frequency is the most consistent. In this research, the SPI developed above is
employed as the safety performance criterion in the subsequent time period. The
test statistic is given as:
n

SCTh,t+1 = SPIq,h,t+1 , h = 1, 2, · · · , H, (14)
q=n−nγ+1
where h is the HSID method index being compared; n is the total number of
roadway segment, γ is the threshold of identified hotspots (e.g., γ = 0.01 corre-
sponds with top 1% of n roadway segments identified as hotspots, and nγ is the
number of identified hotspots).
From the site consistency test, it is shown in Table 1 that the SPI method
outperforms other HSID methods in identifying both of the top 1% and 5% of
hotspots with highest SCT values, 21521.79 and 44251.68, in Period 2, followed
closely by the EB CD (GNM) method. The ARP (NB GLM) performs the worst
in both cases, with the identified hotspots experiencing the lowest number of SCT
values, say, 19034.25 and 42106.73, respectively (although the ARP is based on
reduction potential, so the total count can be misleading).
5.2 Method Consistency Test
The method consistency test (MCT) evaluates a method’s performance by mea-

suring the number of the same hotspots identified in both time periods. It is
assumed that road sections are in the same or similar underlying operational
state and their expected safety performance remains virtually unaltered over
the two analysis periods. With this assumption of homogeneity, the greater the
number of hotspots identified in both periods the more consistent the perfor-
mance of the HSID method. The test statistic is given as:

MCTh = {sn−nγ+1 , sn−nγ , · · · , sn }h,t {sn−nγ+1 , sn−nγ , · · · , sn }h,t+1 , h = 1, 2, · · · , H,
(15)
here, only segments {sn−nγ+1 , sn−nγ , · · · , sn } identified in the top thresh-

old γ are compared. Table 2 shows the number of similarly identified hotspots
identified by alternate HSID methods over the two periods. The SPI method
is superior in this test by identifying the largest number of the same hotspots
in both cases of γ = 0.01 and γ = 0.05, with 124 and 546 roadway segments,
respectively. In other words, the SPI method identified 124 segments in 2011–
2012 that were also identified as hotspots in 2013–2014. The ED CD (GNM),
which performs slightly better than the ED CD (GLM) method, places 2nd with
identifying 118 consistent hotspots (in the case of γ = 0.01) and 472 consistent
hotspots (in the case of γ = 0.01). The ARP (NB GLM) performs worst, with
the lowest number of consistent hotspots identified in the two periods. Again,
294 X. Xu et al.
the SPI method outperforms the other HSID methods. Also shown in Table 2
are differences between percentages (shown in the parentheses) of Column 3
and Column 4 for the eight methods. There is a consistent drop in percentages
as threshold values drop. The explanation is that the top segments suffer from
greater random fluctuations in crashes, and thus the higher is the threshold, the
larger are the random fluctuations and the likelihood of not being identified in
a prior period.
Table 2. Results of method consistency test of various HSID methods
HSID method index (h) HSID method name Parameter settings

γ = 0.01 γ = 0.05
1 SPI 124 (60.7%) 546 (53.5%)
2 PSII 103 (50.5%) 452 (44.3%)
3 CD (NB GLM) 92 (45.1%) 406 (39.8%)
4 CD (GNM) 101 (49.5%) 423 (41.5%)
5 EB CD (NB GLM) 109 (53.4%) 441 (43.2%)
6 EB CD (GNM) 118 (57.8%) 472 (46.3%)
7 ARP (NB GLM) 83 (40.7%) 382 (37.5%)
8 ARP (GNM) 88 (43.1%) 393 (38.5%)
5.3 Total Rank Differences Test

The total rank differences test (TRDT) takes into account the safety perfor-
mance rankings of the road sections in the two periods. The test is conducted by
calculating the sum of the total rank differences of the hotspots identified across
the two periods. The smaller the total rank difference, the more consistent the
HSID method. The test statistic is given as:
n

TRDTh = |R(qh,t ) − R(qh,t+1 )|, h = 1, 2, · · · , H, (16)
q=n−nγ+1
where R(qh,t ) is the rank of segment q in period t for method h. The difference
in ranks is summed over all identified segments for threshold level γ for period t.
Table 3 illustrates that the SPI method is superior in the total rank differences
test. In both the γ = 0.01 and γ = 0.05 cases, the SPI method has significantly
smaller-summed ranked differences, by about 22.6% (in the case of γ = 0.01) and
16.9% (in the case of γ = 0.05) compared with the EB CD (GNM), and by about
75.1% (in the case of γ = 0.01) and 77.4% (in the case of γ = 0.05) compared
with the ARP (NB GLM). This result suggests that the SPI method is the
best HSID method (of the eight evaluated here) for ranking roadway segments
consistently from period to period.
5.4 Total Score Test
The total score test (TST) combines the site consistency test, the method con-
sistency test, and the total rank difference test in order to provide a synthetic
index. The test statistic is given as:

SCTh,t+1 h −minh {TRDTh }
TSTh = 100 3 ( maxh {SCTh,t+1 }
) + ( maxMCT h
h {MCTh }
) + (1 − TRDT
maxh {TRDTh } ) ,
h = 1, 2, · · · , H,
(17)
where the test assumes that the SCT, MCT, and TRDT have the same weight.
The former three tests provide absolute measures of effectiveness, whereas the
total score test gives an effectiveness measure relative to the methods being
compared. If method h performed best in all of the previous tests, the TST value
is equal to 100. If method h performed worst in all of the tests, the TST value
is positive since all three components of the test have a positive value. Indeed,
SCT and MCT, which should be maximized by the HSID methods, are weighted
Table 3. Results of total rank differences test of various HSID methods

γ = 0.01 γ = 0.05
1 SPI 2354 10237
2 PSII 3031 12798
3 CD (NB GLM) 3672 14781
4 CD (GNM) 3298 13587
5 EB CD (NB GLM) 3158 13016
6 EB CD (GNM) 2887 11973
7 ARP (NB GLM) 4123 18167
8 ARP (GNM) 3887 17105
Table 4. Results of total score test of various HSID methods

γ = 0.01 γ = 0.05
1 SPI 100 100
2 PSII 87.89 89.31
3 CD (NB GLM) 78.55 82.07
4 CD (GNM) 85.37 85.83
5 EB CD (NB GLM) 88.63 88.14
6 EB CD (GNM) 93.81 92.12
7 ARP (NB GLM) 70.82 73.82
8 ARP (GNM) 75.16 77.02
296 X. Xu et al.
in relation to the maximum values in the tests, whereas TRDT, which should be
minimized by the HSID methods, is weighted in relation to its difference from
the minimum value in the test. Table 4 illustrates the results of total score test
of the eight HSID methods, in which SPI performed best in both γ = 0.01 and
γ = 0.05 cases, and was followed closely by EB CD (GNM) method with 93.81
score (in the case of γ = 0.01) and 92.12 score (in the case of γ = 0.05). ARP
(NB GLM) performed the worst in both cases, with 70.82 score and 73.82 score
respectively.
Overall, the four tests reveal that the SPI method is the most consistent
and reliable method for identifying hotspots. Although it can only be applied
to roadway segments where the crash data for different levels of severity are
available, with the rapid development of intelligent transportation systems and
data collection technologies, this method could become quite useful in identifying
high-risk road sites. On several criteria, the SPI outperforms other methods by
a wide margin. This evaluation suggests that the SPI method (of the methods
compared) has a potential to become the industry standard.
6 Application on Regional Map Based Analytical

Platform
A regional map based analytical platform was developed on the DRIVE Net sys-
tem to highlight the methodology developed under this project. Ultimately, the
existing safety performance analysis function under the “Safety Performance”
module was expanded. The SPI developed in the preceding is used to color-code
the regional map based on safety performance. The PSII is employed to highlight
potential safety improvements on the map. By combining the two indices on the
regional map, one can easily identify accident hotspots and the key influencing
factors to consider in an improvement package.
The interface of the safety performance module in the regional map based
analytical platform is illustrated in Fig. 1. There are three sub-functions imple-
mented on this panel: Incident Frequency (NB GLM), Estimated Crash Mean
and Potential Safety Improvement Index (ARP NB GLM). The new SPI and
PSII were added as expended safety performance analysis options. As stated
in the previous modeling part of this report, within a selected time range and
corridor, the SPI shows a more comprehensive view of safety performance on a
given corridor. The accident/incident data is from Washington Incident Tracking
System (WITS) database. The SPI level ranges from Level A to Level F, where
Level A (light green) corresponds to the highest safety performance and Level
F (dark red) corresponds to the lowest safety performance expected as shown in
Fig. 1.
The PSII implements the EB method in the modeling part. In this function,
both the historical incident data and the characteristics of the selected corridor
are used as model inputs. The output format still uses the six different colors
representing Level A to Level F to show the potential safety improvement index
on the map, where Level A shows the segment has the least potential to improve
SPI Levels
Level A 1/6 percentile

Level B 2/6 percentile
Level C 3/6 percentile
Level D 4/6 percentile
Level E 5/6 percentile
Level F 6/6 percentile
Fig. 1. SPI level ranges from Level A to Level F in the safety performance module
Fig. 2. An example of the PSII function
its safety, and Level F shows the segment has the most potential to improve its
safety. Figure 2 shows an example of this function.
7 Conclusions
A CCS-based HSID method is developed by extending the traditional EB method
to a GNM-based mixed MNL approach in this paper. A new SPI and a new
PSII are developed by introducing the risk weight factor and compared with
traditional indexes by employing four HSID evaluating methods, including the
site consistency test, method consistency test, total rank differences test, and
the total score test. The test results showed that the new SPI derived by the
298 X. Xu et al.
GNM-based mixed MNL approach is the most consistent and reliable method
for identifying hotspots. Finally, the new CCS-based HSID method was applied
on a regional map based analytical platform.
Acknowledgements. This research was supported by the Youth Program of National

Natural Science Foundation of China (Grant No. 71501137), the General Programs of
China Postdoctoral Science Foundation (Grant No. 2015M572480), the International
Postdoctoral Exchange Fellowship Program of China Postdoctoral Council (Grant No.
20150028), the project of Research Center for System Sciences and Enterprise Devel-
opment (Grant No. Xq16B05), and also funded by Sichuan University (Grant No.
skqy201647). The authors would like to give our great appreciation to the editors and
anonymous referees for their helpful and constructive comments and suggestions, which
have helped to improve this article.
References
1. Cheng W, Washington S (2005) Experimental evaluation of hotspot identification
methods. Accid Anal Prev 37:870–881
2. Cheng W, Washington S (2008) New criteria for evaluating methods of identifying
hot spots. Transp Res Rec 2083:76–85
3. Deacon J, Zegeer C, Deen R (1975) Identification of hazardous rural highway
locations. Transp Res Rec 543:16–33
4. Hauer E (1980) Bias-by-selection: overestimation of the effectiveness of safety coun-
termeasures caused by the process of selection for treatment. Accid Anal Prev
12:113–117
5. Hauer E, Persaud B, Smiley A et al (1991) Estimating the accident potential of
an Ontario driver. Accid Anal Prev 23:133–152
6. Hensher D, Greene W (2003) The mixed logit model: the state of practice. Trans-
portation 30:133–176
7. Laughland J, Haefner L, Hall J et al (1975) Methods for evaluating highway safety
improvements. NCHRP 162, Transportation Research Board
8. Mannering F, Shankar V, Bhat C (2016) Unobserved heterogeneity and the statis-
tical analysis of highway accident data. Anal Methods Accid Res 11:1–16
9. McFadden D, Train K (2000) Mixed MNL models for discrete response. J Appl
Econometrics 15:447–470
10. Montella A (2010) A comparative analysis of hotspot identification methods. Accid
Anal Prev 42:571–581
11. Park B, Lord D, Lee C (2014) Finite mixture modeling for vehicle crash data with
application to hotspot identification. Accid Anal Prev 71:319–326
12. Qu X, Meng Q (2014) A note on hotspot identification for urban expressways. Saf
Sci 66:87–91
13. Stokes R, Mutabazi M (1996) Rate-quality control method of identifying hazardous
road locations. Transp Res Rec 1542:44–48
14. Train K (2003) Discrete Choice Methods with Simulation. Cambridge University
Press, Cambridge
15. Washington S, Haque M, Oh J et al (2014) Applying quantile regression for model-
ing equivalent property damage only crashes to identify accident blackspots. Accid
Anal Prev 66:136–146
16. Ye F, Lord D (2011) Investigation of effects of underreporting crash data on three

commonly used traffic crash severity models multinomial logit, ordered probit, and
mixed logit. Transp Res Rec 2241:51–58
17. Zeng Z, Zhu W, Ke R et al (2017) A generalized nonlinear model-based mixed
multinomial logit approach for crash data analysis. Accid Anal Prev 99:51–65
Comparison Between K-Means and Fuzzy
C-Means Clustering in Network Traffic
Activities
Purnawansyah1 , Haviluddin2(B) , Achmad Fanany Onnilita Gafar3 ,

and Imam Tahyudin4
1
Faculty of Computer Science, Universitas Muslim Indonesia, Makassar, Indonesia
2
Faculty of Computer Science and Information Technology,
Mulawarman University, Samarinda, Indonesia
haviluddin@gmail.com
3
State Polytechnic of Samarinda, Samarinda, Indonesia
4
Information System Department, STMIK AMIKOM, Purwokerto, Indonesia
Abstract. A network traffic utilization in order to support teaching

and learning activities are an essential part. Therefore, the network
traffic management usage is requirements. In this study, analysis and
clustering network traffic usage by using K-Means and Fuzzy C-Means
(FCM) methods have been implemented. Then, both of method were
used Euclidean Distance (ED) in order to get better results clusters.
The results showed that the FCM method has been able to perform
clustering in network traffic.
Keywords: Network traffic · K-Means · Fuzzy C-Means · Clustering
1 Introduction
A bandwidth usage monitoring management in a network traffic at university is
indispensable. Therefore, a recording and analysis of bandwidth usage by net-
work administrators is very beneficial. It aims to make use of the bandwidth
can be well controlled, stable access, and gives users convenient access. There-
fore, mapping or cluster of bandwidth usage in order to support the network
administrator performance analysis is required.
In this study, the cluster method based on intelligence algorithm are pro-
posed in bandwidth usage, such as Self-Organizing Maps (SOM), K-Means,
Fuzzy C-Means, etc. For that reason, these algorithms have been widely used by
researchers in order to solve cluster data problems in a variety of fields, includ-
ing economics [5], supply chain [2], engineering [6], hydrology [1,4], internet and
social media [3,8], pattern recognition [7], and so forth. Many research results
have shown that the algorithms are able to provide accurate information in solv-
ing the clustering problem.

Comparison Between K-Means and Fuzzy C-Means Clustering 301
Afterward, two intelligence algorithms is K-Means and Fuzzy C-Means

(FCM) for cluster bandwidth usage data has been implemented. The purpose
of this study are compared the feasibility of two clustering methods and how it
works in the real world problems particularly subject on network traffic. Thus,
the expected results of network traffic clustering are coordination schemes that
support resource management [8]. Furthermore, this paper will apply two mod-
els, namely K-Means and Fuzzy C-Means (FCM) that have been developed and
compared in order to cluster the network traffic usage. Section 2 describes the
architectures of K-Means and FCM clustering models are proposed. Section 3
describes the analysis and discussion of the results. Finally, conclusions are sum-
marized in Sect. 4.
2 Research Method
In this section, a brief information of K-Means and FCM models are presented.
2.1 Principle of K-Means Method
K-means clustering is an unsupervised learning classified. This algorithm is based

on the determination of the distance between the centroid and the training data.
Then, the number of cluster centroid based on the number desired. Meanwhile,
the initialization centroid randomly generated by considering the data training.
In other words, the centroid should be in the training data space. Then, a couple
of training data is from the attributes of the data patterns to be planned. In each
iteration, the distance of each training data with each centroid to be calculated.
It means that any training data will have a centroid distance. At the same time,
members of the cluster indicated by the smallest distance from the corresponding
centroid. Then, a new centroid value is calculated based on the average value
of each member of the cluster. If the cluster member does not change then the
iteration is stopped, Fig. 1.
The following data clustering techniques using the K-Means algorithm as
follows.
• Determine the number of clusters K.

• The Initialization of K cluster centers can be done randomly and used as
initial cluster centers.
• Allocate all data/objects to the nearest cluster. The proximity of the two
objects is determined by the distance of the object. To calculate the distance
of all the data to the cluster center point using the Euclidean distance theory
formulated as follows:

T(x,y) = (T1x − T1y )2 + (T2x − T2y )2 + · · · + (Tkx − Tky )2 , (1)
where, T(x,y) is distance data of x to the cluster center of y; Tkx is i-data on

attribute data of k; Tky is the center point of j on the attribute of k.
302 Purnawansyah et al.
Fig. 1. K-Means algorithm
• Recalculate the center of the cluster with the new cluster membership. This
is computed by determining the centroid/center cluster.
• Assign each object to put on the new cluster center, if the center of the cluster
changed, back to 3, otherwise clustering is complete.
• Analyze the results in the clustering process.
2.2 Principle of Fuzzy C-Means Method
Other advanced techniques clustering in machine learning model come from the
Dunn in 1973, and improved by Bezdek in 1981, called Fuzzy C-Means (FCM)
clustering was developed by Dunn in 1973, and improved by Bezdek in 1981. In
principle, FCM clustering process is based on a partition of a set of data into a
number of clusters with minimum similarity between different clusters [2]. Since
the introduction of the fuzzy set theory in 1965 by Zadeh, it has been applied
in a variety of fields. FCM is a flexible fuzzy partition that an improvement of
common C-Means algorithm [5]. At FCM, each feature vectors valued between
[0–1] by using the membership function, because FCM is based on the criteria of
distances numbers between clusters. In other words, FCM clustering is based on
Point-Prototype Clustering Model with output centroid most optimal partition.
Where, partition optimization centroid is obtained by minimizing the objective
function. The formula is given by:

N
C
q 2
JFCM (U, V ) = (uij ) (dji ) ,
j=1 i=1
where
U = Fuzzy datasets K-partition;
V = Set of prototype centroid;
V = {v1 , v2 , · · · , vC } ⊂ RP ;
2 2
2 2
(dji ) = xj − vi = xj (row) − vi (row) + xj (col) − vi (col) . (2)
Euclidean Distance between xj and vi ;

X = {x1 , x2 , · · · , xn } ⊂ RP ;
vi = Centroid cluster to i;
uij = Membership level xj in cluster to i;
N = Total data;
C = Total cluster;
q = Fuzzifier parameter, q > 1.
Afterward, all of the objects in each cluster has a certain degree of proxim-
ity or similarity. Meanwhile, the FCM processes consists of five stages. First, to
determine the cluster is set become a center cluster location marker on aver-
age for each cluster. Second, calculate the distance between feature vector (X)
and the centroid vector (V )[X → V ]. In this experiment, Euclidean Distance
(ED) was implemented. Third, Calculate membership level. Fourth, Calculate
new centroid. Lastly, recalculated the new centroid, if criterion between [0–1] is
reached then stop the iteration. In this study, the FCM Clustering algorithm is
as follows:
Step 1. Initialization Vector centroid, vi (prototypes).

Step 2. Calculate the distance between feature vector (X) and the centroid vec-
tor (V )[X → V ]. Feature vectors with the closest distance to one of the
centroid vectors then expressed as a cluster member.
Step 3. Calculate membership level of all feature vectors in all clusters by using
the formula:
1/(q−1)
1
1 2
(dji )
K (dji )2 1/(q−1) K 1 1/(q−1)
uij = = . (3)
k=1 (djk )2 k=1 (djk )2
Step 4. Calculate new centroid using Eq. (4).

N q
j=1 (uij ) Xj
V̂i = N q
. (4)
j=1 (uij )
Step 5. Recalculate the step 4, uij → ûij If, max |uij − ûij | < ε, where ε is
ij
termination criteria between 0 and 1. Then, the iteration process is
stopped. If not go back to step 5. The FCM algorithm can be seen in
Fig. 2.
Fig. 2. Fuzzy C-means algorithm
(1) Initialization Centroid

Centroid is used as a central member of the cluster that expected to be an
average of all the cluster members. If k cluster is formed, the number k cen-
troid is necessary. There are various ways of doing initialization centroid, such
as the random generation and the average spread. For example, interval data
is [mindata· · · maxdata]. Centroid random initialization is done with formula
C = mindata + (maxdata − mindata) × rand (). In this study, the initialization
centroid for K-Means and FCM are used the average spread then the formula as
follows:
cluster number = k
s = (maxdata − mindata) /k
m = mean (data)
C = [(m − s∗ (k)) (· · ·) (m − s∗ (1)) (m) (m + s∗ (1)) (· · ·) (m + s∗ (k))]
(2) Calculated Distance Neighborhoods

Neighborhoods distance is the distance between a centroid to each data that
expressed by Eq. (2). According to [9] “an important step in most clustering
is to select a distance measure, which will determine how the similarity of two
elements is calculated”. Thus, there are some varieties of distance function in
clustering, including Euclidean distance (ED), Manhattan distance, Mahalanobis

distance, and Hamming distance. In this study, K-Mean and FCM clustering
using ED distance.
(3) Generated New Centroid
In this study, generating new centroid both of K-Means and FCM are based on
the average all members cluster values. As an illustrated, for training data is as
follows.
X x1 x2 x3 x4 x5 x5
T = = .
Y y1 y2 y3 y4 y5 y6
Next, the clustering results in an iteration is as follows.
c1 101100
G= = .
c2 010011
Then, the new centroid is as follows.

x1 + x3 + x4 x2 + x5 + x6
c1 = c2 = .
y1 + y3 + y4 y2 + y5 + y6
Meanwhile, in FCM, generating new centroid by using the COA (Center of

Area) formula, Eq. (4).
2.3 Datasets
In this study, 152 days (from January–May 2016) daily network traffic usage of
four client datasets from ICT unit were captured. Then, the data are analyzed
by using MATLAB R2013b. The real dataset can be seen in Table 1 (Fig. 3).
Fig. 3. Plot of network traffic per days

Table 1. The real of daily network traffic data (January–May 2016)
Rectorate Forestry Science Economic No Rectorate Forestry Science Economic

0.341 0.391 0.343 0.332 77 0.39 0.365 0.402 0
0.08 0.195 0.061 0.174 78 0.392 0.271 0.358 0.2
0.069 0.468 0.042 0.212 79 0.385 0.32 0.483 0.225
0.097 0.586 0.048 0.007 80 0.08 0.311 0.237 0.004
0.564 0.687 0.352 0.362 81 0.107 0.205 0.069 0.043
0.429 1 0.68 0.476 82 0.415 0.325 0.508 0.348
0.411 0.705 0.467 0.259 83 0.436 0.348 0.485 0.275
0.423 0.569 0.465 0.212 84 0.433 0.312 0.384 0.291
0.379 0.623 0.364 0.447 85 0.331 0.409 0.358 0.203
0.078 0.496 0.044 0.167 86 0.099 0.28 0.115 0.001
0.057 0.345 0.028 0.007 87 0.04 0 0 0.001
0.659 0.583 0.444 0.41 88 0.065 0 0 0.001
0.642 0.724 0.414 0.384 89 0.401 0.384 0.441 0.319
0.471 0.663 0.48 0.556 90 0.412 0.396 0.473 0.324
0.421 0.597 0.295 0.452 91 0.483 0.393 0.413 0.282
0.354 0.45 0.394 0.771 92 0.441 0.501 0 0.227
0.043 0.372 0.119 0.21 93 0.411 0.308 0.09 0.375
0.059 0.269 0.054 0.313 94 0.081 0.131 0.065 0.007
0.507 0.569 0.506 0.46 95 0.045 0.188 0.039 0.116
0.499 0.51 0.445 0.612 96 0.387 0.182 0.166 0.199
··· ··· ··· ··· ··· ··· ··· ··· ···
··· ··· ··· ··· ··· ··· ··· ··· ···
0.062 0.213 0.034 0 150 0.06 0.257 0.056 0.055
0.386 0.364 0.457 0.001 151 0.082 0.228 0.014 0.029
0.373 0.289 0.264 0 152 0.362 0.259 0 0.232
3 Results and Analysis
This section presents the empirical work and compares the experimental results
of the K-Means and FCM algorithms on highest average usage network traffic
problems. The performances are measured by the objective function value given
by Eq. (1).
3.1 Analysis of K-Means
In this experiment, the scheme used is to classify the training data into 3, 4, and
5 cluster grouping patterns in order to observe the good cluster. The average
data value is used as one of the centroid values. Then, for another centroid value
is determined randomly in the space of data training. The results of K-Means
are shown in Fig. 4.
Fig. 4. Plot of clustering network traffic results using k-means
3.2 Analysis of Fuzzy C-Means
In this experiment, the scheme used is to classify the training data into 3, 4, and
5 cluster grouping patterns in order to observe the good cluster. The average
data value is used as one of the centroid values. Then, for another centroid value
is determined randomly in the space of data training. In this test, the FCM
Fig. 5. Plot of clustering network traffic results using fuzzy C-means
initial parameter values were implemented, (a). Number of clusters 3, 4, and 5;

(b). Maximum iterations = 50; and (c). Stopping Criteria (ξ) = 10−5 . In this
research, applied testing scheme is the amount of usage per day of data network
traffic from the four parts (rectorate, forestry, science, and economic). Then, the
Table 2. Results comparison K-means and FCM methods
Cluster Methods Highest cluster Average of network Highest of network

members number traffic usage traffic usage month
3 KCM 18 2.4076 2
FCM 18 2.4076 2
4 KCM 14 2.5149 2
FCM 17 2.4412 2
5 KCM 14 2.5149 2
FCM 16 2.4703 2
highest, middle and lowest data values are obtained. The results of FCM are
shown in Fig. 5.
In this study, the accuracy level in order to get the centroid value by using
FCM method is quite careful than K-Means method, Table 2.
4 Conclusion
In this paper we presented a comparison of K-Means and FCM methods then
compared the centroid accuracy by using various performance criteria. This
research was used network traffic usage from four units; rectorate, forestry, sci-
ence, and economic. In clustering, the examining of 1 parameter of centroid
value is 3 centroid values. Our experiment showed that the FCM method is
better results analysis in clustering than K-Means. Nevertheless, we have also
concluded that FCM algorithm was slower than K-Means. As future work, an
optimizing methods in order to get good accuracy between centroids is proposed.
References
1. Arroyo A, Herrero A et al (2016) Analysis of meteorological conditions in spain by
means of clustering techniques. J Appl Logic. doi:10.1016/j.jal.2016.11.026.
2. Bai C, Dhavale D, Sarkis J (2016) Complex investment decisions using rough set
and fuzzy c-means: an example of investment in green supply chains. Eur J Oper
Res 248(2):507–521
3. Carvalho L, Barbon S Jr et al (2016) Unsupervised learning clustering and self-
organized agents applied to help network management. Expert Syst Appl 54(C):29–
47
4. Elaal A, Hefny H, Elwahab A (2010) Constructing fuzzy time series model based
on fuzzy clustering for a forecasting. J Comput Sci 6(7):735–739
5. Huang CK, Hsu WH, Chen YL (2013) Conjecturable knowledge discovery: a fuzzy
clustering approach. Fuzzy Sets Syst 221(2):1–23
6. Pandit YP, Badhe YP et al (2011) Classification of Indian power coals using k-means
clustering and self organizing map neural network. Fuel 90(1):339–347
7. Stetco A, Zeng XJ, Keane J (2015) Fuzzy c-means++: fuzzy c-means with effective
seeding initialization. Expert Syst Appl 40(21):7541–7548
8. Sucasas V, Radwan A et al (2016) A survey on clustering techniques for cooperative
wireless networks. Ad Hoc Netw 47(C):53–81
9. Velmurugan T (2014) Performance based analysis between k-means and fuzzy c-
means clustering algorithms for connection oriented telecommunication data. Appl
Soft Comput 19(6):134–146
RDEU Evolutionary Game Model
and Simulation of the Network Group Events
with Emotional Factors
Guoqiang Xiong(B) , Xian Wang, Ying Yang, and Yuxi Liu
School of Business Administration, Xi’an University of Technology,

Xi’an 710054, People’s Republic of China
xgq168@163.com
Abstract. At present, network communication becomes more efficient

and geographical restrictions barely exist. However, the network mass
incidents tend to be multiple and frequent, which are mostly unexpected.
In this paper, we point at studying the limitations of the traditional game
theory in the players’ rational assumptions, and consider the influence
of the irrational factors on the netizens behavior. Based on RDEU the-
ory and evolutionary game theory, a RDEU evolutionary game model of
group events on network is constructed. In addition, we studied the evo-
lution mechanism of game Nash equilibrium under different emotional
conditions and the numerical simulation of the model, which is carried
out by using the numerical simulation method of Matlab. After that, the
evolutionary equilibrium state of the game between the netizens in dif-
ferent emotional states of the network group events is analyzed. Finally,
research results show that emotional factors have a great impact on the
network behavior of group events; when the emotions of netizens are
pessimistic, the game participants are more inclined to confrontational
behavior, and will have both the conflicting herd effect, more prone to
network mass incidents. In this case, it is more likely to produce group
events on network.
Keywords: Network group event · RDEU evolutionary game model ·

Emotion · Simulation
1 Introduction
Generally, the whole Society is in the period of social transformation filled with
various contradictions and social conflicts, including the increasing number of
network mass incidents and the negative influence from that. We often find a
simple network information will attract tens of millions of click rate in a short
period, an infectious emotion may spread out of control like a virus, thus caus-
ing emotional excitement. That is to say, the development of the group events
on networks is often due to the interaction between social and human factors.
312 G. Xiong et al.
The evolution of this interaction has the derivative harmfulness, easily leading
to systemic economic and social crisis. Thus, social stability is facing a severe
challenge. Therefore, keeping in touch with the emotional state of netizens in
time is of great significance to grasp the law of the evolution in the behavior of
netizens group and monitor the occurrence of early warning derivative events.
At present, a majority of researchers of the network mass incidents focused on
the aspects of the network conflict. Morales [5] believed that network mass inci-
dents refer to the large-scale pooling of netizens’ opinions in the network for spe-
cific hotspot problems, thus affecting the real-life group events. Levine [2] argued
that cyberspace disputes and conflicts, which expand the interaction of people,
can activate the public awareness of the society in order to create a new public
realm, and provide new environmental resources for democracy, effectively avoid
autocratic. In addition, there are some research focus on the function of net-
work conflict, the conflict on real society and democratic system. Post [6] pointed
out that the traditional legal control tools are confronting severe network con-
flict challenges. Technicians cannot depends on the traditional management pol-
icy to effectively restrict network environment and reconstruct the rules’ system of
cyberspace. Qiu [7] summarized three types of network mass incidents: First, some
incidents occur in the Internet but exerts substantial influence on the real society;
the second type of incidents happens in the reality with the virtual online organiza-
tion as the intermediary; the third one happens through offline interaction that is
based on the online communication. Some scholars proposed corresponding strate-
gies to solve the problem about unexpected group incidents from different perspec-
tives. From the view of communication studies, Zhu et al. [13] emphasized that the
fundamental effect of faith on the public opinion and the guidance of the “opin-
ion leader” for the direction of public opinion. From the social and psychological
point of view to investigate the social struggle, Sidney et al. [10] argued it was the
existence of a sense of deprivation of the individual that resulted in resentment,
thus forming a collective behavior. It follows that the present scholars’ research
on network unexpected group incidents mostly applies the qualitative, canonical
approaches and comes from the single perspective but seldom establishes math-
ematical models based on the theories of management science to make relevant
research. At present, a few scholars as Liu [3,4], Deng et al. [1], Xie [12], Wang
[11] and Sun [9] made research on traditional unexpected group incidents. Aim-
ing at the conflict evolutionary mechanism of unexpected group incidents, they
established the models of unexpected group incidents and initially revealed the
mechanism of such incidents. Although evolutionary game models of unexpected
group incidents deepened the research concerning the conflict evolutionary mech-
anism by dividing up the social groups into advantaged groups and vulnerable
groups, they did not consider the influence of emotional factor on the state of evo-
lutionary game equilibrium in decision-making in the game.
All in all, the study of the evolution mechanism of the group events on net-
works is still in its infancy, and only qualitative methods have been used to
analyze the causes, types and countermeasures of the events. There are little
research on the evolution mechanism of the group events on networks, especially
RDEU Evolutionary Game Model 313
we lack the direct research on game model of endogenous factors like emotion
to further explore the evolution of the event. Based on the Quiggin’s RDEU
theory (Rank-dependent expected utility) [8], this paper constructs the RDEU
evolutionary game model of the group events on networks and studies the evolve-
ment process of netizens’ strategy selection under different emotional states, and
provides emergency decision support for policy selection of the group events on
networks. There are four sections in this paper. In the second section, we intro-
duce the emotional function which reflects the psychological activities of netizens
and establish the RDEU game model of the behavior mechanism of the group
events on networks. The third section, according to the different emotions of neti-
zens, adopts Matlab’s numerical simulation analysis of evolution equilibrium on
the netizens RDEU game model under six evolutionary scenarios (for example,
netizens of one party are rational, while the other party is pessimistic, netizens
are both pessimistic, etc.). The fourth section is the summary of the full text.
2 The RDEU Evolutionary Game Model among Netizen

Group
2.1 The Method of Constructing the Netizen Game Model
In the network unexpected group incident, the strategy sets of the netizen groups
can be divided into: “fight” (represented by F ) and “peace” (indicated by P ),
the former refers to radical articles or comments while the latter refers to quiet
observation of the course of events without taking any radical actions. In that
game, there are three different situations: (1) When both of the two groups of
netizens adopt the sheep-flock strategy to “F ”; the influence of the incident may
be accordingly expanded, and both of interactive objects may respectively gain
interaction revenues represented by “V ”; extraneous revenues is implied by “S”;
paying the costs is implied by “C”. Then, both sides should pay the costs and
make extraneous and interaction revenues; (2) When two netizen groups adopt
different strategies, the principal party of “P ” can make interaction revenues
without paying “C” while the principal party of “F ” pays “C” and makes “V ”
and “S”; (3) When both groups apply the peaceful sheep-flock strategy, namely,
as onlookers or passers-by, neither of them gain any revenue or deliver any con-
tribution to the incidents, make any “S” or pay any costs, so they spend no
cost and gain no benefit. Table 1 shows the revenue matrix in the game of the
netizen’s sheep-flock effect in the network unexpected group incidents.
Table 1. Revenue matrix in the game of the netizen
Netizen B
Fight (F ) q Peace (P ) 1 − q
Netizen A Fight (F ) p 2V + S − C, 2V + S − C V + S − C, V
Peace (P ) 1 − p V, V + S − C 0, 0
314 G. Xiong et al.
Table 1 shows that the netizens pay a little price when they forward or com-
ment the article in most case. To be concluded, the fight brings more bene-
fits than peace, obviously S > C, so the size relation between parameters is:
2V + S − C > V + S − C > V > 0.
2.2 Establishing the RDEU-Based Evolutionary Game Model for

the Netizen Groups
The dynamic game model was basically built on the basis that the netizens
were entirely rational. However, when the social economic environment and the
decision problems are more complicated, people’s rationality will be obviously
limited. For this reason, in a move to ensure the application value of game analy-
sis, the netizen game of limited rationality must be analyzed. Consequently, the
following part will discuss the evolutionary game model among netizen groups in
a drive to analyze the game under the analytical framework of repeated random-
pair game among netizens of low rationality. Suppose that the netizens adopting
“F ” occupy the proportion of “p” of the netizen group A and the counterparts
adopting “P ” account for (1 − p); the netizens adopting “F ” take up a propor-
tion of “q” of the netizen group B and the counterparts adopting “P ” account
for (1 − q), where p, q ∈ [0, 1].
Network unexpected group incidents have displayed distinctive characteris-
tics from traditional group incidents chiefly in terms of emotional extremity in
these years. Because of lack of authoritative review mechanism in communi-
cation, some features, such as popularization and randomness arose. It is cor-
respondingly difficult for the governmental institutions to control. At the same
time, because of some problems like enlarging gap of wealth, soaring house prices
and inappropriate governmental behaviors, the netizens tend to show sympathy
to the disadvantaged groups. When the incident object relates to social unfair-
ness, the netizens are evidently easier to get agitated or go to extremes. To reflect
the emotional status of both parties of the game in strategy selection, the paper
hereby applies the RDEU theory to include the emotional factors of the netizen
groups into the above game model of the netizen’s sheep-flock effect. Suppose
the emotional function of netizen group A and netizen group B is respectively
wA (p) = pr1 and wB (q) = q r2 and r1 , r2 > 0, called the emotional index of the
netizens respectively.
In the RDEU theory, we can get that the probabilities of different revenues
made by netizen group A and B when they adopt different strategies and the rank
of the revenue value of different strategy portfolio among all the revenue values. It
can be called the rank of corresponding revenue. If the rank of a strategy portfolio
is higher, it is less possibility that the revenues of other strategy portfolios can
exceed it. That is the implication of “high revenue and low probability”. The
corresponding decision weights are calculated with probabilities and ranks of
different revenues, as listed in Tables 2 and 3.
Table 2. The probability distribution, rank and decision weights of netizen group A’s
revenue
The revenue of 2V + S − C (F) V + S − C (F) V (P) 0 (P)

Netizen group A xi
Probability pi pq p(1 − q) (1 − p)q (1 − p)(1 − q)
Rank RPi 1 1 − pq 1−p (1 − p)(1 − q)
Decision weights wA (pq) wA (p) − wA (pq) wA (p + q − pq) − 1 − wA (p + q + pq)
π(xi ) wA (p)
Table 3. The probability distribution, rank and decision weights of netizen group B’s
revenue
The revenue of 2V + S − C (F) V + S − C (F) V (P) 0 (P)

Netizen group A xi
Probability pi pq (1 − p)q p(1 − q) (1 − p)(1 − q)
Rank RPi 1 1 − pq 1−q (1 − p)(1 − q)
Decision weights wB (pq) wB (q) − wB (pq) wB (p + q − pq) − 1 − wB (p + q + pq)
π(xi ) wB (q)
UAF and UAP represent the expected revenues of netizen group A adopting
“F ” and “P ” respectively in the network unexpected group incidents.
UAF = V (pq)r1 + (V + S − C)pr1 ,

UAP = V (p + q − pq)r1 − V pr1 .
The RDEU expected revenues of netizen group A as:
UA (p, q; w) = UAF + UAP = V (pq)r1 + V (p + q − pq)r1 + (S − C)pr1 . (1)
Similarly, UBF and UBP are taken to represent the expected revenues of
netizen group B adopting “F ” and “P ”
UBF = V (pq)r2 + (V + S − C)q r2 ,

UBP = V (p + q − pq)r2 − V q r2 .
The RDEU expected revenues of netizen group B as:
UB (p, q; w) = UBF + UBP = V (pq)r2 + V (p + q − pq)r2 + (S − C)pr2 . (2)
UA is taken to represent the average revenue of netizen group A, then
UA =pr1 UAF + (1 − pr1 )UAP = V p2r1 q r1 − V pr1 (p + q − pq)r1

+ (2V + S − C)p2r1 + V (p + q − pq)r1 − V pr1 . (3)
The limited rationality of the netizen groups indicates that people will find
better strategies by learning in the game instead of finding the best strategy at
316 G. Xiong et al.
the beginning. According to the theory of replicated dynamics for biological evo-
lution, the game player who adopts strategies yielding lower revenues will change
the strategy and turn to (simulate) the strategies yielding higher revenues. For
this reason, the proportion of group members who apply different strategies
may change, then the speed of proportion change and the proportion of specific
strategies have positive relation with the margin of its revenue over the average
revenue. Consequently, the speed of change of “p” of the netizens resorting to
“F ” among the above netizen group A can be shown in the replicated dynamic
equation as follows:
dp
= pr1 (UAF − UA ) = pr1 (1 − pr1 )[V (pq)r1
dt
+ (2V + S − C)pr1 − V (p + q − pq)r1 ]. (4)
At the same time, is taken to represent the average revenue of netizen group
B, so
UB = q r2 UBF + (1 − q r2 )UBP = V pr2 q 2r2 − V q r2 (p + q − pq)r2

+ (2V + S − C)q 2r2 + V (p + q − pq)r2 − V q r2 . (5)
When the netizen group B take the “fight” behavior, the change of speed of
netizen number’s proportion can be represented as the below replicator dynamics
equation as
dq
= q r2 (UBF − UB ) = q r2 (1 − q r2 )[V (pq)r2
dt
+ (2V + S − C)q r2 − V (p + q − pq)r2 ]. (6)
Simultaneous integration of the replicated dynamic equations of netizen

group A and B may help to build the replicated dynamic model for the evo-
lutionary game of netizen group A and B applying changing strategies as time
changes.
dp
dt = p (1 − p )[V (pq) + (2V + S − C)p − V (p + q − pq) ]
r1 r1 r1 r1 r1
(7)
dt = q (1 − q )[V (pq) + (2V + S − C)q − V (p + q − pq) ].
dq r2 r2 r2 r2 r2
3 Numerical Simulation of RDEU-Based Evolutionary

Game Model
Because of the numerous parameters in Formula 7 and the uncertainty of the
parameter values, this part adopts the approach of numerical simulation to fur-
ther discuss the afore-mentioned model in order to reflect how emotions influence
the behaviors of netizens. To facilitate the operation, we suppose V = 6, S = 8
and C = 4 and rewrite the replicated dynamic model into:
dp
dt = p (1 − p )[6(pq) + 16p − 6(p + q − pq) ]
r1 r1 r1 r1 r1
(8)
dt = q (1 − q )[6(pq) + 16q − 6(p + q − pq) ].
dq r2 r2 r2 r2 r2
After that we assume r1 and r2 which the emotional indexes of the netizens
be divided into six kinds of conditions, then the conditions simulation software
named matlab7.0 will be applied to simulate the changes of the strategies of the
network groups in network unexpected group incidents and analyze the different
conditions to reach the state of evolutionary equilibrium. Assuming the initial
ratio of p and q is (0.2, 0.7). When t = 0, the proportion of netizens of group B
applying “F ” is higher than that of group A.
Situation 1: Both sides of netizen groups are rational, namely r1 = r2 = 1.
Fig. 1. The result of simulation when r1 = 1, r2 = 1
Figure 1 shows that the evolutionary result of the experiment when all the
parameters conform to situation 1. The result manifests that when both the
netizen group A and netizen group B are rational, the system will eventually
evolve to be “F ” of group B and “P ” of group A.
Situation 2: One of the netizen groups is rational while the other is optimistic.
r1 = 1 and r2 ∈ (1, +∞) or r1 ∈ (1, +∞) and r2 = 1, suppose r1 = 1, r2 = 1.7
or r1 = 1.7, r2 = 1
Figure 2 describes the evolutionary results of the system when all the para-
meters satisfy to situation 2. The result shows that when netizen group A and
group B are optimistic or group A are optimistic and group B are rational, the
system will finally evolve to be “F ” of Group B and “P ” of Group A.
Situation 3: One party of the netizen groups is rational while the other is
pessimistic. r1 = 1 and r2 ∈ (0, 1) or r1 ∈ (0, 1) and r2 = 1, suppose r1 = 1, r2 =
0.9 or r1 = 0.9, r2 = 1.
Figures 3(a) and (b) indicate that the evolutionary results of the system when
all the parameters conform to situation 3. The result shows that when netizen
group A are rational and group B are pessimistic or group A are pessimistic and
group B are rational, the system will finally evolve to be “F ” of group B and
“P ” of group A.
318 G. Xiong et al.
Fig. 2. (a) The result of simulation when r1 = 1, r2 = 1.7; (b) The result of simulation
when r1 = 1.7, r2 = 1
Fig. 3. (a) The result of simulation when r1 = 1, r2 = 0.9; (b) The result of simulation
when r1 = 0.9, r2 = 1
Situation 4: Both sides of netizen groups are pessimistic, as r1 ∈ (0, 1) and

r2 ∈ (0, 1), suppose r1 = 0.5, r2 = 0.6.
Figure 4 describes that the evolutionary result of the system when all the
parameters conform to situation 4. The result shows that when netizen group A
and B are both pessimistic, the system will finally evolve to generate the sheep-
flock effect that both sides apply “F ” because netizen group A simulates netizen
group B.
Situation 5: Both sides of netizen groups are optimistic, as r1 ∈ (1, +∞) and
r2 ∈ (1, +∞), suppose r1 = 1.2, r2 = 1.3.
Figure 5 depicts the evolutionary result of the system when all the parameters
conform to situation 5. The result shows that when netizen group A and B are
both optimistic, the system will finally evolve to be the “F ” of netizen group B
and the “P ” of netizen group A.
Situation 6: One party of the netizen groups is optimistic while the other is
pessimistic, namely, r1 ∈ (0, 1) and r2 ∈ (1, +∞) or r1 ∈ (1, +∞) and r2 ∈ (0, 1),
suppose r1 = 0.9, r2 = 1.1 or r1 = 1.1, r2 = 0.9
Fig. 4. The result of simulation when Fig. 5. The result of simulation when
r1 = 0.5, r2 = 0.6 r1 = 1.2, r2 = 1.3
Fig. 6. (a) The result of simulation when r1 = 0.9, r2 = 1.1; (b) The result of simulation
when r1 = 1.1, r2 = 0.9
Figure 6(a) and (b) describe the evolutionary results of the system when all
the parameters conform to situation 6. The result shows that when netizen group
A and group B are optimistic or group A and group B are pessimistic, the system
will finally evolve to be of Group B and of group A.
4 Conclusion
Based on the shortcomings of the existing literature on the research of group
events on networks and the new RDEU theory, this paper establishes RDEU
game model of group events on network. Under the six evolutionary scenarios,
it analyzes the extent and condition of influence of different emotions on neti-
zens’ game behavior. This paper overcomes the limitations of the traditional
game model which on the basis of “full rationality” hypothesis of netizens, and
considers the influence and mechanism of endogenous factors such as netizen
emotion on the evolution of the group events on networks. It also reveals the
deeper root of the outbreak and expansion of group events on networks. The
results show that emotional factors have an important influence on the choice of
320 G. Xiong et al.
the game strategies of the group events on networks. When some netizens has
a “pessimistic” emotion, they tend to “antagonistic” behavior, and the higher
the degree of pessimism, the more likely they choice a risk strategy; When some
netizens has a “optimistic” emotion, they are easy to commit “concessions”
behavior; When the netizens are pessimistic in both groups, the event is likely
to evolve into a conflict effect with the herd effect.
The conclusion of this study can provide some reference and inspiration for
the emergency management and prevention and control measures of network
mass incidents. Government departments or social media should increase the
response mechanism for the interests and wishes of netizens, strengthen netizens’
emotion monitoring and analysis and provide psychological intervention and
emotional counseling for them timely.
References
1. Deng Y, O’Brien KJ (2012) Relational repression in China: using social ties to
demobilize protesters. Soc Sci Electron Publishing 215(215):533–552
2. Levine P (2001) Civic renewal and the commons of cyberspace. Nat Civic Rev
90(3):205–212
3. Liu D, Wang W, Li H (2013) Evolutionary mechanism and information supervision
of public opinions in internet emergency. Procedia Comput Sci 17:973–980
4. Liu D, Han C, Yin L (2016) Multi-scenario evolutionary game analysis of evo-
lutionary mechanism in urban demolition mass incident. Oper Res Manage Sci
1:76–84
5. Morales AJ, Borondo J et al (2014) Efficiency of human activity on information
spreading on twitter. Soc Netw 39(1):1–11
6. Post DG (2002) Against ’against cyberanarchy’. SSRN Electron J 17(4):1365–1387
7. Qiu W (2009) The study of the network group events. J Ningbo Radio TV Univ
8. Quiggin J (1982) A theory of anticipated utility. J Economic Behav Organ
3(4):323–343
9. Sun H, Wang X, Xue Y (2016) Stochastic evolutionary game model for unexpected
incidents involving mass participation based on different scenarios. Oper Res Man-
age Sci 4:23–30 (in Chinese)
10. Tarrow SG (1994) Power in movement: social movements and contentious politics.
Contemp Sociol 28:3–12
11. Wang X, Li Y, Sun H (2015) Evolutionary game analysis of unexpected incidents
involving mass participation based scenario inference. J Manage Sci 6:133–143 (in
Chinese)
12. Xie B, Zhang W et al (2014) Evolutionary game analysis and application to mass
emergency under government coordination. Operations Research and Management
Science 5:243–249 (in Chinese)
13. Zhu J, Zhang E, Cui Y (2013) Characteristics and causes of the network group
events. Theory Res 32:96–97 (in Chinese)
Effects of Internet Word-of-Mouth of a Tourism
Destination on Consumer Purchase Intention:
Based on Temporal Distance and Social Distance
Mo Chen(B) and Jingdong Chen
Faculty of Economy and Management, Xi’an University of Technology,

chenmo19921201@foxmail.com
Abstract. The proportion of tourists who referencing the Internet

word-of-mouth increased. Hence, tourism destinations’ online word-of-
mouth has an practically significant effect on the willing of purchase.
Due to an increasing trend about the proportion of tourists referencing
Internet word-of-mouth (IWOM) for, making a study of tourism desti-
nation online word-of-mouth influence on purchase intention has prac-
tical significance. The purpose of this article is to provide marketing
advises in different temporal and social distance. The research methods
of this article is dividing 110 college students into two groups. All the stu-
dents are allowed to read online word-of-mouth about two destinations,
and this is for measuring consumers’ tourism consumption which under
different temporal distance and social distance categories. The conclu-
sion is that quality of word-of-mouth exercise a greater effect for the
consumers when they will travel in the far future. And the quantity of
word-of-mouth influence nearly future purchase intention of consumers
more prominently. For far Social distance between reader and IWOM
witter, purchase intention is more significantly affected by the quality of
word-of-mouth. On the contrary, purchase intention is more significantly
influenced by the quantity word-of-mouth, when they feel they are close.
Keywords: Internet · Word-of-Mouth (IWOM) · Consumers purchase

intention · Temporal distance · Social distance
1 Introduction
The context of traveling getting increasingly popular, prospects is taking on
prosperity and potential is substantial for the traveling sector in China, with
online consumption in doing sightseeing projected to reaching RMB 800 Billion.
In 2014, this country had a transaction scale of RMB 3250 Billion, in which the
E-market of tourism represents 11.3% to 367 Billion, a 45.6% rise compared to
last year.
The reason why Chinese marketing has good developing prospects is great
influence of WOM. As the consequence surveyed by Market research company
322 M. Chen and J. Chen
GfK shows that more than half of Chinese travelers rely on personal contact to
obtain travel information, including friends, family and colleagues, which is same
as online reading. In terms of actual booking, online instead of offline, there are
63% of visitors choosing online payment, which means that they are unavoidable
to read some online information about destinations. 80% of consumers said they
buy, influenced by the Internet word-of-mouth IWOM, implying that researching
the tourism destination IWOM is of practical significance. At the same time,
customer get tourism information platforms and ways, which are more diverse,
because the Internet and mobile phone mobile client widely are used. The IWOM
regarding tourism destinations in platforms such as ctrip, weibo or other tourism
BBS or APP spread more widely, showing it exerts a great influence.
On the one hand, because the time intervals from customers reading IWOM
to the actual purchase is different, leading to their purchase will be different.
Customers will have a psychological judgment when they read online word-of-
mouth. The judgment of the psychological distance between customers and the
author of IWOM will also change the customer’s purchase intention. Therefore,
this paper starts from the difference of the influence of the word-of-mouth in
the different temporal distance and social distance, observes the performance of
consumers’ purchasing intention and explains the influence of online word-of-
mouth.
2 Literature
2.1 Word-of-Mouth Research
Arndt [1] first raised word-of-mouth in marketing research significance and
showed word-of-mouth as a third party’s influences on consumer behavior. And
early studies tended to probe into the reason for word of mouth research [7].
With the development of the Internet and the increase of its scale, word-of-
mouth research started in the Internet-word-of-mouth. In research, Internet-
word-of-mouth (IWOM) [24], electronic word-of-mouth (ewom) [22], and elec-
tronic reputation refer to the same concept. In the study of IWOM, scholars
[6,13] thought the IWOM is that customers or potential customers in the net-
work community or BBS put forward for the product or business the positive
and negative. Dong dahai defined the concept of IWOM [8], and considered
IWOM has three main differences compared to traditional word-of-mouth: 1
Spread more broadly. 2 Connection between the receivers and senders is not
determined [4]. In addition to the strong relationship between circles of friends,
there is also a weak relationship.
3 commercial uncertainty. Due to some rea-
sons such as the anonymity of the Internet behavior, Internet word of mouth
will be controlled by some enterprises or business organizations. So, examining
IWOM is different from traditional marketing and this makes word-of-mouth a
fresh meaning.
From the beginning of the 21st century’s China, research on word-of-mouth
started to be increasing gradually. Domestic study about word-of-mouth for
consumers to buy has two directions. One direction is discovering the emotional
Effects of Internet Word-of-Mouth of a Tourism Destination 323
Online
word of attitude actual purchase
mouth
regulatory
factors
Fig. 1. Word-of-mouth and attitude theoretical model
factors of word-of-mouth. Guo [10] put forward the theoretical model of word-
of-mouth for consumer attitudes, as shown in Fig. 1. After the model raised, A
lot of research around the positive word-of-mouth and negative word-of-mouth.
They believe that the emotion behind IWOM influence actual purchase through
attitude. Baber et al. [2] found that the positive online word-of-mouth can bring
positive change on consumers’ purchase intentions. Bailey [3] proved that neg-
ative reviews have a significant impact on customers decision. Based on the
original model studied, concentrated on regulating variables, which joined the
adjustment factors such as trust and relationship strength. Balaji [22] found
that the relationships among language divergence effect positive word-of-mouth
intentions.
On the other hand, Zuo [20] put forward a model concerning the consumer
purchase intention affected by word-of-mouth’s quantity and quality. The exist-
ing literature has proved quality and quantity of word-of-mouth influencing
intentions of consumers purchase [23]. This studies focus on the transmission
of information. Balajia [4] mentioned information quality in the article about
language divergence in WOM. This article in also refer to interaction quality
and relationship quality.
2.2 Temporal Distance and Social Distance

Research on consumers’ behavior and psychology research on time distance and
social distance in recent years also have attracted more attention. With the
development of the theory of psychological explanation level, researchers began
to define the temporal distance and social distance from the Angle of cognitive.
Reference point means the subject is located in someplace at the moment, the
distance between the location of reference point and object is vague [19]. Time
distant refers to people’s feeling about the interval from an event occurrence
to now [5], while social distant is they feeling about the relationship between
relevant person and them [5].
Many studies probe into social distance and time distance with purchase
intention of literature. Junghye’s [15] conclusions are drawn that with the time
interval between intention formation and the action becoming longer, and the
greater the spatial distance to a destination is, the non-travelers have higher
the probability to change behaviors. Xu [14], through experiment to measure
the temporal distance and social distance for the interaction of the virtues and
vices products, proved discrepancy in customers’ preferences for the product
under different situations. Zhang [21] proved the temporal distance effect on
consumers’ impulse consumption. Zou [25] found that the customers emphasis
desirability or feasibility differently, when the consumer choose financial products
under different temporal distances.
The preliminary study shows that temporal and social distance have an effect
on consumer purchase intention. In short, combined with psychology and word
of mouth is also a relatively new research.
3 The Experimental Research
3.1 Research Hypothesis
Internet word-of-mouth has one of the biggest differences compared with the
traditional word of mouth, whose access is of diversification, so the quantity of
electronic word-of-mouth is substantial. In terms of tourism destination’s IWOM,
consumers can collect and obtain, from several ways like tourism network BBS,
blog, Micro-blog, instant messaging (We-chat, qq), network video, Wikipedia
and travel quiz, tourist IWOM. Liu [16] proved that the movie aspect through
word-of-mouth affects the relationship between the number of product sales and
reputation. The study of the quantity of word-of-mouth, with the development
of the network, from the original single-platform to cross-platform research, at
present, has developed [11] swiftly, so the total quantity of IWOM is measured
from the viewpoint of cross-platform.
The quality of word-of-mouth in essence refers to the content of word of
mouth. In terms of tourist destination of IWOM, word-of-mouth quality eval-
uation is mainly from the correlation of word of mouth to convey information,
professional degrees, and the vitality of comprehensive evaluation on the qual-
ity of word-of-mouth. In the texts of existing research on tourism destination’s
IWOM, online word-of-mouth interesting and vivid impact on the quality of word
of mouth are great, illustrating IWOM which has high quality in the customer
in the heart of higher quality [12].
Comprehensive quality and quantity of word-of-mouth’s related research,
Grant [9], proved that the total effect of word-of-mouth quantity and quality
is the most significant, especially the quality of word of mouth, which is dif-
ferent from the path on the purchase intention. So the theoretical model of its
influence on purchase intention action is (as shown in Fig. 2).
quality of word of mouth

consumers purchase
intention
quantity of word of mouth
Fig. 2. Word-of-mouth and purchase intention model

This basis the paper puts forward the following two assumptions:
Hypothesis 1 : Consumers purchase intention is more significantly affected by
the quality of word-of-mouth in far future and the quantity of word-of-mouth
influence nearly future purchase intention of consumers more prominently.
Hypothesis 2 : When there is far social distance between receivers and senders,
the receivers’ purchase intention is more significantly affected by the quality
of word-of-mouth; When they have nearly social distance, customers’ purchase
intention is more significantly influenced by the quantity word-of-mouth.
3.2 Design of Experiment
(1) Subjects
The participants were senior undergraduate and graduate students, between
the age of 21 to 25 years old, a total of 110 people, is divided into two groups,
each group divided into 55 people. There are 28 women, and 27 men in each
group. Choosing college students as subjects is mainly based on the following
reasons: first, according to the survey of 2013, we found that 85.43% of college
students surveyed said “I like traveling”. And college students often use the
Internet and like to travel. Second, college students and postgraduates with
high homogeneity, can effectively control the contrast group and control group
of demographic heterogeneity.
(2) Experiment Preparation
Experiment on all testers to get on the Internet first hand tourist informa-
tion platform and channel was tested. These platforms for tourist destination’s
IWOM, including Ctrip, Qyer and Ma Feng Wo, the tourism BBS, post tourist
experience. There are 90.9% testers, who collect information from those platform.
89.1% from weibo, 60% form We-chat’s moment and subscription account, 1.8%
from other ways. Due to the W-chat moment’s privacy, word-of-mouth authors
have close social distance with word-of-mouth readers. So that the experience of
travel sites post and micro-blog platform simulation scene, which measures the
participants’ purchase intention in the different temporal and social distance.
First of all, by reading the study of word-of-mouth, the author summarized
main export tablets in quality and quantity of several attributes. Word-of-mouth
quantity evaluation is mainly from three aspects [18]: 1 multiple online platform
to see multiple destinations related IWOM appear; 2 the common platform to
see multiple destinations IWOM; 3 the IWOM of destination has a lot of for-
warding. Measurement of word-of-mouth quality is from the relevance, integrity,
interesting and professional comprehensive evaluation.
On this basis, the author collected the related tourism destination online net-
work of word of mouth. To avoid the destination names, which interfere the sub-
jects of purchase intention, we use A and B represent to tourist destination, and
both are simulating the ancient town tour relevant scenic spots, this ruling out
consumer preferences’ disturbance. The IWOM of Destination A reflect word-
of-mouth quantity, and tourist destination B’s IWOM reflect word-of-mouth
quality. Word of mouth simulates the situation in three network platforms like
weibo, ctrip and Ma Feng Wo. The tester would be shown 12 items IWOM about
A and 3 items IWOM about B. The IWOM about B is 1 item IWOM in each
platform, and the IWOM about A is 4 items IWOM in each platform. At the
same time the experiment obscures the names of related sites.
(3) Experimental Design
Experiment is divided into two steps: the first measuring different temporal
distance and the second experiment measuring different social distance. Measur-
ing subjects for tourism destination A and B different purchase intention, and
the scale basis by using three items, designed by Mr Schiff (Schiffman et al.) [17]:
1 I am willing to travel to the area;
2 I would like to recommend to others the
region tourism; 3 I’d be happy to try to travel to the area.
Before answering questionnaire, subjects need to read the IWOM which sim-
ulates the IWOM’s model of tourism BBS and Micro-blog. Scale takes 7 likert
scale from “strongly disagree” to “strongly agree” assignment 1 to 7, recycling
questionnaire by examining Cronbach’s Alpha = 0.767 > 0.7, certificating valid-
ity within a reasonable range.
Because tourism is different from the retail purchase, it needs to cooper-
ate with holiday, so the writer has been testing the latest holiday time for the
recent test used in measuring temporal distance experiment. One is “weekend”
and another is “one year”, which are two different time distance measurements.
Experiment announced by two word of mouth the identity of senders, 53-years-
old workers and 23-year-old college student, respectively.
3.3 The Results
Experiment 1: using SPSS19.0 for data analysis and processing. In the study of
the demographic variables such as gender, age did not produce any significant
effect, therefore not included in the model. Mixed repeat test analysis of variance,
the BOX ’sM inspection results F (2, 108) = 0.174 > 0.05, co-variance matrix
through the homogeneity test, in the error variance equality Levene test, word-
of-mouth quality P = 0.937, word-of-mouth quantity P = 0.094 were greater
than 0.05 through inspection. Through the following Table 1 descriptive statistics
analysis, it can be seen that word-of-mouth obvious interaction between quality
and quantity exists.
Internet word of mouth and the interaction of the distance and time (Eta =
0.645, p > 0.645) can better account for the variation model. As shown in Table 1
and Fig. 3, in the distant future buying situation, the purchase intention of sub-
jects who read the IWOM of tourism destination B (reflect word-of-mouth qual-
ity), are significantly higher than that of tourism destination A (reflect word-of-
mouth quantity); In near future purchase situations, the purchase intention of
testee who read destination A (reflect word-of-mouth quantity), is significantly
higher than of destination B (reflect word-of-mouth quality). Hypothesis 1 is
validated.
Because of social distance perception differences, the experiment by control-
ling the reviewers with different background information realizes the control of
Table 1. The results of experiment one
Temporal distance Online-word-of-mouth Paired-sample t test

Quality Quantity t-value Degree of
freedom
Mean Standard Mean Standard
deviation deviation
Near future 3.18 1.335 4.35 1.377 3.734∗∗∗ 54
Distant future 4.24 1.319 2.47 1.052 7.251∗∗∗ 54
∗∗∗
is p < 0.001
quantity
Purchase Intention: (1-7)
quality
Near future distant future
Fig. 3. Impact of word-of-mouth quality and quantity on purchase intention of con-

sumers in different temporal distances
social distance. In the social distance situation in the group, we told the partic-
ipating commentators, 53-year-old workers (social identities and differences in
subjects); Then in the social situation in the group closer, we told participating
reviewers, 23-year-old college students (high degree of social identity is similar
to the participants).
Social distance measure takes the following scale: (1) I think of myself and
word-of-mouth authors similar; (2) I think of myself and word-of-mouth authors
psychological distance; (3) I think of myself and word-of-mouth authors belong-
ing to the same group. Adopting likert scale at 7 level produces The result of
social distance is that a 23 of college students (M = 5.26) is significantly higher
than a 53 office worker (M = 3.44), showing that the social distance control
experiment is effective.
BOX’sM inspection results F (2, 108) = 0.100 > 0.05 in Levene test, word-of-
mouth quality P = 0.803, the word-of-mouth quantity (P = 0.07 were greater
than 0.05 through inspection. And the interaction of the Internet word of mouth
and social distance (Eta = 0.531, p > 0.531) can better explain the variation
model.
As shown in Table 2 and Fig. 4, when the social distance is far between the
senders and receivers of IWOM, the subjects’ purchase intention of tourism
destination B (reflect word-of-mouth quality), are significantly higher than that
of tourism destination A (reflect word-of-mouth quantity); When the reader feel
Table 2. The results of experiment two
Social distance Online-word-of-mouth Paired-sample t test

Quality Quantity t-value Degree of
freedom
Mean Standard Mean Standard
deviation deviation
Far social distance 3.55 1.212 4.51 1.611 5.185∗∗∗ 54
Close social distance 4.80 1.311 2.65 1.181 6.885∗∗∗ 54
∗∗∗
is p < 0.001
quantity
Purchase Intention: (1-7)
quality
Far social distance close social distance
Fig. 4. Impact of word of mouth quality and quantity on purchase intention of con-
sumers in different social distances
they have close social distance, which means that they believe that they have a
similar social identity, the purchase intention of destination A (reflect word-of-
mouth quantity), is significantly higher than B ((reflect word-of-mouth quality).
Hypothesis 2 is validated.
4 Results, Discussion and Limitations

In this paper, tourism destination online word-of-mouth can be divided into two
different measures, including the quality and the quantity of word-of-mouth.
This paper, through two experiments, tests the quantity and quality of IWOM
influencing on consumer purchase intention under different temporal distance
and the social distance situation. The conclusion is as follows:
(1) Quality of Internet word-of-mouth for the forward travel consumer buying
may exercise a greater effect and the quantity of IWOM influence on recent
purchase intention of consumers more prominently. This suggests that if the
time interval between reading and buying is far, the customers’ decision-
making depends on IWOM quality. It because the customer’s high level
interpretation will be easily activated, when the cognitive and purchase time
interval is far. That means high levels of explanation of IWOM more easily
by cognitive quality and low levels of explaining IWOM quantity will be

weakened.
(2) The far social distance between the word-of-mouth authors and readers, con-
sumers purchase intention is influenced significantly by the quantity of word-
of-mouth. This suggests that when customers and word-of-mouth authors
are far in social distance, the customers’ decision-making depends on IWOM
quality. It because the customer’s high level explanation is easy to be acti-
vated. That means high levels of explanation of IWOM quality theory is
more easily cognitive and low explaining IWOM quantity will be weakened.
(3) The average of quality of word of mouth influencing on consumers’ will-
ingness to buy is higher than word quantity, which means the high-quality
IWOM have more extensive and more lasting influence.
In this paper, the study of online word-of-mouth, expands the research con-
tents of psychological explanation level about social distance, and also provides
inspiration for tourist destination choice.
First, the relevant tourism destination management on tourism propaganda
should choose who have closer social distance with destination target customers.
At the same time, the high quality tourism strategy and the relevant evaluation
can attract more consumers. The platform of their own tourism official weibo,
also should share more high-quality tourist guides, and, release illustrated weibo.
Second, when it is tourist season, the relevant tourism management depart-
ment can propaganda through multiple platforms, and form a certain scale of
word-of-mouth, to attract more attention from people.
Finally, because the study take the simulation experiment method, in the
simulation of situation and realistic scenarios exist certain differences. In the
future, we should be in the real scene to verify this conclusion.
Acknowledgement. This paper is supported by the Shaanxi Province Social Science

Fund (No. 2016R015) and Research on Promoting to Construct Cultural Powerhousee
by Cultural-oriented Tourism–Based on Customer Marketing (No. 2016R015).
References
1. Arndt J (1967) Role of product-related conversations in the diffusion of a new
product. J Mark Res 4(3):291–295
2. Baber A, Ramayah T et al (2015) Online word-of-mouth antecedents, attitude and
intention-to-purchase electronic products in Pakistan. In: USENIX conference on
USENIX technical conference, pp 307–318
3. Bailey AA (2004) Thiscompanysucks.com: the use of the internet in negative
consumer-to-consumer articulations. J Mark Commun 10(3):169–182
4. Balaji MS, Khong KW, Chong AYL (2016) Determinants of negative word-of-
mouth communication using social networking sites. Inf Manage 53(4):528–540
5. Bar-Anan Y, Liberman N, Trope Y (2006) The association between psychological
distance and construal level: evidence from an implicit association test. J Exper
Psychol Gener 135(4):609
6. Bussière D (2015) Evidence and implications of electronic word of mouth. Springer
7. Chu SC, Kim Y (2011) Study on the relationship between the feature of microblog’s
word-of-mouse and consumer’s brand attitude. Int J Advertising 30(1):47–75
8. Dong D (2012) Contrast analysis of word-of-mouth, internet word-of-mouth and
word-of-mouse. Chin J Manage 3:428–436 (in Chinese)
9. Grant R, Clarke RJ, Kyriazis E (2007) A review of factors affecting online consumer
search behaviour from an information value perspective. J Mark Manage 23(5):519–
533
10. Guo G (2007) The impact of word-of-mouth on consumer’s attitude: a theoretical
model. Manage Rev 6:20–26 (in Chinese)
11. Haixia Y (2015) The distribution of word-of-mouth across websites and online
sales: sensitivity analysis on entropy and network opinion leaders based on artificial
neural network. Econ Manage J 10:86–95 (in Chinese)
12. Haiyan C (2011) Study of communication on electronic word-of-mouth of destina-
tion. Master’s thesis, Wuhan University (in Chinese)
13. Hennig Thurau T, Gwinner KP et al (2004) Electronic word-of-mouth via
consumer-opinion platforms: what motivates consumers to articulate themselves
on the internet. J Interact Mark 18(1):38–52
14. Jinjing X (2015) The role of temporal distance and social distance in the consumer
decision-making: evidence from eye movements. Master’s thesis, Jinan University
(in Chinese)
15. Kah JA, Choongki L, Seonghoon L (2016) Spatial-temporal distances in travel
intention-behavior. Ann Tourism Res 57:160–175
16. Liu Y (2013) Word of mouth for movies: its dynamics and impact on box office
revenue. J Mark 70(3):74–89 (in Chinese)
17. Schiffman LG, Leslie LK (2000) Consumer behavior. Prentice Hall, Upper Saddle
River
18. Schubert P, Selz D (1999) Web assessment - measuring the effectiveness of elec-
tronic commerce sites going beyond traditional marketing paradigms. In: Hawaii
international conference on system sciences, p 5040
19. Trope Y, Liberman N, Wakslak C (2007) Construal levels and psychological dis-
tance: effects on representation, prediction, evaluation, and behavior. J Consum
Psychol 17(2):83–95
20. Wenming Z, Xu W, Chang F (2014) Relation between electronic word of mouth and
purchase intention in social commerce environment: a social capital perspective.
Nankai Bus Rev 04:140–150 (in Chinese)
21. Xuan Z (2012) The influence of temporal distance on consumers’ online impulse
buying behavior. Master’s thesis, Huazhong University of Science and Technology
(in Chinese)
22. Zainal NTA, Harun A, Lily J (2017) Examining the mediating effect of attitude
towards electronic words-of mouth (eWOM) on the relation between the trust in
eWOM source and intention to follow eWOM among Malaysian travellers. Asia
Pacific Management Review
23. Zheng C, Han Q, Wang H (2015) How do paid posters’ comments affect your
purchase intention. Nankai Bus Rev 01:89–97 (in Chinese)
24. Zhu X, Li A (2014) Determinants of consumer engagement in electronic word-of-
mouth (eWOM) in social networking sites. PhD thesis, Guangdong University of
Technology (in Chinese)
25. Zou P, Hao L, Li Y et al (2014) Impacts of perceived temporal distance-based sales
promotion on purchasing behaviors under information asymmetry. J Manage Sci
Analysis and Prediction of Population Aging
Trend Based on Population Development Model
Jiancheng Hu(B)
College of Applied Mathematics, Chengdu University of Information Technology,

jianchenghu@163.com
Abstract. In recent years, with the acceleration of population aging, the

development of population structure is an important factor that influ-
ences economy and social development of China. This paper studied the
future trend of population in Sichuan of China by establishing a popula-
tion development equation. Taken the statistical data of census in 2010
in Sichuan, some important factors, such as survival rate, fertility mode
and gender ratio etc. are considered to estimate the population structure,
especially the proportion of older people, in Sichuan. The model is uti-
lized to predict the population aging trend and aging index in the case of
different total fertility rate and to provide reference for the government
to make corresponding social and economic decisions.
Keywords: Population aging · Population development equation ·

Aging index · Total fertility rate
1 Introduction
Population aging is a phenomenon that occurs when the median age of a country
or region increases due to rising life expectancy and/or declining fertility rates.
Population aging is becoming one of the global problems of the modern times,
but with different features from region to region and from country to country.
Recent declines in fertility rates and increases in life spans are producing a signif-
icant shift in the age distribution of population [4]. Aging of population is closely
associated with increased life expectancy, which was mainly determined by bet-
ter quality health services, new discoveries in pharmaceutics, children diseases
treatment. Recent research by the United Nations reveals that the proportion
of people aged 60 and over is growing faster than any other age group and it
is expected to reach 1 billion by 2020 and almost 2 billion by 2050 (representing
22% of the global population). The proportion of individuals age 80 or over is
projected to rise from 1 to 4% of the global population by 2050 [5]. Figure 1
shows the change in the age structure of the world’s population over time. Each
shaded section reveals the age distribution at one point in time. The elderly
share will be much higher in 2050 than it is now.
332 J. Hu
Fig. 1. World population by age group in 1950–2050
Population aging is one of the important factors that influence social and
economic development of a country. Since different age groups have different
economic needs and productive capacities, a country’s economic characteristics
may be expected to change as its population ages. Therefor correct forecast
on the future population trend has significant guiding to the economic plan of
overall country and local government.
In recent years, many articles have been paying more and more attention to
population aging prediction and family policy to deal with the aging population
problem, such as using gray dynamic model with equal dimension forest theory
[9], overlapping generations model [1,7], logistic population model [2], predator-
prey population model [8,11,12], stochastic differential equations [10] and so
on. In order to analyze and predict the population trend of Sichuan, China,
population development equation model is presented and applied to analyze the
population of China [3,6].
The United Nations’ old standard maintains that a country with more
than10% of its population over 60 years old is an aging society. The new stan-
dard presents the population of 65 years and above people exceeds 7% of the total
population. In fact, according to census of population, there were 88.11 million
population aged over 65 years in Chinese mainland, accounting for 6.96% of the
total population till November, 2016; while the number turned to 130 million
and 10.20% when it came to the population aged over 60 years.
This paper takes the 6th census data in Sichuan as the accordance, and uses a
population development equation model to predict the population development
in Sichuan. The model is utilized to analyse the population aging trend in the
future in Sichuan from a short period, and further predict the long-term popu-
lation development trend and aging population change condition in Sichuan in
the case of different total fertility rate.
Analysis and Prediction of Population Aging Trend 333
2 Population Development Equation

2.1 Continuous Model of Population Development Equation
Let p(r, t) expresses the population density of r years in the year of t, considering
the population in the age interval [r, r + Δr) from year of to t + Δt, survival
population turns into age interval [r + Δt, r + Δr + Δt), and the number of
dead population is μ(r, t)p(r, t)ΔrΔt, which μ(r, t) is mortality rate of r years
population in the year of t. Then we have
p(r, t)Δr − p(r + Δt, t + Δt)Δr = μ(r, t)p(r, t)ΔrΔt (1)
or
p(r + Δt, t + Δt) − p(r, t + Δt) p(r, t + Δt) − p(r, t)
+ = −μ(r, t)p(r, t), (2)
Δt Δt
when time increases Δt, the age also increase Δt, that is Δt = Δr. Let Δt =
Δr → 0 in Eq. (2), then
∂p(r, t) ∂p(r, t)
+ = −μ(r, t)p(r, t). (3)
∂r ∂t
Given the initial and boundary conditions, that is, initial population density
function p(r, 0) = p0 (r) and fertility rate p(0, t) = f (t), the continuous model of
population development equation can be obtained
⎧ ∂p(r,t) ∂p(r,t)
⎪
⎪ + ∂t = −μ(r, t)p(r, t), 0 ≤ r ≤ rm , t ≥ 0
⎨ ∂r
p(r, 0) = p0 (r), 0 ≤ r ≤ rm
(4)
⎪
⎪ p(0, t) = f (t), t≥0
⎩
p(rm , t) = 0, t ≥ 0,
where the rm is the maximum age of population. Equation (4) is difficult to

solve, but in ideal situation, such as stable society circumstance, short time
interval etc., the mortality rate and fertility rate are independent with time t,
the solution of Eq. (4) can be obtained.
2.2 Discrete Model of Populat1ion Development Equation
In order to study the dynamical population of different ages at any time, the
Eq. (4) should be discreted respect to variables r and t. Based on Eq. (1), the r
years population Nr (t) in the year of t can be written
Nr+1 (t + 1) = sr (t)Nr (t), r = 0, 1, 2, · · · rm − 1 (5)
and by virtual of initial condition, also have
N0 (t) = s0 (t)b(t), (6)

334 J. Hu
where sr (t) and s0 (t) are the survival rate of r years population in the year of t
and infant survival rate respectively, which can be calculated by the statistical
data of population census. b(t) is the population born in the year of t. Concerning
the reproductive age interval [r1 , r2 ], b(t) can be described as
r2

b(t) = br (t)kr (t)Nr (t), (7)
r=r1
where br (t) is the fertility rate of childbearing age female at r years old in the
year of t, i.e. the average number of infant born by each r years old female in
the year of t, kr (t) is gender ratio, the ratio of r years old female in the total
population in the year of t.
By representing the population of each age in the year of t as a vector of size
rm , let
T
N (t) = [N0 (t), N1 (t), · · · , Nrm (t)] , (8)
the population development Eqs. (5) and (6) can be written as
N (t + 1) = A(t)N (t) + β(t)B(t)N (t), (9)
where
⎡ ⎤
0 0 ··· 0 0
⎢ s1 (t) 0 · ·· 0 0⎥
⎢ ⎥
⎢ s2 (t) · · · 0 0⎥
A(t) = ⎢ 0 ⎥, (10)
⎢ .. .. . . .. .. ⎥
⎣. . .. .⎦
0 0 0 srm −1 (t) 0
⎡ ⎤
0 · · · 0 b∗r1 (t) · · · b∗r2 (t) 0 · · · 0
⎢0 ··· 0 0 ··· 0 0 ··· 0⎥
⎢ ⎥
B(t) = ⎢ . . . . . . .. . . .. ⎥ , (11)
⎣ .. . . .. .. . . .. . ..⎦
0 ··· 0 0 ··· 0 0 ··· 0
b∗r (t) = s0 (t)sr (t)kr (t)hr (t), r = r1 , · · · , r2 , (12)
and the total fertility rate

r2

β(t) = br (t), (13)
r=r1
which is the average number of infants born by each female in the year of t.
Define the female’s fertility mode hr (t), which is r years old female’s fertility
r2
weighted factor and satisfied r=r 1
hr (t) = 1. Then br (t) can also written as
br (t) = β(t)hr (t). (14)

3 Model Establishment and Solution

As Sichuan has come into aging society, in order to make clearly the future
population aging trend, we use population development equation to predict the
population in a short period and analyze the change trend of future population
aging speed. At the same time, predicting the long-term population development
process of Sichuan in the case of different total fertility rate, and trying to seek
for a relatively good total fertility rate not only can control too fast population
growth but also ease the population aging condition.
3.1 Prediction of Population Structure
As the data information provided by the 6th census in Sichuan is limited, and
the factors that influence population development are various, it is relatively
difficult to get the rule that mortality rate function and fertility mode changes
as the time changes. The population fertility policy will not change in a relatively
short period; in the condition of stable society, we usually suppose that survival
rate sr (t), fertility mode hr (t) and gender ratio kr (t) relatively do not change.
The data published from the 6th census in Sichuan shows the total fertility rate
of women β(t) = 1.075 at present. Under the supposition above, we use the
data of the 6th census in 2010 in Sichuan as the cardinal number to predict the
population in the future six years in Sichuan. The prediction result is shown as
Table 1.
From the prediction results, we can see that the population in Sichuan is still
increasing. After Guangdong, Shandong and Henan, Sichuan is ranked fourth big
province in the whole country population, also aging serious provinces. Although
the data of the census shows the population fertility rate is relatively low, and
the total population still continuously grows.
From Table 2, we can see from the calculation result of population proportion
at each age section that the proportions of aged people and children still con-
tinuously increase, but the proportion of adults continuously decreases. In 2010,
people under the age of 14 years accounts for the total population of 16.97%,
0.37% points higher than the national average level; 15–64 year old population
account for the total population of 72.08%, 2.45% points lower than the national
average; the population aged 65 years account for the province’s total population
of 10.95%, 2.08% points higher than the national average.
3.2 Aging Index
The aging index (AI, sometimes referred to as the elder-child ratio) is a common
indicator of changes in age structure, which is defined as the number of people
aged 65 and over per 100 young people aged 0–14 years [10]. In 2000, only a
few countries (Germany, Greece, Italy, Bulgaria, and Japan) had more elderly
than youth (aging index above 100). To date, aging indexes are much lower in
developing countries than in the developed world, but the proportional rise in the
336 J. Hu
Table 1. Age structure of population by five-year age group, 2010–2016
Age structure 2010 2011 2012 2013 2014 2015 2016

(years)
0–4 4277055 4053597 3948212 4059425 4214668 4781547 5427837
5–9 4277626 4293234 4359772 4421205 4389209 4264325 4041644
10–14 5092442 4832421 4597470 4347174 4310136 4268574 4284142
15–19 6213182 6240489 6147844 5893439 5505725 5080985 4821552
20–24 6360027 6240153 6047303 6201402 6234291 6193317 6220385
25–29 4514128 4779065 5055472 5395750 5886962 6333555 6214086
30–34 4874122 4529418 4569747 4499193 4389343 4489706 4753438
35–39 8077512 7665262 6998152 6248575 5514606 4839660 4497725
40–44 7819390 7982458 8320171 8225243 8268743 7997893 7589384
45–49 5934098 6988192 7385771 7423684 7470576 7707164 7868020
50–54 4142765 3507478 3346675 4030751 4908537 5809492 6838453
55–59 5725272 5805843 5583171 5302779 4637600 4015098 3398585
60–64 4304402 4414855 4770329 5089829 5300649 5467686 5541134
65–69 3295231 3405761 3555960 3645737 3813371 4004605 4106617
70–74 2408632 2489142 2581438 2698645 2870848 2914706 3013052
75–79 1588550 1682718 1733306 1774822 1836206 1960249 2027791
80–84 962223 1006531 1054057 1100412 1124821 1123230 1192039
85–89 399472 423506.4 456082.5 502784.8 519828.3 557905.7 582064.4
90–94 118758 128692.8 137625.8 146182.7 161883.4 172212.2 182358.7
95–99 29414 31314.39 32530.13 34764.69 36467.73 40232.09 43599.68
100+ 3227 1706.79 1944.118 1895.151 2065 2130.435 2332.162
Total 80417528 80501838 80683033 81043692 81396535 82024275 82646242
Source: Authors’ calculations based on data in the 6th population census 2010.
Table 2. Proportion of the population at each age section in total population
Age structure 2010 2011 2012 2013 2014 2015 2016

(years)
0–14 13647123 13179252 12905455 12827804 12914013 13314446 13753624
(%) −16.97% −16.37% −16.00% −15.83% −15.87% −16.23% −16.64%
15–64 57964898 58153213 58224635 58310644 58117031 57934558 57742763
(%) −72.08% −72.24% −72.16% −71.95% −71.40% −70.63% −69.87%
65+ 8805507 9169373 9552943 9905244 10365491 10775272 11149855
(%) −10.95% −11.39% −11.84% −12.22% −12.73% −13.14% −13.49%
Source: Authors’ calculations based on data in the 6th population census 2010
Fig. 2. Aging index of Sichuan and China in 2010–2014
aging index in developing countries is expected to be greater than in developed

countries.
In 2010, the AI in China was approximately 53, meaning that there were 53
people aged 65 and over for every 100 persons aged 0–14 years, while the AI
in Sichuan was approximately more than another 12. As Fig. 2 shown, the AI
in both China and Sichuan increased progressively from 2010 to 2014, and the
increment of AI in Sichuan is more than that of China in studied period from
65(2011) to 80(2014).
3.3 Total Fertility Rate Analysis

In the late-1970s started the family planning policy which implemented, espe-
cially in the urban only child policy, China effectively was suppressing the pop-
ulation growth, the fertility rate dropped continually, and in recent years, the
fertility rate has remained at 1.4%–1.5% level for a long time, which have also
accelerated the aging of population. In order to promote the balanced develop-
ment of population, adhere to the basic state policy of family planning, improve
the population development strategy, China fully implement two-children policy,
that is, a couple can have two children. Two-children policy appropriate to slow
down the population aging process, help to solve the aging population.
From the short term, the family planning policy will not influence the pop-
ulation structure of Sichuan temporarily, however, from the long term, it will
have played a positive role on easing the aging population. In the population
analysis and decision-making, the total fertility rate is an important index and
parameter for the prediction of population. It can be directly used to compare
different periods of women’s fertility level, which has important reference to the
long-term macro decision of the population. In order to study population fertility
policy, the following supposes that the other factors do not change in the future
years, and predicts the population structure of Sichuan in different total fertility
rate. Following, we will suppose that the total fertility rate β(t) is respectively
at 1.5, 2.0 and 2.5, and predict the total population and the proportion of 65+
years old population in the future 50 years in Sichuan. The prediction results are
shown as Figs. 3 and 4.
338 J. Hu
Fig. 3. Prediction of total population in 2010–2065
From Fig. 3 we can find that the total population quantity in the future will
increase continuously in the case of high total fertility rate. When the total fer-
tility rate maintains the low level, the total population has a short term growth,
and soon will decrease steadily. In the case of medium total fertility rate, the
total population will maintain stable in the future 50 years. Meanwhile, the pre-
diction of aging population proportion in the future 65 years is shown as Fig. 4.
Fig. 4. Proportion of population at 65+ years old in 2010–2065
As Fig. 4 shown, in the case of low total fertility rate, the proportion of aging
population will continuously increase and finally exceed 25% in 2060. In the case
of medium total fertility rate, the aging population increase slowly, and stabilizes
around 22% after 30 years. When the total fertility rate is high, the proportion
of aging population is relatively slow, and gradually stabilize and maintain in
the below 20% after 30 years of growth.
From the prediction results, we can find that the relatively high fertility rate
will make the population aging keep at a relatively low level, but will cause
the rapid population growth. For example, considering the proportion of aging
population and the total population in the period of 2020–2040, the proportion
of aging population changes from 16% to 26% and the total population decreases
from 83 million to 77 million in the case of low total fertility rate. Meanwhile,
the proportion of aging population only increases 6% from 16% and the total
population increases 13 million from 83 million in the case of high total fertility
rate. The prediction indicates that only when the total fertility rate keeps at 2.0
can a relatively good result appear.
The rate of population aging may also be modulated by migration. Immi-
gration usually slows down population aging, because immigrants tend to be
younger and have more children. On the other hand, emigration of working-age
adults accelerates population aging. The 6th census shows that in 2010 the pop-
ulation mobility rate of Sichuan was 18.92%, lower than the national average
only 0.59% points. Among them, the emigration rate of Sichuan was 23.24%,
and the immigration rate was 14.59%.
4 Conclusions
This paper predicts the population structure in the future six years in Sichuan
by establishing the population development equation. And the results show that
total population in Sichuan will continuously increase and the aging population
is increasingly aggravated. The population aging is an inevitable population
problem. The aging population increase will be a series of problems to society,
we can not only solve it through population policy but also old-age supporting
policy, and social security policy, and regulating economic structure, etc.
In order to analyze the effect of fertility rate on the aging of the population,
we predict the population structure in the future 50 years in Sichuan in the case
of different fertility rate. And the results indicate that appropriate fertility rate
will help to regulate population structure and ease the population aging. When
the total fertility rate is 2.0, the population development in the future in Sichuan
will be relatively stable and the trend of population aging will slow down and
finally keep at the level of 22%.
Acknowledgements. This research is supported by the Scientific Research Founda-

tion of CUIT (Grant No. KYTZ201425) and Center of System science and Enterprise
Development Research (Grant No. Xq16C01).
References
1. Abdessalem T, Cherni HC (2016) Macroeconomic effects of pension reforms in
the context of aging populations: overlapping generations model simulations for
tunisia. Middle East Dev J 8(1):84–108
2. Alves CO, Delgado M et al (2015) Existence of positive solution of a nonlocal logis-
tic population model. Zeitschrift fr angewandte Mathematik und Physik 66(3):943–
953
3. Cui Q, Lawson GJ, Gao D (1984) Study of models of single populations: develop-
ment of equations used in microorganism cultures. Biotechnol Bioeng 26(7):682–
686
4. DESA (2013) World population aging 2013. Technical report, PD Department of
Economic and Social Affairs (DESA), New York
340 J. Hu
5. DESA (2013) World population prospects: the 2012 revision, key findings and
advance tables. Technical report, PD Department of Economic and Social Affairs
(DESA), New York
6. Liu M, Liao D et al (2013) The correctional model of population development
equation. Model Numer Simul Mater Sci 03(4):139–141
7. Muto I, Oda T, Sudo N (2016) Macroeconomic impact of population aging in
Japan: a perspective from an overlapping generations model. IMF Econ Rev
64:408–442
8. Novozhilov AS (2004) Analysis of a generalized population predator-prey model
with a parameter of normal distribution over individuals of the predator popula-
tion. J Comput Syst Int 43(3):378–382
9. Shi LN, Huang HM et al (2009) Aging population trend prediction based on
the gray recurrence dynamic model with equal dimension. Ludong Univ J 25(4):
315–317
10. Yang F, Zhang LD, Shen JF (2013) Further discussion on population growth mod-
els by stochastic differential equations. Adv Mater Res 705:499–503
11. Zhang Y, Gao S et al (2015) On the dynamics of a stochastic ratio-dependent
predator-prey model with a specific functional response. J Appl Math Comput
48(1–2):441–460
12. Zhou A, Sattayatham P, Jiao J (2016) Analysis of a predator-prey model with
impulsive diffusion and releasing on predator population. Adv Differ Eqn 1:1–18
Evaluation of Progressive Team Intervention
on Promoting Physical Exercise Behavior
Xinyan Guo(B)
Economy and Management Faculty in Chengdu Sport Institute,

sindystory@163.com
Abstract. Focusing on parts of urban residents, which were divided

into two groups, intervention group (165 people) and control group (178
people), a “5As” physical exercise behavior intervention process (Ask,
Access, Advice, Assist and Arrange) was applied to the intervention
group people with auxiliary materials including intervention manuals
and exercise prescriptions. The baseline data and the tracking data of
the two groups were collected and analyzed by descriptive statistics, com-
parison of means, analysis of variance and correlation analysis to study
the effectiveness of intervention. The results of tracking survey showed
that residents had better acceptance for the use of intervention. Before
and after the intervention, there was a certain change about the resi-
dents’ physical exercise behavior and there were significant differences
in the behavior variables (P < 0.05). Physical exercise behavior belongs
to preventive innovation which requires corresponding health education
and intervention. The research adopts a step by step mode to implement
targeted intervention measures, and peer groups make it easier to accept
and prove to be efficient for the individuals. And the research provides
a new idea and method for the physical exercise intervention.
Keywords: Physical exercise behavior · Progressive intervention ·

Tracking survey · Urban resident
1 Introduction
In the modern society, sports and human survival and development are more and
more closely related, its significance and function far exceeded the original con-
cepts and categories. Physical activity can effectively change people’s way of life,
prevent all kinds of chronic diseases, and moderate physical activity is one of the
most simple ways to improve and maintain health [3,17,18]. People pay much
more attention to living environment, living styles and the consequent changes of
morbidity and mortality rates [16]. Researches on health management and promo-
tion show a close relationship between residents’ living style and physical health.
In Japan, less than 30% of Japanese do frequent (usually one year or more)
and regular (30 min per time, twice a week) physical exercises. A survey in
342 X. Guo
Australia showed that 53% of the adults didn’t take regular exercises with their
sitting time being 4.7 h per day or more [10]. High income countries like U.S. and
U.K. showed the same results. According to the survey, in U.K., only 39% male
and 29% female adults were doing exercises regularly, moreover, the participation
rate declined with the change of age, work and family [1,11,12].
There are many exercise behavioral intervention studies. Geoff et al. (2010)
conducted a pre-post survey research to 200 British female undergraduates who
did not attended any physical exercises, and analyzed the benefits and cogni-
tive disorders [13]. This study indicated that interventions should give priority
to solve the disorders of participating physical exercises which can be perceived
by individual. On the other hand, interventions should emphasize this cognition
that physical exercises are beneficial to health, thereby promoting modifications of
cognition and behavior for participating physical exercises. In order to study the
influence of PE class to students’ physical activities, Sallis et al. designed a trial
that it made 338 senior undergraduates to attend randomly intervention credit
courses or control credit courses [19]. The results demonstrated that the total vol-
ume of activities including leisure time, strength exercises and flexible exercises in
female group was increased due to interventions, which facilitated female to partic-
ipate healthy physical exercises. However, there was not significant in male group,
because male students were more like sports than female found in basic survey.
About the applied studied of stage-based intervention, Marcus et al. performed
an exercise intervention the Imagine Action campaign for community volunteers
to research the results between pre-exercises and post-exercises, and this study
indicated that 30% of participants in Contemplation and 60% of participants in
Preparation moved into Action, while 31% in Contemplation moved into Prepara-
tion. This study also indicated that less expensive and intensive exercise interven-
tion can be adopted significantly [14]. Bogdan et al. evidenced the effectiveness of
stage-based intervention in their study which aimed at exercises behavioral inter-
vention for 560 orthotic patients [2]. Other researchers also proved the advantages
of stages-matched intervention to stages-non-matched intervention using other
approaches such as Internet aided intervention [4].
However, to make it function well, coordination and cooperation are needed
between individuals, family, society and various departments (such as health,
education etc.). At the same time, it also depends on the scientific method of
promoting behavior to intervene [5,7].
In this paper, Progressive Intervention is adopted to understand the individ-
ual characteristics, strengthen the awareness of individual self-management and
help individuals form a good behavior habits. Then through baseline data and
tracking data, the effectiveness of intervention was analyzed to provide opera-
tional framework and theoretical guidance for the health promotion, especially
for the “sub health” status of the population.
Evaluation of Progressive Team Intervention on Promoting Physical Exercise 343
2 Objects and Methods

2.1 Participants
Stratified random sampling method was used to select the urban residents partic-
ipants from the fitness clubs, parks, squares, community focus points etc. And
they include volunteer solicitation, recruitment network, and random samples
through the local network. Sample size for the baseline survey was 343. Samples
were randomly divided into two groups: intervention group and control group,
165 in the intervention group, and 178 in the control group (Tables 1 and 2).
Table 1. Gender of respondents (baseline data) N = 343
Gender %
Male 47.1
Female 52.9
Table 2. Age of respondents (baseline data) N = 343
Age ≤18 18–29 30–39 40–49 50–59 ≥60

(%) 6.5 10.8 25.7 28.1 19.2 9.7
To ensure the quality of the intervention, staff training was carried out before
the test to ensure that the intervention process was completely. And all these
works are taken out accordance with the framework of the theory hypothesis.
2.2 Test Instruments

The related testing instruments used in the study include the test scale, the
intervention manual and the experiment site, and so on.
The test scale is mainly used to measure the process and the result of the
residents’ physical exercise behavior. This study integrated the health belief the-
ory and the theory of planned behavior and propose that the exercise behavior
outcome of urban resident can be predicted by some factors including behav-
ioral attitude, body norm, perceived subjective control, perceived susceptibil-
ity, perceived severity, perceived behavior benefit, perceived behavior disorder,
behavior clue and behavior intention. The scale is based on the integration of
relevant theory [8]. A self-designed questionnaire was formed on the combination
of the theory of planned behavior and health belief model, the questionnaire con-
tain 4 items of personal basic information (including gender, age, occupation,
education etc.), perceived susceptibility, perceived severity, perceived barriers,
perceived benefits, behavioral cues, behavioral attitude, subjective norm, per-
ceived behavioral control, behavioral intention and behavior, a total of 33 items.
344 X. Guo
Scale by Li Kete five point scale scoring methods: among them, perceived barri-
ers project assigned the remaining items reverse, positive score, namely “strongly
disagree” for 1 points, “disagree” for 2 points, “not sure” for 3 points, “agreed”
for 4 points, “strongly agree” for 5 points.
The effectiveness of the physical exercise behavior at the present stage
(including the action stage and the intensity of action) and the related char-
acteristics were measured [6].
Based on exercises behaviors traits and stages of residents, the invention
handbook was composed of introduction, interpretation and suggestion. The
content is as follow:
(1) Health, sub-health and diseases
The aim of this part is to provide some essential concepts about their physical
fitness for residents so that lay a good foundation for healthy life style. For
example, health can be divided into three states, namely health, sub-health
and diseases. A state of sub-health is either non-health or non-diseases. The
symptoms of sub-health include lack of energy and other weakness, but these
symptoms cannot be diagnosed by clinical examinations.
(2) Introduction of scientific exercise methods
In the part, the benefits of regular exercises will be introduced, including the
improvement of physical fitness, adjustment of psychological factors and so on.
And then to introduce what is regular exercises, how to control the amount of
exercises, how to warm-up and how to eliminate fatigue quickly.
(3) The common exercise methods aerobic exercise
The most common exercise is aerobic exercise that is to improve the cardio res-
piratory fitness and physical fitness. Furthermore, methods and skills of aerobic
exercise will be introduced in this part.
(4) Other exercise methods
These exercise methods do not need complex techniques, fitness facilities, and
extra costs, at the same time these methods are effective, feasible and popular
for public. For example, jogging, running, bicycle riding, aerobics and fitness
path.
(5) Exercise methods of different population
People should perform suitable exercises based on their physical and psychologi-
cal conditions such as ages, gender and vocation. The main contents of this part
is to provide distinct advices for different population.
2.3 Test Procedure
The study is a quasi-experimental design due to the large sample size and the
certain range of exercise behavior intervention. In order to ensure the rationality
and integrity of data analysis, the empirical research is divided into two stages:
the first stage is baseline survey data analysis, and the second stage is tracking
data analysis. The test procedure is shown in Fig. 1.
Determining test variables: exercise behavior and its

influencing factors
Initial Experiment Determine the experimental group: intervention group
and control group
First measurement
Beginning of the First Baseline data test
Month Setting up the intervention means and Distributing
intervention handbook
Second measurement
End of the Second Month
Tracking data test
Intervention according to the Procedure
Third measurement
End of the Forth
Tracking data test
Month
fifth measurement
End of the Sixth
Tracking data test
Month
Fig. 1. Intervention procedure
Firstly, the baseline survey was performed with the permission of subject indi-
viduals. The questionnaire was distributed to the residents, and the characteristics
of the physical exercise behavior and basic information were measured, then the
experimental group of residents were carried out by the intervention staff.
After six month behavior intervention, behavioral tests were carried out every
two months, four times totally. Both groups use the same questionnaire scale.
The survey was conducted in a narrative way of step-by-step introduction, expla-
nation and recommendations.
3 Intervention Process
3.1 Intervention Evaluation
Firstly, Epi Data was used to establish a database, and all the data were checked
after the entry. Then the modeling and analysis of the related research can be
carried out only after the examination of the normal distribution, in which the
reference value of all samples were within the reference range of criteria. And the
absolute value of Skewness coefficient and kurtosis coefficient should be in the 0
to 1 range, and the S-W test P value is greater than 0.05, which demonstrated
that the formal sample can meet the requirements of normal distribution [15].
Urban residents of Sichuan randomly selected for the survey sample (n =
343), and the two subjects showed no significant difference in the results of
exercise behavior and behavioral characteristics variables (p > 0.05).
3.2 Follow-up
Complete physical exercise behavior intervention points can be summed up in 5
words: Ask, Access, Advice, Assist and Arrange (5As).
346 X. Guo
Ask is to inquire about the individual residents to participate in physical

exercise and understand their health status; assess is to measure individual res-
idents’ intention of participating physical exercise, behavior characteristics and
the behavior of the stage; advice is to provide targeted exercise recommenda-
tions, support and help based on the individual physical exercise behavior; at
last follow-up (shown in Fig. 2).
Ask
Arrange “5As” Access

model
Assist Advice
Fig. 2. “5As” Intervention flow chart of physical exercise behavior
After the beginning of intervention, according to the experimental arrange-

ment, “Arrange” is set up in the four follow-up tests. The main purpose is to
understand if the residents still persist in physical exercises, congratulate the
residents who have maintained physical exercise and encourage them to con-
tinue to persist; and give the affirmation to those who maintained a period of
time and then give up, and encourage them to start exercise behavior again. At
the same time, according to the interview, guidance and help are given, in case
of giving up.
4 Intervention Outcome
Based on the theoretical hypothesis, the exercise behavior intervention exper-
iment is implemented to analyze the changes of physical exercise behavior
between intervention group and the control group before and after the test
period.
Firstly, the slope homogeneity of the covariance analysis is tested. The vari-
ance analysis model were analyzed in which the variables of behavioral charac-
teristics after intervention were dependent variables, the group variables (1 =
intervention group and 2 = control group) were independent variables. The
result showed that the interaction between independent variables and covariates
was not significant (P > 0.05), that is, the assumption of the slope homogeneity
(Table 3).
The result showed that the effectiveness is not significant from previous per-
ceived susceptibility, previous perceived severity, previous perceived benefits,
previous perceived barriers, previous behavioral attitude, previous subjective
norm and previous perceived behavioral control variable (P > 0.05). But some
variables’ effectivenesses are significant (P < 0.05), such as previous behavioral
cues and previous behavioral intention. Results are shown in Table 4.
Table 3. Covariance pre-analysis results (interaction)
Source of variance P value

∗
Group Previous behavioral cues 0.363
Group∗ Previous perceived susceptibility 0.627
Group∗ Previous perceived severity 0.351
Group∗ Previous perceived benefits 0.094
∗
Group Previous perceived barriers 0.126
Group∗ Previous behavioral attitude 0.521
Group∗ Previous subjective norm 0.438
Group∗ Previous perceived behavioral control 0.095
Group∗ Previous behavioral intention 0.766
∗
The difference is significant at 0.05 level.
∗∗
Table 4. Covariance pre-analysis results (Intervention effectiveness)
Source of variance P value

Previous behavioral cues 0.044∗
Previous perceived susceptibility 0.426
Previous perceived severity 0.233
Previous perceived benefits 0.068
Previous perceived barriers 0.091
Previous behavioral attitude 0.084
Previous subjective norm 0.258
Previous perceived behavioral control 0.086
Previous behavioral intention 0.032∗
∗
∗∗
Therefore, the exercise behavior variables of the two groups were compared
and analyzed (Tables 5 and 6). The results indicated that the exercise behavior
of the intervention group was significantly higher than that of the control group,
which means the progressive intervention is effective to promote the formation
of physical exercise.
348 X. Guo
Table 5. Comparative analysis of the effectiveness before and after the intervention
Variable Intervention group Control group P value

Perceived susceptibility 3.35 ± 0.0.583 2.83 ± 0.0.518 0.032∗
Perceived severity 2.89 ± 0.167 2.49 ± 0.0.273 0.050∗
Perceived benefits 3.05 ± 0.0.267 3.68 ± 0.0.415 0.021∗
Perceived barriers 3.58 ± 0.0.326 2.94 ± 0.0.572 0.006∗∗
Behavioral attitude 4.25 ± 0.0.597 3.92 ± 0.0.486 0.006∗∗
Subjective norm 4.15 ± 0.0.278 4.00 ± 0.0.436 0.000∗∗
Perceived behavioral control 4.83 ± 0.0.352 4.70 ± 0.0.492 0.000∗∗
∗
∗∗
Table 6. Difference analysis of adjusted mean
Variable (I) Group (J) Group Mean difference (I−J) P Value

Behavioral cues Intervention group Control group 0.42 0.022∗
Behavioral intention Intervention group Control group 0.36 0.003∗∗
∗ The difference is significant at 0.05 level.
∗∗ The difference is significant at 0.01 level.
5 Discussion
The subject was intervened by progressive intervention, and the effectiveness of

the intervention was confirmed by the comparative analysis of the intervention
group and the control group.
The influence factors of urban residents’ exercising behavior are multidi-
mensional, including 9 factors: behavioral attitude, subjective norm, perceived
behavioral control, perceived susceptibility, perceived severity, perceived ben-
efits, perceived barriers, behavioral cues and behavioral intention. Perceived
susceptibility and perceived severity of physical exercise have effects on per-
ceived benefits and behavior attitude: perceived susceptibility, perceived sever-
ity, behavior barriers and perceived benefits have a positive impact on behavior
attitude; perceived benefits have a negative role in behavior barriers but positive
in behavior attitude; behavior attitude, subjective norm and perceived behav-
ioral control function positively to behavior intention and behavior attitude does
positive role to perceived behavioral control at the same time; behavior inten-
tion, perceived behavioral control, perceived benefits, barriers and cues exerts
positive effects on behavior itself. Based on the survey results, in the baseline
data, the majority of urban residents are aware of the benefits from physical
exercise, but they are not necessarily involved in the fitness. In other words,
the promoting effect which perceived benefits impacted on behavior intention
is still to be strengthened. For there are more worried about, such as ways of
physical exercises and strength, physical training environment, time cost, etc.
In addition, although the residents hold a positive attitude for physical exercise
benefits, but due to the limited understanding of their own physical condition
and the relevant authority on argument, there is no clear subjective feeling for
the contrast effect not to participate in fitness activities, and actively participate
in fitness activities. But the intervention group was improved in these aspects.
In the four follow-up tests, according to the time schedule, the main purpose
set of the “Arrange” is to understand whether the residents are still persist in
this behavior or not [9]. In the process of behavior intervention, the people who
did not participate in physical exercise should be divided into stages to carry out
the work, such as conducting appropriate in-depth interviews with individuals,
giving psychological guidance, emphasizing the importance of exercise behavior
in their consciousness, at the same time, enhancing the confidence of physical
exercise by demonstration role or embodiment contrast, and helping them realize
the change of behavior as early as possible. In order to pursue a better coop-
eration of individual residents, access to information about physical exercise,
intervention staffs may need to make some special arrangements for the environ-
ment of the conversation. For example, put some health handbooks or posters
about physical exercise, display some sports souvenirs on desk, to help residents
receive health intervention, and let them feel about that physical exercise is a
very important work in the process of communication.
6 Conclusions
This study indicated that individual had different psychological cognitive status
and behavior intention level, and faced different difficulties and obstacles. Based
on the survey results, in the baseline data, the majority of urban residents are
aware of the benefits from physical exercise, but they are not necessarily involved
in the fitness. In other words, the promoting effect which perceived benefits
impacted on behavior intention is still to be strengthened. For there are more
worried about, such as ways of physical exercises and strength, physical training
environment, time cost, etc. In addition, although the residents hold a positive
attitude for physical exercise benefits, but due to the limited understanding
of their own physical condition and the relevant authority on argument, there
is no clear subjective feeling for the contrast effect not to participate in fitness
activities, and actively participate in fitness activities. But the intervention group
was improved in these aspects.
National fitness activities is mass fitness activities, to promote physical and
mental health. Urban residents as the main subject of national fitness activi-
ties, their views of the community sports, support and participation in decision-
making, directly impact on the mass sports development. Using progressive inter-
vention mode can make people to clear the relationship between physical exercise
and promote the health, to deepen the understanding of physical exercise behav-
ior. in this study the intervention based physical exercise behavior intervention
points (5As), and the result meant progressive intervention is effective to pro-
mote the formation of physical exercise.
350 X. Guo
References
1. Begg S, Khor S, Bright M (2003) Risk factor impact on the burden of disease and
injury in Queensland. Queensland Burden of Disease and Injury Circular Series
2. Bogdan RC, Biklen SK (1982) Qualitative Research for Education. Allyn and
Bacon, Boston
3. Physical Activity Guidelines for Americans Committee et al (2008) Physical Activ-
ity Guidelines for Americans. US Department of Health and Human Services,
Washington, DC
4. French KE, Nevett ME (1997) Knowledge representation and problem solution in
expert and novice youth baseball players. Res Q Exerc Sport 67(4):386–395
5. Guo X (2015) Application of diffusion of innovation theory on the intervention
research of exercise behaviors in urban residents. Mod Prev Med 31(3):202–208
6. Guo X (2015) Construction and Validation of the Integration Model of Planned
Behavior Theory and Health Belief Model. Springer, Heidelberg
7. Guo X (2016) Examination on the discontinuity of physical exercise behavior stages
and its implication for intervention. J Chengdu Sport Univ 42(3):42–47
8. Guo X, Xu J (2010) Making and testing of the scale to measure the change stage
for the exercising behavior of urban residents based on a survey into some Sichuan
urban communities. J Chengdu Sport Univ 5:71–74
9. Guo X, Zhang W (2017) Stages and Processes of Self Change of Exercise Behavior:
Toward an Integrative Model of Change. Springer, Singapore
10. Harper C, Epidemiologist C (2008) Challenges and opportunities - key challenge 2:
make prevention a key to reducing health inequality. In: The Health of Queenslan-
ders 2008-8211; Prevention of Chronic Disease: Second Report of the Chief Health
Officer, Queensland. Queensland Burden of Disease and Injury Circular Series
11. Department of Health (2005) Choosing activity: a physical activity action plan.
Healthex Specialist
12. Department of Health (2005) Choosing health: making healthy choices easier. Pri-
mary Health Care
13. Lovell GP, Ansari WE, Parker JK (2010) Perceived exercise benefits and barriers
of non-exercising female university students in the United Kingdom. Int J Environ
Res Pub Health 7(3):784–798
14. Marcus BH, Banspach SW (1992) Using the stages of change model to increase the
adoption of physical activity among community participants. Am J Health Promot
AJHP 6(6):424–429
15. Markus KA (2012) Principles and practice of structural equation modeling by rex
b. kline. Struct Equ Model Multidisciplinary J 19(3):509–512
16. Mesbah M (2016) Measurement and analysis of quality of life related to environ-
mental hazards: the methodology illustrated by recent epidemiological studies. Int
J Manage Sci Eng Manage 11(2):1–16
17. World Health Organization (2008) Prevalence of insufficient physical activity, age
15+, age-standardized: both sexes. World Health Organization, Geneva
18. World Health Organization (2013) Bull World Health Organ. World Health Orga-
nization
19. Sallis JF, Calfas KJ (1999) Evaluation of a university course to promote physical
activity: project grad. Res Q Exerc Sport 70(70):1–10
SEM-Based Value Generation Mechanism
from Open Government Data
in Environment/Weather Sector
Xiaoling Song1,2(B) , Charles Shen2 , Lin Zhong3 , and Feniosky Peña-Mora2

1
School of Economics and Management, Nanjing University of Science
and Technology, Nanjing 210094, People’s Republic of China
songxiaoling426@163.com
2
Advanced ConsTruction and InfOrmation techNology (ACTION) Laboratory,
Department of Civil Engineering and Engineering Mechanics, Columbia University,
New York 10027, USA
3
Abstract. Environmental issues are harmful effects of human activities

on the biophysical environment. More and more citizens are engaged in
the open data movement for different purposes, in particular, emerging
private companies are being built on Open Government Data (OGD) to
predict extreme weather events for environmental protection, to satisfy
more social requirements (e.g. job creation) and to make economic prof-
its, simultaneously. Emerging companies are utilizing OGD to generate
values. Building on a synthesis of the OGD literature and established
theories of value generation, we develop a structural equation modelling
(SEM)-based model to explore the causal relationship between OGD and
value generation. This study constructs the conceptual structural equa-
tion model and lays a foundation for the upcoming research.
Keywords: SEM · Value generation · Open government data · Envi-

ronment/weather sector
1 Introduction
Environmental issues are harmful effects of human activities on the biophysical
environment. Major current environmental issues may include climate change,
pollution, environmental degradation, and resource depletion etc. Global warm-
ing, or climate change as it’s now commonly referred to, is the greatest envi-
ronmental threat we’ve ever faced. There is little doubt that climate change is
contributing to the extreme weather disasters we have been experiencing, such as
worsening air quality, searing heat, and more frequent extreme weather events
(e.g. raging storms, ferocious fires, severe droughts, and punishing floods). It
threatens our health, communities, economy, and national security [7], for exam-
ple, the climate change-related heat mortality in the first half of 2012 is a part
352 X. Song et al.
of a deadly trend. In 2011, at least 206 people died from extreme heat, up
from 138 fatalities in 2010 and nearly double the 10-year average, according to
the National Oceanic and Atmospheric Administration [8]. If we don’t do more
to reduce fossil fuel emissions and other heat-trapping greenhouse gases that
are making heat waves more intense, more than 150,000 additional Americans
could die by the end of this century due to excessive heat. Heat-related death is
just one deadly side effect of extreme weather tied to climate change [7]. Other
extreme events also have side effects, such as storms can cause drowning, con-
taminate drinking water and result in outbreaks of infectious diseases; heat and
ozone smog increases respiratory diseases (e.g. asthma) and worsens the health
of people suffering from cardiac or pulmonary disease.
According to NRDC (2015), the global warming has mainly six side effects
on human health, including more vector-borne disease, more water-borne ill-
ness, longer allergy seasons, declining crops, air pollution and more heatwaves
[7]. Therefore, as we work to stop climate change and prevent those associated
impacts, we are also developing policies to minimize human suffering.
At present, worldwide nations have begun taking steps to combat this grow-
ing threat, working toward an international agreement in which every country
on earth plays its part. Many of the world’s largest polluters have stepped up
with significant commitments, amplified by efforts from cities, businesses, sports
leagues, non-governmental organizations (NGOs), and many other individuals
and groups that have responded to the urgent need for a climate action. For
example, since carbon pollution fuels climate change, drives extreme weather,
threatens communities and cuts too many lives short, the U.S. Environmental
Protection Agency has developed the Clean Power Plan, which sets the first
national limits on carbon pollution from power plants and provides states with
the flexibility to meet them.
The NRDC has suggested more solutions to address the global warming
issues, such as setting limits on global warming pollution; investing in grean
jobs and clean energy; driving smarter cars; creating green homes and buildings;
as well as building better communities and transportation networks. These solu-
tions will motivate our lawmakers to quit ignoring climate change and start lim-
iting carbon pollution that is heating our planet and increasing the intensity of
extreme weather. However, the government cannot be expected to do everything
to protect the environment. In cooperating with the private sector as a partner
can make sense based on the principle of partnership and appropriate process.
Opening data is recognized as a driver of efficiency and a vehicle for increasing
transparency, citizen participation and innovation in society. Open data can help
improve food security, healthcare, education, cities, the environment, and other
public and private services that are critical to development. Taiwan environ-
ment protection agency (TEPA) views environmental data as a strategic asset
to protect our environment. At present, more and more citizens are engaged in
the open data movement for different purposes, in particular, emerging private
SEM-Based Value Generation Mechanism from Open Government Data 353
companies are being built on Open Government Data or Open Environmental

Data to predict extreme weather events for environmental protection, to sat-
isfy more social requirements (e.g. job creation) and to make economic profits,
simultaneously.
Today, we are generating and storing more data than at any other time in his-
tory. The public sector and especially the various government branches are one of
the main sources of data. Government data has traditionally been accumulated
in protected repositories and registries as public record and a matter of civil
order. Consequently, over centuries, archived government data has had limited
access via proprietary interfaces and often cumbersome fee-based procedures.
With time going on, governments have started to look into the prospects of pro-
viding open access to their data repositories. The concept of Open Government
Data (OGD) refers to government data defined as “data and information pro-
duced or commissioned by government or government controlled entities” that
are opened up for use and re-use by public and private agents alike [3]. These
data encompass both scientific data about the environment (from Earth Obser-
vation and other fields) and other public sector information, including diverse
topics such as demographics, health and crime. In the current-used terminology,
OGD does not include data that are subject to valid privacy, security or privilege
limitations, as governed by other statutes. Government data sets are an inter-
esting subset of open data because such subsets have already been collected for
specific use, have been paid for by taxpayers, are relevant and offer value beyond
what is captured from the originally intended use. When opened up, government
data become a common, shared resource (i.e., public good) that is provided by
the government. In our digital age, opening up government data to everyone puts
power in the hands of citizens [3]. Making environmental data open is an effec-
tive method to guide the local agencies for policy-making and decision-making,
and to increase the citizen participation for environmental protection.
Under the OGD circumstances, many private companies are set up to make
business values in various sectors, such as business & legal services, data &
technology, education, energy, finance & investment, lifestyle & consumer. It
becomes common to observe that data itself has no value, the value comes from
how data is used and analyzed in particular businesses. Specifically, private
sectors, non-government organizations, businesses and APP developers in the
environment/weather are encouraged to use the datasets already published to
create economic, social values, and in particular environmental values (Sustain-
able Objectives). Many pioneering companies are using OGD as a key resource
to generate values, for example, OGD-driven business models are proposed to
evaluate how OGD is used as a key resource to understand the value generation
process in e-business industry. The Open Data 500 Global Network is an inter-
national network of organizations that seek to study the use and impact of open
data. Coordinated by the Governance Lab (GovLab) the OD500 Global Network
enables participating organizations to analyze open data in their country in a
manner that is both globally comparative and domestically specific. The OD500
Global Network starts from the assumption that only by mapping the use of open
354 X. Song et al.
data within and across countries, can new approaches for understanding the eco-
nomic and social impact of open government data be generated. The Open Data
500 U.S.A is the first comprehensive study of U.S. companies that use OGD
to generate new business and develop new products and services [6]. Similarly,
the Open Data 500 Korea, the Open Data 150 Canada, the Open Data 500
Australia, the Open Data 500 Mexico, the Open Data 500 Italy projects have
been launched. Many companies from different sectors (e.g. healthcare, food and
agriculture, finance etc.) are developed and generate value based on OGD from
department of defense, department of energy etc. Our study show that around
many environmental companies uses OGD to generate value, including the eco-
nomic, social and environmental values.
Despite the potential significance of OGD, emphasized by an abundance of
anecdotal evidence, we could not identify many studies on how OGD will con-
tribute to value generations. To-date, the economic and social impact of open-
data policies remains largely unclear, and there are scant empirical data available
on the effects of the various policy approaches, thus leaving policy makers with-
out the facts they need to assess and improve these policies. Some studies were
conducted to illustrate how data are used as a key resource to enhance their
core values [5,9,10]. However, more studies are needed to figure out how OGD
is used as a key resource to generate values in specific domains, such as the
environment/weather sector.
2 Research Motivation and Approach
Although a lot of anecdotal evidence has demonstrated the potential impor-

tance of OGD, it is still a mystery that how OGD will contribute to value
generations in various sectors. Accordingly, our research on Open Government
Data for Value Generation aims to address the question: How can use of OGD
stimulate value generation in environment/weather companies? Building on a
synthesis of the OGD literature and established theories of value generation,
we develop a structural equation modelling (SEM)-based model to explore the
causal relationship between OGD and value generation. The model is proposed
to provide some guidance for the establishment of more companies to address
more environmental issues. This research is planned in three phases. In the first
phase the global open government data landscape in environment/weather sec-
tor is surveyed using secondary data from Open Data 500 projects. Accordingly,
a conceptual SEM model is to be established for further research. In the second
phase, questionnaires are designed and investigations conducted to collect data
for model calculation and validation. In the third phase of this research, we plan
to propose and implement appropriate OGD-driven companies in other countries
to assist the governments for environmental protection.
This paper, however, only looks into the first phase of the planned research
where it attempts to survey the landscape of OGD-driven environmental
companies in USA, Australia and South Korea and establish a conceptual struc-
tural equation model for further validation.
3 OGD-Driven Companies in Environment/Weather

Sector: The Current Situation
Many countries are entering the mainstream of open data movement. A number
of open data benchmarks have been developed such as the World Bank’s Open
Data Readiness Assessment (ODRA), World Wide Web Foundation’s Open Data
Barometer (ODB, 2015), Open Knowledge Foundation Network’s (2014) Open
Data Index (ODI) and Capgemini Consulting’s (2013) Open Data Economy
(ODE), to name few global and widely used benchmarks. However, each of these
benchmarks serves a different purpose and focus. Susha et al. (2014) suggest
that ODB provides a more comprehensive perspective since it not only includes
measures at various stages like readiness, implementation, and impact but also
highlight the importance of involvement of major stakeholders and challenges
throughout the open data process [9]. According to ODB (2016), a scaled score
is an indication of how well a country is doing against other countries in getting
the basics of open data readiness, implementation and impact [4]. According to
the ODB (2014) technical handbook, When evaluating the ODB score, the open-
ness of environmental data (D14 element) is considered based on the data on
one or more of: carbon emissions, emission of pollutants (e.g. carbon monoxides,
nitrogen oxides, particulate matter etc.), and deforestation. This study take 34
companies from US, Australia and South Korea into consideration due to the fol-
lowing two reasons. On one hand, the ODB rankings (as demonstrated in Fig. 2)
show that US, Australia and South Korea are in the top ten countries in terms
of open data readiness, implementation and impact. On the other hand, due to
the high data openness in US, Australia and South Korea, many environmental
companies participated in the Open Data 500 USA project, the Open Data 500
Australia project, and the Open Data 500 Korea project.
The 34 OGD-driven companies in environment/weather sector are surveyed
and partially shown in Table 1, which present social impact, revenue (economic
impact), and environmental impact of different companies. It is noted that the
34 companies use OGD from various official sources and their main purposes are
to address severe environmental issues. The environmental issues to be addressed
and other missions to be accomplished are also listed.
The aim of this study is to understand the causal relationship of why the
OGD is beneficial for value generation from the environmental, social and eco-
nomic aspects (Table 2).
356
Table 1. OGD-driven companies in environment/weather sector (Part 1)
No Company name City state Environmental Social impacts Revenue Data sources Purposes Websites
country impacts sources
1 AccuWeather State Environmental Environment Subscriptions National (1) Provide weather
College protection, and climate Weather Service, forecasts (2) Deliver http://www.
PA US Climate change change Monitoring enterprise solutions (3) accuweather.com
prevention stations Cooperate relationships
X. Song et al.
between government
weather agencies and
the weather industry
2 CoolClimate Berkeley Environmentally Cleaner living Data analysis Multiple (1) Provide
CA US sustainable environment for clients, government open decision-making tools http://
development Database data sources and programs (2) coolclimate.
licensing, (e.g. Design tailored climate berkeley.edu
User fees for environmental solutions (3) Motivate
web or records, land use low-carbon choices (4)
mobile access information) Design climate actions
and programs
3 Earth Networks Germantown Extreme weather Environment Weather data Neighborhood- (1) Operate weather
MD US (e.g. tornadoes, and climate analysis for level sensors, observation, lightning www.
cyclones), change organizations National Oceanic detection and climate earthnetworks.
Climate change and Atmospheric (green gas) networks com
detection Administration, (2) Enable enterprises
National and consumers to make
Weather Service informed decisions
4 Environmental Milford Environmental Environment Data U.S. (1) Provide robust data
Data Resources CT US problem report and climate collection Environmental (2) Provide smarter http://www.
change outsources, Protection workflow tools for data edrnet.com
Data Agency, Multiple management
management government open
for clients data sources
(e.g.
environmental
records, land use
information)
No Company name City state Environmental Social impacts Revenue Data sources Purposes Websites
country impacts sources
5 Bass Coast Bass Environmentally Citizen Contributions/ Department of (1) Support healthy
Landcare VIC AU sustainable engagement and donations, Environment and resilient http://www.
Network Inc development participation, Government Agencies, Local ecosystems (2) basscoastlandcare.
educational contract, Governments, Deliver sustainable org.au
opportunity, Membership Organizations agricultural and
good governance fees, environmental
Philanthropic management
grants practices
6 Enviro-dynamics Hobart Environmental Citizen Consulting Department of (1) Provide
Pty Ltd TAS AU protection, engagement and Environment environmental www.
Climate change participation, Australian solutions (2) Help enviro-dynamics.
prevention educational Bureau of business steer com.au
opportunity Statistics through the green
tape (3) Support
government to
deliver timely and
sustainable
environmental
policy and practice
outcomes (4)
Provide agricultural
and environmental
data analysis
SEM-Based Value Generation Mechanism from Open Government Data
357
358
Table 2. OGD-driven companies in environment/weather sector (Part 2)
No Company name City state Environmental Social impacts Revenue sources Data sources Purposes Websites
country impacts
7 Parklands Wodonga Environmental Citizen Contributions/ New South (1) Tackle www.parklands-
Albury Wodonga VIC AU protection, engagement and donations, Wales environmental alburywodonga.
Ltd Climate change participation, Government Government challenges (2) org.au
X. Song et al.
prevention educational contract, Care for public

opportunity, Philanthropic land (3) Work
good governance, grants together with
public safety businesses,
neighbors and all
levels of
government
8 Southern New Armidale Environmental Citizen Contributions/ Agriculture, (1) Lead, http://snelcc.
England NSW AU protection, engagement and donations, ABARE, connect and org.au/web3/
Landcare Climate change participation, Membership Environment, enable
prevention educational fees, Mapping, OEH communities to
opportunity, Philanthropic threatened meet sustainable
food access and grants, species living goals
supply Government
grants
9 Ecobrain Geumcheon-gu Environmental N/A Consulting, Data Korea (1) Develop as
Seoul KR protection, analysis for Meteorological an environment www.ecobrainkr.
Climate change clients Administration monitoring com
prevention system (2)
Predict
expansion paths
of pollutants
from plants (3)
Offer
company-specific
contributions
No Company name City state Environmental Social impacts Revenue sources Data sources Purposes Websites
country impacts
10 SBIS Yeongdeungpo- Environmental Citizen Consulting, Data Korea (1) Provide
gu protection, engagement and analysis for Meteorological reliable systems http://www.sbis.
Seoul KR Climate change participation clients Administration for customers co.kr
prevention
11 Softworx Nowon-gu Environmental Consumer Advertising, Korea (1) Offer
Seoul KR protection, empowerment Software Environment information on https://play.
Climate change licensing Corporation air pollution and google.com/
prevention fine dust store/apps/
details?id=com.
softworx.cai
12 w365 Gangnam-gu Environmental Consumer Consulting, Data Korea (1) Provide
Seoul KR protection, empowerment analysis for Meteorological weather www.w365.com
Climate change clients Administration information
prevention
SEM-Based Value Generation Mechanism from Open Government Data
359
360 X. Song et al.
Fig. 1. Top ten countries in the ODB 3rd edition ranking (Source: 2016 ODB global
report)
4 SEM Conceptual Model for Value Generation

OGD-driven companies in environment/weather sector apply OGD as a key
resource in their business to generate sustainability values. Accordingly, it is
necessary to reflect the causal relationship between OGD and the generated
values. In the research, the SEM is proposed as a cause -and-effect diagram to
illustrate how OGD is used to generate values.
4.1 Basic Structural Equation Modeling
Since Bentler’s appeal to apply the technique to handle latent variables (i.e.
unobserved variables) in psychological science, structural equation modeling
(SEM) has become a quasi-routine and even indispensable statistical analysis
approach in the social sciences. The emergence and development of SEM was
regarded as an important statistical development in social sciences in recent
decades and this “second generation” multivariate analysis method has been
widely applied in theoretical explorations and empirical validations in many
disciplines. Structural Equation Models with latent variables are extensively in
measurement and hypothesis testing [1,2]. Compared with other statistical tools
such as factor analysis and multivariate regression, SEM carries out factor analy-
sis and path analysis simultaneously, since it can (1) measure and accommodate
errors of manifest variables (i.e. observed variables); (2) represent ambiguous
constructs in the form of latent variables (i.e. unobserved variables) by using
several manifest variables; and (3) simultaneously estimate both causal relation-
ships among latent variables and manifest variables. In addition, SEM can also
provide group comparisons with a holistic model, resulting in much more vivid
impressions than traditional ANOVA. SEM can also handle longitudinal designs
when time lag variables are involved.
As introduced above, SEM describes and tests relationships between two
kinds of variables – latent variables (LVs) and manifest variables (MVs). Latent
variables cannot be observed directly due to their abstract character. In con-
trast, observed variables contain objective facts and easier to measure. Sev-
eral observed variables can reflect one latent variable. As presented in Fig. 2, a
structural equation model usually consists of two main components, a structural
model and several measurement models. A simple measurement model includes
a latent variable, a few associated observed variables and their corresponding
measurement errors. The structural model consists of all LVs and their interre-
lationships. For model development purposes, some researches aim to validate
their assumptions of a dimensional framework of one or several discriminant LVs,
while others aim to elicit the causal relationship between the LVs. Confirmatory
factor analysis (CFA) with correlating latent variables satisfies the former pur-
pose, while these correlations need to be replaced by directional relationships for
the latter.
Figure 4 provides a simple example of a structural equation model investi-
gating the effect of LVs ξ1 , ξ2 and ξ3 on LVs η1 , η2 and η3 , and where several
MVs are used to represent the LVs. The MVs are shown in rectangles, the LVs in
ellipses, measurement errors in circles and with arrows indicating the direction
of the effects. If directional arrow between ξ and η is replaced by a correlation
two-way arrow, the model is a CFA and its purpose is to test whether MVs can
represent LVs well (i.e. convergent validity) and whether ξ and η are different
(i.e. discriminant validity). The basic concepts and principles of SEM are now
well established with the help of early explorations by researchers, structured
textbooks, well developed soft programs (e.g. LISREL, EQS and AMOS).
Fig. 2. A structural equation model diagram

362 X. Song et al.
4.2 Model Path Analysis
In order to establish a cause -and-effect diagram to illustrate how OGD is used

to generate values, a path analysis is conducted, along with which we decide
necessary latent variables and manifest variables based on a large amount of
literature reviews.
(1) Constructs of OGD that are important in value creation
According to data-intensive science, data is prone to generate value through a
value chain of big data, including value capture, value curation, value analysis
and value visualization. To make sure that decision makers are able to produce
information, knowledge or wisdom by utilizing various data analytics techniques
and technologies. OGD is applied as a core resource for creating economic profits
for themselves, bringing social values for the society and promoting environmen-
tal protection. In order to find out in what aspects OGD is important when firms
use OGD to guide decision making or enhance transparency for value generation.
The following features of data are essential for whether or not they are able to
create values.
1 Data quality & Integrity
OGD exposes data to public view in a way that allows for crowdsourced and
quality control. Poor quality data can mislead for decision for environmental
data-driven companies.
2 Data disaggregation
OGD has many elements of being disaggregated and being highly accessible to
many or even any party that wishes to use it. To some extent, disaggregated data
can provide a better comparative picture of what works and help inform and pro-
mote evidence-based policy-making or decision making in environment/weather
sector.
3 Data timeliness
OGD allows for private-sector and citizen contributions to open datasets that
can make data more timely and relevant. In order to achieve real-time weather
report, the data cycle has to match the decision cycle for better decision-making.
Environmental data timeless in environment/weather sector is quite important
for firms to make decisions, produce new products or provide new services.
4 Data transparency and openness
The promotion of OGD and tightened feedback loops between data users and
data suppliers (namely government or government-governed entities). Publicly-
funded datasets, as well as data on public spending, should be available to other
public ministries or the general public. With high data transparency and open-
ness, data should be made public in ways that encourage greater use and be
complete, machine-readable, freely available for reuse without restrictions, and
transparent about underlying assumptions. Environmental data transparency
and openness in environment/weather sector is quite important for firms to
make decisions, produce new products or provide new services.
5 Data usability and curation

Through OGD roundtables and similar events, data users can provide feedback
on open datasets that makes them more accessible and useful. Data architecture
should place great emphasis on user-centered design and user-friendly interfaces.
Communities should be fostered to develop new tools that can translate raw data
into something meaningful to a broader constituency of non-technical potential
users.
6 Data protection and privacy
The challenge of opening government datasets while anonymizing sensitive infor-
mation on health or environment sets a high bar for developing data privacy
safeguards. Clear international norms and must be developed to protect data.
The data protection is quite important when environmental companies utilize
data in their daily business for weather forecast or policy suggestions.
7 Data governance and independence
The open data charter, open government partnership, and other international
efforts are beginning to provide a governance structure for open data. Data
quality should be protected and improved to ensure they are functionally
autonomous, and independent of political influence.
8 Data resource and capacity
National statistical systems should be established that are capable of producing
high quality statistics in line with global standards and expectations. Under
these circumstances, people can make sure that the companies are able to provide
reliable suggestions and conduct right measures for environmental protection.
(2) Constructs of value creation mechanisms
With rapid development of the society, data has become an important resource
for creating values and we have entered the big data era. Most firms are uti-
lizing big data for creating values and data-intensive science has become a new
discipline. Data, in nature, has no value and data analysts aim to extract val-
ues, information, knowledge and wisdom by applying various techniques learning,
data mining and technologies. It is the job of computer scientist to develop tools,
but it is the practitioners’ job to make decisions, provide new offerings or enhance
transparency for creating values. The complementarity between the superior
computational capabilities provided in the data analytics process and the higher
judgmental capabilities of firm managers is quite important for realizing val-
ues from big data. Our study mainly consider three value creation mechanism
based on literature reviews, and they are guiding decision makings, providing
new offerings and enhancing transparency.
1 Decision making
Decision making is quite important for firms to create values. By making better
decisions based on data analytics, managers are able to achieve more economic
or social values. Data-driven decision-making positively affects value through
generation of new information, knowledge and wisdom.
364 X. Song et al.
2 New offerings
Innovation is the source of value creation in Schumpeter’s economic theory, bring-
ing about novel combinations of resources, new production methods, as well as
new products and services, which, in turn, lead to the transformation of markets
and industries, thus increasing value. Data-driven innovation positively affects
value through generation of new knowledge, new processes, services and prod-
ucts, and new businesses.
3 Transparency & accountability
Most definitions of transparency recognize the extent to which an entity reveals
relevant information about its own decision processes, procedures, functioning
and performance. However, opening access to chosen public documents does not
necessarily contribute to a transparent government. By utilizing OGD for compa-
nies in environment/weather sector, it is important to promote the transparency
and accountability of OGD from the environment sector.
(3) Constructs of values for OGD-driven companies in environment/weather
sector
Two types of value are frequently discussed: economic value, defined as the
worth of a good or service as determined by the market, and social value, which
is created when resources, inputs, processes or policies are combined to gener-
ate improvements in the lives of individuals or society as a whole. The main
difference is that our research emphasizes the environmental impacts of those
companies from the environment/weather sector as the 34 companies make great
contributions to environmental protection and disaster prevention. The environ-
mental, economic and social values are combined to achieve the sustainability
goal of the world, as shown in Fig. 3. The environmental impacts mainly include
pollution reduction, natural resource conservation and climate change resilience.
Fig. 3. Constructs of values

OGD Enablers
Data quality Pollution
& Mechanisms reduction
Integrity
Sustainable Values Natural
resourse
Data
Disaggregation
Decision conservation
Making Climate
Environmental change
Data Timeliness
Impacts resilience
Data
Profit
transparency
&
Openness
Cost savings
Economic
Data usability New offerings
& Impacts
Curation Research &
development
Data protection
& Citizen
Privacy engagement &
participation
Data governance
& Transparency Social
Independence Impacts Education
&
Data source Accountability
Health &
&
Capacity standard of
living
Fig. 4. Path analysis
The economic impacts are reflected in profits, cost savings, and research & devel-
opment. The social impacts are categorized as citizen engagement & participa-
tion, education, and health & standard of living.
As previously discussed, the conceptual SEM model-based value generation
methodology is established in Fig. 1.
5 Conclusions and Recommendations
Intensive research has been conducted to figure out the importance of big data
or massive OGD, among which value chain analysis approach and value stream
mapping method are extensively used. Both technical workers and management
practitioners are key actors in finding out how big data or OGD contributes to
value generations in various sectors. The beginning stage of our research on Open
Government Data for Value Generation aims to address the issue of how OGD
stimulates value generation in environment/weather companies. We attempt to
analyze the causal relationship between OGD and sustainable values driven by
OGD based on a structural equation model.
This paper, however, only looks into the first phase of the planned research
where it attempts to survey the landscape of OGD-driven environmental com-
panies in USA, Australia and South Korea and establish a conceptual struc-
tural equation model for further validation. Further research is needed to further
enrich the current study.
366 X. Song et al.
Acknowledgements. This research was supported by the Program of China Schol-

arship Council (No. 201506240179).
References
1. Bagozzi RP, Yi Y (1988) On the evaluation of structural equation models. J Acad
Mark Sci 16(1):74–94
2. Bagozzi RP, Yi Y (2012) Specification, evaluation, and interpretation of structural
equation models. J Acad Mark Sci 40(1):8–34
3. Foundation OK (2016) Open data index (2016)
4. Fundation WWW (2016) Open Data Barometer, 3rd edn. (2016). http://www.
opendatabarometer.org/report/about/method.html
5. Magalhaes G, Roseira C, Manley L (2014) Business models for open government
data. In: Proceedings of the 8th International Conference on Theory and Practice
of Electronic Governance. ACM, pp 365–370
6. Network TOG (2016) Open data 500 U.S. (2016). http://www.opendata500.com/
7. (NRDC) NRDC (2016) Our stories (2015). https://www.nrdc.org/
8. Oceanic N, (NOAA) AA (2016) Climate data and report (2015). http://www.noaa.
gov/
9. Susha I, Zuiderwijk A et al (2015) Benchmarks for evaluating the progress of open
data adoption. Soc Sci Comput Rev 33(5):613–630
10. Zuiderwijk A, Janssen M (2013) Open data policies, their implementation and
impact: a framework for comparison. Gov Inf Q 31(1):17–29
Impact of Management Information Systems
Techniques on Quality Enhancement Cell’s
Report for Higher Education Commission
of Pakistan
Faraz Ullah Khan and Asisf Kamran(B)
Management and Social Science Department,

Institute of Business and Technology, Karachi, Pakistan
asifkamrankhan@gmail.com
Abstract. The way to achieve high quality education is through

the quality affirmation. Higher education Commission has developed
processes for the evaluation of the scholastics program’s quality. The
department of Quality Assurance has set up Quality Enhancement Cell
which works under indicated Criterions. The point is to compose gath-
ered information to produce useful information for QA clients. At present
QEC is setting Self-Assessment Report through their predefined Per-
forma and manual of models to help scholarly foundations. The funda-
mental issue of setting up this report is the utilization of manual strate-
gies, which is the reason for delivering reports with repetitive errors and
clearly, lumbering to oversee. The proposed solution for this issue is, as
one may have seen in this time, fast move towards automation. MIS
techniques will be applied by utilizing software development tools and
database integration and broadly utilized reporting tools. Utilizing above
mentioned tools and techniques will empower QA and clients to play out
their job viably, reduce the repetitive information, save clients time and
upgraded clarity. A more easy to understand Graphical User Interface
helps client to accomplish more exact data for the decision making which
will help them complete their tasks in impressively less time. This pro-
cedure have a major space for system enhancement system management
more graphical reports more auxiliary changes by the assistance of soft-
ware development tools.
Keywords: Self-Assessment Report · Higher education commission ·

Automation · Quality Enhancement Cell · Quality assurance · Manage-
ment information systems
1 Introduction
Quality is a term which has always be confused with other concepts. The word
quality has a Latin word “quails” its meaning is ‘what sort of’. It has a range
368 F.U. Khan and A. Kamran
of different meanings. Quality is something for you knowingly known. But there
are things that are better than others. A high evaluation accorded to an educa-
tive process [11]. Where it has been demonstrated that, through the process,
the students’ educational development has been enhanced not only have they
achieved the particular objectives set for the course but, in doing so, they have
also fulfilled the general educational aims of autonomy of the ability to partici-
pate in reasoned discourse, of critical self-evaluation, and of coming to a proper
awareness of the ultimate contingency of all thought and action [15].
Through above examinations, you will comprehend that esteem is undefined
and setting focused. It stretches out from meaning “standard” to ‘flawlessness’.
Both are significantly settled in their different qualities operationalize in inde-
pendent, institutions, and, the national practice will be visualize later. “Values
can be justified in as a lowest ‘threshold’ by which the accuracy is judged”
[4]. It is quite possible to have higher education institutions to have a common
understanding in formulating to have teaching and research standards with high
quality.
The administration measurement of value is most likely more likened to
the instructive methodologies. We realize that not at all like physical products,
administrations are vaporous to the degree that they can be expended just the
length of the action or the methodology proceeds. Prominent issue of quality
had been identified in the higher education in the Medium Term Development
Framework (MTDF) [12]. Therefore, to enhance the quality of output and effi-
ciency of the higher education learning systems, a mechanism of establishment
of QECs has been developed by the Quality Assurance Committee to improve
the standards of quality of higher education in an organized way with consis-
tency across institutions. QEC prepared SAR to assess a systematic process of
collecting, reviewing and using relevant quantitative/qualitative data and infor-
mation from multiple and diverse sources about educational programs, for the
purpose of improving student learning, and evaluating academic and learning
standards. With a specific end goal to accomplish the subsequent reports the
way toward finishing SAR is complicated [2]. The fundamental issue of setting
up this report is the utilization of manual techniques, which is the reason for
creating reports with repetitive errors and obviously, unwieldy to oversee. The
answer for this issue is, as one may have seen in this time, quick move towards
automation [13]. Applied MIS techniques by utilizing software development tools
and database integration and generally utilized reporting tools. Utilizing above
mentioned tools and techniques will empower QA and clients to perform their job
effectively, lessen the repetitive data, save clients time and enhanced clarity [3].
1.1 Self-Assessment Report Manual by HEC

(Document Prepared By: Dr. Abdul Raouf )
Criterion: Program Mission, Objectives and Outcomes
Each program must have a mission, measurable objectives and expected out-
comes for graduates. Outcomes include competency and tasks graduates are
expected to perform after completing the program. A strategic plan must be
Impact of Management Information Systems Techniques 369
in place to achieve the program objectives. The extent to which these objec-
tives are achieved through continuous assessment and improvements must be
demonstrated [10].
Standard 1: The program must have documented measurable objec-

tives that support Faculty/College and institution mission statements.
• Document institution, college and program mission statements.
• State program objectives. Program educational objectives are intended to be
statements that describe the expected accomplishments of graduates during
the first several years following graduation from the program.
• Describe how each objective is aligned with program, college and institution
mission statements.
• Outline the main elements of the strategic plan to achieve the program mission
and objectives.
• Provide for each objective how it was measured, when it was measured and
improvements identified and made. Table 1 provides a format for program
objectives assessment.
Table 1. Program objectives assessment
Objective How measured When Improvement Improvement

measured measured identified made
1.
2. - do -
3. - do -
4. - do -
5. - do -
Standard 2: The program must have documented outcomes for grad-

uating students. It must be demonstrated that the outcomes support
the program objectives and that graduating students are capable of
performing these outcomes.
• Describe how the program outcomes support the program objectives. In Table
3 show the outcomes that are aligned with each objective..
• Describe the means for assessing the extent to which graduates are perform-
ing the stated program outcomes/learning objectives. This should be accom-
plished by the following (Table 2):
– Conducting a survey of graduating seniors every semester.
– Conduct a survey of alumni every two years.
– Conduct a survey of employers every two years.
Table 2. Outcomes versus objectives
Program objectives Program outcomes

1 2 3 4
1.
2.
3.
– Carefully designed questions asked during senior projects presentations.

These questions should be related to program outcomes.
– Outcomes examinations.
The data obtained from the above sources should be analyzed and presented
in the assessment report.
It is recommended that the above surveys should be conducted, summarized
and added to the self-study assessment report. Departments should utilize the
results of the surveys for improving the program as soon as they are available.
An example follows:
1.2 Example (Program Objectives-Program Outcomes)
An example of program objectives and program outcomes is given below.

(1) Program Objectives (as developed by the department)
• Foundation
• Skills and Tools
• Awareness and Professional Ethics
1 Objective 1
To provide students with a strong foundation in engineering sciences and design
methodologies that emphasizes the application of the fundamental mathemati-
cal, scientific and engineering principles in the areas of engineering.
2 Objective 2
To provide students with skills to enter the workplace well-prepared in the core
competencies listed below:
• Design and modeling experience

• Open-ended problem solving ability
• Experimental and data analysis techniques
• Teamwork experience
• Oral written and multimedia communication skills
• Experience with contemporary computing systems and methodology
3 Objective 3
To provide students with knowledge relevant to engineering practice, including
ethical, professional, social and global awareness, the impact of engineering on
society, the importance of continuing education and lifelong learning in both
technical and non-technical areas.
(2) Program Outcomes (as developed by the department)

Degree of skills and capabilities that will reflect on their performance as engi-
neers:
• Students shall have an ability to apply knowledge of mathematics science and
fundamental engineering to mechanical engineering problems.
• Students shall have an ability to identify, formulate and solve practical engi-
neering problems.
• Students shall have an ability to design components, processes and systems
to meet desired needs.
• Students shall have an ability to conduct engineering experiments to study
different engineering systems, including various modes of operation, perfor-
mance evaluation, properties of materials and manufacturing techniques, as
well as to use laboratory instruments and computers to analyze and interpret
data.
• Students shall have an ability to use modern tools, techniques, and skills nec-
essary for practicing mechanical engineering including computational tools,
statistical techniques, and instrumentation.
• Students shall have an ability to work in a professional engineering environ-
ment, and to understand the associated economical considerations.
• Students shall have an ability to communicate effectively in written, oral, and
graphical forms, including the use of professional quality visual aids.
• Students shall have an ability to work effectively in teams including multidis-
ciplinary teams to solve engineering problems relevant to their field.
• Students shall have an understanding of the professional and ethical respon-
sibilities of engineers.
• Students shall have an understanding of the impact of engineering on society
and environment.
• Students shall have recognition of the need and an ability to engage in life
long learning of engineering.
• The program outcomes are the by products of the program objectives and
are interrelated. An example of interrelation between the program objectives
and the program outcomes is shown in the following table.
Standard 3: The results of program’s assessment and the extent to

which they are used to improve the program must be documented.
• Describe the actions taken based on the results of periodic assessments.
• Describe major future program improvements plans based on recent assess-
ments.
Table 3. Relationship between Program objectives and Program outcomes
Program objectives Program outcomes

2 3 4 5 6 7 8 9 0
1
2a
2b
2c
2d
2e
2f
3
Legend: Denotes substantial contribution to the objective and
denotes moderate contribution to the objective. Denotes no
contribution to the objective.
• List strengths and weaknesses of the program.

• List significant future development plans for the program.
Standard 4: The department must assess its overall performance peri-

odically using quantifiable measures.
• Present students enrolment (undergraduate and graduate) during the last
three years indicating percentages of honor students, student faculty ratio,
average graduating grade point average per semester, average time for com-
pleting the undergraduate program and attrition rate.
• Indicate percentage of employers that are strongly satisfied with the perfor-
mance of the department’s graduates. Use employer’s survey.
• Indicate the median/average student evaluation for all courses and the % of
faculty awarded excellence in teaching award.
• Present performance measures for research activities. These include journal
publications, funded projects, and conference publications per faculty per
year and indicate the % of faculty awarded excellence in research award.
• Present performance measures for community services. This may include
number of short courses per year, workshops and seminars organized.
• Indicate faculty and students satisfaction regarding the administrative ser-
vices offered by the department. Use faculty and students surveys.
2 Problem Statement
The main problem of preparing this report is the use of manual methods, which is
the cause of producing reports with repetitive errors and obviously, cumbersome
to manage. The proposed solution to this problem is, as one may have observed
in this era, rapid move towards automation.
Fig. 1. Self-Assessment procedure
3 Method of Research
This research is aimed that of the process of Self-Assessment Report gets auto-
mated by using technology, for example through user friendly software then
this will help in enhance the process of QA Reporting for HEC. Therefore this
research is a Descriptive Research in which the method of research is Observa-
tional Research. Since the respondents in this research work in a natural envi-
ronment, this method is convenient in assessing multiple users performing their
jobs.
Using detailed investigation methodology adds a lot for the investigation, for
instance:
• Detailed research reveals difficulties as well as excessive ailments to ensure

that remedial procedures may be instituted. This might assistance in present-
ing ideas for predicaments exactly where progress may be required.
• Detailed exploration facilitates the particular prediction of the future by
results about existing disorders, effects and by persons in the direction of
a particular matter.
• Detailed exploration provides a much better and much deeper comprehension
of any trend by an in-depth study from the trend.
3.1 Research Methodology
Fig. 2. Proposed solution automation of QEC’s SAR for HEC
3.2 Design of Research
In view of these variables and hypothesis, the Research Design followed for this
thesis is Qualitative Research Design. Reasons for choosing this design:
• This design helps in identifying & controlling extraneous factors also.

• This design is efficient in using available resources.
3.3 Techniques & Instrument of Research

Research instruments included a single List of questions. The set of questions
was structured. Various factors relevant to the study have been integrated from
the set of questions. There were 20 issues in which concerned numerous qualities
in addition to considerations. The set of questions was made in view from the
study ambitions. The options from the set of questions derive from Contract by
making use of Liker Range.
• Strongly Agree
• Agree
• Disagree
3.4 Target Population

The sample of the SAR user is selected for the research, those who are work-
ing for QEC. Those individual SAR users who are preparing their report by
manual method. On a whole we have taken almost 61 responses from different
institutions.
3.5 Data Analysis

The purpose of the research is to analyze the aspects which are affecting the
adoption of Automation of QEC’s SAR. The researcher selected the question-
naire for collecting the responses. The average graph of respondents is showing
high requirement for the automation of SAR.
3.6 Research Methodology

The representation of the operation sequence uses job-based encoding [13], and
the length of the chromosome equals the total number of operations. The job
number denotes the operation of each job, and the l − th occurrence of a job
number refers to the l−th operation in the sequence of this job. Figure 1 presents
an example considering a rescheduling with a 4 jobs and 3 machines.
3.7 Impact of Research Question

The study will apply the following impacts, when the SAR will automate.
• Time saving
• Process
• Error Reduction
Figure 3(a) shows the significance of Automation of Self Assessment Report
by high requirement of respondents. Figure 3(b) shows the significance of process
Improvement viewed by respondents. Figure 3(c) shows the significance of Time
saving can be viewed by the positive responses. Figure 3(d) shows the magnitude
of error reduction is shown by the positive responses.
Fig. 3. Questionnaire graphical analysis
4 Conceptual Model
The above given diagram is a conceptual model of all entities participating in
the process of making SAR. This model only shows how the whole process works
currently without having being automated. From this diagram one can start to
figure out how the software can be evolved in the system to provide or achieve
the outcomes which are supposed to be achieved.
The process starts when an institution wishes to approve its program by HEC.
QEC personal sends a request to HEC which in return supplies ‘SAR Manual’
in order to prepare SAR Report. QEC user inserts MISSION OBJECTIVES &
OUTCOMES based on the CRITERION-1 in Self Assessment Report Manual.
Finally report is generated as per the format of HEC.
4.1 Waterfall Models

There are different models used for software life cycle the one which is suitable
in SAR is Waterfall Model. As this model distinguishes the whole process into
distinct phases.
“A software requirements specification (SRS) is a description of a soft-
ware system to be developed, laying out functional and non-functional
Fig. 4. Conceptual model
requirements” [8]. This model is used as a frame work or it provides a specified

format to develop any software in a predefined phases. It starts from gathering
requirements from users and then moves on to produce software design. Later
this design will be implemented after unit testing. The last two phases may not
need very much consideration in SAR.
Fig. 5. Waterfall models
4.2 Languages
For the above purpose, C# & VB Languages have been selected to process
SAR Report. C# & VB is usually a multi-paradigm programming language
encompassing strong typing, essential, declarative, sensible, universal, object-
oriented (class-based), and also component-oriented programming disciplines.
C# & VB is one of the programming languages designed for the Common Lan-
guage Infrastructure.
4.3 Programming Methodology
Object-Oriented Programming has the following advantages over conventional

approaches [7]:
Provides a clear modular structure for programs which makes it good for
defining abstract data types where implementation details are hidden and the
unit has a clearly defined interface.
• Makes it easy to maintain and modify existing code as new objects can be
created with small differences to existing ones.
• Provides a good framework for code libraries where supplied software com-
ponents can be easily adapted and modified by the programmer. This is par-
ticularly useful for developing graphical user interfaces [1].
4.4 Platform
Visual Studio is a complete set of development tools for building C# & VB
applications, XML Web Services, desktop applications, and mobile applications.
Visual Basic, Visual C#, and Visual C++ all use the same integrated develop-
ment environment (IDE), which enables tool sharing and eases the creation of
mixed-language solutions [14].
4.5 Database
A Relational database is a computerized database whose association is focused

around the Relational model of information, as proposed by E.f. Codd in 1970.
This model arranges information into one or more tables (or “relations”) of
columns and sections, with a Primary key for each one line [6]. By and large,
every element sort portrayed in a database has its own particular table, the lines
speaking to occurrences of that substance and the segments speaking to the
characteristic qualities depicting each one occurrence (Education Indicators [9].
Since each one column in a table has its own particular interesting key, pushes in
different tables that are identified with it can be connected to it by putting away
the first line’s one of a kind key as a characteristic of the optional line (where it is
known as a “remote key”). Codd [6] demonstrated that information connections
of subjective unpredictability can be spoken to utilizing this straightforward set
of ideas. The different programming frameworks used to keep up social databases
are known as Relational Database Management Systems (RDBMS).
Essentially all Relational database frameworks use SQL (Structured Query
Language) as the dialect for questioning and keeping up the database.
A good choice would be “Microsoft SQL Server is a relational database man-
agement system developed by Microsoft. As a database, it is a software product
whose primary function is to store and retrieve data as requested by other soft-
ware applications, be it those on the same computer or those running on another
computer across a network (including the Internet)” [17].
Fig. 6. Types of reporting tool
4.6 Reporting Tool
The Visual Studio Report Designer provides a user-friendly interface for creating
robust reports that include data from multiple types of data sources [5] (Fig. 6).
‘Visual Studio’ reports let you slice and dice your data and present it in
detail or summary form regardless of how the data is stored or sorted in the
underlying tables. It offers a great deal of power and flexibility to analyze and
present results [16].
4.7 Software Architecture
User authentication and password protection Managers and QEC directors given
the authorization to change all the aspects of SAR report the director is working
with are also given the authorization to access the whole report but with the
exception of few section which are in read only mode.
Fig. 7. Software design

4.8 Software Design

Is the process by which an agent creates a specification of a software artifact,
intended to accomplish goals, using a set of primitive components and subject
to constraints. Software design usually involves problem solving and planning a
software solution [17]. The figure explains main components of ESAR Design.
Application interacts with the four graphical user interfaces (GUI) sequentially.
GUI component establishes two way communication with database components
(that comprises related tables) once all the required data is stored in database,
user initiates the process of generating report (an ultimate components of ESAR)
this final result will also be saved (Fig. 7).
5 Result
• Figure 2 illustrates the research framework and demonstrate the function of
moderate variable.
• Figure 4 shows the process of Self-Assessment Report through conceptual
model.
• Figure 5 demonstrates the software design of Self-Assessment Report.
• TIME: the process of ESAR will significantly save the user’s time.
• MIS Techniques: the above advantages of results may well be achieved MIS
techniques.
• REDUCTION OF REPEATION ERRORS: using RDBMS.
• FLEXIBILITY IN FUTURE MODIFICATION: can be achieved using OOP
techniques.
6 Conclusions
Quality assurance is a portal to “High quality education”. With a specific end
goal to evaluate the programs quality Higher Education Commission” has cre-
ated predefine Criterions forms. The department of Quality Enhancement Cell
which works under those predefined Criterion. This Criterion is met utilizing
frames with applicable information. Thusly, this will bring about an enhanced
interaction with QA clients utilizing E-SAR. One of the other major advantages
of it will be as better organized report. In this manner, the real issue for QA
clients of following QEC procedures can extensively be all around encouraged
utilizing the proposed solution. Infect, repetitive errors are better managed with
it. A well designed application is most suitable to accomplish the proposed results
for future work. In future, alternative design technique of HCI can be utilized
to furnish client with more upgraded collaboration in SAR making.
In order to discuss the process of QEC’s SAR, we clearly observe that the
whole process is currently based on manual work. There is a dire need to auto-
mate SAR process which reflects from the survey data analysis, collected from
the QEC users. The remaining Criterion of SAR can easily be considered by
applying MIS technology and software development tools in order to automate.
References
1. Afaf (2016) Advantages of OOP. https://www.researchgate.net/post/What is
adventages of Object Oriented Programming
2. Alvi NA, Alam A (2004) Pakistan institute of quality control
3. Arshad M (2003) Attitude of teachers of higher education towards their profession.
Ph.D. thesis, Allama Iqbal Open University, Islamabad
4. Ashcroft K (2005) Criteria standards and quality. https://books.google.com.pk/
books?isbn=113571990X
5. Code EF (1990) Keys and referential integrity. https://www.seas.upenn.
edu:codeblab.com/wp-content/uploads/2009/12/rmdb-codd.pdf
6. Code EF (1990) The relational model for database management: verison 2[M].
Addison-Wesley Longman Publishing co. Inc
7. Coulter N (2015) Advantages and disadvantages of OOP
8. ISO/IEEE 830-1998 (1998) Recommended practice for software requirements spec-
ifications. doi:10.1109/IEEESTD.1998.88286
9. Indicators E (2015) Expenditure on education as % of total government expendi-
ture (%). http://data.worldbank.org/indicator/SE.XPD.TOTL.GB.ZS
10. Institute BS (2015) Learning and growth perspective. http://balancedscorecard.
org/Learning-and-Growth-Perspective
11. James Hutton DP (2010) Corporate governance-good governance in the university
context. http://www.minterellison.com/Pub/NL/201003 HEFd/
12. Khan MA, Usman M (2015) Education quality and learning outcomes in higher
education institutions in pakistan. Springer, Singapore, pp 449–463
13. Michal Pietrzak JP (2015) The application of the balanced scorecard (bsc) in
the higher education setting of a polish university. Online J Appl Knowl Manag
3(1):151–164
14. Microsoft (2008) Introducing visual studio. https://msdn.microsoft.com/en-us/
library/fx6bk1f4(v=vs.90).aspx
15. Mishra S (2007) Concept of quality. http://oasis.col.org/bitstream/handle/11599/
101/QAHE Intro.pdf?sequence=1
16. Ralph P, Wand Y (2017) A proposal for a formal definition of the design concept.
In: Lyytinen K, Loucopoulos P et al (eds) Design Requirements Workshop. LNBIP,
vol 14. Springer, Heidelberg, pp 103–136
17. WikiPedia (2008) Microsoft sql server. https://en.wikipedia.org/wiki/Microsoft
SQL Server
A Priority-Based Genetic Representations
for Bicriteria Network Design Optimizations
Lin Lin1,2,3(B) , Jingsong Zhao1 , Sun Lu1 , and Mitsuo Gen2

1
Dalian University of Technology, Dalian, People’s Republic of China
lin@dlut.edu.cn
2
3
Dalian, People’s Republic of China
Abstract. Network design is one of the most important and most fre-
quently encountered classes of optimization problems. It is a combinatory
field in combinatorial optimization and graph theory. A lot of optimiza-
tion problems in network design arose directly from everyday practice in
engineering and management. Furthermore, network design problems are
also important for complexity theory, an area in the common intersec-
tion of mathematics and theoretical computer science which deals with
the analysis of algorithms. Recent advances in evolutionary algorithms
(EAs) are interested to solve such practical network problems. However,
various network optimization problems typically cannot be solved ana-
lytically. Usually we must design the different algorithm for the different
type of network optimization problem depending on the characteristics of
the problem. In this paper, we investigate the recent related researches,
design and validate effective priority-based genetic representations for
the typical network models, such as shortest path models (node selec-
tion and sequencing), spanning tree models (arc selection) and maximum
flow models (arc selection and flow assignment) etc., that these models
covering the most features of network optimization problems. Thereby
validate that EA approaches can be effectively and widely used in net-
work design optimization.
Keywords: Evolutionary Algorithm (EA) · Shortest path model ·

Spanning tree model · Maximum flow model · Bicriteria network design
1 Introduction
Many real-world problems from operations research (OR)/management science
(MS) are very complex in nature and quite hard to solve by conventional opti-
mization techniques. Since 1960s, there has been being an increasing interest
in imitating living beings to solve such kinds of hard optimization problems.
Simulating natural evolutionary process of human beings results in stochastic
optimization techniques called evolutionary algorithms (EAs) that can often out-
perform conventional optimization methods when applied to difficult real-world
A Priority-Based Genetic Representations for Bicriteria Network Design 383
problems. EAs mostly involve metaheuristic optimization algorithms such as:

genetic algorithm (GA) [18,21], evolutionary programming (EP) [12], evolution
strategies (ES) [36,39], genetic programming (GP) [27,28], learning classifier sys-
tems (LCS) [22], swarm intelligence (comprising ant colony optimization (ACO)
[9] and particle swarm optimization (PSO) [25,26]. Among them, genetic algo-
rithms are perhaps the most widely known type of evolutionary algorithms today
and load more stuff again.
Network design is one of the most important and most frequently encoun-
tered classes of optimization problems [2]. It is a combinatory field in graph
theory and combinatorial optimization. A lot of optimization problems in net-
work design arose directly from everyday practice in engineering and manage-
ment: determining shortest or most reliable paths in traffic or communication
networks, maximal or compatible flows, or shortest tours; planning connections
in traffic networks; coordinating projects; and solving supply and demand prob-
lems. The most of network design problems where even one cost measure must
be minimized are often NP-hard [14]. However, in practical applications, it is
often the case that the network to be built is required to multiobjective. In
this paper, we introduce three kinds of major bicriteria network design mod-
els. These models are the core multi-criteria network design models and exten-
sively used in practical applications. (1) Bicriteria shortest path (bSP) model
is one of the basic multi-criteria network design problems. It is desired to find
a diameter-constrained path between two specified nodes with minimizing two
cost functions. Hansen presented the first bSP model at 1979 [19]. Recently, Hao
and Gen [20] examined an effective multiobjective EDA for bi-criteria stochas-
tic job-shop scheduling problem; Shokrollahpour and Zandieh [47] presented A
novel imperialist competitive algorithm for bi-criteria scheduling of the assembly
flowshop problem. (2) Bicriteria spanning tree (bST) model play a central role
within the field of multi-criteria network modes. It is desired to find a subset
of arcs which is a tree and connects all the nodes together with minimizing two
cost functions. Craveirinha et al. [4] presented a bi-criteria minimum spanning
tree routing model for MPLS/overlay networks, and Steiner [41] Computing all
efficient solutions of the biobjective minimum spanning tree problem.
2 Multiobjective Genetic Algorithms

Optimization deals with the problems of seeking solutions over a set of possible
choices to optimize certain criteria. If there is only one criterion to be taken into
consideration, it becomes to single objective optimization problems, which have
been extensively studied for the past 50 years. If there are more than one criterion
which must be treated simultaneously, we have multiple objective optimization
problems [1,6,42]. Multiple objective problems arise in the design, modeling, and
planning of many complex real systems in the areas of industrial production,
urban transportation, capital budgeting, forest management, reservoir manage-
ment, layout and landscaping of new cities, energy distribution, etc. It is easy
to find that almost every important real world decision problem involves multi-
ple and conflicting objectives which need to be tackled while respecting various
384 L. Lin et al.
constraints, leading to overwhelming problem complexity. The multiple objec-

tive optimization problems have been receiving growing interest from researchers
with various background since early 1960 [23]. There are a number of scholars
who have made significant contributions to the problem. Among them, Pareto
is perhaps one of the most recognized pioneers in the field [11]. Recently, GAs
have been received considerable attention as a novel approach to multiobjective
optimization problems, resulting in a fresh body of research and applications
known as evolutionary multiobjective optimization (EMO).
2.1 Fitness Assignment Mechanism

GAs are essentially a kind of meta-strategy methods. When applying the GAs to
solve a given problem, it is necessary to refine upon each of the major components
of GAs, such as encoding methods, recombination operators, fitness assignment,
selection operators, constraints handling, and so on, in order to obtain a best
solution to the given problem. Because the multiobjective optimization prob-
lems are the natural extensions of constrained and combinatorial optimization
problems, so many useful methods based on GAs developed during the past two
decades. One of special issues in the multiobjective optimization problems is
fitness assignment mechanism. Since the 1980s, several fitness assignment mech-
anisms have been proposed and applied in multiobjective optimization problems
[7,8,13,15,16,24,38,40,49]. Although most fitness assignment mechanisms are
just different approach and suitable to different cases of multiobjective opti-
mization problems, in order to understanding the development of multiobjective
GAs, we classify algorithms according to proposed years of different approaches:
2.2 Performance Measures

Let Sj be a solution set (j = 1, 2, · · · , J). In order to evaluate the efficiency of
the different fitness assignment approaches, we have to explicitly define measures
evaluating closeness of Sj from a known set of the Pareto-optimal set S*. For
example, the following common 3 measures are considered that already used in
different moGA studies [24]. They provide a good estimate of convergence if a
reference set for S* (i.e., the Pareto optimal solution set or a near Pareto optimal
solution set) is chosen.
(1) Number of Obtained Solutions |Sj |
Evaluate each solution set depend on the number of obtained solutions.
(2) Ratio of Nondominated Solutions RNDS (Sj )
This measure simply counts the number of solutions which are members of
the Pareto-optimal set S ∗ . The RNDS (Sj ) measure can be written as follows:
|Sj − {x ∈ Sj |∃r ∈ S ∗ : r ≺ x }|
RNDS (Sj ) = ,
|Sj |
where r ≺ x means that the solution x is dominated by the solution r. The

RNDS (Sj ) =1 means all solutions are members of the Pareto-optimal set S ∗ ,
and RNDS (Sj ) = 0 means no solution is a member of the S ∗ . It is an important

measure that although the number of obtained solutions |Sj | is large, if that the
ratio of nondominated solutions RNDS (Sj ) is 0, it may be the worst result. The
difficulty with the above measures is that although a member of Sj is Pareto-
optimal, if that solution does not exist in S ∗ , it may be not counted in RNDS (Sj )
as a non-Pareto-optimal solution. Thus, it is essential that a large set for S ∗ is
necessary in the above equations.
(3) Average Distance D1R(Sj )
Instead of finding whether a solution of Sj belongs to the set S ∗ or not, this
measure finds an average distance of the solutions of Sj from S ∗ , as follows:
1
D1R (Sj ) = min{drx |x ∈ Sj },
|S ∗ |
r∈S∗
where drx is the distance between a current solution x and a reference solution r
in the 2-dimensional normalized objective space. fi means the objective function
for each objective i = 1, 2, · · · , q.

q

drx =
2
(fi (r) − fi (x)) .
i=1
The smaller the value of D1R (Sj ) is, the better the solution set Sj is. This
measure explicitly computes a measure of the closeness of a solution set Sj from
the set S ∗ .
(4) Reference Set S ∗
For making a large number of solutions in the reference set S*, the first step
calculates the solution sets with special GA parameter settings and much long
computation time by each approach which used in comparison experiments, the
second step combine these solution sets to calculate the reference set S ∗ . In
the future, a combination of small but reasonable GA parameter settings for
comparison experiments will be conducted. Thus, ensure the effectiveness of the
reference set S ∗ .
16, 40 19, 20
2 5 8
14, 23 16, 32 18, 39
18, 30
13, 50 15, 32 17, 25
s t
1 19, 28 16, 27 15, 32 19, 28 1
1 3 6 9 11
17, 28 18, 38 14, 30
15, 38
17, 32 15, 37 17, 37
19, 22 13, 29
4 7 10
Fig. 1. A simple network with 11-node/ 22-edge

386 L. Lin et al.
3 Bicriteria Shortest Path (bSP) Model

Routing problem is one of the important research issues in communication net-
work fields. As B-ISDN standards and ATM networks emerge, modern networks
are expected to provide a wide range of services and guarantee various end-to-end
Quality of Service (QoS). Traditionally, the routing problem has been a single
objective problem having as goal the minimization of either the total distance
or travel time. However, in practical applications of communication network,
create an effective route consider not only cost or travel time, but also require-
ments of social and economic environment. In other words, it is necessary to take
into account that many real-world problems are multiobjective in nature. The
objective functions related to cost, time, accessibility, environmental impact,
reliability and risk are appropriated for selecting the most satisfactory (“best
compromise”) route in many communication network optimization problems [3].
3.1 Mathematical Formulation

Let G = (N, A) be a directed network, consisting of a finite set of nodes N =
1, 2, · · · , n and a set of directed arcs A = {(i, j), (k, l), · · · , (s, t)} joining m pairs
of nodes in N. Arc (i, j) is said to be incident with nodes i and j, and is directed
from node i to node j. Suppose that each arc (i, j) has assigned to it nonnegative
numbers cij , the cost of (i, j) and dij , the transmission delay of (i, j). Let xij
is a 0-1 Decision variables the link on an arc (i, j) A. The bSP problem is
formulated as follows, in which the objectives are to minimizing total cost z1
and minimizing transmission delay z2 form source node 1 to sink node n.
n
n
min z1 = cij xij (1)
i=1 j=1
n n
min z2 = dij xij (2)
i=1 j=1
⎧
n
n
⎨ = 1, i = 1
s.t. xij − xki = 0, i = 2, 3, · · · , n − 1 (3)
⎩
j=1 k=1 = −1, i = n
xij = 0 or 1, (i, j = 1, 2, · · · , n), (4)
where constraint (3), a conservation law is observed at each of the nodes other
than s or t. That is, what goes out of node i, i=1 xij must be equal to what
comes in k=1 xki .
3.2 GA Approaches for bSP Model

How to encode a solution of the network design problem into a chromosome is a
key issue for GAs. In Holland’s work, encoding is carried out using binary strings.
For many GA applications, especially for the problems from network design
problems, the simple approach of GA was difficult to apply directly. There two
special difficulties by using GAs for creating a path: (1) different path contains
variable number of nodes; (2) a random sequence of nodes usually does not
correspond to a path.
Zhang et al. [48] extended this variable-length chromosome for solving the
SPR problem. But crossover may generate infeasible chromosomes that generat-
ing loops in the routing paths. It must be checked that none of the chromosomes
is infeasible at each generation, and is not suitable for large networks or unac-
ceptable high computational complexity for real-time communications involving
rapidly changing network topologies. An example of generated variable-length
chromosome and its decoded path are shown in Fig. 2(a) and (c), respectively
for the directed network shown in Fig. 1.
We proposed a priority-based encoding method. As it is known, a gene in
a chromosome is characterized by two factors: locus, i.e., the position of gene
located within the structure of chromosome, and allele, i.e., the value the gene
takes. In this encoding method, the position of a gene is used to represent node
ID and its value is used to represent the priority of the node for constructing a
path among candidates. A path can be uniquely determined from this encoding.
An example of generated priority-based chromosome is shown in Fig. 2(b). At the
beginning, we try to find a node for the position next to source node 1. Nodes 2,
3 and 4 are eligible for the position, which can be easily fixed according to adja-
cent relation among nodes. The priorities of them are 1, 10 and 3, respectively.
The node 3 has the highest priority and is put into the path. The possible nodes
next to node 3 are nodes 4, 6 and 7. Because node 6 has the largest priority
value, it is put into the path. Then we form the set of nodes available for next
position and select the one with the highest priority among them. Repeat these
steps until we obtain a complete path, (1-3-6-5-8-11). Considering the charac-
teristic of priority-based chromosome, they proposed a new crossover operator,
called weight mapping crossover (WMX) and adopted insertion mutation and
immigration operators.
locus : 1 2 3 4 5 6 locus : 1 2 3 4 5 6 7 8 9 10 11
node ID : 1 3 6 5 8 11 node ID : 11 1 10 3 8 9 5 7 4 2 6
(a) variable-length chromosome (b) priority-based chromosome
path : 1 3 6 5 8 11
(c) decoded path
Fig. 2. An example of chromosome and its decoded path

388 L. Lin et al.
3.3 Experiments and Discussion
In this subsection, the first experiment examines the different chromosome

representations. We compare priority-based encoding with Ahn and Ramakr-
ishna’s variable-length encoding. Then to validate the multiobjective evolution-
ary approaches for solving bSPR problems, we compare our i-awGA with 3
different fitness assignment approaches (spEA, nsGA II and rwGA). All test
network topologies were constructed by Beasley and Christofides [31]. Costs and
delay are represented as random variables depend on the distribution functions.
(1) Different Encoding Methods
In order to evaluate the results of each test, we use the single objective by mini-
mizing total cost, and combine average of the best solutions (ABS). In addition,
we demonstrate the difference among the quality of solutions obtained by vari-
ous GA parameter settings and our auto-tuning strategy. There are 3 kinds of
different GA parameter settings:
Para 1: popSize = 10, pc = 0.3, pM = 0.7, pI = 0.30
Para 2: popSize = 10, pc = 0.5, pM = 0.5, pI = 0.30
Para 3: popSize = 10, pc = 0.7, pM = 0.3, pI = 0.30.
In addition, two different stopping criteria are employed. One of them is the
number of maximum generations, maxGen = 1000. Another stopping criteria
is T = 200. That is, if the best solution is not improved until successive 200
generations, the algorithm will be stopped.
Table 1 shows the ABS of 50 runs by different GA parameter settings with
different genetic representations respectively. As depicted in Table 1, most results
of ABS of 50 runs by priGA with auto-tuning operator proposed are better than
each of the other combinations, except to the test 4, test 6 and test 8.
(2) Different Fitness Assignment
In the second experimental study, we demonstrate the performance comparisons
of multiobjective GAs for solving bSP problems by different fitness assignment
approaches, there are spEA, nsGA II, rwGA and i-awGA. In each GA approach,
priority-based encoding was used, and arithmetical crossover, swap mutation,
immigration and auto-tuning operators were used as genetic operators.
As depicted in Table 2, most results of ABS of 50 runs by i-awGA are better
than each of the other fitness assignment approach. In addition, we do not say
the efficiency of the approach, only depend on the performance measure |Sj |
or RNDS (Sj ). We can have worst results when compared to another run with
a low RNDS (Sj ). Therefore, we show the proposed i-awGA outperform another
approach with the efficiency both of the performance measure |Sj | or RNDS (Sj ).
In Table 3, the values of |Sj | are given as rational numbers, though the value of
|Sj | was defined as the integer number. Because we give an average of |Sj | with
50 runs for comparing the different approaches. Furthermore, the values of |Sj |
increases (or decreases) depended on the characteristic of different testing data.
Table 1. The ABS of 50 runs by different ga parameter settings with different genetic
representations
ahnGA priGA
ID Optimal Para1 Para2 Para3 Para1 Para2 Para3 Auto-timing
1 47.93 47.93 47.93 47.93 47.93 47.93 47.93 47.93
2 210.77 232.38 234.36 244.64 224.82 234.91 228.72 224.09
3 1.75 2.69 2.71 2.83 2.68 2.73 2.79 2.64
4 17.53 37.6 39.43 47.26 36.1 35.3 34.08 34.6
5 54.93 60.77 62.26 65.35 57.26 57.42 58.5 56.87
6 234.45 276.72 288.71 295.77 269.23 268.52 273.16 270.66
7 1.83 2.4 2.66 3.31 2.01 2.27 2.32 1.98
8 22.29 47.29 49.58 57.04 41.48 45.89 44.17 41.9
9 70.97 - - - 72.29 75.74 77.27 70.97
10 218.78 - - - 276.56 276.15 284.85 272.1
11 3.82 - - - 5.85 6.91 6.41 5.78
12 20.63 - - - 60.14 57.52 61.53 52.18
“-” means out of memory error
4 Bicriteria Spanning Tree (bST) Model

Modeling and design of large communication and computer networks has always
been an important area to both researchers and practitioners. The interest in
developing efficient design models and optimization methods has been stimu-
lated by high deployment and maintenance costs of networks, which make good
network design potentially capable of securing considerable savings [33].
Table 2. The ABS of 50 runs by different fitness assignments
|Sj | RNDS (Sj ) DlR (Sj )

ID spEA nsGA rwGA i-awGA spEA nsGA rwGA i-awGA spEA nsGA rwGA i-awGA
1 1.64 1.7 1.64 1.84 1 1 1 1 0 0 0 0
2 5 5.08 4.98 5.64 0.18 0.16 0.22 0.38 0.18 0.23 0.17 0.1
3 3.3 3.04 3.22 3.48 0.91 0.93 0.92 0.91 0 0 0 0
4 7.36 7.4 7.12 7.46 0.04 0.02 0.04 0.04 0.06 0.06 0.05 0.05
5 3.26 3.22 3.12 3.46 1 1 1 1 0 0 0 0
6 1.74 2.4 2.2 1.54 0.28 0.14 0.18 0.3 0.17 0.24 0.22 0.15
7 4.16 3.96 3.66 3.7 0.52 0.59 0.66 0.68 0.4 0.42 0.4 0.05
8 5.9 4.8 5.3 5.16 0.05 0.13 0.07 0.1 1.1 0.89 0.96 0.86
9 1.16 1.24 1.28 1.36 0.99 0.96 0.91 0.99 0 0.01 0.01 0
10 2.6 2.42 2.62 2.3 0.11 0.18 0.16 0.33 1.17 0.76 0.99 0.59
11 2.86 2.9 2.7 3.22 0.31 0.3 0.3 0.43 0.01 0.04 0.01 0
12 5.82 6.02 6.14 6.2 0.03 0.03 0.04 0.05 0.19 0.19 0.2 0.19
390 L. Lin et al.
The Minimum Spanning Tree (MST) problem is one of the best-known net-
work optimization problems which attempt to find a minimum cost tree network
that connects all the nodes in the communication network. The links or edges
have associated costs that could be based on their distance, capacity, quality of
line, etc.
In the real world, the MST is often required to satisfy some additional con-
strain for designing communication networks such as the capacity constraints on
any edge or node, degree constraints on nodes, and type of services available on
the edge or node. This additional constraint often makes the problem NP-hard.
In addition, there are usually such cases that one has to consider simultane-
ously multicriteria in determining a MST, because there are multiple attributes
defined on each edge, has become subject to considerable attention. Almost every
important real-world decision making problem involves multiple and conflicting
objectives [17].
In this paper, we are considering a bicriteria spanning tree (bST) model. The
bST is to find a set of links with the two conflicting objectives of minimizing
communication cost and minimizing the transfer delay and the constraint of
network capacity is met. This problem can be formulated as the multiobjective
capacitated minimum spanning tree (mcMST) problem, and is a NP-hard.
4.1 Mathematical Formulation
The communication network is modeled using an edge-weighted undirected

graph G = (V, E, C, D, W ) with n = |V | nodes and m = |E| edges, cij C = [cij ],
dij D = [dij ] and wij W = [wij ] represent the cost, delay and weight of each
edge (i, j)E, where the variable is restricted to be a nonnegative real number,
respectively. And ui is given weight capacity of each node i. Figure 3 presents a
simple network with 12 nodes and 40 arcs.
3 7 3 7
1 4 8 11 1 4 8 11
2 5 9 12 2 5 9 12
6 10 6 10
(a) a simple network (b) a spanning tree
Fig. 3. A simple network with 12 nodes and 40 edges
The decision variables in the bicriteria communication spanning tree problem

(bSTP) are the minimum cost z1 with minimum delay z2 . Mathematically, the
problem is reformulated as a constrained Minimal Spanning Tree (MST) problem
is as follows:
n
n
min z1 (x) = cij xij (5)
i=1 j=1
n n
min z2 (x) = dij xij (6)
i=1 j=1
n
n
s. t. xij = n − 1 (7)
i=1 j=1
n n
xij ≤ |S| − 1 for any set S of nodes (8)
i=1 j=1
n
wij xij ≤ ui , ∀i (9)
j=1
xij ∈ {0, 1}, ∀i, j. (10)
In this formulation, the 0-1 variable xij indicates whether we select edge (i, j)
as part of the chosen spanning tree (note that the second set of constraints with
|S| = 2 implies that each xij ≤ 1). The constraint (7) is a cardinality constraint
implying that we choose exactly n − 1 edges, and the packing constraint (8)
implies that the set of chosen edges contain no cycles (if the chosen solution
contained a cycle, and S were the set of nodes on a chosen cycle, the solution
would violate this constraint). The constraint (9) guarantees that the total link
weight of each node i does not exceed the upper limit Wi .
4.2 GA Approaches for bST Model

Recently, GA and other Evolutionary Algorithms (EAs) have been used exten-
sively in a wide variety of communication network design problems. For exam-
ple, Mathur [32] Voβ, Stefan et al. [46] gave the EAs for the capacitated MST
problem occurring in telecommunication applications, Ruiz and Duhamel [10,37]
investigated the different encoding methods and gave the performance analyzes
for the degree-constraint MST problem in communication network, respectively.
Liang [29], Torkestani et al. [43], Rezaei et al. [45] investigated the multicrite-
ria communication design problems with GA approaches and other EAs. GAs
were also applied to solve other communication network design problem such
as two graph problem, one-max tree problem and communication spanning tree
problem [44].
We need to consider these critical issues carefully when designing an appro-
priate encoding method so as to build an effective GA. How to encode a spanning
tree T in a graph G is critical for developing a GA to network design problems,
it is not easy to find out a nature representation. We summarized the several
kinds of classification of encoding methods as follows:
Characteristic Vectors-based Encoding: Davis et al. [5], Piggott and
Suraweera [34] have used binary-based encoding method to represent spanning
trees in GAs.
392 L. Lin et al.
Edge-based Encoding: Edge-based encoding is an intuitive representation of

a tree. General, an edge-based encoding requires space proportional to n and
the time complexities is O(m). Edge-based encoding and binary-based encoding
have very similar performance in theory. However, there are 2n(n − 1)/2 possible
values for a tree, and only a tiny fraction of these chromosomes represents feasible
solutions, and weaken the encoding heritability.
Prefer Number-based Encoding: Many researchers have encoded spanning
trees as Prüfer numbers in GAs, called Prüfer number-based encoding for a
variety of problems. These include the degree-constrained MST problems, leaf-
constrained MST problem, and multicriteria MST problems, etc.
Predecessor-based Encoding: Abuali et al. [35] proposed a predecessor-based
encoding that applies to spanning trees. It is a more compact representation of
spanning trees by the predecessor or determinant encoding, in which an arbitrary
node in G is designated the root, and a chromosome lists each other node’s
predecessor in the path from the node to the root in the represented spanning
tree: if pred(i) is j, then node j is adjacent to node i and nearer the root.
Thus, a chromosome is string of length n − 1 over 1, 2, · · · , n, and when such
a chromosome decodes a spanning tree, its edges can be made explicit in time
that is O(nlogn).
PrimPred-based Encoding: We improved the predecessor-based encoding
that adopted Prim’s algorithm in chromosome generating procedure. Prim’s
algorithm implements the greedy-choice strategy for minimum spanning tree.
Starting with an empty tree (one with no edges), the algorithm repeatedly adds
the lowest-weight edge (u, v) in G such that either u or v, but not both, is
already connected to the tree. Considering the characteristic of predecessor-
based encoding, they proposed a new crossover and mutation operators. These
operators offer locality, heritability, and computational efficiency. An example of
generated PrimPred-based chromosome and its decoded spanning tree are shown
in Fig. 4 for the undirected network shown in Fig. 3.
node id k: 2 3 4 5 6 7 8 9 10 11 12
chromosome v(k): 5 1 3 8 9 8 4 5 12 8 11
T = {(1, 3), (2, 5), (3, 4), (4, 8), (5, 8), (5, 9), (6, 9), (7, 8), (8, 11), (10, 12), (11, 12)}
Fig. 4. An example of the Prim Pred-based chromosome
4.3 Experiments and Discussion

In this section, first, the proposed PrimPred-based GA is compared with Prefer
Number-based Encoding, Edge-based Encoding for solving several large scale
minimum spanning tree problems. Afterwards, compare the different fitness
assignment approaches by multiobjective GA for solving bSTP.
(1) Different Genetic Representations

For examining the effectiveness of different encoding methods, we applied Prefer
Number-based encoding method and Edge-based encoding method on 6 test
problems [30]. We are combining the Prefer Number-based encoding with one-
cut point crossover and swap mutation, and combining the Edge-based encoding
using two kinds of mutation operators which included in [48], and for initializing
the chromosomes based on the edge set, we combining the Raidl and Julstrom’s
PrimRST (Prim Random Spanning Tree). Each algorithm was run 20 times
using different initial seeds for each test problems. The GA parameter is setting
as follows:
Population size: popSize = 10;
Crossover probability: pC = 0.30, 0.50 or 0.70;
Mutation probability: pM = 0.30, 0.50 or 0.70;
Maximum generation: maxGen = 1000;
The experimental study was realized to investigate the effectiveness of the
different encoding method; the interaction of the encoding with the crossover
operators and mutation operators; and the parameter settings affect its perfor-
mance. Table 3 gives computational results for four different encoding methods
on six test problems by tree kinds of parameter settings, where, edge-based 1
and edge-based 2 were considered with different mutation operators of Raidl and
Julstrom’s approach [35].
When we compare with columns of the best cost of four encoding methods, it
is possible to see that whereas the Prefer number-based approach is the fastest
than others, it is difficult to consist mostly of substructures of their parents’
phenotypes (poor heritability), and the results is very far from the best one. The
two kinds of mutation is used in edge-based encoding, the second one (depends
on the cost) gives better performance than the first one. For considering the
computational cost (CPU time), because of the LowestCost mutation in proposed
approach, spend the most CPU times to find the edge with lowest cost, it is
always longer than other algorithms. However, our algorithm developed in this
study gives best cost than other algorithms.
(2) Different Fitness Assignment Approaches
The second experimental study, we show performance comparisons of multiob-
jective GAs for solving bSTP by different fitness assignment approaches, there
are spEA, nsGAII, rwGA and i-awGA. The data in test problem was generated
randomly. In each GA approach, PrimPred-based encoding was used, and Prim-
based crossover and LowestCost mutation were used as genetic operators. GA
parameter settings were taken as follows:
Population size: popSize = 20;
Crossover probability: pC = 0.70;
Mutation probability: pM = 0.50;
Stopping criteria: evaluation of 5000 solutions.
We compare i-awGA with spEA, nsGAII and rwGA trough computational
experiments on the 40-node/1560-arc test problem under the same stopping
condition (i.e., evaluation of 5000 solutions). Each algorithm was applied to
394 L. Lin et al.
Table 3. Performance comparisons with different GA approaches.
Test Optimal Prüfernum-based Edge-based1 Edge-based2 PremPred-based

problem solutions
n m pC pM avg CPU-t avg CPU-t avg CPU-t avg CPU-t
1 470 40 780 0.3 0.3 1622.2 72.2 1491.8 1075.2 495.6 1081.4 470 1100.2
0.5 0.5 1624.4 87.6 1355.8 2184.4 505.8 2175 470 2256.4
0.7 0.7 1652.6 134.8 1255.2 3287.4 497.6 3281.4 470 3316
2 450 40 780 0.3 0.3 1536.6 74.8 1458.2 1118.6 471.6 1093.8 450 1160.2
0.5 0.5 1549.2 78.2 1311.4 2190.8 480.2 2175 450 2200.2
0.7 0.7 1564.4 122 1184.4 3287.6 466.4 3262.4 450 3275
3 820 80 3160 0.3 0.3 3880.4 150 3760.2 5037.8 923.2 5059.6 820 5072
0.5 0.5 3830 184.4 3692 10381.2 871 10494.2 820 10440.6
0.7 0.7 3858.2 231.2 3483.8 16034.8 899.2 15871.8 820 15984.6
4 802 80 3160 0.3 0.3 3900.6 131.4 3853 5125 894.6 4934.2 802 5071.8
0.5 0.5 3849.6 206.2 3515.2 10325.2 863 10268.8 802 10365.6
0.7 0.7 3818.4 222 3287.2 16003 868 15965.4 802 15947.2
5 712 120 7140 0.3 0.3 5819.4 187.4 5536.6 15372 871.8 15306.4 712 15790.4
0.5 0.5 5717.2 293.8 5141 31324.8 805.4 30781.4 712 31503.2
0.7 0.7 5801.4 316 5035.2 47519 804.2 47047.2 712 47865.8
6 793 160 12720 0.3 0.3 7434.8 284.4 7050.4 41993.6 1353.6 42418.6 809.6 42628.2
0.5 0.5 7361 421.8 7111.6 87118.8 1061.6 86987.4 793 86828.4
0.7 0.7 7517 403.2 6735 163025 955.4 161862.4 793 154731.2
Table foot note (with superscript)
Table 4. Performance evaluation of fitness assignment approaches for the 40-

node/1560-arc test problem.
# of eval.solut. |Sj | RNDS (Sj ) DlR (Sj)

spEA nsGA II rwGA i-awGA spEA nsGA II rwGA i-awGA spEA nsGA II rwGA i-awGA
50 31.45 30.4 32.6 36.2 0.34 0.31 0.36 0.39 178.85 200.47 182.03 162.57
500 42.4 45.6 43.2 47.6 0.42 0.45 0.4 0.52 162.97 151.62 160.88 157.93
2000 46.6 52.2 45.3 55.5 0.54 0.61 0.58 0.66 118.49 114.6 139.4 92.41
5000 51.2 54.4 50.3 60.7 0.64 0.7 0.62 0.73 82.7 87.65 117.48 77.98
each test problem 10 times and gives the average results of the 3 performance
measures (i.e., the number of obtained solutions |Sj |, the ratio of nondominated
solutions RNDS (Sj ), and the average distance D1R measure). In Table 4, better
results of all performance measures were obtained from the i-awGA than other
fitness assignment approach.
5 Conclusion
In this paper, we investigated with a broad spectrum of bicriteria network opti-

mization models, analyze the recent related researches, design and validate effec-
tive EA approaches for the typical network models: bicriteria shortest path (bSP)
model, bicriteria spanning tree (bST) model, bicriteria network flow (bNF)
model. Focus on the broad applications of proposed evolutionary algorithms
(EAs) to network design with large-scale optimization.
For bSP model, we introduced an efficient genetic representation using the

priority-based encoding method; based on the characteristic of this represen-
tation, we propose a new crossover operator, called weight mapping crossover
(WMX), adopt insertion mutation and immigration operator. We gave effective
analysis of different evolutionary approaches for bSP model dedicated to calcu-
late nondominated paths for the minimum total cost and the minimum transmis-
sion delay. For bST model, we investigated different GA approaches for solving
minimum spanning tree (MST) problems, and introduced a new genetic repre-
sentation using the PrimPred-based encoding method. And we also introduced a
local search technique into GA loops, called a LowestCost mutation operator. For
bNF model, we introduced a new multiobjective genetic algorithm (moGA) to
solve the problem with two conflicting objectives to minimize the total cost and
maximize the total flow simultaneously. We combine the priority-based encod-
ing method with a special decoding to overcome the special difficulties of bNF
model.

References
1. Abbasi S, Taghipour M (2015) An ant colony algorithm for solving bi-criteria
network flow problems in dynamic networks. IT Eng 3(5):34–48
2. Ahuj RK, Magnanti TL, Orlin JB (1993) Network flows. Prentice Hall, New Jersey
3. Climaco JCN, Craveirinha JMF, Pascoal MMB (2003) A bicriterion approach for
routing problems in multimedia networks. Networks 41(4):206–220
4. Craveirinha J, Maco J et al (2013) A bi-criteria minimum spanning tree routing
model for mpls/overlay networks. Telecommun Syst 52(1):1–13
5. Davis L, Orvosh D et al (1993) A genetic algorithm for survivable network design.
In: Proceedings of 5th international conference on genetic algorithms, pp 408–415
6. Deb BK (2010) Optimization for engineering design: algorithms and examples.
Prentice-Hall, New Delhi
7. Deb K (1989) Genetic algorithms in multimodal function optimization. Master’s
thesis, University of Alabama
8. Deb K, Thiele L et al (2001) Scalable test problems for evolutionary multiobjective
optimization. Wiley, Chichester
9. Dorigo M (1992) Optimization, learning and natural algorithms. PhD thesis,
Politecnico di Milano
10. Duhamel C, Gouveia L et al (2012) Models and heuristics for the k-degree
constrained minimum spanning tree problem with node-degree costs. Networks
60(1):1–18
11. Fawcett H (2014) Manual of political economy. Macmillan and co., London
12. Fogel LJ, Owens AJ, Walsh MJ (1966) Artificial intelligence through simulated
evolution. Wiley, New York
13. Fonseca C, Fleming P (1995) An overview of evolutionary algorithms in multiob-
jective optimization. IEEE Trans Evol Comput 3(1):1–16
396 L. Lin et al.
14. Garey MR, Johnson DS (1979) Computers and intractability: a guide to the theory
of np-completeness. W.H. Freeman, New York
15. Gen M (2006) Genetic algorithms and their applications. Springer, London
16. Gen M, Cheng R, Oren SS (2000) Network design techniques using adapted genetic
algorithms. Springer, London
genetic algorithm approach. Springer, London
18. Goldberg DE (1989) Genetic algorithms in search, optimization, and machine
learning. Addison-Wesley Pub. Co., Boston
19. Hansen P (1979) Bicriterion path problems. In: Proceedings of 3rd conference
multiple criteria decision making theory and application, pp 109–127
20. Hao X, Gen M et al (2015) Effective multiobjective EDA for bi-criteria stochastic
job-shop scheduling problem. J Intell Manufact 28:1–13
21. Holland J (1975) Adaptation in Natural and Artificial System. MIT Press, Ann
Arbor
22. Holland JH (1976) Adaptation. In: Rosen R, Snell FM (eds) Progress in theoretical
biology IV
23. Hwang CL, Yoon K (1994) Multiple attribute decision making. Springer, Heidel-
berg
24. Ishibuchi H, Murata T (1998) A multi-objective genetic local search algorithm and
its application to flowshop scheduling. Comput Ind Eng 28(3):392–403
25. Kennedy J, Eberhart R (2011) Particle swarm optimization. In: Proceeding of the
IEEE international conference on neural networks, Piscataway, pp 1942–1948
26. Kennedy J, Eberhart R (2011) Particle swarm optimization. Morgan Kaufmann,
San Francisco
27. Koza JR (1992) Genetic programming, the next generation. MIT Press, Cambridge
28. Koza JR (1994) Genetic programming II (videotape): the next generation. MIT
Press, Cambridge
29. Liang W, Schweitzer P, Xu Z (2013) Approximation algorithms for capacitated
minimum forest problems in wireless sensor networks with a mobile sink. IEEE
Trans Comput 62(10):1932–1944
30. Lin L (2006) Node-based genetic algorithm for communication spanning tree prob-
lem. IEICE Trans Commun 89(4):1091–1098
31. Lin L, Gen M (2008) An effective evolutionary approach for bicriteria shortest path
routing problems. IEEJ Trans Electron Inf Syst 128(3):416–423
32. Mathur R, Khan I, Choudhary V (2013) Genetic algorithm for dynamic capacitated
minimum spanning tree. Comput Technol Appl 4(3):404
33. Medhi D, Pioro M (2004) Routing, flow, and capacity design in communication
and computer networks. Morgan Kaufmann Publishers, San Francisco
34. Piggott P, Suraweera F (1995) Encoding graphs for genetic algorithms: an investi-
gation using the minimum spanning tree problem. Springer, Heidelberg
35. Raidl GR, Julstrom B (2003) Edge-sets: an effective evolutionary coding of span-
ning trees. IEEE Trans Evol. Comput. 7(3):225–239
36. Rechenberg I (1973) Optimieriung technischer Systeme nach Prinzipien der biolo-
gischen Evolution. Frommann-Holzboog, Stuttgart
37. Ruiz E, Albareda-Sambola M et al (2015) A biased random-key genetic algorithm
for the capacitated minimum spanning tree problem. Comput Oper Res 57:95–108
algorithms. In: International conference on genetic algorithms, pp 93–100
39. Schwefel HPP (1995) Evolution and Optimum Seeking: The Sixth Generation.
Wiley, New York
40. Srinivas N, Deb K (1995) Multiobjective function optimization using nondominated

sorting genetic algorithms. IEEE Trans Evol Comput 2(3):221–248
41. Steiner S, Radzik T (2008) Computing all efficient solutions of the biobjective
minimum spanning tree problem. Comput Oper Res 35(1):198–211
42. Steuer RE (1986) Multiple criteria optimization theory computation and appli-
cation Wiley series in probability and mathematical statistics. Probability and
mathematical statistics. Wiley, New York
43. Torkestani JA (2013) Degree constrained minimum spanning tree problem: a learn-
ing automata approach. J Supercomput 64(1):226–249
44. Torkestani JA, Meybodi MR (2012) A learning automata-based heuristic algorithm
for solving the minimum spanning tree problem in stochastic graphs. J Supercom-
put 59(2):1035–1054
45. Torkestani JA, Rezaei Z (2013) A learning automata based approach to the
bounded-diameter minimum spanning tree problem. J Chin Inst Eng 36(6):749–759
46. Voβ S, Martello S, et al (2012) Meta-heuristics: Advances and trends in local search
paradigms for optimization
47. Zandieh M (2011) A novel imperialist competitive algorithm for bi-criteria schedul-
ing of the assembly flowshop problem. Int J Prod Res 49(11):3087–3103
48. Zhang Q, Ding L (2016) A new crossover mechanism for genetic algorithms with
variable-length chromosomes for path optimization problems. Expert Syst Appl
60:183–189
49. Zitzler E, Laumanns M, Thiele L (2001) SPEA2: improving the strength pareto
evolutionary algorithm for multiobjective optimization. In: Evolutionary methods
for design, optimization and control with applications to industrial problems. Pro-
ceedings of the Eurogen 2001, Athens, Greece, September
Study of County Level Government’s
E-Government Efficiency Evaluation in Sichuan
Province Based on DEA
Yu Liu(B) , Jiawei Liu, Zhaohui Zhang, and Liangqing Zhang
School of Political Science and Public Administration,

University of Electronic Science and Technology of China,
laujames work@163.com
Abstract. The county government plays a fundamental role in the Chi-

nese governance. The e-government system of it can directly reflect the
quality and efficiency of its service. Thus, it’s significant to evaluate the
efficiency of the e-government at county level. Nowadays, there’re plenty
of methods to do this kind of evaluation. This paper mainly uses principal
component analysis to establish the index system, and adopts the CCR
and BCC models, two of the data envelopment analysis DEA, to reflect
the current input and output efficiency of 147 counties’ e-government in
Sichuan Province. The analysis reveals that most of the e-governments’
comprehensive efficiency, technical efficiency and scale efficiency are non-
DEA effective in county level governments of Sichuan Province, and the
differences in development among them are quite distinctive. What’s
more, the local governments’ information department should embark
on improving their investment scale and technical efficiency in order to
enhance the efficiency of e-government.
Keywords: E-government · County government · Data envelopment

analysis · Efficiency
1 Introduction
The county government has long been the basis of administration management
and state governance in China. It faces to the grass-roots and serves them
directly. Thus, the efficiency of its operation can greatly influence the over-
all quality of people’s life and production. Nowadays, as an effective way of
improving administrative efficiency and transparency, the e-government has been
praised highly by many countries. And China has also launched the Government
Online Project since 1999. In fact, the construction of e-government within the
county level is a very important part of China’s e-government development, as
well as the terminal node to the provincial level e-government, and governments
should pay great attention to e-government construction at county level for con-
structing mature e-government system structure [7]. To promote the healthy and
County Level Government’s E-Government Efficiency Evaluation 399
sustainable development of a county e-government, an evaluation of its efficiency

which may help improve its service quality is strongly required.
Evaluation systems like the Global E-government Evaluation from the UN
[11], the “eEurope” Strategic Assessment System from the EU [3] and the Overall
e-government Maturity Assessment from the Accenture company [1] are inter-
national representatives. While in China, it also has assessment models like the
CCW Research [12], the China Software Testing Center [4], the CCID Consulting
[5] and the CCTC [2].
According to the actual situation, the E-government efficiency evaluation
index system was constructed from different perspectives and the correspond-
ing analysis methods are used to evaluate the efficiency of the E-government.
Extraordinary representatives: Luo and Shi [10] introduced genetic algorithm
optimized to BP neural network weights and thresholds after reducing the
established evaluation index system by rough set, and established e-government
website evaluation model based on genetic neural network algorithm; Rorissa,
Demissie and Pardo [13] used benchmarking to assess the strengths and limi-
tations of six frameworks for computing e-government indexes; Luna et al. [9]
used data envelopment analysis (DEA) to show how efficient are Mexico state
governments in their use of certain inputs to produce high quality e-government
portals; Zdjelar [17] measured efficiency of Croatian regional government imple-
mentation by Balanced Scorecard Method; Song and Guan [14] utilized a super
efficiency slack-based measure (SBM) model to evaluate the e-government per-
formance of environmental protection administrations in the 16 cities of Anhui
Province; Kao [8] evaluated the e-government of 21 European countries based
on the concept of Pareto optimality and developed a compromise programming
model to rank these nations; Wu and Guo [15] used method of DEA to evalu-
ate e-government performances of 31 provincial government websites in China
and concluded that most of these provincial government websites operated at an
inefficient level and in a bad manner.
To sum up, the achievements of researches on the e-government’s efficiency
provide a useful reference to the study. However, the evaluation of Chinese e-
government’s efficiency has not yet had a uniform index system. Most researches
are focusing on the overall efficiency of the provincial research objects and
few investigate the input-output efficiency of e-government from the perspec-
tive of e-government at county level. Therefore, this paper attempts to use
principal component analysis to build the evaluation index system of county
level e-government’s efficiency, and adopt the DEA model to evaluate the effi-
ciency of e-government in 147 counties (including county level cities) of Sichuan
Province to provide decision-making reference for the promotion of county level
e-government’s efficiency.
The remainder of this paper is organized as follows. In the first half of Sect. 2,
we briefly discuss the principle of principal component analysis and the practical
application value of it. In the second half of Sect. 2, we narrate the construction
of CCR model and BCC model of DEA which will be used for e-government effi-
ciency evaluation. The process of e-government efficiency evaluation is presented
400 Y. Liu et al.
in Sect. 3 which includes the samples selection, data collection, the establishment
of evaluation index system and the model solution. On the basis of the solution,
the analysis of the current situation in Sichuan Province and the suggestions for
the ineffective county level government are given in Sect. 4. Finally, we conclude
with a summary of this paper and prospect about evaluation of e-government in
Sect. 5.
2 The Method of E-Government Efficiency Evaluation

2.1 Principal Component Analysis
On the premise of the least information loss, principal component analysis is a

statistical procedure that uses the thought of dimension reduction to convert a
set of multiple indexes with possible correlation into a few linearly uncorrelated
comprehensive indexes.
The basic principles of principal component analysis are as follows:
We assume that there are p objects that each object is described by n indexes.
In this way, we can constitute a data matrix:
⎛ ⎞
x11 x12 · · · x1n
⎜ x21 x22 · · · x2n ⎟
⎜ ⎟
X=⎜ . .. .. ⎟ = (X1 , X2 , · · · , Xn ). (1)
⎝ .. . . ⎠
xp1 xp2 · · · xpn
X1 , X2 , · · · , Xn are defined as original variable indexes and F1 , F2 , · · · ,

Fm (m ≤ n) are defined as new variable indexes. By using the n vectors
X1 , X2 , · · · , Xn of data matrix to make a linear combination, then we can get:
⎧
⎪ F1 = a11 X1 + a12 X2 + · · · + a1n Xn
⎪
⎨
F2 = a21 X1 + a22 X2 + · · · + a2n Xn
(2)
⎪
⎪ ··· ···
⎩
Fm = am1 X1 + am2 X2 + · · · + amn Xn
Coefficient aij meets:
• Fi is linearly independent with Fj (i = j; i, j = 1, 2, · · · , n);

• F1 has the greatest variance of the linear combination of X1 , X2 , · · · , Xn
mentioned above. F2 has the greatest variance of the linear combination of
X1 , X2 , · · · , Xn which are independent with F1 . The rest could be deduced
in this way. Fm has the greatest variance of the linear combination of
X1 , X2 , · · · , Xn which are independent with F1 , F2 , · · · , Fm−1 (m ≤ n).
New variable indexes F1 , F2 , · · · , Fm (m ≤ n) are called the first, second, · · · ,

mth principal component respectively.
2.2 Data Envelopment Analysis

Based on the relative efficiency concept, the data envelopment analysis is a
non-parametric statistical method employing convex analysis and linear pro-
gramming as tools to calculate the relative efficiency between the same type of
multi-input and multi-output decision making units. The relative efficiency is
used to evaluate whether the decision-making unit is effective.
(1) CCR Mode
Assume that there are n decision-making units (hereinafter referred to as
DMU), each DMU has n inputs and m outputs. The input vector of DM Uj
is Xj = (x1j , x2j , · · · , xmj )T ≥ 0, Xj ∈ R+ while output vector is Yj =
(y1j , y2j , · · · , ysj )T ≥ 0, Yj ∈ R+ Building a model with a non-Archimedes infi-
nitely small to evaluate the relative efficiency of the DMUj :
min ⎧θ − ε(eT1 s− + eT2 s+ )
⎪ n
⎪
⎪ λj xj + s− = θx0
⎪
⎪
⎪
⎨
j=1
n
(3)
s.t. λj yj − s+ = y0
⎪
⎪
⎪ j=1
⎪
⎪
⎪ λ ≥ 0, j = 1, 2, · · · , n
⎩ +j
s ≥ 0, s− ≥ 0
ε is non-Archimedes infinitely small. s− is the slack variable while s+ is the
remaining variable. Assuming that Eq. (3) has the best solution θ0 , λ0 , s−0 , s0
+
−
for DMUj0 . If θ0 < 1, DMUj0 will be non-DEA effective. If θ0 = 1 and s0 or
0 = 0, DMUj0 will be weakly DEA effective. If θ0 = 1 and DMUj0 will be DEA
s+
effective.
(2) BCC Model
BCC model is the extension of
fixed scale benefit data envelopment analysis.
n
Adding the convexity condition j=1 λj = 1 to the linear programming under
fixed scale benefit model (CCR model) is the BCC model:
min ⎧θ − ε(eT1 s− + eT2 s+ )
⎪ n
⎪
⎪ λj xj + s− = θx0
⎪
⎪
⎪
⎪
j=1
⎪
⎪ n
⎪
⎨ λj yj − s+ = y0
j=1 (4)
s.t. n
⎪
⎪ λj = 1
⎪
⎪
⎪
⎪ j=1
⎪
⎪
⎪
⎪ λ ≥ 0, j = 1, 2, · · · , n
⎩ +j
s ≥ 0, s− ≥ 0.
3 The Process of E-Government Efficiency Evaluation

3.1 Samples, Indexes and Data Collection
The decision-making units in DEA should have the same type of characteristics.
In order to evaluate the relative effectiveness of e-government at the county level
402 Y. Liu et al.
in Sichuan Province, this paper selects the 147 districts and counties (including
county level cities) in Sichuan Province as the objects of the research.
In accordance with the requirements of the DEA method, we need to estab-
lish an index system of input and output based on the comprehensive consider-
ation of region informatization level, economic development, cultural and edu-
cational level and other factors. There are nine categories of input indexes in
this paper which specifically include fixed telephone subscribers, the number of
mobile phone users, the value of the tertiary industry income, GDP, total fis-
cal revenue, governmental investment in science and education, the proportion
of urban population, the number of students in colleges and universities, the
number of colleges and universities. These index data are mainly obtained from
“Sichuan Province Statistical Yearbook” (2014) [16] and “The Statistics Bulletin
of the National Economy and Social Development” (2013) issued by the various
districts and counties (including county level cities).
To select proper output indicators, we referenced the objective indexes
derived from CCID consulting assessment of government website performance
and The United Nations for the global objective of e-government assessment.
This study revolves around the authoritative evaluation results to screen the
indexes. The five main evaluation indexes consist of government information
publicity, work services, public participation, the website management, applica-
tion of new technology. Output index data are mainly obtained from “The total
Report of the Sichuan Government Website Performance Evaluation” (2013) [6].
3.2 Principal Component Analysis
The principal component analysis of 9 input indexes and 5 output indexes data
were conducted by SPSS software (Tables 1 and 2).
According to extracts of the principal component to calculate the principal
component load, the formula is:
√
lij = p(zi , xj ) = λi aij (i = 1, 2, · · · , k, k as main composition, j = 1, 2, · · · , 147)
(5)
Then we can get the principal component index data of inputs and outputs.
Table 1. The result of PCA of the input indexes
Principal component analysis Eigenvalue

F1 5.963
Table 2. The result of PCA of the output indexes
Principal component analysis Eigenvalue

E1 2.469
E2 1.004
Table 3. The results of e-government input and output efficiency evaluation (1)
City or Autonomous County CCR BCC Scale

prefecture efficiency
Chengdu City Jinjiang District 0.919 1 0.919
Qingyang District 0.581 0.581 0.999
Jinniu District 0.557 0.582 0.957
Wuhou District 1 1 1
Chenghua District 0.832 0.995 0.836
Longquanyi District 0.627 0.627 0.999
Qingbaijiang District 0.226 0.226 1
Xindu District 0.583 0.868 0.672
Wenjiang District 0.464 0.465 0.999
Dujiangyan City 0.318 0.318 0.998
Pengzhou City 0.261 0.262 0.999
Qionglai City 0.17 0.17 0.999
Chongzhou City 0.193 0.385 0.5
Jintang County 0.18 0.18 0.999
Shuangliu County 0.673 0.673 1
Pixian County 0.587 1 0.587
Dayi County 0.185 0.187 0.991
Pujiang County 0.226 0.226 0.999
Xinjin County 0.318 0.77 0.413
Guangan City Guangan District 0.201 0.202 0.998
Linshui County 0.179 0.18 0.995
Wusheng County 0.157 0.158 0.996
Yuechi County 0.158 1 0.158
Huayun City 0.139 0.14 0.997
Bazhong City Bazhou District 0.185 0.185 1
Pingchang County 0.153 0.153 0.996
Tongjiang County 0.144 0.144 0.997
Nanjiang County 0.175 0.175 0.998
Meishan City Dongpo District 0.257 0.258 0.999
Pengshan District 0.138 0.218 0.633
Renshou County 0.221 0.221 0.998
Qingshen County 0.088 0.113 0.777
Hongya County 0.092 0.1 0.922
(Continued)
404 Y. Liu et al.

Tibetan Autonomous Kangding County 0.199 0.212 0.936
Prefecture of Garze Luding County 0.146 0.213 0.686
Jiulong County 0.117 0.218 0.537
Dawu County 0.088 0.159 0.551
Litang County 0.095 0.152 0.622
Daocheng County 0.132 0.26 0.507
Yajiang County 0.104 0.204 0.511
Liangshan Autonomous Xichang City 0.341 0.342 0.998
Prefecture Dechang County 0.116 0.146 0.793
Huili County 0.162 0.162 0.999
Huidong County 0.12 0.137 0.876
Ningnan County 0.123 0.151 0.817
Jinyang County 0.077 0.145 0.529
Butuo County 0.083 0.16 0.516
Zhaojue County 0.096 0.159 0.599
Leibo County 0.124 0.18 0.689
Ganluo County 0.105 0.181 0.582
Yuexi County 0.098 0.148 0.666
On account of that the result of the principal component analysis is negative

number, and DEA model requires that the data is not negative, so the results
of the principal component calculated by poor transformation normalized to [1,
10] interval. The range transform formula is:
nnew = 1 + 9 × [(n − nmin )/(nmax − nmin )], (6)
where nmax and nmin represent the maximum and minimum values of the column
data respectively (Tables 3 and 4).
3.3 The Solution of the DEA Model
According to the principal component analysis method, the main components

of input and output index data are obtained. After that, we use the DEAP 2.1
to solve 147 counties’ (including districts’ and cities’) input-output efficiency
of e-government with CCR and BCC model, and we obtain the comprehensive
efficiency, pure technical efficiency and scale efficiency of all counties’ (including
districts’ and cities’) e-government. All results are shown in Table 5.

Yibin City Cuiping District 0.386 0.387 0.999
Nanxi District 0.166 0.19 0.874
Yibin County 0.184 0.185 0.995
Jiang’an County 0.153 0.154 0.995
Gongxian County 1 1 1
Junlian County 0.151 0.168 0.903
Xingwen County 0.159 0.164 0.967
Pingshan County 0.106 0.161 0.659
Ya’anCity Yucheng District 0.423 0.424 0.998
Hanyuan County 0.171 0.145 0.848
Shimian County 0.117 0.151 0.774
Lushan County 0.114 0.152 0.747
Nanchong City Shunqing District 0.346 0.347 0.999
Gaoping District 0.134 0.134 0.998
Xichong County 0.124 0.124 0.996
Nanbu County 0.16 0.161 0.998
Yilong County 0.141 0.142 0.998
Yingshan County 0.147 0.148 0.995
Peng’an County 0.107 0.108 0.998
Luzhou City Jiangyang District 0.282 0.283 0.998
Naxi District 0.261 0.262 0.999
Luxian County 0.216 0.217 0.998
Gulin County 0.137 0.137 0.998
Dazhou City Tongchuan District 0.277 0.278 0.998
Dachuan District 0.234 0.234 1
Xuanhan County 0.267 0.267 1
Kaijiang County 0.166 0.171 0.969
Dazhu County 0.276 0.277 0.996
Quxian County 0.25 0.251 0.997
Wanyuan City 0.196 0.409 0.48
Aba Tibetan and Qiang Maerkang County 0.171 0.189 0.903
Autonomous Prefecture Jinchuan County 0.089 0.151 0.594
Xiaojin County 0.152 0.181 0.841
Aba County 0.088 0.161 0.55
Ruoergai County 0.101 0.132 0.771
(Continued)
406 Y. Liu et al.

Hongyuan County 0.128 0.218 0.587
Rangtang County 0.085 0.175 0.485
Wenchuan County 0.124 0.145 0.858
Lixian County 0.096 0.159 0.599
Maoxian County 0.108 0.147 0.735
Songpan County 0.114 0.177 0.645
Jiuzhaigou County 0.103 0.133 0.775
Heishui County 0.095 0.163 0.585
Guangyuan City Lizhou District 0.248 0.249 1
Zhaohua District 0.128 0.271 0.473
Chaotian District 0.089 0.115 0.777
Wangcang Distric 0.18 0.183 0.98
Qingchuan County 0.17 0.232 0.732
Jiange County 0.19 0.196 0.967
Cangxi County 0.191 0.192 0.996

Prefecture efficiency
Suining City Chuanshan District 0.26 0.26 0.997
Anju District 0.127 0.146 0.871
Shehong County 0.181 0.209 0.867
Pengxi County 0.109 0.114 0.957
Daying County 0.121 0.134 0.905
Neijiang City Central District 0.353 0.58 0.609
Dongxing District 0.27 0.319 0.846
Zizhong County 0.216 0.367 0.588
Longchang County 0.245 0.302 0.81
Weiyuan County 0.211 0.407 0.518
Leshan City Central District 0.399 0.4 0.998
Shawan District 0.18 0.191 0.943
Wutongqiao District 0.154 0.154 0.998
Jinkouhe District 0.102 0.152 0.67
Qianwei County 0.186 0.187 0.997
(Continued)

Prefecture efficiency
Jingyan County 0.122 0.145 0.839
Jiajiang County 0.153 0.153 0.997
Muchuan County 0.129 0.175 0.733
Emeishan City 0.261 0.262 0.999
Ebian Yi 0.1 0.148 0.674
Autonomous County
Mabian Yi 0.107 0.162 0.664
Autonomous County
Ziyang City Yanjiang District 0.206 0.207 0.998
Jianyang City 0.236 0.236 0.998
Mianyang City Fucheng District 0.554 0.555 1
Youxian District 0.219 0.219 0.998
Zitong County 0.12 0.141 0.85
Santai County 0.204 0.204 0.998
Yanting County 0.195 0.201 0.969
Anxian County 0.133 0.14 0.954
Pingwu County 0.145 0.156 0.93
Jiangyou City 0.261 0.262 0.998
Beichuan Qiang 0.106 0.14 0.756
Autonomous Region
Zigong City Ziliujing District 0.368 0.369 0.997
Yantan District 0.168 0.198 0.849
Fushun County 0.224 0.225 0.996
Panzhihua City East District 0.396 0.402 0.984
West District 0.227 0.564 0.403
RenHe district 0.127 0.138 0.92
Miyi County 0.14 0.159 0.883
Yanbian County 0.199 0.23 0.864
Deyang City Jingyang District 0.312 0.312 0.998
Guanghan City 0.205 0.205 0.998
Shifang City 0.19 0.191 0.998
Mianzhu City 0.169 0.17 0.996
Zhongjiang County 0.184 0.185 0.998
Luoyang County 0.124 0.134 0.92
408 Y. Liu et al.
4 Analysis of the Result

Through the analysis of CCR, it appears that Wuhou District in Chengdu and
Gongxian County in Yibin have achieved a full effectiveness of comprehensive
efficiency (CCR = 1.000) among 147 counties, districts, or cities of Sichuan
Province, which indicates that they have obtained a fully efficient outcome.
However, two districts, Jinjiang and Chenghua in Chengdu, are in the position
of a light level of non-DEA effective degree (0.700 ≤ CCR < 1.000). For those
of a light level of non-effective degree, the development of electronic adminis-
tration requires focusing more on the control and management of details and
proper distribution of the investment, as there’s still some room for improve-
ment. Meanwhile, most of counties (including districts and cities) are in a severe
level of non-DEA effective degree (CCR ≤ 0.500), which suggests that most
counties’ e-government development in Sichuan province don’t work out. The
overall development gap is huge, and most of objects have not achieved the com-
prehensive effectiveness. The input of the e-government construction has not
acquired the desired results. On the one hand, the technical inefficiency leads
to the comprehensive ineffectiveness. On the other hand, the low scale efficiency
also affects the improvement of the comprehensive efficiency. Therefore, the gov-
ernment should make a huge effort to find out the route to suit the situation, so
as to surmount the problem fundamentally. The counties and districts with the
lowest CCR, Jinyang County and Butuo County in Liangshan, Rangtang County
in Tibetan Qiang Autonomous Prefecture of Aba, Qingshen county in Meishan,
Dawu County in Tibetan Autonomous Prefecture of Garze, Aba County in Aba
Tibetan and Qiang Autonomous Prefecture, correspond to their relatively weak
national economy and social development, as well as the influence of little cap-
ital invested by the government on information construction. Thus, problems
should be solved and development should be sped up in the future construction
of e-government.
The pure technical efficiency results obtained by the BCC model show that
Jinjiang district, Wuhou District and Pixian County of Chengdu are with pure
technical effective (BCC = 1.00). It indicates that their services systems of
e-government run well, and the resources invested can develop productivity effi-
ciently. The pure technical efficiency of 132 counties, districts or cities are lower
than 0.5, and it appears that these local governments’ technical outcomes are
not enough. The last 5 counties, districts or cities with the lowest BCC succes-
sively are Hongya County in Meishan, Peng’an County in Nanchong, Qingshen
County in Meishan, Pengxi County in Suining, Chaotian District in Guangyuan.
It indicates these governments didn’t distribute the resources legitimately, and
they need more technological innovation to improve the service abilities of e-
government from technology level.
From the point of view of scale efficiency, there are 9 counties or dis-
tricts of Sichuan Province in full scale efficiency. They are Wuhou District,
Qingbaijiang District and Shuangliu County in Chengdu, Gongxian County
in Yibin, Dachuan District and Xuanhan District in Dazhou, Fucheng District
in Mianyang, Lizhou District in Guangyuan and Bazhou District in Bazhong.
These governments input resources in e-government are proper, and returns of

scale remain unchanged. The rest 138 counties, districts or cities are not in
full scale efficiency, it manifests that these governments waste some resources
during the construction of e-government. In the future, the governments need to
increase the investment in information technology and science and education, for
speeding up the process of urbanization and informatization. What’s more, the
governments should allocate resources more reasonably to improve the level of
government information disclosure, services and websites management. Besides,
governments should make more and more people participate in the construction
of e-government in order to take better use of the resources.
5 Conclusions
Due to the wave of the informatization and popularization of the network, the
construction of e-government in the human condition is getting better and bet-
ter. Governments at all levels are also vigorously investing money and human
resources for the construction of e-government to promote the informatization.
To guide the construction of e-government, how to accurately evaluate and sci-
entifically observe the efficiency of e-government is a hotspot and difficulty in the
field of current e-government. Although there are a lot of foreign standards being
references, different countries may have different national conditions. The real
effective e-government construction has to be combined with the actual situation.
According to the views of scholars and organizations at home and abroad, based
on observation and thinking, we use different indexes and principal component
analysis and DEA model to evaluate the e-government construction efficiency of
all districts and counties of Sichuan Province. We hope to be able to scientifically
evaluate the comprehensive efficiency of various districts and counties of Sichuan
Province in the construction of e-government. On the whole, the local govern-
ment’s information department should focus on improving their investment scale
and technical efficiency to enhance the efficiency of e-government.
However, this paper also has some limitations. Firstly, we take both input
and output into consideration which also have been rotated, extracted and stan-
dardized. Therefore, this paper couldn’t use the projection algorithm to get the
improvement solutions which give the target value and scope for the non-DEA
effective. Secondly, DEA focus on the relative efficiency. It means that different
evaluation range may have different result. So how to choose the evaluation range
is a project worth exploring. In later study, we will commence on diminishing
the limitations above and optimize the solution.
Acknowledgements. We would like to express our sincere appreciation to the review-

ers for their insightful comments, which have greatly aided us improving the quality
of the paper. This research was supported by the Soft Science Research Project of
Chengdu (Grant 2016-RK00-00041-ZF).
410 Y. Liu et al.
References
1. Accenture (2006) E-government leadership: high performance and maximum value.
http://www.accenture.com
2. CCCT (2015) General report on performance evaluation of Chinese government
website in 2015. http://ccct.mofcom.gov.cn/article/zt jxpg2015
3. Commission E (2006) i2010–a European information society for growth and
employment. http://www.docin.com/p-28146066.html
4. CSTC (2013) 2013 Chinese government website performance evaluation index sys-
tem. http://www.cstc.org.cn/templet/default/show nbkw.jsp?article id=123668&
id=1619
5. CSTC (2015) 2015 Chinese government website performance evaluation results.
http://www.ccidnet.com/2015/1209/10063631.shtml
6. CSTC SPegnoc (2014) The total report of the Sichuan government website perfor-
mance evaluation 2014
7. Guo X, Zou J (2015) A summary of the current situation and pattern of
e-government development in China’s rural areas. E-Government 1:90–96 (in
Chinese)
8. Kao C (2015) Evaluation and improvement of e-government: the case of European
countries. In: International conference on E-democracy & E-government, pp 1211–
1216
9. Luna DE et al (2010) Using data envelopment analysis (DEA) to assess govern-
ment web portals performance. In: Proceedings of the 13th annual international
conference on digital government research, pp 107–115
10. Luo D, Shi Y (2010) Evaluation on e-government websites based on rough set
and genetic neural network algorithm. In: Proceedings of the third international
symposium on computer science and computational technology, pp 124–128
11. Nations U (2014) United nations e-government survey 2014: E-government for the
future we want
12. Research C (2004) Description of government website evaluation index system.
http://dc.ccw.com.cn
13. Rorissa A, Demissie D, Pardo T (2011) Benchmarking e-government: a comparison
of frameworks for computing e-government index and ranking. Gov Inf Q 28:354–
362 (in English)
14. Song M, Guan Y (2014) The electronic government performance of environmental
protection administrations in Anhui province. Technol Forecast Soc Change 96:79–
88 (in Chinese)
15. Song M, Guan Y (2015) Measuring e-government performance of provincial govern-
ment website in China with slacks-based efficiency measurement. Technol Forecast
Soc Change 39:25–31 (in Chinese)
16. Sichuan Provincial Bureau of Statistics SPB (2014) Sichuan province statistical
yearbook 2014
17. Zdjelar R (2013) Measuring success of implementation county development strat-
egy by using balanced scorecard method. In: International convention on informa-
tion & communication technology electronics & microelectronics, pp 1211–1216
Comparing Visitors’ Behavior Through Mobile
Phone Users’ Location Data
Masahide Yamamoto(B)
Faculty of Foreign Studies, Nagoya Gakuin University, Nagoya, Japan

myama@ngu.ac.jp
Abstract. In recent years, so-called “big data” have been attracting the
attention of companies and researchers. This study aims to identify the
number of visitors of each period and their characteristics based on the
location data of mobile phone users collected by the mobile phone com-
pany. The study sites of this survey are tourist destinations in Ishikawa
Prefecture and Toyama city, including Kanazawa city, which became
nationally popular after the Hokuriku Shinkansen opened in 2015. The
opening of the Hokuriku Shinkansen brought more visitors to many areas.
However, it also led to fewer visitors in some areas. The positive effect
was remarkable in Kanazawa.
Keywords: Mobile phone · Location data · Tourism
1 Introduction
In Japan, many people have struggled to promote tourism in their regions to

vitalize their local economies. The low-cost carriers have boosted competition in
the transportation industry, and domestic transportation costs have declined in
several regions. Therefore, tourism is likely to become increasingly important to
local economies. Initially, Japan’s tourism industry suffered significant volatility
in demand depending on the season and day of the week. In addition, there was
significant loss of business opportunities because of congestion during the busy
season.
To cope with such volatility, various events have been held to eliminate the
seasonal gap. Many events are newly launched. To date, it has been difficult to
accurately grasp the extent to which these events attract visitors and the types of
people who visit. However, by employing the recently provided Information and
Communication Technology (ICT) services, it is possible to verify the number
and characteristics of visitors to a particular event.
This study attempts to identify the number of visitors in different periods and
their characteristics based on the location data of mobile phone users collected
by the mobile phone company. In addition, it also attempts to demonstrate an
alternative method to more accurately measure the number of visitors attracted
by an event.
412 M. Yamamoto
2 Method
This study used “MOBILE KUUKAN TOUKEIT M ” (mobile spatial statistics)
provided by NTT DoCoMo, Inc. and DoCoMo Insight Marketing, Inc. to collect
the location data of mobile phone users in order to count the number of visi-
tors at specific tourist destinations and examine their characteristics. MOBILE
KUUKAN TOUKEIT M is statistical population data created by a mobile phone
network. It is possible to estimate the population structure of a region by gender,
age, and residence using this service of a particular company.
Table 1. Survey areas and regional mesh codes
Survey areas Regional mesh code Type of codes

Kanazawa Kanazawa Station 5436-6591-2 1/2
Kenrokuen 5436-6572 + 5436-6573-1, 5436-6573-3 Tertiary, 1/2
Higashi Chayagai 5436-6583-3 1/2
Nanao Wakura Hot Springs 5536-5703 Tertiary
Nanao Station 5536-4757 Tertiary
Kaga Yamanaka Hot Springs 5436-2299, 5436-2390 Tertiary
Wajima Wajima 5636-0772 Tertiary
Toyama Toyama Station 5537-0147-1 1/2
Note: A regional mesh code is a code for identifying the regional mesh, which is substantially
divided into the same size of a square (mesh) based on the latitude and longitude in order
to use it for statistics. The length of one side of a primary mesh is about 80 km, and those
of secondary and tertiary meshes are about 10 Km and 1 km respectively.
The sites studied in this survey are tourist destinations in Ishikawa Prefecture
and Toyama city, including Kanazawa city, which became nationally popular
when the Hokuriku Shinkansen (high-speed railway) opened in 2015. Moreover,
the locations and characteristics of the individuals obtained herein are derived
through a non-identification process, aggregation processing, and concealment
processing. Therefore, it is impossible to identify specific individuals.
The survey areas are presented in Table 1 and Fig. 1. A regional mesh code
is a code for identifying the regional mesh. It stands for an encoded area that is
substantially divided into the same size of a square (mesh) based on the latitude
and longitude in order to use it for statistics. With regard to regional mesh,
there are three types of meshes: primary, secondary, and tertiary. The length of
one side of a primary mesh is about 80 km, and those of secondary and tertiary
meshes are about 10 Km and 1 km respectively.
In addition, split regional meshes also exist, which are a more detailed
regional division. A half-regional mesh is a tertiary mesh that is divided into
two equal pieces in the vertical and horizontal directions. The length of one side
is about 500 m. Furthermore, the length of one side of a quarter and 1/8 regional
meshes is about 250 m and 125 m respectively.
Comparing Visitors’ Behavior Through Mobile Phone Users’ Location Data 413
Fig. 1. Survey areas
This study analyzed the location data collected from NTT DoCoMo, Inc.
to consider the effect of the opening of the Hokuriku Shinkansen on the survey
areas.
3 Previous Research
Previous tourism marketing research has primarily focused on the ways service
promises are made and kept, mostly generating frameworks to improve man-
agerial decisions or providing insights on associations between constructs [4].
Big data have become important in many research areas, such as data mining,
machine learning, computational intelligence, information fusion, the semantic
Web, and social networks [3]. To date, several attempts have been made to use
large-scale data or mobile phone location data in tourism marketing studies.
Most studies dealing with big data in tourism were published after 2010.
Fuchs et al. [5] presented a knowledge infrastructure that has recently been
implemented at the leading Swedish mountain tourism destination, Åre. Using
a Business Intelligence approach, the Destination Management Information Sys-
tem Åre (DMIS-Åre) drives knowledge creation and application as a precondition
of organizational learning at tourism destinations. Xiang et al. [9] tried to apply
big data to tourism marketing. The study aimed to explore and demonstrate the
utility of big data analytics to better understand important hospitality issues,
namely, the relationship between hotel guest experience and satisfaction. Specif-
ically, the investigators applied a text analytical approach to a large number of
414 M. Yamamoto
consumer reviews extracted from Expedia.com to deconstruct hotel guest expe-

riences and examine the association with satisfaction ratings. These studies are
similar to this study in that they attempted to utilize big data in tourism. How-
ever, the research methods and objectives of these studies are different from that
of the present study.
Studies on using mobile phone location data for tourism surveys can be traced
back to 2008. Ahas et al. [2] introduced the applicability of passive mobile posi-
tioning data for studying tourism. They used a database of roaming location
(foreign phones) and call activities in network cells: the location, time, random
identification, and country of origin of each called phone. Using examples from
Estonia, their study described the peculiarities of the data, data gathering, sam-
pling, the handling of the spatial database, and some analytical methods to
demonstrate that mobile positioning data have valuable applications for geo-
graphic studies. Japan Tourism Agency conducted a similar study using inter-
national roaming service in December 2014 [1].
Since the creative work of Ahas et al. [2], several studies employing location
data have emerged. Liu et al. [7] investigated the extent to which behavioral
routines could reveal the activities being performed at mobile phone call loca-
tions captured when users initiate or receive voice calls or messages. Using data
collected from the natural mobile phone communication patterns of 80 users
over more than a year, they assessed the approach via a set of extensive exper-
iments. Based on the ensemble of models, they achieved prediction accuracy of
69.7%. The experiment results demonstrated the potential to annotate mobile
phone locations based on the integration of data mining techniques with the
characteristics of underlying activity-travel behavior.
Alternative related studies have also been conducted. Gao and Liu [6]
attempted to examine the methods used to estimate traffic measures using infor-
mation from mobile phones, accounting for the fact that each vehicle likely con-
tains more than one phone because of the popularity of mobile phones. Steen-
bruggen et al. [8] used mobile phone data to provide new spatio-temporal tools
for improving urban planning and reducing inefficiencies in current urban sys-
tems. They addressed the applicability of such digital data to develop innovative
applications to improve urban management.
As described above, this study surveyed previous related research. Among
those studies, the present study could be characterized as similar to Ahas et
al. [2]. However, Ahas et al. [2] is based on results obtained by analyzing data
roaming activity. Mobile phone users in the study are obviously limited. There-
fore, whether the knowledge gained applies to the average traveler is not clear.
In the present study, I analyzed data provided by NTT DoCoMo, Inc., which is
the largest mobile phone service provider in Japan. Therefore, their data should
be more reliable in that the parameter is quite large.
4 Results
In general, the number of visitors has been increasing since the Hokuriku
Shinkansen was launched on May 14, 2015, with the exception of Nanao station.
Fig. 2. Visitor transitions at Kanazawa Fig. 3. Visitor transitions at Toyama

station station
Fig. 4. Visitor transitions at Ken- Fig. 5. Visitor transitions at Higashi

rokuen Chayagai
It should be noted that these “visitor” also include the residents living there,
because the data cannot exclude them. Of course, I tried to exclude residential
areas as much as possible when I specified the regional mesh codes. However, it
was rather difficult to do that, because the mesh codes are square-shaped.
First, I compared the results of two larger cities, Kanazawa and Toyama (see
Figs. 2 and 3). Both these cities have a station at which the Hokuriku Shinkansen
stops. It should be noted that Kanazawa city and Toyama attracted more visi-
tors in the afternoons, whereas Wakura and Wajima, which are located on the
Noto Peninsula, had more visitors in the mornings (8:00 a.m.–9:00 am). Visitors
to Toyama Station demonstrated approximately the same trend as those visit-
ing Kanazawa Station. However, there were fewer visitors on holidays than on
weekdays in Toyama.
I then examined the data of Kanazawa city (see Figs. 4 and 5). There were
three survey areas in this city: Kanazawa station, Kenrokuen Park and Higashi
Chayagai.
416 M. Yamamoto
Fig. 6. Visitor transitions at Wajima Fig. 7. Visitor transitions at Nanao

station
Fig. 8. Visitor transitions at Wakura Fig. 9. Visitor transitions at

hot springs Yamanaka hot springs
Wajima is famous for its Asaichi (morning market). Presumably, visitors

enjoy shopping at the Noto Shokusai market near Nanao Station during the
daytime, move on to Wakura hot springs later in the day, and then return to the
morning market the next day (see Figs. 6, 7 and 8).
Wajima is also the setting for a TV drama “Mare”, which was broad-
casted nationwide from April to September 2015. Visits to Wajima have slightly
increased in 2015. The increase in visits could be attributed to this TV drama
rather than the opening of the Hokuriku Shinkansen, because Nanao attracted
fewer visitors in 2015 than in 2014 in spite of being a better location and nearer
to Kanazawa.
Regarding Yamanaka hot springs, there was no significant difference in visits
between different periods. Some visitors might have spent more than one night
in Yamanaka hot springs (see Fig. 9).
Although both Wakura and Yamanaka hot springs are a little far from
Kanazawa, their results were contrary. The TV drama might have increased
the number of tourists in Wakura (Figs. 10 and 11).
70 70
60 60
50 50
40 40
30 30
20 20
Kenrokuen 15 Wajima 15
3,000 2,000 1,000 0 1,000 2,000 400 200 0 200 400

female male female male
Fig. 10. Visitors’ gender distribution at Kenrokuen and Wajima (12:00 a.m.–1:00 p.m.
on holidays in October 2015)
70 70
60 60
50 50
40 40
30 30
20 20
Wakura
15 Yamanaka
15
250 150 50 50 150 200 150 100 50 0 50 100 150

Fig. 11. Visitors’ gender distribution at Wakura hot springs and Yamanaka hot springs
(12:00 a.m.–1:00 p.m. on holidays in October 2015)
70 70
60 60
50 50
40 40
30 30
20 20
Kanazawa 15 15
Toyama
1,000 500 0 500 1,000 400 200 0 200 400

Fig. 12. Visitors’ gender distribution at Kanazawa station and Toyama Station (12:00
a.m.–1:00 p.m. on holidays in October 2015)
418 M. Yamamoto
Kanazawa city (particularly Kenrokuen) attracted a variety of visitors,

including many female visitors.
On the other hand, many elderly people (over 60 years old) visited the hot
springs and the Noto Peninsula. I presume that the local people account for a
large proportion of these visitors.
A comparison of visitors at Kanazawa Station with those at Toyama Station
found that Kanazawa Station attracted visitors from a wider area of Japan.
Mobile phone users around Kanazawa Station were from 235 municipalities,
including Ishikawa Prefecture, whereas those at Toyama Station were from 43
municipalities, including Toyama Prefecture (12:00 a.m.–1:00 p.m. on holidays
in October 2015) (Fig. 12).
Although the Hokuriku Shinkansen stops at both stations, the results suggest
that Kanazawa has been more successful so far in attracting visitors (see Table 2).
The number of visitors from Tokyo (gray column) increased in both the cities.
Table 2. Visitors’ residential distribution at Kanazawa Station and Toyama Sta-

tion(12:00 a.m.–1:00 p.m. on holidays in October 2014 and 2015)
Kanazawa station Toyama station

Residence 2015 2014 Residence 2015 2014
Toyama, Toyama Pref. 131 147 Kanazawa, Ishikawa Pref. 67 60
Takaoka, Toyama Pref. 93 94 Fukui, Fukui Pref. 18 13
Fukui, Fukui Pref. 91 80 Setagaya-ku, Tokyo 16 n/a
Myoko, Niigata Pref. 47 n/a Suginami-ku, Tokyo 15 n/a
Nanto, Toyama Pref. 34 31 Yamagata, Yamagata Pref. 14 n/a
Setagaya-ku, Tokyo 34 17 Takayama, Gifu Pref. 14 n/a
Imizu, Toyama Pref. 34 42 Nakamura-ku, Nagoya, Aichi Pref. 13 n/a
Oyabe, Toyama Pref. 33 39 Minami-ku, Niigata, Niigata Pref. 13 n/a
Sakai, Fukui Pref. 32 30 Gifu, Gifu Pref. 12 n/a
Omachi, Nagano Pref. 30 n/a Hakusan, Ishikawa Pref. 12 20
Ota-ku, Tokyo 28 12 Ota-ku, Tokyo 12 n/a
Tsubame, Niigata Pref. 28 n/a Himeji, Hyogo Pref. 12 n/a
Bunkyo-ku, Tokyo 27 n/a Nagano, Nagano Pref. 12 n/a
Nagano, Nagano Pref. 24 n/a Nerima-ku, Tokyo 11 n/a
Hiratsuka, Kanagawa Pref. 24 n/a Shinagawa-ku, Tokyo 11 n/a
Sanda, Hyogo Pref. 23 n/a Hida City, Gifu Pref. 11 10
Tonami, Toyama Pref. 23 26 Joetsu, Niigata Pref. 11 n/a
Suginami-ku, Tokyo 22 16 Kawaguchi, Saitama Pref. 10 n/a
Nerima-ku, Tokyo 22 10 Adachi-ku, Tokyo 10 n/a
Tsu, Mie Pref. 22 n/a Takatsuki, Osaka 10 n/a
Note: The figures above do not include visitors from Ishikawa Pref. for Kanazawa Station and
Toyama Pref. for Toyama Station. Numbers less than ten are represented as “n/a”
Despite the fact that Toyama is nearer to Tokyo than Kanazawa, the latter
successfully attracted more visitors from Tokyo.
Regarding the two prefectures, Ishikawa and Toyama, they both do have
many wonderful tourist attractions. However, as for the two cities, it seems that
Kanazawa is more attractive to tourists.
5 Conclusion
This study attempted to identify the number of visitors at two points in time
at various places in Japan and their characteristics using the location data of
mobile phone users collected by the mobile phone company.
As explained above, the opening of the Hokuriku Shinkansen increased the
number of visitors to many areas. However, it also led to fewer visitors in some
other areas. Its positive effect was remarkable in Kanazawa.
Numerous events have been recently held in Japan to attract visitors. In
addition to using the “MOBILE KUUKAN TOUKEIT M ”, combining it with
other ICT services, such as Google Trends, can help better predict the number
of visitors at new events. Specifically, by combining the “MOBILE KUUKAN
TOUKEIT M ” and the transition of the search results for a particular tourist des-
tination, it would be possible to more accurately predict the number of tourists.
If we can realize more accurate demand forecasting, it would be possible to
optimize the necessary goods and number of non-regular employees in advance.
Moreover, understanding consumers’ characteristics beforehand could enable us
to optimize the services, which could influence customer satisfaction.
Acknowledgements. This work was supported by JSPS KAKENHI Grant Number

JP15K01970. The “MOBILE KUUKAN TOUKEIT M ” and logo are trademarks of
NTT DOCOMO, Inc.
References
1. Agency JT (2014) Keitaidenwa kara erareru ichijouhou tou wo katuyousita hounichi
gaikokujin doutaichousa houkokusho [foreign visitors dynamics research report uti-
lizing mobile phone location information]. Technical report, Japan Tourism Agency
2. Ahas R, Aasa A et al (2008) Evaluating passive mobile positioning data for tourism
surveys: an estonian case study. Tourism Manage 29(3):469–486
3. Bello-Orgaz G, Jung JJ, Camacho D (2016) Social big data: recent achievements
and new challenges. Inf Fusion 28:45–59
4. Dolnicar S, Ring A (2014) Tourism marketing research: past, present and future.
Ann Tourism Res 47:31–47
5. Fuchs M, Höpken W, Lexhagen M (2014) Big data analytics for knowledge gener-
ation in tourism destinations - a case from Sweden. J Destination Mark Manage
3(4):198–209
6. Gao H, Liu F (2013) Estimating freeway traffic measures from mobile phone location
data. Eur J Oper Res 229(1):252–260
7. Liu F, Janssens D et al (2013) Annotating mobile phone location data with activity
purposes using machine learning algorithms. Expert Syst Appl 40(8):3299–3311
420 M. Yamamoto
8. Steenbruggen J, Tranos E, Nijkamp P (2015) Data from mobile phone operators: a

tool for smarter cities? Telecommun Policy 39(3–4):335–346
9. Xiang Z (2015) What can big data and text analytics tell us about hotel guest
experience and satisfaction? Int J Hosp Manage 44(44):120–130
Research on Geo/Geo/1 Retrial Queue
with Working Vacation Interruption
and Nonpersistent Customers
Mingcong Wu1 , Yong Huang1(B) , Yang Song2 , Liang Zhao1 , and Jian Liu1
1
huangyong scu@aliyun.com
2
Department of Mathematics, Nanjing University of Aeronautics and Astronautics,
Nanjing 210016, People’s Republic of China
Abstract. Queueing phenomenon is prevalent in all areas of society and

a reasonable queueing design can improve service efficiency and customer
satisfaction. In order to adapt to the new requirements of web service sys-
tem, this paper studies a Geo/Geo/1 retrial queue with working vacation
interruption and nonpersistent customers. Firstly, a series of assumptions
about the queueing system are put forward and the corresponding tran-
sition probability matrix is obtained. Then the stationary condition of
the queueing system is derived. After that, the stationary distribution
and the performance measures are obtained by using the matrix-analytic
method. Finally, numerical analysis is conducted to discuss the effect of
parameters on performance measures, furthermore, the performance of
the queueing model is optimized to obtain the best parameters and the
minimum cost.
Keywords: Web service · Discrete-time queueing system · Retrial ·

Working vacation interruption · Nonpersistent customers
1 Introduction
With the development of the internet, the internet-based web service industry
is growing explosively, and web service of higher quality is required. The basic
application structure of web service is shown in Fig. 1. In the system, users
can request services according to their own needs through the internet. After
their service requests are passed to web server, the web server will interact with
application server and database server for different requests, and process various
requests properly. Finally, the results of the requests will be returned to the users.
The function of the application server is to guarantee the business logic in the
activities and coordinates the information exchange among the users. The task
of the database server is to complete query, storage, modify and other functions
of web service data in the database.

422 M. Wu et al.
Web service system needs to use the adapted queueing model in dealing with
the order of the users, but the traditional queueing model cannot accurately deal
with this scene. When the arriving user finds the server is temporarily unable to
provide service, he will leave the server and repeat requests after a while, which is
called retrial strategy. Considering the utilization of server, managers often allow
server to serve at a relatively low rate when there are no users in the system,
but when the number of the users requesting service increases in the system, the
service rate needs to return to a higher level immediately, which is called work-
ing vacation interruption strategy. Besides, users who conduct business on the
internet tend to give up receiving service and leave the service system because
they don’t want to wait too long, which is called nonpersistent customers strat-
egy. Scholars have studied the above three cases separately, but there are no
studies of the queueing model which possess these three strategies at the same
time. This paper tries to study the combination of these three strategies, and
solve the construction of the queueing model of web service system under these
conditions.
This paper is organized as follows. In Sect. 1, relevant literatures are reviewed.
Section 2 is dedicated to formulate a queueing model to solve the problem men-
tioned above. In Sect. 3, a series of numerical analyses are presented and the
performance of the model is optimized. Section 4 gives some conclusions.
2 Literature Review
Queueing theory studies the working process of random service system. It orig-
inated from the telephone conversation in the early 20th century. In 1909–1920,
Erlang, the Danish mathematician and electrical engineer, studied the problem
of telephone conversation using the method of probability theory, which created
this subject of applied mathematics. Queueing system is divided into continuous-
time queueing system and discrete-time queueing system. Compared with the
former, the research of discrete-time queueing system started relatively late and
Fig. 1. The basic application structure of web service

Research on Geo/Geo/1 Retrial Queue 423
the early research results about the discrete-time queueing system are few. But
with the development of computer and communication in which time is slotted,
the discrete-time queueing system has become one of the research hotspots of
queueing theory in recent years. Meisling [10] made a pioneering work on the
discrete-time queueing system, and then a series of studies have deepened and
developed the analysis of this classical model. Kobayash and Konheim [3] stud-
ied the application of the discrete-time queueing system in the field of computer
communication network, and pointed out that the discrete-time queueing sys-
tem is more suitable for the modeling and analysis of computer networks, which
has played a positive role in the research and application of the discrete-time
queueing system.
Retrial queueing system is an important part of queueing theory research.
The characteristic of retrial queueing system is that customer who requests ser-
vice will enter a queue called orbit and retry after a random time when he finds
all servers are busy. Kosten [4] first proposed a retry service system. Yang and
Li [18] were the first to study a discrete-time retrial queue, they studied the
steady-state queue size distribution of the discrete-time Geo/G/1 retrial queue.
Then retrial queueing system is widely used in telephone switching systems,
computer networks and communications networks. So far, these research areas
are still vibrant. Some of the latest developments regarding retrial queueing
systems are Dimitriou [2], Rajadurai [13] and so on. If the server is occupied,
the customer retrying from the orbit may leave the queueing system or return
to the orbit, which is called nonpersistent customers phenomenon. Palm [11]
first proposed the issue of nonpersistent customers. Then, many scholars intro-
duced nonpersistent customers into various queueing systems. Liu and Song [9]
introduced nonpersistent customers to a Geo/Geo/1 retrial queue with work-
ing vacations. Phung-Duc [12] introduced two types of nonpersistent customers
into a M/M/c/K (K ≥ c ≥ 1) retrial queues. These models are also widely
used in real life. In fact, the introduction of nonpersistent customers is of great
significance to imitate the impatience characteristic of customers in reality.
The queueing systems with vacation and working vacation have always been
a research hotspot. Queueing system with vacation is an extension of the classi-
cal queueing system, which allows server not to serve customers at certain time
called vacation. For more information about queueing system with vacation,
see Tian and Zhe [16]. If server does not completely stop serving on vacation,
but rather serve at a slower service rate than the rate of regular busy period,
this vacation policy is called working vacation. Servi and Finn [14] first intro-
duced the queueing system with working vacation when they studied the model
of communication network. They introduced the working vacation mechanism
into the M/M/1 queue. Then, many scholars introduced working vacation into
various queueing system. Tian et al. [17] studied a discrete-time queueing sys-
tem with working vacation. Li and Tian [7] studied a discrete-time Geo/Geo/1
queue with single working vacation. On the basis of working vacations, working
vacation interruption strategy was introduced, which means that once a certain
indicator (such as the number of customers in the system) reaches a certain
424 M. Wu et al.
value on working vacation, server can stop working vacation and return to nor-
mal service level immediately. Li and Tian [6] first proposed working vacation
interruption strategy. They introduced working vacation and vacation interrup-
tion into M/M/1 queue, then they studied the corresponding discrete-time queue
[5]. Recently, studies on working vacation interruption are continued. Li et al. [8]
introduced working vacation interruption into discrete-time Geo/Geo/1 retrial
queue, Gao et al. [15] studied an M/G/1 queue with single working vacation and
vacation interruption under Bernoulli schedule.
Previous studies can’t address the real-world scenes of web service that are
mentioned. So, this paper considers introducing working vacation interruption
and nonpersistent customers to discrete-time Geo/Geo/1 retrial queue to provide
the solution.
3 Method
In order to deal with retrial, working vacation interruption and nonpersistent
customers phenomena in web service, this section constructs an adapted queue-
ing model and studies the queueing system. Section 3.1 put forwards a series of
assumptions about the queueing model (for convenience of study, assume that
the number of server is one) and obtains the corresponding transition probabil-
ity matrix for subsequent derivation. The research of queueing system needs to
be based on the stability of the system, so Sect. 3.2 derives the condition under
which the system is stable, and the rate matrix R for the derivation of stationary
distribution. The main purpose of the research of queueing system is to study
the states and the indexes when the system is stable, so Sect. 3.3 derives the
stationary distribution and some performance measures using matrix-analytic
method.
3.1 Model
The Geo/Geo/1 retrial queue with working vacation interruption and nonper-
sistent customers is assumed as follows:
First of all, the rules of service and customer arrival are assumed. Assume the
beginning and ending of service occur at slot division point t = n, n = 0, 1, · · · .
The service time Sb in a regular busy period follows a geometric distribution with
parameter μb . The service time Sv in a working vacation period follows a geomet-
ric distribution with parameter μv . Note that 0 < μv < μb < 1. Assume customer
arrivals occur at the slot t = n, n = 0, 1, · · · . Inter-arrival time, which is an inde-
pendent and identically distributed sequence, follows a geometric distribution
with parameter during a regular busy period and a geometric distribution with
parameter λb during a working vacation period. Also Note that 0 < λv < λb < 1.
Then, assume the rules of retrial and nonperisistent customers strategies.
Customer from the orbit of infinite size requests retrial at the slot t = n, n =
0, 1, · · · . And the time between two successive retrials follows a geometric dis-
tribution with parameter α. If the customer arrival and retrial occur at the
same instant when the server is not occupied, we assume that the arriving cus-
tomer receives the service. And, due to the introduction of nonpersistent cus-
tomers strategy, suppose the probability that the retrial customer leaves system
is q(0 < q < 1), and the probability that retrial customer returns to the orbit is
1 − q.
Finally, working vacation and vacation interruption are assumed. The server
begins a working vacation each time when there is no customer in the system,
i.e., the server is free and there is no customer in the orbit. Assume the beginning
and ending of working vacation occur at the slot t = n+ , n = 0, 1, · · · . At the
end of each working vacation, the server begins a new vacation only if there
is no customer in the system. And the vacation time V follows a geometric
distribution with parameter θ, (0 < θ < 1). If the service is complete at slot
t = n− and there is customer in the system at slot t = n+ , the server will stop
the working vacation immediately and return to the normal service level, i.e.,
working vacation interruption strategy.
In this paper, for any real number x ∈ [0, 1], denote x = 1 − x. Assume
that interarrival time, service time, and working vacation time are mutually
independent.
After completing the above assumptions, the occurring order of the random
events can be described in Fig. 2 using number axis.
Fig. 2. The occurring order of the random events
Meanwhile, the structure of the queueing model in web service can be shown
in Fig. 3. User accesses the web server through the internet. He joins the orbit
if the web server is busy, or else he receives service immediately and leaves the
system after the service is completed. The retrial user from the orbit receives
service immediately if the web server is free, otherwise he may leave the system
or come back to the orbit.
Let Qn be the number of customers in the orbit at the slot n+ , and Jn the
state of server at the slot n+ . There are four possible states of the server as
follows:
(1) The state Jn = 0 denotes the server is free in a working vacation period at
the slot n+ .
(2) The state Jn = 1 denotes the server is busy in a working vacation period at
the slot n+ .
(3) The state Jn = 2 denotes the server is free in a regular busy period at the
slot n+ .
426 M. Wu et al.
Fig. 3. The structure of the queueing model in web service
(4) The state Jn = 3 denotes the server is busy in a regular busy period at the
slot n+ .
Then, {Qn , Jn } , n ≥ 0, is a Markov chain with state space Ω = {(k, j), k ≥ 0,

j = 0, 1, 2, 3}. Note that, the probability that the server is free in a regular busy
period is zero when there is no customer in the orbit.
Using the lexicographical sequence for the states, the transition probability
matrix of {Qn , Jn } can be written as:
⎛ ⎞
B1 B0
⎜ A2 A1 A0 ⎟
⎜ ⎟
P =⎜ A2 A1 A0 ⎟,
⎝ ⎠
.. .. ..
. . .
where
⎛ ⎞
⎛ ⎞ − −
0 0 0 0
⎜ λv λv θ 0 λv θ ⎟
⎜ − − − ⎟ ⎜ − − − − ⎟
⎜ 0 μv λv θ 0 μv λv θ ⎟ ⎜ μv λv μ−v λv −θ μ λ μ λ θ ⎟
B0 = ⎜ ⎟ , B 1 = ⎜ 0 v v + v v ⎟,
⎝0 0 0 0 ⎠ ⎜ 1 0 0 0 ⎟
− ⎝ ⎠
− − −
0 0 0 μb λb
μb λb 0 0 μb λb + μb λb
⎛ ⎞
0 0 0 0
⎜ − − − − − − − ⎟
⎜ 0 μv λv (α +α q ) θ 0 μv λv (α +α q )θ ⎟
A0 = ⎜ ⎟,
⎝0 0 0 0 ⎠
− − −
0 0 0 μb λb (α +α q )
⎛
− −− − −
⎞
−− −− − −
⎜ λv α θ λv α q θ +λv α θ λv α θ λv α θ + λv α q θ ⎟
⎜ − − − − − − − − − − − − − − ⎟
⎜ 0 μv λv (α +α q ) θ + μv λv αq θ μv λv α (μv λv + μv λv θ)(α +α q ) + μv λv αqθ ⎟
⎜
A1 = ⎜ ⎟,
− − − − ⎟
⎜ 0 0 λb α λb (α +α q ) ⎟
⎝ ⎠
− − − − − − −
0 0 μb λb α (μb λb + μb λb )(α +α q ) + μb λb αq
⎛ ⎞
− − − −
⎜0 λv αq θ + λv α θ 0 λv αqθ + λv αθ ⎟
⎜ − − − − − − ⎟
⎜0 μv λv αq θ 0 (μv λv θ + μv λv )αq + μv λv α⎟
A2 = ⎜
⎜ −
⎟.
⎟
⎜0 0 0 λb αq + λb α ⎟
⎝ ⎠
− − −
0 0 0 (μb λb +μb λb )αq + μb λb α
It can be seen from the block structure of the transition probability matrix
that is a quasi-birth and death (QBD) process.
3.2 Stability Condition and Rate Matrix R

In this section, after using the matrices A0 , A1 and A2 to get the matrix A,
derive the stability condition of the queueing system using the matrix A.
Let A = A0 + A1 + A2 , then,
⎛ ⎞
− −− − −− − − −
⎜ λv α θ α θ +λv α θ λv α θ αθ + λv α θ ⎟
⎜ − − − − − − ⎟
⎜ 0 μ v θ μ λ α μ θ + μ λ + μ λ α ⎟
A=⎜ ⎜ v v v v v v v ⎟
− − − ⎟.
⎜ 0 0 λ α λ + λ α ⎟
⎝ b
− −
b b ⎠ −
−
0 0 μb λb α μb +μb λb + μb λb α
Theorem 1. The QBD process {Qn , Jn } is positive recurrent (stable) if and

− − − −
only if μb (λb α −1)(αq − λb ) < μb (λb +λb q)α.
Proof. Since A is reducible, according to Theorem 1 in Latouche and Ramaswami
[1], know that the QBD process is positive recurrent if and only if
⎛ ⎞
−
0 λ αq + λ α 0 0
v⎝ −
b
−
b ⎠e > v − − − e,
−
0 μb λb αq + μb λb α + μb λb αq 0 μb λb (α +α q )
where v is the stationary probability vector of

⎛ ⎞
− − −
⎝ λb α − − −
λb + λb α
−
⎠,
μb λb α μb +μb λb + μb λb α
e is a column vector with all elements equal to one. By solving
⎧ ⎛ ⎞
− − −
⎪
⎪
⎨ ⎝ λb α λb + λb α ⎠=v
v − − − − ,
⎪
⎪ α
μb λb μb +μb λb + μb λb α
⎩
ve = 1
get
− − − −
μb λb α 1 − λb α
v=( − − − −
, − − − −
).
1 + μb λb α − λb α 1 + μb λb α − λb α
428 M. Wu et al.
Then take v into the inequation above, after a series of algebraic operations, the
− − −
QBD process {Qn , Jn } is positive recurrent if and only if μb (λb α −1)(αq − λb ) <
−
μb (λb +λb q)α, i.e., the stability condition of the queueing system.
Next, before deriving the stationary distribution of the system, need to obtain
the minimal non-negative solution R which satisfies the matrix quadratic equa-
tion
R = R2 A2 + RA1 + A0 . (1)
− − − −
Theorem 2. If μb (λb α −1)(αq−λb ) < μb (λb +λb q)α, the matrix equation above
has the minimal non-negative solution
⎛ ⎞
0 0 0 0
⎜ 0 r1 r2 r3 ⎟
R=⎜ ⎝0 0 0 0 ⎠,
⎟
0 0 r4 r5
where
− − − − − − − √ − − − −
1 − μv λv (α +α q ) θ − μv λv αq θ − D μv λv α r1 + μb λb α r3
r1 = , r2 = ,
− − − − −
2 μv λv αq θ 1− λb α
− − −− −
M r1 2 + N r1 + λv (α +α q )θ(1 − λb α)
μv
r3 = − − −
,
(1 − λb α)[1 − L − F (r1 + r5 ) − (λb αq + λb α)r4 ] −G
− − − − −
μb μb λb λb (α +α q ) α
r4 = ,
− − − − − −
μb λb αq − μb λb α α q + μb λb α + μb λb αq
− − − − −
μb λb (α +α q )(1 − λb α)
r5 = ,
− − − − − −
μb λb αq − μb λb α α q + μb λb α + μb λb αq
2
D = [μv λv (α + αq)θ + μv λv αqθ − 1]2 − 4μv 2 λv λv αq(α + αq)θ ,
− − − − − − − −
M = (μv λv αqθ + μv λv α + μv λv αq)(1 − λb α) + μv λv α(λb αq + λb α),
− − − − − − − − − − −
N = [(μv λv + μv λv θ)(α +α q ) + μv λv αqθ](1 − λb α) + μv λv λb (α +α q ) α ,
− − − − −
L = (μb λb + μb λb )(α +α q ) + μb λb αq,
− − −
F = μb λb αq + μb λb α + μb λb αq,
− − − − − − −
G = μb λb α r1 (λb αq + λb α) + μb λb λb (α +α q ) α .

R11 R12
Proof. Notice the structure of A0 , A1 and A2 , assume R = , where
0 R22
R11 , R12 and R22 are all 2 × 2 matrices. Taking R into Eq. (1), have
⎧ ⎛ ⎞ ⎛− ⎞
−
⎪
⎪
− −−
α
−−
q
−−
α
⎪
⎪ 0 λ αq θ + λ αθ ⎠ + R11 ⎝ v θ λ λ α θ +λ θ
R11 = R11 2 ⎝ ⎠
v v v v
⎪
⎪ − − − − −
⎪
⎪
−
α
− −
q
− −
⎪
⎪ 0 μ λ
v v αq θ 0 μ λ
v v ( +α ) θ + μ λ
v v αq θ
⎪
⎪
⎪
⎪ 0 0
⎪
⎪ +
⎪
⎪ −
0 μv λv ⎛
− − −
(α +α q ) θ
⎪
⎪ ⎞
⎪
⎪ −
⎪
⎪
⎪
⎪ 0 λ αqθ + λ αθ
R12 = R11 2 ⎝ ⎠
v v
⎪
⎪ − − −
⎪
⎪
⎪
⎪ 0 μ v λv αqθ
⎛ + μ v λ v α + μ v λ v αq ⎞
⎪
⎪ −
⎪
⎪
⎪
⎪ 0 λ αq + λ α
⎪ + (R11 R12 + R12 R22 ) ⎝ − − ⎠
b b
⎪
⎪ −
⎪
⎪ 0 μ λ αq + μ λ α + μ λ αq
⎪
⎪ ⎛ − b b b b b b ⎞
⎪
⎨ − − −
λ α θ λ α θ + λ α q θ
+ R11 ⎝ ⎠
v v v
− − − −
⎪
⎪ − −
q
−
⎪
⎪ μ
⎛ v− v λ α (μ λ
v v + μ λ
v v )( α +α ) + μ λ
v v ⎞αqθ
⎪
⎪
⎪
⎪ − − −
q
⎪
⎪ ⎝ b−λ α λ b (α +α ) ⎠
⎪ + R12
⎪
⎪ − − − − − −
⎪
⎪ μ λ α (μ λ + μ λ )( α +α q ) + μ λ αq
⎪
⎪
b b b b b b b b
⎪
⎪ 0 0
⎪
⎪
⎪
⎪ + − − −
⎪
⎪ 0 μv λv ⎛ (α +α q )θ
⎪
⎪ ⎞
⎪
⎪ −
⎪
⎪ 0 λ αq + λ α 0 0
⎪
⎪ R22 = R22 2 ⎝
b b ⎠+
⎪
⎪ − − −
μ
−
λ α
− −
+α q )
⎪
⎪ 0 μb λb αq + μb λb α + μb λb αq 0 b b (
⎪
⎪ ⎛ ⎞
⎪
⎪ − − − −
⎪
⎪ λb α λb (α +α q )
⎪
⎪ + R22 ⎝ ⎠
⎪
⎩ − − − − − − −
μb λb α (μb λb + μb λb )(α +α q ) + μb λb αq

0 0 0 0
From the first and third equations, getR11 = and R22 =
0 r1 r4 r5
respectively.
Then,
taking R11 and R22 into the second equation, finally obtain
0 0
R12 = are as shown in Theorem 2.
r2 r3
3.3 Stationary Distribution and Performance Measures

− − − −
If μb (λb α −1)(αq − λb ) < μb (λb +λb q)α, Let {Q, J} be the stationary limit of
the {Qn , Jn }. And denote
πk = (πk0 , πk1 , πk2 , πk3 ), k ≥ 0,

πkj = P {Q = k, J = j} = lim P {Qn = k, Jn = j}, (k, j) ∈ Ω.
n→∞
Note that, π02 = 0.

430 M. Wu et al.
− − − −
Theorem 3. If μb (λb α −1)(αq − λb ) < μb (λb +λb q)α, the stationary distribu-
tion of Q, J is given by
⎧
⎪
⎪ πk0 = 0, k≥1
⎨ π = π r k, k≥0
k1 01 1
k−1 r3 r4 k−1 k−1 k−1 (2)
⎪
⎪ π k2 = π 01 (r2 r1 + r5 −r1 (r5 − r1 )) + π 03 r4 r5 , k≥1
⎩ r3 k k k
πk3 = π01 r5 −r1 (r5 − r1 ) + π03 r5 , k ≥ 0,
where
−1
1 + r2 r3 (1 + r4 ) 1 + r4
π00 = 1 + ( + )Y + U ,
1 − r1 (1 − r1 )(1 − r5 ) 1 − r5
− − − − − −
λv θ λv (1 − λv θ ) − μv λv λv αq θ r1
Y = ,U = ,
− − − − − − − − − − − − −
1 − μv λv θ − μv λv αq θ r1 μb λb (1 − μv λv θ − μv λv αq θ r1 )
− − − − − −
λv θ λv (1 − λv θ ) − μv λv λv αq θ r1
π01 = π00 , π03 = π00 .
− − − − − − − − − − − − −
1− μv λv θ − μv λv αq θ r1 μb λb (1 − μv λv θ − μv λv αq θ r1 )
Proof. Using the matrix-analytic method [1], have
πk = (πk0 , πk1 , πk2 , πk3 ) = π0 Rk = (π00 , π01 , 0, π03 )Rk , k ≥ 0. (3)
Taking
⎛ ⎞
0 0 0 0
⎜ 0 r k
r r k−1
+ r3 r4 k−1
− r1 k−1 ) r3 k
− r1 k ) ⎟
Rk = ⎜
1 2 1 r5 −r1 (r5 r5 −r1 (r5 ⎟,k ≥ 1
⎝0 0 0 0 ⎠
0 0 r4 r5 k−1 r5 k
into Eq. (3), then Eq. (2) is obtained. Meanwhile, π0 satisfies the following two
equations
π0 (B1 + RA2 ) = π0 , (4)

−1
π0 (I − R) e = 1. (5)
From Eq. (4), get π01 and π03 as shown in Theorem 3. Then, taking π01 , π02 = 0
and π03 into Eq. (5), obtain π00 as shown in Theorem 3. So far, Theorem 3 is
proved.
Using the stationary distribution derived above, some performance measures
can be obtained as follows.
(1) The probability that the server is busy
∞
∞
1 − r5 + r3 1
Pb = πk1 + πk3 = π01 + π03 .
(1 − r1 )(1 − r5 ) 1 − r5
k=0 k=0
(2) The probability that the server is free

∞
∞
r2 − r2 r5 + r3 r4 r4
Pf = πk0 + πk2 = π00 + π01 + π03 .
(1 − r1 )(1 − r5 ) 1 − r5
k=0 k=1
(3) The probability that the server is in a regular busy period

∞
∞
r2 + r3 − r2 r5 + r3 r4 1 + r4
Pr = πk2 + πk3 = π01 + π03 .
(1 − r1 )(1 − r5 ) 1 − r5
k=0 k=0
(4) The probability that the server is in a regular busy period

∞
∞
1
Pv = πk0 + πk1 = π00 + π01 .
1 − r1
k=0 k=0
(5) The mean number of customers in the orbit

∞

E[L] = k(πk1 + πk2 + πk3 )
k=1
2
(r1 + r2 )(1 − r5 ) + r3 r4 (2 − r1 − r5 ) + r3 (1 − r1 r5 ) r4 + r5
= π01 2 2 + π03 2.
(1 − r1 ) (1 − r5 ) (1 − r5 )
(6) The mean number of customers in the system

∞
∞

∼
E[L] = kπk2 + (k + 1)(πk1 + πk3 ) = E[L] + Pb .
k=1 k=0
3.4 Empirical Research

This section provides some numerical examples to illustrate the effect of varying
parameters on the performance measures of the system. Assume λv equals 0.3,
− − −
λb equals 0.6 and μb equals 0.8. Under the condition that μb (λb α −1)(αq −λb ) <
−
μb (λb +λb q)α, plot Figs. 3, 4, 5, 6, 7, 8, 9 and 10, which illustrate the effect of
vacation period rate θ, retrial rate α, departure rate q and service rate μv on
the performance measures Pb and E[L].
Remark 1. Considering the constraint of the stationary condition and the display
of the graph, the values of the parameters are the most suitable.
(1) The numerical analysis of Pb

Figures 4, 5, 6 and 7 illustrate the effect of parameters on the probability that the
server is busy. The effect and the reasons which cause the effect are discussed.
Figure 4 illustrates the effect of θ on the probability that the server is busy.
Figure 4(a) is for different α, Fig. 4(b) is for different q and Fig. 4(c) is for differ-
ent μv . Obviously, whether in Fig. 4(a), (b) or (c), the probability Pb decreases
432 M. Wu et al.
0.64 0.65
0.65
=0.5 q=0.1 v=0.2
0.64 0.645
=0.7 0.62 q=0.3 v=0.3
0.63 =0.9 q=0.5 0.64 v=0.4
0.6 0.635
0.62
0.61 0.63
0.58
Pb
Pb
Pb
0.6 0.625
0.56
0.59 0.62
0.58 0.54 0.615
0.57 0.61
0.52
0.56 0.605
0.55 0.5 0.6

0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95
,when v=0.3, b=0.6, v=0.4, b=0.8,q=0.1 ,when v=0.3, b=0.6, v=0.2, b=0.8, =0.7 ,when v=0.3, b=0.6, b=0.8, =0.5,q=0.1
(a) (b) (c)
Fig. 4. (a) The probability that the server is busy versus θ for different α (b) The
probability that the server is busy versus θ for different q (c) The probability that the
server is busy versus θ for different μv
0.75 0.64 0.66
q=0.5 0.63
=0.6 v=0.2
q=0.7 =0.7 0.64 v=0.4
0.7
q=0.9 0.62 =0.9 v=0.6
0.62
0.61
0.65
0.6 0.6
Pb
Pb
Pb
0.6 0.59
0.58
0.58
0.55 0.56
0.57
0.56 0.54
0.5
0.55
0.52
0.45 0.54
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.4 0.5 0.6 0.7 0.8 0.9 1 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
,when v=0.3, b=0.6, v=0.4, b=0.8, =0.7 ,when v=0.3, b=0.6, v=0.2, b=0.8,q=0.2 ,when v=0.3, b=0.6, b=0.8, =0.5,q=0.2
(d) (e) (f)
Fig. 5. (d) The probability that the server is busy versus α for different q (e) The
probability that the server is busy versus α for different θ (f) The probability that the
server is busy versus α for different μv
0.7 0.7 0.62
0.68
=0.4 v=0.2 =0.6
0.68 0.61
=0.6 v=0.5 =0.8
0.66 =0.8 0.66 v=0.7 0.6 =0.9
0.64 0.64
0.59
0.62 0.62
0.58
Pb
Pb
0.6
Pb
0.6
0.57
0.58 0.58
0.56
0.56 0.56
0.54 0.54 0.55
0.52 0.52 0.54
0.5 0.5 0.53

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.1 0.15 0.2 0.25 0.3 0.35 0.4
q,when v=0.3, b=0.6, v=0.2, b=0.8, =0.7 q,when v=0.3, b=0.6, b=0.8, =0.7, =0.4 q,when v=0.3, b=0.6, v=0.2, b=0.8, =0.7
(g) (h) (i)
Fig. 6. (g) The probability that the server is busy versus q for different α (h) The
probability that the server is busy versus q for different μv (i) The probability that the
server is busy versus q for different θ
gradually with the rate θ increasing regardless of the change of α, q and μv . What
causes this trend is the fact that the average service rate becomes smaller with
θ increasing. Note that the effect of θ on Pb becomes weaker with μv increasing.
Meanwhile, because Pb + Pf = 1, the probability that the sever is free Pf should
show the opposite trend of change (the same below).
Figure 5 illustrates the effect of α on the probability that the server is busy.
Figure 5(d) is for different q, Fig. 5(e) is for different θ and Fig. 5(f) is for
0.6 0.58 0.63
=0.4 0.57
q=0.5 =0.6
0.625
0.58 =0.6 q=0.7 =0.7
=0.8 0.56 q=0.9 0.62 =0.8
0.56 0.55 0.615
0.54 0.61
0.54
Pb
Pb
0.53
Pb
0.605
0.52
0.52 0.6
0.5 0.51 0.595
0.5 0.59
0.48
0.49 0.585
0.46 0.48 0.58

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35
v,when v=0.3, b=0.6, b=0.8, =0.7,q=0.5 v,when v=0.3, b=0.6, b=0.8, =0.7, =0.4 v,when v=0.3, b=0.6, b=0.8, =0.8,q=0.1
(j) (k) (l)
Fig. 7. (j) The probability that the server is busy versus μv for different α (k) The
probability that the server is busy versus μv for different q (l) The probability that the
server is busy versus μv for different θ
different μv . Among these three figures, Pb decreases with the rate α increas-
ing regardless of the change of α, q and μv . This is because the number of the
customers who give up receiving service increases with μv increasing.
Figure 6 illustrates the effect of q on the probability that the server is busy.
Figure 6(g) is for different α, Fig. 6(h) is for different μv and Fig. 6(i) is for
different θ. Obviously, Pb decreases with the rate q increasing among these three
figures. This is because the number of the customers who leave the system due
to the failure of retrial increases with q increasing.
Figure 7 illustrates the effect of v on the probability that the server is busy.
Figure 7(j) is for different α, Fig. 7(k) is for different q and Fig. 7(l) is for different
θ. Obviously, Pb decreases with the rate μv increasing. This is because the average
service rate increases with μv increasing. From the change of the other three
parameters, this trend is consistent. Besides, the effect of μv on Pb also becomes
weaker with θ increasing.
(2) The numerical analysis of E[L]
Figures 8, 9, 10 and 11 illustrate the effect of parameters on the mean number
of customers in the orbit. The effect and the reasons that cause the effect are
also discussed.
0.3 0.65 0.45
=0.5 q=0.2 v=0.2

0.6
=0.7 q=0.4 v=0.3
0.25 =0.9 0.55 q=0.6 0.4 v=0.4
0.5
0.2 0.45 0.35

E[L]
E[L]
E[L]
0.4
0.15 0.35 0.3
0.3
0.1 0.25 0.25
0.2
0.05 0.2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
,when v=0.3, b=0.6, v=0.4, b=0.8,q=0.5 ,when v=0.3, b=0.6, v=0.4, b=0.8, =0.5 ,when v=0.3, b=0.6, b=0.8,q=0.5, =0.5
(m) (n) (o)
Fig. 8. (m) The mean number of customers in the orbit versus θ for different α (n) The
mean number of customers in the orbit versus θ for different q (o) The mean number
of customers in the orbit versus θ for different μv
434 M. Wu et al.
2.5 1 2.5
q=0.4 0.9
=0.1 v=0
q=0.6 =0.5 v=0.4
2 q=0.8 0.8 =0.9 2 v=0.8
0.7
1.5 0.6 1.5
E[L]
E[L]
E[L]
0.5
1 0.4 1
0.3
0.5 0.2 0.5
0.1
0 0 0
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
,when v=0.3, b=0.6, v=0.4, b=0.8, =0.6 ,when v=0.3, b=0.6, v=0.4, b=0.8,q=0.4 ,when v=0.3, b=0.6, b=0.8, =0.6,q=0.4
(p) (q) (r)
Fig. 9. (p) The mean number of customers in the orbit versus α for different q (q) The
mean number of customers in the orbit versus α for different θ (r) The mean number
of customers in the orbit versus α for different μv
0.3 1
1.4 q=0.5 =0.3
0.95
=0.5 q=0.7 =0.5
1.2 =0.7 0.25 q=0.9 0.9 =0.7
=0.9 0.85
1
0.2 0.8
0.8
E[L]
E[L]
0.75
E[L]
0.6 0.15 0.7
0.65
0.4
0.1 0.6
0.2 0.55
0 0.05 0.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
v,when v=0.3, b=0.6, b=0.8, =0.3,q=0.1 v,when v=0.3, b=0.6, b=0.8, =0.6, =0.5 v,when v=0.3, b=0.6, b=0.8, =0.6,q=0.1
(s) (t) (u)
Fig. 10. (s) The mean number of customers in the orbit versus μv for different α (t)
The mean number of customers in the orbit versus μv for different q (u) The mean
number of customers in the orbit versus μv for different θ
Figure 8 illustrates the effect of θ on the mean number of customers in the

orbit. Figure 8(m) is for different α, Fig. 8(n) is for different q and Fig. 8(o) is
for different μv . With the rate θ increasing, the time during which the sever is in
a working vacation period is reduced, then the average service decreases, which
leads to the decreasing trend of the mean number of customers E[L] as shown in
the figures. It is worth noting that the effect of θ on E[L] becomes weaker with
μv increasing, which is consistent with the effect of θ on Pb in the Fig. 4(c).
Figure 9 illustrates the effect of α on the mean number of customers in the
orbit. Figure 9(p) is for different q, Fig. 9(q) is for different θ, Fig. 9(r) is for
different μv . For the same q, θ and μv , E[L] decreases with α increasing. It
is because the number of the customers who leave the orbit due to success or
failure of retrial increase with the retrial rate α increasing. Furthermore, with α
increasing, the effect of it on E[L] becomes weaker.
Figure 10 illustrates the effect of μv on the mean number of customers in
the orbit. Figure 10(s) is for different α, Fig. 10(t) is for different qand Fig. 10(u)
is for different θ. Obviously, no matter how α, q and θ change, E[L] decrease
with μv increasing. This is because the average service rate increases with μv
increasing. From Fig. 10(u), we can see that the effect of μv on E[L] becomes
1.4 1.4 0.5

=0.5 v=0.4 =0.3
0.45
1.2 =0.7 1.2 v=0.6 =0.5
=0.9 v=0.8 0.4 =0.7
1 1 0.35
0.3
0.8 0.8
E[L]
E[L]
E[L]
0.25
0.6 0.6
0.2
0.4 0.4 0.15
0.1
0.2 0.2
0.05
0 0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
q,when v=0.3, b=0.6, v=0.4, b=0.8, =0.7 q,when v=0.3, b=0.6, b=0.8, =0.5, =0.7 q,when v=0.3, b=0.6, v=0.4, b=0.8, =0.7
(v) (w) (x)
Fig. 11. (v) The mean number of customers in the orbit versus q for different α (w)
The mean number of customers in the orbit versus q for different μv (x) The mean
number of customers in the orbit versus q for different θ
smaller with θ increasing, which is consistent with the effect of μv on Pb in the

Fig. 7(l).
Figure 11 illustrates the effect of q on the mean number of customers in the
orbit. Figure 11(v) is for different α, Fig. 11(w) is for different v and Fig. 11(x) is
for different θ. With q increasing, the number of customers who leave the system
due to failure in retrial increases, which leads to the decreasing trend of E[L]
as shown in the figures. Besides, the effect of q on E[L] becomes weaker with
q increasing. And the lager α is, the smaller the effect of q on E[L] becomes,
which means that the departure rate of customers will diminish the impact on
the mean number of customers in the orbit when the load of the orbit becomes
lighter.
(3) Performance optimization
Using the performance measures derived in Sect. 3.3, this section designs a cost
function, i.e., the average total cost of the system per unit time, and establishes
an optimization model on the basis of the cost function. Then we use the opti-
mization model to optimize the queueing model established in Sect. 3.1 under
the environment of web service system, and obtain the optimal parameters and
the minimum average total cost per unit time.
To design the cost function, the cost structure is defined firstly. Assume there
are two costs in the system, one is the waiting cost, and the other is the service
cost. Assume the average waiting cost of each customer per unit time is c, the
average service cost of each server in a regular busy period per unit time is ar ,
and the average service cost of each server in a working vacation period per unit
time is av Then, the average total cost of the system per unit time M can be
expressed as
∼
M = cE[L] + ar Pr + av Pv , (6)
436 M. Wu et al.
i.e., cost function. Based on Eq. (6), establish an optimization model as

∼
min = cE[L] + ar Pr + av Pv
⎧
⎪ − − − −
⎪
⎪ μ (λ α −1)(αq − λb ) < μb (λb +λb q)α
⎨ b∼ b
s.t. E[L] ≥ 0
⎪
⎪ 0 < Pr , P v < 1
⎪
⎩
0 < μv < μb < 1, 0 < θ, α < 1.
Then, consider using the optimization model above to optimize the Geo/Geo/1
retrial queue with working vacation interruption and nonpersistent customers
established in Sect. 3 under the environment of web service system. Assume the
average waiting cost of each customer per unit time is 5, the average service cost
of each server in a regular busy period per unit time is 10, the average service
cost of each server in a working vacation period per unit time is 6, the probability
that the user requests service during a regular busy period is 0.6, the probability
that the user requests service during a working vacation period is 0.3, the retrial
rate is 0.6 and the probability that the user gives up receiving service due to the
failure of retrial is 0.2. Taking these parameters into the optimization model and
using LINGO to solve it, get μv = 0.277, μb = 0.739, θ = 0.125, M = 11.359. The
result means that the average total cost of the system per unit time is 11.359,
i.e., the minimum sum of the average waiting cost and the average service cost
per unit time, now, the service rate in a regular busy period and the service
rate in a working vacation period are 0.739, 0.277 respectively, and the vacation
period rate θ is 0.125. That is to say, in order that the performance of the web
service system is the best, managers should adjust the actual service capabilities
of the server to guarantee that the service time in a regular busy period follows
a geometric distribution with parameter 0.739, the service time in a working
vacation period follows a geometric distribution with parameter 0.277 and the
working vacation time follows a geometric distribution with parameter 0.125.
4 Conclusions
With the development of the internet, the web service industry has put forward
many new requirements to the design of queueing model. This paper provides a
queueing solution for web service. After presenting three practical problems that
web service encounters in real life, a queuing model which can deal with these
problems is constructed and a series of necessary researches are carried out on the
model. Then, based on the results of the researches, the numerical analysis and
model optimization are conducted, which provide some valuable consultations for
the web service management under the corresponding circumstances. In addition,
since the model proposed in this paper is a single server, which is somehow
discrepant with the reality of multi-server web service, the number of services
can be pending in the future research to better cope with the actual situation.
References
1. Burr T (2001) Introduction to matrix analytic methods in stochastic modeling.
Technometrics 95(3):1379
2. Dimitriou I (2015) A retrial queue for modeling fault-tolerant systems with check-
pointing and rollback recovery. Comput Ind Eng 79:156–167
3. Kobayashi H, Konheim AG (1977) Queueing models for computer communications
system analysis. IEEE Trans Commun 25(1):2–29
4. Kosten L (1947) On the influence of repeated calls in the theory of probabilities of
blocking. De Ingenieur 49(1):947
5. Li J, Tian N (2007) The discrete-time GI/Geo/1 queue with working vacations
and vacation interruption. Appl Math Comput 185(1):1–10
6. Li J, Tian N (2007) The M/M/1 queue with working vacations and vacation inter-
ruptions. J Syst Sci Syst Eng 16(1):121–127
7. Li J, Tian N (2008) Analysis of the discrete time Geo/Geo/1 queue with single
working vacation. Qual Technol Quant Manage 5(1):77–89
8. Li T, Wang Z, Liu Z (2012) Geo/Geo/1 retrial queue with working vacations and
vacation interruption. J Appl Math Comput 39(1):131–143
9. Liu Z, Song Y (2013) Geo/Geo/1 retrial queue with non-persistent customers and
working vacations. J Appl Math Comput 42(1):103–115
10. Meisling T (1958) Discrete-time queuing theory. Oper Res 6(1):96–105
11. Palm C (1953) Methods of judging the annoyance caused by congestion. Tele
4(189208):4–5
12. Phung-Duc T (2014) Multiserver retrial queues with two types of nonpersistent
customers. Asia Pac J Oper Res 31(2):144–171
13. Rajadurai P (2015) Analysis of M [X]/G/1 retrial queue with two phase service
under bernoulli vacation schedule and random breakdown. In J Math Oper Res
7(1):19–41
14. Servi LD, Finn SG (2002) M/M/1 queues with working vacations (M/M/1/WV).
Perform Eval 50(1):41–52
15. Tao L, Zhang L, Xu X (2013) An M/G/1 queue with single working vacation and
vacation interruption under Bernoulli schedule. Appl Math Model 37(3):1564–1579
16. Tian N, Zhang Z (2006) Vacation queueing models: theory and applications.
Springer, New York
17. Tian N, Ma Z, Liu M (2008) The discrete time Geom/Geom/1 queue with multiple
working vacations. Appl Math Model 32(12):2941–2953
18. Yang T, Li H (1995) On the steady-state queue size distribution of the discrete-time
Geo/G/1 queue with repeated customers. Queueing Syst 21(1):199–215
Pan-Tourism Urbanization Model Based
on System Dynamics: A Case Study of Barkam
Lin Hu, Lu Gan(B) , and Qian Yang
College of Architecture and Urban-Rural Planning, Sichuan Agricultural University,

ganlu soarpb@sicau.edu.cn
Abstract. With the rapid development of urbanization, tourism-driven

urbanization process has become an important development strategy in
China. There are many examples of the tourism urbanization, however,
the integration with industries is scarce. In order to describe the char-
acterization of the industrial clusters system, this paper uses SPSS [8]
to analyze the data. And it also uses DYNAMO simulation language
and Vensim [2] to establish the pan-tourism urbanization model based
on the theory of industrial cluster and system dynamics. Simultaneously,
taking Barkam city in the northwestern Tibetan area of Sichuan as an
example, this paper analyzes the operation mechanism of pan-tourism
urbanization mode to find the optimal system structure, which can pro-
vide corresponding scientific basis and theoretical support for promoting
the development of urban and rural areas, improving the level of new
urbanization and building new towns.
Keywords: System dynamics · Pan-tourism · Industrial cluster · Sim-

ulation
1 Introduction
In recent years, the model of new urbanization has been widely concerned.
Tourism industry and related industries develop interactively in order to promote
the new urbanization and the coordinated development of new rural construc-
tion. Therefore, the model of pan-tourism industry integration and industrial
cluster development is an effective way to promote the development of urban-
ization, which is proposed by experts and scholars at home and abroad.
Some scholars have made some research ideas. Alexandrova [1] aims to review
tourism cluster initiatives in Russia, particularly, tourism cluster formation
processes. Jackson [6] takes four towns in Australia as an example, exploring
the effectiveness of correlation factors of these four regions, and studying the
competitive advantage of cluster theory. However, they only concentrate on the
research of the existing tourism industrial clusters evaluation and focus on using
descriptive language to assess or predict which lacks a practical and objective
evaluation system.
Pan-Tourism Urbanization Model Based on System Dynamics 439
Jackson [5] probes into the development of regional tourism clusters, and he
finds the correctness in theory of the Porter’s [12] model. Tourism cluster may
become a way to convert comparative advantage into competitive advantage and
make better use of existing tourism resources in China. It can be more adapted to
the characteristics of Chinese region. Qian [13] questions Mullins’s [10] classical
theory of tourism urbanization, and he shows the idea that the process of urban-
ization driven by modern tourism is not the product of post-modernization of
urban cultural expression, however, it is the product of tourism-related services
based on relatively standardized regulations. The two scholars focus on the com-
parative analysis of relevant theoretical models and existing data. The research
model they put forward has positive effects on the development of domestic
tourism, but they ignore the systematic description of tourism urbanization.
This paper is organized as follows. In Sect. 2, the pan-tourism urbanization is
introduced and then we proposed a differential pan-tourism system model with
industrial clusters theory. In Sect. 3, on the basis of pan-tourism system with
the actual situation in Barkam, we investigate the relevant data, including the
number of tourists, tourism revenue and other data about urbanization. Once we
have carried out regression analysis, we can get simulation results for pan-tourism
urbanization system. And we analyze the results and give recommendations. In
Sect. 4, we make a conclusion about the effectiveness of such a model and discuss
the shortcomings of this model.
2 Modeling
2.1 Concept Model
System dynamics is a science of dynamic complexity of the system developed by
the Professor Forester of MIT, which emphasizes the overall consideration of the
system and the interaction of each composition of the system. It is designed to
deal with the problems of high order, nonlinear and multiple feedbacks in com-
plex system [17]. Based on the method of qualitative and quantitative analysis
and the system dynamic simulation experiments, which can forecast the dynamic
change, the decision makers can observe the simulation results and take different
measures in various contexts.
In this paper, the process of system dynamics modeling can be divided as fol-
lows: firstly, according to the characteristics of problems, we realize the purpose
of modeling and determine the system boundaries to avoid the system archi-
tecture which will be too large or too small; secondly, we analyze the system
structure to divide the system hierarchy and sub modules, determine the causal
feedback loop and define the interaction mechanism among elements; thirdly,
through the regression analyzing of the data, we establish the system dynamic
mathematical model completely; finally, we analyze the simulation results which
are obtained from the model running and evaluate the model in order to identify
the problems and improve the model.
Based on the feedback control theory and the concept of industrial cluster
development, this paper constructs the dynamic analysis model of pan-tourism
440 L. Hu et al.
Fig. 1. Flow chart for research process
urbanization system to find out the basic reasons, key factors, changing rules
and internal relations further. The research ideas are shown in Fig. 1.
2.2 System Elements Determination

System dynamics believes that the fundamental power and the direct cause of
the whole socio-economic development come from the inner system, which deter-
mines the behavior pattern and characteristics. Besides, the system boundaries
define the minimum number of units. Only defining the boundaries can the sys-
tem solve structural problems [15]. Traditional tourism is only a single industry
with simple form, while the pan-tourism urbanization model is the integration
of pan-tourism industry development.
Pan-tourism industrial clusters have three main modes, such as the feed-
back, promotion and interaction modes [4]. The first one is the feedback model,
referring to the non-profit service industries, which provide the tourism indus-
try with resourceful and environmental infrastructures. They receive economic
benefits from tourism and make themselves developing stably and healthily with
tourism. The second promotion model means industries which cannot integrate
with the tourism to produce new products or formats, but their development
will be promoted when tourism developed vigorously. For instance, the lodg-
ing, catering and transportation industry can supply support elements for the
tourism while the wholesale and retail trade, real estate, warehousing postal
services, construction and industry which offering corollary facility to tourism.
The third one is the interaction mode, referring to integration of agriculture
and tourism, which can form a new tourism products and formats, rich tourism
content and enhance the value of tourism products. Not only further does the
mode develop the new tourism market, but also it promotes the industry and
tourism industry to develop collaboratively.
The development of tourism industry leads to the development of related
industries and forms pan-tourism industrial clusters, therefore it brings about the
agglomeration of population and elements, and promotes the process of urban-
ization in local area. Urbanization is the process of the evolution of the economic
structure, mode of production, life style and social concept after rural population
and various factors constantly gathered to the city. The level of urbanization is
usually measured by the proportion of urban population to total population.
Actually, urbanization is not just a regional change in the nature of the
population. The connotation of urbanization is very abundant which means a
regional change process based on the non-agricultural population and non - agri-
cultural economy [16], and it’s overgeneralization if the urbanization level only
measured by the population ratio. This paper references to Fang [3], Sun [14]
and other scholar’s defining of urbanization degree and combines with the devel-
opment characteristics of pan-tourist urbanization to define urbanization degree
of pan-tourism urbanization system by using two indicators of urbanization and
economic urbanization.
The population urbanization rate is usually used to measure the level of
urbanization, and it is an important symbol to measure a country’s level of
economic development [7]. The integration development of tourism and related
industries influence the changes of the industrial structure, reduce the employees
of the primary industry continuously and increase the employees of the second
and tertiary Industry. In addition, the urban employment increases through the
transfer of labor, while gradually realizing the transformation of farmers’ iden-
tity, then completing the population in situ urbanization.
The economic urbanization rate reflects the non-agricultural process of the
overall economic structure, characterized by the proportion of the second and
third industries gradually increased while proportion of primary industry grad-
ually decreased. Tourism drives pan-tourism industrial clusters, whose output
values are calculated from the second and third industrial output values, loca-
tion entropies and the contribution rates [18]. The economic urbanization rates
are the proportion of cluster output values to GDPs.
442 L. Hu et al.
2.3 System Causality Analysis

Pan-tourism urbanization system includes three subsystems: tourism subsystem,
industrial cluster subsystem and urbanization subsystem. This paper chooses
the relevant variables from the perspective of pan-tourism industrial cluster and
urbanization, and analyzes the system elements from the perspective of tourism
driven urbanization.
The pan-tourism industrial clusters and urbanization interactive development
system consist of three basic feedback loops: firstly, tourism drives pan-tourism
cluster through industrial integration, increasing the cluster output values and
improving the level of urbanization; secondly, industrial cluster increases employ-
ment in the second and third industries. The transfer of the labor also increases
the urban population and raises the urbanization level; thirdly, the development
of region economy and the increase of urban population promote urban con-
struction, thus enhancing tourism reception capacity and developing the tourism
industry. The causal relationship is shown in Fig. 2.
Tourist amount
+ -
Pollution
+
Tourism +
development Level of three
Tourism reception industry development
+
+ +
Pan - tourism
industry +
development
The change of
+ employment structure
Level of Industrial Cluster
+
urbanization +
+
Migration
Fig. 2. Pan-tourism urbanization causality graph
POV GDP
<CPV> STE ICAF

ICSU SOV EUR
TLE
ICTA CAPOV SCR
ICCS
UP <Time>
IOV COV
ICTT
TIOV
PUR TP ICTL CPV
ICPT LCOV
SLE
TR ICTW
<EUR> TIR ICET IPT WROV
REOV TCR
TN ICTE
TIA TRA
PIOV
TOV
ICTP
TRAV
OSOV
ICTTIR
ICTO
Fig. 3. Pan-tourism urbanization system flow diagram

Table 1. Meaning of the main variables and influence factors
Notation Variable name Unit

TN Number of tourists 104 persons
TR Tourism revenue 104 yuan
TIR Tourist increase rate No dimension
TIA Tourist increase amount 104 persons
TRA Tourist reduction amount 104 persons
TRAV Tourism revenue added value 104 yuan
PIOV The primary industry output value 104 yuan
SOV The secondary industry output value 104 yuan
TOV The tertiary industry output value 104 yuan
SLE Secondary industry location entropy No dimension
TLE Tertiary industry location entropy No dimension
SCR Secondary industry contribution rate No dimension
TCR Tertiary Industry contribution rate No dimension
EUR Economic urbanization rate No dimension
CPV Cluster production value 104 yuan
STE Employment in the second and third industries 104 persons
UP Urban population 104 persons
TP Total population 104 persons
PUR Population urbanization rate No dimension
IPT Impact index of pollution on tourists No dimension
CAPOV Characteristic agricultural products output value 104 yuan
IOV Industrial output value 104 yuan
COV Construction industry output value 104 yuan
TIOV Transportation industry output value 104 yuan
LCOV Lodging and catering industry output value 104 yuan
WROV Wholesale and retail output value 104 yuan
REOV Real estate output value 104 yuan
PIOV Postal industry output value 104 yuan
OSOV Other service industry output value 104 yuan
ICAF Influence coefficient of CAPOV on POV No dimension
ICTA Influence coefficient of TR on CAPOV No dimension
ICTI Influence coefficient of TR on IOV No dimension
ICTC Influence coefficient of TR on COV No dimension
ICTT Influence coefficient of TR on TOV No dimension
ICTL Influence coefficient of TR on LCOV No dimension
ICTW Influence coefficient of TR on WROV No dimension
ICTE Influence coefficient of TR on REOV No dimension
ICTP Influence coefficient of TR on PIOV No dimension
ICTO Influence coefficient of TR on OSOV No dimension
ICTT Influence Coefficient of TN on TRAV No dimension
ICET Influence coefficient of EUR on TRAV No dimension
ICPT Influence coefficient of PUR on TN No dimension
ICCS Influence Coefficient of CPV on STE No dimension
ICSU Influence Coefficient of STE on UP No dimension
444 L. Hu et al.
2.4 System Flow Diagram and Parameter Equation
(1) System Flow Diagram

As shown in Fig. 3, two flow variables, three rate variables and multiple aux-
iliary variables are selected to construct the system flow diagram based on the
causal feedback map. The meaning of the main variables and influence factors
in this model is shown in Table 1.
(2) Index System and Parametric Equation
The parametric equations are determined according to the logical relation-
ship between system variables. There are two level variables named number of
tourists and tourism revenue, and the three rate variables are tourist increase
amount, tourist reduction amount and tourism revenue added value. The auxil-
iary variables are logically related to the level and rate variables. The constants
include the influence coefficients, the contribution rates, the location entropy,
and the total population expressed by the table function. The main parameters
and equations are shown in Table 2.
Table 2. Main variables and equations
Variable type Variable name Equation

Level variable TN INTEG (TIA − TRA, Initial value)
TR INTEG (TRAV, Initial value)
Rate variable TIA TN × TIR
TRA TN × IPT
TRAV ICTTIR × TN
Auxiliary TIR ICETEUR+ICPTPUR
variable
CAPOV TR × ICTA
IOV ICTI × TR
COV ICTC × TR
TIOV ICTT × TR
LCOV ICTL × TR
WROV ICTW × TR
REOV ICTE × TR
PIOV ICTP × TR
OSOV ICTO × TR
POV ICAF × CAPOV
SOV IOV + COV
TOV Regression equation
CPV SLE × SOV × SCR + TLE × TOV × TCR
GDP POV + SOV + TOV
EUR CPV/GDP
TEP ICCS × Ln(CPV)
UP Regression equation
PUR UP/TP
Constant SLE, TLE, SCR, TCR No dimension
TP Table function
3 Case Analysis
3.1 Data Sources
In this paper, Barkam city in northwest Tibetan area of Sichuan is selected as the
study object. The urbanization mode driven by tourism has been well reflected
at here. Based on the actual situation of Barkam, the pan-tourism urbaniza-
tion system is simulated [9]. The primary data is collected from Barkam’s sta-
tistical yearbook and Sichuan Province statistical yearbook. Although there are
some uncertain values, such as the total population and urban population, which
are strongly affected by the social environment and natural disasters, it can be
explained reasonably to make the model reflect the general trend of pan-tourism
system.
The simulation step length of this model is set to 1 year, and the simulation
year is from 2011 to 2025. The initial values for 2011 are determined based on the
Barkam’s statistical yearbook: the number of tourists was 505,800; tourism rev-
enue is 434.48 million Yuan. Table 3 shows the values of parameters determined
by regression analysis, table function and other methods.
Table 3. Parameter’s value
Parameter Value Explanation

SLE 0.354 (Regional secondary or tertiary industry
output/Regional GDP)/(national secondary
or tertiary industry output/GDP)
TLE 1.654
SCR 0.15 The contribution degree of urban secondary
and tertiary industry to regional GDP
TCR 0.75
ICET 0.172 Regression equation:
ICPT 4.85 TIR = (0.172099×EUR−4.85×PUR)+2.4092
ICT 0.964, 0.996, 0.486, Regression equation:
1.164, 2.094, 0.151 TOV = 0.964 × TIOV + 0.996 × LCOV
+0.486 × WROV + 1.164 × REOV
+2.094 × PIOV + 0.151 × OSOV
ICSU 0.42, −1.788 Regression equation:
UP = −1.788 × STE + 0.42 × STE2 + 4.616
ICCS, ICTTIR, ICTA, 0.514, 101.277, 0.225, Regression analysis determines the parameters
ICTI, ICTC, ICTT, 0.298, 0.068, 0.384,
ICTL, ICTW, ICTE, 0.316, 0.081, 0.154,
ICTP, ICTO 1.662, 0.788
IPT 0.04 The extent of the impact of pollution on the
reduction of visitors
TP Table function With Lookup
(([(2011,5.9)-(2015,5.9)],(2011,5.9),
(2012,5.9),(2013,5.9),
(2014,5.9),(2015,5.9))
446 L. Hu et al.
3.2 Validity Test

(1) Historical Inspection
In this paper, the comparison of simulation results with actual data is con-
ducted to test the effectiveness of the model. The simulated data, actual data
and error rate for 2011–2015 are shown in Table 4. These data show that the
error rate is kept within the 10% controllable range, which indicates the model
can forecast the running state of the real system with good fitting degree.
(2) Operation inspection
The step length of these fitting values of tourism revenue is 1 year, and thus
we select the step length of 0.5 years and 0.25 years to make comparison. It
can be seen that curve of tourism revenue of Barkam changes implicitly under
different step lengths of simulation. And the average error rate is less than 10%,
which means the analogue system is stable (see Fig. 4).
3.3 Simulation Results

(1) Tourists and Tourism Revenue
Figures 5 and 6 show the simulation results of the tourists and the tourism
revenue in Barkam City. The number of tourists of Barkam will increase steadily
Table 4. Historical data and simulation results analysis
Variable name Value type 2011 2012 2013 2014 2015

T N/104 persons Actual value 50.58 64.57 77.3 85.74 92.02
Simulation value 50.58 60.8498 73.1725 86.4152 99.9342
Error rate/% 0 −5.761 −5.34 0.787 8.601
T R/104 yuan Actual value 43448 57725 69642 76572 84565
Simulation value 43448 57023.2 64819.7 73864.2 84249.9
Error rate/% 0 −1.216 −6.924 −3.536 −0.373
250000 1year
0.5year
0.25year
Tourism revenue/10 yuan
200000
4
150000
100000
50000
2012 2014 2016 2018 2020 2022 2024 2026
Year
Fig. 4. Barkam’s tourism revenues under different step simulation

180 180
160 160
Number of tourists/104persons
Number of tourists/104persons
140 140
120 120
100 100
80 80
60 60
40 40
2012 2014 2016 2018 2020 2022 2024 2026 2012 2014 2016 2018 2020 2022 2024 2026
Fig. 5. Barkam’s number of tourists Fig. 6. Barkam’s tourism revenue simu-

simulation lation
from 2011 to 2023, reaching 1.5 million by 2019 and having a peak at the end
of 2023. After that, there will be a slight decline. The speculative reason is that
pollution caused by the industrial clusters may affects the tourists’ enthusiasm
to travel to Barkam, while the development of the pan-tourism industry is driven
by tourism. At the same time, Barkam’s tourism revenue is also growing steadily
and it will exceed 20 billion Yuan by 2022. We can see that the booming tourism
industry of Barkam will be in a steady upward trend in the next decade from the
simulation results. Therefore, in order to maintain the vitality of tourism indus-
try and promote the development of industry clusters to propel urbanization, the
government should pay more attention on the tourism industry meantime to con-
centrate on the impact of resource depletion and pollution from local industrial
development [11].
(2) Simulation Analysis of Economy
As the Fig. 7 shown, the three industrial output values grow steadily in the
period from 2011 to 2025. It is speculated that tertiary industry output values
of Barkam will exceed 3 billion Yuan, while the GDP will exceed 4 billion Yuan.
Contrasting the primary and secondary industries, the tertiary industry out-
put values increased significantly. The main reason is that tourism promote the
development of related industries, such as transportation, lodging and catering
industry and related social public services, so that the tertiary industry become
the leading industry.
On the contrary, the growth rate of the primary industry is very slow, presum-
ably due to the destruction of the local ecological environment and the obstruc-
tion of rural construction in the process of realizing urbanization. Local gov-
ernments should take corresponding measures to strengthen the linkage of the
tourism industry with the primary and secondary industry, promote the devel-
opment of agricultural tourism and industrial tourism vigorously, extend the
chain of business opportunities strenuously, and create an advantage platform
to promote economic development effectively.
448 L. Hu et al.
600000
Primary industry output value
Secondary industry output value
500000 Tertiary industry output value
GDP
Output value/10 yuan

400000
4
300000
200000
100000
0
2012 2014 2016 2018 2020 2022 2024 2026
Year
Fig. 7. Barkam’s output value simulation

0.54
3.4
0.53
Urban population/10 persons
3.3
0.52
Urbanization rate/%
3.2
4
0.51
3.1
3.0 0.50
2.9 0.49
2.8 0.48
2012 2014 2016 2018 2020 2022 2024 2026 2012 2014 2016 2018 2020 2022 2024 2026
Year Year
Fig. 8. Barkam’s urban population sim- Fig. 9. Barkam’s urbanization rate sim-
ulation ulation
(3) Population Simulation

The simulation results of the employment population, the urban population
and the urbanization rate are shown in Figs. 8 and 9. According to simulation
results, the number of urban population will reach 33,000 while the urbanization
rate of population will reach 53% by 2025, and both of them are in a steady rising
state.
The main reason for the increase is the development of local tourism, driving
the development of the tertiary industry such as services and circulation indus-
tries. These industries require low technical and easy to realize the labor force
transfer.
Furthermore, the process of rural population to non-agricultural population is
accelerated, and the in-situ urbanization of farmers is achieved which constantly
improve the level of urbanization. From the data we can see that although the
level of urbanization has increased year by year, the growth rate is relatively
slow, mainly due to the fact that Barkam is located in the Tibetan area of
northwest Sichuan, and the residents have long been engaged in agriculture and
animal husbandry. The process of local urbanization has been speeded up by
tourism, but some remote farmers still use the traditional agricultural mode,
which affects the local urbanization process.
550000
0.924
500000
Economic urbanization rate/%

450000 0.922
Cluster output/10 yuan
400000
0.920
4
350000
300000 0.918
250000
0.916
200000
150000 0.914
100000
0.912
2012 2014 2016 2018 2020 2022 2024 2026 2012 2014 2016 2018 2020 2022 2024 2026
Year Year
Fig. 10. Barkam’s cluster output simu- Fig. 11. Barkam’s economic urbaniza-
lation tion rate simulation
(4) Cluster Simulation

Cluster output value and economic urbanization rate simulation results
shown in Figs. 10 and 11.
This paper refers to the definition of cluster output value by scholar Xiaoye
Zhou [18], whose economic meaning is the development of secondary and ter-
tiary industry in the region, relating to the entire country. As the Fig. 10 shows,
the cluster output increased year by year from 2011 to 2025, which will exceed 5
billion Yuan in 2025. The stable increase in cluster output values means that the
local secondary and tertiary industries are booming, which will lead to overall
economic development of Barkam. Economic urbanization rate is defined as the
proportion of cluster output values of GDP. The current urbanization rate of
Barkam city has reached more than 90%, which shows that local industrial clus-
ters have reached a high level. Figure 11 speculates that the Barkam’s economic
urbanization rate will be improved steadily in the next decade.
It can be concluded that the tourism industry of Barkam will develop rapidly
in the next decade and the output values of industrial clusters will increase
steadily. The urban GDP and urban population are also increasing. With the
development of industrial clusters driven by tourism, the urbanization rate of
Barkamis gradually increased, and the level of urbanization has laid a foundation
for the local tourism development and industrial clusters. Meanwhile the develop-
ment of tourism in Barkam has higher requirements for the whole urban system
from the external to the internal connection service, which promotes the devel-
opment of local construction, transportation, lodging catering and social public
services, etc. The direct effect of tourism on the evolution of local urbanization
has increased the attractiveness of cities and towns, and accelerated the process
of Urbanization. In addition, we find that the growth rate of local agricultural
output value is smaller than other two industries, which is not conducive to the
coordinated development of various industries and the optimization of industrial
structure. Therefore, the local government should maintain the smooth devel-
opment of tourism and pay attention to the integration of tourism and related
450 L. Hu et al.
industries so that the development of the local three industries will be promoted
harmoniously.
4 Conclusion
In this paper, the quantitative relationship was analyzed by SPSS, and the
DYNAMO language and Vensim are used to establish the system dynamic model
of pan-tourism urbanization. The three subsystems about tourism, industrial
cluster and urbanization are interrelated through the visitor numbers, tourism
revenue, and clusters output value and employment population. The interactive
development relationship among the three reflects the industrial cluster driven
by tourism development and the process of rural labor transfer. This model
systematically analyzes pan-tourism urbanization operation mechanism, and it
provides the corresponding scientific basis and theoretical support to promote
urban and rural development, improve the level of characteristic urbanization
and the construction of new towns. Because the selected city of Barkam has
certain particularity, the system elements may not be comprehensive. In addi-
tion, as the difference of local development and economic level, the pan-tourism
urbanization model may not be suitable for all cities and towns, so it is necessary
to adjust specific elements and coefficients in a specific place in order to make a
more scientific analysis.
Acknowledgements. This research was supported by the Key Funds of Sichuan

Social Science Research Institution “System Science and Enterprise Development
Research” (Grant NO. Xq15B09), the Youth Funds of Sichuan Provincial Education
Department (Grant NO. 14ZB0014), and the Humanities Social and Sciences Research
Funds of Education Ministry (Grant NO. 15XJC630001).
References
1. Alexandrova A, Vladimirov Y (2016) Tourism clusters in russia: what are their
key features the case of vologda region. Worldwide Hospitality Tourism Themes
8(3):346–358
2. Eberlein RL, Peterson DW (1992) Understanding models with vensim. Eur J Oper
Res 59:216–219
3. Fang C, Jing W (2013) A theoretical analysis of interactive coercing effects between
urbanization and eco-environment. Chin Geogr Sci 23(2):147–162
4. Gimzauskiene E, Duoba K et al (2015) Tourism clusters in eastern poland - analysis
of selected aspects of the operation. Procedia Soc Behav Sci 213:957–964
5. Jackson J (2006) Developing regional tourism in china: the potential for activating
business clusters in a socialist market economy. Tourism Manag 27(4):695–706
6. Jackson J, Murphy P (2006) Clusters in regional tourism an australian case. Ann
Tourism Res 33(4):1018–1035
7. Jian ML, Chi FL (2016) A qualitative study of urbanization effects on hotel devel-
opment. J Hospitality Tourism Manag 29:135–142
8. Liu RX, Kuang J et al (2001) Principal component regression analysis with SPSS.
Comput Methods Programs Biomed 71(2):141–147
9. Matinzadeh MM, Koupai JA et al (2017) Development of an innovative integrated

model for the simulation of nitrogen dynamics in farmlands with drainage systems
using the system dynamics approach. Ecol Model 347:11–28
10. Mullins P (1991) Tourism urbanization. Int J Urban Reg Res 15(3):326–342
11. Ning B, He Y (2007) Tourism development and water pollution: case study in
Lijiang Ancient Town. China Population Resour Environ 17(5):123–127
12. Porter M (1990) Competitive Advantage of Nations. Free Press, New York
13. Qian J, Feng D, Zhu H (2012) Tourism-driven urbanization in china’s small town
development: a case study of zhapo town, 1986–2003. Habitat Int 36(1):152–160
14. Sun PJ, Song CL et al (2013) Non-coordination in china’s urbanization: assessment
and affecting factors. Chin Geogr Sci 23(6):729–739
15. Surya B (2016) The processes analysis of urbanization, spatial articulation, social
change and social capital difference in the dynamics of new town development in
the fringe area of makassar city (case study: In metro tanjungbunga area, makassar
city). Procedia Soc Behav Sci 227:216–231
16. Tan Y, Xu H et al (2017) A study of best practices in promoting sustainable
urbanization in china. J Environ Manag 193:8–18
17. Zhao B, Tang T, Ning B (2017) System dynamics approach for modelling the
variation of organizational factors for risk control in automatic metro. Saf Sci
94:128–142
18. Zhou XY, Fu DM et al (2016) Research on interaction between industrial clusters
and urbanization based on system dynamics. J Shenyang Univ Technol 9(1):47–52
(in Chinese)
A Study on the Three Different Channel Models
in Supply Chain Under the Network
Externalities
Jinjiang Yan, Jiankai Xing(B) , Kai Zhu, and Ke Liu
Business School, Sichuan University,

mukaimu@163.com
Abstract. With the rapid developments of Internet, many manufactur-

ers and retailers choose not only the offline channel but also the online
channel. According to the choices of different channels, this paper models
three types of channel models in supply chains. Considering the presence
of the network externalities, the degree of substitution between online
and offline channels and consumers’ preference for two entities (the man-
ufacture and the retailer), we study the effect of retailer’s choice to online
channel and the differences among the three different channel models.
Our main findings are that compared to the first two models, the sales
price of unit product will be smaller and in contrast the market size
will be bigger in the third model. Also, when the entities make the best
strategies, the retailer’s profit and the total profit of the channel can
reach the maximum. The study shows that the introduction of online
channel can create more social values, which is why the dual-channel
model of retailers is so widespread. In the end, a numerical example is
given to verify the conclusions.
Keywords: Network externality · Supply chain · Online channel ·

Offline channel
1 Introduction
With the rapid development of the Internet, on the one hand, online sales have
become a mode in which the manufacturers and retailers can increase sales. On
the other hand, more and more products have gradually affected by the network
externality. The introduction of online channel for the manufacturer can increase
their sales and avoid the large retailers’ monopoly of the offline channel. At the
same time, it can provide opportunities for the retailer to expand the market
and increase sales. Many well-known manufacturers (such as Lenovo, Apple, etc.)
chose to introduce the online channel based on the offline channel. However, a lot
of famous retailers (such as Alibaba, Wal-Mart Stores, etc.) had also introduced
the online channel one after another. Nevertheless, most studies of dual-channels
A Study on the Three Different Channel Models in Supply Chain 453
focused on the choice of the manufacturer, the internal channel coordination and
channel pricing, but ignored the impact of network externalities and the choice
of retailers in dual-channel situation.
The network externality means that the customer can obtain more utilities
when more suppliers are willing to join in the network and opposite situation
is similar. In making strategies for channel markets, however, sellers (including
the manufacturer and the retailer) have typically relied on assumptions and par-
adigms that apply to business without network effects. Therefore, the strategy
that they make may not match the economics of their changing industries. Many
literatures studied the pricing strategy for the dual-channel. Liu [13] pointed that
compared to the traditional single channel, when the manufacturer brings in the
online channel, pricing strategies will have different effects on the channel. Leider
[9] studied experimentally bargaining in a multiple-tier supply chain with hor-
izontal competition and sequential bargaining between tiers. They found that
the structural issue of cost differentials dominates personal characteristics in
explaining outcomes, with profits in a tier generally increasing with decreased
competition in the tier and increasing with decreased competition in alternate
tiers. Naeem, Charles and Rafay [3] considered a dynamic problem of joint pric-
ing and production decisions for a profit-maximizing firm that produced multiple
products. They presented a solution approach which is more general than pre-
vious approaches that require the assumption of a specific demand function.
Chuang and Yang [7] modeled the deteriorating goods pricing strategy with
reference price effects and solved the dynamic problem by obtained theoretical
results.
Some scholars carried out their research from the aspect of dual-channel
coordination. Shamir [14] studied the operational motivation of a retailer and
showed that by making forecast information publicly available to both his man-
ufacturer and to the competitor, a retailer is able to credibly share his forecast
information-an outcome that cannot be achieved by merely exchanging informa-
tion within the supply chain. Yang and Zhuo etc. studied a two-period supply
chain that consists of a retailer and a supplier and found that a revenue shar-
ing contract significantly affects the retailer’s payment behavior and supplier’s
wholesale price [16]. Zhong, Xi and Jing [17] analyzed the coordinating mecha-
nisms for a single-period supply chain comprising of one supplier and one retailer.
They pointed that the transfer payment plus a contract could mitigate the down
side risk effect on the supply chain performance. Yue and Allen, etc. [15] studied
a supply chain consisting of one supplier and n retailers and pointed the pric-
ing strategy under different situations. Wenyi et al. [6] presented an integrated
model of marketing choices and supply chain design decisions and developed the
closed-loop supply chain. Yulan and Stein, etc. reviewed operational models in
service supply chain management (SSCM) and examined the definitions of SSCM
[6]. Panda et al. [12] analyzed coordination of a manufacturer-distributer-retailer
supply chain, where the manufacturer exhibited corporate social responsibility
[12]. They built a manufacturer-Stackelberg game setting to propose a contract-
bargaining process and resolved channel conflict. Jing, Hui and Ying [5] analyzed
454 J. Yan et al.
the channel effects of four kinds of dual-channels’ pricing strategies. Aimin and
Liwen analyzed the stochastic demand, joint promotion, the price competition
and coordination problems between the manufacturer and the retailer [1]. Ravi
and Rajib [11] studied the co-opetition between differentiated platform in two-
sided markets. In their study, they highlighted the importance of technology.
They pointed that collaboration might provide incentives for a dominant plat-
form. In this paper, we consider the platform/channel ownership is independent,
not belonging to the seller or the consumer.
Some scholars focus on the study of network externalities for the competi-
tive strategy. Blocher [4] showed that amplifying network externalities among
mutual funds could explain substantial flow-based effects documented in the lit-
erature. Liu [10] identified the valid mechanism for the alternative range of profit-
sharing contracts and analyzed the effect of product substitutability coefficient
and network externalities on the alliance and profit-sharing contract. Hagiu and
Spulber [8] presented first-party content and coordination in two-sided markets
and they found that the strategic use of first-party content by two-sided plat-
form was driven by two key factors: the nature of buyer and seller expectations.
Edward and Anderson researched the platform performance investment in the
presence of network externalities [2]. They carried out a full analysis of three
distinct settings: monopoly, price-setting duopoly and pricing-taking duopoly.
They concluded that the conditions under which offering a platform with lower
performance but greater availability of content can be a winning strategy.
These literatures have analyzed the pricing and coordination strategies of the
dual-channels and the influence of network externalities on pricing and competi-
tive strategy of enterprises respectively. We can conclude that the main problem
of pricing and coordination is to coordinate the distribution of benefits between
manufacturers and retailers in the structure of the manufacturer’s choice of dual
channel. The network externality can increase the strength of the sharing econ-
omy, bring about a positive cycle effects, and provide guidance for pricing and
coordination of enterprises. However, from the two aspects of researches, there
exists the following deficiencies.
The retailer’s behavior can also have a great impact on the dual-channel in
the supply chain system. But most of the studies about the dual-channel and
the network externalities are limited to the behavior of the manufacturer. In
addition, the network externalities also have a great influence on the pricing and
decision making. Many literatures about the dual-channel have largely ignored
the effect of network externalities on consumers and the whole supply chain
system. Many studies about the network externalities simply consider the effect
of network externalities on the single channel mode, not using it to dual-channel
mode.
With the premise that both entities’ (including the manufacturer and the
retailer) optimal profits must greater than zero, this paper considers the prod-
ucts with the network externalities and builds three channel models under the
condition that the retailer is the leader in the structure models. This paper names
them as single channel model (SCM), the dual-channel model of manufacturers
(DCMM) and the dual-channel model of retailers (DCMR) respectively. First,

this paper analyzes the optimal product pricing and strategy for the manufac-
turer and the retailer under three different channel models. Then we analyze the
differences among the three models and their effect on the local supply chain.
At last, this paper uses the numerical examples to demonstrate the conclusions.
2 Model Assumptions and Notations Description
Assumed that the manufacturer only produces a kind of product with network
externality. Considering that the sale period is not too long, consumers only buy
the goods from the manufacturer or the retailer on the channel and there is no
intersection between the consumption of different channels.
The retailer is the leader in the structure in all models. In the single channel
model, firstly, according to the retailer’s own marginal profit m1 , the manu-
facturer decides the wholesale price ω1 to maximize their own profits. Then,
the retailer adjusts its revenue to maximize its own marginal profit according
to the wholesale price. Further, the retailer sets the sales price of the product
p1 = ω1 + m1 . In the dual-channel model of manufacturers, firstly, according to
the retailer’s own marginal profit m2 , the manufacturer decides the wholesale
price ω1 and the sale price p22 on online channel to maximize their own profits.
Second, the retailer adjusts its revenue to maximize its own margin according
to the wholesale price. At last, the retailer sets the sales price of the product.
Similar to the dual-channel model of manufacturers, in the dual-channel model
of retailers, the manufacturer decides the wholesale price and the sale price ω2
on online channel p22 . The main difference is that the retailer needs to decide
the sale price on online channel p22 .
Consumers can buy goods through different channels, but they will have
different utilities and the estimated value of the product on offline or online
channels will also be different. For example, In the physical store, consumers
can immediately get product information and then purchase products. While
through the network channel, consumers can get more products no matter how
far it is. With the development of modern logistics, waiting time will become very
short. So we denote the degree of substitution or monopoly between online and
offline channels by β (0 ≤ β < 1), so 1 − β indicates the degree that the online
channel cannot monopolize. While refers to the manufacturer’s credit tendency
in the mind of the consumer, so does the 1 − γ.
In the single channel model, the consumer’s estimation value of the product
on the offline channel is V11 ; in the dual-channel model of manufacturers, the
consumer’s estimation value of the product on the offline and online channels are
V21 , V22 ; in the dual-channel model of retailers, the consumer’s estimation value
of the product on the offline channel is V31 , while the consumer’s estimation
value of the manufacturer’s and the retailer’s product on the online channel are
V32 , V33 .
The consumer will select the channel only when their utilities are greater
than zero. Without loss of generality, assume the customer is heterogeneous
456 J. Yan et al.
and subject to uniform distribution. In model one, θ11 refers to the consumers’
preference distribution on offline channel. It represents the consumer’s measure
for each potential transaction; In model two, we use θ21 instead of θ11 and denote
the consumer’s preference distribution for manufacturers of online channel by
θ22 . In model three, we use θ31 , θ32 instead of θ21 , θ22 and denote the consumer’s
preference distribution for manufacturers of online channel by θ32 .
The sales cost of unit product for the manufacturer on offline channel can be
interpreted by c1 , while for the retailer is c2 . Because the cost of online channel
is lower than the offline channel, so define k1 as the manufacturer’s allocation
efficiency relative to the offline channel, while for the retailer is k2 . The smaller
the k, the lower the cost. To be more representative, we can get k ∈ (0, 1).
α1 expresses the strength of network effects of consumers who select retailer’s
offline channel; α2 expresses the strength of network effects of consumers who
select manufacturer’s online channel; α3 expresses the strength of network effects
of consumers who select retailer’s online channel. Without loss of generality,
α ∈ (0, 1).
In the single channel model, A11 represents the largest potential market
demand on offline channel, Nr1 denotes the number of retailers on offline chan-
nel. In the dual-channel model of manufacturers, use A21 , Nr2 instead of A11 ,
Nr1 , A22 represents the largest potential market demand on online channel, Ns2
denotes the number of manufacturers on online channel. Due to the widening of
the market, we can conclude A21 + A22 > A11 . At the same time, because the
manufacturer adds the online channel, and appeal some consumers, so it will
inevitably lead to a decrease in the number of retailers. So Nr2 ≤ Nr1 . In the
dual-channel model of manufacturers, similar to model two, A31 can instead of
A21 ; A32 represents the largest potential market demand for the manufacturer
on online channel; A33 represents the largest potential market demand for the
retailer on online channel; denotes the number of retailers on online channel;
Ns3 denotes the number of manufacturers on online channel; Nr4 denotes the
number of retailers on offline channel; Similarly, A32 + A33 > A22 , Ns3 ≤ Ns2
and Nr4 ≤ Nr2 .
3 The Model
3.1 The Single Channel Model (SCM)
Based on the above assumptions, we study the single channel model firstly. The
supply chain mode is shown in Fig. 1.
Fig. 1. The single channel mode (SCM)

Assume that the utility function of the consumer on the single channel is:
U11 = V11 + Nr1 θ11 α1 − p11 . (1)
According to Eq. (1), for this channel, the consumer’s individual rationality
(participation) constraint is:
U11 ≥ 0. (2)
So
p11 − V11
θ11 ≥ . (3)
Nr1 α1
As a result, according to Eq. (3), the number of consumers who join the
retailer’s offline channel is:
1
(Nr1 α1 − p11 + V11 )A11
D11 = A11 dθ = . (4)
p11 −V11
N α
Nr1 α1
r1 1
And it must satisfy:
D11 ≥ 0. (5)
(1) The manufacturer’s profit is:
Nr1 α1 − ω1 − m1 + V11
ΠM 1 = (ω1 − c1 )D11 = A11 (ω1 − c1 ) . (6)
Nr1 α1
Proposition 1. The manufacturer’s profit function is a concave function of the
wholesale price, so when the manufacturer’s profit reaches the maximum, the
price’s optimal value is:
Nr1 α1 − m1 + c1 + V11
ω1∗ = . (7)
2
(2) The retailer’s profit of offline platform is:
Nr1 α1 − m1 − c1 + V11
ΠR1 = (p11 − ω1 − c2 )D11 = A11 (m1 − c2 ) . (8)
2V11 α1
Proposition 2. The retailer’s profit function is a concave function of the earn-
ing of unit product on offline channel, so when the retailer’s profit reaches the
maximum, the margin’s optimal value is:
Nr1 α1 − c1 + c2 + V11
m∗1 = . (9)
2
458 J. Yan et al.
In the single channel model, the final optimal values for these parameters are
as following:
Nr1 α1 + 3c1 − c2 + V11
ω1∗ = , (10)
4
3Nr1 α1 +c1 + c2 + 3V11
p∗11 = , (11)
4
2
∗ (Nr1 α1 − c1 − c2 + V11 ) A11
ΠM 1
= , (12)
16Nr1 α1
2
∗ (Nr1 α1 − c1 − c2 + V11 ) A11
ΠR1 = , (13)
8Nr1 α1
2
3(Nr1 α1 − c1 − c2 + V11 ) A11
Π1∗ = ΠM
∗
1
∗
+ ΠR1 = . (14)
16Nr1 α1
3.2 The Dual-Channel Model of Manufacturers (DCMM)

Based on the above assumptions and the first model, we study the dual-channel
model of manufacturers next. The supply chain mode is shown in Fig. 2.
Fig. 2. The dual-channel model of manufacturers (DCMM)
As consumers have their own preferences for online and offline channels,
assume that the utility function of the consumer on the offline channel and
online channel respectively are:
U21 = V21 + Nr2 θ21 α1 − p21 − (1 − β)V21 V22 , (15)
U22 = V22 + Ns2 θ22 α2 − p22 − βV21 V22 . (16)
On the basis of Eqs. (15) and (16), the consumer’s individual rationality
(participation) constraints for these two channels are:
U21 ≥ 0, (17)
U22 ≥ 0. (18)
So
p21 + (1 − β)V21 V22 − V21
θ21 ≥ , (19)
Nr2 α1
p22 + βV21 V22 − V22
θ22 ≥ . (20)
Ns2 α2
As a result, according to Eqs. (19) and (20), the numbers of consumers who
join retailer’s offline channel and online channel respectively are:
Nr2 α1 − p21 − (1 − β)V21 V22 + V21

D21 = A21 , (21)
Nr2 α1
Ns2 α2 − p22 − βV21 V22 + V22
D22 = A22 . (22)
Ns2 α2
And they must satisfy:
D21 ≥ 0, (23)
D22 ≥ 0. (24)
The manufacturer decides the wholesale price and the online channel’s price
of unit product to maximize their own profits. The manufacturer’s profit is:
ΠM 2 = (ω2 − c1 )D21 + (p22 − k1 c1 )D22 . (25)
Proposition 3. The manufacturer’s profits function is a concave function of

the wholesale price and the online channel’s price of unit product, so when the
manufacturer’s profit reaches the maximum, the optimal values of the prices are:
Ns2 α2 − βV21 V22 + k1 c1 + V22
p∗22 = , (26)
2
Nr2 α1 − (1 − β)V21 V22 − m2 + c1 + V21
ω2∗ = . (27)
2
At the same time, the retailer’s decision-making is the same as the single
channel model. The retailer’s profit of offline platform is:
ΠR2 = (p21 − ω2 − c2 )D21

m2 − c2 Nr2 α1 − (1 − β)V21 V22 − m2 − c1 + V21
= A21 . (28)
2 Nr2 α1
Proposition 4. The retailer’s profits function is a concave function of the earn-
ing of unit product on offline channel, so when the retailer’s profit reaches the
maximum, the margin’s optimal value is:
Nr2 α1 − (1 − β)V21 V22 − c1 + c2 + V21

m∗2 = . (29)
2
So in the dual-channel model of manufacturers, the final optimal values for
these parameters are as the following:
Nr2 α1 − (1 − β)V21 V22 + 3c1 − c2 + V21

ω2∗ = , (30)
4
3Nr2 α1 − 3(1 − β)V21 V22 + c1 + c2 + 3V21
p∗21 = , (31)
4
460 J. Yan et al.
Ns2 α2 − βV21 V22 + k1 c1 + V22

p∗22 = , (32)
2
2
∗ [Nr2 α1 − (1 − β)V21 V22 − c1 − c2 + V21 ]
ΠM 2
=A21
16Nr2 α1
2
(Ns2 α2 − βV21 V22 − k1 c1 + V22 )
+ A22 , (33)
4Ns2 α2
2
∗ [Nr2 α1 − (1 − β)V21 V22 − c1 − c2 + V21 ]
ΠR2 =A21 . (34)
8Nr2 α1
Finally, consider the prerequisite: the manufacturer’s profit in this model
should be greater than in model one (Because only in this condition, the manu-
facturer will choose to introduce online channel). So for the manufacturer, their
optimal profit must be satisfied:
ΠM 2 ≥ Π M 1 . (35)
2 2
Let (Nr1 α1 −c 1 −c2 +V11 ) A11
16Nr1 α1 = C1 , A21 [Nr2 α1 −(1−β)V21 V22 −c1 −c2 +V21 ]
16Nr2 α1 = C2 and
βV21 V22 − k1 c1 + V22 = C3 . So according to formula (35), we can get:

4(C1 −C2 ) 2
A22 + 2Ns2 C3 + [ 4(CA1 −C 2)
+ 2Ns2 C3 ] − 4Ns2 2 C
3
α2 ≥ 22
2 . (36)
2Ns2
3.3 The Dual-Channel Model of Retailers (DCMR)
In the end, we analyze the dual-channel model of retailers. According to the

above analysis, the supply chain mode is shown in Fig. 3.
Fig. 3. The dual-channel model of retailers (DCMR)
In the case of this model, consumers have not only different choices of chan-
nels, but also preferences for the entities in the channels. Therefore, the net util-
ities to a consumer on the retailer’s offline platform, the manufacturer’s online
platform, the retailer’s online platform are given by:
U31 = V31 + Nr4 θ31 α1 − p31 − (1 − β)V31 (V32 + V33 ), (37)

U32 = V32 + Ns3 θ32 α2 − p32 − βγV31 (V32 + V33 ). (38)
Similarly, the numbers of consumers who join retailer’s offline channel, man-
ufacturer’s online channel and retailer’s online channel respectively are:
Nr4 α1 − p31 − (1 − β)V31 (V32 + V33 ) + V31

D31 = A31 , (39)
Nr4 α1
Ns3 α2 − p32 − βγV31 (V32 + V33 ) + V32
D32 = A32 , (40)
Ns3 α2
Nr3 α3 − p33 − β(1 − γ)V31 (V32 + V33 ) + V33
D33 = A33 . (41)
Nr3 α3
According to the actual situation, they must satisfy:
D31 ≥ 0, (42)
D32 ≥ 0, (43)
D33 ≥ 0. (44)
Similar to the model two, the manufacturer decides the wholesale price and
the online channel’s price of unit product to maximize their own profits. The
manufacturer’s profit in this model is:
ΠM 3 = (ω3 − c1 )D31 + (p32 − k1 c1 )D32 . (45)
Proposition 5. The manufacturer’s profits function is a concave function of

the wholesale price and the online channel’s price of unit product, so when the
manufacturer’s profit reaches the maximum, the optimal values of these prices
are:
Ns3 α2 − βγV31 (V32 + V33 ) + k1 c1 + V32
p∗32 = , (46)
2
Nr4 α1 − (1 − β)V31 (V32 + V33 ) − m3 + c1 + V31
ω3∗ = . (47)
2
Different from the model two, the retailer not only decides the earning of unit
product on offline channel, but also the online channel’s price of unit product.
The retailer’s total profit in this model is:
ΠR3 = (p31 − ω3 − c2 )D31 + (p33 − k2 c2 )D33 . (48)
Proposition 6. The retailer’s profits function is a concave function of the earn-

ing of unit product on offline channel and the online channel’s price of unit
product, so when the retailer’s profit reaches the maximum, the optimal values
are:
Nr4 α1 − (1 − β)V31 (V32 + V33 ) + c2 − c1 + V31
m∗3 = , (49)
2
Nr3 α3 − β(1 − γ)V31 (V32 + V33 ) + k2 c2 + V33
p∗33 = . (50)
2
462 J. Yan et al.
Therefore, in the dual-channel model of retailers, the final optimal values for
these parameters are as the following:
Nr4 α1 − (1 − β)V31 (V32 + V33 ) − c2 + 3c1 + V31
ω3∗ = , (51)
4
3Nr4 α1 − 3(1 − β)V31 (V32 + V33 ) + c2 + c1 + 3V31
p∗31 = , (52)
4
Ns3 α2 − βγV31 (V32 + V33 ) + k1 c1 + V32
p∗32 = , (53)
2
Nr3 α3 − β(1 − γ)V31 (V32 + V33 ) + k2 c2 + V33
p∗33 = , (54)
2
2
∗ [Nr4 α1 − (1 − β)V31 (V32 + V33 ) − c1 − c2 + V31 ]
ΠM 3
=A33
16Nr4 α1
2
[Ns3 α2 − βγV31 (V32 + V33 ) − k1 c1 + V32 ]
+ A34 , (55)
4Ns3 α2
2
∗ [Nr4 α1 − (1 − β)V31 (V32 + V33 ) − c1 − c2 + V31 ]
ΠR3 =A33
8Nr4 α1
2
[Nr3 α3 − β(1 − γ)V31 (V32 + V33 ) − k2 c2 + V33 ]
+A35 . (56)
4Nr3 α3
Finally, consider the prerequisite: the manufacturer’s profit in this model
should be greater than in model one and the retailer’s profit in this model should
be greater than in model two. So for the manufacturer, their optimal profit must
be satisfied:
ΠM 3 ≥ Π M 1 . (57)
2
Let A33 [Nr4 α1 −(1−β)V3316N
(V34 +V35 )−c1 −c2 +V33 ]
r4 α1
= C4 and βγV31 (V32 + V33 ) +
k1 c1 − V32 = C5 . So:

2
4( C1A−C
32
4
) + 2N s3 C 5 + [4( C1A−C
32
4
) + 2Ns3 C5 ] − 4Ns3
2 C
5
α2 ≥ 2 . (58)
2Ns3
At last, for the retailer, their optimal profit must be satisfied:
Let β(1 − γ)V31 (V32 + V33 ) + k2 c2 − V33 = C6 . According to formula (57)

2
8( C2A−C
33
4
) + 2N r3 C6 + [8( C2A−C
33
4
) + 2Nr3 C6 ] − 4Nr3
2 C
6
α3 ≥ 2 . (59)
2Nr3
Comparing the three models and when these prerequisites are met, we can
get the following conclusions. With manufacturers and retailers choose the online
channel in turn, it can be seen that the price of unit product on offline channel
gradually reduces. In model three, the online channel’s price of unit product
for the manufacturer and the retailer both lower than the online channel’s price
in model two. With the introduction of online channel, the wholesale price is
gradually reduced and the number of consumers who join retailer’s offline channel
will decrease. As the retailer chooses the online channel, the number of consumers
who join manufacturer’s online channel will also decrease, but the total number
of consumers will increase. Compared to the model one, when the manufacturer
chooses the online channel, their profit gets to the maximum, then when the
retailer uses online channel, the manufacturer’s profit reduced, but is higher
than in the model one. In model two, when the manufacturer chooses the online
channel, the retailer’s profit gets to the minimum, then when they choose the
online channel, their profits will increase and higher than in model one. Along
with the introduction of the online channel, the overall profit of the channel will
increase.
4 Numerical Example
This section presents numerical simulations to clarify the findings which is not
easy to understand. First, we maintain the following assumptions over the simu-
lations. For the case of the single channel model, in order to achieve profitability,
some variables need to meet some restrictions: A11 ≥ p11 ≥ ω1 ≥ c1 . In model
two, in order to prevent retailers ordering goods from online channels and be
able to make profit, they need to meet A21 ≥ p21 ≥ ω2 ≥ c1 , p22 ≥ ω2 ≥ c1 and
A22 ≥ p22 ≥ k1 c1 . Similar to model two, in model three, A31 ≥ p31 ≥ ω3 ≥ c1 ,
p32 ≥ ω3 ≥ c1 , A32 ≥ p32 ≥ k1 c1 , p33 ≥ ω3 ≥ c1 and A33 ≥ p33 ≥ k2 c2 .
According to formula A11 ≥ p11 ≥ ω1 ≥ c1 , A21 ≥ p21 ≥ ω2 ≥ c1 , A31 ≥
p31 ≥ ω3 ≥ c1 and formulas (2), (21), (22), (42), (43) and (44), the prerequisite is:
c1 + c2 + (1 − β)V31 (V32 + V33 ) − V31

α1 ≥ . (60)
Nr4
According to formula p22 ≥ ω2 ≥ c1 , A22 ≥ p22 ≥ k1 c1 , p32 ≥ ω3 ≥ c1 ,
A32 ≥ p32 ≥ k1 c1 and formulas (21), (22), (36), (42), (43), (44) and (57), the
prerequisite is:
Nr4 α1 + (β − 1 + 2βγ)V31 (V32 + V33 ) + (3 − 2k1 )c1 − c2 + V31 − 2V32
α2 ≥ .
2Ns3
(61)
According to formula p32 ≥ ω3 ≥ c1 , A32 ≥ p32 ≥ c1 , p32 ≥ p32 ≥ k1 c1

and formulas (42), (43) and (44), and controlling other variables constant, the
prerequisite is:
2Nr3 α3 − Nr4 α1 + V31 (V32 + V33 ) + (2k2 + 1)c2 − 3c1 + 2V33 − V31
β≤ . (62)
(3 − 2γ)V31 (V32 + V33 )
So according to previous assumptions, these parameters’ values can be set:
A11 = 16, V11 = 0.35, α1 = 0.7, c1 = 0.5, c2 = 0.3, Nr1 = 10, A21 = 14, A22 = 13,
V21 = 0.5, V22 = 0.37, α2 = 0.6, β = 0.7, k2 = 0.5, Nr2 =, Ns2 = 0.6, A32 = 10,
A31 = 13, A32 = 12.5, V31 = 0.6, V31 = 0.6, V32 = 0.4, V33 = 0.5, γ = 0.6,
464 J. Yan et al.
k2 = 0.1, α3 = 0.5, Ns3 = 5, Ns3 = 5.5 and Nr4 = 7. According to formulas

(60), (61) and (62), and controlling other variables constant, these prerequisites
are 0.052 ≤ α1 ≤ 1, 0.569 ≤ α2 ≤ 1, and 0 ≤ β ≤ 0.782. So the prices of unit
product on different channels can be shown in Fig. 4.
Fig. 4. The prices of unit product
As shown in Fig. 4, under the premise that manufacturers and retailers

can choose the online channel, we can conclude that the price of the retailer’s
unit product on offline channel gradually decrease; the manufacturer’s and the
retailer’s online channel’s prices of unit product are also lower than in single
channel model. That means consumers are very favorable when the manufac-
turer and the retailer both choose the online channel.
Fig. 5. The wholesale price of unit product

Fig. 6. The number of consumers
Fig. 7. The number of consumers
Compared the three models, the wholesale price of unit product is gradually
reduced with the introduction of online channel. It’s benefit for the retailer
(Figs. 6 and 7).
With the increase of the network externality, the number of consumers who
join the offline and online channels are all increasing in the three modes. Com-
pared the three modes, due to the introduction of online channel, some con-
sumers transfer to online channel. So the number of consumers who join the
offline channel will decrease. When the retailer chooses to introduce the online
channel, the number of consumers who join the manufacturer’s online channel
will also decrease. As shown in Fig. 7, the total number of consumers is increas-
ing with the increase of network externalities and decreasing with the degree of
substitution or monopoly between online and offline channels. With the introduc-
tion of online channel, the total number of consumers is increasing in turn. The
466 J. Yan et al.
Fig. 8. The profit
total number of consumers can increase with the introduction of online channel.
At the same time, consumers have more channels to choose, manufacturers and
retailers have more demand for products, and thus these behaviors can increase
consumption and create more social values.
We can see from Fig. 8, with the increase of the network externalities all
profits are increasing. In model one, the profit of the manufacturer is the lowest.
In model two, as the manufacturer chooses the online channel their profits reach
the maximum. In model three, as the retailer chooses the online channel, their
profits somewhat reduce but higher than in model one. On the contrast, in model
one, the profit of the retailer is lower than in model three but higher than model
one. In model two, as the manufacturer chooses the online channel, their profits
reach the minimum. In model three, as the retailer chooses the online channel,
their profits increase and reach the maximum. The total profit is increasing with
the introduction of online channel. That means when manufacturers and retailers
both choose to use online channel, they will create more social values.
5 Conclusions
This paper considers the choice of the retailer based on the choice of the manufac-
turer under the network externalities. We focus on the influence when the entities
have different choices. By contrasting the single channel model, the dual-channel
model of manufacturers and the dual-channel model of retailers, the introduction

of online channel is benefit to the local supply chain.
Our study of channel performance under the affecting of the network exter-
nalities has yielded several important insights that should be of managerial inter-
est. There are certain conditions for the manufacturer’s and retailer’s selection
of the online channel. When manufacturers or retailers introduced the online
channel first, the other one will make the same choice. With the introduction of
online channel, the whole price will decrease but the number of consumers and
the total profit will increase. That is why the dual-channel model of retailers is
so widespread.
In developing our models, there are a few simplifying assumptions. For exam-
ple, consistent with many literatures, we only study a single-period model and
assume the consumer can only choose one channel. While we believe the intu-
ition of our results applies to cases with different assumption, future research
might investigate how such assumptions (e.g., independent of retailers’ utility)
affect the conditions in the proposition. Finally, future research could divide the
channel into monopoly channel and competitive channel, so we can further dis-
cuss the choice of the manufacturer and the retailer between two channels or
among three channels. There also might conduct a comprehensive analysis of
manufacturers’ or retailers’ decisions which include different first-party content,
technology and market share.
Acknowledgements. This research is funded by The Soft Science Research Project

of Sichuan Province (No. 2016ZR0042).
References
1. Ai-Min YU, Liu LW (2012) Competing and coordination strategies for the dual
channel under stochastic demand and cooperative promotion. J Ind Eng Eng
Manag 26(1):151–155
2. Anderson EG, Parker GG, Tan B (2014) Platform performance investment in the
presence of network externalities. Inf Syst Res 25(1):152–172
3. Bajwa N, Sox CR et al (2016) Coordinating pricing and production decisions for
multiple products. Omega 64:86–101
4. Blocher J (2016) Network externalities in mutual funds. J Finan Mark 30:1–26
5. Chen J, Zhang H, Sun Y (2012) Implementing coordination contracts in a manu-
facturer stackelberg dual-channel supply chain. Omega 40(5):571–583
6. Chen W, Kucukyazici B et al (2016) Supply chain design for unlocking the value
of remanufacturing under uncertainty. Eur J Oper Res 247(3):804–819
7. Dye CY, Yang CT (2016) Optimal dynamic pricing and preservation technology
investment for deteriorating products with reference price effects. Omega 62:52–67
8. Hagiu A, Spulber D (2013) First-party content and coordination in two-sided mar-
kets. Manag Sci 59(4):933–949
9. Leider S (2016) Bargaining in supply chains. Manag Sci 62:3039–3058
10. Liu X (2016) Contracting for competitive supply chains under network externalities
and demand uncertainty. Discrete Dyn Nat Soc 2016(1):1–9
11. Mantena R, Saha RL (2012) Co-opetition between differentiated platforms in two-
sided markets. J Manag Inf Syst 29(2):109–140
468 J. Yan et al.
12. Panda S, Modak NM et al (2015) Channel coordination and profit distribution in

a social responsible three-layer supply chain. Int J Prod Econ 168:224–233
13. Shamir N (2015) Pricing strategy in dual channel under the leadership of retailers.
Chin J Manag Sci 6(6):91–98
14. Shamir N (2016) Public forecast information sharing in a market with competing
supply chains. Manag Sci J Inst Oper Res Manag Sci 62:2994–3022
15. Xie Y, Tai AH et al (2016) Pricing strategy for a two-echelon supply chain with
optimized return effort level. Int J Prod Econ 182:185–195
16. Yang H, Zhuo W et al (2016) Two-period supply chain with flexible trade credit
contract. Expert Syst Appl 66:95–105
17. Yao Z, Xu X, Luan J (2016) Impact of the downside risk of retailer on the supply
chain coordination. Comput Ind Eng 12(102):340–350
Construction of Restoration System for Old
Books Written in Braille
Yuko Shimomura1 , Hiroyuki Kawabe1 , Hidetaka Nambo2(B) , and Shuichi Seto3

1
Kinjo University, Ishikawa, Japan
2
Kanazawa University, Kanazawa, Japan
nambo@blitz.ec.t.kanazawa-u.ac.jp
3
Kinjo College, Ishikawa, Japan
Abstract. Braille books follow a process of deterioration different from

ordinary printed books. Braille is a reading letter and visually impaired
people touch the boss with the finger’s belly. Therefore, in Braille books
that are frequently read, Braille is dirty with aged use, holes open, and
they collapse. In old Braille books, braille collapses and deteriorates due
to pressure from left and right books in the library. Therefore, in this
research, we convert Braille books into machine-readable electronic data.
First, a Braille book is scanned by an image. Next, we detect braille
by image recognition technology one by one from braille page. And we
classify and identify Braille. Furthermore, we correct errors such as mis-
detection of braille and misidentification. Finally, we save the result as
character code.
Keywords: Braille · Old boo · Restoration · Deep learning
1 Introduction
In this section we briefly explain Braille and degraded braille.
1.1 Braille and Braille Library

The origin of Braille is old. In 1825, Braille [3] invented Braille in France. In
1879, Megata introduced Braille to Japan. In 1890, Ishikawa devised Japanese
braille. This is the beginning of Japanese Braille. Since then Braille has been
used as a character for the blind.
Braille consists of six embossed or flat points. These six points constitute
one cell with two by three shape, which represents one character or a modifier of
63 kinds. With the modifier, characters more than 64 kinds are represented. In
Japanese braille, each of voiceless sounds is represented by a plain cell and other
characters such as syllabic nasals, voiced consonants, semi- voiced consonants,
numerals, alphabets and so on are represented by a modifier and subsequent
cell(s). In Fig. 1, Japanese Braille has a blank space for each clause like English.
470 Y. Shimomura et al.
Some Japanese Braille are constructed with a prefix as last three examples in
Fig. 1. Though Japanese is vertical writing or horizontal one, Japanese braille is
horizontal writing only.
Fig. 1. Japanese Braille
The most primitive tool for writing braille is the slate and stylus. To use it,
the user presses the tip of the stylus down through the small rectangular hole to
make braille dots. Therefore the user is required to punch the braille dots with
mirror image in reverse order. Nowadays, for publishing a book in braille, texts
are input according to the braille grammar by using a braille text editor, are
revised, and are printed in braille, as Fig. 2.
Fig. 2. Normal dot and mirrored one
In 1988, IBM opened “Tenyaku Hiroba (Field of Braille Transcription)”,

that was the Braille information network system, and digitization of braille had
become widespread. The books in braille before the IBM’s project are stored as
they are without being converted into electronic data. Even if they have high
storage value, the braille books are bulky, so braille libraries in various places
are struggling to save. Each library has been forced to decide whether to discard
them in recent years or to leave them.
In the earlier days when the Braille code was still in the research phase and
there was no Windows, no personal computer, no internet, Shimomura et al. [1]
studied coding of Braille.
Braille books are stored in the braille libraries throughout the country. The
braille libraries registered in National Association of Institutions of Information
Service for Visually Impaired Persons have about 100 libraries. Braille books.
Braille books are borrowed from the braille libraries and public libraries via the
braille libraries. In addition, many braille books are provided from the WEB
braille library.
1.2 Degraded Braille
Braille is a tactile reading system. One reads tangible points with a finger’s belly.
Figure 3 is a typical page of Braille book. The braille books that are frequently
Construction of Restoration System for Old Books Written in Braille 471
Fig. 3. Page of Braille book
read are dirty, holes are opened in dots, and dots collapse. Figures 4 and 5 show
fresh dots and a part of a page that both normal cells and mirror images of
cells, respectively. Figures 6, 7, 8 and 9 show the degraded braille cells, where
the collapsed cell, the hole opened one, the dirty one and the distorted one is
presented, respectively. Tears, creases and stains in pages are shown in Figs. 10,
11 and 12, respectively.
Fig. 4. Cell Fig. 5. Normal and mirrored cell
Shimomura et al. [2] also studied restoration of the degraded Braille by

shadow of Braille using fuzzy theory. In that study, the extraction of Braille
was made by hand and computer programs executed the determination of exis-
tence of Braille. In this project, we extract Braille from scanned page images,
which contain the degraded Braille, by computer programs, recognize them and
restore the pages.
Fig. 6. Collapsed cell Fig. 7. Hole opened cell
Fig. 8. Dirty cell Fig. 9. Distorted cell
Fig. 10. Tears Fig. 11. Creases Fig. 12. Stains
2 Restoration System
In this subsection we explain our project of the restoration system for old books
written in Braille.
(1) Page scanning.

(2) Determination of whether scanned dot is normal or mirrored.
(3) Extraction of cells and classification into 63 categories.
(4) Error correction by using a scanning redundancy, that one cell is read twice
as the normal image and the mirror image.
(5) Error correction by using Braille grammar.
(6) Interpretation Braille into Japanese.
(7) Error correction by using Japanese grammar.
As noted above, the system has two main components, a machine learning system
for recognition of Braille and an error correction system.
Fig. 13. Normal dot Fig. 14. Mirrored dot Fig. 15. Background
image image image
2.1 Preparation of Braille Images for Machine Learning

To detect degrade cells by machine, we must prepare many image of degrade
cells. We have already scanned many the old braille books with resolution of 200
dpi, extracted each cell into 54 × 36 pixel by hand, and have obtained about
15000 cell images. These images were classified into 63 normal categories and 63
mirror ones. Next, we cut each call images into 6 dot areas by a tiny application
program and got about 45000 dot images and 45000 flat (non dot) ones. The
followings are the normal image, the mirrored dot one and the background one
for the machine learning (Figs. 13, 14 and 15).
2.2 Dot Recognition by OpenCV
OpenCV is a famous computer vision library and has functions of object detec-
tion. Some detectors such as face, eye, mouth and so on are prepared, and one
also can make detectors for arbitrary objects. To make the object detector, one
prepares many images one wants to detect and execute the machine learning.
In the machine learning, characteristics are extracted from the many images of
the object, and the machine learns the characteristics. The set of images and
characteristics learned by the machine is called a cascade classifier. OpenCV has
some algorithm to make the classifier. In our system, we adopt train the cascade
routine with LBP characteristic to search dots in scanned page.
As the cells are regularly aligned on paper, by using the gradient of dots
or the edge of paper we can correct the gradient of page. The distribution of
detected dots determines row lines in page.
2.3 Scanning of Old Braille Books
It is necessary to treat the old Braille books carefully because of the degradation
of cells in them as mentioned above. Therefore we do not adopt flatbed scanners
but adopt a noncontact type of scanner, Fujitsu ScanSnap SV600, which operates
by taking an elevated view (see Fig. 16), with resolution of 600 dpi and grey scale.
If a Braille book is printed on both side and all pages are scanned, all Braille
printed is scanned twice. They are read as the normal image and the mirror
image. This gives us a redundancy for interpretation of cells.
The next step is the dot detection by OpenCV mentioned above. The result
is given in Figs. 17 and 18.
Fig. 16. ScanSnap SV600
Fig. 17. Dot detection by OpenCV Fig. 18. Close up of detected dots
2.4 Normal and Reverse Dot Classification by Deep Neural

Network
To detect degrade cells by machine, we must prepare many image of degrade

cells. We have already scanned many the old braille books with resolution of 200
dpi, extracted each cell into 54 × 36 pixel by hand, and have obtained about
15000 cell images. These images were classified into 63 normal categories and 63
mirror ones. Next, we cut each call images into 6 dot areas by a tiny application
program and got about 45000 dot images and 45000 flat (non dot) ones.
The deep learning is the one of technique of machine learning based on a
deeply structured and hierarchical neural network. This technique is applied on
the image recognition, the sound recognition and so on. In our project, we use
this technique.
As usual Braille books are printed on both side, we must distinguish between
the normal dots and the mirror ones. The both cell image are give again in
Figs. 19 and 20. We have already obtained many dot images for them, so we use
the images as reference data for the dot classification in the deep neural network.
This procedure gives us a normal dot distribution and a mirrored distribution.
By converting the mirrored distribution into left-side right, we obtain the second
candidate of the normal dot distribution.
Fig. 19. Normal cell Fig. 20. Mirror cell
2.5 Cell Classification

As Braille is regularly aligned, the coordinate of dots tells us the starting point
of cells. From this point, the dot images are sequentially analyzed, recognized
and classified as cells by the deep neural network learned with 15000 cell images.
Figures 20, 21, 22, 23, 24, 25 and 26 are binary aligned cell images. According
to the classification, a character code is assigned to the cell image and is stored
into database with a page number, a line number in the page, a word number in
the line and a character number in the word. If the cell image has no character
code, a flag of scanning error is also stored in the database.
2.6 Error Correction by Scanning Redundancy

When a Braille book is printed on both side and all pages are scanned, all pages
become to be scanned twice. The character codes obtained from the normal
images and ones from mirrored images are stored in the database and the codes
obtained from normal and mirror images, for examples both shown in Figs. 16
and 17, are identical each other. By collating these codes and finding discrepan-
cies, we have possibility to correct the scanning error.
2.7 Error Correction by Braille Grammar

Braille has own grammar. The character codes stored in the database are ana-
lyzed according to the grammar. In Japanese Braille, for example;
Fig. 21. “1xxxxx” cell Fig. 22. “x2xxxx” cell Fig. 23. “12xxxx” cell
Fig. 24. “1234 × 6” cell Fig. 25. “12345x” cell Fig. 26. “123456” cell
A postposition “ha” is translated into “wa”.

There is not the postposition “ha” following numeric characters.
A long sound written with kana character is translated into a macron and so
on. When unacceptable sequence of characters is found, a flag of Braille gram-
matical error is stored in the database. If a probable code can be presumed
by the grammar, the code is also stored as a candidate in the database.
2.8 Error Correction by Japanese Grammar
According to Japanese grammar, we check the sequence of code, set flags of

Japanese grammatical error for ungrammatical expressions and correct the codes
to probable codes if possible. Typical grammatical errors are as follows;
• A sequence of punctuation marks;

• A sequence of contracted sounds;
• Non-correspondent parentheses Incorrect sonant marks.
2.9 Output and Correction by Hand
Finally the codes are converted into ink-spots expressing Braille and are output
(See Fig. 27). For codes with the error flag, the ink-spots and candidates of
Japanese character are output, and are corrected by hand.
Fig. 27. Ink-spots expressing Braille
3 Concluding Remarks
In this paper, we explain our project to restore old Braille books by converting
Braille books into machine-readable electronic data. The Braille book is scanned
by an image. With the machine learning Braille are detected by image recogni-
tion technology, classified and identified. Furthermore, the error corrections are
executed. Finally, the machine-readable character codes are stored.
Shimomura and colleagues have studied Braille 35 years before. The resump-
tion of our study of Braille is because a request by the Japan Braille Library
and Ishikawa Braille Library. In recent years, Braille books have been converted
into electronic data and stored, and printed by computer processing. However
the previous books are not converted into electronic data and are left as Braille
books as it is. There are Braille books where the original is not found and with-
out the original. The request was to convert Braille books into digital data before
they became degraded and they became unreadable. Responding to this request,
we resumed our research. We want to quickly restore degraded Braille books.
Acknowledgement. We would like to thank the staff of the following organizations

for enabling us to research the actual condition of old Braille books: Ishikawa Associ-
ation for Providing Facilities for the Visually Impaired and Japan Braille Library.
References
1. Shimomura Y, Mizuno S (1981) Translation of machine readable catalog/cataloging
into braille and vice versa. Bull Kanazawa Women’s Junior Coll 23:111–117
2. Shimomura Y, Mizuno S, Hasegawa S (1983) Automatic translation of machine read-
able data into braille. The Institute of Electronics, Information and Communication
Engineers, IEICE Technical report ET82-9:57–62
3. National Association for Providing Facilities for the Visually Impaired (2003) Braille
transcription guide, 3rd edn
Enterprise Operation Management
Discrimination of Personal Opinions
from Commercial Speech Based
on Subjectivity of Text
Yoshio Seino1 and Takahiro Hayashi2(B)

1
Graduate School of Science and Technology, Niigata University, 8050,
Ikarashi-2-no-cho, Nishi-ku, Niigata-shi, Niigata 950-2181, Japan
2
Institute of Science and Technology, Niigata University, 8050, Ikarashi-2-no-cho,
Nishi-ku, Niigata-shi, Niigata 950-2181, Japan
hayashi@ie.niigata-u.ac.jp
Abstract. This paper presents a method for discriminating personal

opinions from commercial speech. Today many personal opinions such
as complaints about a particular product can be found on the Web.
Mining useful information from these opinions is important for a wide
range of applications. Personal opinions on the web, however, are often
contaminated with commercial speech. Commercial speech is generated
by companies and individuals for the intent of making a profit. The data
cleaning process that discriminates personal opinions from commercial
speech is important for obtaining useful results in opinion mining. As a
data cleaning method, we propose a language-independent method for
discriminating personal opinions from commercial speech based on sub-
jectivity of text. Assuming that subjective words frequently occur in per-
sonal opinions rather than commercial speech, we define the subjectivity
score of each word. Estimating the total subjectivity score of a given text,
the proposed method identifies whether the given text expresses per-
sonal opinions or commercial comments. From experiments using texts
datasets written in different languages, we have found over 90% of per-
sonal opinions can be correctly discriminated from commercial speech by
the proposed method.
Keywords: Personal opinions · Commercial speech · Data mining ·

Data cleaning · Information filtering
1 Introduction
Gathering personal opinions such as complaints about a particular product is

an important task for a wide range of applications. For example, in marketing,
it helps identifying the potential revenues, judging the success of a new product
launch, improving the quality of customer service, reducing the development cost
of a next product, etc. Today, many personal opinions can be found on websites

482 Y. Seino and T. Hayashi
of individuals, social networking services like Twitter and customer review pages
in E-commerce sites. Owing to the importance of personal opinions on the Web,
many opinion mining techniques have been proposed [3,5–9]. One of the major
topics in opinion mining is to identify the polarity of a given text whether the
expressed opinion in a text is positive, negative, or neutral.
However, texts collected from web sites are often contaminated with com-
mercial speech. Commercial speech is generated by companies and individuals
for the intent of making a profit, which is essentially different from personal
opinions. Therefore, differentiating between personal opinions and commercial
speech is important for obtaining useful results in opinion mining.
In our previous study, we proposed a method for discriminating personal
opinions from commercial speech [1,2]. In the previous study, we assumed that
subjective expressions such as negative meaning words like “bad” and emoticons
like “:)” are more likely to be used for describing personal opinions than commer-
cial speech. Under the assumption, we modeled the total subjectivity score of a
given text for discriminating between personal opinions and commercial speech.
However, in the previous method, the dictionary of subjective expressions was
generated by human labor and it was strongly dependent on the Japanese gram-
mar. It is, therefore, difficult to apply the method to other languages.
In this paper, we propose a language-independent method for discriminating
personal opinions from commercial speech. The proposed method focuses on the
statistical difference of appearance ratio of each word between personal opinions
and commercial speech corpora. Based on the difference, the subjectivity of each
word is estimated. Assuming that words having high subjectivity are frequently
used in personal opinions, the proposed method judges whether a given text is
personal opinions or commercial speech.
The rest of the paper is organized as follows. In Sect. 2, we explain the outline
of related works. In Sect. 3, we describe the details of the proposed method. In
Sect. 4, we explain experiments using corpora written in different languages. In
the section we show that the proposed method can accurately discriminate the
two types of documents independently from languages.
2 Related Works
The process of opinion mining can be divided into three phases: data collection,
polarity analysis and post processing such as visualization of mining results. In
this paper we focus on the data collection phase. In many studies on opinion
mining aims at analyzing the polarity (positive, negative or neutral) of a given
sentence or text [3,5–9]. Text corpora using in these studies are collected from
various sources such as social networking service sites like Twitter and the cus-
tomer review pages of E-commerce sites. These studies implicitly assume that all
the analyzing documents are regarded as personal opinions. However, the texts
gathered from these sources are often contaminated with commercial speech.
There is obvious difference between minds of writers in personal opinions and
Discrimination of Personal Opinions from Commercial Speech 483
commercial speech. Therefore, for people who want to collect real public impres-
sions about subjects, personal opinions should be discriminated from commercial
speech.
In our previous study [1,2], we proposed a method for discriminating personal
opinions from commercial speech. In the previous method, in order to discrim-
inate the two types of documents, we focused on the four kinds of subjective
expressions: negative meaning expressions, sentence-final particles, interjections,
and specific symbols such as emoticons. Measuring the appearance ratios of these
expressions, we defined the subjectivity score of a given text. Using the scoring
model, we judged texts having high subjectivity scores as personal opinions.
However, the four kinds of subjective expressions were selected by human labor
and these words are strongly dependent on the Japanese Grammar. Therefore,
it is difficult to apply the previous method to other languages.
In this paper, we propose a language-independent method, where the dictio-
nary of subjective expressions is automatically generated by analyzing two kinds
of corpora: texts of personal opinions and ones of commercial speech. Based on
statistical difference of appearance ratio of each word between the two corpora,
the subjectivity score of each word is defined. Subjective words frequently occur-
ring in personal opinions can be automatically detected. Since the scoring model
is independently from languages, the proposed method can be applied to any
languages by just chaining corpora.
3 Proposed Method
In order to discriminate personal opinions from commercial speech, the proposed
method focuses on the difference of appearance ratios of words between two text
corpora: corpus Cp and corpus Cn . Here, corpus Cp consists of text documents
describing personal opinions and corpus Cn consists of text documents describing
commercial speech. Using the two corpora, the subjectivity score of word w is
defined as
s (w) = sp (w) − sn (w) , (1)
where sp (w) and sn (w) are defined as
rp (w)
sp (w) = , (2)
rp (w) + rn (w)
rn (w)
sn (w) = , (3)
rp (w) + rn (w)
here, rp (w) and rn (w) are the appearing ratios of word w in the corpus Cp and
Cn , respectively. rp (w) and rn (w) are respectively defined as follows:

d∈C f (w, d)
rp (w) = p
, (4)
d∈Cp w ∈W (d) f (w , d)

d∈Cn f (w, d)
rn (w) =
, (5)
d∈Cn w ∈W (d) f (w , d)
here, W (d) is the set of words occur in document d, and f (w, d) is the frequency
of word w in document d.
Since subjective words like negative-meaning words are more frequently used
in personal opinion rather than commercial speech [1,2], sp (w) is expected to
be larger than Sn (w) if word w is a subjective expression. The subjective scores
of such words are expected to be positive (s(w) > 0). On the other hand, the
subjective scores of words used in commercial speech tend to be negative.
Eliminating extremely-common words from a document is an important pre-
process for text mining because these words give little meaningful information.
In order to detect such common words, the document frequency of each word in
a corpus is calculated. If word w occurs in over 99% documents in both corpus
Cp and corpus Cn , the frequencies fp (w) and fn (w) are set as 0s.
Based on the subjective scores of words, the total subjectivity of a document
d can be calculated as follows:
1
S (d) = s (w) , (6)
|W (d)|
w∈W (d)
where W (d) is the set of words which appear in document d. If the subjectivity
of a document is positive (S(d) > 0), the proposed method judges the text
as personal opinions; if otherwise (S(d) ≤ 0), the method judges the text as
commercial speech.
4 Experimental Results
In order to confirm that the proposed method works independently from lan-
guages and subjects, we conducted 6 kinds of tasks (T1–T6) using different
corpora. As corpus Cp (personal opinions) and corpus Cn (commercial speech)
used in each task, we collected texts from different websites. Table 1 shows the
source websites, the number of documents in each corpus, the language of the
texts, and the topic of the texts. As shown in the table, in task T2, T3 and T4,
the texts in corpus Cp and corpus Cn (personal opinions and commercial speech)
were collected from the same websites. In these websites, documents are catego-
rized into user reviews and professional reviews. In these tasks, the user reviews
were assigned as the texts of corpus Cp (personal opinions) and the professional
reviews were assigned as the texts of corpus Cn (commercial speech). There is
obvious difference between minds of writers of the two types of documents. Pro-
fessional writers tend to emphasize positive comments about subjects for their
sponsors and seldom report negative comments about them. On the other hand,
personal writers report any comments about subjects. For people who want to
collect real public impression about subjects, user reviews would be more useful
than professional comments. Therefore, professional reviews were regarded as
commercial speech in the tasks.
The purpose of the experiments is to confirm whether the texts in corpus
Cp (personal opinions) and the text in corpus Cn (commercial speech) were
Table 1. The details of corpora used in each task
Task Language Topic of documents Sources of corpora

T1 Japanese Product Cp : 1980 Amazon.co.jp
(https://www.amazon.co.jp/)
Cn : 1975 Impress Watch
(http://www.watch.impress.co.jp/)
T2 Japanese Movie Cp : 1496
(http://eiga.com/)
Cn : 1706
(http://eiga.com/)
T3 English Product Cp : 4900 CNET.com
(https://www.cnet.com/)
Cn : 3803 CNET.com
(https://www.cnet.com/)
T4 English Movie Cp : 990 Rotten Tomatoes
(https://www.rottentomatoes.com/)
Cn : 999 Rotten Tomatoes
(https://www.rottentomatoes.com/)
T5 French Product Cp : 1147 Amazon.fr
(https://www.amazon.fr/)
Cn : 953 CNET France
(http://www.cnetfrance.fr/)
T6 French Movie Cp : 692 SensCritique
(https://www.senscritique.com/)
Cn : 3093 Critikat
(http://www.critikat.com/)
correctly classified by the proposed method. In addition, in order to clarify how

the classification with the proposed method will generalize to an independent
dataset, we used the technique of 10-fold cross validation [4]. That is, the texts in
each corpus are divided into 10 subsets. By using 9 subsets as a training set, the
subjectivity score of each word is defined by Eq. (1). Based on the subjectivity
scores obtained with the training set, the classification is conducted for the texts
in the remaining one subset. The training and classification are repeated 10 times
by changing the combination of subsets for the training and test sets. Finally,
the average performance across the 10 rounds is calculated.
For each round, the performance of classification results is evaluated with
the F -value. The F -value is defined as follows:
2P R
F = . (7)
P +R
The F -value is the harmonic mean of the precision P and the recall R. The
precision P and the recall R are defined as
|Drelevant ∩ Dclassified |
P = , (8)
|Dclassified |
|Drelevant ∩ Dclassified |
R= , (9)
|Drelevant |
where Drelevant is the set of relevant texts, i.e., the texts of personal opinions in
the test set, and Dclassified is the set of documents classified as personal opinions
by the proposed method.
Table 2. The performances (F -values) of the proposed method in each task
Task Maximum Minimum Average

T1 0.95 0.93 0.94
T2 0.93 0.88 0.90
T3 0.99 0.98 0.98
T4 0.99 0.97 0.98
T5 1.00 0.99 0.99
T6 0.98 0.96 0.97
The experimental results are shown in Table 2. In the table, the minimal,
maximal and average F-values across the 10 rounds are shown for each task.
From the table, we can confirm that in all the tasks (T1–T6), the high classifi-
cation performances (F̄ ≥ 0.90) were obtained. These results indicate that the
proposed method can discriminate personal opinions from commercial speech
independently from languages and subjects.
Figure 1 shows the distributions of subjectivity scores of the texts in corpus
Cp and corpus Cn in used in each task. As shown in the figure, personal opinions
and commercial speech can be clearly separated by the subjectivity score. These
results indicate the effectiveness of the proposed scoring.
We analyzed frequently occurring words in personal opinions and commercial
speech, respectively. Table 1(a) and (b) shows typical words in personal opinions
and commercial speech, respectively. We have found that the words listed in
Table 1(a) have high subjectivity scores and the words in Table 1(b) have nega-
tive subjectivity scores. As shown in the table, in personal opinions, subjective
expressions, sentence-final particles (in Japanese), emoticons (in Japanese), and
first-person singulars (in English and French) frequently occur. On the other
hand, in commercial speech first-person plurals (in English and French) and
neutral expression such as words just explaining the specification of a product.
While negative-meaning expressions such as “bad” frequently occur in personal
opinions regardless of its language, the appearance ratios of these words are
Fig. 1. The distributions of the subjectivity scores of the texts in corpus Cp (personal
opinions) and corpus Cn (commercial speech) used each task
extremely lower in commercial speech. Sentence-final particles occur at the end

of a sentence expressing linguistic modality. We have found sentence-final par-
ticles have high subjectivity scores, which indicate these words are useful for
differentiating between personal opinions and commercial speech. Sentence-final
particles are used in Asian languages including Japanese and Chinese. Further
experiments applying the proposed method to other languages including Chinese
texts remain as a future work (Table 3).
In the Indo-European languages including English and French, the subjects of
sentences are helpful for differentiating personal opinions and commercial speech.
For example, first-person singular is used in personal opinions but first-person
plural is often used in commercial speech. These characteristics cannot be found
Table 3. Words frequently occurring in personal opinions and commercial speech
in Japanese. This is because in Japanese the subject of a sentence is often absent.

Although there is different usage of words in each language, with the proposed
scoring method, informative words helping for decimation of personal opinions
and commercial speech can be found automatically regardless of languages.
5 Conclusion
In this paper, we have presented a language-independent method for discriminat-

ing personal opinions from commercial speech. From the results of experiments
using corpora written in different languages, we have confirmed that with the
proposed method personal opinions can be effectively separated from commercial
speech. Further experiments using other languages remain as a future work.
Opinion mining contributes to improve various phases in product-life cycle
management such as improving the quality of customer service, creating new
products, identifying the potential revenues, reducing the development cost, etc.
The proposed method can be used as a method for data cleaning processing
of opinion mining. Combining the proposed method with other post processing
techniques in opinion mining such as sentiment analysis and visualization, we
are planning to develop a comprehensive framework for opinion mining.
References
1. Hayashi T, Abe K, Onai R (2008) Retrieval of personal web documents by extract-
ing subjective expressions. In: International conference on advanced information
networking and applications - workshops, pp 1187–1192
2. Hayashi T, Abe K, Roy D, Onai R (2009) Discrimination of personal web pages by
extracting subjective expressions. Int J Bus Intell Data Min 4(1):62–77
3. Khan FH, Bashir S, Qamar U (2014) Tom: Twitter opinion mining framework using
hybrid classification scheme. Decis Support Syst 57(3):245–257
4. Kohavi R (1995) A study of cross-validation and bootstrap for accuracy estimation
and model selection. In: International joint conference on artificial intelligence, pp
1137–1143
5. Rao S (2016) A survey on sentiment analysis and opinion mining. In: International
conference on advances in information communication technology & computing,
p 53
6. Ravi K, Ravi V (2015) A survey of opinion mining and sentiment analysis. J Knowl
Based Syst 89(C):14–46
7. Selvam B, Abirami S (2013) A survey on opinion mining framework. Int J Adv Res
Comput Commun Eng 3(9):3544–3549
8. Vinodhini G, Chandrasekaran RM (2012) Sentiment analysis and opinion mining:
a survey. Int J Adv Res Comput Sci Softw Eng 2:282–292
9. You Q (2016) Sentiment and emotion analysis for social multimedia: Methodologies
and applications. In: ACM on multimedia conference, pp 1445–1449
Shift in the Regional Balance of Power from
Europe to Asia: A Case Study of ICT Industry
Zahid Latif1(B) , Jianqiu Zeng1 , Shafaq Salam1 , Zulfiqar Hussain1 , Lei Wang1 ,
Nasir Jan2 , and Muhammad Salman3
1
School of Economics and Management, Beijing University of Posts and
Telecommunications, Beijing 100876, People’s Republic of China
zahid25latif@yahoo.com
2
School of Economics, Beijing Normal University, Beijing 100088,
3
Department of Economics, Gomal University, Khyber Pakhtunkhwa, Pakistan
Abstract. During the last couple of decades, ICT sector became the
most innovative service sector that affected the living standards of human
beings all over the world. In the beginning of the 21st century, some of the
Asian countries made reforms in the ICT sector and spent an enormous
amount for the progress of this sector. On the other hand, developed
countries in the European Union (EU) faced different crises which badly
affected the dissemination of this sector. Consequently, EU countries lost
their hegemony in the field of information technology and resultantly,
some of the emerging Asian countries like China, India, and South Korea
got supremacy over the EU in this field. Currently, these countries have
a strong IT infrastructure, R&D sector, IT research centers working for
the development of ICT. Moreover, this paper investigates reasons for
the shifting of the balance of digital power from Europe to Asia.
Keywords: Information and communication technology (ICT) ·

Infrastructure · R&D · One belt · One road · Game changer · Supremacy
1 Introduction
The global economy is facing many challenges as advanced economies are declin-
ing and emerging economies continue to expand their grounds. The level and
duration of these developments are unprecedented. Also, recent developments in
the field of information and telecommunication technology (ICT) such as cloud
computing affects the trend of economies. We can define the trend as a shift in
the regional balance of power. As the world’s economic balance of power is shift-
ing quickly, this trend has been accelerated due to the global recession. China
overtakes the United States as the world’s largest economic power and India
enters the race of these two countries as another big economy. After the great
Shift in the Regional Balance of Power from Europe to Asia 491
global recession, the world’s balance of economic power in term of gross domes-
tic product (GDP) was gradually shifted to the South and East [9]. Currently,
the western industrialized countries are trying to get the growth momentum
after the recession tragedy but they haven’t fully recovered yet. On the other
hand, the developing nations including Asian countries suffered comparatively
less than the industrial countries and recovered rapidly after the recession. In the
next years, the developing Asian countries accelerated their industrial growth,
particularly in information and communication sector [12,25].
A steep decline has seen with the occurrence of 9/11, which resulted in the
steady attrition of economic certainty that finally concluded with the great
geopolitical obstruct of the western financial crisis [1,7]. It has now become
the truth of the 21st century that the western world (Europe) is rapidly los-
ing its superiority in the field of ICT and is replaced by a new international
system shaped by geographical entity known as Asia [7]. Asia has the greater
most number of internet users around the world and it covers almost half of
the whole world. On the basis of internet usage and population statistics, Asia
covers 45.7% of the world internet users, while rest covers the remaining 54.3%
(See Table 1 and Fig. 1).
Table 1. Analysis of internet usage in Asia and rest of the world
Region Population % Population Internet users Penetration % Internet

Est. 2016 of the world on 30/06/2016 (% Population) users
Asia 4,052,652,889 55.20% 1,846,212,654 45.60% 50.20%
Rest of world 3,287,506,603 44.80% 1,829,612,159 55.70% 49.80%
World total 7,182,406,565 100% 3,035,749,340 42.30% 100%
Source: Internet World Stats, (www.internetworldstats.com/stats.htm)
Fig. 1. Internet users in the world by regions

492 Z. Latif et al.
The above analysis shows the internet penetration and usage in Asia and
rest of the world. The figures clearly describe the shift of technology towards
Asia. Although due to low GDP and poverty, many people in Asian countries do
not have access towards internet but the figures show richer tendency for these
countries. China has one of the biggest communication network in Asia, as well
as in the whole world. The ICT semiconductor industry has become one of the
key industry of People Republic of China and has a huge potential to do business
in ICT. This industry has undergone a rapid change of growth and development
during the past decade [15]. The ICT sector in Asia includes goods and services
that process, transmit or receive information. It includes technologies such as
hardware, software, computer services, microelectronics, e-learning, e-business,
e-health, multimedia, as well as emerging technologies such as photonics, fixed
and mobile network convergence, life sciences, environmental sciences, Internet
of Things, mobile internet, cloud computing and digital imaging [26].
2 Research Background
2.1 Literature Review
According to realist approach, the potential for clashes among great powers is
an old story in the international system [4,8,19]. European Commission research
scholars in EU Monthly Magazine predicted about the multi-polar world in 2025.
It described that it is likely for the world to become truly multi-polar and reflect-
ing the new balance of power, and the loss of US leadership. It further predicted
that if the US remains the first military power, the scientific and technological
advancement of some Asian states in new irregular war tactics like cyber-war
and cyber-attacks will weaken the hegemony of US in information and communi-
cation technology [4,6,11,23]. Some authors give much importance to ICT even
four decades before and predicted that the rise of a post-industrial society would
be information centered and marked by a shift from production to service job
[18,27,31].
The world’s economic balance of power is shifting rapidly from North to
South, and the trend has been accelerated by the global recession [1,9,30].
Kenichi Ohmae developed a concept of a “Triad”, he predicted that the world
economy including the information technology will be led by the United States,
Japan, and the European Union [13]. But in present circumstances, the concept
of Kenichi seems to have become concealed by a new order consisting of China,
United States, and India in the field of ICT. Uri Dadush presented a model in a
chapter of book “Handbook of Emerging Economies”, in which he argued that
the shift of technological balance from North to South is increased after the
global recession. He considered technology as the most important element than
the other factors for economic growth of both groups [16,24].
2.2 Conceptual Framework of ICT

The conceptual framework can be observed from the Fig. 2 which classifies eco-
nomic and social benefits of ICT as well as the poverty outcomes. It clarify the
concept that how ICT can promote economic and social benefits in a society.
Fig. 2. A conceptual framework of ICT and development
The diffusion of ICTs, like other innovative technologies, proceed with three
particular stages introduction, growth and maturity [17]. The ICT skills and
infrastructure vary from country to country and is analyzed according to these
three stages.
Fig. 3. ICT diffusion at three stages
Figure 3 shows that if a country developed, not only the mass of ICT users
and practitioners increased but the necessary infrastructure for ICT must be
494 Z. Latif et al.
developed. During the growth stage the diffusion rate is fast and is continuous
until the maturity has been attained. The conceptual model of ICT diffusion
will able the government and policy makers to adjust their goals and strategies
in order to move from one stage to another stage. The proposed model will help
the policy makers to improve the ICT skills and infrastructure in a way that it
will reach to the next stages of the model.
3 A Comparative Analysis of Emerging Asian Economies

and Europe
The recent development in ICT in the Asian region has accelerated the GDP
growth of emerging Asian countries. Furthermore, the Asian region receives mas-
sive investments in the field of ICT, including both local and foreign investments.
Most of the Asian countries attract massive investment through the sector of
ICT. In this regard, the People Republic of China is on the top among the other
emerging states followed by Russia, Brazil, and India. China is the biggest sup-
plier of ICT equipment, while India has developed IT city called Silicon City
(Bangalore).
The share of ICT in GDP has been decreasing gradually from the last several
years in Europe, which dropped from 6.6% in 2009 to 5.9% in 2013. While the
situation in emerging countries (mostly Asian and Asia pacific countries) in
2009 have accounted for third of the telecom services revenue and in 2013 it
received almost 40% of the total world telecom revenue [3,5]. As a result, the
gap in the growth rate and development of ICT between the developed countries
and emerging countries became wider in 2013. In 2013, the balance in power
regarding ICT kept on shifting to emerging countries continuously and become
shared 80% of global growth contributing almost more than third of the world
ICT market [29]. The comparison of ICT market growth and GDP are shown n
the Table 2.
Table 2. Comparison of growth rates in advanced and emerging Asian countries
2010 2011 2012 2013 2014

Advanced Digiworld market 2.60% 1.30% 1.00% 1.10% 2.40%
economies growth
GDP growth 4.20% 3.70% 3.10% 2.60% 3.90%
Emerging Digiworld market 9.70% 10% 7% 7.70% 7.90%
Asian growth
economies
GDP growth 15.70% 13.20% 9.30% 10.10% 10.30%
4 Factors Influencing the Transfer of Digipower to Asia

The following are the some key reasons that cause the shift of balance of ICT in
the favor of Asia from Europe.
• Single Telecommunication market (Monopoly)

• Fragmentation of digital market in Europe
• Influence of financial crisis in early years of 21st century
• Complex ICT Regulations that cause decrease in investment
• Economic Crisis
• Coalition against “war on terror”
Table 3. Global digiworld market by region
Billion 2012 2013 2014 2015 2018

North America 1073 1114 1166 1215 1349
Europe 924 925 929 955 1034
Asia/Pacific 993 1056 1123 1196 1411
Latin America 256 272 291 307 361
Africa/Middle-East 193 207 224 239 285
World 3439 3574 3732 3913 4439
The Table 3 shows the breakdown of digiworld in the region, it also have
some predictions about the ICT market globally. The Fig. 4, also give a graphical
picture of the percentage share of ICT market contribution throughout the world.
The Figure clearly shows that the emerging Asian countries surpass the Europe
in ICT market share following US.
4.1 Transfer of Digipower to Asia
Some scholars also state that the rise of Asia, particularly China will enable Bei-
jing to pressurize US and the other western countries, not only through military
power, but in the field of technology as well, which will lead to an “open and
intense geopolitical rivalry”. China will rise and become a new power through-
out Asia [10,22,28]. According to the analysis of European Commission Joint
Research Centre, China, Korea and Taiwan are the three highly specialized coun-
tries in ICT manufacturing, who have continued to strengthen their positions in
the global ICT market [14]. China is the most obvious power on the rise in the
present world scenario regarding ICT. In this study, China is the subject matter,
because it not only belongs to the group of emerging economies growing very
rapidly [2]. China took the lead among other regional states due to both its spe-
cialization in manufacturing of IT equipment and its economic size. Moreover,
496 Z. Latif et al.
Fig. 4. Breakdown of global ICT market by region
Fig. 5. Shift in the regional balance of power from Europe to Asia (ICT)
India and other emerging Asian states, which are now promoting their ICT sec-
tor, and can possibly surpass few major western states in the upcoming decades.
The transfer of digital power ICT from West to Asia is getting velocity and
may soon change the perspective for dealing with the global challenges. West is
already aware of Asia’s increasing potency in ICT.
The key lesson learnt from the international point of view is that, China,
Korea and Taiwan having a large ICT manufacturing sector, have an important
potential for growth, particularly, as the important resources like expenditures
and qualified staff. In other words they have strengthened the ICT and R&D
sector. While the US has dominance over EU due to its high productivity level
of manufacturing as well as high R&D resources in order to retain its competi-
tiveness in global ICT market [20,21]. The detailed comparison and contribution
of Asian countries and Europe in the world market (in ICT sector) are given in
Fig. 5.
5 Conclusion
The facts and figures outlined in this paper show that the balance of power in
information and communication technology (ICT) is drifting slowly and grad-
ually towards Asia from Europe. Asia is a densely populated region with more
than 2.8 billion people, living in a land fertile for IT and communication indus-
try, with cheap labor and capital, rich raw resources and tremendous absorbing
capacity for new technologies and developments. All these factors are molding
the situation in favor of emerging Asian countries and shifting the regional bal-
ance of power in ICT towards Asia. The development of ICT has been the main
objective of most of the Asian countries in changing the global environment
towards the Asian region in terms of communication and technology. New Silk
Road is an example of this strategy. Besides, US as the ICT giant has its own
strategies towards occupying the information and communication market. This
will require the sacrifice of core national interests in the short run. At the same
time, China and other Asian emerging countries can get the dominancy in ICT
sector only if these could ensure the interests of other countries according to
their own ways in the long run.
References
1. Altman RC (2009) The great crash, 2008: a geopolitical setback for the west.
Foreign Aff 88(1):2–0014
2. Bannister F, Connolly R (2014) ICT, public values and transformative government:
a framework and programme for research. Gov Inf Q 31(1):119–128
3. Bastarrica F (2014) Mergers and acquisitions HR index. SSRN Electron J. http://
dx.doi.org/10.2139/ssrn.2484775
4. Buzan B (2013) China and the US: comparable cases of ‘peaceful rise’ ? Chin J Int
Polit 6(2):109–132
5. Chesbrough HW (2006) The open innovation model: implications for innovation in
Japan
6. Coen D (2009) Business lobbying in the European Union. Lobbying the European
Union Institutions Actors & Issues
7. Cox M (2012) Power shifts, economic change and the decline of the west? Int Relat
26(4):369–388
8. Cox M, Booth K, Dunne T (1999) The interregnum: controversies in world politics
1989–1999. Cambridge University Press, Cambridge
9. Dadush U, Stancil B (2010) The world order in 2050
10. Friedberg AL (2000) Arming China against ourselves. Commentary 108(1):27–33
11. Halliday F (1996) Book review: Michael COX, US foreign policy after the cold war:
superpower without a mission? Millenn J Int Stud 25(1):174–175 (London: Royal
institute of international affairs and pinter, 1995, 148 pp, no price given)
12. Hawksworth J, Tiwari A (2011) The world in 2050: the accelerating shift of global
economic power: challenges and opportunities. PwC-PricewaterhouseCoopers LLP,
London
13. Hodara J, Ohmae K (1987) Beyond national borders: reflections on Japan and the
world. Foreign Aff 65(5):1103
498 Z. Latif et al.
14. Hoge JF (2004) A global power shift in the making: is the united states ready?
Foreign Aff 83(4):2–7
15. Kunigami A, Navas-Sabater J (2010) Options to increase access to telecommuni-
cations services in rural and low-income areas. Access & Download Statistics
16. Looney RE (2014) Handbook of emerging economies
17. Mamaghani F (2010) The social and economic impact of information and commu-
nication technology on developing countries: an analysis. Int J Manage 18(7):79–79
18. Masuda Y (1980) The information society: as post-industrial society. Institute for
the Information Society
19. Mearsheimer JJ (2001) The tragedy of great power politics. Foreign Aff 80(6):173
20. Nepelski D, Prato GD (2012) Internationalisation of ICT R&D: a comparative
analysis of Asia, the European Union, Japan, United States and the rest of the
world. Asian J Technol Innov 20(2):219–238
21. Nepelski D, De Prato G, Stancik J (2011) Internationalisation of ICT R&D. Insti-
tute for Prospective Technological Studies, Joint Research Centre, European Com-
mission
22. Ross RS (2006) Balance of power politics and the rise of China: accommodation
and balancing in East Asia. Secur Stud 15(3):355–395
23. Ross RS (2012) The geography of the peace: East Asia in the twenty-first century.
Int Secur 23(4):81–118
24. Schaub M (2009) Foreign investment in China-entry, operation and exit strategy.
CCH Hong Kong Limited
25. Singh A (1997) Financial liberalisation, stockmarkets and economic development.
Econ J 107(442):771–782
26. Sudan R, Bank W (2010) The global opportunity in it-based services: assessing
and enhancing country competitiveness. World Bank Publications, pp 214–219
27. Toffler A (1991) Power shift: Knowledge, wealth, and violence at the edge of the
21st century. Powershift Knowl Wealth Violence Edge Century 5(4):91–92
28. Williams F (1989) Measuring the information society 4(1):82–83
29. Wohlers M, Giansante M et al. (2014) Shedding light on net neutrality: towards
possible solutions for the Brazilian case. In: International Telecommunications Soci-
ety Conference, pp 87–99
30. WorldBank (1997) Global economic prospects and the developing countries. World-
Bank
31. Zhao Y (2014) Communication, crisis, & global power shifts: an introduction. Int
J Commun 8(2):26
The Empirical Evidence of the Effect
on the Enterprises R&D from Government
Subsidies, Political Connections
and Rent-Seeking
Dongliang Cai1(B) , Zhen Yang2 , Wu Jiang1 , and Qiuhua Xu1

1
School of Finance, Southwestern University of Finance and Economics,
caidl@swufe.edu.cn
2
Department of Finance, Hanqing Advanced Institute of Economics and Finance,
Renmin University of China, Beijing 100872, People’s Republic of China
Abstract. This study investigates the impact of rent-seeking behav-

ior, political connections, and government science and technology (S&T)
subsidies on the research and development (R&D) of private listed enter-
prises. Based on empirical data of China’s private listed enterprises
between 2008 and 2012, this paper performs validation at three levels
and obtains suggestive conclusions. First of all, the purpose to obtain
more S&T subsidies motivates private listed enterprises to take rent-
seeking activities which can facilitate them in establishing political con-
nections (PCs). Secondly, rent-seeking activities could eventually result
in crowing-out effect of the R&D input, especially for the local PCs.
Thirdly, the side effect of rent-seeking is more significant for enterprises
located in provinces of high corruption levels, which weakens the avail-
ability of government S&T subsidies. The improper intervention of gov-
ernment distorts the market resources allocation.
Keywords: Rent-seeking · Political connections · Government subsidy ·

Research and development
1 Introduction
During the “supply-side reform”, China’s economic development mode is

transforming from “factor-driven” to “innovation-driven”, so it is critical to
strengthen the R&D of enterprise innovations. But funding and policy restraints
make it difficult for small and medium-size private enterprises to carry out
project research. Since the process of research and innovation is a long high-risk
investment period, it is necessary for governments to provide various incentives
and subsidies so as to promote enterprises R&D. However, through the eco-
nomic transition, private listed enterprises, especially small ones, suffered from
500 D. Cai et al.
“explicit” or “hidden” discrimination in policies especially in terms of financial

resources and government subsidies.
Plenty researches have recognized that enterprise can promote development
by political connections, benefiting from the government subsidies, financing
convenience, and tax preference [2,6,21]. Since PCs can bring huge economic
benefits, enterprises are intent to ingratiate with government officials by rent-
seeking activities for the purpose of getting more S&T subsidies [17]. There
are many studies investigating the effects of the PCs and government subsidies
on the enterprise performance. Wu [22] finds the state-owned enterprises can
receive more subsidies than private enterprises, which points out that PCs can
help enterprises obtain S&T subsidies. Meuleman and Maeseneire [16] found the
government subsidies promote enterprises to innovate. Xu [25] analyze the signal
effect of the government subsidies. But on the other side, S&T subsidies will
crowd out the enterprise R&D input [26]. Some empirical results indicate that
political connections are more likely to serve a binding function [28]. However,
beyond that, it still lacks detailed discussions of how PCs and S&T subsidies
affect the R&D input, fairly clear explanations of why the S&T subsidies failed
to make effects, and intuitive evidences of the crowding-out effect of rent-seeking
corruptions on R&D input.
In order to enrich the research of this field, this paper designs the three
levels of verification: first, to test if rent-seeking activities help maintain the
PCs; second to verify if it will crowd out the enterprise R&D input; lastly, to
investigate if corruption levels in enterprise provinces will make a difference and
whether it will ruin the utility of S&T subsidies.
The contributions of this study are as follows. Firstly, it provides empirical
evidence of the economic consequences of establishing political connections and
obtaining S&T subsidies on the enterprise R&D input. Secondly, it enriches the
explanation for the failure of government subsidies. It is found that although
rent-seeking activities facilitate establishing PCs and makes it easier to get gov-
ernment S&T subsidies, it reduces the utility of government subsidies. Thirdly,
this study also provides microcosmic evidence that government S&T subsidies
may distort the social resource allocation under certain conditions.
2 Theoretical Analysis and Research Hypothesis

2.1 Rent-Seeking and Political Connections
Political connections mean close connections between the internal personnel of
the enterprise and government officials, or participation in political affairs. How-
ever, political connections do not mean corruption [6]. The existing financial and
legal systems in China still cannot fully meet the various needs of the transition
of economic development patterns, and the distribution of resources is mainly
driven by the government, which all lead to the current situation that enterprises
scramble for more opportunities by seeking for informal ways under the devel-
opment pressure [1,15]. Under the special period of “adjusting structure and
transition of patterns”, and increasingly drastic market competition in China,
The Empirical Evidence of the Effect on the Enterprises R&D 501
enterprises are trying to gain more financial resources or policy welfares through
such “recessive capitals”, namely the PCs [23,29,30].
In these situations, enterprises in China must have strong motivation to
establish PCs by rent-seeking [4]. Therefore, the following hypothesis is pro-
posed:
Hypothesis 1 : Rent-seeking can help enterprises establish PCs.
2.2 Rent-Seeking and Enterprise R&D
Now, the formal institutions are weak in China [5]. Corruption and rent-seeking
distort fiscal and monetary policy [3,8]. Entrepreneurs are more likely to spend
time and resources on rent seeking rather than on productive activities in order to
influence governments [4]. Under weak institutional contexts, enterprises have to
satisfy government’s requirements, which aggravates burden on enterprises [14].
In this condition, government subsidies can deteriorate corporate performance.
Quevedo [7] proposed that government S&T subsidies have a negative effect on
the R&D expenditure, namely the crowding-out effect. Huang [9] made a case
study on Chinese listed companies between 2001 and 2007, finding that PCs
have a promoting effect on the performance of SOEs, but negative impact on
private listed enterprises.
Following the normal logic of enterprise behavior, the following hypothesis is
proposed:
Hypothesis 2 : Enterprises are intent to establish PCs by rent-seeking for obtain-
ing more government S&T subsidies, but it will lead to crowding-out effect on
R&D input.
2.3 S&T Subsidies and Enterprise R&D
As is recognized, government S&T subsidy are important for enterprises, which

can alleviate market failures in R&D activities [11,24]. Levin and Reiss [13]
studied the relationship between government subsidies and R&D output on the
micro level, finding that government subsidies have a remarkable role in promot-
ing R&D. Lee and Cin [12] drew the same conclusions by considering different
country samples and selecting panel data for several years. The government R&D
subsidies contribute to both green innovations [19] and sustainable innovations
[10]. Yet, there are opposite opinions. Zhang [27] found subsidies, in long and
short terms, have significant positive effect on the financial performance of wind
energy manufacturing companies. In renewable energies sector, subsidy policies
could cause adverse consequences to a certain extent [20]. The current economic
transition period in China, is combined with a serious rent-seeking phenom-
enon and ethical risks, so the establishment of PCs may consume a tremendous
amount of manpower and material resources for enterprises and thus fail to
improve the performance.
502 D. Cai et al.
Therefore, the following hypothesis is proposed.

Hypothesis 3 : In provinces with serious corruptions, the rent-seeking results in
more significant crowding-out effect, frustrating the availability of government
S&T subsidies.
3 Research Design and Sample Selection
3.1 Sample
The original sample contains the non-financial private A-share companies listed
in Shanghai and Shenzhen Stock Exchange during 2008–2012. The final sample
keeps 361 companies listed 1614 totally after basic sample processing.
The major data used in the empirical part includes: the detailed government
subsidies data, PCs data, basic enterprise data, post crime data at the provincial
level and the excess administration expense data. The data was derived from
the CSMAR and WIND database, the China inspection yearbook and China
statistical yearbook. Most of them need manual calculation.
3.2 Model Setup and Variable Definition
To test Hypothesis 1, we set up the following probit regression model
PPoli,i,t = α + β1 × Renti,t + β2 × Xi,t + εi,t . (1)
To illustrate the robustness of conclusions we also set up the following logistic

model:
PPoli,i,t
log = α + β1 × Renti,t + β2 × Xi,t + εi,t . (2)
1 − PPoli,i,t
Here, PPoli,i,t is the probability that company i established political connections
in year t. Correspondingly, Polii,t represents the dummy variable of PCs in year
t while Loca represents the dummy variable of central PCs. Then, referring
to the treatment methods proposed by Yu et al. [17], the participation in the
politics of private listed companies is depicted by whether the directors and
senior executives of listed companies have held a position in the National People’s
Congress (NPC), Chinese People’s Political Consultative Conference (CPPCC)
or government sectors. Specifically, P olii,t is defined as 1 if the directors or senior
executives have been the members of the NPC or CPPCC, or held a position at
a government sector, otherwise it is 0. Locai,t is the same way defined for the
local NPC, CPPCC, or local government sectors.
Renti,t represents the excess administration expense of company, a proxy
variable of rent-seeking, which should be calculated as with a regression process
introduced by RichardsonM [18]. The excess administration expense is the
in-sample forecast residuals of the regression between the ratio (management
cost/operating income) with a series of factors based on the historical data.
To verify Hypothesis 2, we set up the following model:
log Devlpexpi,t = α + β1 × log SciSubsibyi,t + β2 × Renti,t + β3 × Xi,t + εi,t ,

(3)
where log Devlpexp is the logarithm of corporate R&D investment;
log SciSubsidy represents the enterprises’ S&T subsidies from government.
To verify Hypothesis 3, we set up the following model:
log Devlpexpi,t = α + β1 × log SciSubsibyi,t + β2 × Renti,t

(4)
+ β3 × P olii,t + β4 × Xi,t + εi,t .
On the basis of model (3), model (4) adds the PCs variable Poli, and the
sub-sample regression for local PCs companies. In addition, the regression is
divided into different groups according to the corruption level in the provinces.
The corruption level is depicted by the corruption and bribery cases registered
every 10 thousand public officers, which is exactly the ratio of registered post
criminal cases at the provincial level to the number of civil servants.
In the model (1) & (2), X is a multiple vector of control variables includ-
ing: the logarithmic company size (log Asset), rate of asset (ROA), debt-to-
assets ratio(Dbassrt), establishment years (EstAge), the logarithm of top three
executive compensations (LogMane3Pay), the duality of CEO and chairman
Table 1. Descriptive statistics of variables from groups with different PCs (10,000
yuan)
Companies without PCs Companies with PCs Companies with local PCs
Mean sd Med Mean sd Med Mean sd Med
S&T subsidies 2.62 6.52 0 4.54 31.21 0 4.75 33.4 0
Operating 3037.4 3993.23 1607.68 6495.62 13721.8 2260.71 5423.21 12418.35 1929.38
income
R&D input 45.08 89.64 14.06 77.18 248.51 16.44 59.19 241.09 12.15
Overheads 187.82 224.04 119.54 298.49 554.39 136.26 247.64 451.23 114.78
Top three 1.18 1.17 0.9 1.34 1.36 1.05 1.16 0.9 0.97
executive
compensation
Excess −0.01 0.12 −0.01 0 0.15 −0.01 0 0.07 −0.01
administration
expense
Top three 1 0.97 0.75 1.27 1.61 0.88 1.04 1.03 0.79
directors’
compensation
The proportion 0.37 0.05 0.36 0.36 0.05 0.33 0.36 0.04 0.33
of independent
directors
The duality of 0.19 0.39 0 0.18 0.38 0 0.24 0.43 0
CEO and
chairman
Corruption 24.08 5.67 24.31 24.52 6.78 24.31 24.4 7.34 24.31
level
504 D. Cai et al.
(Chairman&CEO is defined as 1 for duality), the proportion of indepen-

dent directors (IndeDirRatio), year control variable, industry control variable
(Industry is defined as 1 if the company belongs to the national key support,
highly regulated or monopoly industries, otherwise 0 [17].
3.3 Descriptive Statistics

It can be seen from Table 1, politically connected companies obtain more govern-
ment S&T subsidies in average which is consistent with the previous literatures
view that political connections help obtain the government subsidies. At the
same time, the politically connected enterprises have relatively higher overheads
and excess administration expense, indicating that rent-seeking may exist.
Table 2. Descriptive statistics of variables between groups with different excess admin-
istration expense
Low excess High excess

administration expense administration expense
Mean sd Med Mean sd Med
S&T subsidies 5.87 24.78 0 10.47 51.32 0
(10,000 yuan)
T&D input 44.51 107.41 16.34 148.96 359.71 41.77
(10,000 yuan)
Overheads 192.9 230.08 130.62 531.86 773.28 287.51
(10,000 yuan)
R&D 0.21 0.23 0.11 0.23 0.26 0.13
input/Overheads
Operating income 3903.8 6600.95 2171.7 11630.4 18952.16 4605.86
(10,000 yuan)
Overheads/ 0.07 0.04 0.06 0.13 0.27 0.07
Operating income
Total asset 5773.42 10141.05 2899.96 11278.23 17493.17 5611.22
(10,000 yuan)
Top three executive 1.18 0.89 0.96 1.54 1.83 1.14
compensation
(10,000 yuan)
Top three directors’ 1.1 1.28 0.8 1.36 1.9 0.93
compensation
(10,000 yuan)
The proportion of 0.37 0.05 0.33 0.36 0.05 0.33
independent
directors
The duality of CEO 0.16 0.37 0 0.13 0.33 0
and chairman
On the basis of Table 2, though the enterprises with high excess adminis-
tration expense can obtain more S&T subsidies, they do not show any sig-
nificantly higher ratio of R&D input/overheads but get higher ratio of the
overheads/operating income. It may be caused by the crowding out effect from
the rent-seeking behaviors, namely the existence of corruption and rent-seeking
reduce the efficiency of the use of government S&T subsidies.
4 Empirical Results and Analysis

4.1 Rent-Seeking Behavior and Political Connections
Table 3 contents 4 regression results of the probit model and logistic model
used to test the influence of excess administration expense on the total polit-
Table 3. Test results of the hypothesis 1
Variable PPoli log(PPoli /(1 − PPoli ))

Probit Model Logistic Model
Rent 0.319∗∗ 0.222∗ 0.541∗∗ 0.378∗
(0.127) (0.128) (0.224) (0.225)
L1Rent - 0.018∗ - 0.031
- (0.010) - −0.021
logAsset −0.088∗ −0.104∗ −0.158∗ −0.187∗
(0.053) (0.058) (0.092) (0.101)
ROA −0.378 −0.468 −0.635 −0.756
(0.25) (0.299) (0.394) (0.506)
Dbassrt −0.088 −0.23 −0.158 −0.411
(0.169) (0.19) (0.28) (0.314)
EstAge 0.018 0.014 0.032 0.026
(0.012) (0.014) (0.022) (0.026)
logMane3Pay −0.151∗∗ −0.126 −0.258∗∗ −0.211
(0.07) (0.077) (0.119) (0.131)
Chairman& CEO −0.239∗∗ −0.259∗∗ −0.392∗∗ −0.415∗
(0.113) (0.126) (0.19) (0.213)
IndeDirRatio −2.529∗∗∗ −1.539 −4.059∗∗ −2.384
(0.96) (1.067) (1.608) (1.784)
cons 5.337∗∗∗ 5.150∗∗∗ 9.115∗∗∗ 8.757∗∗∗
(1.425) (1.538) (2.45) (2.649)
Year Yes Yes Yes Yes
Industry Yes Yes Yes Yes
R2 0.052 0.055 0.052 0.054
N 925 727 925 727
∗∗∗
Note: Robust standard errors in parentheses, Significant at
0.01, ∗∗ Significant at 0.05, ∗ Significant at 0.1.
506 D. Cai et al.
ical connections. The regression results of probit model reflect a significantly

positive coefficients of Rent and the first-lagged Rent as well. It suggests that
excess administration expense can help enterprises establish political connec-
tions, which is companied with hysteresis effect. To maintain the robustness of
the conclusion, the logistic model is conducted at the same time and the results
are basically the same as the probit models, which shows a robust estimating
conclusion.
4.2 Rent-Seeking Behavior and Enterprise R&D Input

Based on the results of the Hausman test and W ald test, corresponding regres-
sion models are applied to test the influences of government S&T subsidies and
excess administration expenses on enterprise R&D input. Table 4 contents the
regression results of groups with different political connections. Comparing the
regression in columns 1 and 2, the coefficient of Rent is significantly negative
for companies with PCs, indicating that high excess administration expense
will probably reduce the R&D input. Overall, test results of the Hypothesis 1
draw the conclusion that enterprises are intend to establish political connections
through rent-seeking which makes it easier to acquire government S&T subsidy.
According to the motivation of establishing PCs, we can infer that the pursuance
of government S&T subsidies will stimulate more rent-seeking activities which
eventually result in crowding-out effect on the primary business performance.
Under the fiscal decentralization system in China, local governments are
empowered with great administrative power and relatively weak financial power
since the tax reform. When local governments try to obtain more resources
through extra-budgetary way with less supervision, it is prone to create rent-
seeking corruptions between local governments and enterprises. At the same
time, the local governmental officials have strong discretionary power in sci-
ence subsidy and financial support (including rewards) which has no explicit
restriction in laws or regulations. So the difference between enterprises with and
without local political connections are further investigated as regression results
in columns 3 and 4. The variable Rent has a significantly larger negative coeffi-
cient than total sample in columns 1, which suggests that local PCs will cause
higher rent-seeking cost and worse impact on the enterprise’s performance.
4.3 R&T Subsidies and Corporate R&D

Following the Hausman test and LR test, corresponding regression models are
applied to test the hypothesis 3 and results are listed in Table 5. Compared to
the low corruption group, high corruption group in columns 1 and 2 showed dis-
tinctly negative effect of variable Rent. The figures illustrate that rent-seeking
behavior will lead to substantial crowding-out impact on the R&D of enter-
prises in provinces with serious corruptions. On the other hand, coefficients of
log SciSubsidy are almost significantly negative in columns 1 and 2, but posi-
tive in columns 3 and 4. Thus for high corruption group, the government S&T
subsidies probably cause negative impact on the enterprise R&D. It is because
enterprises have to invest great manpower and resources for rent-seeking, which
will fail the fulfill use of S&T subsidies and crowd out scientific research innova-
tion input. In addition, the coefficient of Loca is significantly negative in column
2, suggesting that the local PCs will strengthen such side effect, which is consis-
tent with the results in Table 4.
Table 4. Test results of the Hypothesis 2
Variable log Delpexp

Companies Companies Companies Companies
with PCs without with local without
PCs PCs local PCs
log SciSusidy with PCs without local PCs local PCs
0.073∗∗∗ PCs 0.01 0.075∗∗
(0.027) 0.063 (0.047) (0.032)
(0.217)
Rent −3.427∗∗∗ 11.83 −5.131∗∗∗ −3.757∗
(1.057) (7.729) (1.182) (2.178)
log Asset 0.485∗∗∗ 0.818 0.352 0.617∗∗∗
(0.119) (2.199) (0.229) (0.127)
ROA 4.009∗∗ 8.815 0.956 3.860∗∗
(1.636) (8.911) (3.319) (1.815)
Dbassrt −1.718∗∗ 1.274 0.735 −2.201∗∗∗
(0.725) (3.023) (1.213) (0.782)
EstAge 0.004 0.674 −0.056 0.002
(0.016) (1.09) (0.062) (0.016)
log Mane3Pay 0.457∗∗∗ −0.459 0.658∗∗ 0.512∗∗∗
(0.163) (0.716) (0.301) (0.182)
ChairmanCEO −0.105 −1.427∗∗∗ 0.704 0.356
(0.23) (0.243) (0.482) (0.33)
IndeDirRatio −5.961∗∗ −0.782 3.087 −4.518∗
(2.89) (8.063) (5.342) (2.639)
cons 2.829 −8.068 0.121 −1.098
(3.674) (25.669) (7.705) (3.363)
r2 0.371 0.785 0.526 0.395
N 192 49 64 177
∗∗∗
Note: Robust standard errors in parentheses, Significant at 0.01,
∗∗
Significant at 0.05, ∗ Significant at 0.1.
508 D. Cai et al.
Table 5. Test results of the hypothesis 3
Variable logDelpexp
High corruption Low corruption
∗∗∗
log SciSubsidy −0.044 −0.119 0.106∗∗ 0.095∗
(0.032) (0.046) (0.058) (0.066)
Rent −6.180∗∗∗ −5.567∗∗∗ −0.020 −0.115
(0.818) (1.051) (1.859) (2.053)
log Asset 0.595∗∗ 0.507∗∗ 0.611∗∗∗ 0.855∗∗∗
(0.232) (0.247) (0.204) (0.250)
ROA 3.687 −1.426 5.704∗ 3.225
(3.745) (2.402) (3.713) (3.424)
Dbassrt −2.181 −1.514 −0.651 −0.966
(1.741) (1.533) (1.207) (1.209)
EstAge 0.034 0.011 0.009 0.024
(0.054) (0.066) (0.018) (0.024)
log Mane3Pay 0.994∗∗∗ 0.948∗∗∗ 0.159 0.084
(0.324) (0.280) (0.300) (0.319)
ChairmanCEO −0.364 0.204 0.000 −0.057
(0.703) (0.732) (0.479) (0.409)
IndeDirRatio 1.163 −2.755 −4.325 −5.091
(2.960) (2.324) (5.482) (5.076)
Poli −0.175 - 0.257 -
(0.393) - (0.379) -
Loca - −1.136∗∗∗ - 0.412
- (0.430) - (0.449)
cons −8.135 −2.900 1.771 −2.137
(7.067) (7.119) (5.826) (6.425)
R2 0.505 0.584 0.479 0.434
N 62 62 82 82
Note: Robust standard errors in parentheses, ∗∗∗ Significant at 0.01,
∗∗
Significant at 0.05, ∗ Significant at 0.1.
4.4 Robustness Analysis

To guarantee the robustness of conclusions, we build logistic and probit models.
Considering the possible endogenous problems, the lagged excess administra-
tion expense is added into the test of Hypothesis 1, As for hypothesis 2 and
3, we use the current and previous budget of technology spending from enter-
prise‘s local government as the instrumental variable to conduct two-stage least
squares regressions, the exogeneity of the government budget avoid the two-way
causality problem. The results are basically consistent, indicating the reliability
of conclusions.
5 Research Conclusions
We find evidence that rent-seeking behavior facilitate private listed enterprises in
establishing PCs and thus obtaining more S&T subsidies. However, it is proved
that rent-seeking activities will finally result in crowing-out effect of the R&D
input, especially for the enterprises located in the province of higher corruptions.
For such companies, the PCs are closely related to the S&T subsidies, which
breeds severe rent-seeking corruption and weakens the availability of subsidies.
Our results support the idea that PCs can result in rent-seeking corruption,
because enterprises are motivated to ingratiate with government officials by rent-
seeking activities for the purpose of getting more S&T subsidies. And it turns out
that the crowding-out effect of such activities will eventually fail the availability
of subsidies. The PCs and S&T subsidies cannot effectively promote corporate
performance or improve the development of the whole society, but only serve
for a few entrepreneurs and politicians. The government should regulate the
allocation of government subsidies, and have a reasonable control over local
government powers to improve the efficiency of the whole market and stimulate
the enthusiasm of enterprise innovation.
Acknowledgements. This research was supported by the project of Research Center

for System Sciences and Enterprise Development (Grant No. Xq16C03).
References
1. Allen F, Qian J, Qian M (2005) Law, finance, and economic growth in china. J
Financ Econ 77(1):57–116
2. Claessens S, Feijen E, Laeven L (2008) Political connections and preferential access
to finance: the role of campaign contributions. J Financ Econ 88(3):554–580
3. Dimakou O (2009) Bureaucratic corruption and the dynamic interaction between
monetary and fiscal policy. Eur J Polit Econ 40:57–78
4. Du J, Mickiewicz T (2015) Subsidies, rent seeking and performance: being young,
small or private in China. J Bus Ventur 31(1):22–38
5. Estrin S, Korosteleva J, Mickiewicz T (2013) Which institutions encourage entre-
preneurial growth aspirations? J Bus Ventur 28(4):564–580
6. Faccio M (2006) Politically connected firms. Am Econ Rev 96(1):369–386
7. Garcla-Quevedo J (2004) Do public subsidies complement business R&D? A meta-
analysis of the econometric evidence. Kyklos 57(1):87–102
8. Hessami Z (2014) Political corruption, public procurement, and budget composi-
tion: theory and evidence from OECD countries. Eur J Polit Econ 34(6):372–389
9. Huang MC, Wu CC (2016) Facts or fates of investors’ losses during crises? Evidence
from reit-stock volatility and tail dependence structures. Int Rev Econ Financ
42:54–71
10. Ketata I, Sofka W, Grimpe C (2015) The role of internal capabilities and firms’
environment for sustainable innovation: evidence for germany. R&D Manage
45(1):60–75
11. Kleer R (2010) Government R&D subsidies as a signal for private investors. Res
Policy 39(10):1361–1374
510 D. Cai et al.
12. Lee EY, Cin BC (2010) The effect of risk-sharing government subsidy on corporate
R&D investment: empirical evidence from korea. Technol Forecast Soc Change
77(6):881–890
13. Levin RC, Reiss PC (1984) Tests of a schumpeterian model of R&D and market
structure. University of Chicago Press, pp 175–208
14. Marquis C, Qian C (2013) Corporate social responsibility reporting in China: sym-
bol or substance? Organ Sci 25(1):127–148
15. Mellahi K et al (2016) A review of the nonmarket strategy literature: toward a
multi-theoretical integration. J Manage 42(1):143–173
16. Meuleman M, Maeseneire WD (2012) Do R&D subsidies affect SMEs’ access to
external financing? Soc Sci Electron Publishing 41(3):580–591
17. Minggui Y, Yafu H, Hongbo P (2010) Political connection, rent seeking and the
effectiveness of fiscal subsidies by the local government. Econ Res J 3:65–77
18. Richardson S (2006) Over-investment of free cash flow. Rev Acc Stud 11(2):159–189
19. Schiederig T, Tietze F, Herstatt C (2012) Green innovation in technology and
innovation management-an exploratory literature review. R&D Manage 42(2):180–
192
20. Shen J, Luo C (2015) Overall review of renewable energy subsidy policies in China-
contradictions of intentions and effects. Renew Sustain Energy Rev 41:1478–1488
21. Su ZQ, Fung HG et al (2014) Cash dividends, expropriation, and political connec-
tions: evidence from China. Int Rev Econ Financ 29(1):260–272
22. Wu A (2016) The signal effect of government R&D subsidies in China: do ownership
matter? Technol Forecast Soc Change 117:339–345
23. Xu C (2011) The fundamental institutions of China’s reforms and development. J
Econ Lit 49(4):1076–1151
24. Xu E, Xu K (2013) A multilevel analysis of the effect of taxation incentives on
innovation performance. IEEE Trans Eng Manage 60(1):137–147
25. Xu K, Huang KF, Xu E (2014) Giving fish or teaching to fish? An empirical
study of the effects of government research and development policies. R&D Manage
44(5):484–497
26. Yu F, Guo Y et al (2016) The impact of government subsidies and enterprises’
R&D investment: a panel data study from renewable energy in China. Energy
Policy 89:106–113
27. Zhang H, Li L et al (2014) Political connections, government subsidies and firm
financial performance: evidence from renewable energy manufacturing in China.
Renew Energy 63(1):330–336
28. Zhang J, Marquis C, Qiao K (2016) Do political connections buffer firms from or
bind firms to the government? A study of corporate charitable donations of Chinese
firms. Organ Sci 27:1307–1324
29. Zhang Y (2015) The contingent value of social resources: entrepreneurs’ use of
debt-financing sources in western China. J Bus Ventur 30(3):390–406
30. Zhou W (2013) Political connections and entrepreneurial investment: evidence from
China’s transition economy. J Bus Ventur 28(2):299–315
Scenario-Based Location Arc Routing Problems:
Introducing Mathematical Models
Alireza Amini1 , Reza Tavakkoli-Moghaddam2(B) ,

and Sadoullah Ebrahimnejad3
1
School of Industrial Engineering, South Tehran Branch,
Islamic Azad University, Tehran, Iran
2
School of Industrial Engineering, College of Engineering,
University of Tehran, Tehran, Iran
tavakoli@ut.ac.ir
3
Department of Industrial Engineering, Karaj Branch,
Islamic Azad University, Karaj, Iran
Abstract. A location arc routing problem (LARP) is an important issue

that finds the best locations of depots and routing simultaneously. It
deals with a routing problem, in which demands are on arcs instead of
nodes. Additionally, parameters may not be deterministic in real prob-
lems. Thus, this paper addresses an uncertain LARP through developing
a deterministic mathematical model regarding to the respective litera-
ture and employing two scenario-based approaches. The objectives of the
model are to minimize the maximum regret and minimize the mean and
deviation of the objective function value (OFV). A numerical example
is generated and the results analyze the performance of scenario-based
models.
Keywords: Location arc routing · Mathematical model · Scenario-

based · Regret
1 Introduction
An important cost of the most of companies is their logistic costs. It can be
reduced through appropriate planning and designing supply chains. In fact, dis-
tribution networks in supply chains are highly important [9]. The location prob-
lem aims at finding optimal locations for facilities or centers. In fact, one of
the most important factors in the success of a production unit is determining
appropriate locations of sites and also, on a smaller scale, production facilities.
Therefore, finding optimal or near-optimal solutions is essential. Vehicles rout-
ing, on the other hand, is among the most challenging issues in supply chain
management. These two issues affect the cost of networks and supply chains.
Hence, the location-routing problem (LRP)is defined to determine the location
and routing decisions in at the same time. It is a combination of both loca-
tion and routing optimization problems, whose decisions are made at the same
512 A. Amini et al.
time. In recent years, several LRPs were presented. Considering the LRP, better
logistics could be designed [4].
In general, routing issues may be defined on a graph of number of nodes and
arcs. The LRP deals with routing problems where demands are on nodes. One
related issue which has been considered in recent years by some researchers is the
routing problem when demands belong to arcs instead of nodes. Arc routing is
a special type of routing problems. In this content, researchers have taking into
account the conditions and restrictions in real applications, a variety of models,
and methods for solving these problems. Ghiani and Laporte [5] addressed a
problem and solved it based on the concept of the rural postman problem (RPP)
as one the earliest studies on arc routing problems. There are many researches
in this filed. For example, Pia and Filippi [8] addressed a capacitated arc routing
problem (CARP) to consider a waste collection problem and Beullens et al. [1]
employed some heuristics for a periodic CARP.
Fig. 1. LRP versus LARP
The location arc routing problem (LARP) is emerged when arc routing deci-
sions are made along with location ones. The difference between the LRP and the
LARP is displayed by Fig. 1. Waste collection, mail delivery, and telecommuni-
cation network design is among most well-known problems in which LARP could
be addressed [7]. Also, some node routing problems could be considered as an arc
routing one. For instance, when there are supermarkets as the nodes of demand,
one can convert a node routing to an arc routing problem; because the street
with N supermarkets is an arc with an aggregated demand from them. It might
reduce the complexity of the problem. It means that a LARP could be defined
Scenario-Based Location Arc Routing Problems 513
for a lot of real applications. Firstly, Levy and Bodin [2] considered a LARP by
which a mail delivery problem is taken into account. Then, Ghiani and Laporte
[6] employed some heuristics to deal with a LARP. Lopes et al. [11] developed a
mathematical model for LARP. In addition, Hashemi and Seifi [3] developed two
mathematical models for single-depot and multiple-depot LARPs, respectively.
One the latest works is what Riquelme-Rodrlguez et al. [10] presented which
deals with a LARP with inventory constraints.
The paper on hand addresses a LARP by developing a mathematical model.
Also, according to the uncertain nature of the data and information in the real
world, we intend to build our research based on this fact. In our best of knowledge
there is a gap in considering uncertainties in LARP. So, two scenario-based
approaches are employed in the developed mathematical model. Section 2 defines
the problem and notations and develops a deterministic mathematical model
for LARP. Section 3 deals with two scenario-based approaches and modifies the
mathematical model for each. In Sect. 4 a numerical example is analyzed based
on the mathematical models. Finally, we conclude the paper.
2 Deterministic LARP Model
There is a complete graph G = (V, A) where V and A denotes the set of vertices
and set of arcs, respectively. V is consisted of K as the set of potential depots and
R as the set of extremities of arcs, V = R ∪ K. Each arc denoted by (i, j) which
means that the arc begins form vertex i and ends at j. There is a given number
of homogenous vehicles for each potential depot in a set called P . Then each
depot is able to have |P | separate tours. Some arcs have positive demands and
the demand of arc(i, j) equals the demand of arc(j, i). This means that both of
them are not required to be serviced and it is sufficient to meet the demand of one
of them. The problem is finding optimal depots among the potential ones, assign
vehicles and arcs to them, and traverse those arcs in an optimum routing. It is
assumed that arcs with positive demands must be serviced once by a vehicle and
also split delivery of demands is not allowed. Number of deadheaded traversing
in each arc is not limited. In addition, tours must begin from a depot and end to
the same one. In the deterministic model are parameters are given in advance.
Following notations are introduced to develop the mathematical model.
Parameters:
Fk : Establishment cost of depot k;

Ckij : Servicing cost of required arc (i, j) by depot k;
Hkij : Deadheading cost of required arc (i, j) by depot k;
W : Hiring cost of each vehicle;
Dij : Demand of arc(i, j) ∈ R;
Q: Vehicle’s capacity;
Bk : Capacity of depot k.
514 A. Amini et al.
Variables:
xpk
ij : Binary variable which equals 1 if required arc(i, j) is served by vehicle p
of depot k;
pk
yij : Number of deadheaded traverses in arc(i, j);
pk
fij : Departed flow form arc(i, j);
vij : Binary variable which equals 1 if arc(i, j) has a positive demand;
qpk : Binary variable which equals 1 if vehicle p of depot k is established;
gpk : Departed flow form depot k by vehicle p;
zk : Binary variable which equals 1 if depot k is established.
Regarding the assumptions and notations the mathematical model for the
deterministic LARP is developed as below. As a matter of fact, this model is
inspired by what is developed in [3].
pk pk
k k
min θ = zk Fk + xij Cij + yij Hij + qpk W (1)
k i,j,p,k i,j,p,k p,k

s.t. (xpk
ij + xpk
ji ) = vij , ∀(i, j) ∈ R (2)
k∈K,p∈P

(xpk pk
ij + xji ) = 0, ∀(i, j) ∈ K (3)
k∈K,p∈P

(xpk pk
ij + yij ) − (xpk pk
ji + yji ) = 0, ∀j ∈ V, ∀p ∈ P, ∀k ∈ K (4)
i∈V i∈V
qpk ≤ zk , ∀p ∈ P, ∀k ∈ K (5)
pk pk
(xkj + ykj ) = qpk , ∀p ∈ P, ∀k ∈ K (6)
j∈R

(xpk pk
jk + yjk ) = qpk , ∀p ∈ P, ∀k ∈ K (7)
j∈R
xpk pk
ij + yij ≤ M qpk , ∀(i, j) ∈ V, ∀p ∈ P, ∀k ∈ K (8)
xpk
+
ii
pk
yii = 0, ∀i ∈ V, ∀p ∈ P, ∀k ∈ K (9)
pk
yij = 0, ∀p ∈ P, ∀k ∈ K (10)
i∈K,i=k
xpk
ij ≤ Dij , ∀(i, j) ∈ R, ∀p ∈ P, ∀k ∈ K (11)
Dij ≤ M vij , ∀(i, j) ∈ R. (12)

qpk ≤ |P | , ∀k ∈ K (13)
p∈P

gpk ≥ Dij xpk
ij − M (1 − qpk ), ∀p ∈ P, ∀k ∈ K (14)
i,j

gpk ≤ Dij xpk
ij + M (1 − qpk ), ∀p ∈ P, ∀k ∈ K (15)
i,j
gpk ≤ Qqpk , ∀p ∈ P, ∀k ∈ K (16)

gpk ≤ Bk , ∀p ∈ P, ∀k ∈ K (17)
p∈P
pk
pk

fij − fji = Dji xpk
ji , ∀(i, j) ∈ V, ∀p ∈ P, ∀k ∈ K (18)
i i i
pk pk pk
fij ≤ Q(xij + yij ), ∀(i, j) ∈ V, ∀p ∈ P, ∀k ∈ K (19)
pk
fkj = gpk , ∀j ∈ V, ∀p ∈ P, ∀k ∈ K (20)
j
pk
fik = 0, ∀i ∈ V, ∀p ∈ P, ∀k ∈ K (21)
j
xpk
ij , vij , zk , qpk ∈ {0, 1}, ∀(i, j) ∈ V, ∀p ∈ P, ∀k ∈ K (22)
pk
yij ≥ 0&Integer, ∀(i, j) ∈ V, ∀p ∈ P, ∀k ∈ K (23)
pk
fij , gpk ≥ 0, ∀(i, j) ∈ V, ∀p ∈ P, ∀k ∈ K. (24)
Objective function (1) minimizes the total costs of opening depots, traversing,
and hiring vehicles. Constraint (2) guarantees that required arc(i, j) must be
served by one vehicle form one depot in one direction, i.e. from i to j or j to i.
Constraint (3) ensures that traveling between depots is not allowed. Constraint
(4) represents the continuity of tours, and Constraint (5) is in charge of vehicle
assignments to depots could be considered if that depot is opened. Constraint (6)
indicates that one arc must leave the depot to one arc of its tour at the beginning
of the tour and, on the other hand, Constraint (7) shows that arc must enter
the depot form one arc of its tour at the end of the tour. In Constraint (8) it is
mentioned that xpk pk
ij and yij can be equal to 1 if they are assigned to the vehicle
p of depot k, where M is a big sufficient positive number.
Constraint (9) guarantees that there is not any arc between one vertex and
itself. Constraint (10) declares that tours are not allowed to end at a different
depot from which they begin. Also, Constraints (11) and (12) find the arcs with
positive demands. Regarding Constraint (13), maximum number of assigned
vehicles to each depot cannot exceed the total number of vehicles. Constraints
(14) and (15) measure the total amount of demands for each vehicle of each
depot if it is opened. Constraints (16) and (17) regard the capacity limitations
for opened depots and selected vehicles. Constraint (18) guarantees that the
existing flows from arc(i, j) equal the entering ones to that arc minus its demand
if there is one servicing traveling for that arc in that tour. Also, Constraint (19)
516 A. Amini et al.
shows that flows of arc(i, j) are zero if there is not any traversing. Although,
flows are not required to be determined, Constraints (18) to (21) ensure not
to have sub-tours. Finally, Constraints (20) and (21): existing flow form each
opened depot by each vehicle equals the total amount of demands of that tour
and entering flow to that depot is zero.
3 Scenario-Based LARP
Regarding the real problems in the world, data might not be exactly given. So,
many problems have uncertain parameters. The paper on hand aims at employ-
ing two scenario-based approaches to consider uncertainties: (i) Minimization of
maximum regret, (ii) Minimization of mean and deviation of objective function
value. The scenario-based LARP is called SLARP in this paper. So the first
Approach is named SlARPR and the second one is named SlARPM D . Also, let
s denotes the index of scenarios form its set S from now on.
3.1 SlARPR Mathematical Model

This approach begins with employing the LARP model to solve the prob-
lem under each scenario separately. Then an integrated mathematical model
is required to obtain a unique solution which meets all requirements of all sce-
narios. Developing the integrated mathematical model, following variables are
introduced.
rmax : Maximum regret;
pk
fijs : Departed flow form arc (i, j) under scenario s;
gpks : Departed flow form depot k by vehicle p scenario s.
In fact, despite of other variables, flows are dependent on scenarios. Let θs∗
indicates the optimum value for objective function under scenario s ∈ S. In
addition θs is the obtained OFV for scenario s ∈ S when the integrated model
is solved, Eq. (25). Then, the regret of scenario sS is the difference between θs∗
and θs . θs ≥ θs∗ for each scenario s ∈ S.
⎛ ⎞
pk pk
θs = ⎝ zk Fks + k
xij Cijs + k
yij Hijs + qpk Ws ⎠ , ∀s ∈ S. (25)
k i,j,p,k i,j,p,k p,k
Considering LARP model, Eq. (25), and new variables, following model is
developed as the integrated model for SLARPR .
min θR = rmax , (26)
s.t. rmax ≥ θs − θs∗ , ∀s ∈ S (27)
xpk
ij ≤ Dijs , ∀(i, j) ∈ R, ∀p ∈ P, ∀k ∈ K, ∀s ∈ S (28)
Dijs ≤ M vij , ∀(i, j) ∈ R, ∀s ∈ S (29)

gpks ≥ Dijs xpk
ij − M (1 − qpk ), ∀s ∈ S (30)
i,j

gpks ≤ Dijs xpk
ij + M (1 − qpk ), ∀s ∈ S (31)
i,j
gpks ≤ Bks , ∀p ∈ P, ∀k ∈ K, ∀s ∈ S (32)

gpks ≤ Bks , ∀p ∈ P, ∀k ∈ K, ∀s ∈ S (33)
pk pk
fijs − fjis = Dijs xpk
ij , ∀j ∈ V, ∀p ∈ P, ∀k ∈ K, ∀s ∈ S (34)
i i i
pk
fijs ≤ Qs (xpk
ij
pk
+ yij ), ∀(i, j) ∈ V, ∀p ∈ P, ∀k ∈ K, ∀s ∈ S (35)
pk
fijs = gpks , ∀(i, j) ∈ V, ∀p ∈ P, ∀k ∈ K, ∀s ∈ S (36)
j
pk
fijs = 0, ∀(i, j) ∈ V, ∀p ∈ P, ∀k ∈ K, ∀s ∈ S (37)
i
pk
fijs , gpks ≥ 0, ∀(i, j) ∈ V, ∀p ∈ P, ∀k ∈ K, ∀s ∈ S. (38)
Constraints (2–10), (13), (22), (23).

Objective function (26) minimizes the maximum regret among all scenarios
measured by Constraint (27). Other constraints are similar to those presented
for the LARP model. Some of them (i.e., Constraints (28) to (38)) are required
to be considered for each scenario separately.
3.2 SlARPM D Mathematical Model

This approach is similar with SlARPR in beginning with solving problems under
scenarios separately and takes all of them into account in an integrated model.
Its main difference with SlARPR is in its objective function. Also, SlARPM D
needs to know the probability of occurrence of each scenario, Ps . SlARPM D
2
aims at minimizing the mean
of OFV, θ, and its deviation form, σθ , where
θ = s Ps θs and dθ = s Ps |θs − s Ps θs |. So, this approach minimizes
θ + dθ in its integrated model. It is worth mentioning that deviation could be
differently important for the different decision makers. Therefore, the weight
λ
is introduced the integrated model minimizes θ + dθ which is
for it. Now
P θ
s s s + λ P |θ
s s s − s Ps θs |. The deviation term results in having a non-
linear objective function. As a result to non-negative variables are employed to
linearize the objective function: e1s and e2s . Now the term s|θ − s s s |
P θ
in objective function is replaced with e1s + e2s when θs − s Ps θs equals to
e1s − e2s . Following model is the integrated model for SlARPM D .

min θM D = Ps θ s + λ Ps (e1s + e2s ), (39)
s s

s. t. θs − Ps θs = e1s − e2s , ∀s ∈ S. (40)
s
Constraints (2–10), (13), (22), (23), (27-38).

518 A. Amini et al.
4 Numerical Example
A numerical example is considered to analyze developed models. This exam-
ple consists of three potential depots, two vehicles for each, and six extremi-
ties for arcs. We assume that there is a link between each pair of extremities
and between each potential depot and each extremity. Potential depots belong
to set {1, 2, 3} and extremities are in set {4, 5, 6, 7, 8, 9}. In this example arcs
there are positive demands between extremities (4, 5), (4, 7), (4, 9), (5, 7), (5, 8),
(6, 7), (6, 8), (6, 9), and (8, 9). Arcs with positive demands must be served in
one direction. For instance, one of arcs 4-5 and 5-4 must be chosen for meet-
ing the demand between extremities 4 and 5. Five parameters of the model
are generated with patterns W ∼ Uniform(100, 200), F ∼ Uniform(400, 800),
k k
Cij ∼ Uniform(50, 100), Hij ∼ Uniform(10, 50), and Dij ∼ Uniform(100, 1000).
In addition, parameters of capacities are generated regarding the demands in a
way that infeasibility does not occur.
Also, five scenarios are defined. Here, scenario S1 has the lowest amount of
demands among all scenarios. Then, we assume that this scenario has the lowest
amount of capacities and costs. So, scenario S5 has the highest amount for
each parameter. All parameters of a scenario are generated by above-mentioned
patterns through dividing the ranges of each parameter. For example, range
(100,200) is assumed for W . This range is divided to five ranges to generate the
scenarios. Thus, ranges (100,120), (121,140), (141,160), (161,180), and (181,200)
are the ranges of the uniform distribution function of W for scenarios S1 to S5,
respectively. This process is employed for each parameter.
Table 1. Optimal depots and routings for each scenario
Scenarios Depot Vehicle Sequence Deadheaded

S1 3 1 3-7-6-9-8-3 3-7, 8-3
3 2 3-5-7-4-5-8-6-4-9-3 3-5, 6-4, 9-3
S2 2 1 2-7-4-9-8-6-2 2-7, 6-2
3 1 3-8-5-7-6-9-3 3-8, 9-3
3 2 3-4-5-3 3-4, 5-3
S3 2 1 2-4-9-6-8-5-8-9-2 2-4, 5-8, 9-2
2 2 2-6-7-4-5-7-2 2-6, 7-2
S4 1 1 1-8-5-7-4-9-6-1 1-8, 4-9, 6-1
3 1 3-7-6-8-9-4-5-3 3-7, 5-3
S5 3 1 3-6-9-4-5-4-7-5-8-3 3-6, 5-4, 8-3
3 2 3-9-8-6-7-3 3-9, 7-3
The mathematical models are solved by GAMS optimization software.

Table 1 presents the optimum locations of depots and routings for each scenario,
separately. Also, Table 2 presents the optimum objective function value (OFV)
Table 2. Values of objective functions
λ P set θs θR θM D
S1 S2 S3 S4 S5 λ=0 λ = 0.5 λ=1
S1 - - 1168.00 - - - - - - - -
S2 - - - 2137.00 - - - - - - -
S3 - - - - 1757.00 - - - - - -
S4 - - - - - 2545.00 - - - - -
S5 - - - - - - 2244.00 - - - -
θR - - 1811.00 2201.00 2533.00 2802.00 3263.00 1019.00 2522.00 3554.00 4586.00
θM D 0 α1 1752.00 2169.00 2520.00 2753.00 3298.00 1054.00 2498.40 2713.56 2928.72
θM D 0 α2 1752.00 2169.00 2520.00 2753.00 3298.00 1054.00 2243.70 2455.32 2666.94
θM D 0 α3 1750.00 2161.00 2531.00 2763.00 3296.00 1052.00 2326.50 2508.00 2689.50
θM D 0 α4 1752.00 2169.00 2520.00 2753.00 3298.00 1054.00 2497.40 2637.62 2777.84
θM D 0 α5 1752.00 2169.00 2520.00 2753.00 3298.00 1054.00 2656.90 2823.56 2990.22
θM D 0 α6 1752.00 2169.00 2520.00 2753.00 3298.00 1054.00 2765.90 2978.74 3191.58
θM D 0.5 α1 1752.00 2169.00 2520.00 2753.00 3298.00 1054.00 2498.40 2713.56 2928.72
θM D 0.5 α2 1752.00 2169.00 2520.00 2753.00 3298.00 1054.00 2243.70 2455.32 2666.94
θM D 0.5 α3 1752.00 2169.00 2520.00 2753.00 3298.00 1054.00 2327.10 2505.36 2683.62
θM D 0.5 α4 1752.00 2169.00 2520.00 2753.00 3298.00 1054.00 2497.40 2637.62 2777.84
θM D 0.5 α5 1757.00 2173.00 2531.00 2747.00 3297.00 1053.00 2657.40 2821.16 2984.92
θM D 0.5 α6 1811.00 2201.00 2533.00 2802.00 3263.00 1019.00 2773.40 2974.96 3176.52
θM D 1 α1 1752.00 2169.00 2520.00 2753.00 3298.00 1054.00 2498.40 2713.56 2928.72
θM D 1 α2 1752.00 2169.00 2520.00 2753.00 3298.00 1054.00 2243.70 2455.32 2666.94
θM D 1 α3 1752.00 2169.00 2520.00 2753.00 3298.00 1054.00 2327.10 2505.36 2683.62
θM D 1 α4 1752.00 2169.00 2520.00 2753.00 3298.00 1054.00 2497.40 2637.62 2777.84
θM D 1 α5 1757.00 2173.00 2531.00 2747.00 3297.00 1054.00 2657.40 2821.16 2984.92
θM D 1 α6 1740.00 2188.00 2542.00 2759.00 3290.00 1046.00 2769.00 2977.40 3185.80
of each scenario. We can find that scenarios do not have same results to each
other. It means that the value of parameters affects the optimum decisions even
when the arc with positive demands are same. Also it is found that if traversing
costs of each tour are dependent to its depot, decision makers cannot simply
declare that having multiple tours with one depot is better than having multiple
depots. For example, although scenarios S1, S3, and S5 prefer to use two tours
for the opened depot and not to open an extra one, S4 opens two depots with
one tour for each while it can assign another vehicle to each of them and remove
the other one. In fact, parameters of the problem indicate the best solution.
SlARPR and SlARPM D models have been taken into account after obtain-
ing optimal solutions of scenarios. Regarding the objective function of SlARPM D ,
a value for λ is required. This example assumes three different levels for λ: 0,
0.5, and 1. Also, this model needs to know the probability of each scenario in
advance. Considering different conditions, six sets, called P set , are given for the
following probabilities of scenarios: α1 = {0.2, 0.2, 0.2, 0.2, 0.2}, α2 = {0.4,
0.2, 0.2, 0.1, 0.1}, α3 = {0.2, 0.4, 0.2, 0.1, 0.1}, α4 = {0.1, 0.2, 0.4, 0.2, 0.1}, α5 =
{0.1, 0.1, 0.2, 0.4, 0.2}, α6 = {0.1, 0.1, 0.2, 0.2, 0.4}, where each set is defined as
{Ps1 ,Ps2 ,Ps3 ,Ps4 ,Ps5 }. Table 2 indicates the OFV of models under different values
for λ and set of probabilities and all equivalent OFVs for scenarios and other mod-
els. For instance, the optimum value of θR are fixed for scenarios and SlARPM D
520 A. Amini et al.
Table 3. Optimal depots and routings for the SlARPR and two cases of the SlARPM D
Model Depot Vehicle Sequence Deadheaded

SLARPR 1 1 1-4-9-6-7-5-1 1-4, 5-1
2 1 2-6-8-9-2 2-6, 9-2
2 2 2-7-4-5-8-2 2-7, 8-2
SLARPM D (λ = 0, α1) 1 1 1-8-5-7-4-5-1 1-8, 5-1
3 1 3-9-8-6-7-3 3-9, 7-3
3 2 3-4-9-6-3 3-4, 6-3
SLARPM D (λ = 1, α6) 1 1 1-6-9-8-5-4-1 1-6, 4-1
3 1 3-9-4-7-5-3 3-9, 5-3
3 2 3-8-6-7-3 3-8, 7-3
model their OFVs are measured. It is worth mentioning that we assume that sce-
narios have same probabilities when we want to fix the SlARPR optimum value of
decision variables in SlARPM D .
It can be figured out that each OFV is in its lowest level when its respective
model is optimized. Obviously, some other problems reached the same value. For
∗
example, θR and it equals to the equivalent value for θR when SlARPM D model
is optimized with λ = 0.5 and α6. Also, one can find some optimum decisions
in Table 3. It shows the effect of parameters and models in optimum solutions.
Minimizing the maximum regret of scenarios, depots 1 and 2 are opened while
depots 1 and 3 are opened when SlARPM D model is considered. In addition, λ
and P set affect the routing of arcs. Minimizing total costs, the models prefer not
to deadhead arcs as much as possible. Most deadheading traverses relate to the
linking arcs between depots and extremities.
5 Conclusion
This paper aimed at considering uncertainties in a location arc routing prob-

lem (LARP). First, a deterministic mathematical model was developed. Then,
two scenario-based approaches were employed to deal with the uncertainties of
parameters. The bi-objectives were to minimize the maximum regret (SLARPR )
and minimize the mean and deviation of the objective function value (OFV)
(SLARPM D ). It was required to find the optimal solution of each scenario in
both approaches. The SLARPR measured the distance between the OFV of each
scenario to its optimal value (i.e., the regret of each scenario), and minimized
the maximum one. On the other hand, the SLARPM D minimizes the mean of
OFVs of scenarios and their deviations from their optimum OFVs. A numeri-
cal example was generated to demonstrate the performance of models. Further
studies may take the other approaches of uncertainty into account in the LARP.
References
1. Beullens P, Muyldermans L, Cattrysse D et al (2003) A guided local search heuristic
for the capacitated arc routing problem. Eur J Oper Res 147(3):629–643
2. Bodin L, Levy L (1989) The arc oriented location routing problem. INFOR Inf
Syst Oper Res 27(1):74–94
3. Doulabi SHH, Seifi A (2013) Lower and upper bounds for location-arc routing
problems with vehicle capacity constraints. Eur J Oper Res 224(1):189–208
4. Drexl M, Schneider M (2014) A Survey of the Standard Location-Routing Problem.
Publications of Darmstadt Technical University Institute for Business Studies
5. Ghiani G, Laporte G (1999) Eulerian location problems. Networks 34(4):291–302
6. Ghiani G, Laporte G (2001) Location-arc routing problems. Opserach 38(2):151–
159
7. Nagy G, Salhi S (2007) Location-routing: issues, models and methods. Eur J Oper
Res 177(2):649–672
8. Pia AD, Filippi C (2006) A variable neighborhood descent algorithm for a real
waste collection problem with mobile depots. Int Trans Oper Res 13(2):125–141
9. Prodhon C, Prins C (2014) A survey of recent research on location-routing prob-
lems. Eur J Oper Res 238(1):1–17
10. Riquelme-Rodrı́guez JP, Gamache M, Langevin A (2016) Location arc routing
problem with inventory constraints. Comput Oper Res 76:84–94
11. Rui BL, Fea Plastria (2014) Location-arc routing problem: heuristic approaches
and test instances. Comput Oper Res 43(3):309–317
The Impact of Industrial Structure Change
on Economic Growth
Hongxia Liu(B) , Zongtang Xie, and Xin Song
School of Economics, Northwest Minzu University,

Lanzhou 730030, Gansu, People’s Republic of China
shirleygnd@163.com
Abstract. Taking case of Hebei Province, by building the regression

model, this paper estimates and measures the rationalization and high-
level of industrial structure effect on economic growth. The results show
that the rationalization of industrial structure and advanced industrial
structure can bring positive effects on economic growth. The effect of
industrial structure rationalization on economic growth is greater than
advanced industrial structure. In addition, the relationship between the
rationalization of industrial structure and economic growth is basically
stable in Hebei, while the relationship between advanced industrial struc-
ture and economic growth is instability. Finally, this paper puts forward
some suggestions to optimize the industrial structure to promote eco-
nomic growth.
Keywords: Industrial structure change · Economic growth, Rational-

ization of industrial structure · Advanced industrial structure
1 Introduction
Industrial structure and regional economic growth are interrelated and pro-
mote each other, the upgrading of industrial structure will promote the regional
economic growth, regional economic growth process accompany with industrial
structure change, reasonable evolution of industrial structure has become impor-
tant symbol of economic upgrading and modernization, which can promote the
potential power of regional economic growth. Domestic and foreign scholars make
a lot of research on the industrial structure and economic growth. In the late
1960s, Kuznets explored the process of economic growth in the United States
from 1948 to 1966, believing that industrial structural change was an impor-
tant driving force of economic growth. In the 1980s, Chenery et al. found that
economic growth was unbalanced, the production factors would flow from the
low-yield sector to the high-yield sector, and the effect of industrial structure
can drive economic growth [1]. At the beginning of this century, Peneder argued
that the differences in productivity and its changing trends among production
sectors would lead to the flow of the production factors in various sectors, the
The Impact of Industrial Structure 523
resulting structural dividend could cause economic growth [6]. Perez and Free-
man indicated the inherent relationship between industrial structure and the
whole economic and social development model and the level [2]. Jorgenson dis-
cussed the origin of US economic growth and its international comparison, and
revealed the role of the evolution of industrial structure [5]. Gan et al. took use
of stochastic frontier production function to research and find that industrial
structure and institutional factors could directly affect the economic scale, but
also through the indirect impact of resource allocation function, and then they
acted on the efficiency of output to promote regional economic growth [3].
2 Index Selection
2.1 Measurement of Industrial Structure Rationalization
Rationalization of the structure refers to the coupling relationship of input-
output in, reflecting the rational use of resources and the degree of coordination
among the various industries. The Theil index was originated from the entropy
concept in information theory, which was used by Theil to calculate the degree
of income gap in 1967. It is also scientific to use this index to measure the
rationalization of the structure. Referring to the relevant research of Chunhui
Gan et al. [3], this paper adopts the Theil index, the formula as following:
n n
Yi Yi LI Yi Yi Y
TL = ln / = ln / ,
i=1
Y Y L i=1
Y LI L
T L is the Theil index, Y is GDP, Yi is output value of various industries, L is

total employment in a region, Li is number of employment in various Industry.
T L is equal to 0, the industrial structure is reasonable, T L is not 0, the economic
deviates from the equilibrium state, the industrial structure is unreasonable. The
closer the index value is 0, the more reasonable the structure is.
2.2 Measurement of Advanced Industrial Structure

The sophistication is the dynamic process of industrial structure, evolving from
low-level to high-level. In this paper, the ratio of third industry and the second
industry (TS) is used to measure the high grade, which can not only reflect the
service-oriented tendencies of the structure, but also clearly indicate the develop-
ment direction of industrial structure. Thus, using TS to measure the advanced
industrial structure is rationality and scientific. The calculation formula is:
Y3
TS = ,
Y2
where T S is output ratio of the third and second industrial, Y3 is output value
of the third industrial, Y2 is output value of the second industrial. When T S
rises, it shows economic is the evolution of services, industrial structure is in the
continuous upgrading.
524 H. Liu et al.
2.3 Measurement of Economic Growth
This paper chooses the Gross Domestic Product (GDP) as a measure of economic
growth. Gross Domestic Product (GDP) refers to the market value of the final
products and services which a country or a region uses production factors to
produce in a given period. It not only reflects a country’s economic performance,
but also can reflect national strength and wealth.
3 Model Construction
In order to further explore the impact of industrial Structure rationalization and

advanced Industrial Structure on the economic growth of Hebei Province, this
paper uses multiple regression model, the model is set to
ln Gi = β1 + β2 ln T Li + β3 ln T Si + ui .
In the model, G is regional GDP, T L is the Theil index, referring to the ratio-
nalization of industrial structure, T S is output ratio of the third and second
industrial, to measure advanced industrial structure. In order to eliminate het-
eroscedasticity, each variable is taken as a natural logarithm, and u is the random
error term, β1 , β2 , β3 is parameter.
4 Empirical Analysis of Industrial Structure Change on

Economic Growth in Hebei Province
4.1 The Characteristics of Industrial Structure Change in Hebei
Province
This paper selects the relevant data from 1978 to 2012 in Hebei, and applies
Eviews7.0 software to process the relevant data. The calculated results of indus-
trial structure rationalization and advanced industrial structure are shown in
Table 8.
Table 1 shows: Theil index mean is 0.2324, the standard deviation is 0.0517,
the maximum and minimum are 0.3508, 0.1466. In 28 years, degree of indus-
trial structure rationalization kept steady. T S appears overall upward trend, its
mean is 0.6596, standard deviation is 0.0709, the maximum and minimum are
0.8160,0.5001, showing volatility rise (Table 2).
Figure 1 shows that from 1985 to 2012, T L changes from initial higher than
0.2 to later less than 0.2, showing that industrial structure is gradual in the
reasonable trend. At the same time, T S changes from the initial 0.4 and 0.6 to
0.6 and 0.8 later, industrial structure is a high-level forward in Hebei Province.
In addition, it should be emphasized that the correlation of T L and T S is not
strong (as shown in Table 1), and the change trend of them is not the same.
Thus, the interaction of T L and T S need not to be considered when analyzing
the impact industrial structure change on economic growth (Table 3).
Table 1. Indicators values of industrial structure rationalization and advanced mea-

surement in Hebei
Year T L TS Year T L TS
1985 0.2252 0.5001 1999 0.2275 0.6943
1986 0.2268 0.511 2000 0.2518 0.6777
1987 0.2317 0.5012 2001 0.2409 0.7069
1988 0.2679 0.6667 2002 0.2383 0.7383
1989 0.2831 0.671 2003 0.2455 0.7139
1990 0.2706 0.7249 2004 0.2156 0.6607
1991 0.3481 0.816 2005 0.2216 0.6337
1992 0.3398 0.7821 2006 0.2219 0.6375
1993 0.3508 0.6383 2007 0.1924 0.6388
1994 0.2551 0.648 2008 0.1969 0.6063
1995 0.2023 0.6769 2009 0.179 0.6773
1996 0.1858 0.6533 2010 0.1713 0.6653
1997 0.197 0.6502 2011 0.1667 0.6462
1998 0.2065 0.6626 2012 0.1466 0.6702
Table 2. Descriptive statistics
N Maximum Minimum Mean Standard deviation

VAR0001 28 0.3508 0.1466 0.2324 0.0517
VAR0002 28 0.816 0.5001 0.6596 0.0709
Table 3. Pearson correlation test of T L and T S
TL TS
TL Pearson correlation 1 0.094
Sig.tow-tailed - 0.693
N 28 28
TS Pearson correlation 0.094 1
Sig.tow-tailed 0.693 -
N 28 28
4.2 Regression Analysis Results
(1) Stationary Test of Time Series

In reality, most of the time series variables are non-stationary, which may lead
to pseudo-regression and invalid result in the regression analysis [4,7,8]. There-
fore, before testing the co-integration relationship between GDP and industrial
structure in Hebei, this paper adopts unit root to test stationary of time series.
526 H. Liu et al.
Fig. 1. Change trend of T L and T S
Table 4. Unit root test results
Variable Stationary difference order t-Statistic Test critical values (10%)

lnGDP level one −3.316089 −3.233456
lnTL level one −4.016404 −3.233456
lnTS level one −4.341045 −3.254671
ADF test results show that the t-test values of lnGDP, ln T L, ln T S are less
than the critical value at the significance level of 10%, indicating that the null
hypothesis is rejected at least confidence level of 90%, lnGDP and ln T L, ln T S
are integrated of order on. Thus, the long-term relationship between lnGDP and
ln T L, ln T S can be co-integration test.
(2) Co-integration Test
In the co-integration test, the Johansen maximum likelihood method is used
to test the co-integration relationship among variables. For a number of non-
stationary time series, if one linear combination is stationary, there exists a
long-term equilibrium co-integration relationship among these variables. The
trace statistic and the maximum eigenvalue statistic can be used to test. The
test results are shown in Table 5:
Table 5. Co-integration test results (trace test)
Hypothesized trace = 0.05

No. of CE(s) Eigenvalue Statistic Critical value Prob.∗∗
∗
None 0.860236 68.96358 29.79707 0
∗
At most 1 0.363917 19.76856 15.49471 0.0107
At most 2∗ 0.28703 8.457905 3.841466 0.0036
Note: ∗ indicates that the statistical significance is at 5% level, the
same below. ∗∗ indicates that the statistical significance is at 1%
level.
Table 6. Co-integration test results
Hypothesized max-eigen = 0.05

No. of CE(s) Eigenvalue Statistic Critical Value Prob.∗∗
∗
None 0.860236 49.19502 21.13162 0
∗
At most 1 0.363917 11.31066 14.2646 0.0059
At most 2∗ 0.28703 8.457905 3.841466 0.0036
Note: ∗ indicates that the statistical significance is at 5% level, the
same below. ∗∗ indicates that the statistical significance is at 1%
level.
The results of co-integration test in Tables 5 and 6 show that the trace sta-
tistic and the largest eigenvalue statistic are larger than the critical value at
the significance level of 5%, indicating that the hypothesis can be rejected at the
95% confidence level, indicating that there is co-integration relationship between
the variables. Thus, there are more than two co-integration relationships among
the three variables (lnGDP, ln T L, ln T S) at the significance level of 5%.
(3) Vector Error Correction Model

The above test shows that there exists co-integration relationship or long-term
stability relationship among the variables. On this basis, the VEC model will be
established to reflect the long-term static and short-term dynamic relationship.
The results are as follows:
Table 7. Vector error correction model
Error correction D(lnGDP) D(ln T L) D(ln T S)

CointEq1 0.009515 −0.000401 0.011106
−0.00351 −0.00666 −0.0055
[2.71014] [−0.06021] [2.01862]
D(lnGDP(−1)) 0.231887 −0.763078 −0.239289
−0.18074 −0.34268 −0.28324
[1.28297] [−2.22681] [−0.84484]
D(lnT L(−1)) 0.217572 0.041751 −0.072049
−0.10565 −0.20031 −0.16557
[2.05929] [0.20843] [−0.43517]
D(lnT S(−1)) −0.093157 0.507951 0.146124
−0.14085 −0.26704 −0.22072
[−0.66140] [1.90214] [0.66204]
C 0.117807 0.099722 0.046109
−0.0301 −0.05707 −0.04717
[3.91384] [1.74742] [0.97754]
528 H. Liu et al.
Table 7 shows the estimated values of the parameters of the VEC model,
where the value of CointEq1 represent the coefficient estimate of the error cor-
rection term. While the error equation can be derived from Table 4:
CointEq1 = D(ln GDP) + 0.23 ln GDP(−1) + 0.22 ln T L(−1) − 0.09 ln T S(−1) + 0.12.
This equation is deformed to the following equation:
D(ln GDP ) = CointEq1 − 0.23 ln GDP (−1) − 0.22 ln T L(−1) + 0.09 ln T S(−1) − 0.12.
This error equation shows: when ln GDP changes each additional unit, the
impact of ln GDP (−1) on it is −0.23 units, the impact of ln T L(−1) on it is
−0.22 units, the impact of ln T S(−1) on it is 0.09 units. In addition, Theil index
and GDP are negative correlation, lnT L rises one unit, D(ln GDP ) will drop
0.22 units; and TS and GDP show positive correlation, ln T S rises one unit,
D(ln GDP ) rises 0.09 units. This shows that the economy in Hebei Province
will grow with the industrial structure rationalization and the advanced indus-
trial structure, but its influence coefficient shows that the influence of industrial
structure rationalization on the economy is higher than that of the advanced
industry structure on economy in Hebei.
(4) Granger Causality Test
Table 8 shows: ln GDP and ln T L exist causal relationship, there is a mutually
reinforcing relationship between industrial structure rationalization and regional
economic growth; while advanced industrial structure and economic growth are
only a single causal relationship, economic growth can promote advanced indus-
trial structure, and advanced industrial structure can not promote economic
growth. GDP in Heibei increased from 39.675 billion RMB in 1985 to 2.65751
trillion RMB in 2012, while the output value of the secondary industry and
tertiary industry in Hebei Province has maintained a growing trend, indicating
that Hebei’s economic growth can promote industrial structure more reasonable.
From 1985 to 2012, the second industry in Hebei still dominates in economy. But
as time changes, the proportion of the first industry declines, the proportion of
tertiary industry continues to rise, the industrial structure is more rationalized.
This shows the stability of economic growth is mainly due to the rationalization
Table 8. Granger causality test results
Null hypothesis Obs F-Statistic Prob

InT L is not the Granger cause of lnGDP 25 4.90145 0.0116
Ln GDP is not the Granger cause of lnTL 9.17695 0.0007
lnT S is not the Granger cause of lnGDP 25 0.95535 0.435
Ln GDP is not the Granger cause of lnTS 3.77428 0.0291
LnT S is not the Granger cause of lnTL 25 1.41804 0.2702
lnTL is not the Granger cause of lnTS 0.76488 0.5285
of the structure, but advanced industrial structure is only one aspect of economic
growth effect, not to promote sustained economic growth.
5 Conclusion
This paper studies the impact of industrial structure dynamic change on

economic growth from the angles of industrial structure rationalization and
advanced industrial structure, and draws the conclusion that there is a long-term
equilibrium trend between industrial structure change and economic growth,
There is a stable relationship between the industrial structure rationalization
and economic growth, while the instability of advanced industrial structure is
more prominent; the industrial structure is the most important factor for the
economic growth in Hebei. Evolution of industrial change in Hebei is in line with
the general law, which will gradually move towards higher level.
Acknowledgements. This research was supported by the National Social Science

Foundation of People’s Republic of China (Grant NO. 11XMZ069), Scientific and Tech-
nological Project of Gansu Province of People’s Republic of China (NO. 148ZCRA012)
and the Special Fund of Fundamental Scientific Research Business Expense for Higher
School of Central Government (Northwest Minzu University).
References
1. Chenery HB, Robinson S, Syrquin M et al (1986) Industrialization and growth: a
comparative study, vol 4, pp 591–596
2. Freeman P (2000) Structural crises of adjustment, business cycles and investment
behavior. Tech Change Econ Theor 38(6):38–66
3. Gan C, Zheng R, Yu D (2011) The impact of china’s industrial structure change on
economic wth and volatility. Struct Change Econ Dyn 5(4):4–18 (in Chinese)
4. Ghosh SK (2015) The effect of imperfect production in an economic production
lotsize model. Int J Manag Sci Eng Manag 10(4):288–296
5. Jorgenson DW, Ho MS, Stiroh KJ (2003) Lessons from the us growth resurgence. J
Policy Model 25(5):453–470
6. Peneder M (2003) Industrial structure and aggregate growth. Struct Change Econ
Dyn 14(4):427–448
7. Tiemei G (2009) Economic analysis methodology and modeling - eviews application
and example. Beijing Tsinghua Univ 1:126–165
8. Tripathi RP (2016) Economic ordering policies under credit financing for deteriorat-
ing items with stock-dependent demand rate using discounted cash flow approach.
Int J Manag Sci Eng Manag 12:111–118
Disclosure Behavior of Annual Reports of Listed
Companies Under Digital Worship
Xiaojing Xu(B)
Chongqing Technology and Business University,

Chongqing 400067, The People’s Republic of China
34144333@qq.com
Abstract. Digital worship is widely used in China. Does this phenom-

enon also affects the China’s stock market? Does Digital Worship influ-
ence the choice of disclosure time of listed companies’ annual reports?
Through the empirical analysis for stock market of China, it founds that
the listed companies’ behavior of timing of annual reports released is
affected by the short-term influence of digital worship, which is a new
market phenomenon. The research of this paper has reference significance
for investment decision-making, risk management and understanding the
behavior of listed companies and investors.
Keywords: Digital worship · Annual report disclosure · Auspicious day
1 Introduction
Digital worship refers to the number of people to avoid or love [2]. The digital
worship is widely used in China, such as the choice of telephone numbers, busi-
ness opening date selection and so on. In this digital influence, people think that
figure “8” seems to be “fatten” homophonic, meaning fortune. People think that
the number “4” seems to be “dead” homophonic, and quite avoided. Such as
the 2008 Beijing Olympic Games elected in August 8th at 8: 8 pm to open, and
Alibaba Group was originally scheduled for August 8th , 2014 to be on the United
States stock market. Many studies have proved that there is a phenomenon of
worshiping number 8 in China’s stock market. In this paper the question posed
is whether this digital worship exists in the timing of listed company’s annual
report disclosure? If so, does it have a significant impact on the market? Specifi-
cally, on the one hand, if listed companies are affected by the number of worship,
whether they will choose date with the lucky number to publish annual report.
On the other hand, if the investors’ behavior is also affected by the digital wor-
ship, whether the listed companies will make good use of selecting the auspicious
figures to disclose annual reports to get more investors’ attention? If the digital
worship has a impact on the disclosure behavior, it will be a new market vision.
The research of this paper has reference significance for investment decision-
making, risk management and understanding the behavior of listed companies
Disclosure Behavior of Annual Reports of Listed Companies 531
and investors. The article will use the method of empirical analysis, with the
date ending in number 8 and number 4 as the starting point to demonstrate
whether the listed companies behavior of timing of annual report disclosure
is affected by the digital worship. The remainder of the paper is structured as
follows: The second part briefly reviews the literature. The third section describes
the data and makes assumptions. The fourth part introduces the econometric
model and empirical analysis. The fifth part summarizes the full text and makes
recommendations.
2 Related Literature
Cao, Li et al. [2] studied China is a very particular about the “homophonic cul-
ture” of the country, 8 homonyms are “fatten”, meaning fortune. And 4 homo-
phonic is “dead”, by a lot of people deliberately avoided. Xu [8] pointed out that
the Chinese people believe that the numbers can predict good and bad. With
social development, the number will have a new reference meaning and more
distinct meaning, the number “8” is repeatedly subject to people’s heat Hold-
ing. Shen [6] studied the symbolization and illusion of money, and discussed the
influence of digital worship on modern wealth. In the stock market, Zhao and
Wu [9] selected 569 non-financial stocks in Shanghai stock market as the object
of study, and got code ending with 8 stock price high, its price-earnings ratio on
IPO day and the following year higher than other mantissa stocks. Philip and
Jason [1] found that the stock valuation of Shanghai and Shenzhen stock market
there is a number of worship under the cluster effect, the price median with
8 appears twice the frequency of 4, because the Chinese inherent values that
the number 8 is auspicious number, 4 is unlucky numbers. The above document
proves that the stock code and the price contains the auspicious number will
have an impact on the stock price. Liu, Li and Yang [4] found that auspicious
number preference will significantly affect the relationship between the system-
atic risk of stock collapse and the expected rate of return. The above literature
proves that there is a phenomenon of digital worship in many areas of Chinese
stock market, and it has significant influence on the stock price and the expected
return rate. However the number of worship will affect the information disclosure
of listed companies and investors decision-making behavior, there is no relevant
literature, which is the research direction of this paper.
Human decision-making process is not based on purely rational, but there
will be a series of cognitive bias [5]. In a given context, people’s decision-making
behavior depends not only on the problem itself, but also on the way the problem
is expressed. Han and Xu [3] found that listed companies significantly lower or
early disclosure of the disclosure of annual reports will receive more attention
to investors, which reveals the listed companies through the choice of time to
disclose the annual report of an important channel for market value management.
Zhou and Huang [10] found that the disclosure of our annual report on the stock
market appeared “week preference” and “get together preferences”. To disclose
the good news on Tuesday, the stock excess return is higher. But the disclosure
532 X. Xu
of bad news on Saturday, the stock is also higher than the excess rate of return.
The stock has a positive impact on excess returns which the company chose the
time when a number of annual reports published at the same time to disclose
the bad news, in order to avoid the attention of investors. Wang and Wang [7]
studied the impression management theory of information disclosure strategy of
listed companies, and analyzed how listed companies can maximize their own
benefits through management information disclosure. Thus, listed companies
have strategy in the choice of disclosure of the timing, and they choose to publish
are going to catch the chance of timing to attract and distract the attention of
investors. This paper attempts to prove whether the listed companies use the
digital worship effect and select the mantissa is the auspicious number of dates
in order to achieve the effect of investor concern. In short, digital worship and
disclosure strategy belong to the category of behavioral finance, and it is also a
kind of market vision on the stock market.
3 Data and Assumptions

We collect and analyze the 2011–2015 Shanghai and Shenzhen A-share annual
report disclosure date (a total of 12,610 samples) were statistics, the results
shown in Fig. 1.
Fig. 1. 2011–2015 Annual report announcement date mantissa chart
As shown in the above figure, the number which listed companies choose to
publish the annual report in the date of mantissa 8 is the most, accounting for
15.01%. But the date selected in the mantissa of 4 is the least, accounting for
4.72%. The results shows that the date listed companies choose to publish the
annual report is Imbalance. There is a phenomenon of digital worship. However,
how much dose this phenomenon impact the stock market? In particular, if the
investor has the behavior of digital worship, the listed company chooses the
auspicious date to disclose the annual report in order to get more attention
from the investors when it chooses the disclosure time of the annual report. How
to measure the annual report after the release of more investors are concerned
about? This paper selects the 5-day yield index, which reflects the short-term
price volatility of the stock market. If the five-day rate of return is greater than
zero, then the stock received the attention of investors and investment, so that
the stock prices up in the short term. Conversely, the opposite is true.
In order to facilitate the study, this paper chooses the date of the mantissa
8 as the auspicious date and the date of the mantissa 4 as the unlucky date to
analyze whether the digital worship on the date will affect the disclosure time
of the listed company. So this paper puts forward the hypothesis,
Hypothesis 1 (H1): If it choose the date of the mantissa 8 to publish annual
report, the five-day stock returns will have a positive effect.
Hypothesis 2 (H2): If it choose the date of mantissa 4 to publish annual
report, the five-day return has a reverse effect.
On the other hand, listed companies are also under the digital worship of
praying auspicious, especially the annual report is bad news. They may choose
to disclose the annual report in the auspicious date so that corporate earnings
have a positive effect. The article sets that when the net profit margin is less
than or equal to last year’s net profit margin is bad news, when the net profit
margin higher than last year’s net profit margin is good news. Therefore,
Hypothesis 3 (H3): If the listed company which has good news in annual
report select the date of mantissa 8 to disclose annual report, not only have
a positive effect on the company’s five-day yield, but also have higher five-day
yield compared to other days.
Hypothesis 4 (H4): If the listed company which has good news in annual
report select the date of mantissa 8 to disclose annual report, not only have
a positive effect on the company’s five-day yield, but also have higher five-day
yield compared to other days.
4.1 Data Selection and Data Sources
This data selection follows these principles: (1) As the annual reports always
publish in the next year’s January to April, so financial data selected from 2010
to 2014 annual report. And the yield selected from 2011 to 2015 annual report
after the publication of five-day yield. Since the 2014 annual report is published
in 2015, so the annual report data is from 2014, and the yield data is collect in
2015 after the release of annual report. (2) Removing some of the outliers and
missing values; (3) The independent variables of the maximum and minimum of
3% flattening treatment. All the data in this article comes from the database of
Guotai’an. The empirical analysis is based on Eviews6.0.
4.2 Variable Selection

(1) Dependent Variable
Five-day rate of return (Rit ) = P1P−P
2
2
× 100%.
P1 is the annual report after the fifth trading day closing price. P2 is the
annual report published on the day closing price (if the annual report is published
after the end of the trading or in the non-trading day, we select the next trading
day closing price.
534 X. Xu
(2) Independent variable

Assets-liability ratio (X1it ) = Total liabilities/Total assets × 100%.
Operating profit margin (X2it ) = Operating profit/Total business income ×
100%.
Total assets growth rate (X3it ) = Total assets growth/Total assets at the
beginning of the year × 100%.
Net cash flow per share from operating activities (X4it ) = Net cash flow
from operating activities/Total share capital.
(3) Virtual Variable
This paper sets up the virtual variable to study whether the annual report can
attract investors’ attention and can cause the significant change of the 5-day
rate of return. In order to study the annual report on the end of the mantissa of
8 or 4 on the impact of the five-day rate, we set as following:
Firstly, the annual report published on the end of 8 is 1, not the end of 8 is
0, with DV1 it said. Secondly, Announcement date to 4 at the end is 1, not the
end of the 4 is 0, with DV2 it said. Thirdly if the listed company’s net profit
growth rate is less than or equal to last year is bad news, then the net profit
growth rate is greater than last year is good news. The good news is 1, the other
is 0, with DV3 it said. Finally, Annual bad news is published 1, the other is 0,
with DV4 it said.
4.3 Build Models

This article refers to the requirements of the panel data model, the model is as
follows:
Model 1 Rit = X1it + X2it + X3it + X4it + DV1it + µit ,
Model 2 Rit = X1it + X2it + X3it + X4it + DV1it DV3it + µit ,
Model 4 Rit = X1it + X2it + X3it + X4it + DV2it + µit ,
Model 6 Rit = X1it + X2it + X3it + X4it + DV2it DV4it + µit .
Rit is the five-day rate of return, X1it is on behalf of the asset-liability ratio,
X2it it represents operating profit margin, X3it represents the total assets growth
rate, X4it represents the net cash flow per share of operating activities, DVkit (k =
1, 2, 3, 4) represents the dummy variables, µit represents the random error term.

During the inspection period, the five-day average yield is about 0.2297%, greater
than zero. This shows that after the publication of the annual report, the major-
ity of listed companies received the attention and recognition of the market, and
stock prices rise. The maximum value is about 48.2540% and the minimum value
is −50.2257%. This shows that the annual report after the release of the stock
market response is very rapid, and investors can judge the tendency of markets
(Table 1).
Table 1. Variable descriptive statistics
Rit X1it X2it X3it X4it

Average value 0.2297 45.6519 8.564 16.8504 0.3557
Median 0 45.8901 7.064 10.5624 0.2712
The maximum 48.254 92.0426 42.0043 118.4737 2.4154
Minimum value −50.2257 6.6931 −30.0346 −20.5382 −1.2789
The number of observation 8805 12508 12273 11521 12508
The number of cross sections 2433 2487 2451 2438 2487
Table 2. Empirical results
Announcement date mantissa Announcement date mantissa

is 8 combinations is 4 combinations
Model 1 Model 2 Model 3 Model 4 Model 5 Model 6
X1it 0.0108∗∗∗ 0.0089∗∗∗ 0.0088∗∗∗ 0.0111∗∗∗ 0.0081∗∗∗ 0.0073∗∗∗
X2it −0.0077∗∗∗ −0.0055∗∗∗ −0.0054∗∗∗ −0.0006 −0.0026 −0.0042*
X3it −0.0025∗∗ −0.0040∗∗∗ −0.0045∗∗∗ −0.0048*** −0.0046∗∗∗ −0.0054∗∗∗
X4it 0.0342* 0.0333∗∗∗ 0.0427∗∗∗ −0.0192 0.0268 0.0126∗∗∗
DV1it 0.9320∗∗∗ - - - - -
DV1it DV3it - 1.0161∗∗∗ - - - -
DV1it DV4it - - 0.8512∗∗∗ - - -
DV2it - - - −1.6410∗∗∗ - -
DV2it DV3it - - - - −1.5512∗∗∗ -
DV2it DV4it - - - - - −1.8321∗∗∗
C −0.1687∗∗∗ 0.0798∗∗∗ 0.0876∗∗∗ 0.0026 0.1321∗∗∗ 0.2121∗∗∗
R2 22.01% 19.89% 88.11% 15.89% 2.43% 1.65%
F statistic 417.53∗∗∗ 351.01∗∗∗ 10201.19∗∗∗ 279.48∗∗∗ 31.98∗∗∗ 24.78∗∗∗
Note: ∗ , ∗∗ , ∗∗∗ respectively means significant at 10%, 5%, 1% level
4.5 Empirical Analysis
In this paper, a hybrid panel data model is used. Using the cross-section weight-
ing method, the empirical results are shown in Table 2.
4.6 Result Analysis
Based on the empirical results in Table 2, the analysis of the various research
hypotheses is as follows:
Firstly, the DV1 coefficient of model 1 is 0.9320, which indicates that the
listed company has positive effect on its 5-day rate of return on the date of the
ending date of 8, and compared to other dates, which can improve the 5-day
yield of 0.9320 units. So hypothesis 1 proves.
536 X. Xu
Secondly, as can be seen from model 2, the good news is better than other
dates by 1.0161 units of the five-day rate of return on the date of tail number 8.
So the good news announced in the date of tail number 8 can get more investors
attention. So Hypothesis 2 proves.
Thirdly, from Model 3 we can see that listed companies can increase the yield
of 0.8512 units by 8 days, which means that bad news can weaken the negative
effect of the stock market when the date is 8. So Hypothesis 4 gives evidence.
Finally, in the announcement date mantissa is 4 combinations, the model four
DV2 coefficient is −1.6410. It indicated that the listed company in the date
with tail number of 4 published annual report will reduce the 1.6410 unit rate
of return. So Hypothesis 2 gives evidence. From Model 5, we can see that the
good news is released on the date ending on the 4th , and the yield is reduced
by 1.5512 units compared with other dates. From Model 6, we can see that the
bad news is released on the date ending on the 4th, which will reduce the 1.8321
unit yield compared with other dates.

The influence of digital worship on the stock market belongs to the category of
behavioral finance. It is also a kind of market vision and a hot topic in the field
of finance research. From the perspective of digital worship, this paper analyzes
the motivations of time choice of disclosure of listed companies’ annual reports,
and draws the following conclusions.
First of all, the annual disclosure of listed companies by the impact of China’s
digital culture, there is a certain tendency to auspicious. The number of listed
companies which select the date with mantissa of 4 to publish annual report is
the lowest. But the number of listed companies which select date with mantissa
of 8 is the highest.
Second, whether the annual report is bad news or good news of listed compa-
nies, choosing the mantissa of the auspicious date to publish the annual report
has a positive effect on five-day yield, and the positive effect of the good news
is more obvious. Whether the annual report is good news or bad news of the
listed companies, choosing the date mantissa of 4 to publish annual report has
a reverse effect on five-day rate of return. This fully shows that digital worship
behavior exists in the listed companies when they choose the time to announce
annual report. While listed companies which has the bad news to use digital
worship in the market, and choose a auspicious day to publish annual report in
order to dilute the fundamentals of company unfavorable factors. On the other
hand, investor behavior is also affected by the number of worship.
Last but not the least, we give the suggestions. The paper chooses the 5-day
yield index as an empirical analysis, which is a short-term indicator. Therefore,
listed companies in order to continue to receive the attention and recognition
of investors should improve the quality of annual reports. Investors should also
be aware of the impact on trading behavior which comes from the cultural and
psychological factors, and finally make more rational investment decisions.
Acknowledgements. This paper is supported by The Ministry of Education of

China and Doctoral Program of Specialized Research Project Fund Project (No.
20120184110021).
References
1. Brown P, Mitchell J (2004) Culture and stock price clustering: evidence from the
peoples’ republic of China. Soc Sci Electron Publ 16(1C2):95–120
2. Cao B (2015) Developmental characters and reason analysis of number preference.
Adv Psychol 05(11):640–647
3. Han G, Xu H (2016) Operating performance of listed companies, disclosure of
annual report and investors’ concerns. China’s Econ Probl 4:17–34
4. Li L (2015) Cognitive bias of compound proportion relation and its interpretation
verification. PhD thesis, East China Normal University, Shanghai
5. Shen G (2016) Financialization, symbolism and illusion of money. J Wuhan Univ
Sci Technol (Soc Sci Ed) 18:478–482
6. Shengyao Liu LY, Yunhong Y (2016) Collapse of Chinese stock market systematic
risk and investor behavior preference. Finan Res 2:55–70
7. Wang X, Wang Y (2006) Theoretical basis of information disclosure strategy of
listed companies. Auditing Econ Res 2:84–87
8. Xu L (2012) A comparison of connotations of Chinese and western digital cultures.
J Zhejiang Univ Sci Technol 24:115–127
9. Zhao WF (2009) Asset price vision under digital worship. Econ Res 6:129–141
10. Zhou J, Huang D (2011) Investors’ limited attention and time choice of listed
companies’ announcements. Secur Mark Guide 5:53–60
Energy Finance in Promoting Renewable Energy
Sustainable Development in China:
The Channels and Means of Financing
Lei Xu1(B) , Yanfei Deng2 , Yuan Yuan3 , and Karen Mancl4

1
School of Economics, Xihua University,
leihsu@163.com
2
School of Management, Southwest University for Nationalities,
3
China Power Construction Corporation,
Sichuan Electric Power Design and Consulting Co., Ltd.,
Chengdu, People’s Republic of China
4
College of Engineering, Ohio State University, Columbus, USA
Abstract. As global climate warming and the continued use of fos-

sil energy are unsustainable, it has become imperative to accelerate
the development of renewable energy. As one of the largest developing
countries, this is especially true for future sustainable development in
China; therefore, focusing on renewable energy can significantly reduce
its dependence on imported fossil fuels. With policy support, China’s
renewable energy industries have had rapid development and are now at
the international forefront level in some fields. However, the renewable
energy industries are facing challenges, particularly in terms of energy
finance. Basically, the renewable energy industry in China is facing a con-
tradictory dilemma because of funding shortages and inefficient invest-
ment strategies, due in part to government-centered renewable energy
investment and financing. In view of this, the renewable energy financ-
ing and the corresponding polices in China are discussed in detail in this
paper. Issues of energy finance are also analyzed.
Keywords: Energy finance · Renewable energy · Financing channels ·

Financing means · Sustainable development
1 Introduction
Renewable energy (RE) has become the fundamental direction and core content
of the global energy transformation. As renewable energy is environmentally
friendly, low-carbon energy is the main force behind the technological revolu-
tion in energy and has become a major emerging industry growth point. China
enacted a series of polices to promote sustainable RE development to increase

Energy Finance in Promoting Renewable Energy 539
investment in RE industries and to gradually eliminate the obstacles to RE sus-

tainable development. These government actions have improved the RE subsidy
mechanism and incentives to ensure that renewable energy generation maintains
a steady upward development trend. As China is only in the middle stages of
industrialization and urbanization, energy demands continue to grow. The aver-
age annual RE installed capacity is expected to reach 2600 kilowatts or more
every year in the future. There is a need, for example, for new investment in
the solar RE industry of around 190 billion CNY to meet the goals of annual
new installed capacity of 20 million kilowatts. Wind and solar power develop-
ment together needs future annual investments of at least 400 billion CNY,
with a total RE investment of 2 trillion CNY required for the “13th Five-Year”
period [10].
International experts have now categorized RE into traditional and new, with
the former referring to giant hydropower and directly burnt biomass and the
latter referring to small hydropower, solar energy, wind energy, biomass energy,
geothermal energy and others [3]. The renewable energies discussed in this article
fall into the latter category. There are abundant RE resources in China; therefore,
to make good use of these RE resources, China needs to promote the development
of large new energy resources. Seeking to increase RE use and deployment to
encourage a move towards sustainable development, governments have developed
and implemented RE support policies. However, as the renewable energy sector
grows, it is facing increasingly more complex investment, financing and policy
challenges. Because of the initial high costs of renewable energy projects, market
development requires focused policy support. If RE financing mechanisms are
incomplete or poorly focused, renewable energy project financing has higher
costs and attracts less investment; therefore, innovative investment and financing
mechanisms are needed to break through the RE financing bottleneck and the
continued reliance on government guidance and incentives and to encourage
financial market innovations and policy incentives. Existing studies have found
that renewable energy laws and regulations as well as incentive policies are an
important driving force in the rapid RE development in China [1,5] such as
incentive policies, financial subsidy policies, tax rebates and exemption policies,
preferential feed-in tariff policies and technical support policies. China’s current
renewable energy policies include assistance with development plans, industrial
guidance, technical support, legal responsibility and popularization [4,11]. Based
on a questionnaire, Lam et al. found that the most important drivers of RE
development in China were government financial assistance, technical support,
feed-in-tariff policies and tax incentives.
2 The Driving Forces of RE Development

Strategically, it is important for China to develop domestic renewable energy
manufacturing and promote RE applications for several reasons. First, the pro-
motion and development of RE can ensure that China achieves its goal of 15%
non-fossil energy by 2020 and 20% by 2030. Second, RE gives developing coun-
tries an important opportunity to catch up with developed countries in terms of
540 L. Xu et al.
RE production capacity and to improve their energy structures [9]. Therefore,

with these goals in mind, it is important to assess China’s renewable energy
financing and policies issues to understand how such large scale investments in
renewable energies have been achieved, the major renewable power generation
investment and financing channels, the types and effectiveness of Chinese govern-
ment financing incentive mechanisms, the characteristics of China’s renewable
energy and whether the current policy mechanisms are suitable for the future
development of renewable energy. These issues are fully discussed in this paper.
Seeking to satisfy the growing energy needs and to reduce greenhouse gas
emissions, it is important that more RE resources be harnessed. However,
the wider consumption of RE resources is closely related to financial support.
Government financial support mechanisms and appropriate finance policies are
important to enhance RE sustainable development. It has been widely recog-
nized that renewable energy policies, such as incentive policies financial subsidy
policies, tax rebates and exemption policies, preferential feed-in tariff policies,
technical support policies, development plans, industrial guidance and technical
support, price incentives, legal responsibility and popularization, are essential for
the promotion of renewable energy development. The renewable energy market
share has been expanding and in recent years, China’s renewable energy has had
significant development because of the implementation of the Renewable Energy
Law and related policies. Existing studies have shown that incentive policies have
been an important driving force behind the rapid RE development in China [8].
3 RE Financing Channels in China
The Energy foundation and the National Development and Reform Commission
predicted that from 2005 to 2020, China’s needed to invest 18 trillion CNY into
the energy sector, with new energy, energy saving and environmental protection
measures needing about 7 trillion CNY [6]. Therefore, government subsidies
and investment alone is clearly insufficient, so full use needs to be made of
market financing using financial channels to resolve the funding shortages. It
has been pointed out that the development of appropriate financing channels
and instruments for both end users and the industry was one of the drivers for
increased investment [2]. This issue is analyzed in this part of the paper.
3.1 Bank Financing
Bank financing, such as policy-related loans and commercial finance, is an impor-

tant channel for the renewable energy industry. To implement the country’s new
energy industry policy, the National Development Bank strengthened financial
support for the new energy industry. Currently, bank financing plays a significant
role in facilitating investment into the renewable energy sector. The World Bank,
the Asian Development Bank and other international financial organizations pro-
vide low interest concessional loans to support investment into Chinese energy
sources, transportation, environmental protection, public utilities and other non-

profit projects. Of these, energy saving and emissions reduction have attracted
the most finance. The International Bank for Reconstruction and Development
also provides market-based (hard) loans and guarantees for Chinese renewable
energy projects to increase the share of clean energy.
3.2 International Financing
International financing sources have played a major role as incubator funds for
the development of rural renewable energy in China, with many renewable energy
projects in China having been financed by multilateral and bilateral organi-
zations. In China the internationally funding projects have included: “Capac-
ity building for the rapid commercialization of renewable energy”, a “China
renewable energy development project”, a “solar village project”, a “wind power
project in China’s Hubei” and a “clean energy research project”. International
financing channels for China’s renewable energy sector include international
financial institution loans, inter-government loans, non-government loans and
foreign investment, overseas Chinese investment in the territory, the Clean Devel-
opment Mechanism (CDM) and the Carbon Abatement Fund.
3.3 Public-Private Partnerships
Public-private partnerships involve cooperation between a public sector author-

ity and the private sector for the financing of revitalization, management or
infrastructure maintenance projects including renewable energy projects. Pri-
vate and public capital partnerships are generally joint ventures in which the
entrepreneurs and the government work together to provide a faster and more
efficient implementation of infrastructure projects. Also, because of the private
capital injection, the assets that the public sector would otherwise invest in
infrastructure projects in the traditional way can now be redirected towards cre-
ating social wealth. Additionally, there is a growing awareness that cooperation
with the private sector in such public-private partnership projects has a num-
ber of advantages: (1) acceleration, (2) more rapid implementation, (3) reduced
whole life infrastructure provision costs, (4) better risk allocation, (5) better
incentives to perform, and (6) improved service quality.
3.4 Government Financing
Government financing also plays an important role in financing renewable energy

in China, with the government taking a leading position in renewable energy sec-
tor financing. Government finance has been the most important channel for the
financing of renewable energy projects through the implementation of national
programs and the supply of grants to subsidize the capital costs of equipment,
loans, loan guarantees and tax incentives. The Chinese government has a played
a significant leading role in financing rural renewable energy through national
542 L. Xu et al.
initiatives and the provision of financial incentives. The Chinese Township Elec-
trification Program was the first nationally based implementation of renewable
technologies to supply electricity to rural populations and is one of the largest
renewable energy-based rural electrification programs in the world. The develop-
ment of such projects demonstrates the robust and sustainable renewable energy
infrastructure in China, especially for solar energy.
3.5 Capital-Market Financing

Capital-market financing is a channel for new energy enterprises to raise funds
as it can effectively solve fund shortage problems in a relatively short period of
time. However, China’s multi-level capital market is still somewhat behind as the
capital market, stock market, bond market and investment products are not yet
rich enough. China’s Shanghai and Shenzhen stock exchange markets are very
similar in terms of their organization, listing standards, trading patterns and
regulatory bodies. The most direct way to utilize bonds for RE financing would
be for RE companies to issue corporate bonds. China has had several successful
case-studies for raising investment funds for the RE sector. Green bonds have
emerged as a potential financing channel for RE, in which the issuer commits
to the use of bond proceeds only for environmentally-friendly products. The key
differentiating feature is that a green bond is marketed as green to investors and
their proceeds are restricted to green projects. Issuing a green bond also increases
the visibility of the bond to conventional investors. The ability to highlight the
green attributes of green bonds could also improve overall publicity and improve
an organization’s image, thereby further broadening their access to capital.
4 RE Financing Means in China

Financing means, such as corporate financing, project financing, Third party
financing and financing leases are relatively backward in China compared to
developed countries [7].
4.1 Project Financing

Project financing has become one of the most important financing means for
Chinese state-owned enterprises. A common project financing method is the
Build-operate-transfer (BOT) method, with concessions having been granted to
attract foreign and private investment. BOT was introduced as an alternative
project financing model to motivate private investment participation. The govern-
ment and power supplier sign a power purchase agreement, and to ensure that the
RE power plant supplies a timely transmission of electricity, the project company
receives payment for the power products from the power supplier. The renewable
energy project BOT structure is shown in Fig. 1. In addition, the project company
can apply for government authorized special renewable energy funds or can apply
for financial subsidies. This additional income (including government subsidies)
can assist the project company to cover the costs of the renewable energy plants
or to repay bank loans or syndicated loan principal and interest.
Fig. 1. BOT structure for renewable energy project
4.2 Financing Leases

Financing leases are a significant financing means for the development of China’s
renewable energy sector. Large scale renewable energy projects are less attrac-
tive because of the massive initial investment needed and the uncertainty risks.
Lease financing programs refer to the relationship between the enterprises selling
the RE equipment and a financial institution, for which the financial institution
provides financing for the purchase of the equipment. The typical structure is
depicted in Fig. 2. Finance leases for large-scale renewable energy power genera-
tion projects are initiated through commercial capital leases for the purchase of
power generating equipment and technology, which is then recouped from elec-
Fig. 2. The FL structure for an RE project

544 L. Xu et al.
tricity users by the project operators to pay for the lease financing. In this way, a
project sponsor can reduce their capital pressure as well as reduce the project’s
financial risk.
4.3 Third Party Financing

Third party financing (TPF) deals with customer sources, installation, engineer-
ing, maintenance and financing services for the RE system on the host customer’s
properties. The advantages of TPF contracts are that the source of funds is guar-
anteed and a number of services are provided which minimize commitment risk.
TPF can effectively aid in the development of RE technologies in developing
countries and rural areas with less investment. The third-party ownership model
is shown in Fig. 3.
Fig. 3. Third-party ownership model
4.4 Venture Capital

Venture capital or private equity investment facilitates the provision of funding
to firms in the industry that may have unproven or high risk technologies. From
an investment point of view, almost all new energy projects have a corresponding
capital risk, such as solar energy and wind energy related technological research
and development projects. Venture capital generally involves a private placement
of funds to assist in the establishment of emerging small and medium-sized RE
enterprises which tend to have high risk/high return characteristics. At present,
China’s venture capital investment in new energy sectors is still in a trial stage;
therefore, the investment is more cautious and the investment funds are often
quite limited.
5 Issues of Renewable Energy Finance

This paper has closely examined optimization and innovation in RE finance
issues to ensure RE sustainable development in China. From this close examina-
tion, there are several outstanding issues that need to be resolved to guarantee
the continued long-term sustainable success of RE in China. The main issues are
as follows:
(1) Incomplete financing systems for RE projects. Generally, investment options

have focused on the industrialization of RE technologies only rather than
on innovation. Second, domestic enterprises are not investing enough and
have not been sufficiently integrated into RE development chains. However,
for the RE true sustainable development, investment chains for renewable
industry sectors and investment integration is necessary to ensure that the
investments contribute to overall long-term sustainability.
(2) Lack of policy coordination and consistency. As there appear to be dif-
ferences in the goals of each of the RE sectors, it has been difficult to
ensure consistency in policy frameworks making it difficult to develop a long-
term and effective policy system to support renewable energy”s sustainable
development. Further, as energy issues encompass federal, state and local
governments and several different departments such as energy, science and
technology and agriculture, there has been inconsistency and a lack of coor-
dination in policy development across these boundaries, ultimately eroding
state macro-control and resulting in wasted investments due to duplication.
(3) Weakness Financial subsidies for renewable energy investment projects are
much lower than for other sectors and China’s support for renewable energy
through taxation methods is fairly weak, with the actual taxation benefits
for most renewable energy projects being similar to or of less benefit than
conventional energy tax benefits.
(4) Inadequate investment in technical R&D and weak independent RE devel-
opment innovation. At the moment, China lacks a powerful platform to
support technological research. What is needed is a clear long-term continu-
ous and rolling R&D development investment plan, with guaranteed funding
to support R&D. Most core technologies in most domestic enterprises come
from abroad, which will ultimately affect the sustainable development of the
RE sector.
6 Conclusion
Investment, financing and corresponding policies have played a significant role in
sustainable RE development. Investment, financing, finance means and related
laws, regulations and major incentive mechanisms in China have progressed sig-
nificantly in recent years. Various incentive mechanisms such as the RE prefer-
ential fiscal policies and R&D fund support for the commercialization of tech-
nologies has gone some way to decreasing the development costs of renewable
energy projects. In the future, joint financing and policies and support mech-
anisms for renewable power generation need to be further strengthened. Fiscal
and tax incentives have significantly relieved the financial burdens on renew-
able energy power generation enterprises and preferential tax policies for VAT,
income tax and import duties for the renewable energy power generation indus-
try as well as strong financial subsidies have provided invaluable support for
renewable power projects. Tariff incentives, in particular, have ensured reason-
able profits for renewable energy power generation enterprises. However, invest-
ment and financing renewable energy policies need to be strengthened further
546 L. Xu et al.
as China’s RE financing methods and development policies have some inherent

disadvantages. For the future, it is necessary to develop a range of coordinated
strategies to ensure long-term efficient developments for the RE industry sector.
In brief, this paper provides a comprehensive analysis of China’s RE sustainable
development. The major financing methods and identified incentive strategies
and approaches can assist researchers and industry practitioners gain a bet-
ter understanding of how RE sustainable development can be promoted by the
Chinese government. The lessons learned from this China study can also be
compared with or used to formulate innovative financing methods and incentive
polices for RE sustainable development in other countries.
Acknowledgements. This research is supported by the Program of National Nat-

ural Science Foundation of China under Grant (No. 71503216) and (No. 71401136),
the Social Science Planning Project of Sichuan Province under Grant (No. SC15C051),
the Central Universities Project under Grant (No. 2016NZYQN13), the Technology
Finance and Entrepreneurship Financial Research Center Project under Grant (No.
JR201507), and the System Science and Enterprise Development Research Center
Project under Grant (No. xq15c02).
References
1. Abolhosseini S, Heshmati A (2014) The main support mechanisms to finance
renewable energy development. Renew Sustain Energy Rev 40:876–885
2. Bobinaite V, Tarvydas D (2014) Financing instruments and channels for the
increasing production and consumption of renewable energy: Lithuanian case.
Renew Sustain Energy Rev 38:536–545
3. Cao X, Kleit A, Liu C (2016) Why invest in wind energy? Career incentives and
chinese renewable energy politics. Energy Policy 99:120–131
4. Center NRE (2016) The national renewable energy development report
5. Geng W, Ming Z et al (2016) China’s new energy development: status, constraints
and reforms. Renew Sustain Energy Rev 5:885–896
6. Koseoglu NM, van den Bergh JC, Lacerda JS (2013) Allocating subsidies to R&D
or to market applications of renewable energy? Balance and geographical relevance.
Energy Sustain Dev 17:536–545
7. Ming Z, Song X et al (2013) New energy bases and sustainable development in
China: a review. Renew Sustain Energy Rev 20:169–185
8. Ng TH, Tao JY (2016) Bond financing for renewable energy in Asia. Energy Policy
95:509–517
9. Shen J, Luo C (2015) Renewable and sustainable energy reviews. Overall review
of renewable energy subsidy policies in China—Contradictions of intentions and
effects. Renew Sustain Energy Rev 41:1478–1488
10. Yang XJ, Hu H et al (2016) China’s renewable energy goals by 2050. Environ Dev
20:83–90
11. Zhao ZY, Zuo J et al (2011) Impacts of renewable energy regulations on the struc-
ture of power generation in China—a critical analysis. Renew Energy 36:24–30
The Study on Factors Affecting the Attraction
of Microblog – An Empirical Study Based on
Sina Microblogging Site
Yue He, Yue Zhang(B) , Min Xiao, and Lingxi Song

luner.zhang@foxmail.com
Abstract. New media like microblogging site becomes increasingly pop-

ular in China nowadays. Microblogging site has become an important
social platform for information communication, and also an important
marketing channel for enterprises. However, marketing messages are
often submerged in a large amount of information on microblogging sites,
so how to get more attention for their marketing microblogs becomes a
concern of enterprises. Based on the data from Sina Weibo, we put for-
ward “Attraction Index of Microblog” to measure how much attention
a microblog has attract. Then we used several quantitative methods like
correlation analysis and association rules to find the factors affecting the
attraction of microblog, in order to provide some guiding ideas for mar-
keting users to improve the attraction of their microblogs. With empir-
ical studies, we found that microblogs published by authenticated users
are more attractive. Whether the microblog has media information does
not affect the attraction of microblog. The more fans and microblogs an
authenticated user has; the more attractive the user’s microblogs are.
For a non-authenticated user, the number of fans somehow affects the
attraction of their microblogs. The microblogs which are published at
7–10 am or 19–21 pm are more attractive.
Keywords: Microblog marketing · The attraction index of microblog ·

Influence factor · Association rules
1 Introduction
Sharing and transferring information, from useful knowledge to gossip, is an

essential part of social life. In modern society, this needs are met by the
social media, which allow people to share information more easily and rapidly.
Microblogging site has gained tremendous popularity since their introduction in
2005, with the number of registered users growing exponentially [12]. Twitter, a
microblogging service, is one of the most popular social network services [11].
548 Y. He et al.
However, in China, Sina Weibo, similar with Twitter, is currently the most
widely used microblogging platform with the most number of users. As of June
2016, the size of micro-blog users was 242 million, and the utilization rate is
34%, which has been gradually recovered compared with the end of 2015 [4].
The widespread use of Sina Weibo has attracted enterprises to pay attention to
microblog marketing. Enterprises need to use microblog for branding promotion
and products marketing.
With the rapid development of social network based on Web2.0, micro-blog
services as an important social platform, which has the advantage of informa-
tion sharing has attracted the participation of many network users. With the
rapid development of Internet and wide application in many fields, most of the
enterprises have begun to use the Internet for marketing activities of enterprises,
micro-blog can focus on his fans and fans, which can realize real-time informa-
tion sharing, business information dissemination. As an independent media on
social media, the commercial value of micro-blog’s business from the micro-blog
platform, a large number of users and huge traffic. Enterprises can use micro-
blog platform, and according to their own characteristics to implement a variety
of marketing activities. Therefore, the paper hopes to find what kind of micro-
blog is more likely to get people’s attention, and help enterprises to make better
social marketing.
2 Related Work
Since Twitter become people’s important social networking platform and the
important marketing channel of enterprises, there is a growing body of research
on Twitter. The characteristics of Twitter and its applications are studied in
various fields. Kwak et al. [18] studied the topological features of Twitter, and
indicated that Twitter serves more as a news medium than a social network. Fis-
cher and Reuber [5] use an inductive, theory-building methodology to develop
propositions regarding how effectuation processes are impacted when entrepre-
neurs adopt Twitter. Chen [3] studied how active Twitter use gratifies a need to
connect with others and the characteristics of user behavior.
There are many scholars have studied the user behavior characteristics on
social network like Twitter, and also studied the effect of user behavior on user
influence, information spread and information credibility. Cha et al. [1,2] stud-
ied the effects of user behavior characteristics such as number of fans and users’
authoritativeness on microblog diffusion, and found that microblogs of opinion
leaders, celebrities, politicians and entrepreneurs are more widespread. Räbiger
and Spiliopoulou [16] proposed a complete framework for supervised separation
between influential and non-influential users in a social network, finding that
there are predictive properties associated with the activity level of users and
their involvement in communities. These studies are focus on the influence of
user identity (such as user occupation and user’s authoritativeness) and user
behavior characteristics (such as number of fans and contact degree between
users and fans). They discuss about few influence of contents of tweet. Besides,
The Study on Factors Affecting the Attraction of Microblog 549
user influence, information credibility and information spread effect have a cer-
tain correlation with popularity or attraction of message, but they are not all
the same. Messages of influential users or of credibility may be not attractive.
And attractive messages may not have good spread effect (i.e., a tweet receives
many comments and likes, but fewer users’ retweeting. We may say this tweet is
attractive, but with less spread effect).
The largest microblogging website in China, Sina Weibo, born in 2009, also
attracts much attention from researchers. In a comparison of Twitter and Sina
Weibo, Gao et al. [7] analyzed the textual features, topics and sentiment polar-
ities of posts for two microblogging websites, revealing significant differences
between them. Liu et al. [14] analyzed the spatial distribution pattern of Chinese
tourism by microblog, and revealed the factors which affect the spatial distrib-
ution of tourism. Gao Jie [6] makes some discussion on the celebrity microblog
advertisement in Sina micro-blog. Combined with the features of government
microblog, Xie et al. [13] constructs a theoretical model, and use structured ques-
tionnaire to validate the model. Yu et al. [17] examined the key trending topics
on Sina Weibo, compared them with the observations on Twitter. They found
that, on Sina Weibo, trends emerge almost completely attributed to reposts of
entertainment content such as jokes and images, while on Twitter, the trends
are always due to current global events and news stories.
There is also some research about user behavior of popular microblogs on
microblogging sites. Qiu [9] qualitatively analyzed the influencing factors on
popular microblog from the five aspects of disseminator, dissemination contents,
audiences, dissemination channels and dissemination skills. Liu et al. [15] found
that source trustworthiness, source expertise, source attractiveness, and the num-
ber of multimedia have significant effects on the information retweeting. Guan,
Gao and Yang [8] found that male users are more likely to be involved in hot
events, and messages that contain pictures and those posted by verified users
are more likely to be retweet, while those with URLs are less likely. Lun et
al. [10] found that retweeting and commenting are distinct types of microblog-
ging behaviors. Retweeting aims to disseminate information in which the source
credibility and posts’ in formativeness play important roles, whereas commenting
emphasizes social interaction and conversation in which users’ experience and
posts’ topics are more important. Most of them think that a popular microblog
is retweeted many times or commented by many people, few of them consider
the number of “likes” as the factor to measure the popularities of the microblog.
On the basis of previous studies, this paper combine the forwarding number,
the number of comments and points like the number of calculation of “micro-blog
attention” score to measure micro-blog attention degree. And through correla-
tion analysis, association analysis to analyze the factors affecting micro-blog
attention.
550 Y. He et al.
3 Data Collection and Methods

3.1 Data Collection
We randomly extracted 107,328 original microblogs from Sina Weibo with the
LocoySpider 8.0. The microblogs were posted from March to April 2015. The
data include information about the user (including user’s nickname, the num-
ber of fans, followers (who the user follows) and microblogs, whether the user
is authenticated) and microblog (including microblog content, the number of
people who like it, the number of comments, retweeting times, whether with
multimedia information, posting time).
Because the data has different dimension or different order of magnitude,
which will affect the comparison and analysis of the data, so data standardization
is needed first. Data is standardized between the interval [0, 1] with the SoftMax
standardization method. The formula to standardize xi to xinew is as follows:
1
xinew = . (1)
1 + e−xt
The xt in the formulae (1) is calculated as:
xi − x̄
xt = . (2)
σ
3.2 MAI (“Attraction Index of Microblog”)

Liu et al. [15] measured the attraction of microblog with retweeting times. Lun
et al. [8] used number of comments to measure microblog attraction. On Sina
Weibo, users can not only “retweet” and “comment” a microblog, but also “like”
a microblog. If a user agrees with or appreciates a microblog, he/she may push
the button “like”. A microblog receives more “like”, which means the microblog
attracts more attention and love. Therefore, we believe that “like” can also
reflect the microblog attention. So the number of “like” should be added into
“Attraction Index of Microblog” model.
Compared with “like” and “comment”, retweeting is more unique in the
three kinds of user behaviors. Retweeting show that a user not only likes this
microblog, but is more willing to share with more people. Microblog through
retweeting layer upon layer, its mode of transmission is fission, which leads to
exponential growth of the number of users it attracts. Therefore, the retweeting
times in “Attraction Index of Microblog” model is in the form of exponential
function.
According to the characteristics of these three parameters, the number of
“like” and comments and the retweeting times should be combined to evaluation
microblog attraction. So “Attraction Index of Microblog” was built:
MAI = (αM Li + βM Ci ) eM Fi . (3)
M Li , M Ci and M Fi respectively represent the number of “likes”, comments

and retweeting times which are standardized. α, β respectively represent the
coefficients of the number of “likes” and comments. In order to determine the

coefficient α, β, we used the method of partial correlation analysis. Based on the
microblog data, we calculated the partial correlation coefficients of the number
of “likes” and comments with retweeting times respectively, which are as the
value of α, β.
Partial correlation analysis [10] is used to analyze the partial correlation
coefficients of two variables among multiple variables under the control of the
effect of other variables, in order to represent the degree of linear relationship.
Suppose there are x1 , x2 , x3 three variables, it is required to calculate the
partial correlation coefficients between the variables x1 and x2 , with the exclud-
ing the impact of variable x3 . The partial correlation coefficient is denoted as
r12,3:
r12 − r13 r23
r12,3 = . (4)
1 − r13
2 1 − r23
2
In formula (4), rij represents the simple correlation coefficient of variable xi

and xj , and i, j = 1, 2, 3.
The results of partial correlation analysis with SPSS19.0 are shown in Table 2
Table 1. The partial correlation coefficients of the number of “likes” and comments
with retweeting times respectively.
Number of “likes” Number of comments

Retweeting times 0.457 0.523
Table.1 shows that the coefficients of the number of “likes” α = 0.457 and
the coefficients of the number of comments β = 0.52.
Eventually, “Attraction Index of Microblog” formula was:
MAI = (0.457M Li + 0.523M Ci ) eM Fi . (5)
3.3 Association Rules
Microblog with pictures, video and other multimedia information will bring more
visual or auditory stimuli, so these kinds of microblogs more attractive. If the
user is authenticated, whose microblog are more reliable and authoritative. So, it
is more likely to be commented or retweeted by others, the microblog attraction
may also be higher.
We used Apriori model in Clementine 12.0 to analyze the relationship
between MAI and microblog multimedia information, users’ authoritativeness
respectively. We set the support level with 5%, and the confidence level with
10%.
552 Y. He et al.
Association rules is used to find out the relevance hidden in the database,
and expressed in the form of rules. It is shaped like an implication “X ⇒ Y ”.
The associational rule refers to the sets of attribute-values, which frequently
appeared in data set recognition, also named as frequent item-sets. Using these
frequent sets, association relation rules process is set up. In association rules,
count (X) is the number of tuples that object set D contains set X. Support for
set X is:
count (X)
Support (X) = . (6)
|D|
Confidence is the proportion of the number of tuples that contains set X and
Y and the number of tuples that contains set X:
count (X ∩ Y )
Confidence (X ⇒ Y ) = . (7)
count (X)
Apriori algorithm proposed by Agrawal in 1994 is a relatively simple and

general algorithm. The algorithm has two key steps: the first is to find all frequent
item sets, and the second is to generate strong association rules.
4 Results
4.1 Descriptive Statistics of MA
For the 107328 microblogs, according to the formula (5) we calculated each
microblog’s MAI. To limit the value of MAI within [0, 1], we used formula (8)
to standardize the value of MAI, then the final MAI of each microblog was got.
xi − xmin
xnew = . (8)
xmax − xmin
The average of MAI is 6.62×10−3 . The minimum value is 0, and the maximum
value is 1. The MAI of most microblogs was low. The number of microblogs
whose MAI is equal to 0 is 65270, accounting for 60.8% of the total number of
microblogs. The number of microblogs whose MAI is greater than the average
MAI is 14726, accounting for 13.7% of the total number of microblogs. The MAI
distribution is shown in Fig. 1. Due to the small value of MAI, we expanded the
value 100 times to display (because the microblogs whose MAI greater than 0.1
is few, Fig. 1 is not in full).
According to the MAI, microblogs were divided into two categories: attractive
microblogs and unattractive microblogs. Microblogs whose MAI is greater than
the average MAI are attractive microblogs, while Microblogs whose MAI is less
than the average MAI are unattractive microblogs. The following analysis is
based on this.
Fig. 1. The MAI distribution situation
4.2 Microblog Multimedia Information and User Authoritativeness

The results of association rules analysis is shown in Table 2:
Table 2 shows:
• Whether the microblog with multimedia information does not affect the
microblog attraction.
• Whether the user is authenticated has effect on the microblog attraction.
The microblogs posted by authenticated users are more attractive than those
posted by non-authenticated users.
Table 2. The extraction results of association rules of MAI
Y X support (X) confidence (X ⇒ Y )

With multimedia Attractive 13.72% 72.36%
Attractive With multimedia 74.52% 13.43%
Authenticated user Attractive 13.72% 21.49%
Attractive Authenticated user 8.69% 34.21%
The distribution of MAI of the microblogs respectively post by authenticated

users and non-authenticated users is shown in Fig. 2. As we can see, the MAI of
microblogs posted by non-authenticated users are lower the authenticated users.
It also confirmed whether users’ authoritativeness is a key factor affecting the
microblog attraction.
4.3 Information of User Behavior

The Spearman correlation coefficients between MAI and the number of user’s
fans, followees and microblogs respectively are shown in Table 3.
554 Y. He et al.
Fig. 2. The distribution of MAI of the microblogs respectively post by authenticated

users and non-authenticated users
Table 3. The Spearman correlation coefficients between MAI and the number of user’s
fans, followers and microblogs respectively
The Spearman correlation coefficient N fana N folb N mcbc

MAI Authenticated user 0.522∗∗ 0.113 0.241∗∗
Non-authenticated user 0.271∗∗ 0.074 0.097
a
: N fan: The number of fans. b : N fol: The number of followees. c :
N mcb: The number of microblogs. ∗∗ : Correlation is significant at
the 0.01 level (2-tailed).
Table 3 shows that for different users, the effect of user behavior on microblog
attraction is different. For authenticated users, the number of fans have a great
effect on the microblog attraction, with the correlation coefficient reached 0.522,
followed by the number of microblogs, and the number of followees has little
effect. For non-authenticated users, the number of fans has less effect on the
microblog attraction, and the number of followees and microblogs has almost no
effect.
4.4 Posting Time

Generally speaking, most people’s leisure time is the two period of time at noon
and in the evening, so we can guess that microblogs posted in the two period of
time are more than other time. So, how about attractive microblogs? Are there
also some regularities the posting time of attractive microblogs follow?
In order to study the effects of microblog posting time on microblog attrac-
tion, we analyzed the posting time of attractive microblogs. We calculated
the number of total microblogs and attractive microblogs per hour, and com-
pared the two distributions of posting time of total microblogs and attractive
microblogs.
Due to the large gap of the absolute number of total microblogs and attrac-
tive microblogs, it is not suitable to compare the absolute value directly. So,
we calculated the proportion of total microblogs and attractive microblogs per
hour respectively. For example, the proportion of total microblogs per hour is
calculated by dividing the number of total microblogs per hour by the number
of all microblogs (107328), so as the proportion of attractive microblogs. The
posting time distribution of total microblogs and attractive microblogs is shown
in Fig. 3.
As Fig. 3 shows, the posting time distribution of total microblogs conforms
to the guess that microblogs posted at noon (11:00–13:00) and in the evening
(21:00–23:00) are more than that posted at other time. And there are few
microblogs posted between 1:00–7:00 am. We can also find that attractive
microblogs posted between 8:00–11:00 and 19:00–21:00 are more than that at
other time, and the fluctuation is more obvious.
Fig. 3. Posting time distribution of total microblogs and attractive microblogs
In order to observe the posting time distribution of attractive microblogs

vividly, the number of attractive microblogs per hour was divided by the number
of total microblogs per hour. The proportion of them is shown in Fig. 4, which
will eliminate the interference of regularities of the posting time distribution of
total microblogs.
From Fig. 4, we can obviously see that the microblogs posted at 8:00–11:00
and 19:00–21:00 are more attractive. Besides, comparing the posting time of total
microblogs and attractive microblogs, we can find that attractive microblogs
are posted earlier about two hours than total microblogs. According to our
real-life, we can explain that most people at the leisure time (11:00–13:00 and
21:00–23:00) are likely to see the microblogs posted 2 or 3 h in advance. So,
these microblogs are more likely to be read, increasing the possibility to be
retweeted/comment/liked, resulting in becoming attractive microblogs.
556 Y. He et al.
Fig. 4. Proportion of the attractive microblogs within all microblogs per hour
5 Summary and Discussion

Through the empirical research, we found the factors affecting the microblog
attraction as follows:
• Users’ authoritativeness. If the user is authenticated, he/she is more credible
and popular, so his/her microblogs also can obtain others’ trust and concern,
which makes the microblogs attractive.
• The information of user behavior. For the authenticated users, factors affect-
ing the microblog attraction is the number of fans, followed by the num-
ber of microblogs. The more the number of fans or microblogs, the stronger
the microblog attraction is. For non-authenticated users, factors affecting the
microblog attraction is the number of fans, but it does not affect the microblog
attraction as much as the number of authenticated users’ fans does.
• Microblog posting time. The study shows that attractive microblogs are
mostly posted between 8–11 a.m. and 7–9 p.m., so microblogs posted in these
two periods of time would attract more attention.
Our study possesses certain innovation and pertinence. We put forward
“Attraction Index of Microblog” to definitely measure how much attention the
microblog has attract. And we used several quantitative data analysis methods
to analyze how these factors affect the microblog attraction. The findings of our
study also have some practical implications. Our study provides some guiding
ideas for Weibo users to improve the attraction of their microblogs. Weibo users
especially advertisers who aim to market their product and disseminate their
corporate culture through microblogging sites, for example, must try to obtain
authentication to attract more fans. And they must keep posting microblogs to
provide enough messages. It is not necessary to insert media information like
pictures or videos into microblogs. It is also important to post microblog in the
two period of time, which can make the microblog be seen by more people.
6 Limitations and Directions for Future Studies

This paper has certain innovation, but there are still shortcomings and limi-
tations, which are worthy of exploration and improvement for future research,
mainly in the following points:
• We only studied the effects of microblog post form and the user information to
the microblog attraction, not considering the influence of microblog content.
So in the following research, we can continue to study the effects of microblog
content to its attraction, including the microblog sentiment, and topic types
etc.
• For the analysis of each factor affecting the microblog attraction, we only
used a method. However, different methods as the principle or perspective is
different, may lead to different results. Therefore, in the future study, we can
analyze the same problem with different methods, and compare the difference
of results to choose the more appropriate methods and results.
• The current study examined microblogging behavior using Sina Weibo in the
Chinese context. Other social networking sites like Twitter, Facebook have
much similarity to microblogging site, but in many ways, they are different,
such as different technology, different culture, and different user behaviors
between different platforms. Therefore, it would be unacceptably risky to con-
clude that the findings of this study represent a cross-contextual evaluation
of microblogging behavior. Future researchers should make extensive efforts
to explore the difference of user behavior between different social networking
platforms.
Acknowlegement. The authors acknowledge the financial support National Natural

Science Foundation Project of China, project (No. 71372189), and Sichuan University
Fundamental Research Funds for the Central Universities Project (No. skqy201406).
The author is grateful to the anonymous referee for a careful checking of the details
and for helpful comments that improved this paper.
References
1. Cha M, Haddadi H et al (2010) Measuring user influence in twitter: the million
follower fallacy. In: International conference on weblogs and social media, ICWSM
2010, Washington, DC, USA, May 2010
2. Cha M, Benevenuto F et al (2012) The world of connections and information flow
in Twitter. IEEE Trans Syst Man Cybern Part A Syst Hum 42(4):991–998
3. Chen GM (2011) Tweet this: a uses and gratifications perspective on how active
twitter use gratifies a need to connect with others. Comput Hum Behav 27(2):755–
762
4. CNNIC (2016) The 38th statistical report on internet development in China. Tech-
nical report. http://www.cnnic.net.cn/ (in Chinese)
5. Fischer E, Reuber AR (2011) Social interaction via new social media: (how)
can interactions on twitter affect effectual thinking and behavior? J Bus Ventur
26(1):1–18
558 Y. He et al.
6. Gao J (2016) Research on the advertising value of celebrity micro-blog. PhD thesis,
Jilin University, Changchun (in Chinese)
7. Gao Q, Abel F et al (2012) A comparative study of users’ microblogging behavior
on Sina Weibo and Twitter. Springer, Heidelberg
8. Guan W, Gao H et al (2014) Analyzing user behavior of the micro-blogging website
Sina Weibo during hot social events. Phys A Stat Mech Appl 395(4):340–351
9. Jin Q (2013) An analysis of the influencing factors on popular micro-blog. PhD
thesis, Zhengzhou University (in Chinese)
10. Kim JM, Jung YS et al (2011) Partial correlation with copula modeling. Comput
Stat Data Anal 55(3):1357–1366
11. Ko J, Kwon HW et al (2014) Model for Twitter dynamics: public attention and
time series of tweeting. Phys A Stat Mech Appl 404(24):142C149
12. Li F, Du TC (2014) Listen to me - evaluating the influence of micro-blogs. Decis
Support Syst 62(2):119–130
13. Liang CC (2015) Factors influencing office-workers’ purchase intention though
social media: an empirical study. Int J Customer Relat Mark Manage 5:109–116
(in Chinese)
14. Liu DJ, Jing HU et al (2015) Spatial pattern and influencing factors of tourism
micro-blogs in China: a case of tourism Sina micro-blogs. Sci Geogr Sin 35(6):717–
724
15. Liu Z, Liu L, Li H (2012) Determinants of information retweeting in microblogging.
Internet Res 22(4):443–466
16. Räbiger S, Spiliopoulou M (2015) A framework for validating the merit of proper-
ties that predict the influence of a Twitter user. Expert Syst Appl 42(5):2824–2834
17. Yu L, Asur S, Huberman BA (2011) What trends in Chinese social media? arXiv
preprint arXiv:107:3522
18. Yu S (2014) Microblogging becomes the first listed Chinese social media. Accessed
People’s daily: http://tech.sina.com.cn/i/2014-04-18/09309329858.shtml
A Genetic Algorithm Approach for Loading
Cells with Flow Shop Configuration
Patrick Gannon and Gürsel A. Süer(B)
Department of Industrial and Systems Engineering,

Ohio University, Athens, OH 45701, USA
suer@ohio.edu
Abstract. This paper proposes a three-phase methodology for worker

allocation and flowshop scheduling in a multistage manufacturing envi-
ronment. A case study using a shoe manufacturing plant is examined.
The proposed methodology consists of genetic algorithm and mathemat-
ical models. Phase 1 allocates workers in between operations in two sep-
arate stages in the manufacturing process. Phase 1 evaluates how work-
ers are allocated between and within the manufacturing stages. Phase
2 uses worker allocations from phase 1 to perform cell loading based
on machine-level-based similarity using genetic algorithms. Four differ-
ent Genetic Algorithm approaches are proposed and evaluated. Further-
more, the impact of different manpower levels are also studied. Phase 3
schedules products within cells to optimize makespan, total tardiness, or
number of tardy jobs.
Keywords: Genetic algorithms · Cell loading · Flowshop scheduling ·

Worker allocation
1 Introduction
The goal of this paper is to provide a methodology to most effectively utilize
available manpower and optimize the scheduling of products with regards to
multiple performance measures. The proper utilization of resources is crucial
to success and survivability of any company and the proposed methodology
seeks to improve utilization of resources in multi-stage labor-intensive manu-
facturing environments. Using a shoe manufacturing plant as a case study, this
paper addresses the problems of manpower allocation, cell loading, and flowshop
scheduling. The shoe manufacturing plant has several cell groups, consisting of
a lasting cell (LC), a rotary machine cell (RMC), and a finishing cell (FC) as
shown in Fig. 1. The LC and FC are labor intensive manufacturing cells consist-
ing of several sequential operations. A simplified version of the RMC is used in
this paper, so the majority of the focus of this paper is on the LC and FC. There
are fewer operation types in the LC and FC than there are workers available,
560 P. Gannon and G.A. Süer
so several operators can be assigned to each operation. With correct manpower

allocation, a balance can be achieved and the cells’ output can be maximized.
Considering the worker allocation and capacity requirements cell loading can be
performed. By carefully assigning products to cells, the benefits of cellular man-
ufacturing, such as reductions in flowtime, setup times, and work-in-progress
inventories, can be realized. Once products are assigned to cells, scheduling
is performed in each cell group independently. Scheduling is necessary when
resources are limited and need to be optimally utilized for several competing
tasks.
Fig. 1. Shoe manufacturing system
1.1 Products
This plant produces shoes with a variety of designs, colors, and materials. The
shoes are available in two sole types, full shot or midsole. Of these two sole types,
the shoes can be made using either PVC or TPR and are available in three colors
(black, honey, and nicotine). The breakdown of this product structure is shown
in Fig. 2.
The demand for shoes is based on their size. The sizes range from size 5 to
size 15 and demand is greatest closest to middle of the size range. The demand
curve shown in Fig. 3 demonstrates the relationship between size and demand.
1.2 Rotary Machine
The Rotary Machine Cell (RMC) is the bottleneck of the manufacturing cell. The
RMC has six available positions, each of which can process a pair of shoes. The
Rotary Machine also has a tank to hold the material to be injected. Unprocessed
shoes are loaded into the machine and injected with either PVC or TPR in order
A Genetic Algorithm Approach for Loading Cells 561
Fig. 2. Product types
Fig. 3. Typical shoe demand by size
Fig. 4. Rotary injection molding machine
to form the soles of the shoes. A setup period is needed to switch materials or
colors to be injected since the machine possesses only one tank to hold the
material.
The Rotary Machine consists of a loading/unloading station, an injection
station, and four cooling stations. Figure 4 illustrates the setup of the rotary
machine. A shoe’s processing time on the Rotary Machine is influenced by the
mold type (full shot or mid-sole), the size of the shoe, and whether it is designed
for males or females. The average injection molding time for full shot shoes is
0.356 min for male shoes and 0.3 min for female shoes. Midsole shoes require less
processing, having average injection molding times of .286 min for male shoes and
.25 min for female shoes. Table 1 shows injection times for the various designs
at the available sizes. Times are classified as Fullshot (FS), Midsole (MS), male
(M) and Female (F).
Table 1. Injection Times
FS/MS M/F Different sizes

1 2 3 4 5 6 7 8 9
5 5.5 6 6.5 7 7.5 8 8.5 9
FS M - 0.224 0.247 0.269 0.281 0.289 0.290 0.301 0.306
FS F 0.154 0.207 0.216 0.229 0.23 0.237 0.308 0.338 0.356
MS M - - 0.160 0.162 0.203 0.239 0.24 0.242 0.245
MS F 0.109 0.116 0.183 0.190 0.196 0.260 0.261 0.290 0.305
FS/MS M/F Different sizes
10 11 12 13 14 15 16 17 18
9.5 10 10.5 11 11.5 12 13 14 15
FS M 0.32 0.342 0.372 0.426 0.428 0.448 0.460 0.499 0.556
FS F 0.368 0.395 - 0.424 - 0.440 - - -
MS M 0.247 0.273 0.293 0.311 0.349 0.398 0.401 0.405 0.415
MS F 0.31 0.342 - 0.344 - 0.347 - - -
Shoes loaded into the rotary machine cannot be unloaded until all shoes in
the machine have been processed. This leads to shoes with lower processing times
spending more extra time in the machine relative to shoes with higher processing
times. This extra time leads to extra material being injected into their molds.
Because of this, it is beneficial to minimize variations in the processing times of
shoes that are to be loaded together. However, this problem has been addressed
in prior works [6,12,16], so for the purpose of this paper, one size will be assigned
at a time in the rotary machine.
1.3 Molds
Different molds are needed for processing shoes depending on size and type (Full
Shot or Mid Sole). The molds carry a prohibitive cost, limiting their availability.
As shoes of varying size and type are processed in the RMC, the molds have
to be constantly changed in order to process the corresponding shoe. Also, the
molds have to be cleaned when the material or color is changed. The problem
considering a limited number of molds for each size has been addressed previously
in the literature. For this study, it is assumed that there are no mold availability
restrictions.
2 Literature Review
This paragraph will briefly discuss some prior cell loading and manpower deci-
sion work done in the literature. Süer, Saiz, Dagli, and Gonzalez [14] expanded
research in the cellular control area to multi-cell environment, which necessitates
cell loading considerations. Süer [13] developed a two-phase methodology to find
optimal manpower allocation and cell loads simultaneously. An extension of this
work was done by Süer [13]. This work add lot-splitting considerations to the two
phase approach to find optimal manpower allocation and cell loads and found
that lot splitting is advantageous when setup times are negligible. Akturk [1]
addressed the production planning problem in cellular manufacturing systems.
By using capacity constraints to evaluate the impact of cell loading decisions on
lower levels, they were able to more accurately portray the cellular manufactur-
ing system and its operation. Saad [9] developed a multi-objective optimization
technique to load and schedule cellular manufacturing systems using simulation
and a taboo search algorithm. Babayigit [2] analyzed the problems of manpower
allocation and cell loading with a mathematical model and genetic algorithm,
further extending previous works by Süer [13,14].
Süer [15] proposed a three-phase methodology to address cell loading and
product sequencing in labor-intensive cells. Optimal manpower allocation and
cell loads are found in the first two stages and the third stage treats product
sequencing as a travel salesman problem to minimize intra-cell manpower trans-
fers. Some scheduling work done in the literature is discussed in this paragraph.
Nakamura [8] considered group scheduling on a single stage to minimize total
tardiness and found that scheduling problems of moderate size can be ordered to
reduce the number of schedules to be searched. Sinha [10] found that a desired
level of throughput and optimum work-in-progress in a cell can be achieved
through sequencing, reduced batch size, and period batch control. Gupta [5]
developed an optimization algorithm and heuristic procedures to schedule jobs
on two identical machines. The objective was to find a schedule with optimal total
flowtime that gives smallest makespan. Kim [3] used a restricted taboo search
algorithm to schedule jobs on parallel machines in order to minimize the max-
imum lateness of jobs. Gupta [4] studied single machine group scheduling with
sequence-independent family setup times in order to minimize total tardiness.
The heuristics developed were shown to be effective in minimizing total tardiness.
Subramanian [11] focused on cell loading and job scheduling at a shoe man-
ufacturing plant. The objective of this research was to develop heuristics for cell
loading and combine these with simple scheduling methodologies to load all jobs
without exceeding capacity. This capacity was set using a MaxCap methodol-
ogy, which is the maximum capacity to be loaded. This methodology is critical
in determining the number of cells required and how the load is spread across
different cells in the facility. An interesting finding from this research is that
the optimal result from the cell loading stage may not remain optimal after the
cell scheduling stage. Urs [7] developed three scheduling heuristics using basic
machine scheduling methodology. The heuristics aimed to minimize makespan in
the rotary machine scheduling. The best performing heuristic was the minimum
difference in cycle time heuristic. This research stands out from prior works by
assuming a limited number of molds for each size, meaning multiple sizes can be
run on the Rotary Molding Machine simultaneously. Mese [7] studied cell load-
ing and family scheduling in a cellular manufacturing environment. This study
is distinguished from prior research by giving every job in a family an individual
due date. A focus of this study was the tradeoff between meeting due dates of
jobs and reducing total setup time. Setup times are reduced if all jobs of a family
are scheduled together, but this could cause delays and tardiness in other jobs in
other families. A way to offset this is to allow family splitting, but this increases
setup times and in return may increase the number of tardy jobs. The author
used mathematical modeling and a genetic algorithm to solve this problem. The
mathematical model was slow and impractical to use for larger problems. The
GA developed was both effective in finding optimal or near optimal solutions
and efficient.
Süer [16] developed a three phase methodology to perform cell loading and
scheduling in a shoe manufacturing company. This research used three family
definitions (sub-families, families, and superfamilies) in the cell loading process.
These different family definitions are used at different stages of the planning and
scheduling process. Superfamily definition was used to determine the number of
cells of each type. Family definitions are used after product-to-cell assignments.
These families allow the number of setups to be minimized. Finally, the subfamily
definition is used in the implementation of heuristics before cell loading begins.
This research also used the MaxCap methodology used by Subramanian [12]. A
valuable conclusion from this research is that the best post loading results did
not give the best post scheduling results. This demonstrated the need to consider
that isolating a single level of a multilevel problem may not result in the best
solution for the overall problem. Only the injection molding cell is covered in this
research, leaving further research to be done regarding the lasting and finishing
cells in the shoe manufacturing company.
3 General Methodology
The three-phase methodology proposed to solve the problem is described in the
following sections.
Phase 1 - Manpower Allocation
Thirty-five workers are split between the LC and FC. The six worker splits
evaluated are 15/20, 16/19, 17/18, 18/17, 19/16, 20/15, where 15/20 represents
15 workers available for allocation in the LC and 20 available in the FC. The
workers allocated to the LC and FC are then allocated among the operations
within those cells. A mathematical model is used to optimally allocate manpower
in order to maximize production rates. The model is run separately for each
product at each worker level in both the LC and FC.
Phase 2 - Cell Loading
Products are assigned to cells in order to maximize the machine-level based
similarity among products in the cells and also minimize the number of cells
opened. A genetic algorithm is used to assign products to cells.
Phase 3 - Flowshop Scheduling

After jobs have been assigned to cells, flowshop scheduling is performed in
each cell to minimize one of three performance measures. The performance mea-
sures considered are makespan, total tardiness, and number of tardy jobs. The
methodology is summarized in Table 2.
Table 2. Summary of General Methodology
Hase Method Objective Cells Purpose

1-Worker Mathematical Maximize production LC, FC Assign workers to stations in
allocation model rate LC & FC
2-Cell Genetic Maximize machine level LC, FC Assign products to cells,
loading algorithm similarity between minimize machines required
products within cells
3-Flowshop Mathematical Minimize makespan, LC1-RC1-FC1 Determine sequence of
scheduling model number of tardy LC2-RC2-FC2 products along
products, total with start LCn-RCn-FCn and
tardiness, number of completion time independently
machines
4 Manpower Allocation Phase

In phase 1, workers are allocated to stations in the LC and FC, considering
various worker levels, in order to maximize production rates. The model is solved
for each product at each worker level and in both the LC and FC.
The index, parameters and decision variables are defined below:
Index:
j : Operation
Parameters:
R : Production rate for the product;
tj : Unit operation time for operation j;
uj : Maximum number of operators for operation j;
W : Total number of workers available
Variable:
mj : Number of workers assigned to operation j
Objective Function:
max Z = R (1)

1
s. t. mj − R ≥ 0, j = 1, 2, · · · , s (2)
tj
mj ≤ Uj , j = 1, 2, · · · , s (3)
s
mj ≤ W (4)
j=1
mj integer and positive for all j

R positive.
Equation (1) shows the objective function of the mathematical model, which
is to maximize the production rate. The relationship between number of workers
at a station and the production rate is determined in Eq. (2), ensuring enough
workers are assigned to each operation to meet the desired production rate. Eq.
(3) establishes an upper limit on the number of workers allowed at an operation.
Equation (4) ensures that the total number of workers assigned to the stations
does not exceed the total number of workers in the system. An example will be
used to explain each phase of the problem. An example featuring 10 products
each with five operations in both the LC and FC will be used. Unit processing
times are given in Table 3. The mathematical model described earlier in this
section was run using these processing times. The optimal worker allocations for
product 1 at all worker levels are shown in Table 4. Table 5 shows optimal worker
allocations for all 10 products at the 15/20 worker level.
Table 3. Unit processing times (minutes/unit) for the example problem
Product Lasting cell Finishing cell

Operations Operations
1 2 3 4 5 1 2 3 4 5
1 1.41 1.36 0.76 0.65 0.39 0.44 0.41 1.49 0.54 1.15
2 0.80 0.78 0.85 0.76 0.59 1.23 1.10 0.32 0.63 1.17
3 0.79 1.45 1.13 1.98 0.53 1.07 0.43 0.91 1.21 1.12
4 0.31 0.54 2.79 0.66 0.95 0.77 0.99 0.95 1.47 0.33
5 1.26 0.43 1.39 1.28 1.40 0.58 0.68 1.49 0.28 0.40
6 0.73 0.57 1.04 0.78 0.94 2.18 0.43 0.42 1.72 0.23
7 0.57 0.78 1.23 0.19 0.36 0.60 0.44 1.12 0.81 1.00
8 0.43 1.42 1.07 0.25 0.79 1.64 2.05 0.70 0.31 2.55
9 0.87 0.61 1.04 0.87 0.68 0.48 0.59 0.26 0.15 0.41
10 1.98 0.62 1.25 1.25 0.80 1.39 2.10 0.62 0.63 1.17
Table 4. Worker allocation combinations for product 1
Product Worker level Lasting cell Finishing cell

Operation Prod. rate Operation Prod. rate
(U/min) (U/min)
1 2 3 4 5 1 2 3 4 5
1 15/20 4 4 3 2 2 2.84 2 2 7 3 6 4.55
1 16/19 5 4 3 2 2 2.94 2 2 7 3 5 4.35
1 17/18 5 5 3 2 2 3.08 2 2 6 3 5 4.03
1 18/17 5 5 3 3 2 3.55 2 2 6 2 5 3.7
1 19/16 6 5 3 3 2 3.68 2 2 6 2 4 3.48
1 20/15 6 6 3 3 2 3.95 2 2 5 2 4 3.36
Table 5. Worker allocation for all products at 15/20 worker combination
Product Worker level Lasting cell Finishing cell

Operation Prod. rate Operation Prod. rate
(U/min) (U/min)
1 2 3 4 5 1 2 3 4 5
1 15/20 4 4 3 2 2 2.8369 2 2 7 3 6 4.5455
2 15/20 3 3 3 3 3 3.5294 5 5 2 3 5 4.065
3 15/20 2 3 3 5 2 2.069 4 2 4 5 5 3.7383
4 15/20 1 2 7 2 3 2.509 4 4 4 6 2 4.0404
5 15/20 3 1 4 3 4 2.3256 3 4 8 2 3 5.1724
6 15/20 3 2 4 3 3 3.1915 8 2 2 7 1 3.6697
7 15/20 3 4 5 1 2 4.065 3 2 6 4 5 4.5455
8 15/20 2 5 4 1 3 3.5211 4 6 2 1 7 2.439
9 15/20 3 2 4 3 3 3.2797 5 6 3 2 4 9.7561
10 15/20 5 2 3 3 2 2.069 5 7 2 2 4 3.1746
The mathematical model was solved using ILOG OPL 6.3. The optimal pro-
duction rates are determined in both the LC and FC at the given worker level. For
example, product 2 at the 15/20 worker level for the Lasting Cell requires three
workers at each operation and results in a production rate of 3.52 units/minute.
5 Cell Loading Methodology

Cell loading is performed using a genetic algorithm approach to maximize the
similarity among products while minimizing the number cells opened. Similarity
between jobs is measured using the machine level-based similarity coefficient
between products i and k developed by Süer and Ortega [14]. The coefficient is
calculated using Eq. (5). A sample calculation using the worker allocations in
the lasting cell for products 1 and 2 is shown in Eq. (5).
s
j=1 min(mij , mkj )
MLB − SCik = s , (5)
j=1 max(mij , mkj )
min(4, 3) + min(4, 3) + min(3, 3) + min(2, 3) + min(2, 3)
MLB − SC12 =
max(4, 3) + max(4, 3) + max(3, 3) + max(2, 3) + max(2, 3)
13
= = 0.76.
17
This calculation is performed for each pair of jobs to create the similarity
matrix used in the genetic algorithm approach. A similarity matrix was generated
to evaluate similarity between jobs at each worker level. The matrix generated
for the example problem is shown in Table 6.
Table 6. Product similarity coefficients
Products 1 2 3 4 5 6 7 8 9 10
1 - 0.76 0.67 0.5 0.58 0.67 0.76 0.67 0.67 0.76
2 - - 0.76 0.58 0.76 0.88 0.67 0.67 0.88 0.76
3 - - - 0.5 0.58 0.67 0.58 0.58 0.67 0.67
4 - - - - 0.58 0.67 0.58 0.58 0.67 0.5
5 - - - - - 0.88 0.58 0.58 0.88 0.67
6 - - - - - - 0.67 0.67 1.00 0.76
7 - - - - - - - 0.76 0.67 0.58
8 - - - - - - - - 0.67 0.5
9 - - - - - - - - - 0.76
10 - - - - - - - - - -
The capacity requirements are calculated using the production rates found
in phase 1 and demand. Each product’s demand and capacity requirements are
shown in Table 7. Capacity requirements are shown in minutes and as fraction
of the total minutes available in each cell. Also, capacity requirements are deter-
mined for both the LC and FC and cells are loaded so that neither is over the
allowable utilization, 100% for the example problem. An example solution for
the problem are shown in Table 8. Cell 1 is assigned products 10, 7, 6, and 2.
The utilization in the LC is 0.88 and the utilization for FC is 0.71.
Table 7. Capacity requirements
Product Demand Lasting cell Finishing cell

Production Capacity Capacity Production Capacity Capacity
rate requirement requirement rate requirement requirement
(min) (as fraction of (min) (as fraction of
2400 min) 2400 min)
1 1863 2.84 656.70 0.27 4.55 409.86 0.17
2 2147 3.53 608.32 0.25 4.07 528.17 0.22
3 2291 2.07 1107.30 0.46 3.74 612.85 0.26
4 1328 2.51 529.29 0.22 4.04 328.67 0.14
5 2300 2.33 988.99 0.41 5.17 444.67 0.19
6 1627 3.19 509.79 0.21 3.67 443.36 0.18
7 1389 4.07 341.70 0.14 4.55 305.58 0.13
8 2010 3.52 570.84 0.24 2.44 824.11 0.34
9 1601 3.28 488.15 0.20 9.76 164.10 0.07
10 1409 2.07 681.01 0.28 3.17 443.84 0.18
Table 8. Cell loading results for the example problem
Products LC utilization FC utilization

Family 1 [10,7,6,2] 0.88 0.71
Family 2 [8,3,1] 0.97 0.77
Family 3 [9,5,4] 0.83 0.4
5.1 Flowshop Scheduling Methodology

The objective of the mathematical model for flowshop scheduling is to sequence
the jobs in each part family such that the desired performance measure
(makespan, total tardiness, or number of tardy job) is minimized.
Production rates found in phase 1 are used to determine processing times for
the LC and FC. The processing times for the RMC are determined by the shoes
mold type and size. The indices, parameters and decision variables are described
below.
Index:
u : Cell index;
j : Job index
Parameters:
Pij : Processing time of job j in cell i;
dj : Due date of job j;
r : Small number j;
M : Large number
Decision variable:
yij : Start time of job j in cell i;
cij : Completion time of job j in cell i;
tj : Tardiness of job j;
wj : 1 if job j is tardy, 0 otherwise;
zji : 1 if job j is processed before job i in cell 1, 0 otherwise;
bji : 1 if job j is processed before job i in cell 2, 0 otherwise;
gji : 1 if job j is processed before job i in cell 3, 0 otherwise;
T T : total tardiness variable;
h : makespan variable
Objective functions:

min N1 = wj + r × tj , j = 1, 2, · · · , n (6)
min N2 = T T (7)
min N3 = H. (8)

s. t. tj = T T, j = 1, 2, · · · , n (9)
cij ≤ h, j = 1, 2, · · · , n (10)
yi+1,j − yij ≥ pij , j = 1, 2, · · · , n (11)
cij − yij ≥ pij , j = 1, 2, · · · , n (12)
cij − ≥ dj , j = 1, 2, · · · , n (13)
M × wj ≥ tj , j = 1, 2, · · · , n (14)
yij − yij+1 + M × zij ≥ pij , j = 1, 2, · · · , n (15)
yij+1 − yij − M × zij ≥ −M + pij , j = 1, 2, · · · , n (16)
yij − yij+1 + M × bij ≥ pij , j = 1, 2, · · · , n (17)
yij+1 − yij − M × bij ≥ −M + pij , j = 1, 2, · · · , n (18)

yij − yij+1 + M × gij ≥ pij , j = 1, 2, · · · , n (19)
yij+1 − yij − M × gij ≥ −M + pij , j = 1, 2, · · · , n. (20)
The various objective functions which can be used depending on the desired
performance measure are shown in Eqs. (6), (7), (8), which seek to minimize
number of tardy jobs, total tardiness, and makespan, respectively. Eq. (9) sums
the jobs tardiness to define total tardiness. Eq. (10) establishes the relationship
between the jobs’ completion time and its makespan, ensuring that makespan is
equal to the completion time of the last job. Eq. (11) ensures that a job has to
finish processing in its current cell before it can start in the following cell. In a
similar manner, Eq. (12) ensures that a job must complete processing in the final
cell before it can be labeled complete. The relationship between completion times,
due dates, and tardiness is established in Eq. (13). Equation (14) assigns a value
of either zero, for on time jobs, or one, for tardy jobs. Equations (15) and (16)
perform a pairwise comparison between jobs in cell 1 to determine which job is
processed first. This comparison is done between all jobs and Eq. (17) through
Eq. (20) show the comparison being performed in the second and third cells.
The schedules for the part families from the example with regards to
makespan are shown in Table 9. For example the sequence for part family 1
is 2-7-6-10 and the makespan is 53. A Gantt chart is shown for part family’s
schedule with regards to makespan in Fig. 5.
Table 9. Example schedules to minimize makespan
Schedule Makespan
Part Family 1 2 7 6 10 53
Part Family 2 8 3 1 - 60
Part Family 3 5 4 9 - 50
Fig. 5. Gantt chart for family 1
6 Genetic Algorithms Approaches for Cell Loading

A genetic algorithm for cell loading is proposed in this section. Chromosome rep-
resentation, crossover, and mutation strategies used are discussed. The genetic
algorithm has two crossover strategies and the option to perform mutation before
crossover. This created four general strategies; (1) order-based crossover before
mutation, (2) position-based crossover before mutation, (3) order-based crossover

after mutation, and (4) position-based crossover after mutation. Within each of
the four general strategies, ten combinations of crossover probability and muta-
tion probability were tested. Five trials were performed for each crossover and
mutation probability combination in each of the general strategies.
The chromosome initially is created by randomly ordering the products. The
cells are then loaded sequentially from the randomly ordered list. When the
utilization in either the lasting cell or finishing exceeds a maximum allowed
utilization, that cell is closed and the next cell is opened. This continues until all
products are assigned to a cell. An example of the chromosome representation
is shown in Table 10. The fitness function is calculated using Eq. (21).
n
n n

max Z = Sik xik − (pxkk ). (21)
i=1 k=1 k=1
Table 10. Example chromosome representation
Product sequence
1 2 3 4 5 6 7 8 9 10
Product 2 4 1 5 6 8 3 9 10 7
LC Utilization 0.24 0.3 0.28 0.17 0.22 0.2 0.2 0.15 0.23 0.26
FC Utilization 0.27 0.3 0.17 0.3 0.25 0.22 0.18 0.15 0.23 0.18
Cell 1 1 1 2 2 2 2 3 3 3
Total LC Util 0.24 0.54 0.82 0.17 0.39 0.59 0.79 0.15 0.38 0.64
Total FC Util 0.27 0.57 0.74 0.3 0.55 0.77 0.95 0.15 0.38 0.56
Two crossover strategies are used in the genetic algorithm, order-based

crossover and position-based crossover, and these strategies are discussed in the
ensuing paragraphs. Position-based crossover is performed by selecting a set of
positions at random from one parent chromosome and passing the values at
these positions to the child. The missing genes are then selected and filled in
the child chromosome in the order they appear in the other parent chromosome.
If there are an even number of products, (#Products)/2 is used to determine
the number of random positions selected, (#Products-1)/2 is used for an odd
number of products. Order-based crossover is performed by selecting a string of
genes from one parent and passing the string onto the child chromosome. The
remaining genes are then filled in the order they appear in the other parent
chromosome. Examples of position-based crossover and order-based crossover
are shown in Figs. 6 and 7, respectively. The arrows show how genes are passed
from the parent chromosome to the child chromosome.
The mutation strategy used is reciprocal exchange mutation. This is per-
formed by using random numbers to select two positions in the chromosome and
swapping the values in these positions. An example of this mutation strategy is
shown in Fig. 8.
Parent 1 8 1 7 2 6 3 5 4
Parent 1 4 2 3 6 5 8 7 1
Parent 2 2 5 8 7 6 3 4 1
Parent 2 2 7 6 5 1 3 8 4
Child 1 4 2 3 6 7 5 8 1 Child 1 8 1 7 2 5 6 3 4
Child 2 2 7 4 6 1 3 5 8 Child 2 1 2 8 7 6 3 5 4
Fig. 6. Position-based Crossover Fig. 7. Order-based Crossover
Fig. 8. Mutation Example
After crossover and mutation have been performed, products are reassigned
to cells in the same manner as the original chromosome to ensure the maxi-
mum utilizations are not exceeded. Selection is by selecting the top chromosomes
among both the parent and child chromosomes to move onto the next generation.
This study also investigates the effect of performing mutation before crossover.
Classical genetic algorithms flow is crossover before mutation. By reversing the
traditional flow, it is possible better parent chromosomes will be created and per-
forming crossover with the better parent chromosomes will result in improved
offspring chromosomes. Selection is done in this study by selecting the best
ranked chromosomes from all the parent and offspring chromosomes for the next
generation.
7 Results and Conclusion

Scheduling is done on the part families found using the four GA strategies. The
scheduling results for the worker levels are shown in Tables 11, 12, 13, 14, 15
and 16 below. The bolded and italicized values in the tables below indicate that
those are the best result in the particular performance measure in that worker
level.
At the 15/20 worker level given in Table 11, GA Strategy 2 either outper-
formed or tied with respect to MS, TT and number of machines required whereas
GA Strategy 4 provided the best result in nT. At the 16/19 worker level shown
in Table 12, GA Strategy 4 dominated all other strategies. At the 17/18 worker
level presented in Table 13, no strategy dominated others. However, GA Strategy
3 outperformed others with respect to total tardiness and number of machines

while GA Strategy 4 outperformed in makespan. At worker level 18/17 given
in Table 14, GA Strategy 3 dominated in all scheduling performance measures.
However, this strategy requires a significant increase in number of machines rela-
tive to the other strategies. This is because an additional part family is opened in
this strategy. At worker level 19/16 shown in Table 15, no strategy demonstrated
dominance over the rest. GA Strategy 2 and GA Strategy 4 performed equally
well. At worker level 20/15 presented in Table 16, GA Strategy 3 dominated in
makespan and tied for dominance in nT and number of machines required. Also,
GA Strategy 3 provided the lowest makespan among all worker levels in this
problem. GA strategy 4 dominated in total tardiness.
The results are summarized in Tables 17 and 18. In terms GA strategies,
GA Strategy 3 was the overall best GA Strategy (3 out of 7 times). This was
followed by GA Strategies 2 and 4 (2 out of 7 times). GA Strategy 1 never made
the list. The variation in MS was 58-61 and nT was 8-9 whereas it was much
Table 11. Worker allocation 15/20 scheduling summary
Solution strategy 15/20 results

GA1/MM GA2/MM GA3/MM GA4/MM
MS 60 60 60 60
nT 9 9 10 8
TT 145 137 139 139
# Machines 276 274 274 277

MS 61 61 60 60
nT 9 9 9 8
TT 124 124 123 122
# Machines 285 285 273 272

MS 62 62 62 60
nT 10 9 9 9
TT 122 127 120 126
# Machines 286 277 276 279
Table 14. Worker Allocation 18/17 Scheduling Summary

MS 66 64 61 63
nT 10 10 9 10
TT 157 147 135 159
# Machines 256 247 275 254
Table 15. Worker Allocation 19/16 Scheduling Summary

MS 60 59 60 59
nT 9 8 9 8
TT 111 116 111 116
# Machines 279 279 279 279

MS 60 60 58 61
nT 9 9 8 8
TT 128 128 117 116
# Machines 275 275 270 278
Table 17. Summary of manpower level results
Manpower levels
15/20 16/19 17/18 18/17 19/16 20/15
Overall best GA2 GA4 GA3 GA3 GA2/GA4 GA3
GA strategy
Best MS 60 60 60 61 59 58
Best nT 8 8 9 9 8 8
Best TT 137 122 120 135 111 116
Best # Machines 274 272 276 247 279 278
Table 18. Summary of performance measure results
Performance measure Best so far Manpower levels GA strategies

MS 58 20/15 GA3
nT 8 20/15, 19/16, 16/19 GA3/GA4, GA2/GA4, GA4
TT 111 19/16 GA1/GA3
# Machines 247 18/17 GA2
higher in terms of TT (111–137) and number of machines (247–279) as shown

in Table 17. Finally, the best values of performance measures obtained in this
experimentation are summarized in Table 18. We can safely say that unbalanced
manpower levels in LC and FC produced better results, in other words, worker
levels 20/15, 19/16, and16/19. One can actually say that higher manpower level
in LC seemed to work better. One again, GA Strategy 3 produced better results
along with GA Strategy 4.
References
1. Akturk MS, Wilson GR (1998) A hierarchical model for the cell loading problem
of cellular manufacturing systems. Int J Prod Res 36(7):2005–2023
2. Babayiǧit C (2003) Genetic algorithms and mathematical models in manpower
allocation and cell loading problem [electronic resource]. Master’s thesis, Ohio Uni-
versity
3. Chang OK, Shin HJ (2003) Scheduling jobs on parallel machines: a restricted tabu
search approach. Int J Adv Manuf Technol 22(3):278–287
4. Gupta JND, Chantaravarapan S (2008) Single machine group scheduling with fam-
ily setups to minimize total tardiness. Int J Prod Res 46(6):1707–1722
5. Gupta JND, Ho JC (2001) Minimizing makespan subject to minimum flowtime on
two identical parallel machines. Comput Oper Res 28(7):705–717
6. Huang J, Süer GA, Urs SBR (2012) Genetic algorithm for rotary machine schedul-
ing with dependent processing times. J Intell Manuf 23(5):1931–1948
7. Mese E (2009) Cell loading and family scheduling for jobs in a shoe manufacturing
company. Master’s thesis, Ohio University
8. Nakamura N, Yoshida T, Hitomi K (1978) Group production scheduling for mini-
mum total tardiness part (i). IIE Trans 10(2):157–162
9. Saad SM, Baykasoglu A, Gindy NNZ (2002) A new integrated system for loading
and scheduling in cellular manufacturing. Int J Comput Integr Manuf 15(1):37–49
10. Stnha RK, Hollier RH (1984) A review of production control problems in cellular
manufacture. Int J Prod Res 22(5):773–789
11. Subramanian AK (2004) Cell loading and scheduling in a shoe manufacturing
company. Master’s thesis, Ohio University
12. Süer G, Subramanian A, Huang J (2009) Heuristic procedures and mathematical
models for cell loading and scheduling in a shoe manufacturing company. Comput
Ind Eng 56(2):462–475
13. Süer GA (1996) Optimal operator assignment and cell loading in labor-intensive
manufacturing cells. Comput Ind Eng 31(1–2):155–158
14. Süer GA, Saiz M et al (1995) Manufacturing cell loading rules and algorithms for
connected cells. Manuf Res Technol 24(24):97–127
15. Süer GA, Cosner J, Patten A (2008) Models for cell loading and product sequencing
in labor-intensive cells. Comput Ind Eng 56(1):97–105
16. Süer GA, Ates OK, Mese EM (2014) Cell loading and family scheduling for
jobs with individual due dates to minimise maximum tardiness. Int J Prod Res
52(19):5656–5674
Research on the Competitive Strategy
of Two Sided Platform Enterprises
Based on Hotelling Model
Zeming Wang(B) and Yinpeng Guo
School of Management, Northwest University for Nationalities,

Lanzhou 730030, People’s Republic of China
641881800@qq.com
Abstract. In this paper, to take the group buying platform as an exam-

ple, we use Helloing model to analyze the influence the pricing and profit
of the platform enterprises under the different homing behaviors of users.
The analysis platform of enterprise competitive strategy from the per-
spective of pricing and profit, we show that compared with network exter-
nalities, the influence of differences is bigger about platform enterprises;
changed the competitive strategy to improve customer loyalty can bring
more profits; mutual cooperation between the platform of enterprise will
increase the enterprise platform competitiveness. We hope to provide ref-
erences for the competitive strategic choices of the platform enterprises.
Keywords: Group purchase platform enterprise · Hotelling model ·

Bilateral platform enterprise · Competitive strategy
1 Introduction
In recent years, bilateral market is the industrial organization and the theory of
competitive strategy research. Its unique business platform competition strate-
gies have attracted the attention of scholars. The earliest group buying websites
appeared in United States (U.S. buy network Groupon). It is mainly the use
of consumers to get a larger discount psychology, and take shape an effective
network interaction. Then, group buying to earn service fees from businesses.
Influenced by this operating mode Chinese group purchase network develop-
ment rapid after 2010. From the group purchase network to the mobile terminal
APP application, Meituan, Public comment network, Juhuasuan, Baidu Nuomi,
Wowo Group and other large professional group buying provide abundant prod-
ucts that involved in almost every aspect of people’s life such as food, clothing,
cosmetics beauty, fitness, hotel, tourism and so on. Then consumers are attracted
by lower prices and larger discounts. With the increase of the number of con-
sumers, more and more merchants registered in platform of enterprise. Volume
of business brings in profit for the platform enterprise. According to the Chinese
group purchase market statistics in June 2015 shows that by the end of June
578 Z. Wang and Y. Guo
2015, Chinese network group purchase transaction volume reached 16.74 billion
yuan, compared to the same period the growth of 182.5%, the number of peo-
ple who participate in the group reached 250 million person-times, compared to
the same period the growth of 161.8%. As a typical enterprise platform, group
purchase platform competition is increasingly fierce. Everything is not going
smoothly. Due to the diversity of businesses and consumers, some businesses
and consumers will choose a platform for business, others will be registered in
enterprises with different network platform. The choice behavior of different busi-
nesses and consumers will influence the pricing direction and profit of enterprise
platform. Research the choice behavior of consumers and businesses to develop
a reasonable pricing mechanism and competitive strategy for different types of
choice is of great significance for the development of the group buying platform.
2 Related Works
Entering twenty-first Century, the bilateral market theory has attracted wide
attention with the rapid development of the information economy. As the leading
theory of industrial organization, its importance has been increasing. The earliest
examples of two-sided markets can be traced back to the “New York sun” led the
so-called “penny movement”, all the journalists no longer need to pay the cost
to buy a newspaper, the newspaper operator is the main profit from advertising
revenue [1]. Rochet and Tirole [7] thought bilateral market is described that have
two independent edge, and its ultimate benefit from the platform transaction.
They established a bilateral market platform model that has become research
basis of bilateral market theory. Cheng and Sun [6] put forward that bilateral
market is a dumbbell type structure, the participating parties B expressed as
consumers, and the participating parties S as businesses, as shown in Fig. 1.
Fig. 1. Two sided market simplified schematic
Chen [4] proposed the network group purchase system, the network group
purchase has carried on the more comprehensive description (Fig. 2). As a bilat-
eral platform for enterprises, group buy site is also a two different types of
participants through the platform to trade. It is the specific bilateral market
dumbbell type structure too.
Through the study a lot of literature, most of scholars think that the bilateral
market has the following characteristics. Firstly, the platform has two different
types of user (businesses and consumers) at least in the market, and provides
Competitive Strategy of Two Sided Platform Enterprises 579
Fig. 2. Network group purchase system
different products or services for different types of user. Or make a deal for
different types of user through the platform. Secondly, various types of user
interaction platform, they need each other. Platforms help them to complete the
transaction and improve the efficiency of transaction. Thirdly, the existence of
network externalities between different types of client platform, which increased
the number of user on the platform in one side (for example increasing the
number of consumers) will increase the value of platform for other types of
user [5].
In recent years, with the advent of the era of big data, there has been a new
platform of bilateral market enterprise research. Bardey, Helmuth and Lozach-
meur [2] studied competition in two-sided markets with a common network exter-
nality rather those than with the standard inter-group effects. Behringer and
Filistrucchi [3] showed that a two-sided monopolist may find it short-run profit-
maximizing to charge a price below marginal cost on one side of the market.
Hence showing that the price is below marginal cost on one side of a two-sided
market cannot be considered a sign of predation. They then argue for a two-sided
Areeda-Turner rule that takes into account price-cost margins on both sides of
the market. Roger [8] studies duopoly in which two-sided platforms compete in
differentiated products in a two-sided market. Zhu [10] talked about the plat-
forms that is from the one-sided market to the one-sided + two-sided market,
as an example of JD.com. Scholars from different angles discussed the bilateral
market competition relationship, research more and more comprehensive. This
paper will discuss the competitive relationship of bilateral market based on the
user ownership relationship.
3 Put Forward Problems

In the bilateral market, users (merchants and consumers) will consider their
own interests. Some consumers and businesses only register a group purchase
platform, or choose only one platform to make a deal. This user’s behavior
belongs to the single homing. Other consumers and businesses are registered or
select more than one platform to consumption. This user behavior belongs to
multi homing. Users behavior present in more than one group purchase platform,
because its ownership behavior is not unified. Poolsombat and Vernasca thought
in the bilateral market this behavior called partial multi-homing that one part
of user belongs to the single homing and other part of user belongs to multi
homing. In reality, the users both ends of platform companies such as most of
group purchase platform belong to partial multi-homing. Taking into account
the factors of customer loyalty, homing behaviors of user are divided into three
types: the type of one is the consumers and businesses all belong to single homing
behavior; the type of two is the consumers or businesses belonging to partial
multi-homing, and other businesses and consumers belong to single homing;
types of three consumers and businesses are belonging to partial multi-homing.
According to the characteristics of the bilateral market and external factors that
influence the bilateral market, this paper studies the following problems.
(1) Owing to the difference between two sides of group platform and impact
on network externalities, leading to businesses and consumers have differ-
ent homing behaviors. The type of users’ homing behavior in the difference
between the group buying platforms can have an effect on great difference.
Whether it can reduce the impact of network externalities, obtain a larger
profit for group buying platforms.
(2) Businesses and consumers can freely choose the ownership of the group pur-
chase platform. Belonging to different behavior of businesses and consumers
will affect the group purchase platform enterprise pricing. Because of con-
sumers and businesses of different homing types, group purchase platform
pricing may be different.
(3) In the two-sided market can choose freely under the condition of different
attribution behavior of businesses and consumers. Corporate profits will
affect by consumers and businesses of different types of ownership. As a
result, group purchase platform gains may be different.
(4) In order to develop itself, group purchase platform may often take different
strategy. For example they use market segmentation to attract different con-
sumers and businesses. By taking the difference strategy, ownership behav-
ior of consumer businesses may be changed. Ownership direction guiding of
users may bring more profits for the platform of enterprise development.
4 The Model of Competition

Hotelling model is an important analytical model in the theory of industrial
organization. Amstrong [9] is an earlier study on the pricing strategy of platform
competition in two-sided markets by using the Hotelling analysis framework. In
this paper, we also learn from this analysis framework to study competition
strategy of the group buying platform.
First of all, we define symbols for the convenience of research and then discuss
classification. In this model, service cost of the group buying platform to provide
users is ignored.
Notation:
i, j: Two platform enterprises with a competitive relationship;
k, m: Businesses and consumers are located on both sides of the group buying,
and obey the uniform distribution on the line k, m = 1, 2;
vki , vkj : The user register exchange on the group-buying platform i and platform j
pik : Group-buying platform pricing for the users;
t: User to buy platform unit transportation cost (its economic meaning of
the difference between the two platforms);
β: Network externalities between the users;
v0 : Users obtain basic utility from platform i or platform j based on the net
utility of the platform, if all the users will be registered on the platform
to have transaction at least, the v0 tends to infinity;
nik , njk : The number of single homing user group K on the group buying platform
i and platform j;
Nki : The number of single homing and the number of multi homing on the
group buying platform i of users user group K;
π: The profit of group buying, equal to the number of users of the price
multiplied by the price;
ss: Both of two sides are single homing
sm: One side is single homing, another side is partial multi-homing;
mm: Both of two sides are partial multi-homing
4.1 Both of Two Sides are Single Homing

In this model, two group purchase platform enterprise that have competitive rela-
tion are supposed, and located at the ends of the line segment [0,1]. Businesses
and consumers are users of two sides of group purchase platform enterprise.
Through the comparison net income of group buying platform 1 and group buy-
ing platform 2, users choose a high net income platform registered and traded.
As shown in Fig. 3, making Hotelling model, a user side K to group purchase
platform 1 distance is x, and to the group purchase platform 2 distance was
1 − x. Obviously, the group purchase platform profit equals based utility minus
the paid price, minus the cost of transportation, and add users to the group
purchase website external effect.
Thus we get vk1 and vk2 :
vk1 = v0 − p1k − tx + βn1m , vk2 = v0 − p2k − t1 − x + βn2m .
A and B is homing zero point on the both of two sides of two group purchase
platforms, and in this two points the users obtain same profit. The number of
users in the platform n1k = x, n1m + n2m = 1, so:
n1k = 1/2 − 1/2t(p1k − p2k ) + β/2t(2n1m − 1).

Fig. 3. The user of both of two sides are single homing
Then get the solution:

t(p11 −p21 ) β(p12 −p22 )
n11 = 1
2 − 2(t2 −β 2 ) − 2(t2 −β 2 )
t(p11 −p21 ) β(p12 −p22 )
n12 = 1
2 − 2(t2 −β 2 ) − 2(t2 −β 2 ) .
Because both of two platforms are single homing n11 + n12 = 1, n21 + n22 = 1,
get n21 and n22 .
The profit function of the group buying platform is:

π1 = p11 n11 + p12 n12
π2 = p21 n21 + p22 n22 .
Based on symmetry p11 = p21 = p1 , p12 = p22 = p2 , p1 , p2 are got by derivation,

as follows, group purchase platform pricing is: pss ee
1 = p2 = t − β.
Both of sides of the market accounted for 1/2. In the condition of t > β,
profits of the platform 1 and platform 2 have maximum value. So group purchase
platform profit is: π ss = t − β
4.2 One Side is Single Homing, Another Side

is Partial Multi-homing
Set one side to the side of the single side (Fig. 3 has been described in Fig. 1),
part of the multiple homing of 2, the situation as shown in Fig. 4:
Fig. 4. On both sides of the platform users in the two competing platforms belonging
to the range
In the two sides the total number of users is 1. So Nk1 + n2k = 1, Nk2 + n1k = 1.
User of single homing on the platform 1 net utility is: vk1 = v0 − p1k − tx + βN21 ,
K side users of single homing on the platform 2 profit as follows:
vk1 = v0 − p1k − tx + βN21 .

β
The number of users on the platform 1 is: n11 = 12 − 2t
1
(p11 − p21 ) + 2t (N21 − N22 )
Users on the platform 1 of the net utility of n12 , which is 2 of the user’s
distance from the platform 1 of the distance x: x = n12 = t−β t + (p2 )/t + β/tn1 .
2 1
Users to the platform 2 of the net utility of 1 − n22 , which is 2 of the user’s
distance from the platform 1 of the distance y: y = 1−n22 = β/t−(p12 )/t−β/tn21 .
Simultaneous solution:
⎧
t(p11 −p21 ) β(p11 −p21 )
⎪
⎪ n1
= 1
− 2 −β 2 ) − 2(t2 −β 2 )
⎨ 1 2 2(t
β β(p11 −p21 ) β2 2t2 −β 2
n12 = 1 − − − 2t(t2 −β 2 ) p2
1
− 2t(t2 −β 2 ) p2
2
⎪
⎪
2t 2(t2 −β 2 )
⎩ n2 = 1 − β
−
β(p11 −p21 )
− β 2
β 2
2t(t2 −β 2 ) p2 − 2t(t2 −β 2 ) p2 .
1 2
2 2t 2(t2 −β 2 )
In the condition of t > β, the net utility of the group buying platform 1
and group buying platform 2 can get the maximum value, by p11 = p21 = p1 ,
p12 = p22 = p2 , group buying platform pricings are: psm sm
1 = (t − β )/t, p2 = 0.
2 2
Under the “SM” homing conditions, the numbers of users are:

1 β
n1sm
1 = n2sm
1 = , n1sm
2 = n2sm
2 =1− .
2 2t
According to the formula, the profit for the group purchase platform: π sm =
t2 −β 2
2t .
4.3 Both of Two Sides are Partial Multi-homing

From the first two sections, we can know the “mm” belonging to the number of
users under the conditions of a single:

p2
n1k = 1 + tk − βt n2m
p1
1 − n2k = − tk + βt n1m .
Solution to get the user’s number of single user on two platforms:
⎧ 1 t β t
⎪
⎪ n1 = t+β − t2 −β 2 p2 + t2 −β 2 p1
1 2
⎪
⎨ n1 = t − β p1 + t p2
2 t+β t2 −β 2 1 t2 −β 2 2
t β β
⎪
⎪ n1 = t+β − t2 −β 2 p1 + t2 −β
2 1
2 p2
2
⎪
⎩ 2 t β β
n2 = t+β − t2 −β 2 p2 + t2 −β 2 p21 .
1
Under the circumstances, p11 = p21 = p1 , p12 = p22 = p2 we can get the price of
the group buying platform:
t−β
pmm = pmm =β .
1 2
2t − β
The number of single belonging to the two group purchase platform is:
2t2 − β 2
n1mm = n2mm = n1mm = n2mm = .
1 1 2 2
(t + β)(2t − β)
Similarly, in the case of t > β, the existence of the maximum profit is:
2tβ 2 (t − β)
π mm = .
(t + β)(2t − β)2
5 Analysis of the Model

5.1 The Condition of Maximum Profit
In the condition of t > β, above of three types of homing are existence of the
maximum profit. That is, the difference between the group buying platforms is
greater than the network externalities for the maximum value of the platform.
These pricings and profits show in Table 1.
Table 1. Three types of users buy site pricing and profits
Type of homing Pricing Profits

ss pss ss
1 = p2 = t − β π ss = t − β
2 2
t −β t2 −β 2
sm psm
1 = t
, ps2 m = 0 π sm = 2t
β(t−β) 2tβ 2 (t−β)
mm p1 = pmm
mm
2 = 2t−β
π mm = (t+β)(2t−β) 2
The difference of platform enterprises is helpful to eliminate the influence of

the uncertainty of network externalities on the profit of the group purchase plat-
form. In other words, if the differences of between the two group buying platforms
are enough small, the impact of network externalities may be increased. Then,
businesses and consumers are more likely to increase multi-homing behaviors in
the platforms, and the uncertainty of platform profit margin is also increased. If
there is a big difference between the two platforms of the competition behavior,
it can avoid the influence of network externalities on its profitability.
5.2 Different Homing Different Pricing

Compared to the three types of homing structure under the platform pricing:
t−β β(t−β)
pss sm ss sm ss
1 − p1 = −β t , due to t > β, apparently, p1 < p1 . p1 − p1 = −
sm
t <
mm ss sm
0, get p1 < p1 < p1 .
pmm
1 < pss
1 < p1
sm
shows that the bilateral market platform can also by
attract consumers by increasing consumer quantity and obtaining consumer loy-
alty influence of the numbers of registers of businesses at group buying platform.
As a result, platform through charge higher fees to businesses to obtain profits.
This conclusion is consistent with the general pricing rules of the group buying
platform and the basic characteristics of the bilateral market. There is also differ-
ence pricing strategies for the platform enterprises. Adopted differential pricing
method is also one of the strategies of competition between enterprises.
5.3 Different Homing Different Profit

Comparison of three types of group purchase websites belonging to profit π ss −
2
π sm = (t−β) s
2t , obviously π s > π
sm
.
[(t + β)(2t − β) + 2tβ][(t − β)(2t + β)]
π sm − π mm = (t − β) > 0, by π ss > π sm > π mm .
2t(t + β)(2t − β)2
Group buying platform can get the highest profit in this condition that users
on both of sides belong to single homing. The users on one side belong to single
homing, and on another side belong to partial multi-homing take the second
place. The behavior of users on both of two sides belong to partial multi-homing
is reduced the platform profits. On the one hand, users on the one side belong
to single homing and on the another side belong to partial multi-homing can
put up the platform enterprise pricing, on the other hand, both sides of the user
are single homing can obtain most profitable. This conclusion seems contrary
to the conclusion of the 2. In fact, through use of various initiatives to prevent
users’ multi-homing behaviors, the group purchase platforms enhance loyalty of
user for the platform. It is the internal incentive factors for the development of
platform enterprises.
5.4 Influence of Difference

Considering the effect of the t for a variety of homing structure and profit pricing
∂psm 2 ∂pmm ∂π sm β2 ∂π ,mm
platform ∂t1 = 1 + βt2 > 0 ∂t 1 β
= (2t−β) 2 > 0 ∂t = 12 + 2t2 > 0, ∂t =
2β 2 [t2 (3β−2t)+β 3 ] ,mm
(t+β)2 (2t−β)3 , when t > β, ∂π∂t number of well over 0, simple that when
,mm
2 ∂π ,mm
β ≥ 23 t, ∂π∂t > 0, when β≤ 3 t,
well t sign of uncertainty.
∂t
From the above of that, users on both of sides belong to single homing and
users on one side belong to single homing and another side belongs to partial
multi-homing are show that the difference of group buying platform is influence
greater than network externalities. In these two cases, pricing and profit of plat-
form will be higher. When users on both of sides belong to partial multi-homing
the platform profit is uncertain. This conclusion shows that platform may not be
able to get higher profits. If they only consider the unilateral profits and higher
prices in the process of competition, they need a reasonable price for platform
competition.
6 Conclusion
The users’ homing behavior is a complicated economic behavior, which has an
important influence on the development of competitive strategy. Based on the
group buying platform as an example, from the complexity of social competition

in the market, we sum up the characteristic of group purchase platform and
discuss the pricing and profit enterprise platform strategy in the different user
behavior attribution by making use of the mathematical model. According to
the above research conclusions, this paper puts forward some conclusions:
(1) The platform should be based on the needs of users to develop differentiated
strategies, personalized competitive strategies to meet the needs of the type
of customer needs, and to avoid the impact of network externalities that
bring the risk to the platform.
(2) On the online group buying platform Multi homing behavior of users will
reduce the fees charged to merchants. Forming malignant price competition
and rapid expansion also appeared a lot of homogeneous of group buying
platforms such as lashou.com. At last it lost competitiveness. In the course
of the operation of the platform, we should pay attention to both businesses
and consumers on both sides of the users and try to improve the customers’
loyalty for the development of the platform and enterprise profitability.
(3) Users only choose a group buying platform can enhance the competitive
advantage of group buying platform. Behaviors of consumers and businesses
can get a distinct competitive advantage. Such as the group purchase web-
site, the more good reputations the more consumer users, which can attract
more business enterprises to cooperate with the site, forming a virtuous
circle and crowding out other competitors.
(4) Effect of multi homing behavior interoperability platform users on the plat-
form can weaken the profit. Cooperation with a number of group purchase
website payment platform is not only good for consumer convenience but
also attract more business cooperation. Although the consumer is multi hom-
ing behavior, in the view of platform, it is a single ownership for the group
purchase website. As a whole, the interconnection platform can weaken own-
ership of consumer behavior, improve the competitiveness of the platform
enterprise.
In this paper, we use mathematical model to calculate and compare the

competitive price and the profit of the net purchase platform under different
ownership types. Because group buying platform is a typical bilateral platform,
the study of bilateral platform has a representative significance, to reflect the
actual situation of the operation process of the platform. But the deficiency
lies in the failure of the general suitability of the model to draw the relevant
conclusions. In the future research, we will be in-depth mining related theory, as
early as possible to improve the model of universal verification.
References
1. Amery E, Amery M, South LR (2009) American Journalism History: The History
of Mass Media Interpretation, 9th edn. Renmin University of China press, Beijing
2. Bardey D, Cremer H, Lozachmeur JM (2014) Competition in two-sided markets

with common network externalities. Rev Ind Organ 44(4):327–345
3. Behringer S, Filistrucchi L (2015) Areeda-turner in two-sided markets. Rev Ind
Organ 46(3):287–306
4. Chen S (2011) Group buying trends in China. http://wenku.baidu.com/view/
81d2ad7ca26925c52cc5bf6a.html
5. Chen Y (2014) Research on the business model of platform enterprises in the
bilateral market. PhD thesis, Wuhan University, Wuhan
6. Cheng G, Sun W (2006) The operation mechanism of industry and industry regu-
lation based on bilateral market perspective. Int Financ Res Bank Card 2(1):39–46
7. Rochet JC, Tirole J (1999) Cooperation among competitors: some economics of
payment card associations. CEPR Discuss Pap 33(4):549–570
8. Roger G (2016) Two-sided competition with vertical differentiation. J Econ 120:
1–25
9. Shang X, Chen H (2009) Competitive strategy and regulations of platform firms
from the perspective of bilateral market. Ind Econ Res 10(6):2461–2472
10. Zhu X, Song B et al (2016) Platforms-from the one-sided market to the one-sided
+ two-sided market. Bus Trends Digit Era 1:43–61
Research on Tourists Experience
in Traditional Culture Festival Activities
Based on the Importance-Performance
Analysis Research Method
Lu Li1,2(B)
1
Tourism School of Sichuan University, Chengdu 610065, People’s Republic of China
342920007@qq.com
2
School of Economics and Management of Chengdu Technological University,
Abstract. Chinese excellent traditional culture is an important source

of the party in governing the country. Xi Jinping, emphasized the histori-
cal influence and significance of Chinese traditional culture several times,
and gave it new connotation of the time. Represented by Chengdu Tem-
ple Fair in recent years, the traditional culture festival activities were
not only for the protection and inheritance of traditional culture, but
also became the organic composition part of the local tourism festival
activities. In this paper, we stand in the perspective of tourist experi-
ence and take Chengdu Temple Fair as the research object. It carried
on the questionnaire design and survey to understand of tourists focus
and satisfaction. The IPA methods help us to identify the strengths and
weaknesses and give improvement suggestions.
Keywords: Tourists experience · Importance-performance analysis

(IPA) · Cultural festival activities
1 Introduction
The former researches on traditional culture festival activities mostly focus on

culture itself, such as cultural value, cultural inheritance, cultural exchange, and
so on. However, in the tourism activities, tourist experience has been considered
as one of the most important embodiment of product quality. At the same time,
the tourism experience measurement is also the emphasis and difficulty of such
research. Therefore, the attention to visitors experience is the crucial factor to
make the sustainable development of traditional culture festival activities.
From the perspective of management and marketing, tourist experience is
the tourist satisfaction. From the perspective of managers, the tourist expe-
rience is formed by a travel destination. At present, tourist satisfaction mea-
surement model in the industry are mainly applied with American consumer
Research on Tourists Experience in Traditional Culture Festival Activities 589
satisfaction index (ACSI), service quality (SERVQUAL), as well as Importance

- performance Analysis method (IPA) [3]. IPA analysis method has been applied
to enterprise management and marketing, service quality, competitive analysis,
and other fields.
This paper is to study the situation of the festival activities in the tourist
experience through the IPA analysis method. It analyses the degree of satisfac-
tion and importance on quadrant distribution. The suggestions for such activities
in the future development and quality improvement are also proposed.
Firstly, this paper carries on the literature review of tourism experience and
tourist satisfaction, cultural festivals and tourism experience, satisfaction analy-
sis theory to the literature review. It demonstrates the necessity of carry out
the research in this field and the feasibility of using the IPA analysis. Then,
it focuses in the perspective of tourist experience and takes Chengdu Temple
Fair as the research object. It carried on the questionnaire design and survey to
understand of tourists focus and satisfaction. Finally, it uses the IPA methods
to identify the strengths and weaknesses of traditional culture festival activities
and give improvement suggestions.
2 Literature Review
2.1 Tourist Experience and Satisfaction
Research on tourist experience is an important and difficult problem. Tourist
experiences have both individual difference and the common performance. The
differences mean that the scope, content, depth of each tourists experience are
not the same. But if we take the point of tourist groups, their experiences have
a certain commonality and show similar regularity.
In theory, the ultimate goal of travel is in order to obtain high quality tourism
experience. Therefore, tourism destination and their managers try to achieve the
aim through offering high quality tourist experience [2].
From the perspective of management and marketing, tourist experience is the
tourist satisfaction. From the perspective of managers, the tourist experience is
formed by a travel destination. Thus tourist is the main source of information for
destination management. Tourist evaluations of tourism destination constitute
the important feedback. Managers take the corresponding management improve-
ment and marketing strategies according to the feedback information [16].
This shows the current research situation for all visitors’ experience.
Although the ontology of tourists experience has not reached consensus, most
scholars formed the understanding of the different Angle of view. They showed
their own understandings of tourist experience from different aspects. They also
correspondingly gives the measure method and management tools of tourists
experience [1,11,15].
2.2 Culture Festival Activities and Tourists Experience

External studies on festival tourism started earlier. Meyer [14] classified tourism
festivals according to the festival theme, and made the comparative study of
590 L. Li
festival tourism between the United States and Canada. Gatz [7], Hoyle [10]
studied the problem of festival tourism marketing. Mayfield and Crompton [13]
discussed the marketing concept of festival organizers.
It shows that the researches mainly focus on the influence of the large
tourism festival, particularly in terms of festival tourism and city relationships
[6]. Chinese scholars that studied tourism festival activities focus on the following
aspects, including the impact on the destination, the festival tourism operation
models, the development present situation and countermeasures, perception of
festival tourism for destination residents and tourists, etc.
2.3 IPA Analysis Method
Importance-performance Analysis method, namely IPA method, is widely used

for its intuitive, easy to operate and understand. The method is studied out by
Martilla and James in 1977 [12].
Since the early 1990s, the IPA analysis began to be widely used in service
industry. Its application scope mainly includes service satisfaction, product per-
formance, regional competition (appeal) and macro tourism policy, etc.
On the basis of understanding the multidimensional and multifaceted prop-
erties to the satisfaction, IPA analysis method got rapid development in the
field of marketing research. It can transmit evaluation message to management
through customer evaluation of service product multiple attribute. IPA method
asked respondents to evaluate various evaluation factors in all kinds of measures
from two aspects of importance and performance.
Fig. 1. IPA matrix
When it is used for satisfaction measurement, performance means the satis-

faction evaluation [3]. On the concrete application, first of all, customer under-
standing the importance of product attributes is to be investigated. Then, we
shall understand customer perception of product attribute performance. The

overall average of the two will be showed on the two-dimensional coordinate axis
[8] (Fig. 1). By the graphic method, managers can accurately grasp the manage-
ment status, so as to determine the different management area easily.
For the satisfaction evaluation of tourist experience, attribute of tourist prod-
ucts is transmitted into a multilevel index system. The perception of tourism
product shows the importance and performance. The status from the survey is
put in the I-P model diagram, which can help managers to find the problem
and the corresponding management methods to solve the problem. However,
the traditional I-P model only presented the direction of management, failure
to put forward a measure of the management to improve. Hollenhorst proposed
to improvement model-indicator performance estimate (IPE) [9]. IPE model not
only to helps managers to understand the importance of each index, but also
realize that deviation degree of each index and make its corresponding manage-
ment action.
3 The Empirical Study

Chengdu Temple Fair is selected as a case for empirical research. Chengdu is
the capital of Sichuan province, of Southwest China. It is also one of the most
important economic centers, transportation and communication hubs in Western
China. Chengdu’s earliest temple fair is held before the Qin dynasty period
more than two thousand years ago. Since 2005, the Chengdu temple fair festival
began to hold in Wuhou Shrine till today. The festival takes public benefit as the
core and reveals the local cultural characteristics. Chengdu temple fair festival
highlights three kingdoms culture theme, and constantly absorbed all kinds of
fresh elements. It gradually formed the culture brand and has a wide influence
nowadays. The temple fair was held for years continuously. It has received visitors
over tens of millions of people in 11 years.
“The Spring Festival to the Temple fair”, it has become a New Year cus-
toms of Chengdu area people. Through the successfully held for many years,
Chengdu Temple Fair improved year by year. It becomes not only a regional
cultural brand, but also one of the most famous Chinese traditional Spring Fes-
tival temple fair. Thus, Chengdu temple fair is an appropriate and meaningful
case for empirical research. It is both culture and festival tourism activities rep-
resentative.
3.1 Research Design
The purpose of this study was to apply IPA analysis on festival tourists expe-
rience research. Chengdu temple fair festival is selected as empirical research
object. The most prominent five factors were extracted in the cultural festivals.
Visitors of the target group random questionnaires were distributed during the
2016 Chengdu temple fair. SPASS19.0 software and IPA analysis method was
use to analyze tourist experience and satisfaction.
592 L. Li
This study used the traditional IPA method to analyze the important factors
of tourists’ satisfaction at first. Then the importance and satisfaction evaluation
correlation detection was carried forward. Thirdly, it made the IPA raw data
transformation in the way proposed by Deng [4,5]. Extended importance was
calculated by satisfaction evaluation. Each factor was analyzed through trans-
form data. Finally, the two analysis results were compared to get corresponding
conclusions and suggestions.
3.2 Survey Design

The questionnaire is made up of three parts. The first part is the survey on tourist
satisfaction of cultural festivals elements. On the basis of literature review, the
author consulted the expert group of tourism research. According to the specific
condition of Chengdu Temple Fair, the tourist experience in cultural festival
activities is gained in five elements. They are “The Three Kingdoms culture
theme features”, “traditional folk performances and activities”, “facilities and
services”, “fashion elements and innovation”, and “ticket price setting”. Using
Likert Scale, we ask the respondents to comment on the five aspects of cognitive
performance. In addition, the overall satisfaction perceived evaluation of the
elements is added into the survey, so that subsequent calculation and analysis
on the importance of extension.
The descriptions of the completely positive statements are made for each ele-
ment. Respondents are asked to give evaluation ranges from fully agree (5 points)
to completely disagree (1 point). So that visitors’ satisfaction ranges from fully
satisfied (5 points) to completely dissatisfied (1 point). The second part is the
survey on the importance of the evaluation factors in cultural festivals activity.
In this part, the visitors are asked to the rank the importance by their demands
for each element. This design is to reduce the mutual interference of satisfac-
tion and importance evaluation. The third part is the survey of demographic
characteristics.
3.3 Data Collection and Analysis

From February to March in 2016, this research made questionnaire investigation
at the scene of the Chengdu Temple Fair. A total of five interviewers investi-
gated the visitors in the rest area by random survey. This survey parted out
220 questionnaires, 204 valid questionnaires, and the effective rate of recycling
questionnaire was 93%. During the survey, the interviewers required each visitor
to score on evaluation factors and overall experience satisfaction. Then, accord-
ing to the personal value, visitors were asked to make importance ranking for
evaluation factors.
After the completion of the questionnaire, the results of the survey were input
to SPSS 19.0 database. Data calculation and analysis were done with the help
of software. First of all, we made descriptive statistical analysis of the general
situation of the respondents. Then, the five key elements of the evaluation of
satisfaction and importance of average were calculated. IPA analysis diagram
was drawn. The specific steps of research are listed below. The first step, the
five elements of satisfaction mean value was determined. Then, according to the
importance of evaluation data, we calculated the percentage of the importance,
the average value. We made a horizontal axis and vertical axis, the horizontal axis
represents the degree of satisfaction axis, the dashed vertical axis represents the
importance. The second step, we confirmed the corresponding position location
of 5 evaluation factors in four quadrants according to their importance and
satisfaction of the actual average. The third step, according to the characteristics
of the various factors in different quadrant, we put forward the corresponding
countermeasures and suggestions.
In addition, in order to eliminate the correlation between importance and sat-
isfaction factors, many scholars suggested using extended importance to replace
self-statement importance. Two kinds of caliber is also analyzed in this paper.
Scored for indirect importance, natural logarithmic is computed by partial cor-
relation coefficient of overall satisfaction and satisfaction with all the elements
in the way of Wei-jaw Deng, namely extended importance score.
Table 1. Demographic characteristic of tourists
Item Number Percent

Gender Male 104 51.2
Female 100 48.8
Age ≤17 6 2.9
18–30 153 75
31–45 24 11.8
46–60 11 5.4
≥61 10 4.9
Education Primary school and below 2 2
Junior high school 14 6.9
High school 43 21.2
Undergraduate 138 67.5
Graduate and above 5 2.5
Monthly income (RMB) ≤3000 21 10.3
3000–5000 63 30.9
5000–8000 64 31.4
8000–10000 24 11.8
≥10000 29 14.2
Traveler generating region Chengdu city 83 40.7
Non - Chengdu city 121 59.3
594 L. Li
4 Research Results
4.1 Tourists’ Demographic and Tourism Characteristics
Demographic characteristics showed the 204 visitors polled, basically reached
the same level of the male to female ratio, which accounted for 51.2% of men,
accounted for 48.8% of women. Levels of the age focused on range 18 to 30 years
old, accounted for 75%. Level of education was mostly for college/university
level, 67.5%. Family income more focused on 3000–8000 interval, accounted for
62.3%. Tourist’s origin was in a slightly better non Chengdu city, at 59.3%.
Please refer to Table 1 for relevant information.
This study also surveyed the tourists travelling characteristics, including
travel patterns, the purpose of the visit, travel expenses, etc. In view of tra-
ditional culture festival activities, we also added the play frequency as a survey
element. From the statistical results, it can be seen that the first time visitors
accounted for 59.6%, but there are still close to half for many times. Chengdu
Temple Fair revisit rate is higher, to a certain extent, which also reflects the
regional influence of the cultural festival activities. Travel patterns were given
priority to with family and friends travel together, accounted for 42.2% and
42.2% respectively. For visit purpose, relax accounted for up to 66.7%, the second
is the scenery which is 46.6%. Travel expense (including tickets, transportation,
catering and shopping, etc.) average of about 300 Yuan. Relevant information
as shown in Table 2.
Table 2. Travel characteristics of tourists
Item Number Percent

Times 1 121 59.6
2 39 19.2
3 24 11.8
>3 20 9.4
Travel patterns Alone 18 8.8
With family 86 42.2
With colleagues or friends 63 30.9
With lovers 36 17.7
Purpose Relax 136 66.7
Family reunion 30 14.7
Sightseeing 95 46.6
Photography 9 4.4
Parents-child campaign 13 6.4
Friends meeting 23 11.3
Others 17 8.8
Expense Average 332.43 (RMB)
4.2 IPA Analysis of the Importance and Satisfaction
(1) On the basis of tourists’ self-statement importance

The results showed that Chengdu temple fair investigated into five elements.
Tourists’ self-statement importance showed traditional folk performances and
activities, Three kingdoms culture theme, fashion elements and innovation, sup-
porting facilities and services, and ticket price setting in turn (See details in
Table 3).
Fig. 2. IPA analysis diagram of self-statement importance and satisfaction
Table 3. Ranking of Tourists’ self-statement importance and satisfaction
Ranking of Elements (percent) Importance (5 credits) Satisfaction Ranking of

tourists’ satisfaction
self-statement
importance
1 The Three 0.139 4.1 1
Kingdoms culture
theme features
2 Traditional folk 0.137 4.08 2
performances and
activities
3 Fashion elements 0.115 3.68 5
and innovation
4 Ticket price setting 0.09 3.83 4
5 Supporting facilities 0.087 3.96 3
and services
Thus, we can draw IPA analysis diagram of self-statement importance and

satisfaction. It takes tourists self-statement importance as the horizontal axis,
596 L. Li
satisfaction as the longitudinal axis, and the mean average of each dimension as
the midpoint of four quadrants, as shown in Fig. 2. The 2 elements situated in the
first quadrant means “Keep up the good work”, namely whose satisfaction and
importance are both higher. They are element 1 “Traditional folk performances
and activities” and 2 “The Three Kingdoms culture theme features”. The ele-
ment situated in the second quadrant means “Possible overkill”, namely whose
importance is lower but satisfaction is higher. Element The element situated in4
“Supporting facilities and services” is like that. The 2 elements situated in the
third quadrant show “Low priority”, namely whose importance and satisfaction
are both lower. They are element 3 “Fashion innovation and innovation” and 5
“Ticket price setting”.
(2) On the basis of extended importance
According to the results of the questionnaire, the mean value of overall tourist
satisfaction for Chengdu temple fair is 4.00. Then, we calculated extended impor-
tance of various factors accordance with the method of Deng [4] (See Table 4).
Table 4. Ranking of extended importance and satisfaction
Ranking of Elements Extended Satisfaction (5 Ranking of

extended importance credits) satisfaction
importance (The method
of Wei-jaw
Deng)
1 The Three 0.139 4.1 1
Kingdoms culture
theme features
2 Traditional folk 0.137 4.08 2
performances and
activities
3 Fashion elements 0.115 3.68 5
and innovation
4 Ticket price setting 0.09 3.83 4
5 Supporting facilities 0.087 3.96 3
and services
As a result, IPA analysis diagram of extended importance and satisfaction

was drawn. The basic method is same as the former, but we need to change self-
statement importance data into extended importance, as shown in Fig. 3. The 2
elements situated in the first quadrant means “Keep up the good work”, namely
still the elements 1 “Traditional folk performances and activities” and 2 “The
Three Kingdoms culture theme”. The element situated in the second quadrant
is “Possible overkill”, which implied lower importance but higher satisfaction.
The elements 5 “Supporting facilities and services” showed that. The element
Fig. 3. IPA analysis diagram of extended importance and satisfaction
situated in the third quadrant means “Low priority”, which implied extended
importance and satisfaction are both lower. It is the elements 5 “Ticket price
setting”. Element 3 “fashion elements and innovation” is situated in the fourth
quadrant, namely lower satisfaction but higher importance, which means “Con-
centrate Here”.
(3) Comparison two kinds of IPA analysis conclusion
The results showed that two kinds of IPA analysis conclusion were basically the
same. Slight differences were in two aspects. Firstly, on the ranking of extended
importance, “Three kingdoms culture theme” exceed the “Traditional folk per-
formances and activities”, in the first place. Extended importance of “Support-
ing facilities and services” was below “Ticket price setting”, which rowed at the
bottom. Secondly, by the IPA analysis diagram, we can find that in extended
importance analysis, the “Fashion elements and innovation” element was situ-
ated in the fourth quadrant, which means “Concentrate Here”.
From the two kinds of IPA analysis conclusion, it can be seen that the overall
tourist satisfaction of Chengdu temple fair is good. Core competitive elements
were traditional folk activities and the Three Kingdoms culture theme features.
Tourists’ attention and satisfaction on the both elements is high. We need to
“Keep up the good work” on the two. And for the evaluation of supporting facil-
ities and services, it was different from traditional scenic spots. Tourists showed
lower attention and higher satisfaction. The element was situated in the “Possi-
ble overkill” quadrant. The experience length during the festival activities and
distance of tourist source region can be related reasons. Ticket price is located
in the range of the low awareness and low satisfaction, which embodies the low
sensitivity of tickets present price. In addition, the “fashion and innovation ele-
ments” is situated in different quadrants of different analysis methods. On the
one hand, it reflected the innovation elements had a substantial impact to overall
satisfaction. On the other hand, it also indirectly showed the high attention to
the traditional culture. Visitors still hope the traditional festival activities can
reflect more pure original flavor and essence of traditional culture.
598 L. Li
5 Conclusions and Proposal

This study used the IPA analysis to do case investigation and data analysis for
tourists experience factors of traditional culture festival activities. We selected
Chengdu Temple Fair for empirical research. The self-statement and extended
importance were both used to make IPA analysis comparison. Elements need to
be focus on and improved were obtained to improve tourist experience evaluation
and satisfaction in such cultural festival activities in the future.
First of all, the traditional culture festival activity is a kind of traditional
cultural heritage value and means, and also part of the tourism activities. From
visitors experience the evaluation results, the majority of visitors still focus on
“traditional culture” and “characteristic theme culture” to this kind of festival
activities. Chengdu Temple Fair designed a lot of rich and colorful activities,
such as opening ceremonies, play god party activities, performing arts activities,
theme Lantern Festival, cultural exhibitions, lectures on culture, folk custom
activity, interactive games, traditional arts and crafts exhibition, snack sales and
other major sections. Visible, the key factor to improve the satisfaction degree
of the tourists is to find the core of culture and enhances the core resources of
the “traditional culture” in such festival activities.
Secondly, if we take Chengdu Temple Fair as an example, they paid attention
to not only the Spring Festival traditional culture in local region, but also its
own characteristics and advantages, namely the Three Kingdoms theme culture.
In the temple fair, Three kingdoms culture penetrated into all aspects of the
festival, from activity forms to content. It has been on full display and exten-
sion. Chengdu Temple Fair is different from other temple fairs of the nation’s
by conducting a lot of activities, such as the Three Kingdoms hero theme day,
Three kingdoms culture lamp exhibition, Three kingdoms culture exhibition,
the Three kingdoms interactive games, the Three kingdoms cartoon show, the
Three kingdoms characters in theme and ambience and so on. The Three king-
doms theme features have been the most prominent, attractive features and core
competitiveness of Chengdu temple fair. From the research results, the attention
and satisfaction of visitors to the Three Kingdoms culture theme are also one of
the best.
Thirdly, the IPA analysis tends to a wide range of applications in the indus-
try due to its simple and easy to interpret. The premise of the analysis request
variables on the two dimensions of importance and satisfaction evaluation are
independent of each other, and each variable is linear correlation between sat-
isfaction and overall satisfaction. However, in reality survey, the two assump-
tions are impossible to meet. The main reason is that, on the one hand, the
self-statement importance will be inevitably affected by the respondents’ satis-
faction perception. On the other hand, a certain degree change of satisfaction
will not bring the change of overall satisfaction [3]. As a result, this research
used two caliber of the self-statement satisfaction and extended satisfaction to
do IPA analysis. Managers can make countermeasures for selective in the premise
of contrast both the analysis results.
The main limitations of this study is because of selecting importance sorting

method for evaluation, we selected limited number of elements from several big
aspects. More comprehensive and detailed evaluation and analysis of cultural
festivals can be made by kinds of tourist experience factors. These can also
become the content of the related research in the future.
Acknowledgements. This study was supported by a grant from Higher School Cul-
tural and Social Science Fund of Sichuan Education Department (to DENG Qingnan)
(No. WHCY2016B02).
References
1. Arnould EJ, Price LL (1993) River magic: extraordinary experience and the
extended service encounter. J Consum Res 20(1):24–45
2. Bongkoo L, Shafer CS (2002) The dynamic nature of leisure experience: an appli-
cation of affect control theory. J Leisure Res 34(3):290–310
3. Chen X (2013) The modified importance-performance analysis method and its
application in tourist satisfaction research. Tourism Tribune 28(11):59–66
4. Deng WJ, Lee YC (2007) Applying kano model and ipa to identify critical service
quality attributes for hot springs hotel in peitou. J Qual 14(1):99–113
5. Deng WJ, Pei W (2007) The impact comparison of likert scale and fuzzy linguistic
scale on service quality assessment. Chung Hua J Manag 8(4):19–37
6. Fengying L, Lucang W (2007) Literature review of festival tourism research. For-
ward Position 7(8):33–35
7. Getz D (1997) Event management & event tourism. Ann Tourism Res 10:22–40
8. Haemoon O (2001) Revisiting importance-performance analysis. Tourism Manag
22(6):617–627
9. Hollenhorst S, Gardner L (1994) The indicator performance estimate approach to
determining acceptable wilderness conditions. Environ Manag 18(6):901–906
10. Hoyle LH (2002) Event marketing: how to successfully promote events, festivals,
conventions, and expositions. Wiley, New York
11. Manfredo MJ, Driver BL, Brown PJ (1983) A test of concepts inherent in expe-
rience based setting management for outdoor recreation areas. J Leisure Res
15(3):263–283
12. Martilla JA, James JC (1977) Importance-performance analysis. J Mark 41(1):77–
79
13. Mayfield TL, Crompton JL (1995) The status of the marketing concept among
festival organizers. J Travel Res 33(4):14–22
14. Meyer P (1970) Festival: USA and Canada. Ives Washbum, New York
15. Williams DR (1989) Great expectations and the limits to satisfaction: a review of
recreation and consumer satisfaction research. In: Outdoor Recreation Benchmark
1988: Proceedings of the National Outdoor Recreation Forum, pp 422–438
16. Yu Xy, Zhu GX, Qiu H (2006) Review on tourist experience and its research
methods. Tourism Tribune
How Cross Listing Effects Corporate
Performance: Measurement by Propensity Score
Matching
Yingkai Tang1 , Huang Huang1 , Hui Wang1 , Yue Liu2 , and Jinhua Xu3(B)
1
2
Urban Vocational College of Sichuan, Chengdu 610064, People’s Republic of China
3
School of Management, Guangdong University of Technology,
Guangzhou 510520, People’s Republic of China
Xujinhua@gdut.edu.cn
Abstract. With the wave of overseas-listed companies returning to A-

share market, this paper, taking the perspective of corporate governance,
uses Propensity Score Matching (PSM) to measure the difference of oper-
ating performance between cross-listed companies and non-cross-listed
companies, and uses Structural Equation Model and path analysis to
examine the impact mechanism and specific effect of cross listing on
operating performance. The results show: (1) Compared with non-cross-
listed companies, those companies cross-listed in both A and H-share
market, have witnessed a worsening of operating performance, a signifi-
cant increase of ownership concentration and liabilities to assets (L/A)
ratio, and a significant decrease of asset liquidity. (2) Cross listing has
affected operating performance mainly through debt structure, asset liq-
uidity, less through ownership structure. The results of this paper can
provide the theoretic and empirical basis for future cross listing of com-
panies in China.
Keywords: Cross listing · Propensity Score Matching (PSM) · Struc-

tural equation model · Corporate governance · Operating performance
1 Introduction
As the economic globalization accelerates and international capital market
becomes more and more integrated, many Chinese companies listed overseas
have returned to the domestic A-share market, and got cross-listed. Meanwhile
in recent years, domestic and foreign scholars have paid more attention to the
change of corporate performance after cross listing. However, empirical results
proved cross listing cannot improve corporate performance. Instead, foreign
researches showed that after cross listing, corporate internal and external perfor-
mance had gone down [8,15,26]. Researches on the change of corporate internal
How Cross Listing Effects Corporate Performance 601
environment after cross listing, showed cross listing can improve the stock liq-
uidity, reduce the agency cost [6,7], relax financing constraints and reduce the
cost of financing constraints [22]. In the past, researches on cross listing often
focused on the effect of cross listing on corporate performance, then according
to whether the company had been cross listed, set dummy variables for OLS
regression, ignoring the systematic bias; simultaneously, there was selective bias
in samples of cross-listed companies and non-cross-listed companies. No relevant
research on impact mechanism of cross listing has been found.
This paper, based on mass samples of China’s A-share market, uses Propen-
sity Score Matching (PSM) to examine the difference of operating performance
between cross-listed and non-cross-listed companies, and effectively avoid sys-
tematic bias in the study of cross listing. And then, this paper uses Structural
Equation Model to explore the impact mechanism of cross listing on operating
performance.
The structure of this paper is as follows: the first section is an introduction;
the second section will be literature review; the third section will be research
design; the fourth section will be empirical results; the last section will be con-
clusion and suggestions.
2 Literature Review
Jaffee and Russell [12] and Stiglitz and Weiss [27] pointed out that there was
information asymmetry in credit market, companies were able to complete
projects, but banks did not know about this. Most Chinese listed companies
get external financing from the credit market, where they encounter serious
financing constraints. Therefore, cross listing has always been regarded as an
important signal that a company focuses on protecting investors, strictly dis-
closes information and standardizes corporate governance, which can improve
public awareness and information environment of the company [2], increase the
level of information disclosure in the market, and strengthen investor protec-
tion [3,28], thereby reducing the degree of information asymmetry, and relaxing
financing constraints. Researches of Lins et al. [18] showed that cross listing
can loosen corporate financing constraints, reduce financing costs, and reduce
the dependence on internal cash; Doidge et al. [7] thought that compared with
non-cross-listed companies, cross-listed companies can expand their financing
channels, and prevent management from appropriation of corporate resources;
in China, empirical results of Pan and Dai [23] also showed that after cross list-
ing, corporate investments can be much less sensitive to cash flow, so financing
constraints can be effectively loosened. Empirical studies of Liu [19] showed that
returning to A-share market can ease corporate financing constraints to a certain
extent.
602 Y. Tang et al.
With less financing constraints, companies would reduce cash holdings. Gu

and Sun [10] pointed out that companies with serious financing constraints had
reserve cash to prevent financial risks; Myers [21] pointed out that enterprises
holding with a certain amount of cash holdings can seize every investment oppor-
tunities, and also can avoid information asymmetry caused by financing con-
straints. Zhang et al. [30] found that cross listing can enhance the awareness
of investor protection, and then reduce the company’s cash holdings. However,
according to Agency Theory and Free Cash Flow Theory, too much cash hold-
ings can facilitate management to seek personal gain, and increase haphazard
investments. La Porta [24] found that with too much cash holdings, management
is prone to damage investors’ interests by transfer prices, overinvestment, and
high salaries. Kalcheva [14] pointed out that too much cash holdings can harm
corporate performance and market value.
On the other hand, the mitigation of financing constraints will inevitably
lead to the expansion of credit financing. The Pecking Order Theory argues that
corporate debt financing outperforms equity financing. Jensen [13] pointed out
that debt, as a governance mechanism, can constrain behaviors of major share-
holders and management. Berger [4] argued that higher debt ratio can reduce
agency costs associated with external equity. However, there are many other
factors affecting corporate performance. Aslan and Kumar et al. [1] found that
debt financing did not restrain internal agency, but facilitated major sharehold-
ers to appropriate corporate resources. The Margaritis and Psillaki [20] pointed
out that the relationship of L/A ratio and operating performance presented an
inverted U curve, while the ownership concentration was positively correlated
to operating performance. Korajczyk and Levy [16] pointed out that without
financing constraints, companies would prefer equity financing during periods of
economic expansion, and debt financing during periods of economic contraction,
and vice versa.
Busaba et al. [5] found when overseas-listed companies return to A-share
market in China, they experience poorer post-issuance stock and operating per-
formance in comparison to companies purely listed in China. Meanwhile, compa-
nies from less-developed markets would take advantage of the enhanced visibility
and prestige associated with their foreign listing to issue shares domestically at
inflated prices and favorable terms, and to raise greater proceeds than they can
efficiently use. On the basis of Busaba’s studies, Kot and Tam [16] pointed out
that companies returning to A-share market should be regarded as the valuation
and timing of the whole A-share market. Fernandes and Giannetti [9] made sam-
ple survey on cross-listed companies worldwide, and drew the same conclusion.
In summary, most scholars held the view about cross listing that after cross
listing, operating performance would be damaged, while due to information dis-
closure, regulatory management and the expansion of financing environment,
corporate financing constraints can be mitigated, and corporate governance level
can be improved.
3 Research Design
3.1 Research Methodology and Procedures
In order to examine the effect of A+H cross listing, it is necessary to further

consider the endogeneity of cross listing. Traditional multiple regression analysis
might get biased and inconsistent results; and ordinary matching method cannot
solve the problem of endogeneity. This paper uses PSM proposed by Rosenbaum
and Rubin [25] to examine the effect of A+H cross listing. With PSM, propensity
score of each company is calculated, and the most comparable control samples
are matched with samples of A+H cross-listed companies in multiple dimensions;
and then the difference between the two groups can show the net effect of A+H
cross listing. Therefore, PSM can effectively reduce the selective bias. The specific
steps are as follows:
(1) Get propensity scores
Propensity score refers to the conditional probability of a company to get
A+H cross-listed.
P (X) = Pr [D = 1 |X ] = E [D |X ] . (1)
D is the research variable, standard samples are divided into AH group and
control group, if a company is cross-listed, D = 1, otherwise D = 0; P is proba-
bility of a company to get A+H cross-listed, that is, the propensity score; X is
the influencing factor of cross listing, that is, the matching variable.
Then, as what Dehejia did, this paper uses Logit binary regression model to
estimate:
P S (Xi ) = P (Xi ) = Pr [Di = 1 |Xi ] = exp (βXi ) / (1 + exp (βXi )) . (2)
P S is the propensity score of Company i to get cross-listed; D is the research

variable, if cross-listed, D = 1, otherwise D = 0; is the cumulative distribution
function of logical distribution; X is the matching variable, which affects cross
listing of a company, β is the parametric variable.
(2) Choose the matching method
After getting propensity scores, sample matching is conducted between AH
group and control group. Because of the consistency of propensity score, it is
difficult to find samples with the same propensity score. Therefore, this paper
examines the effect of A+H cross listing by k-nearest neighbors matching.
(3) Conduct balance test
Before getting matching results, it is necessary to test its effectiveness by
“hypothesis of commonality” and “hypothesis of independence”. “Hypothesis of
commonality” requires that control samples should be matched with samples
of cross-listed companies by propensity scores. “Hypothesis of independence”
requires that there should be no significant difference in the matching variable
between AH group and control group.
604 Y. Tang et al.
(4) Calculate the average treatment effect

The average treatment effect (ATT) of AH group and control group, that is,
the net effect of A+H cross listing on the research variable, is calculated with
the formula Eq. (3). If ATT is statistically significant, then cross listing does
have effect on the research variable.
AT T =E [Y1i − Y0i |Di = 1 ] = E {E [Y1i − Y0i |Di = 1, P (Xi ) ]}
= E {E [Y1i |Di = 1, P (Xi ) ] − E [Y0i |Di = 0, P (Xi ) |Di = 1 ]} . (3)
Y is the outcome variable, which is the relevant indicator measuring the effect of
A+H cross listing; Y1i and Y0i represent the outcome variable when a company
is cross-listed and not cross-listed respectively.
3.2 The Design and Definition of Variables

In case of the measurement bias caused by single indicator, this paper chooses
return on total assets (ROA) and Tobin’s Q ratio (TBQ) as proxy variables of a
company’s operating performance, chooses total profitability and market value to
measure its operating performance, and ownership concentration (GQJZD), L/A
ratio (DEBT) and asset liquidity (XJZCB) to measure corporate governance.
Drawing on previous researches, this paper chooses size of asset (SIZE), L/A
ratio (DEBT), Tobin’s Q ratio (TBQ), fixed asset ratio (GDZCB), debt per share
(MGFZ), ownership concentration (GQJZD) and types of enterprises (QYXZ)
as matching variables of logistic regression (Logit) (Table 1).
Table 1. The design and calculation of variables
Names of variables Symbols of Calculation of variables

variables
Return on total assets ROA Net return/Average total assets
Tobin’s Q ratio TBQ Company’s market value/Replacement cost
of assets
Liquidity ratio XJZCB Cash assets at the end of the period/Owner’s
equity at the end of the period
Size of assets SIZE Natural logarithm of total assets
L/A ratio DEBT Total debt/Total assets
Fixed asset ratio GDZCB Fix assets/Total assets
Debt per share MGFZ Total debt at the end of the period/Paid-in
capital at the end of the period
Ownership concentration GQJZD Sum of squares of the biggest three share-
holders’ shareholding ratios
Types of enterprises QYXZ State-owned enterprises is represented by 1;
otherwise, 0
3.3 The Screening of Samples

This paper chooses 58 companies cross-listed in both A and H share market
before 2008 as samples of treatment group, and all companies listed in A share
market between 2006 and 2014 as initial samples of control group; all samples are
screened following three principles: (1) listed financial companies are excluded;
(2) companies specially treated and particularly transferred are excluded;
(3) companies whose data is missing or abnormal are excluded. Most data of
this paper comes from CSMAR Solution, some data comes from corporate annual
reports, and the sample size of this research is 14448 companies. All the data is
processed with Stata11.2.
4 Empirical Results
4.1 The Screening of Matching Models
In order to make effective matching, regression results of five Logit models are
listed in Table 2. Regression results show that the larger size is, and the higher
Tobin’s Q ratio or the market value is, the more a company is willing to be
cross-listed; the higher fixed asset ratio is, the less current asset holdings a com-
pany has, then the higher demand for external financing is, therefore, the higher
fixed asset ratio is, the more a company is willing to be cross-listed; the higher
ownership concentration is, the more a company is willing to be cross-listed,
but after cross-listing, the power of controlling shareholders will undoubtedly
be diluted, therefore, considering the change of balance of shareholder power
after cross listing, we find that cross listing can reduce the power of controlling
shareholders, but nearly double the shareholding ratio of the second and third
largest shareholders, as a result, ownership concentration index is significantly
positive, indicating that major shareholders other than controlling shareholders
have great power to decide whether to be cross-listed or not; L/A ratio and
debt per share have negative effect on the tendency for cross listing; at the same
time, state-owned enterprises (SOEs) are more likely to choose cross listing, as
Huang [11] found that, because of China’s special capital market environment,
leaders of SOEs try to promote their own reputation and benefit their political
career through overseas listing; meanwhile, the total asset growth rate has a
negative effect on the tendency for cross listing, because controlling sharehold-
ers with strong growth ability reject cross listing for fear of the dilution of their
power [5].
As what Lian [17] did, this paper, considering the significance of variables in
Logit regression, uses Psedo-R2 value and AUC value, the area under the ROC
curve, to measure the effect of Logit regression model [29], and all Psedo-R2
values of 5 Logit regression models are more than 0.38, AUC values more than
0.9. When conducting PSM with Logit regression, if AUC value is more than
0.8, then indicators of the regression model are suitable. In this paper, all AUC
values are more than 0.9, so indicators are suitable. Considering Psedo-R2 values
and the significance of variables, Model 4 is chosen as the matching model.
606 Y. Tang et al.
Table 2. The design and calculation of variables
Variables Model1 Model2 Model3 Model4 Model5

∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗
SIZE 1.2905 1.3790 1.2828 1.3503 1.3604∗∗∗
−26.65 −28.74 −26.57 −26.55 −26.83
∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗
DEBT −1.8016 −1.4628 −1.7414 −1.3864 −1.4171∗∗∗
(−5.16) (−3.97) (−5.03) (−3.74) (−3.80)
TBQ 0.1169∗∗∗ 0.1230∗∗∗ 0.1221∗∗∗ 0.1212∗∗∗ 0.1138∗∗∗
−3.68 −3.31 −3.93 −3.24 −2.92
MGFZ −0.0153∗∗∗ −0.0145∗∗∗ −0.0158∗∗∗
(−4.76) (−4.52) (−4.87)
CWGG 0.0001
−0.02
ZZCZZL −1.1183∗∗∗ −0.9897∗∗∗
(−3.79) (−3.32)
GDZCB 0.5561∗∗ 0.5441* 0.5156* 0.5156*
−2 −1.95 −1.85 −1.85
YYSRZZL −0.0015
(−0.92)
CWFYL −0.075
(−0.12)
GQJZD 0.8975∗∗ 0.8374∗∗ 0.6849* 0.7494*
−2.23 −2.07 −1.68 −1.86
QYXZ 0.4158∗∗∗ 0.3591∗∗∗ 0.3926∗∗∗ 0.3326∗∗∗ 0.3690∗∗∗
−3.6 −3.04 −3.31 −2.79 −3.14
CONSTANT −31.763∗∗∗ −34.709∗∗∗ −32.487∗∗∗ −34.219∗∗∗ −34.303∗∗∗
Psedo-R2 0.3829 0.3858 0.3838 0.3865 0.3858
AUC 0.907 0.902 0.907 0.905 0.905
OBS 14448 14448 14448 14448 14448
Notes: (1) ∗ , ∗∗ , ∗∗∗ represent significance at the level of 10%, 5% and 1% respectively.
(2) The indicator of AUC represents the area under the ROC curve.
4.2 Matching Effect Test

Before getting matching results, balance test was conducted by “hypothesis of
commonality” and “hypothesis of independence”. Figure 1 is the probability dis-
tribution of propensity scores of AH group and control group before and after
matching. It shows that before matching, there is a significant difference in the
probability distribution of propensity scores between the two groups, and the dis-
tribution center of control group is significantly higher than that of AH group.
If the difference between the two groups is directly compared, there must be
large deviation in the results. However after matching, the probability distribu-
tion curve of propensity scores of control group shifts significantly to the right,
and the difference of the probability distribution of propensity scores between
the two groups significantly decreases, indicating that the matching significantly
amends the deviation of the probability distribution of propensity scores, the
matching is effective, and hypothesis of commonality is verified.
Fig. 1. Probability distribution of propensity scores of AH group and control group

before and after matching
Hypothesis of independence requires that after matching, there is no signif-

icant difference in variables between AH group and control group. As shown
in Table 3, after matching, all standard deviations sharply decrease, and all
absolute standard deviations are less than 10%, indicating that the matching
achieves good results. After matching, t-test results show that there is no signif-
icant difference between AH group and control group. Above all, hypothesis of
balance is verified.
4.3 Univariate Test
After the first phase of PSM, we match cross-listed companies with non-cross-
listed companies, and use univariate test to examine their operating performance
and corporate governance. In Panel A, all samples in treatment group and control
group are compared, and results are as follows: after cross listing in both A
and H − share market, the company’s cash asset ratio significantly decreases
by 0.0406, and ownership concentration significantly increases, indicating that
after cross listing, major shareholders increase their stakes in case of the dilution
of their power. L/A ratio increases by 0.0398, indicating that after cross listing,
it is easier to get external debt financing. Meanwhile, this paper chooses ROA
and TBQ as proxy variables of corporate value, which decrease by different
degrees after cross listing. In Panel B, samples of state-owned enterprises in
treatment group and control group are compared, and results show that cash
asset ratio slightly decreases in the significance level of 10%, while none of TBQ,
608 Y. Tang et al.
Table 3. Matching variables and balance test
Names of variables Mean Standard devi- Decrease t-test

ation (%) of standard
deviation (%)
Treatment Control group t p-value
group
SIZE Before matching 24.685 21.779 178 49.59 0
After matching 24.685 24.553 8.1 95.5 1.08 0.28
TQB Before matching 1.0989 2.055 −57.9 −9.76 0
After matching 1.0989 1.2167 −7.1 87.7 −1.16 0.763
DEBT Before matching 0.5854 0.46009 61 12.42 0
After matching 0.5854 0.5751 5 91.7 0.82 0.413
GQJZD Before matching 0.2895 0.16866 94.1 20.67 0
After matching 0.2895 0.28644 2.4 97.5 0.3 0.763
GDZCB Before matching 0.2976 0.24184 28 6.68 0
After matching 0.2976 0.31491 −8.7 69 −1.15 0.25
MGFZ Before matching 14.522 4.8996 39.7 20.79 0
After matching 14.522 14.263 1.1 97.3 0.13 0.899
QYXZ Before matching 0.4087 0.24021 36.6 8.28 0
After matching 0.4087 0.41087 −0.5 98.7 −0.07 0.947
Note: ∗ , ∗∗ , ∗∗∗ represent significance at the level of 10%, 5% and 1% respectively.
Table 4. The effect of cross listing in A and H-share market
Variables ROA TBQ XJZCB GQJZD LEV

Panel A Treatment group VS Control group
Treatment group (AH = 1) 0.0371 1.0988 0.1098 0.2894 0.5854
Control group (AH = 0) 0.0492 1.3462 0.1504 0.254 0.5456
Treatment group-control group 0.0121 0.2474 0.0406 −0.0354 −0.0398
T-value 3.0436∗∗∗ 2.2772∗∗ 5.2667∗∗∗ −3.2978∗∗∗ −2.6576∗∗∗
Panel B state-owned enterprises:
treatment group vs control
group
Control group (AH = 0) 0.0446 1.3158 0.1348 0.3106 0.5729
Treatment group-control group 0.0039 0.0958 0.0216 0.0134 −0.0193
T-value 0.6567 0.4833 1.8230∗ 0.8631 −0.8669
Panel B Private Enterprises: Treatment group VS Control group
Control group (AH = 0) 0.0526 1.3692 0.1622 0.2112 0.525
Treatment group-control group 0.018 0.3541 0.0548 −0.0729 −0.0557
T-value 3.4034∗∗∗ 2.9402∗∗∗ 5.4136∗∗∗ −5.1239∗∗ − 2.7719∗∗∗
Note: ∗ , ∗∗ , ∗∗∗ represent significance at the level of 10%, 5% and 1% respectively, and T-value
represents the statistics of t in the univariate test.
ROA, ownership concentration and L/A ratio passes the significance test. In
Panel C, samples of private enterprises in treatment group and control group
are compared, the change trend of each indicator is in accordance with that in
Panel A, and all indicators are significant at the level of 1%. However, after
cross listing, ownership concentration and L/A ratio of private enterprises are
apparently higher than that in Panel A, cash asset ratio decreases more than that
in Panel A, ROA and TBQ also decrease more than that in Panel A, indicating
that after cross listing, major shareholders of private enterprises have stronger
awareness of the risk of ownership dispersion, are more willing to increase their
stakes, and are more prone to expand the ratio of debt financing, and to use
financial leverage to gain profits (Table 4).

Table 5 is the descriptive statistical analysis of samples after matching. The mean
of market values is 39 billion yuan (e2 4.3864), indicating that these companies
are big in size, which reflects the choice of overseas-listed companies; their L/A
ratios are relatively higher, and the mean of L/A ratio is 56.94%; the mean of
return on total assets is 0.0419, and TBQ, 1.1603, which reflect corporate operat-
ing performance; the mean of sum of squares of the biggest three shareholders’
shareholding ratios is 0.2752, indicating that the ownership concentration of
China’s listed companies is relatively higher, which reflects the possibility of the
practice of what one person says counts. Relevant variables are winsorized.
Variables Mean Standard Minimum P25 P50 P75 Maximum

deviation
ROA 0.0419 0.0542 −0.2969 0.0126 0.0331 0.067 0.3113
TBQ 1.1603 1.1849 0.0541 0.4165 0.7628 1.4607 6.5051
XJZCB 0.1261 0.1067 0.0029 0.0506 0.0979 0.1662 0.6283
SIZE 24.3867 1.8899 19.6146 23.1212 24.4368 25.4237 30.6755
DEBT 0.5694 0.2045 0.0581 0.423 0.575 0.7307 0.9642
GDZCB 0.2902 0.216 0.0011 0.1033 0.2323 0.4801 0.8621
MGFZ 11.5098 15.0831 0.514 2.8933 6.4708 12.9131 77.8973
GQJZD 0.2752 0.147 0.0057 0.1669 0.2659 0.3746 0.7994
QYXZ 0.4174 0.4934 0 0 0 1 1
4.5 Path Analysis of Operating Performance

(1) The establishment of conceptual model of cross listing’s effect on operating
performance
610 Y. Tang et al.
Above all, cross listing will worsen operating performance. This paper will
explore the impact mechanism of crossing listing on operating performance from
debt structure, asset liquidity and ownership structure.
• Debt Structure. It is certain that changes of financing environment can affect
a company’s asset structure, and then operating performance.
• Asset Liquidity. The mitigation of financing constraints after cross listing can
improve the company’s cash holdings, and then affect operating performance.
• Ownership Structure. After cross listing, financing market is larger, ownership
structure changes, and corporate governance is improved, and then operating
performance is affected.
(2) Path effect analysis
Using Structural Equation Model to perform an iterative scheme for maxi-
mum likelihood with the conceptual model of Fig. 2, and path analysis results
are shown in Fig. 3. The probability of the whole model in adaptability test, is
P = 0.902 > 0.1, which does not pass the significance test and accept the null
hypothesis, indicating that the theoretical model and sample data are adapt-
able. Other adaptability indicators are RMSEA = 0.000 < 0.05, CFI = 1.000,
TLI = 1.025, IFI = 1.002, RFI = 1.000, NFI = 1.000, indicating that adaptabil-
ity of the model is great.
1 The Effect of Cross Listing on Operating Performance
As shown in Fig. 3, cross listing has no significant direct effect on proxy
variables of operating performance ROA and TBQ. Results of significant path
effect are listed in Table 6.
Fig. 2. Conceptual model of cross listing’s effect on operating performance
Fig. 3. Coefficient map of path analysis

Table 6. The effect of cross listing on corporate value
Types of path Paths Results

Path of ROA ah-lev-roa 0.040 × (−0.111) = −0.0044
ah-xjzcb-roa −0.031 × 0.108 = −0.0033
ah-gq-roa 0.035 × 0.026 = 0.0009
ah-lev-xjzcb-roa 0.040 × (−0.120) × 0.108 = −0.0005
ah-gq-xjzcb-roa 0.035 × (−0.128) × 0.108 = −0.0005
Total effect −0.0078
ah-lev-tbq 0.040 × (−3.122) = −0.1249
ah-xjzcb-tbq −0.031 × 2.308 = −0.0632
Path of TBQ ah-lev-xjzcb-tbq 0.040 × (−0.120) × 2.038 = −0.0098
ah-gq-xjzcb-tbq 0.035 × (−0.128) × 2.038 = −0.0091
Total effect −0.207
As shown in Table 6, the total effect of path of ROA is −0.0078, indicating

that after cross listing in both A and H-share market, corporate operating per-
formance turns down, specifically, after cross listing, debt structure and asset
liquidity change in different degrees, affecting operating performance. The total
effect of path of TBQ is −0.2070, indicating that after cross listing, corporate
market value is negatively affected, and decreases sharply. The coefficient of ah-
lev is 0.040, indicating after cross listing, because of the relaxation of financing
environment, companies are more inclined to increase debt financing to get rapid
expansion; the coefficient of ah-xjzcb is negative, indicating after cross listing,
companies reduce their cash holdings to lower the agent cost; the coefficient of
ah-gq is significantly positive, indicating after cross listing, ownership concentra-
tion increases, as market environment is more complex, and major shareholders
are more willing to increase their stakes to further consolidate their power.
As shown in ah-lev-roa and ah-lev-tbq, cross listing can significantly improve
debt capacity, indicating cross listing can enhance companies’ ability to obtain
external financing, however, the increase of L/A ratio cannot bring about growth,
but become an obstacle to the development. As shown in ah-gq-roa, after cross
listing ownership concentration increases, which improves corporate value, but
with limited effect. As shown in ah-xjzcb-roa and ah-xjzcb-tbq, after cross listing,
cash asset ratio sharply decreases, and low cash asset ratio has negative effect
on corporate value. Simultaneously, as shown in indirect path, after cross listing,
ownership concentration and L/A ratio also lead to the significant decrease of
cash holdings, and then the decrease of corporate value.
2 Path Decomposition of the Effect of Cross Listing on Corporate Value
Empirical researches on A+H cross-listed companies in China, choosing ROA
and TBQ as proxy variables of corporate value, show that after cross listing,
internal assets structure is not improved. Specifically, when ROA is the proxy
variable of corporate value, path of debt contributes 51.6% to the decrease of
612 Y. Tang et al.
corporate value after cross listing, path of cash flow, 34.0%, and path of owner-
ship, 14.4%. When TBQ is the proxy variable of corporate value, path of debt
structure contributes 65% to the decrease of corporate value, path of cash flow,
30.5%, and path of ownership, 4.5%, which is relatively lower (Table 7).
Table 7. Path decomposition of the effect of cross listing on corporate value
Proxy variables Types of path Paths Effect Contribution Ratio of contri-

degree bution degree
Path of ROA Path of debt ah-lev-roa −0.0044 0.0045 46.40%
ah-lev-xjzcb-roa −0.0005 0.0005 5.20%
Total −0.0049 0.005 51.60%
ah-gq-roa 0.0009 0.0009 9.20%
Path of ownership ah-gq-xjzcb-roa −0.0005 0.0005 5.20%
Total 0.0004 0.0014 14.40%
Path of cash flow ah-xjzcb-roa −0.0033 0.0033 34.00%
Total effect −0.0078 0.0097 100%
Path of TBQ Path of debt ah-lev-tbq −0.1249 0.1249 60.30%
ah-lev-xjzcb-tbq −0.0098 0.0098 4.70%
Total −0.1349 0.1349 65%
Path of ownership ah-gq-xjzcb-tbq −0.0091 0.0091 4.50%
Path of cash flow ah-xjzcb-tbq −0.0632 0.0632 30.50%
Total effect −0.207 0.207 100%
Note: Contribution degree is indicated by absolute effect, without differentiating positive effect
and negative effect.
5 Conclusion
This paper discusses the development status of companies cross-listed in both A
and H-share market in China, and examines the effect of cross listing on oper-
ating performance by using Propensity Score Matching (PSM), and then uses
Structural Equation Model and path analysis to explore the impact mechanism
and specific effect of cross listing on operating performance from paths includ-
ing debt structure, asset liquidity and ownership structure. The results show:
(1) Compared with non-cross-listed companies, those companies cross-listed in
both A and H-share market, have witnessed a worsening of operating perfor-
mance, a significant increase of ownership concentration and liabilities to assets
(L/A) ratio, and a significant decrease of asset liquidity. (2) Cross listing has
affected operating performance mainly through debt structure, asset liquidity,
less through ownership structure. Main contributions of this paper are as follows:
(1) using Propensity Score Matching (PSM) to solve the problem of endogeneity
in the research of cross listing; (2) further analyzing the impact mechanism and
specific effect of cross listing on corporate performance.
Suggestions are as follows: Firstly, A+H cross listing cannot improve oper-
ating performance of Chinese companies, because most A+H cross-listed com-
panies choose to be listed in H-share market first, and then in A-share market,
however, compared with the capital market in Hong Kong, the capital market
in the mainland of China is not completed enough, thus after A+H cross list-
ing, companies’ governance environment has not been improved. Therefore, the
top priority is to improve the capital market system. Secondly, domestic listed
companies should be encouraged to go out into the developed capital market for
financing, utilizing good markets in developed countries, so as to improve the
domestic market in an indirect way; after cross listing, companies’ ownership
concentration and ratio of debt financing significantly increase, indicating debt
financing becomes more important, and then companies face greater investment
risk. As path of debt is the main factor worsening operating performance, so
companies should adjust their ratio of debt financing to a reasonable range, in
order to avoid excessive leverage risk; after cross listing, with less financial con-
straints, companies tend to reduce their cash holdings, and then to make blind
investments. As decrease of asset liquidity is also a main factor reducing cor-
porate profitability, companies should properly hold liquidity assets in case of
need.
Acknowledgements. The paper is funded by Ministry of Education (AACZXX-01)

and Sichuan University (SKGT201602, HZRK201402).
References
1. Aslan H, Kumar P (2016) Controlling shareholders, ownership structure and bank
loans
2. Baker H, Nofsinger J, Weaver D (2002) International cross-listing and visibility. J
Finan Quant Anal 37:495–521 (in Chinese)
3. Barzuza M (2006) Lemon signaling in cross-listing
4. Berger A, Di Patti E (2013) Computation of discharge through side sluice gate
using gene-expression programming. Irrig Drain 62:115–119 (in Chinese)
5. Busaba W, Guo L et al (2015) The dark side of cross-listing: a new perspective on
cross-listing. J Bank Finan 57:1–16 (in Chinese)
6. Coffee J (2002) Racing towards the top? The impact of cross-listings and stock
market competition on international corporate governance. Columbia Law Rev
102:1757–1831 (in Chinese)
7. Doidge C, Karolyi G, Stulz R (2004) Why are foreign firms that are listed in the
U.S. worth more. J Finan Econ 71:205–238 (in Chinese)
8. Durand R, Gunawan F, Tarca A (2006) Does cross-listing signal quality. J Contemp
Acc Econ 2:48–67 (in Chinese)
9. Fernandes N (2014) On the fortunes of stock exchanges and their reversals: evidence
from foreign listings. J Finan Intermediation 23:157–176 (in Chinese)
10. Gu N, Sun J (2009) Financing constraints, cash flow volatility and corporate pre-
cautionary cash holdings. J Bus Econ 4:73–81 (in Chinese)
11. Hung M, Wong T, Zhang T (2012) Political considerations in the decision of chinese
soe to list in Hong Kong. J Acc Econ 53:435–449 (in Chinese)
614 Y. Tang et al.
12. Jaffee D, Russell T (1976) Imperfect information, uncertainty, and credit rationing.
Q J Econ 90:651–666 (in Chinese)
13. Jensen M (1986) Agency costs of free-cash-flow, corporate finance, and takeovers.
Am Econ Rev 76:323–329 (in Chinese)
14. Kalcheva I, Lins K (2007) International evidence on cash holdings and expected
managerial agency problems. Rev Finan Stud 20:1087–1112 (in Chinese)
15. King M, Segal D (2005) Are there longer horizon benefits to cross-listing untangling
the effects of investor recognition, trading and ownership
16. Kot H, Tam L (2016) Are stock price more informative after dual-listing in emerg-
ing markets evidence from Hong Kong-listed chinese companies. Pacific-Basin
Finan J 36:31–45 (in Chinese)
17. Lian Y, Su Z, Gu Y (2011) Evaluating the effects of equity incentives using PSM:
evidence from china. Front Bus Res China 5:266–290 (in Chinese)
18. Lins K, Strickland D, Zenner M (2005) Do non-US firms issue equity on us stock
exchanges to relax capital constraints? J Finan Quant Anal 40:109–133 (in Chinese)
19. Liu X, Tian M, Zhang C (2016) Has returning to a-share market eased the corpo-
rate financing constraints? Based on the analysis of cash-cash flow sensitivity
20. Margaritis D, Psillaki M (2010) Capital structure, equity ownership and firm per-
formance. J Bank Finan 34:621–632 (in Chinese)
21. Myers S, Majluf N (1984) Corporate financing and investment decisions when firms
have information that investors do not have. J Finan Econ 13:187–221 (in Chinese)
22. Pagano M, Roel A, Zechner J (2002) The geography of equity listing: why do
companies list abroad? J Finan 57:2651–2694 (in Chinese)
23. Pan Y, Dai Y (2008) Dual listing and financing constraints-an empirical evidence
based on chinese companies with ‘a+h’ dual-listing. China Ind Econ 5:139–149 (in
Chinese)
24. Porta R, Lopez-de Silanes F et al (2002) Investor protection and corporate valua-
tion. J Finan 57:1147–1170 (in Chinese)
25. Rosenbaum P, Rubin D (1983) The central role of the propensity score in obser-
vational studies for causal effects. Biometrika 70:41–55 (in Chinese)
26. Sarkissian S, Schill M (2009) Are there permanent valuation gains to overseas
listing. Rev Finan Stud 22:371–412 (in Chinese)
27. Stiglitz J, Weiss A (1981) Credit rationing and markets with imperfect information.
Am Econ Rev 71:393–410 (in Chinese)
28. Wojcik D, Clark G, Bauer R (2004) Corporate governance and cross-listing: evi-
dence from European companies
29. Zhang H, Liu J (2015) Does high tendency for cash holdings hurt corporate value
of private enterprises? A study based on financing constraint hypothesis or free
cash flow hypothesis. Macroeconomics 5:100–109 (in Chinese)
30. Zhang J, Cheng Z, Zhang J (2011) The impact of a cross-listing on the level and
the value of cash holdings-evidence from Chinese listed companies. J Shanxi Finan
Econ Univ 11:108–115 (in Chinese)
Factors Affecting Employee Motivation Towards
Employee Performance: A Study on Banking
Industry of Pakistan
Abdullah Khan1 , Shariq Ahmed2 , Sameer Paul3 ,

and Syed Hasnain Alam Kazmi3,4(B)
1
Khadim Ali Shah Bukhari Institute of Technology,
Karachi, Islamic Republic of Pakistan
abdullah@kasbit.edu.pk
2
Meezan Bank Ltd., Karachi, Islamic Republic of Pakistan
3
Mohammad Ali Jinnah University, Karachi, Islamic Republic of Pakistan
4
Southwest Jiaotong University, Chengdu, China
hasnain.alam@jinnah.edu
Abstract. The study provides the cognitive support for developing

employee’s motivation level at optimized level through consciousness
towards applied approach and unrealistic ideas. The diligence of the
research is to present the factors that help to motivate the employ-
ees to perform their job well. Whereas, the study restrains empirical
and descriptive analysis of variables which comprehensibly highlight the
significant relationship among the independent variable and dependent
variable; the close-ended questionnaire is constructed to compute the
statistical data. The study identifies the strong relationship among the
defined factors that enhance the employee’s motivation level in bank-
ing sector. Although the research is also aware the management and
the managers of banks about the most preference factors that improves
the employee motivation in the banks. This study valued the integrating
knowledge, experience, action to groom skilled employees in the banks.
Moreover, the main contribution of the research evaluating the role of
the motivational factors in banking industry.
Keywords: Motivation · Empowerment · Job environment · Banking ·

Employee performance algorithm
1 Introduction
In the most current scenario employees have became the primary strength of
any business where the employees deliver their continuous effort to drive orga-
nizations’ decisions into action to accomplish its objectives. Whereas, the moti-
vational factor regarding the employees is becoming a part of organizational
strategy. However, motivation is an aspect that helps an individual to select
or deselect the job, to continue and work proficiently during his/her job [16].
616 A. Khan et al.
Further, the motivational research proposes to facilitate the individuals to get

motivated through the target to satisfy their unconvinced desires; consequently
they apply their endeavor hoping to be satisfied with their needs. Conversely,
modern studies’ regarding the motivation doesn’t highlight the Basel desire con-
ception, but in addition they highlight the objective, equality, and values [24].
However, the after effect from the motivated workforce provides the less employ-
ees, turnover, retention and high productivity rate, which leads the organization
towards growth and development [13].
Further, the motivation explains as a dynamic strength among individual
which inspire the employees to do the incredible work to achieve the goal that
regulates to accomplish want or anticipation [19]. It is a complex subject of
human behavior which differs from person to person; therefore, varied individuals
will be stimulated in unique ways [12]. Individual’s behavior will state that what
encourages them to implement the effect of both capacity and inspiration [19].
Spiral inspiration can be characteristic and inappropriate [9]. A few theories of
inspiration were created to recognize the variables of inspiration that impact
human conduct in distinctive ways.
While in conventional management approach the idea is that “Manager recog-
nize the job optimum” is prevailing in current management. In modern concep-
tion the representative’s inspiration will be extremely enormous particularly for
service businesses. While, the resource provokes the representatives, that analyst
separated them into two gatherings; outside and inside components [10]. How-
ever, external elements, there will be inspiration implies such as meeting expecta-
tions condition, pursuing, organization’s picture, work assurance, advancement,
social environment, and position. In terms of learning organizations, when seeing
outer ecological components quick alteration and difficulty to be measured in a
firm national and global struggle, these organizations taking choices to profound
impact in organizations’ by keeping up its presence.
The method individual responds towards organizational working condition
depends over job reliability, job stress, working hours. Thus interpersonal associ-
ations imitate satisfaction in correspondence with a subordinate. Job satisfaction
is completely prejudiced by range of professional characteristic such as organi-
zational ethnicity, interpersonal affairs, worker empowerment, cooperation, and
self-apparent accountability [11]. Further, employees assess their working abil-
ity by receiving rewards furthermore identify additional fulfillment by getting
fair pay in exchange of their efforts within the organization [8]. Fair payment
for their efforts can effect in motivation and fulfillment for invulnerable work-
force. An individual value system is being studied while creating the awareness
of rewards and compensation, whereas individual’s values of receiving pride for
their efforts to support and increase their work level while it’s related to their
educational period or from salaried time line [7]. Individual at its initial level
of its career pays high values to intrinsic and extrinsic feature related to their
job. Moreover, individuals with the startup of their careers valued the rewarding
towards their work assessment and environment further the expert and experi-
enced individuals with high potential and opportunities also provide values. As
Factors Affecting Employee Motivation Towards Employee Performance 617
the different researcher consider the motivation as the function of compensation

and belief towards their jobs; because of which the high rewarding employees
having the high satisfaction in respect to their job [11].
2 Literature Review
2.1 Motivation
Motivation is mostly differentiate as one of the most influential predictors of indi-
vidual behavior and a key analyst of performance for essential aspect of behavior
[22]. Hence, it is not shocking that motivation show in a range of regulation paper
[15], scholars and executives have vast attention in accepting person motivation to
use communal medium and tools on behalf of companies [14]. Motivational aspect
(such as success, appreciation, accountability, effort, progression, individual devel-
opment) are associated to the application contented though hygiene aspects (such
as corporation strategy and management, interpersonal relatives, working envi-
ronment, pay, job safety, position, reimbursement) are associated to the employ-
ment background [6]. Herzberg measured hygiene aspects as extrinsic features,
which have an unsettling outcome on the staff job approach and create them even-
tually displeased in their profession when the needs aren’t sufficiently meet. Fur-
ther, motivational aspects are recognized as intrinsic feature, which make staff ful-
filled when needs are meet, conversely, do not formulate them disgruntled in the
deficiency of these [21]. Further the study affirms that motivation exists merely
when individual appreciates an optimistic correlation that endeavor direct to work
performance and work performance directs to incentives [6]. Moreover, the objec-
tive locale theory of motivation, where the theory emerge from the proposal of
anticipation theory as a target locale is an essential instrument, which work as an
“instant valve of individual act” [17] that direct the individuals towards attaining
the objective. The objective influences the performance by straight concentration,
organizing endeavor, rising determination and inspiring plan. Further the earlier
research [3] offer a perceptive of an individual attitude, prospect, plan and aims in
life. Moreover, this research helps business to considerate their employees’ objec-
tive in both their qualified and private life, whilst identifying the need to stimu-
late them. At the Similar time, they seem to improve the participation, reliability,
and assurance of the staff to their businesses. Positively, accepting the features
that influence the motivation of staff is an huge dispute for business, leaders and
executives [23]. Especially the people worth different aims in life. People might
have diverse needs and wants, ethics, ambition, objective, and prospect in life [1].
A motivational presumption effort to clarify the aspect that have straight or cir-
cuitous pressure on motivation and business performance, such as worker stimulus
and other motivational factors [23].
2.2 Job Environment

The environment where the job performs in their work surroundings can be diverse
from those fully contented challenging and unsafe for the life and health of employ-
ees. Whereas, the complex operational surroundings can be prejudiced by:
618 A. Khan et al.
• Peripheral aspect that comprises environment, meteorological condition, tem-

perature, clamminess, breeze, illumination in the office and intrusion, gases
and other harmful factors.
• Individual aspect that contain sexual characteristics and maturity of the
employee, exhaustion, repetitiveness, adverse stance throughout exertion.
• Another aspect associated with the business of production towards the period
of the working hour, working plan, working occasion, work speed, extreme
injure, etc. further, the work with complicated operational circumstances
may achieve only the workforce who convene precise requirements in stip-
ulations of maturity, gender, education, strength, corporeal and psychologi-
cal circumstances and psycho-physiological and emotional capacity. Complex
operational environment manipulates employees’ work performance.
Further, it is significant that staffs are qualified and knows how to perform
job related duties with the help of paraphernalia as insufficient utensil usage can
affect in mishap or divergences in recital no issue that how a lot and utensil was
appropriate. Learning of the individual must also be provided oriented to the
good use of defensive tools and individual safety [4].
2.3 Benefits
Benefits can affect the Employee performance in many ways. The benefits are not
frequently issued towards assessment and are consequently cheaper to achieve
during an employer throughout the market [2]. Therefore, cheaper the benefits
ought to amplify employee performance. Further the benefits can proceed as an
alternate for salary. To inspect the employer investigation facts and establish
the employee’s reduced the salary; once numerous benefits had obtained to the
individual after a few years [2]. Moreover, the employees sight the benefits and
salary as alternatives, disposed to provide salary in trade for extra benefits. While
the benefits set as a significant part of employee compensation correspondence
and simply acted as organize in different researches, but not as the key issues of
analysis.
Furthermore, in the literature the reward shape towards the benefits that
individual obtain from their work [11], and the important part of a worker job
approach such as business assurance, incentive and employee performance [12].
Thus, in any business rewards in cooperate as a significant role in structuring
and supporting the obligation between the workers that guarantee an average job
performance and employee loyalty. As, the business entity substitute premise,
employee penetrates corporate with the exact set of proficiency towards requir-
ing goals, and anticipate in revisit a respectable working conditions where the
employees assist to use their skills and satisfy and attain the desired goals [18].
The rewards enhance the height of effectiveness and efficiency of the individual
towards their jobs and in place of outcome there is an improvement of organiza-
tional performance [18].
2.4 Empowerment Concept
Empowerment as initially is an important and vital concept by which personnel

are capable to effect and have regulator over judgments that influence over their
productivity in their job. Moreover, this perception support workers to workout
determined regulator or consultant of their effort situation tangibly, communally,
ethnically and spiritually over the assets owed to them [5]. Partaking egalitari-
anism is extra essential idea or belief that qualities enlarging on. Rendering to
[25] sharing was careful to be an inherent quality of equality and thus a clarifi-
cation to suppression. This idea contains the methods and conclusions to inspire
development and autonomy.
2.5 Recognition
“Recognition explains as the appreciation to the employees for the level of per-
formance, and success or an influence to achieve goal. It can be intimate or com-
munity, fundamental or official. It is continuously in tally to pay.” [20]. However
the employees also need recognition. Persons like to distribute the celebration of
their success with others and have to be recognized in the organization. Whereas
the needed is satisfied it works to be an excellent motivator. Further, if employ-
ers depend on reward only to recognize influence and success it is most probable
that the employee’s goals will become altered to protect the pay and nothing
more while this will lead to a besmirched culture of the organization. Thus using
recognition correctly it will be cost-effective way of increasing success and allow
employees to feel intricate in the corporation culture [20].
3 Problem Statement
Today the banks are facing the problem of low productive employees in all the
sectors. However, in banking industry the lower level of motivating employees are
in number due to the nature of their work. However, this study assist to find out
the relationship among the following factors Benefits, Job environment, Empow-
erment, Recognition, to extract the best known factor among the employees to
enhance their motivational level. The study will investigate the impact level for
the above mentioned aspects on employee motivation of their job and their after
cause on their moral satisfaction and their individual performance in banks.
3.1 Research Question/Objective
The following are the objectives of the research:
• To extract the best known motivational factors that supports the moral inten-
tion of individual to perform.
• Which factor is highly motivates employees in the banking industry?
• Do the motivational aspects enhance the individual performance?
620 A. Khan et al.
4 Hypothesis
4.1 Motivational Factors
Hypothesisa1 (Ha1 ): There is a significant impact of Benefits over employee Moti-

vation in the Pakistani Banks.
Hypothesisa2 (Ha2 ): There is a significant impact of Job environment over
employee Motivation in the Pakistani Banks.
Hypothesisa3 (Ha3 ): There is a significant impact of Empowerment over
employee Motivation in the Pakistani Banks.
Hypothesisa4 (Ha4 ): There is a significant impact of Recognition over employee
Motivation in the Pakistani Banks.
4.2 Motivation and Individual Performance

Hypothesisa5 (Ha5 ): There is a significant relationship among motivation and
individual performance in the Pakistani Banks.
Hypothesis05 (H05 ): There isn’t a significant relationship among motivation and
individual performance in the Pakistani Banks.
5 Methodology
The research focuses on descriptive research design. Whereas, the precarious

connection among the answers about discovering out the connection of Benefits,
Job environment, Empowerment, Recognition, that help to enhance the moti-
vational level of individual also to optimize their work performance and as well
as its effect on organization. Furthermore explanatory study helps us to explain
the relationship among Dependent and independent variables. The descriptive
study helps to find out the line of intervention on the factors affecting the indi-
vidual motivation. However, the job satisfied employee’s leads the organization
towards the improvement in employee performance and customer satisfaction.
This study disquiet with discover how, when, what, why questions concerning
about the factors of motivation and clarification of distinct variables may lie in
the study which also ornate the modern association among variables that how
the delimit factors have their impact over the employee performance.
However, the research depends upon two areas, i.e. the base use which is
used for the tactical attitude which trusts over secondary data, comprising sem-
inar proceedings, books, intellectual journals, webinars and articles. Moreover,
the research sturdily depends over primary data source by accompanying the
survey on employee motivation to collect material about diverse aspects which
expressively influence over the employee job performance in banking industry.
The banks included in this study are Meezan Bank, Bank Al-Habib, Albaraka,
and Faysal Bank. The demographic aspect included age, gender, salary range,
job position, service period, and last promotion that received in their organiza-
tion cover in the questionnaire. Structured Close ended questionnaire is being a
part of my study. The questionnaire is the primary tools to gather data for my
dissertation. Further human comportment is coherent which rely upon reasons,
however all individuals has their own explanation for precise concern they are
attaining from convinced area and services even by exploiting certain individual
skills, this is the key reason for picking questionnaire as the primary to of data
collection.
5.1 Research Model

The motivated employees are the most important assets for the organization.
Although it is the considerate perceptional preference of employees being rightly
treatment, the functional loom is being centered towards the inference of an
employee’s motivation on executive success (Fig. 1).
Fig. 1. Conceptual model
Moreover, the employee motivation is in the interest to the executives since

it has a reciprocal outcome for every individual and every other thing in orga-
nization. Further, in banks employee’s motivation is the main factor due to the
nature of job and higher employee turnover rate and stress. Further testing will
I identify the effect of factors on employee motivation.
6 Results and Discussions

6.1 Employee Preference over the Factors Affects Employee
Motivation
The Analysis of the motivational factors that is most preferable by the employees
working in the banking sector industry. However, the factor state below indicates
that the maximum preferable factor is used to increase the level of individual
motivational (Table 1). Further the bar-chart indicate that Benefits share the
ratio of 18% to enhance employee motivational level as well as the interpersonal
skills of an individual has 13% of its contribution (Fig. 2).
Although the most affective factors that helps to improve the employee inner
motivational level is the individual empowerment the employee decision making
authority and its contribution among the other factors is almost 22%. However,
job environment, recognition and employee growth also have their impact on
employee motivational level over their job.
622 A. Khan et al.
Table 1. Preferable factors that improves employees motivation
Motivational factor Frequency Percentage

Benefits 79 26.1589404
Job environment 51 16.88741722
Empowerment 95 31.45695364
Recognition 77 25.49668874
Total 302 100
Fig. 2. Preferable factors that improves employees motivation
6.2 Multi Regression Empowerment, Benefits, Job Environment,

Recognition and Motivation
The regression equation between empowerment, recognition, growth & develop-

ment with motivation:
Motivation(y) = f (empowerment, recognition, benefits, job environment),

y = b1 x1 + b2 x2 + b3 x3 + b4 x4 + Error,
where b1 x1 signifies the intercept of empowerment x, b2 x2 signifies the intercept

of recognition x, b3 x3 signifies the intercept of benefits b4 x4 signifies the intercept
of job environment (Table 2).
Although the table of regression shows the value of R about 0.942 which iden-
tifies the strong relationship among independent variable (empowerment, recog-
nition, benefits and job environment) and dependent variable Employee motiva-
tion. Whereas, R square shows co-efficient of simple determination which is 88%
variance in empowerment, recognition, benefits and job environment explained
by motivation. The F-test identifies the rationality of the research model and the
model fitness. The significance of F-test is computed where the F calculated is
greater than F tabulated which is 0.05. Thus, the null hypothesis H01 , H02 , H03
and H04 has been rejected followed by the results and the results also identifies
the significant connection between empowerment, recognition, job environment,
benefits and employee motivation.
Table 2. Please write your table caption here
6.3 Regression Analysis Between Employee Motivation and

Employee Performance
The regression equation between motivation and employee performance:
Employee performance(y) = f (motivation).
The regression outcome identifies the relation among employee motivation and
performance which is being further specifies in the below mention table (Table 3).
Table 3. Regression analysis between employee motivation and performance
In Table 3, the outcomes indicates the positive connection among motivation

and employee performance is and identify that motivation is also a factor to
enhance the employee work performance. However the result shows the regression
R 0.722 and adjusted Rsquare is 0.522 which shows the significant relationship.
Fcalculated is greater than Ftabulated which is 0.05. Furthermore the result rejects
the null hypothesis H05 .
624 A. Khan et al.
6.4 Conclusion and Managerial Implication
The study specifically identifies the importance of motivation in the environment

of the banking industries and specifies the factors that improve the employee
motivation leveling banks. Whereas, the people spent their life working in the
banks, which highlighted the point to improve the motivation level of employees
towards their job. This research conducted to keep the motivation level high in
banks. Further, the results identify the factors (benefits, recognition, empower-
ment, and job environment) helps to improve the employee motivation level. Fur-
ther, each factor shows its level of importance. Moreover the study also explains
the satisfactory results from the banks employees about the define factors can
enhance employee motivation towards their jobs. The study illustrates that moti-
vated employees satisfied with their job more delicately and serve the organiza-
tion and customer, especially in the banks where the employees is directly linked
with the customer. Whereas the results shows that the factors not only enhance
the employee motivation as well as the factors also enhances their moral and
social behavior that tends to improve their work performance.
References
1. Ahmad K, Fontaine R (2011) Management from Islamic perspective. Pearson,
Malaysia Sdn Bhd (Person custom publishing)
2. Baughman R, Dinardi D, Holtz-Eakin D (2003) Productivity and wage effects of
“family-friendly” fringe benefits. Int J Manpow 24(3):247–259
3. Bhatti K (2015) Impact of Islamic piety on workplace deviance. PhD thesis, Inter-
national Islamic University Kuala Lumpur, Malaysia
4. Bratton J (2012) Human resource management: theory and practice. Palgrave
Macmillan, Basingstoke
5. Fetterman DM, Wandersman A (2005) Empowerment evaluation principles in prac-
tice. Eval Pract 15:1–15
6. Griffin RW (2012) Management eleventh edition. Cengage Learning, Hampshire
7. Hui MK, Au K, Fock H (2004) Reactions of service employees to organization-
customer conflict: a cross-cultural comparison. Int J Res Mark 21(2):107–121
8. Igalens J, Roussel P (1999) A study of the relationships between compensation
package, work motivation and job satisfaction. J Organ Behav 20(7):1003–1025
9. Jones G, George J (2011) Essentials of contemporary management 11(2):81–86
10. Jones PH, Chisalita C, Veer GCVD (2005) Cognition, technology, and work: spe-
cial issue on collaboration in context: cognitive and organizational artefacts. Cogn
Technol Work 7(2):70–75
11. Kalleberg AL (1977) Work values and job rewards: a theory of job satisfaction.
Am Sociol Rev 42(1):124
12. Kressler H (2004) Motivate and reward: performance appraisal and incentive sys-
tems for business success. Palgrave Macmillan, New York
13. Lai C (2009) Motivating employees through incentive programs. Jyvaskyla Univer-
sity of Applied Sciences
14. Leftheriotis I, Giannakos MN (2014) Using social media for work: losing your time
or improving your work? Comput Hum Behav 31(31):134–142
15. Levin MA, Hansen JM (2008) Clicking to learn or learning to click: a theoretical
and empirical investigation. College Stud J 42:665–674
16. Lin PY (2008) The correlation between management and employee motivation in
sasol polypropylene business, South Africa. Polypropylene Business
17. Locke EA, Latham GP (1991) A theory of goal setting and task performance. Acad
Manag Rev 15:367–368
18. Mottaz CJ (1988) Determinants of organizational commitment. Hum Relat
41(6):467–482
19. Mullins LJ (2011) Management and organisational behaviour. Oxford University
Press, Oxford
20. Robbins SP (2005) Organizational behavior. Prentice Hall, Upper Saddle River
21. Simons T, Enz CA (1995) Motivating hotel employees: beyond the carrot and the
stick. Cornell Hospitality Q 36(1):20–27
22. Steers RM, Shapiro DL (2004) The future of work motivation theory. Acad Manag
Rev 29(3):379–387
23. Sulaiman M, Ahmad K (2014) The perspective of Muslim employees towards moti-
vation and career success
24. Wang Y (2004) Observations on the organizational commitment of Chinese employ-
ees: comparative studies of state-owned enterprises and foreign-invested enter-
prises. Int J Hum Resour Manag 15(4–5):649–669
25. Wong S, Siu V, Tsang N (1999) The impact of demographic factors on hong kong
hotel employees’ choice of job-related motivators. Int J Contemp Hospitality Manag
11(5):230–242
How Content Marketing Can Help the Bank
Industrial: Experience from Iran
Sima Zomorodian and Yi Lu(B)

luyiscu@163.com
Abstract. In today’s market, Content marketing is one of the most effi-

cient and successful ways you can do that, but only if you know how to
use it correctly. We tried in this article to find, are there any relation-
ship between sensitive content, quality content, new content and content
marketing? The study had been designing a questionnaire consists of 37
questions in Likert scale to help us finding our hypotheses clearly, after
analyzed the data of questionnaires with SPSS software we fund our three
hypotheses and for accept or reject them we use another questionnaires
with 40 questions in Likert scale and among 550 randomly selected reg-
ular customers of Bank Mellat in Tehran (the capital of Iran) and 400
roperly filled questionnaires were collecting. Then analyzed all the ques-
tionnaire with Amos software. The result show all of those items(sensitive
content, quality content, new content) can have effect on the customer.
So an empirical investigation to determine factors influencing on con-
tent marketing in banking industry because in this area there are a lots
of brands and customers that can use all of them very easy so they
should find different way for attract customers. The result can help the
bank industrial to improve their content marketing and increase their
customer.
Keywords: Marketing · Content marketing · Export company ·

Customer
1 Introduction
Content marketing is the process of creating high-quality, valuable content to
attract, inform, and engage an audience, while also promoting the brand itself.
Content marketing is a kind of Marketing that use words, sounds and photos to
improve the knowledge of customer, introduce their brands and new products.
Content marketing is about giving away information to build relationships and
earn trust, but gating some of your best content is an acceptable and valuable
practice. In fact, this is not the only benefit of content marketing, another utility
of this kind of marketing is costs less than other types of marketing. Content
marketing is a new topic in Iran. Perviously, in bank industrial, they just tried
to give better services to customers and hope they introduction their brands. We
How Content Marketing Can Help the Bank Industrial 627
are trying to help the bank industrial for using content marketing to increase
their customer. Content marketing is comparable to what media companies do
as their core business, except that in place of paid content or sponsorship as a
measure of success, brands define success by ultimately selling more products
or services [13]. Content marketing is the marketing and Business Process for
creating and distributing relevant and valuable content to attract, acquire, and
engage a clearly defined and understood target audience - with the objective
of driving profitable customer action [10]. “Strong brands are based on a story
that communicates who is the company, is to communicate what you really are”
[2]. Therefore, the content marketing should be based on the company’s values.
Also, without considering the quality of the content, which is the most important
part of digital marketing, the choice of the frequency of promotion and of the
right social media plays a significant role in the success of the content marketing
campaign. When building such a strategy, it is important to always have in mind
all the social business strategy factors. All content has to work together; all the
groups need to work together. There are 6 essential components of content mar-
keting strategy: creation - curation - optimization - social media - amplification -
analysis [1]. Content marketing aids in brand recognition, trust, authority, cred-
ibility, loyalty and authenticity. Content marketing can help accomplish these
tasks for a variety of constituencies, and on several levels: for the organization
it represents, for a company’s products and services, and for the employee who
represent the business or service [8]. In the development of a content marketing,
there are numerous opportunities to be more relevant and effective. Planning
content that’s meaningful to the customers you’re trying to engage according of
content [9]. Content marketing is anything an individual or an organization cre-
ates and/or shares to tell their story. What it isn’t: A warmed-over press release
served as a blog post. It is conversational, human and doesn’t try to constantly
sell to you. It also isn’t a tactic that you can just turn on and off and hope
that will be successful. It has to be a mindset that is embraced and encouraged.
You’ve got to start thinking like a publisher and use that to plan and execute
your entire marketing plan which content of any variety should be a part [3].
Making professionally produced creative content available online is proving to
be a high-risk business, because of market fragmentation, high development and
production costs and the need to fund as yet unprofitable new services from the
declining revenue streams of “traditional” analogue and physical distribution
[12]. The marketing management philosophy which holds that achieving orga-
nizational goals depends on determining the needs and wants of target markets
and delivering the desired level of satisfaction more effectively and efficiently
than competitors [7]. The forms of content marketing are constantly changing as
new tools to create, publish and share that content are launched and others are
shut down. Enhancements and new functionality are added to content publishing
tools every day, which means the tools you are using to create, publish and share
content today might not be the tools you are using tomorrow [5]. The paper has
two main research objectives: (1) Study the influential factors on content mar-
keting (2) Study the influential factors on content marketing in Bank Mellat.
628 S. Zomorodian and Y. Lu
The paper is structured as follows: literature review that show our hypothesis
come from previous study. Research methodology that show our method for sur-
vey the elements. Result that show why we accept hypothesis. Discussion that
show the final result and give some suggestion for future research.
2 Literature Review
In today’s market, content marketing can solve some problems for company.
Actually in Iran this kind of marketing is new and we try in this article to show
how this kind of marketing can help to the bank industrial for attract more cus-
tomers. So we focus on previous studies to find some elements related to content
marketing and then use questionnaire to collected data from customers, after
that we analyze the data and make the hypothesis. Again we use questionnaire
to collected data to accept or reject the hypothesis. Content marketing is defined
as a marketing process of creating and properly distributing the content in order
to attract, make communication with, and understand other people so that they
can be motivated to do beneficial activities [4]. More specific solutions are out-
lined in content marketing strategies ‘where I focus on quality of the content,
its creation, distribution and evaluation’ [1]. In addition to monitoring mentions
and share, engaging with people who responded to the content can be a very
powerful way to spread your reach and to connect with potential prospects or
industry stakeholders [10]. The explanation is that “When a brand uses specific
words or stories that resonate with a consumer, they can dig deeper into who
they are as a consumer. By utilizing content marketing, brands can cater cam-
paigns and stories around buying patterns and personalities” [2]. Base on those
researches we make this hypothesis: Quality of content has effect on export com-
panies customers. To be relevant to your audience and create a powerful brand
you must win their trust and admiration. With the creation of valuable con-
tent you build interest that transforms into lasting relationships [11]. Due to
the characteristics of these emerging technologies, the digital content market is
growing rapidly and traditional content providers face service transformation
decisions. While a majority of the previous technology adoption studies have
focused on the viewpoints of users and customers, cost reduction, or electronic
channel related technologies, in this research we analyze the emerging technol-
ogy adoption decisions of competing firms for providing new content services
from a strategic perspective. Utilizing game theoretical models, we examine the
effects of market environments (technology cost, channel cannibalization, brand
power, brand extension, information asymmetry and market uncertainty) on
firms’ adoption decisions [6]. Base on those researches we make this hypothe-
sis: Sensitive of content has effect on export companies customers. However, by
implicitly restricting it’s focus to a buyer’s perspective, the resource-based and
relational views also leave the question of which resources and competencies add
greater value-for-customer largely unanswered [14]. According to some authors,
the contribution of different benefits to a relationship’s value will vary along the
relationship life-cycle. Specifically, the supplier’s know-how of the supply mar-
ket, the adaptation and improvement of extant products and the development
of new products become increasingly relevant in explaining the value of a rela-

tionship from the customer’s perspective as the life-cycle advances [14]. Base on
those researches we make this hypothesis: New of content has effect on export
companies customers.
For finding the elements related to content marketing in the bank industrial,
we find some elements from previous studies use questionnaire in the different
branch of Mellat bank and the customer answer them after that we analyze
the data. The questionnaire consists of 40 questions divided into two sections:
the first, consisting of 4 multiple choice questions, allow us to obtain general
information about the sample (gender, age, years of service and Education level).
The second, measured by a Cronbach’s Alpha analyze:
• Has quality of content effect on bank industrial customers?
• Has sensitive content effect on bank industrial customers?
• Has new content effect on bank industrial customers?
The area of this research is Tehran. The capital of Iran. Statistical Society
of this research is the customers of Mellat bank in 2016. Scale of five options
to describe the items listed in the questionnaire of “very large impact” to “very
little impact” is used. A total of 370 questionnaires were collected at the end of
the questionnaire, as shown in Table 1.
Table 1. The elements
Element Questionnaires The percentage

New content sharing 0.623 Less than 0.001 Accept
Brand 0.742 Less than 0.001 Accept
Content creation 0.701 Less than 0.001 Accept
4 Result
In the first analysis, factor analysis was exploratory and in the second one con-
firmatory. There are plans to test certain hidden factors beyond the variables.
In this case, orders are expected variables. In the researchers to test hypothe-
ses related to a particular factor structure of the action. In this analysis, the
researchers tried to obtain empirical data based on a model that assumes a
relatively small number of parameters, describing explain or justify. Reliability
Statistics: as shown in Table 2.
KMO and Bartlett’s Test: as shown in Table 3.
The SPSS software for data Factor and AMOS software for factor analysis
was used, the result is as shown in Fig. 1.
Table 2. Reliability Statistics
Cronbach’s Alpha N of items

.909 28
Table 3. KMO and Bartlett’s Test
Kaiser-Meyer-Olkin measure of sampling adequacy 0.895

Approx Chi-Square 2076.635
Bartlett’s test of sphericity df 190
Sig 0.0000
Fig. 1. Factor analysis result
(1) Quality of Content

Exploratory factor analysis in the first order for quality of content, and finally,
the three most important variable component factors identified in accordance
with the following table can be significant.as shown in Table 4.
(2) Sensitive Content
Exploratory factor analysis in the first order for sensitive content, and finally,
the three most important variable component factors identified in accordance
with the following table can be significant, as shown in Table 5.
(3) New Content
Exploratory factor analysis in the first order for new content, and finally, the
three most important variable component factors identified in accordance with
the following table can be significant, as shown in Table 6.
Table 4. Quality of content
The index Standard coefficient Significant level Acceptance or rejection situation

New content sharing 0.623 Less than 0.001 Accept
Brand 0.742 Less than 0.001 Accept
Content creation 0.701 Less than 0.001 Accept
Table 5. Sensitive content
The index Standard coefficient Significant level Acceptance or rejection

situation
The uncertainty of the market 0.586 Less than 0.001 Accept
The effects of market environment 0.492 Less than 0.001 Accept
The power of the brand 0.606 Less than 0.001 Accept
5 Discussion
In content marketing the customers are not presented with a direct suggestion,
instead they take another issue and information that is useful for them. At the
end they trust the content provider. Content marketing also can use with another
kinds of marketing.
In this paper, we examine customer opinion to analyze some elements that
can be effective on content marketing. In Iran this kind of marketing is new so
we try to show the bank industrial that they can use it to improve their business.
Our results show that there are elements related to content marketing. How-
ever, they are sensitive content, quality of content and new content. In con-
clusion, In today’s marketing, content marketing due to lower fee than other
marketing methods can be very effective. In the banking industry in Iran due to
the existence of different banks and brand diversity is very high, customers in
the choice of the Bank are having doubts.
The most simple and low-cost way to attract customers in order to help
create content provides the services of the Bank and the Bank’s competitive
advantages compared to other banks to inform customers. In this context must
be carefully that creating content is purely informational aspect and do not have
directly tried to attract customers, because it may have a reverse and does not
help to attract customers.
Table 6. New content
The index Standard coefficient Significant level Acceptance or rejection

situation
The development of new products 0.545 Less than 0.001 Accept
The life cycle of the product 0.714 Less than 0.001 Accept
The perspective of customer 0.810 Less than 0.001 Accept
After content attract trust of a group of customers, this group begin to

do mouth to mouth advertising, this action does not have charge for Bank.
According to the result of the study, New content can improve the knowledge
of customer and can help them to make a good decision. Must be careful in
creating content do not use imitation method because the duplicate content to
help not only our business, but does not cause a lack of quality in the customer’s
mentality.
To create the content need a team to be active and knowledgeable in this
field, that continuously target market observation and based on customer needs
to start creating content.
Content marketing is the present and especially the future of marketing and
inherently to marketing in general. Therefore, a content marketing strategy can-
not be successful without having a quality of content. Higher quality of content
created as compared to other competitors is the number of people who can trust
our business will increase. In today’s world most people unwilling to spend time
for unnecessary information. So the creation of quality content is one of the crit-
ical components in creating the content. Also when we improve the quality of
content we can make a competitive advantage and this is improve our chance
to have more customer. In the uncertainty market we should be careful about
which kind of content is necessary for customer.
In creating the content should be think about specific time like New year.
Create the appropriate content at any time, can be very useful in the success of
the business. So teams that create content need to be aware of the point and
make a suitable content for customers.
Content marketing is still in developmental stage. There is a need to consider
content for bank’s industrial. There is an opportunity for banks to use content
marketing more and more for finding new customer.
For future research there are some suggestions. We can do this research on
another industrial. Also evaluation of the factors influencing the content mar-
keting customer in a bank-centered and compare it to other banks. Factors influ-
encing the relation between detection of content-driven marketing and customer
satisfaction.
Acknowledgements. This research is supported by the Program of Social Science

Foundation of Sichuan (Grant No. SC16C010). The authors are also grateful to Sichuan
University for providing research funding (No. 2016SCU11036; No. skqy201640).
References
1. Augustini M (2014) Social media and content marketing as a part of an effec-
tive online marketing strategy. PhD thesis, Doctoral dissertation, Diploma thesis,
Masaryk University: Faculty of informatics, Brno
2. Baltes LP (2015) Content marketing-the fundamental tool of digital marketing.
Bull Transilvania Univ Brasov Econ Sci Ser V 8(2):111
3. Elisa R, Gordini N (2014) Content marketing metrics: theoretical aspects and
empirical evidence. Eur Sci J 10(34):92–104
4. Forouzandeh S, Soltanpanah H, Sheikhahmadi A (2014) Content marketing

through data mining on facebook social network. Webology 11(1):1
5. Gunelius S (2011) Content marketing for dummies. Wiley, New York
6. Jin BH, Li YM (2012) Analysis of emerging technology adoption for the digital
content market. Inf Technol Manage 13(3):149–165
7. Kotler P (2009) Marketing management: a South Asian perspective. Pearson Edu-
cation India
8. Lieb R (2011) Content marketing: think like a publisher-How to use content to
market online and in social media. Que Publishing, Upper Saddle River
9. Odden L (2012) Optimize: how to attract and engage more customers by integrat-
ing SEO, social media, and content marketing. Wiley, Hoboken
10. Patrutiu Baltes L (2015) Content marketing-the fundamental tool of digital mar-
keting. Bulletin of the Transilvania University of Brasov Series V: Economic Sci-
ences 8(2):111–118
11. Peltekoğlu DFB, Peltekoğlu YI (2016) The digitalization of communications:
towards content marketing. Temmuz, pp 776–780
12. Rosati E (2013) Originality in EU copyright: full harmonization through case law.
Edward Elgar Publishing, Cheltenham
13. Ström C, Pöysti S et al (2015) Developing an event marketing strategy using digital
media: case: data group jyväskylä
14. Zerbini F, Golfetto F, Gibbert M (2007) Marketing of competence: exploring the
resource-based content of value-for-customers through a case study analysis. Ind
Mark Manage 36(6):784–798
The Establishment and Application of AHP-BP
Neural Network Model for Entrepreneurial
Project Selection
Ke Wu1(B) and Xiaofeng Li2

1
Department of Enterprise Management, School of Business, Sichuan University,
wuke@scu.edu.cn
2
Department of Management Science, School of Business, Sichuan University,
Abstract. It is even the most difficult and critical step in starting a

business. Unfortunately, prior entrepreneurial project selection (EPS)
methods are barely satisfactory. In this paper, we constructed AHP-BP
neural network mode based on the traditional BP neural network. It
effectively reduces the input dimensions of traditional BP neural net-
work, significantly raises its learning speed, and improves the prediction
accuracy. In the end, empirical study shows that this model is able to
fully absorb the tacit knowledge of expert assessment, reduce subjec-
tive randomness, which provides important references to entrepreneurs
in starting a business, thus it is worthy being popularized and applied.
Keywords: Entrepreneurial project · Analytic hierarchy process · BP

neural network · Risk decision
1 Introduction
Entrepreneurial activities have shown increasingly close relationship with job
creation in both developed and developing countries, and have become a strong
driving force to world’s economic development. Meanwhile, entrepreneurship has
occupied a prominent position on the agendas of policymakers and researchers
[14]. In complex environments, Triple Helix has been operationalized in differ-
ent ways, spaces, and contexts where those agents are transforming their roles
in the development and strengthening of national innovation and entrepreneur-
ial ecosystems [3]. Unfortunately, at early start, failure is always going along
with those entrepreneurial projects. According to 2015 Annual Report of GEM
(Global Entrepreneurship Watch), the huge success made by 20%–30% startups
in China is at the cost of 70–80% failures. Even in developed countries, 35% star-
tups failed in the first year, almost half of them disappeared in five years [15].
Numerous reasons exist but mainly because entrepreneur started with an inap-
propriate project. For instance, Nabil provided a starting point for a stronger
The Establishment and Application of AHP-BP Neural Network Model 635
theoretical grounding of research that goes beyond the traditional interpretation

of entrepreneurial failure and opens new avenues of research opportunities to
explore and compare the different configurations that have emerged and identify
the possible dynamism and trajectories among these configurations [7]. How-
ever, how to use scientific methods to select appropriate projects is particularly
important for entrepreneurs [11].
Currently, scholars are dedicated to studying the theory and method of EPS,
and made some achievements [4,6,8,12,19,21]. These studies could be divided
into two categories: first, qualitative analyses based on theories, including SWOT
analysis, brainstorming and vision method. These studies has certain feasibility
but underper forms in practice. Secondly, quantitative analyses using mathe-
matical models, including income constraint method, utility function method,
risk ranking method and fuzzy evaluation method. These methods are barely
satisfactory in practice because of subjective factors. Therefore, how to choose
a right entrepreneurial project has become a difficult problem in management.
This paper is organized as follows. In Sect. 2, the evaluation index system of
EPS is built. In Sect. 3, the traditional BP artificial neural network is introduced.
In Sect. 4, we combined AHP and BP neural network to construct the AHP-BP
neural network model of EPS. In Sect. 5, the empirical research based on 21
entrepreneurial projects shows that the model is scientific and practical.
2 Evaluation Index System

Entrepreneurial Project Selection (EPS) means choosing the right entrepreneur-
ial projects at the beginning of business. There are a lot of factors influencing
EPS, subjectively and objectively. Thus, we need to follow certain principles and
methods when establishing scientific index system of EPS. The principles are as
follows:
(1) Goal oriented principles. The index system should be designed according
to the needs of entrepreneurs, and keep in line with policy guidance, and
it should be able to evaluate the market prospects of the entrepreneurial
project.
(2) Scientific principle. Evaluation indicators must be able to reflect and evaluate
the entrepreneurial projects correctly.
(3) Systematic principle. The indicators should be able to reflect the horizontal
and vertical relationship between different levels.
(4) Operational principles. The selected indexes should be easy to practice. The
index system should be simplified, be able to provide qualitative and quan-
titative evaluation to projects.
According to the principles above, the factors that influence the normal oper-
ation of the enterprise are classified into several parts. By sending questionnaires,
interviewing experts and referring to literature at home and abroad [2,17,20,22],
the frame is subdivided, supplemented and deleted in detail. Finally, the index
system is built, as is shown in Table 1.
636 K. Wu and X. Li
Table 1. The evaluation index system of entrepreneurial project
Target layer A Criterion layer Index layer Weight of Comprehensive

(weight) Index weight
The evaluation Environmental Industry environment 0.3946 0.026
index system of competitive-
Industrial policy 0.2119 0.014
entrepreneurial ness
project (0.066) Legal environment 0.2043 0.0135
Regional advantage 0.1892 0.0125
Technical com- Advanced technology 0.3946 0.1827
petitiveness
(0.463)
Technical complexity 0.2449 0.1134
Technical reliability 0.1374 0.0636
Technical substitutability 0.2082 0.0964
Technical accumulation degree 0.015 0.0069
Market com- Market demand 0.3187 0.0303
petitiveness
(0.095)
Market share 0.0579 0.0055
Competitor capability 0.1749 0.0166
Market entry barriers 0.2843 0.027
Cost-effective of products 0.1642 0.0156
Financial com- Capital requirement 0.2362 0.0158
petitiveness
(0.067)
Financial strength 0.1378 0.0092
Fund-raising ability 0.3422 0.0229
Internal rate of return 0.2044 0.0137
Financial management ability 0.0795 0.0053
Internal com- Practical experience 0.1074 0.0332
petitiveness
(0.309)
Leadership ability 0.3311 0.1023
Entrepreneurial skills 0.254 0.0785
Entrepreneurial knowledge 0.1223 0.0378
Social resources 0.1853 0.0573
3 BP Neural Network Architecture and Algorithm
3.1 Overview of AHP Methods
The analytic hierarchy process (AHP), is formally proposed by operational

research experts Thomas L. Satty (T.L. Satty) in the middle of 1970s [9]. AHP
decomposes a complex problem into various components, and group these factors
into an orderly hierarchical structure. The relative importance of the factors is
determined by pair wise comparison, and the overall order of the relative impor-
tance is determined by the comprehensive human judgment. The appearance
of AHP brings great convenience to the decision makers in solving the deci-
sion problem which is difficult to be quantitatively described, it can be used in
almost any scientific field. The detailed calculation steps of AHP method can be
found in literature [1]. But Analytic Hierarchy Process, which analyzes complex
decisions by organizing the problems into a multilayer hierarchic structure, is
a simple yet popular decision technique used extensively in every decision field
and is inadequate to handle the uncertain decision making problems [13].
3.2 Algorithm of Traditional BP Neural Network

BP (Back Propagation) neural network is a kind of error back propagation train-
ing algorithm for the multilayer feed forward network, it is one of the most widely
used neural network model [10], proposed by a group of scientists led by Rumel-
hart and McCelland in 1986. BP neural network reflects many basic features of
human brain. It is a highly complex nonlinear dynamic learning system, espe-
cially suitable for processing imprecise and fuzzy information.
Traditional BP neural network is a kind of non feedback feed forward network
[18]. It is composed of input layer, hidden layer and output layer. A typical
topology of three layer forward BP network is shown in Fig. 1. The first layer
consists of n input nodes. Each of the n input nodes is connect to each of the
r nodes in the hidden layer. The r output nodes of the hidden layer are all
connected to each of the m node in the output layer.
Fig. 1. Typical three layers BP network structure
The connection weights of input nodes i and hidden nodes j is wij , the
connection weights of hidden node j and the output node k is vkj , the thresh-
old of the hidden layer node is, the threshold of the output layer node is θj .

The model is θk provided with N learning samples (XP , YP )(p = 1, 2, · · · , N ).
XP = (xp0 , xp1 , · · · , xp(n−1) )T is the input vectors for learning sample P ,
YP = (yp0 , yp1 , · · · , yp(n−1) )T is the ideal output vector for learning sample P .
In input layer, let the input equals output, that is:
Opi = xpi (i = 0, 1, · · · , n − 1). (1)
The operating characteristic of the middle hidden layer node is:
n−1

netpj = wji Opi − θj , (2)
i−0
Opj = f (netRj )(j = 0, 1, · · · , r − 1). (3)

The operating characteristic of the output layer node is:
r−1

netpk = vkj Opj − θk , (4)
j−0
Opk = f (netpk )(k = 0, 1, · · · , m − 1). (5)

638 K. Wu and X. Li
Opi , Opj , Opk are the output of input layer node, hidden layer node and
output layer node separately. netpj , netpk are the input of hidden layer and
output layer separately. Sigmoid function is used as transfer function f (·): f (x) =
1/(1 + e−x ).
For convenience, in the Eqs. (2), (4), let:

wjn = −θj , Opn = 1, vkr = −θk , Opr = 1.
Equations (2) and (4) become:

n
1
netpj = wji Opi Opj = f (netpj ) = , (6)
i−0
1 + e−netpj
r
1
netpk = vkj Opj Opk = f (netpk ) = . (7)
j−0
1 + e−netpk
The learning process of the input and output samples of the traditional BP
network is the adjustment of error between the actual output and sample output
of network weights, it is actually a nonlinear optimization problem uses the most
common gradient descent method in mathematical programming:
∂E ∂E
w(n + 1) = w(n) − b · v(n + 1) = v(n) − b · .
∂w ∂v
In the above equity, b refers to learning rate. In the traditional BP algorithm,
b is constant. The value of b directly affects the convergence and convergence
rate of the algorithm.
3.3 Improvement of Traditional BP Neural Network Algorithm

Traditional BP neural network has many disadvantages. The steps of the tradi-
tional BP neural network which was improved by HE Chang-zheng etc. [5] are
as follows:
Step 1. Based on actual problem, set the number of input and output layer
nodes, and initialize the parameters, including the learning accuracy ε,
the prescribed number of iterative steps M0 , he upper limit of hidden
layer nodes R, learning parameters b, momentum coefficient a, the initial
number of R is 1. The admissible region of the initial weight H is divided
into N equal parts β1 β2 · · · βN .
Step 2. Input the learning samples and make the values of sample parameters
[0, 1].
Step 3. Give random numbers between −1 and 1 to the initial weighting matrix
βi (βi ⊂ H)(i = 1, 2, · · · , N ).
Step 4. Train the network with the traditional BP method.
Step 5. Judge whether or not the number of iterative steps is exceeding M0 . If
yes, continue. If no, turn to Step 7.
Step 6. If r + 1 exceed the upper limit of R, if yes, end, no, turn to Step 3.
Step 7. Determine whether the accuracy of learning meets the requirements, if
yes, continue, if no, turn to Step 4.
Step 8. Call the deletion and merge algorithm of hidden nodes, if there is a
deletion or merger, turn to Step 4, otherwise, algorithm completes.
4 BP-AHP Model Construction
Table 1 indicates that 24 factors affect the choice of entrepreneurial projects. By

using all of them as the input variables, the complexity will obviously increase
and we may lose the calculation precision. Fortunately, AHP could address this
issue. Firstly, acquire the important evaluation indicators by AHP, and use them
as input variables of the BP network. Secondly, train the improved BP algorithm.
When fitting with other conventional methods, the accuracy of AHP is often not
as good as the BP algorithm. Therefore, we combine these two methods organi-
cally, so as to enhance the problem solving ability of artificial neural network in
the complex nonlinear problem.
4.1 Determine the Input Variables of BP Network Using AHP
Based on the analysis of the factors influencing EPS, a multi-level hierarchical

structure can be established (see Table 1). Then, construct the judgment matrix
and determine the weight of the index. Invite experts to conduct in-depth analy-
sis and reach a consensus on the importance of each index, use the 9 − 1 scale
methods to establish the various layers of λmax of each judgment matrix and its
corresponding characteristic vector are obtained respectively. Finally, the feature
vectors are normalized to get the comprehensive weight of each evaluation index,
see Table 1, CR = 0.016 < 0.10, consistency test passed.
The results show that important indicators affect the choice of entrepreneur-
ial projects are: Industry Environment (x1 ), Advanced Technology (x2 ), Techni-
cal Complexity (x3 ), Leadership Ability (x4 ), Technology Substitutability (x5 ),
Entrepreneurial Skills (x6 ), Technical Reliability (x7 ), Social Resources (x8 ),
Entrepreneurial Knowledge (x9 ), Practical Experience (x10 ), Market Demand
(x11 ), Market Entry Barriers (x12 ), Fund-Raising Ability (x13 ), Competitor
Capability (x14 ), Capital Requirement (x15 ). The above 15 indicators have 90%
or more impact on EPS; they are input variables of the BP network. Other
indicators are less influential, can be deleted.
Denote Vj (j = 1, 2, · · · , n), V = {V1 , V2 , · · · , Vn } represents the com-
ment set of entrepreneurial projects. V = excellent, fine, ordinary, bad =
A1, A2, A3, A4 = 1, 2, 3, 4.
In order to evaluate the entrepreneurial projects objectively, we set up a
team to select evaluation indicators of entrepreneurial projects. The team mem-
bers include 3 entrepreneurs, 3 venture capitalists, 3 business management
experts. The experts involved
9 in the evaluation are required to grade the indexes.
Then, denote vk = 19 i−1 vui , vk stands for the evaluation value of index, xk ,
640 K. Wu and X. Li
i(i = 1, 2, · · · , 9) stands for the 9 experts. vu1 stands for the grade given by
experts, u1 ∈ [0, 1] stands for “poor”, u2 ∈ (1, 2] stands for “bad”, by analogy,
u5 ∈ (4, 5] stands for “excellent”.
4.2 Construction of BP-AHP Model
According to the structure of BP neural network, the BP-AHP model includes

three layers:
(1) Input Layer. First of all, assign the evaluation of the index selected by the
AHP method; and they are used as the input layer of the neural network
variables. From what has been stated above, 15 input layer variables belong
in the neural network.
(2) Hidden Layer. According to the improved learning algorithm of BP neural
network proposed by TIAN Jing-wen etc. [16], set the number of hidden layer
nodes as 1, then, the network will learn by itself until we get the appropriate
number of nodes.
(3) Output Layer. The comments of entrepreneurial project evaluation are
defined as two categories: “choose it” and “give it up”, these can be rep-
resented by the output vectors (1, 0), (0, 1). So, the number of output layer
nodes is 2.
AHP and BP neural network methods are combined to create the BP-AHP
model of EPS, and the basic steps of the algorithm are as follows:
Step 1. Simplify the evaluation index of EPS, and the important index is
selected as the input layer of neural network.
Step 2. Set the number of neural network output layer nodes, and initialize the
network parameters (including the given learning accuracy ε, provisions
iteration number M0, the limit of hidden nodes r, learning parameter
η. The initial hidden node number is set to 1.)
Step 3. Input the learning samples and make the values of sample parameters
[0, 1].
Step 4. Give random numbers between −1 and 1 to the initial weighting matrix.
Step 5. Train the network with the modified BP method, and determine the
weight matrix between each layer.
Step 6. Judge whether or not the number of iterative steps is exceeding the
prescribed. If yes, end; If no, go back to step 5 and continue learning.
Step 7. Collect values of the indexes and process these data to make them [0, 1].
Step 8. Input the processed data to the trained BP neural network and get the
output.
Step 9. According to the output results, combined with the entrepreneurial
project index reviews set, choose the right entrepreneurial projects.
5 Empirical Research
We choose 21 entrepreneurial cases from Sichuan University in the past 3 years,
of which 10 projects have been successful, 11 entrepreneurial projects ended
in failure. Use the BP-AHP model to evaluate entrepreneurial projects selected.
The cases are: computer training (Q1 ), online digital products (Q2 ), development
and production of management software (Q3 ), computer DIY and maintenance
(Q4 ), website construction and service (Q5 ), computer cleaning services (Q6 ),
water saving car washing (Q7 ), reward supervision social software (Q8 ), online
customized travel (Q9 ), advertisement designing (Q10 ), anime 3D printing (Q11 ),
a beverage shop (Q12 ), fruit fast delivery (Q13 ), creative souvenirs design (Q14 ),
fast dry cleaning service (Q15 ), household service (Q16 ), cozy cafe (Q17 ), creative
home decoration (Q18 ), children’s art training (Q19 ), architectural design (Q20 ),
online second hand market (Q21 ). Q1 , Q2 , Q3 , Q5 , Q6 , Q10 , Q12 , Q16 , Q17 , Q19 ,
Q21 failed in the first half year; others operate smoothly. We used the first 16
entrepreneurial projects as training samples of the BP - AHP model, the last 5
items as prediction evaluation samples.
5.1 BP-AHP Neural Network Training
In the process of quantitative evaluation of the selection of venture capital

projects, 9 experts were invited to set up the quantitative evaluation group
of the index system. According to the evaluation index of the evaluation team,
the evaluation team can be used to evaluate the evaluation index of the entre-
preneurial project 1 to the venture capital project (xk ), and the evaluation value
of the evaluation index xk is obtained
1
9
vk = vui .
9 i−1
The output of successful projects are (1, 0), the output of failed projects are
(0, 1). The BP neural network architecture is built by 15-1-2 (The number of
input layer nodes is 15, the number of hidden layer nodes is 1, the number of
output layer nodes is 2). We initialize the network (the upper limit ε = 0.0002,
learning rate η = 0.5, Inertia Parameter a = 0.1), Make xk (k = 1, 2, · · · , 15) ∈
[0, 1] by dividing them with 10, and input the processed data to the BP network
model. Then the network is trained by the modified BP learning algorithm and
the network architecture becomes 15-9-2. At the same time, we get the optimized
network weight matrix.
5.2 Entrepreneurial Project Selection
Now, it is time to evaluate Q17 , Q18 , Q19 , Q20 , and Q21 with the trained neural
network. Give values to the indexes xk of the 5 entrepreneurial projects, Make
xk ∈ [0, 1] by dividing them with 10. Input the processed data to the trained
642 K. Wu and X. Li
BP neural network and get the outputs as shown in the second row of Table 2.
According to the principle of maximum membership degree, we can determine
the evaluation results of each project; see the third row of Table 2.
Table 2. BP neural network output
Entrepreneurial BP neural network Evaluation result The actual operation of

projects output the business project
Q17 (0.0129, 0.9965) Give it up Failed
Q18 (0.9847, 0.0355) Choose it Succeed
Q20 (0.9798, 0.0032) Choose it Succeed
It can be seen from Table 2, the projects which evaluation is “Give it up”
failed in actual operation; the “Choose it” ones succeed. The predictions of BP
neural network completely accord with the practical ones, which indicate that,
this risk early warning model feasible and effective.
6 Conclusions
EPS is a comprehensive prediction problem affected by many factors, so it is
important to choose appropriate evaluation method. According to the charac-
teristics of entrepreneurial projects, this paper analyzes the factors that affect
EPS, and constructs the evaluation model of EPS based on AHP and BP neural
network subjectively and objectively. This model can not only extract the main
attributes of entrepreneurial projects, reduce the input variables of the neural
network, but also reduce the complexity of the neural network and the training
time, improve the generalization ability, reasoning ability and classification abil-
ity. Therefore, the model is feasible and effective, and it has important reference
value in guiding the entrepreneur to choose the right project.
Acknowledgments. This research was supported by the National Social Science

Foundation of China (No. 16BGL024).
References
1. Brennan MJ, Schwartz ES (1985) Evaluating natural resource investments. J Bus
58(2):135–157
2. Deng Z, Yang P, Yang X (2015) The way of college students’ entrepreneurial project
selection. China Manag Informationization 18(21):235–236 (in Chinese)
3. Guerrero M, Urbano D (2016) The impact of triple helix agents on entrepreneurial
innovations’ performance: an inside look at enterprises located in an emerging
economy. Technol Forecast Soc Change
4. Hagan MT, Demuth HB, Beale M (1996) Neural Network Design. PWS Publishing
Company, Thomson Learning, Boston
5. He C, Li X, Yu H (2002) The new improvement of bp neural network and its
application. Math Pract Theory 32(4):555–561
6. Jain R, Ali SW (2013) A review of facilitators, barriers and gateways to entrepre-
neurship: direction for future research. South Asian J Manage 20:122–163
7. Khelil N (2015) The many faces of entrepreneurial failure: insights from an empir-
ical taxonomy. J Bus Ventur 31(1):72–94
8. Kumaraswamy MM, Rahman MM (2002) Risk management trends in the construc-
tion industry: moving towards joint risk management. Eng Constr Architectural
Manage 9(2):131–151
9. Lockett G, Stratford M et al (1986) Modelling a research portfolio using ahp-a
group decision process. R&D Manage 16(2):151–160
10. Lu X (2010) The main strategies for college students to choose the entrepreneurial
projects. Innov Entrepreneurship Educ 6:21–23 (in Chinese)
11. Nielsen SL, Lassen AH (2011) Images of entrepreneurship: towards a new catego-
rization of entrepreneurship. Int Entrepreneurship Manage J 8(1):35–53
12. Pena I (2002) Intellectual captital and business venture success. J Intellect Capital
3:180–198
13. Ren P, Xu Z, Liao H (2016) Intuitionistic multiplicative analytic hierarchy process
in group decision making. Comput Ind Eng 101:513–524
14. Schelfhout W, Bruggeman K, Maeyer SD (2016) Evaluation of entrepreneurial
competence through scaled behavioural indicators: validation of an instrument.
Stud Educ Eval 51:29–41
15. Schwab K (2015) The Global Competitiveness Report, 2014–2015. World Economic
Forum, Geneva
16. Tian J, Gao M (2006) Artificial Neural Network Algorithm Research and Applica-
tion. Beijing Institute of Technology Press (in Chinese)
17. Urban B, Nikolov K (2013) Sustainable corporate entrepreneurship initiatives: a
risk and reward analysis. Technol Econ Dev Econ 19:383–408
18. Wen L, Li L, Liao S (2013) Analysis on influencing factors of college students’
entrepreneurial project selection. Project Manage Technol 11(6):89–92 (in Chinese)
19. Xie Q, Yuan X (2006) A review of the evaluation system of venture investment
projects. Sci Technol Manage Res 8:182–187 (in Chinese)
20. Yan Y (2011) Application of fuzzy comprehensive evaluation method in the
selection of college students’ entrepreneurial projects. Project Manage Technol
9(12):51–55 (in Chinese)
21. Zhang N, Yan P (1998) Neural Networks and Fuzzy Control, 1st edn. Tsinghua
University Press, Beijing (in Chinese)
22. Zhu Y (2014) Selection of college students’ entrepreneurial projects from the per-
spective of the long tail theory. Innov Entrepreneurship Manage 5(6):52–54 (in
Chinese)
Improving the User Experience and Virality
of Tourism-Related Facebook Pages
Ayako Sawada1(B) and Taketoshi Yoshida2

1
Hokuriku Gakuin Junior College, Ishikawa, Japan
sawada@hokurikugakuin.ac.jp
2
Japan Advanced Institute of Science and Technology, Ishikawa, Japan
Abstract. The percentage of people using social network services

(SNSs) exceeds 60% in Japan; SNSs are being used by people of var-
ious age groups. Recently, the number of companies that use Facebook
Pages to advertise their activities and the number of organizations that
use them to disseminate information have been increasing. In tourism,
disseminating local information on Facebook Pages and increasing the
attractiveness of the pages are expected to lead to the promotion of
regional tourism industries. In this study, the current state of the use
of tourism-related Facebook Pages was examined. The engagement rate
and the number of fans were used as an indicator of the attractiveness
of Facebook Pages. An increasing the attractiveness of Facebook Pages
was examined through emotion analysis of user comments.
Keywords: Facebook page · Regional development · User comments ·

Emotion analysis
1 Introduction
Along with the spread of smartphones, the number of people using social network
services (SNSs) has been increasing. The number of monthly active users of
Facebook worldwide (those who use Facebook one or more times per month)
reached 1.79 billion as of 30 September 2016 [1]. In addition, it is reported that
the number of monthly active users of Instagram is 0.6 billion (as of 15 December
2016) [4], that of Twitter is 0.313 billion (as of 30 June 2016) [12], and that of
Line is 0.2184 billion (as of 31 March 2016) [6]. According to the 2015 White
Paper on Information and Communications in Japan published by the Ministry
of Internal Affairs and Communications [5], the SNSs used in Japan over the
last one year include Line (37.5%), Facebook (35.3%), and Twitter (31.0%).
The number of people using SNSs generally tends to decrease with increasing
age: approximately 50% of people aged 20 or younger, less than 40% of people in
their 30’s and 40’s, and more than 20% of people aged 60 or older use Facebook.
It is reported that Facebook is being used by people of various age groups.
With this background, a variety of research projects on Facebook have been
Improving the User Experience and Virality 645
carried out. Wilson et al. classified 412 sociology papers on Facebook into five
categories: descriptive analysis of users, motivations for using Facebook, identity
presentation, the role of Facebook in social interaction, and private information
disclosure [13].
Recently, both the number of companies that use Facebook Pages to advertise
their activities and the number of organizations that use them to disseminate
information have increased. Ohara et al. clarified the characteristics of photos
that increase the responses of users of Facebook Pages of fast-food companies
by principal component analysis [8]. In tourism, disseminating local information
on Facebook Pages and increasing the attractiveness of pages are expected to
lead to the promotion of regional tourism industries. By multiple regression
analysis, Sabate et al. investigated the factors that increase the attractiveness of
Facebook Fan Pages (currently, Facebook Pages) of five travel agencies in Spain
using the numbers of “Likes” and comments regarding each post as indices of
attractiveness [9]. The results indicate that the number of “Likes” is affected
by the presence of video images and photos in the post and that the number of
comments is affected by the presence of photos and the time period of posting.
However, the methods of using Facebook are rapidly advancing and most recent
posts almost always include photos or video images.
Research on Facebook and its usefulness for those involved in the promotion
of regional tourism industries is limited, although how-to books on posting on
Facebook Pages [3] and marketing strategies using Facebook [10] have been pub-
lished. Consequently, Sawada et al. [11] analyzed the content of tourism-related
posts on Facebook Pages with high attractiveness and clarified the following.
The responses of users of Facebook Pages offering information related to Christ-
mas, flowers, and foods were high, whereas those for pages offering information
on cultural events tended to be low. The responses of users of Facebook Pages
offering real-time information, such as information on the start of the bloom-
ing of cherry blossoms, tended to be high. In addition, it was found that the
responses of users were affected by the type of photos posted.
To increase the attractiveness of Facebook Pages, the analysis of not only
the posts of organizations but also the feelings of users that cause them to post
comments need to be clarified. The purpose of this study is to analyze the current
state of the use of Facebook Pages in tourism in Japan and the sentiments behind
the comments on Facebook Pages with high attractiveness, and to clarify the
feelings of the users who posted comments and the factors behind the increase
in the attractiveness of these pages.
In Sect. 2, the current state of the use of Facebook Pages offering regional
tourism information is analyzed in terms of the number of fans, management,
and the responses of users to find the characteristics of Facebook Pages with high
attractiveness. In Sect. 3, user comments on these Facebook Pages are analyzed.
In Sect. 4, the factors behind the increase in the attractiveness of pages are
discussed on the basis of the results of analyzing user comments. Section 5 is the
conclusion.
646 A. Sawada and T. Yoshida
2 Facebook Pages in Tourism

Facebook navi (http://f-navigation.jp/) is the only navigation site in Japan
recognized by Facebook, Inc. In this study, 842 Facebook Pages registered in
the Facebook Page ranking in the travel category of Facebook navi were exam-
ined. The data were collected on 10 January 2016. In this section, Facebook
Pages offering regional tourism information were selected and the relationship
between the number of fans and organization managing the site was clarified.
Then, Facebook Pages with high attractiveness were determined on the basis of
the responses of users and the number of fans.
Fig. 1. Content of Facebook Pages in the travel category
2.1 Contents of Facebook Pages
The authors browsed the 842 target pages in the travel category and classi-
fied them depending on their contents. Figure 1 shows the results. Among the
842 pages, 30% of the pages are managed by accommodation facilities, 14% by
tourist facilities such as aquariums and museums, 7% by travel agencies, and
3% by transportation organizations. Some of these pages offer regional tourism
information such as festivals, events, and the start of the autumn foliage season.
However, these pages are basically used as a public relations tool of these facil-
ities and organizations. Other pages disseminate regional tourism information
rather than advertising the activities of organizations. The management of these
pages varies from companies and individuals to volunteer groups. Twenty-one
percent of the pages (178 pages) focus on information about particular regions;
3% focus on information about specific topics such as one-day hot spring trips,
famous places for flowers, and Buddhist statues; and 6% focus on information
about foreign countries. In this study, we targeted the 178 pages focusing on
information about particular regions.
2.2 Pages Offering Regional Tourism Information
(1) Number of Fans

The users clicking the “Like” button of a Facebook Page are registered as fans of
the page. The postings on the page appear in the News Feed of the user, and the
user can browse the various types of information on a Facebook Page on which
they are registered as fans. The number of users who have clicked the “Like”
button is “the number of fans”. Figure 2 shows the number of fans of the 178
Facebook Pages offering regional tourism information. The number of fans of
the top-ranking page exceeds 70,000. There are 19 pages with more than 10,000
fans. There are 143 pages with less than 5,000 fans, and 72 of the 143 pages with
less than 1,000 fans. For most of the pages, the number of fans is limited and the
number of users browsing the pages offering regional tourism information is low,
suggesting that the effect of these pages on the promotion of regional tourism
industries is limited.
Fig. 2. Number of fans of the 178 Facebook Pages offering regional tourism information
(2) Management
Figure 3 shows the classification of the management of the 178 pages offering
regional tourism information. Among the pages, 36, 24, and 9% of the Face-
book Pages are managed by individuals, private companies, and local govern-
ments, respectively. Others include regional tourist associations in the form of
general incorporated associations, incorporated nonprofit organizations (NPOs),
and volunteer groups.
Figure 4 shows the classification of the management of Facebook Pages with
more than 10,000 fans among the 178 pages. Considering all the Facebook Pages
offering regional tourism information, more than 30% of the pages are managed
by individuals, as shown in Fig. 3; however, only 11% (2 pages) of Facebook Pages
with more than 10,000 fans are managed by individuals, as shown in Fig. 4.
An organization that manages one of the Facebook Pages with more than
40,000 fans was interviewed on 30 October 2015. The management consisted of
a total of nine members including one young staff member and eight people who
provide regional information. They have adopted the management rule that the
frequency of posting is one or two posts per day. The interviewee said that finding
Fig. 3. Management of Facebook Fig. 4. Management of Facebook

Pages offering regional tourism infor- Pages with more than 10,000 fans
mation
new material every day is troublesome work even though there are nine members.
When a Facebook Page is managed by individuals, finding raw material and
collecting information require much time and expense, leading to difficulty in
posting at a high frequency.
(3) Facebook Pages with High Attractiveness
Table 1 shows a summary of 19 Facebook Pages with more than 10,000 fans.
Facebook Pages for various areas from Hokkaido to Okinawa attract many fans,
although 2 pages on Yokohama are also included in the list. For these 19 Face-
book Pages, posts, the presence or absence of photo and video images, the pres-
ence or absence of links to other homepages, the number of “Likes” for each
post, the number of “Shares” of each post, and the number of user comments
were obtained between January 1 and December 31, 2015, by Graph API.
Engagement rate is an index for evaluating the attractiveness of pages on
SNSs. The conventional engagement rate of Facebook Pages is given by
Conventional engagement rate = (Number of “Likes + Number of

comments + Number of “Shares of (1)
the post)/Number of fans.
However, conventional engagement rate defined in this way decreases as the

number of fans increases [2]. Therefore, Facebook, Inc. defined a new engagement
rate as
New engagement rate = (Number of “Likes + Number of comments
+ Number of “Shares + Number of people who
have clicked on the post)/Number of reaches.
(2)
This was released in January 2015 or later but the exact date of introduction
was not publicized. The new engagement rate cannot be calculated by third
parties, other than the management, because the reach count and click count of
the page are used in the calculation. Therefore, in this study, the conventional
engagement rate calculated using Eq. (1) was used to determine the engagement
rate per fan in Table 1. According to the Facebook Engagement Survey 2014
[2], the mean engagement rate of Facebook Pages with ≤ 10, 000 and < 50, 000
fans is 1.996% and that of Facebook Pages with ≤ 50, 000 and < 100, 000 fans
is 1.499%. The mean engagement rate of Facebook Pages in the travel/leisure
category is 1.56% (mean number of fans: 30,642).
The number of Facebook Pages with engagement rates of < 1%, ≥ 1 and
< 2%, ≥ 2 and < 3%, ≥ 3 and < 4%, and ≥ 8 and < 9% per fan is 5, 2, 4,
7, and 1, respectively. The Facebook Page on Shirakawa-go accepts posts from
users, and the number of posts is extremely high. The engagement rate per post
is small, as is the engagement rate per fan.
In this study, both the engagement rate per fan and the number of fans are
used as indices of attractiveness. Three Facebook Pages with an engagement rate
of ≥ 3% and a large number of fans, namely, the Facebook Pages of “Kyushu
Tourism Information”, “Akita Vision”, and “Yokohama Tourism Information”,
were selected from Table 1 as Facebook Pages with high attractiveness to be
analyzed.
3 Analysis of User Comments
In this section, the user comments on the pages with high attractiveness were
classified into ten types to clarify the feelings that cause users to post comments.
3.1 Analysis Method
For the Facebook Pages of “Kyushu Tourism Information”, “Akita Vision”, and
“Yokohama Tourism Information”, user comments collected over a period of one
year (2015) were analyzed to avoid seasonal bias.
First, engagement rates of all the posts for the three pages in 2015 were calcu-
lated. We read all the user comments for the posts with the top 100 engagement
rates (hereafter, user comments for the top 100 posts) and the user comments
for the posts with the lowest 100 engagement rates (hereafter, user comments for
the lowest 100 posts) and classified the user comments in accordance with the
feelings that caused users to post comments. For this classification, ten feelings
in the Dictionaryon Expression of Feeling [7]; namely, “joy (pleasure)”, “anger
(unpleasantness)”, “sadness (crying)”, “fear (apprehension)”, “shame (humil-
iation)”, “liking (longing, nostalgia)”, “disgust (loathing, regret)”, “sensation
(inspiring, excitement)”, “ease (calm)”, and “surprise (amazement)”. Table 2
shows examples of user comments classified into these ten feelings. Examples
1–9 are classified as one of the ten feelings. When multiple feelings are included
in one comment, as in example 10, the user comment is classified as multiple
feelings. Comments including the intension of “share” in the sentence are classi-
fied as “share”. The comments that do not belong to any of the ten feelings are
classified as “others”. Stamps without any comments are classified as “stamp”.
Table 1. Facebook Pages with more than 10,000 fans
Rank Facebook page Management Number Engagements number Engagement rate Number of
name of fans posts between
Jan. 1-Dec. 31
2015
“Likes” “Shares” Comments Per post Per fan
1 Kyushu General 70,105 856,559 38,515 5,493 2,427 3.5% 371
Tourism incorporated
Information association
2 Shirakawa-go Private 67,049 1,125,034 44,538 6,147 450 0.7% 2,612
company
3 Akita Vision Local 44,780 621,495 26,001 4,637 1,513 3.4% 431
government
4 The Heart of Local 43,673 27,622 553 151 44 0.1% 651
Osaka-Visit government
Osaka Japan
5 Yokohama Public interest 42,655 651,163 23,136 3,370 1,433 3.4% 473
Tourism incorporated
Information foundation
6 Yokohama Cooperative 32,748 149,274 4,468 1,054 499 1.5% 310
China Town association
7 Gooood Place!! Private 32,566 236,962 7,080 1,297 1,141 3.5% 215
in Shiga company
8 Otaru Fan Private 22,830 31,289 27 9 870 3.8% 36
company
9 Exchange of Private 18,592 99,287 5,473 436 408 2.2% 258
Kyoto company
Information
-Kyoto Now-
10 Hyogo Tourism Local 16,470 18,832 529 115 81 0.5% 239
Guide government
11 Okinawa Diving Private 13,870 1,520 18 13 111 0.8% 14
company
12 Okayama Great Public interest 12,547 76,720 3,081 721 323 2.6% 249
Spot Net incorporated
association
13 Fukuoka No Individual 12,472 4,241 51 28 103 0.8% 42
Machi
14 I LOVE Individual 12,267 62,946 3,432 797 1,050 8.6% 64
TOKUSHIMA
15 Hokkaido Fan Private 12,053 179,615 10,050 1,129 305 2.5% 625
Magazine company
16 Let’s Visit Nara Private 11,857 40,560 1,590 157 188 1.6% 225
company
17 Kamakura Private 11,275 58,557 1,301 671 348 3.1% 174
Block company
18 Web magazine Private 11,026 58,428 2,468 362 286 2.6% 214
[Shikoku company
Tairiku]
19 Nagasaki Local 10,657 58,284 2,647 384 383 3.6% 160
Tourism government
Promotion
Section
3.2 Analysis Results

Figures 5 and 6 show the classifications of feelings underlying the user comments
for the top 100 posts and those underlying the user comments for the lowest 100
posts, respectively. The total number of user comments for the top 100 posts
classified as one of the feelings, “share”, “stamp”, or “others” is 6,924. The
percentage of “liking (longing, nostalgia)” was the highest (25.9%), followed by
“sensation (inspiring, excitement)” (17.9%). The total number of user comments
for the lowest 100 posts classified as one of the feelings, “share”, “stamp”, or
Table 2. Examples of classifications of feeling of user comments
User comment Classification

∧∧
Example 1 I will go there next week ( ). I can’t wait Joy
Example 2 It’s a waste of electricity Anger
Example 3 I can’t visit to take photos because of work this Sadness
year (T T)
Example 4 I’m worried about fainting because of the high Fear
temperature
Example 5 I love Yokohama! Liking
Example 6 Oh, my god. I can’t take holidays during silver Disgust
week because of work
Example 7 It’s a beautiful night scene!!! Sensation
Example 8 I feel relaxed. I’ll enjoy the night scene when I go Ease
back to my hometown
Example 9 5000, wonderful! Surprise
Example 10 It’s really beautiful!! I’d like to visit there and Sensation, Liking
ride it!
Example 11 I will share the page “Share”
Example 12 Can I get it in Tokyo? “Others”
“others” is 1,200. The percentage of “liking (longing, nostalgia)” was the highest
(29.8%), followed by “share” (11.3%) and “sensation” (8.2%). There are no user
comments for either the top or the lowest 100 posts classified as “shame”.
The numbers of the user comments for the top and lowest 100 posts in each
feeling classification were compared. As shown in Table 3, we developed a 2 × 2
contingency table to summarize the number of user comments classified as one
of the feelings and the total number of user comments classified as other feelings,
“share”, “stamp”, and “others”, to carry out Fisher’s exact test. A contingency
table was developed for each feeling in the test. Table 4 shows the summary
of the data and results. A significant difference was observed between the user
comments for the top and lowest 100 posts for six feelings, “share”, “stamp”,
and “others”. The underlines in Table 4 indicate which of the top or lowest 100
posts has a higher number of user comments than the expected value for the
classification with significant difference.
3.3 Tendency of User Comments Classified as Each Feeling
The number of user comments for the top and lowest 100 posts classified as
“joy” is small. Although a significant difference was observed between the two,
the content of comments is an expression of gratitude, such as “thank you for the
information on xxx”, “I’m looking forward to attending xxx” for the announce-
ment of an event, and “I’m looking forward to eating xxx” for seasonal food.
Fig. 5. Classification of feelings of user Fig. 6. Classification of feelings of user

comments for the top 100 posts. comments for the lowest 100 posts.
Table 3. Contingency table summarizing the number of particular feelings for the user
comments for the top and lowest 100 posts.
Top Lowest
Feeling X a b
Feelings other than feeling X c d
a: The number of user comments for the top
100 posts classified as feeling X
b: The number of user comments for the lowest
100 posts classified as feeling X
c: The number of user comments for the top
100 posts classified as feeling other than X
d: The number of user comments for the lowest
100 posts classified as feeling other than X
The number of user comments for the top and lowest 100 posts classified
as “anger” is small. Although a significant difference was observed between the
two, there are few user comments expressing anger and unpleasantness.
The number of user comments for the lowest 100 post classified as “sadness”
is small. For the top 100 posts, most of the user comments classified as “sadness”
express sadness that they cannot visit a place because of work despite a desire
to do so or that photos of their hometown cause a longing to return.
The number of user comments for the lowest 100 posts classified as “fear”
is small. Examples of the user comments for the top 100 posts include concern
regarding going to a place introduced in the post and the fear of unconventional
designs and ideas in the post.
There are no user comments classified as “shame” for either the top or the
lowest 100 posts.
The number of user comments classified as “liking” is the largest among the
user comments for both the top and lowest 100 posts. Although a significant
difference was observed between the two, the content of comments is an expres-
sion of liking, such as “I like xxx hot spring” and “I love xxx!”, the expression
Table 4. Number of comments for the top and lowest 100 posts classified into each
feeling category and the results of Fisher’s exact test
Feeling Top Lowest p value of Fisher’s exact test

Joy 211 54 0.013∗
Anger 5 6 0.002∗
Sadness 78 6 0.045∗
Fear 59 10 1.000
Liking 1793 357 0.006∗
Disgust 133 36 0.021∗
Sensation 1237 98 0.000∗
Ease 35 8 0.515
Surprise 165 39 0.088
“Share” 533 135 0.000∗
“Stamp” 62 79 0.000∗
“Others” 2613 372 0.000∗
∗
Fisher’s exact test: p < 0.05.
The underlines indicate which of the top or lowest 100 posts
has a greater number of user comments than the expected
value. Also, a significant difference was observed between the
user comments for the top and lowest 100 posts.
of longing, such as “I would like to visit xxx once in the future”, and the feeling
of nostalgia because of overlapping old memories with the post. People often
express liking when they read posts related to their past experience and knowl-
edge, such as a previously eaten food, a place they would like to visit in the
future, and memories of childhood.
The user comments for the top and lowest 100 post classified as “disgust”
mostly express feelings of regret, such as “I would like to go but I cannot”. Only
a few user comments express undesirable and unfavorable feelings.
The number of user comments for the top 100 posts classified as “sensation”
tended to be large, and a significant difference was observed between users com-
ments for the top and lowest 100 posts. Among the user comments for the top
100 posts, the number of user comments classified as “sensation” is the second
most numerous after those classified as “liking”. Users are moved by photos of
beautiful scenes and excited by photos of delicious-looking food for both cate-
gories of posts.
“ease” is small. Users feel relief or healed upon viewing the photos posted.
Regarding “surprise”, for user comments on the top 100 posts, users are
surprised at the scale and impact of photos or amazed by the beauty of photos.
User comments for the lowest 100 posts express surprise because they found
unexpected content in the post, such as “I never imagined that such a place
existed there”.
Some users making comments classified as “share” feel empathy with the
post and express their intent to share the post. However, users simply click the
share button (a simple comment expressing the intension as is often the case). A
significant difference was observed between the user comments for the top and
lowest 100 posts. The number of user comments classified as “share” is greater
for the lowest 100 posts. We request that users related to the management and
the fans of the Facebook Page share posts regardless of their feelings toward the
contents of posts.
“stamp” is small. A significant difference was observed between the two, and the
number of user comments for the lowest 100 posts classified as “stamp” tended
to be larger.
The number of user comments classified as “others” is largest for both the
top and lowest 100 posts. Examples of user comments include questions, such as
example 12 in Table 2, reports of past experience, such as “I have visited there”
and “I ate it yesterday”, and expressions of intent and plans, such as “I will visit
there next” and “I will go to eat there again this year”. We suggest that the
comments reporting past experiences are triggered by the feeling of liking the
post. The comments expressing intent and plans are possibly triggered by the
feeling of sensation (users are moved or excited about the post) and the feeling
of joy. However, we cannot detect the true feeling of users, and we classified user
comments that do not include the direct expression of feelings as “others” in
this study.
4 Discussion
In this section, directions for increasing the attractiveness of Facebook Pages for
those involved in the promotion of regional tourism industries are discussed by
referring to actual examples of user comments for which a large number of user
comments expressed the feelings of users.
4.1 User Comments for the Top 100 Posts Classified as Liking
Example 1 is a post related to seasonal fish. Users report that the fish is deli-
cious and describe the characteristic taste of the fish during breeding season.
Example 2 is a post reporting the arrival of the season for tourist boats with
kotatsu, a small table having an electric heater underneath and covered by a
quilt. Comments 2-1 and 2-2 indicate that users have enjoyed this type of boat
before, recommend it to others, and express their intent to experience it again.
The user of comment 2-3 had experienced the boat without kotatsu and expresses
a willingness to experience the boat with kotatsu this season.
Among the user comments for the top 100 posts, those classified as “liking”
are the most numerous. To increase the attractiveness of Facebook Pages for
tourist spots and food, it is indispensable to offer posts of famous tourist spots
and food that many users like, want to visit, and experience. For well-known
information (tourism resources), timely posts considering the best season to

enjoy the resources will contribute to the improvement of the engagement rate
through links between the information and the lifestyle and memories of users.
Example 1. Seasonal fish
• Comment 1-1 (“liking”) I ate that fish. It was delicious!

• Comment 1-2 (“liking”) I’ve eaten this fish! It had many eggs and was deli-
cious.
Example 2. Cruise on a boat with kotatsu
• Comment 2-1 (“liking”) The boat with kotatsu is only in winter and was
quaint. It was really nice (∧ ∧ )/!
• Comment 2-2 (“liking”) We ate steamed eel after enjoying a boat with
kotatsu!! I’d like to visit again.
• Comment 2-3 (“liking”) I visited there about 5 years ago because it’s my
husband’s hometown(∧ o∧ )/. It was December, but the boat didn’t have a
kotatsu. I’d like to try it when I visit again(o∧∧ o).
4.2 User Comments for the Top 100 Posts Classified as Sensation
Example 3 is a post related to the illumination of the buildings and gardens

of temples and shrines of a certain region. Users posted comments 3-1 and 3-2
because they were moved by the beautiful photos, although it is not clear whether
the users actually had visited the place or not. The user posting comment 3-3
was moved by the beautiful photos and posted a comment although the user had
not visited the place.
The user comments classified as “liking” were mostly posted in response
to objects that users already knew, whereas the user comments classified as
“sensation” were posted by users who were moved or excited by the photos
regardless of their previous knowledge. This finding suggests that the content
and quality of the attached photos are important and that posting of photos
that may induce sensation and excitement in users is effective when posting
information (tourism resources) on sites that people may not be aware of.
Example 3. Illumination of buildings and gardens of temples and shrines
• Comment 3-1 (“sensation”) It’s really beautiful!

• Comment 3-2 (“sensation”) The illumination of a temple in Fukuoka prefec-
ture is beautiful.
• Comment 3-3 (“sensation”) It’s wonderful. I’ve been there once before, but I
didn’t know that it was such a beautiful and wonderful place.
4.3 User Comments for the Lowest 100 Posts Classified as Liking
Example 4 is a post related to local sweets. From comments 4-1 and 4-2, we can
see that users feel nostalgia for their favorite sweets. In contrast, comments 4-3
and 4-4 indicate that the sweets are not famous; only a small number of people
know about the sweets. Example 5 is a post related to a local restaurant. From
the post and comments 5-1 and 5-2, it is clear that the users know the restaurant
introduced in the post and agree on its attractiveness. From comments 5-3 and
5-4, it is clear that the users would like to visit the restaurant and eat the dishes
described in the post.
Among the user comments for the lowest 100 posts, the number of comments
classified as “liking” was the largest (357). The users who have eaten the sweets
or visited the restaurant before, as shown in Examples 4 and 5, are thought to
feel empathy for the posts and make comments because of a feeling of nostalgia
and the memory of the food and service, although users who did not know
the contents of the posts showed no interest in and did not respond to the
posts. In addition, users related to the region are interested in the posts because
information new to them is provided in the posts. This finding suggests that the
posts about places and foods familiar to those from the region or to the local
people, even though not famous, can attract a certain amount of responses from
users, leading to an increase in the attractiveness of Facebook Pages.
Example 4. Local sweets

• Comment 4-1 (“liking”) I love the sweets! This reminds me of the sweets
made by my mother.
• Comment 4-2 (“liking”) It brings back old memories(∗∧ ∧ ∗). I want to eat it
again!
• Comment 4-3 (“others”) I grew up in Akita, but don’t know these sweets.
• Comment 4-4 (“others”) I don’t know them, either.
Example 5. Local restaurant

• Comment 5-1 (“liking”) I’ve frequently visited this restaurant! The ham-
burger, rice omelet, and demiglace sauce there are delicious!
• Comment 5-2 (“liking”) I loved the dishes (≥∇≤)
• Comment 5-3 (“liking”) They look delicious. I want to eat them.
• Comment 5-4 (“liking”) I’d like to visit this restaurant!
5 Conclusion
In this study, we carried out an analysis of the feelings expressed in user com-
ments on Facebook Pages with many fans offering regional tourism information,
and clarified the feelings of the users who posted the comments and the factors
behind the increase in the attractiveness of these Facebook Pages.
Compared with websites, Facebook Pages can be easily started and have
no startup costs. Therefore, the number of pages managed by individuals is
large. When focusing on Facebook Pages with many fans, the percentage of
Facebook Pages managed by individuals is small; the Facebook Pages managed
by organizations such as regional tourist associations and companies offering
regional tourism information are the mainstream. However, such pages are not
always actively accessed by many visitors. The information explained in this
paper may be useful for increasing the attractiveness of Facebook Pages.
The engagement rate of the Facebook Page “I Love Tokushima” (Table 1)
is extremely high (8.6%). The number of posts per year is 64, which means
an annual posting frequency of 1.2. According to Locowise Ltd. (the UK), the
engagement rate of Facebook Pages with 1-4 posts per week is the highest. The
survey by Locowise Ltd. targeted all Facebook Pages, not limited to those of the
tourism industry. In future studies, we will investigate the relationship between
the frequency of posting and the engagement rate of Facebook Pages on tourism
to determine the optimal frequency of posting.
Acknowledgments. This study was supported by Japan Society for the Promotion
of Science (JSPS) KAKENHI (Grant Number 15K00476).
References
1. Facebook (2017) Statistics of facebook. http://newsroom.fb.com/company-info/
2. Facenavi (2017) Facebook engagement survey 2014. http://facebook.boo.jp/
facebook-engagement-survey-2014
3. Inc Comnico (2014) Getting Fans! Know-How of Posting on Facebook. Shoeisha
4. Instagram (2017) Statistics of instagram. https://www.instagram.com/press/
5. Ministry of Internal Affairs and Communications (2017) 2015 white paper on infor-
mation and communications in japan. http://www.soumu.go.jp/johotsusintokei/
whitepaper/ja/h27/pdf/
6. Line (2017) January-march 2016 business report. https://linecorp.com/ja/pr/
news/ja/2016/1347
7. Nakamura A (1993) Dictionary on Expression of Feelings. Books Tokyodo, Tokyo
8. Ohara S, Kajiyama T, Ouchi N (2015) Extraction of characteristics of photos to
increase the engagement rate of facebook pages of companies analysis of photos of
products of fast-food companies. IEICE Trans A J98–A(1):41–50
9. Sabate F, Berbegal-Mirabent J, Cañabate A et al (2014) Factors influencing pop-
ularity of branded content in facebook fan pages. Eur Manag J 32(6):1001–1011
10. Saito S (2014) Facebook Marketing [Business Technique] Technique of Obtaining
Valuable “Like”. Shoeisha
11. Sawada A, Yoshida T, Murakami K (2016) Factors behind increasing attractiveness
of tourism-related facebook pages. In: Proceedings for the 13th Workshop of the
Society for Tourism Informatics, pp 17–20
12. Twitter (2017) Twitter usage/company facts. https://about.twitter.com/ja/
company
13. Wilson RE, Gosling SD, Graham LT (2012) A review of facebook research in the
social sciences. Perspect Psychol Sci 7(3):203–220
Meta-analysis of the Factors Influencing
the Employees’ Creative Performance
Yang Xu1 , Ying Li1(B) , Hari Nugroho2 , John Thomas Delaney3 , and Ping Luo1
1
liyinggs@scu.edu.cn
2
Department of Sociology, Gedung F. Faculty of Social and Political Sciences,
Universitas Indonesia, Depok 16424, Indonesia
3
Kogod School of Business, American University, 4400 Massachusetts Ave,
Washington, DC 20016-8044, USA
Abstract. Innovation has become a popular topic in society. The cre-

ative performance fluctuates with the employees’ mood caused by the
working environment. Many factors have been suggested to be affecting
the employees’ creative performance. The paper answered the questions
as “What are the factors that affect the employees’ creative performance
through employee behavior? What is the proportion (weight) of each of
these determinants?” Furthermore, this study resolves this problem via
the meta-analysis, which refers to do systematic quantitative analysis
by former studies to summarize the research results of creative perfor-
mance over the last ten years, and to ascertain the factors influencing
the creative performance as well as to determine the correlation between
them. Suggestion having been proposed including the further investi-
gation of the interaction between the factors impacting the employees’
creative performance as well as explore the corresponding effects. And
it is necessary to note that changing one factor may eliminate the effect
of several factors, which may cause negative change in the staff creative
performance and there is joint efforts needed in terms of enhancing the
employees’ creative performance.
Keywords: Factors · Creative performance · Meta-analysis
1 Introduction
Creative performance refers to the products, processes, methods, and ideas cre-
ated by the employees, which are novel, practical, and valuable to an organiza-
tion. Although creativity varying from person to person cannot be generalized,
it is not fixed. In other words, creativity is constantly fluctuating. Likewise,
the creative performance fluctuates with the employees’ mood caused by the
working environment. What are the factors affecting the employees’ creative
performance? Do they have positive or negative impacts on the staff? Which of
these factors has more profound influence and which has less? These issues are
Meta-analysis of the Factors Influencing the Employees’ Creative 659
considered by both business managers and numerous scholars. By surveying past

scholarship on creative performance, the paper finds that there are similarities
and differences between the researches. Specifically, some examine the impact
of internal factors on the employees’ creative performance; some explore the
external factors; some investigate the employees’ creative performance affected
by both internal and external determinants, while others study several other
influencing factors. All these researches, however, are not comprehensive. What
are the factors that affect the employees’ creative performance through employee
behavior? What is the proportion (weight) of each of these determinants? The
paper resolves this problem via the meta-analysis, hoping to summarize the
research results of creative performance over the last ten years, and to ascer-
tain the factors influencing the creative performance as well as to determine the
correlation between them.
1.1 Related Concepts
Creative performance refers to the products, processes, methods and ideas cre-
ated by the employees, which are novel, practical, and valuable to an orga-
nization. Work passion, a core characteristic of self-identification, refers an
employee’s strong love for a certain occupation, which he or she considers vital
and into which this individual tends to invest time and effort.
Autonomous orientation refers to an individual’s desire to act, based on his
or her interests and self-values. More precisely, this individual considers incidents
as opportunities and challenges, and he or she will take active actions to seize
these opportunities and to shoulder the responsibilities, showing strong internal
and extrinsic motivations.
Control orientation refers to an individual’s desire to act, based on other
individuals’ attitudes rather than his or her own ideas, because he or she is
vulnerable to the external environment.
Mastery approach goal orientation refers to the development process in which
an individual tends to concentrate on the abilities, learning, and tasks. Per-
formance approach goal orientation refers to an individual’s emphasis of the
performance-related matters.
Performance avoidance goal orientation refers to an individual’s avoidance of
the performance-related matters.
Cognitive re-evaluation refers to an individual’s control of the emotional
response through changing the impact of events on personal awareness.
Inhibition of expression refers to an individual’s control of the emotional
response by suppressing the emotional expression.
Feedback consistency refers to the consistency between an employee’s expec-
tation of the boss’s feedback on his or her work performance and the actual
feedback this individual receives.
660 Y. Xu et al.
1.2 Literature Review

(1) Overview of Foreign Researches
As early as the 20th century, the academic circle started researching on the
creativity and creative performance. Analyzing the impact of leadership style
and support on the employees’ creative performance through empirical studies,
Huang et al. [4], argued that charismatic and supportive leaders will facilitate
the creative performance. Based on the investigation of the relationship between
the employers and the employees, their research results provide references for
the managers. The above researchers, however, do not consider the impacts of
the employees’ personal factors on the creative performance. Ulger [7] examined
the impact of employees’ participation in the creative processes and the pleasure
and challenge of creative tasks on their creative performance through empirical
researches. They claim that the more staff participation and the more intriguing
and challenging the task, the better the creative performance will be. Moreover,
the time of the staff participation process, during which multiple factors may
contribute to the change of staff creativity, will determine the level of the final
creative performance. Nevertheless, these scholars leave the second point for
future research. Chong and Ma [1] explored the influence of the personal factors
of employees, such as motivation, mood and demand, on the creative perfor-
mance via empirical researches, he suggest that the motivations and demands of
employees generally have positive impacts on their creative performance. Fur-
thermore, in terms of mood, the negative emotions have an adverse effect on
their creative performance and vice versa. Their researches assist readers to
understand what affects the employees’ creative performance from the internal
perspective, but they fail to consider the external factor.
(2) Overview of Domestic Researches
Likewise, being interested in this topic, Chinese researchers conduct a host
of studies. Using the empirical research, Zhang [8], Zhang et al. [9], Zhang et
al. [10], Tang et al. [6], examine the impacts of organizational environment (for
instance, working time and pressure, organizational situation, and leadership)
on the employees’ creative performance. They conclude that the staff creativity
is positively correlated with the working time and pressure as well as leadership
styles; the influence of organizational situations varies from person to person.
Song [5], Zhang et al. [9], Dong et al. [2], analyzing the impacts of the employees’
emotions on their creative performance, concur that the negative emotions have
adverse impacts on creative performance, and vice versa. Due to other factors in
these studies, their research data differ moderately. Similar to foreign scholars,
the above researchers only investigate one aspect that affects the staff creativ-
ity. Gong et al. [3], exploring the influence of employees’ motivations on their
creative performance via the empirical research and scenario analysis, ascertain
that the creative performance is inhibited by the workers’ extrinsic motivations
and control motivations but promoted by the employees’ other motives. Their
final results differ because of various classifications of the motivations.
Overall, Chinese and foreign scholars have investigated numerous factors
affecting the creative performance, and their conclusions, which differ slightly
in each essay, are roughly the same. In terms of research methods, these papers,
however, principally adopt empirical research and scenario simulation. As for
the investigations of the influencing factors, they largely highlight one of them.
Based on the literature above, this paper utilizes a different approach– the meta-
analysis, seeking to examine the factors that impact the employees’ creative
performance holistically.
1.3 Introduction to the Meta-analysis
(1) Definition and Characteristics of the Meta-analysis

The Meta-analysis is widely used in disciplines, such as medicine, psychology,
economics, and management. In these fields, numerous scholars, for the most
part, study the same problem, but the final results may be different due to the
process control, error, and sample selections. Therefore, they fail to reach a clear
conclusion. It is of necessity and significance to analyze the results achieved by
other scholars through the Meta-analysis.
Meta-analysis was first proposed by American scholar Glass when he gave a
speech at the American Educational Research Association in 1976. According to
Glass, meta-analysis refers to a statistical analysis that combines the results of
multiple individual studied to resolve uncertainty when reports disagree.
Nonetheless, there is no unified definition of the Meta-analysis. This paper
adopts the definition from the book titled Conducting Meta-Analysis through
Stata: Meta-analysis is a statistical approach that examines the results of the
individual studies of the same topic, analyzing the similarities and the sources
of difference between these research outcomes. Meta-analysis and systematic
evaluation are different concepts, even though there are similarities between
them. Systematic evaluation refers to a rigorous evaluation and review of an
issue via a certain search strategy, in the process of which the Meta-analysis
may or may not be used.
There are some salient features (advantages) of the Meta-analysis, compared
with the individual study: (i) to improve the statistical efficacy; (ii) to evalu-
ate the results quantitatively and comprehensively, resolving uncertainty when
reports disagree; (iii) to evaluate the hypothesis, discovering the deficiencies of
the previous researches; (iv) to analyze numerous literature data simultaneously,
without being affected by the number of references; (v) Using the individual
research results as the original data, Meta-analysis is a higher level of logical
analysis.
(2) Advantages and Disadvantages of the Meta-analysis
Each research method has its advantages and limitations, and the paper
presents a brief introduction to the merits and demerits of the Meta-analysis.
The Meta-analysis is used extensively because of four advantages. First, the
Meta-analysis is more accurate than the significance test in terms of the esti-
mation and reliability analysis of the effect values. Second, the scholars may
discover the shortcomings of the individual researches via the Meta-analysis,
thereby avoiding the flaws. Third, the results of individual studies are aggregated
662 Y. Xu et al.
through the Meta-analysis; hence, more objective explanations are conducted,

and more reliable and accurate conclusions are researched. Fourth, new research
directions and suggestions will be put forward after the Meta-analysis, which
will provide references for further researches.
Nevertheless, like other methods, there are some flaws in the Meta-analysis.
First, it takes a long period of time to conduct the Meta-analysis, due to its
requirement for a large amount of literature. Thus, the workload is heavy. Second,
it is difficult to find all the literature on a particular issue; hence, researchers
cannot conduct a comprehensive statistical analysis. Last, it is difficult to ensure
that the thinking and structure of the existing studies are consistent, that the
qualities of these essays are high, and that the data is complete. Therefore,
not all the documents are useful, and the amount of the literature may vary
significantly among different research groups.
2 Research Method
2.1 Literature Retrieval
Although creative performance is a relatively new topic, scholars at home and

around the world have conducted numerous researches. The paper firstly utilizes
the keyword retrieval, the main method of literature retrieval, to search articles
on the creative performance, employee behavior, and the factors affecting the
employees’ creative performance. Secondly, according to the references of the
retrieved essays, the paper searches the related documents in each database,
including the doctoral dissertations and journal articles published after 2010.
The literature retrieval is conducted at the library of Sichuan University. Both
Chinese databases (such as the CNKI, Wan) and foreign databases (for instance,
Science-Direct, and Springer Link) are used in this research paper. Manual search
is not required, since electronic magazines are available in the information age,
and the databases above provide all the materials needed in this study.
2.2 Inclusion and Exclusion Criteria for Literature
(1) Literature Inclusion Criteria

First, the literature should be published or unpublished between 2000 to
2015.
Second, the literature should be associated with creative performance.
Third, the data should be relatively complete, including the values of corre-
lation coefficients, such as T, F, and P. The articles may not contain all these
data, but they should provide sufficient figures that are needed in this research.
Fourth, the research object should be explicit, describing the sample size,
age, and gender.
Fifth, the factors for creative performance should be quantitatively deter-
mined.
Sixth, the questions listed in the scales should be expressed clearly.
Table 1. An overview of the selected literature
Number Author Year Influencing factor Research Effective

object sample size
1 Zhang Jian 2013 The employees’ emotions, Employees 222
et al. orientations, etc.
2 Zhang Jian 2013 Time pressure, creative Freshmen 60
et al. personality, etc.
3 Dong Li et 2012 The employees’ emotions, Freshmen 273
al. orientations, etc.
4 Zhang Jian 2010 Independent motivation, control Employees 150
et al. motivation, etc.
5 Tang 2012 Leadership style, team Researchers 209
ChaoYing recognition, etc.
et al.
6 Song Yahui 2015 Passion for work Employees 162
et al.
7 Gong 2015 Feedback, goal and personality, Employees 159
Zhenxing et etc.
al.
8 Wang 2013 Time pressure College 300
Haocheng students
et al.
9 Li Xiulin et 2012 The employees’ emotions College 91
al. students
10 Zhang Jian 2003 Working environment, motivation, Employees, 459
needs, etc. graduates
11 Gu Rui 2015 Leadership style, etc. Health care 101
et al. workers
12 Pamela 2004 The employees’ creativity, Managers 12
Tierney leadership support, working
experience, job complexity, etc.
13 Bartol 2010 Participation in the creative Employees 86
Kathryn process, etc.
14 Kwok 2014 Work motivation, etc. Employees 112
Leung
15 Yu-Shan 2015 Environmental factors College 160
Chan students
16 Kani Ulge 2016 Creative thinking, etc. College 216
students
17 Dong gun 2016 The employees’ motivations and College 143
An personality, etc. students
18 Jin Nam 2004 Personal factors, organizational College 430
Choi environment, etc. students
19 Jan Kratzer 2004 Team communication, team Employees 243
recognition, etc.
20 Eric Chong 2010 Individual factors, supervision, Employees 350
and working environment, etc.
21 Wasilu 2013 Working attitude, etc. Employees 70
Suleiman
22 L. Olsson, S 2012 The quantity of managers, Managers, 137
leadership style, etc. employees
664 Y. Xu et al.
(2) Literature Exclusion Criteria

First, the articles largely about the literature review of this topic should be
excluded.
Second, repeated reports should be excluded.
Third, the literature without relevant correlation coefficients or data should
be excluded.
Fourth, the literature without explicit scale measurement questions should be
excluded, since these articles cannot appropriately describe the research objects.
(3) An Overview of the Selected Literature
In the initial screening process of the Meta-analysis by using Stata, all lit-
erature adopts an empirical research with available data. According to different
dependent variables, the selected articles are grouped into six categories: the
employees’ emotions (3 articles), time pressure (2 articles), creative personal-
ity (4 articles), work motivation (6 articles), leadership and teamwork (6 arti-
cles), and feedback consistency (2 articles). This paper examines which factor
affects the employees’ creative performance; hence, there is no causal relation-
ship between these documents. In other words, the data in each article should
be considered separately. After the searching and screening process, the selected
articles are shown in Table 1: Considering the large data used in the Meta analy-
sis via Stata (a software), it is inconvenient to list all of them in the table above.
In the calculation process, however, the paper, according to a calculation method
named the overall effect size, will select the complete and appropriate data from
these research articles as the study data.
3 Research Results
3.1 Description of the Research Data
(1) Overall Effect Value

First of all, according to the above-mentioned method, the paper calculates and
sorts the data extracted from the selected research articles. This study contains
a total of nine research articles, each of which contain multiple influencing fac-
tors. Hence, a total of 25 effect values are obtained. Next, the paper presents a
description of the selected research articles and effect values.
The selected articles with various effect values are divided into different types
according to the reference sources. Table 2 illustrates the type and quantity of
the research articles containing different effect values.
In Table 2, the “Journal” shows the effect value extracted from the journal
articles, the largest proportion of which is 80%. “Thesis” indicates the effect
value extracted from the doctoral dissertations and master’s theses, which is
8%. The effect value extracted from the research reports is 12%. Figure 1 clearly
illustrates these data.
Second, the paper calculates all effect values listed in the literature. This
study mainly considers the factors that affect the creative performance of
employees; therefore, all the determinants with data integrity mentioned in the
Table 2. The quantity of effect values from different article types
Frequency Percentage Effective percentage

Journal 20 80% 80%
Thesis 2 8% 8%
Research report 3 12% 12%
Total 25 100% 100%
Fig. 1. Distribution of the quantity of effect values from different article types
selected documents are included in the table, even though they may influence
each other. The specific data are shown in Table 3.
The distribution of the value of R and Z is clearly illustrated in Fig. 2, a line
graph. In terms of the value of R, the average is 0.35; the standard deviation
is 0.201; the maximum value is 0.71; the minimum value is −0.058; and the
skewness is −0.219, which demonstrates negative skewness. As for the value of
Z, the average is 0.62; the standard deviation is 0.410; the maximum value is
1.506; the minimum value is −0.086; and the skewness is 0.421, indicating pos-
itive skewness (See Table 4). Eventually, the paper, calculates the overall effect
value of all literature, which is 0.551, generally indicating the positive correla-
tion between the variables of the selected research articles and the employees’
creative performance.
3.2 Homogeneity Test
The methods for homogeneity test include Q test, H test, chi-square test, forest
plot. Using the Q test, the paper first calculates the Q value through the following
equations:
k 2
Q= wi(ri − R) , (1)
i=0

R= wiri/ wi, (2)
where k represents the number of the effect values of the selected literature; r is
the effect value; and w denotes the proportion of the effect value. The calculated
values of R and Q are 0.313575684 and 0.040220883, respectively.
666 Y. Xu et al.
Table 3. Related data for the selected literature
Number Factor R value Z value Sample size

1 Positive mood 0.228 0.301 495
2 Negative mood 0.054 0.082 495
3 Autonomous orientation 0.182 0.284 495
4 Control orientation 0.193 0.303 222
5 Features of time and pressure 0.33 0.543 60
6 Psychological satisfaction 0.419 0.718 60
7 Executive support 0.266 0.427 150
8 Need for autonomy 0.172 0.268 150
9 Capability requirement 0.243 0.387 150
10 Relationship requirement 0.277 0.447 150
11 Internal motivation 0.477 0.843 150
12 Motivation for identification 0.442 0.766 150
13 External motivation −0.058 −0.086 150
14 Control motivation −0.04 −0.06 150
15 Working motivation 0.395 0.669 150
16 Team performance 0.71 1.506 209
17 Team identity 0.509 0.917 209
18 Charismatic leadership 0.449 0.781 209
19 Concentration 0.585 1.11 162
20 Mastery approach goal orientation 0.485 0.861 162
21 Performance approach goal orientation 0.548 1.013 162
22 Performance avoidance goal orientation 0.348 0.577 162
23 Cognitive reevaluation 0.707 1.495 162
24 Inhibition of expression 0.431 0.743 162
25 Feedback consistency 0.33 0.543 159
Fig. 2. Distribution of related data for the selected literature

Table 4. Description of statistical data of the selected literature
N Min Max Mean Standard Skewness Kurtosis

deviation
Standard Standard
error error
R value 25 0 1 0.35 0.201 −0.219 0.464 −0.158 0.902
Z value 25 0 2 0.62 0.41 0.421 0.464 0.184 0.902
Valid N (list state) 25
The Meta results in Table 3 show that the 25 effect values are homogeneous.
The Q test follows the chi-square distribution with a degree of freedom K − 1;
hence, when the value of K is 25, the chi-squared value is 28.241 at a significance
level of 0.05. In other words, if Q = 0.040220883 < X0.02524
2
= 28.241, then the
selected research articles are homogeneous.
3.3 Bias Test
Bias test is considered essential in the meta-analysis. The publication bias refers
to the different opportunities for publication and corresponding impacts on the
results caused by the researchers, evaluators, editors’ preferences of the direction
and intensity of the research in the submission, acceptance, and publication
process of an article. According to the book Conducting Meta-Analysis through
Stata, there are multiple approaches to testing the publication bias, such as the
funnel plot, Egger linear regression method, and Begg rank correlation method.
This paper uses the funnel plot, which is the most commonly used visualized
method for a qualitative measurement of publication bias, first proposed by Light
et al., in 1984. Figure 3 shows the funnel plot derived from the Meta-analysis of
Table 3.
Fig. 3. The effect value of the selected literature

668 Y. Xu et al.
According to Fig. 3, there is no publication bias among the articles and theses
used in this study, since the funnel plot is relatively symmetric.
4 Conclusions and Suggestions
It can be seen from overall effect values in Table 3 that when wholly considering
the influencing factors (positive mood, negative mood, autonomous orientation,
control orientation, time pressure, psychological satisfaction, supervisor support,
autonomous need, relationship motive, internal motivation, external motivation,
control motivation, working motivation, team performance, team identity, charis-
matic leadership, concentration, mastery approach goal orientation, performance
approach goal orientation, performance avoidance goal orientation, cognitive re-
evaluation, inhibition of expression, feedback consistency), they are positively
correlated to the employees’ creative performance (R = 0.551). To enhance the
creative performance, the managers may improve the employees’ intrinsic moti-
vation through incentive policies or through the executives’ encouragement and
support. Every employee, for the most part, will be affected by these factors;
hence, this study, to some extent, may serve as a reference for the scholars or
managers. Nevertheless, due to the possible interaction between these factors,
their relationship with the employees’ creative performance may be weak and
accurate, which is the limitation of this paper. The author hopes that the future
scholars of this field may clearly analyze this interaction.
We proposed suggestion as following: First, the future researchers may inves-
tigate the interaction between the factors impacting the employees’ creative per-
formance as well as explore the corresponding effects. The reason is because
adjusting or changing one of the factors may indirectly influence other deter-
minants, which contributes to a great change in the creative performance of
the employees. Meanwhile, changing one factor may eliminate the effect of sev-
eral factors, which may cause negative change in the staff creative performance.
Second, the future researchers may examine the proportion of the effect of the
individual factor on the creative performance. It assists managers to prioritize
incentives so that the enterprise may maximize the profits at the lowest cost.
Overall, joint efforts are needed in terms of enhancing the employees’ creative
performance.
Acknowledgements. This research was supported by Sichuan Social Science Plan-

ning Program Key Project (Grant No. SC16A006), System Science and Enterprise
Development Research Center Project (Grant No. Xq16C12) and Sichuan University
Central University Research Fund project (Grant No. skqy201653), Soft Science Project
of Sichuan Science and Technology Department, (No. 2017ZR0033) and Sichuan Social
Sciences Research 13th Five-Year annual planning project, (NO. SC16A006).
References
1. Chong E, Ma X (2010) The influence of individual factors, supervision and work
environment on creative self-efficacy. Creativity Innov Manag 19(3):233–247
2. Dong L, Zhang J, Li X (2012) Relationship between employee’s affect and creative
performance under the moderation of autonomy orientation standard (in Chinese)
3. Gong Z, Zhang J, Zheng X (2013) Study on feedback-seeking influence of creative
performance and motive mechanism (in Chinese)
4. Huang L, Krasikova DV, Liu D (2016) I can do it, so can you: the role of leader
creative self-efficacy in facilitating follower creativity. Organ Behav Hum Decis
Process 132:49–62
5. Song Y, He L et al (2015) Mediate mechanism between work passion and creative
performance of employees. J Zhejiang Univ (Sci Ed) 42(6):652–659 (in Chinese)
6. Tang C, Liu Y, Wang T (2012) An empirical study of the relationship among charis-
matic leadership style, team identity and team creative performance in scientific
teams. Sci Sci Manag S&T 33(10):155–162 (in Chinese)
7. Ulger K (2016) The creative training in the visual arts education. Thinking Skills
Creativity 19(32):73–87
8. Zhang J (2003) Study on situational factors of employees creative performance
and motive mechanism. Master’s thesis, Capital Normal University, Beijing (in
Chinese)
9. Zhang J, Dong L, Tian Y (2010) Improve or hinderinfluence of affect on creativity
at work. Adv Psychol Sci 6:955–962 (in Chinese)
10. Zhang J, Song Y et al (2013) Relationship between employees’ affect and cre-
ative performance the moderation role of general causality orientation. Forecasting
32(5):8–14 (in Chinese)
The Impact of Mixed Ownership Reform
on Enterprise Performance–An Empirical Study
Based on A-Share Listing Corporation in China
Sheng Ma(B) , Wenjie Li, and Sicheng Yan
Institute of Economics and Management,

Chengdu College, Chengdu, People’s Republic of China
478767236@qq.com
Abstract. This paper, with data of 486 listed companies from 2003 to
2014 over the time span of 12 years, finds that the proportion of non state-
owned shares in the Fixed-Effect Model is negatively correlated with the
profitability of the company, and it is positively correlated with the com-
pany’s development capacity; The proportion of circulating stocks has a
significant positive impact on the corporate debt solvency and devel-
opment ability; The proportion of the largest shareholder’s shareholding
and the proportion of the top ten shareholders are significantly correlated
with the total asset growth rate, while separation rate of two rights and
total asset growth rate are negatively correlated. Moreover, by using the
Difference-in-Difference Model to test the reform policies, the regression
results find that the restructuring has a significantly positive impact on
corporate profitability and development capacity. The consistency of two
methods confirms that the mixed ownership reform policy has a positive
role in promoting the profitability, debt solvency and development capa-
bility of state-owned listed companies.
Keywords: Mixed ownership reform · Non state-owned shares · Enter-

prise performance
1 Introduction
The core reforms of state-owned enterprise (SOE) on the three levels are mainly:
state-owned assets supervision, state-owned enterprise equity, state-owned enter-
prise operation. Among the three, SOE equity is the core of state-owned enter-
prise restructuring. The main core of SOE equity is to actively develop a mixed
ownership economy, allowing more state-owned economy and other ownership
economy to develop a mixed ownership economy, allowing mixed ownership econ-
omy to implement employee holdings, and finally the formation of co-ownership
of capital and labor Therefore, the exploration of the relationship between the
different ownership structure of various property nature and the performance
of enterprises in the first restructuring process and the “second revision” is a
The Impact of Mixed Ownership Reform on Enterprise Performance 671
great guidance for the reform of SOE, and is conducive to a comprehensive

understanding of the impact of restructuring on corporate performance.
In theory, the restructuring of SOE will improve the ownership structure
of listed companies, improve the governance structure of listed companies, and
then is conducive to improving the performance of SOE. This paper attempts
to explore the impact of mixed ownership reform on corporate performance
and analyze the policy implications of restructuring policy-making in the state-
owned enterprises and non-state-owned enterprises, and, through theoretical and
empirical analysis, tries to have some reference value to the participants and
researchers of SOE reform.
2 Literature Review
2.1 Summary of Relevant Literature Abroad
Researches of foreign scholars on corporate restructuring in the background of

the privatization began in the 1980s, and later expanded to a number of coun-
tries, including the restructuring of China’s SOE [3,5,6]. Djankov and Murrell
use the regression method of econometrics to make an empirical research on the
enterprises in the transition economy and carry on the detailed model analysis
[2]. Megginson and Nash studied the political, public institutions and economic
factors in the state-owned enterprises privatization by shares issuance and sale
on the public capital markets, conducted a comprehensive discussion of private
ownership and supported the view that national policy and legal environment
impact financial decisions [9]. Many foreign scholars have make related researches
on the economic policy of corporate governance, such as Pagano and Volpin [10].
2.2 Summary of Relevant Literature in China
Since reform and opening up, China has been experiencing a slow process of
economic transition. From the mid and late 1990s, due to the establishment
of the modern enterprise system emphasized by the Third Plenary Session of
the 14th CPC, the inefficiency of SOE drew the attention of the majority of
Chinese scholars to the researches on the efficiency of SOE. Yao [15] based on
the data analysis of the third industrial census, concluded that the rise of non-
state economic components may promote the overall level of China’s industrial
improvement. Liu using the 1995 national industrial census data, found that
among the variety of ownership enterprises, SOE has the lowest efficiency [7].
Liu and Li used [8] the data from 451 competitive enterprises (1994–1999) to
make an empirical research and obtain the negative impact of state ownership
on corporate performance, and conclude that non-state capital is positively cor-
related with firm performance. During this period, many scholars have studied
the impact of ownership changes on corporate performance [4,13]. Song Ligang,
Yao [12] also found a clear time trend, the effect of restructuring is most sta-
ble for those enterprises with a moderate length of restructuring history and
672 S. Ma et al.
for the enterprises that implemented restructuring between 1997 and 1999. Bai
and Tao [1] used the panel data of the national industrial enterprises from 1998
to 2003 to analyze the economic and social benefits of the SOE restructuring,
and concluded that the economic benefits of the state-controlled enterprises are
better while the social benefits of the non-state-owned enterprises are better,
and that the restructuring effect will continue for some time. Yang et al. [14]
found that although the number of employees decreased after the restructuring
of collective enterprises, the wage benefits per capita and taxation increased sig-
nificantly, indicating that in general the restructuring of collective enterprises
had a positive effect on social welfare. Sheng [11] used the multiplier method
based on the tendency scoring matching to analyze the micro-data of Chinese
industrial enterprises from 1999 to 2007, and concludes that the marketization
and the introduction of competition mechanism makes the reform of SOE play
a role in promoting social welfare. In the ongoing process of enterprise restruc-
turing, some scholars use the Double-differential Model to compare the effects
of the restructuring policy. Li and Qiao adopted the Double-differential Model
for China’s industrial data from 1999 to 2006, and found that the economic
performance of state-owned enterprises improved significantly in 2003, and that
the overall economic performance of SOE improved. Chen and Tang based on
the national industrial data from 1999 to 2007, study the social burden and the
policy burden on the enterprises with the Double-differential Model, and find
that the mixed ownership reform can reduce the policy burden of SOR, and that
the reform efficiency of monopoly mixed ownership is higher than that of the
competitive industry.
3 An Empirical Study on Enterprise Performance by

Mixed Ownership Reform
3.1 Data Sources
This paper uses the 486 listed companies listed in China’s A-share market from
2003 to 2014, including all the industry classification of the SFC in 2014, with a
total of 5832 effective observations, and set the dummy variable for the industry
and the three economic zones. According to the “State Council’s work report
on the state-owned enterprise reform and development” published in 2012, more
than 90% of the state-owned enterprises have completed the shareholding reform.
3.2 Definitions of Variables
(1) Corporate Performance Indicators

We use the company’s financial indicators to measure the corporate perfor-
mance indicators. 1 Profitability is replaced by ROA of total assets. The return
of the total assets(ROA) is the ratio of the net profit to the total assets of
the enterprise over a certain period of time. The higher the value of ROA, the
higher the profitability of the enterprise. 2 Debt solvency is measured by the
Debt Asset ratio (DAR), which is the percentage of total liabilities at the end of
the period divided by the total amount of assets. It is an important measure of
the level of corporate liabilities and the degree of risk. 3 The growth capacity
is measured by the sustainable growth rate (SGR) and the total asset growth
rate (TAGR). The SGR is the highest growth rate that can be achieved by the
non-issuance of new shares and maintaining the current operating efficiency and
financial policy. So, this indicator represents a suitable pace of development The
growth rate of the total assets is the ratio of the total assets increase of the year
to its total assets at the beginning of the year, so it reflects the growth of the
assets. The growth of the assets is an important aspect of the development of
the enterprises. Enterprises of high growth rate can maintain the steady growth
of assets.
(2) Explain the Variables
1 Mixed ownership restructuring index: the proportion of non-state-owned
shares (nonNSOS). Because the GTA CSMAR database contain the capital
structure of all the listed companies including the total number of shares and
the number of state-owned shares, so the proportion of non-state shares can be
calculated.
2 The proportion of circulating capital shares (LTBL): the ratio of the cir-
culating shares to the total shares of the company. The greater the proportion
of circulating shares, the more the stock reflects the true value of the company.
3 Proportion of sponsor shares (POP) is the ratio of the total number of
sponsor shares to the total capital shares of the company. The sponsor shares
refer to the special shares offered by the listed company to the founder(s) of the
company.
4 The proportion of the largest shareholder holdings (POFLS): the ratio
of the number of shares held by the largest shareholder to the total number of
shares.
5 The proportion of the top ten shareholders holding (POTTS): the ratio
of the number of shares held by the top ten shareholders to the total number of
shares.
6 The separation rate of two rights (SRTR): the difference between the
control of the actual controller of the listed company and his or her ownership.
We use the Herfindahl index to measure the concentration of the industrial
market. The higher the Herfindahl index, the higher the degree of market con-
centration, and the higher the degree of monopoly. Other explanatory variables
include total assets, total liabilities, paid-in capital, gross operating income, total
operating costs, equity multiplier, equity ratio, flow ratio, quick ratio, total num-
ber of shareholders, and income tax.
As the listed companies that we have selected basically involve in the SFC
industry classification of 2014, and these listed companies registered in the three
major economic zones of the eastern and central regions. Therefore, for the
study of the industry, several major industry categories are focused on, such as
(1) electricity, heat, gas and water production and supply, (2) manufacturing,
(3) real estate and (4) wholesale and retail. For the study of the regions, the three
674 S. Ma et al.
major economic zones in the eastern, central and western regions, two dummy
variables are set as (East) for the East region and (West) for the West region.
Of the 486 listed companies selected, 268 are located in the eastern region, 114
are in the central region and 106 are in the western region.
3.3 Design of the Measurement Model and Empirical Results

(1) The Impact of Restructuring on Corporate Performance of Listed Compa-
nies C An Analysis of Fixed-effect Model.
First of all, in order to study the impact of restructuring on the performance
of listed companies, that is, the changes of the proportion of non-state equity
in the overall equity changes, we need to control other factors that affect the
performance. Both observable and unobservable factors can affect the corporate
performance, so the observable factors are added to the regression model; for
the unobservable factors that are not observable but will not vary over time, we
can use the Fixed-Effect Model for their estimation, so the Fixed-effect Model
is constructed as follows:
Yit = α + β1 nonNSOSit + βi i = 26Xit + γit controlit + εit . (1)
Yit is the performance index of the listed company i in the year t, LTBLit
is theproportion of circulating capital shares of the company i in the year t,
6
and i=2 Xit are other important explanatory variables. The five explanatory
variables are explained as follow: LTBLit is the proportion of circulating capital
shares of the company i in the year t, POPit is the ratio of the total number
of sponsor shares to the total capital shares of the company i in the year t,
POFLSit is the ratio of the number of shares held by the largest shareholder to
the total number of shares of the company i in the year t, POTTSit is the ratio
of the number of shares held by the top ten shareholders to the total number
of shares of the company i in the year t, and SRTRit is the separation rate of
two rights the company i in the year t. controlit is the other control variables
used to improve the accuracy of the study. ε it is a random perturbation term
and α, β1 , β2 , β3 , β4 , β5 , β6 are the coefficients to be estimated. The results of
model 1 are listed in Table 1.
From Table 1 on the results of restructuring effects, we find that: with the
non-state-owned shares increasing, the profitability declines, sustainable growth
rate decreases, but the total asset growth rate increases significantly. The pro-
portion of circulating shares has a positive impact on the total assets growth
rate and Debt Asset ratio. The proportion of circulating shares is correlated to
the total asset growth rate at the significance level of 1%, and the proportion of
circulating shares is correlated to Debt Asset ratio at the significance level of 5%.
(2) An Analysis of the Effectiveness of Mixed Ownership Reform Policy on Cor-
porate Performance of Listed Companies C a Difference-In-Difference Model
Analysis.
In order to further effectively separate the impact of mixed ownership reform
on the performance of listed companies, the Difference-In-Difference can effec-
tively overcome the selective deviation of the sample in the regression process.
Table 1. Impact of restructuring on corporate performance of listed companies (Fixed-

effect Regression).
Profitability Debt solvency Development Total asset

Return of the debt asset ratio capacity growth rate
total assets (DAR) Sustainable (TAGR)
(ROA) growth rate
(SGR)
Proportion of Non −0.100∗∗∗ −0.12 −2.029∗∗∗ 0.217∗∗
state-owned shares (0.0374) (0.0826) (0.6470) (0.0863)
Proportion of 0.0109∗∗ 0.0363∗∗∗ 0.0643 0.0134
circulating shares (0.0048) (0.0107) (0.0836) (0.0112)
Proportion of 0.0341∗∗ −0.00746 0.912∗∗∗ −0.0247
sponsor shares (0.0170) (0.0376) (0.2950) (0.0393)
Proportion of −0.000233 0.00125 0.00641 0.00367∗∗∗
largest shareholder (0.000517) (0.00114) (0.00895) (0.00119)
holdings
Proportion of top 0.000611 0.00321∗∗ 0.00412 0.00381∗∗∗
ten shareholders (0.000574) (0.00127) (0.00992) (0.00132)
holding
The separation −0.000218 −0.00216 −0.00559 -0.00433∗∗∗
rate of two rights (0.000684) (0.00151) (0.0118) (0.00158)
The Herfindahl −0.0252 −0.352∗∗ −0.247 −0.331∗∗
index (0.0664) (0.147) (1.15) (0.153)
Constant −0.0659 0.924∗∗ −1.417 −1.002∗∗
Intercept term (0.174) (0.385) (3.014) (0.402)
Observations 4,111 4,111 4,110 4,107
R-squared 0.067 0.38 0.011 0.048
Note: The values in parentheses are the standard deviation. The superscripts ∗∗∗ , ∗∗ and ∗
indicate that the estimated coefficients are significant at the 1%, 5% and 10% levels respec-
tively.
Measuring and comparing the impact of restructuring policies on listed compa-

nies requires a more definite restructuring time point. From the micro level, we, in
the process of handling the enterprise sample data, have recognized that China’s
restructuring process is a gradual reform, will always go through a number of
steps or stages in order to achieve the final restructuring goals So the restructur-
ing time point remains difficult to define from the micro level. At the macro level,
the reform of state-owned enterprises has gone through five stages The emphasis
on deepening the reform and on the development of mixed ownership economy
began in the fourth stage with the “Decisions of the CPC Central Committee on
Several Issues Concerning the Improvement of the Socialist Market Economic
System” reviewed and approved by the Third Plenary Session of the 16th CPC
Central Committee. Then the State Council SASAC has put forward two notices:
“General Office of the State Council forwarded the SASAC Notice on Regulating
676 S. Ma et al.
the Reform of State-owned Enterprises” and “Opinions on Further Standardizing

the Reform of State-owned Enterprises”. Most of the equity nature changes of
the listed companies from the CSMAR database occurred around 2006, and the
number of publications on the restructuring of China’s state-owned enterprises
at the China National Knowledge Infrastructure (CNKI) reached its peak point
in 2006. So in our literature the restructuring time point is roughly determined
on this basis. Since the targets of SOE reform process are state-owned enter-
prises, and non-state enterprises are not involved in, so Difference-In-Difference
model can be built. The specific procedures are as follow:
The data sample is divided into two groups: one is the state-owned enterprises
labeled as D1i = 1 (known as the test group, i.e., enterprise i is a state-owned
enterprise), and the other group is non-state-owned enterprises (including pri-
vate and foreign capital) labeled as D1i = 0 (Known as the control group, i.e.,
enterprise i is a non-state-owned enterprise). Therefore, in the implementation
of SOE restructuring policy, it is clear that only the SOE as the test group are
affected, so the impact of restructuring policy on SOE can be expressed as E
(Y |D1i = 1), and the impact of restructuring policy on non-state-owned enter-
prises can be expressed as E(Y |D1i = 0). So we can express the net effect of
restructuring policy on the performance of SOE as:
E(Y |D1i = 1) − E(Y |D1i = 0). (2)
In order to explore the dynamic impact of the restructuring process, we

carry out the time difference on this basis. Because the restructuring policy
at different times will have differences, the test group and the control group
will change with the policy time. One period labeled as D2i = 1 is after the
implementation of the restructuring policy, the other period labeled as D2i = 0
is before the implementation of the restructuring policy. Therefore, we must also
compare the performance changes before and after the restructuring policy, in
which the performance impact after the implementation of restructuring policy
can be expressed as E(Y |D2i = 1), and the performance impact before the
implementation of restructuring policy can be expressed as E(Y |D2i = 0), so
we can get the impact of restructuring policy at the time level:
E(Y |D2i = 1) − E(Y |D2i = 0). (3)
So in order to measure the policy impact changes of the test group and the
control group at the same time level, we can make the following adjustment to
get the difference caused by restructuring policies:
Table 2. An analysis of the effect of mixed ownership reform policy on the performance
of listed companies (difference-in-difference model regression).
Profitability Debt solvency Development Total asset

capacity growth rate
(TAGR)
Return of the Debt Asset Sustainable
total assets ratio (DAR) growth rate
(ROA) (SGR)
D1 before and 0.00598 0.0269 −0.397∗ 0.141∗∗∗
after restructuring
−0.0139 −0.0307 −0.241 −0.0323
D2 test groupor −0.0460∗∗∗ 0.00209 0.091 −0.0201
Control group (0.0177) (0.0392) (0.307) (0.0412)
D3 Interactive item 0.0566∗∗∗ −0.0345 −0.0455 0.0852∗
(D1*D2) coefficient (0.0206) (0.0455) (0.357) (0.0479)
of Policy effect
nonNSOS −0.095∗∗∗ −0.125 −1.893∗∗∗ 0.171∗∗
Portion of Non (0.0371) (0.082) (0.643) (0.0863)
state-owned shares
LTBL 0.0110∗∗ 0.0371∗∗∗ 0.0591 0.0166
Proportion of (0.00476) (0.0105) (0.0826) (0.0111)
circulating shares
POP 0.0309∗ −0.00356 0.871∗∗∗ −0.025
Proportion of (0.017) (0.0376) (0.295) (0.0396)
sponsor shares
POFLS −0.00029 0.000958 0.00697 0.00337∗∗∗
Proportion of (0.000513) (0.00113) (0.00889) (0.00119)
largest shareholder
holdings
POTTS 0.000712 0.00313∗∗ 0.00499 0.00395∗∗∗
Proportion of top (0.000572) (0.00127) (0.00992) (0.00133)
ten shareholders
holding
SRTR −0.000392 −0.00223 −0.0056 −0.0043∗∗∗
The separation (0.000681) (0.00151) (0.0118) (0.00158)
rate of two rights
h −0.0281 −0.295∗∗ −0.5 −0.286∗
The Herfindahl (0.0656) (0.145) (1.139) (0.153)
index
Constant −0.0407 0.727∗∗ −0.864 −1.215∗∗∗
(0.167) (0.369) (2.889) (0.388)
Observations 4,111 4,111 4,110 4,107
R-squared 0.7013 0.3809 0.0339 0.1518
goodness of fit
Note: The values in parentheses are the standard deviation. The superscripts ∗∗∗ , ∗∗ and ∗
indicate that the estimated coefficients are significant at the 1%, 5% and 10% levels respec-
tively.
678 S. Ma et al.
{E(Y |D1i = 1) − E(Y |D1i = 0)} − {E(Y |D2i = 1) − E(Y |D2i = 0)}. (4)
The net effect of this restructuring policy not only measures the impact of
the policy before and after its implementation, but also measures the policy
differences between the test group and control group. So we have this regression
model of Difference-In-Difference (Model 2):
Yit = α + β1 D1i + β2 D2i + β3 D3i + βi i = 49Xit + γit controlit + εit . (5)
Yit is the performance index of the listed company i in the year t, and D1i
is the dummy variable between groups, where D1i = 1 is the test group, D2i is
the time dummy variable; and D3i is the interactive item. D3i = D1i × D2i ; β3
is the Difference-in-Difference statistics, that is, the differences brought about
by policy. Since other variables in Model Two are derived from the Model One,
so the definition and interpretation of these variables in Model Two are not
repeated here. For brevity we only list the regression results of several variables
that related to the proportion of non-state equity, and the results and discussion
of other variables are omitted. The results of Model (2) are shown in Table 2:
Table 2 shows the regression results for Model (2). The regression results
between the companies that implement mixed ownership reforms and those that
do not implement mixed ownership reforms show that difference-in-difference
statistics for the Interactive item D3 is very significant for the total net profit
margin, and is relatively significant for the total assets growth rate. In addi-
tion, the difference-in-difference statistics β3 coefficient is significantly positive,
indicating that compared to the enterprise without reform, the enterprise with
reforms have greatly improved their profitability and development capability.
The effect of the mixed ownership restructuring policy is not significant to the
difference-in-difference statistics β3 coefficient in the regression of the debt sol-
vency, which indicates that the effect of the mixed ownership reform policy is
not reflected. From Table 2, we can also see that with the increase in the pro-
portion of non-state-owned shares, the decline in profitability and sustainable
growth rate is relatively alleviated compared with the results in Table 1, indicat-
ing that the restructuring policy has played a role in improving the profitability
of enterprises.
4 Conclusions and Suggestions

The empirical conclusion is as follows:
(1) With the increase in the proportion of non-state-owned shares, the total
asset net profit margin decreases, and the sustainable growth rate decreases, but
the total assets growth rate increases significantly; the proportion of circulating
shares has a positive impact on the total assets growth rate and Debt Asset ratio;
Both the proportion of the largest shareholders holding and the proportion of
the top ten shareholders holding are positively correlated with the total asset
growth rate, while the separation rate of two rights is negatively correlated with
the total assets growth rate. The vertical comparison indicates that the increase
in the proportion of non-state equity can improve the development capacity of

enterprises, increase its debt solvency and reduce its profitability.
(2) The policy effect of Mixed ownership reform: Compared with enterprises
that have not been reformed, the enterprises that are reformed have a greater
improvement in profitability and development capability. The policy effect of
mixed ownership reform is not reflected in the debt solvency the results of other
explanatory variables in the Difference-in-difference Model are in good agreement
with the regression results of the Fixed-effect Model, which indirectly confirm
that the mixed ownership reform policy has a positive role in promoting these
financial measure of state-owned listed companies such as profitability, debt sol-
vency and development capability.
References
1. Bai D, Lu J, Tao Z (2006) An empirical study on the effect of state-owned enterprise
reform. Econ Res 8:4–13 (in Chinese)
2. Djankov S, Murrell P (2002) Enterprise restructuring in transition: a quantitative
survey. J Econ Lit 40(3):739–792
3. Frydman R, Gray C, Hessel M et al (1998) When does privatization work? The
impact of private ownership on corporate performance in the transition economies.
Q J Econ 114(4):1153–1191
4. Hu Y, Song M, Zhang J (2005) The relative importance and interrelationship of
the three theories of competition, property right and corporate governance. Econ
Res 9:44–57 (in Chinese)
5. Kang SW, Lee KH (2016) Mainstreaming corporate environmental strategy in
management research. Benchmarking 23(3):618–650
6. Kim J, Song HJ et al (2017) The impact of four CSR dimensions on a gaming
company’s image and customers’ revisit intentions. Int J Hospitality Manage 61:73–
81
7. Liu X (2000) The impact of the ownership structure of china’s industrial enterprises
on the efficiency difference - an empirical analysis of the census data of national
industrial enterprises in 1995. Econ Res 2:17–25 (in Chinese)
8. Liu X, Li L (2005) An empirical analysis of the impact of restructuring on corporate
performance. China Ind Econ 3:5–12 (in Chinese)
9. Megginson WL, Nash RC, Netter JM et al (2004) The choice of private versus
public capital markets: evidence from privatizations. J Finan 59(6):2835–2870
10. Pagano M (2005) The political economy of corporate governance. CSEF Working
Pap 95(4):1005–1030
11. Sheng D (2013) The restructuring of state-owned enterprises, the degree of compe-
tition and social welfare - a study based on the corporate cost mark-up percentage.
Economics (Q) 4:1465–1490 (in Chinese)
12. Song L, Yao Y (2005) The impact of restructuring on corporate performance. Soc
Sci China 2:17–31 (in Chinese)
13. Wu C, Li D (2005) Enterprise behavior in mixed market. Donyue Tribune 1:38–47
(in Chinese)
14. Yang Z, Lu J, Tao Z (2007) Political asylum and reform: research on the restruc-
turing China’s collective enterprise. Econ Res 5:104–114 (in Chinese)
15. Yao Y (1998) Effects of non-state ownership on technological efficiency of industrial
enterprises in China. Econ Res 12:29–35 (in Chinese)
Online Impulse Buying on “Double Eleven”
Shopping Festival: An Empirical Investigation
of Utilitarian and Hedonic Motivations
Umair Akram1 , Peng Hui1 , Muhammad Kaleem Khan1 ,

Muhammad Hashim2(B) , Yixin Qiu1 , and Ying Zhang1
1
2
Faculty of Management Sciences, National Textile University,
Faisalabad 37610, Pakistan
hashimscu@gmail.com
Abstract. “Double Eleven” online shopping festival has become the

most celebrated and largest online shopping event in China. This phe-
nomenon originates the need to explore impulse buying behavior in rela-
tionship of double eleven festival. This study aims to investigate the
structural relationship between hedonic, utilitarian web browsing and
online impulse buying behavior on “Double Eleven” shopping festival.
Data was collected from 426 online shoppers in six different districts
of Beijing. In order to assess the measurement model and to test the
hypothesis structural equation modeling was utilized by using AMOS
21. Additionally, Reliability, Discriminant and convergent validity was
used for the results of proposed model. The results of this study found
that utilitarian and hedonic web browsing positively influence on online
impulse shoppers with regard to “Double Eleven” shopping festival in
China. To the best of our knowledge, this study is pioneer, aiming at
exploring the dynamics of online impulse buying behavior with utilitar-
ian and hedonic predictors with regard to biggest online shopping festival
in China. This study provides a deep understanding of e-shoppers toward
online impulse behavior in the field of Chinese e-commerce industry that
can be generalized to other countries also. Implications for scholars and
e-tail managers of our study are discussed.
Keywords: E-commerce · “Double Eleven” shopping festival · Online

impulse buying · Utilitarian web browsing · Hedonic web browsing
1 Introduction
With the growing field of e-commerce, online shopping trend is becoming popular
across the world. China has become largest e-commerce economy in the world.
Consumers always prefer to buy goods online especially on occasions. Most of
Online Impulse Buying on “Double Eleven” Shopping Festival 681
the time when they make decision about online shopping they act impulsively.
One click ordering, easy access to products, absence of delivery, rich information
about product and time saving these all factors leads to buy impulsively online.
There are many factors that influence on impulse buying behavior. Impulse pur-
chase, is explained as compelling, hedonically complex and unplanned buying
behavior [35]. With the tremendous growth of e-commerce and rapid develop-
ment of information technology, online impulse purchasing has become an epi-
demic. It is also noteworthy that about 40% of all online shoppers spending
is attributable to online impulse purchasing [24,36]. Online shopping setting
is more encouraging to impulse buying behavior than its offline complement.
Online shopping environment liberates consumers from the constraints (e.g.,
social pressure from the staff, inconvenient store locations, limited operating
hours, and it need much time) that they might experience during physical shop-
ping events [13].
100000 60.00%
50.30% 50.00%
47.90%
75000 45.80%
42.10% 68826
64875 40.00%
10,000 Persons
38.30%
61758
34.30% 56400
50000 28.90% 51310 30.00%
45730
22.60%
38400 20.00%
16%
25000 10.50%
8.50% 29800
10.00%
21000
11100 13700
0 0.00%
2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015
The Number of Internet Users Internet Penetration Rate
Fig. 1. The size of Chinese internet users and penetration rate
According to the CNNIC report in December 2015 China had 688 million
netizens, up 39.51 million over the previous year. The internet penetration rate
reached 50.3%, up 2.4% points from the end of 2014 (see Fig. 1) [8]. The survey
result shows that in 2015, the mobile phone was the most popular device used
for internet access by new internet users, by 71.5% of them, up by 7.4% points
from the end to 2014. Among the new internet users in 2015, 46.1% were under
19, and 46.4% were students, and entertainment and communication were their
two biggest reasons to access the internet. Online impulse purchase is defined
as an immediate and sudden online buying with no preshopping intentions [31].
Previous studies reported that unplanned buying account for up to 60% of all
purchases [19,27] and according to [18,21] 40% to 60% of impulse purchases
depending on product category. A number of studies have been done in offline
setting with impulse buying while a little attention focused on online impulse
buying. Recently in an online buying setting, scholars have studied how to better
682 U. Akram et al.
appeal to impulse shoppers to take benefit of the behavior which has helped brick
and mortar retailers flourish for decades [9,23,36,37]. Irrespective of context, a
main purpose in retailing is to increase the attraction to improve sales [6,21].
Due to the pervasiveness and practical implications of impulse purchase, retailers
have focused significant efforts on facilitating the behavior [9,23]. This study is
not only beneficial for online retailers, but it also has future directions and
guidelines for scholars who are having been studied on impulse buying behavior
and produced number of studies in last decade.
In this regards, it is important to understand the comparative benefits of
buying online over offline shopping. Therefore, retailers should be aware of the
motivations behind the consumers’ online buying process. Motivations are an
essential factor to define the individual’s behavior. They stem unmet needs and
represent, through concrete actions, the benefits which people hope to achieve
[33]. Two types of motivational factors influence on online shopping i.e., Utilitar-
ian and Hedonic motivations [26]. While making essential consumption decisions,
it is reasonable to create a difference between utilitarian and hedonic motives.
These motives construct two aspects of attitudes with respect to behavioral acts.
Hedonic motivations are related with experiential or emotional aspects which
make the pleasured shopping experience while utilitarian motivation aspects
associated with rational, functional, practical, economic or extrinsic benefits. In
this research, we focus on how these two motivations influence on online impulse
shopping.
Table 1. 2009–2016 Alipay “Double Eleven” day trading volume (unit: 100 million
yuan)
Year 2009 2010 2011 2012 2013 2014 2015

Turnover 0.5 2.5 53 181 350.19 571.12 912.17
Source: Taobao official website.
In China, online shopping market has shown great potential due to rapid
growth of Chinese netizens and online consumers (see Fig. 1). In order to achieve
the competitive advantages a number of businesses joining the e-commerce indus-
try. They must understand consumer needs and try to meet their demands, they
should make a marketing strategies (e.g., low price product, augmented services,
speedy transaction process, quality of products) to earn profit and attract neti-
zens. In this regards, the “Double Eleven” online shopping festival has an emerg-
ing trend. Taobao introduced “Double Eleven” shopping festival on November
11, 2009 on “singles’ Day”. “Double Eleven” online shopping festival is most
popular among Chinese netizens/e-commerce industry and it is largest com-
mercial activity. Early 2016, the festival has been successfully held for seven
times. This event has carried out vast economic benefits, which have increased
year by year, and every year revive the record single-day net buying business
amount (see Table 1). Aforesaid studies about online impulse buying behavior
have shown that there is a great opportunity to identify how online impulse
purchase effected due to this event i.e. “Double Eleven” because authors best
knowledge no study has been conducted in this context. Only a few researchers
have studied to examine the “Double Eleven” model successful factors from dif-
ferent perspectives. Liu [24] reported that this setting can effectively motivate
to online shoppers to create impulse buying behavior (buying behavior that goes
beyond what has been planned).
This study provides a better understanding about motivational factors that
influence on online impulse buying adoption in China with the perspective of
“Double Eleven” shopping festival. A little work has been investigated on rela-
tionship between motivational factors and online impulse buying, but no sin-
gle study has yet investigated these factors in the context of “Double Eleven”
shopping event. The Current study fill this gap by incorporated the conception
of motivational factors i.e. Utilitarian and Hedonic on online impulse buying
behavior in the context of “Double Eleven” shopping activity. Aforementioned
discussions, the purpose of this manuscript is threefold: (1) to highlights the
major findings in study regarding “Double Eleven” shopping festival, (2) to
investigate the relationship between online impulse buying and utilitarian moti-
vations with the regard of Double Eleven shopping activity (3) to explore the
relationship between hedonic motivations and online impulse buying with the
regard to “Double Eleven” shopping activity. The rest of the manuscript is orga-
nized as follows. First we discuss core concepts of online impulse buying and
motivational factors i.e., utilitarian and hedonic and their relationship, based
on their connection we develop the hypotheses and draw research model. The
next section, we test the hypotheses based on data collection and data analysis.
Finally, last section presents the research outcomes, managerial implications and
future direction.
2 Online Impulse Buying

Online impulse purchasing is an immediate, unplanned, compelling and sudden
purchase behavior while shopping experience [1,2]. Consumers usually tend to
make immediate and unintended purchases while making online buying [20] their
intention about online buying might be related to website complicity and simplic-
ity [38]. In this context, Sharma [34] demonstrated that online buying is driven
by low cognitive control or spontaneous behavior and consumers’ emotions. They
argued that consumer’s impulse buying behavior is driven by appealing objects,
which influence them to make purchase without considering financial and other
factors of the online buying. Verhagen et al. [36] and Dholakia [9] mentioned that
online consumers are more likely to buy impulsive as compared to offline. Many
consumers to overspend due to online transactions because the virtual procedure
does not feel like spending money [11]. Online marketing stimuli make buying
spontaneously and allow online buyers to be less risk-averse [25]. E-impulse influ-
enced by many motivational factors i.e. utilitarian and hedonic web browsing and
other factors such as web site quality, sales promotion, ease of payment.
684 U. Akram et al.
3 Utilitarian and Hedonic Motivations

According to the scholars [25,28] Utilitarian and hedonic factor both influence on
impulse purchase. Impulse purchasing and different kind of browsing are effort-
less feelings [34]. Online shoppers motivate by searching due to the benefits such
as entertainment, uniqueness and fun [15]. For specific products, online impulse
buying occurs due to emotional and hedonic browsing [30]. Internet facilitates
browsing the online merchandise for informational purpose (i.e., utilitarian) and
recreational i.e., hedonic browsing [25]. Consumers usually act as impulsively
when they are taking online decisions due to easy buying e.g., one click order,
absence of delivery efforts and easy access to product [36]. Park [9] confirmed
that utilitarian and hedonic browsing has a relationship with impulsiveness in the
context of online buying of apparel products. Additionally, Gohary [14] demon-
strated that effect of hedonic and utilitarian browsing on impulse behavior which
points towards the importance of hedonic and utilitarian browsing for impulse
purchasing over the internet [12,36]. Finally, Rezaei [32] confirmed that utilitar-
ian and hedonic browsing positively influence on online impulse buying. Based
on above studies the following hypotheses regarding the relationship between
motivational factors and online impulse buying is derived:
H1 (HWB → OIB): Hedonic web browsing positively relate with online impulse
buying in the context of “Double Eleven” shopping events.
H2 (UWB → OIB): Utilitarian web browsing positively relate with online
impulse buying in the context of “Double Eleven” shopping events.
Fig. 2. A structure model for Hedonic web browsing, Utilitarian Web browsing and
Online impulse buying
Mathematical Model
OIB = β0 + β1 (HWB) + β2 (UWB) + ε.

The aim of this study is to find the motivational factors effect on OIB on “Dou-
ble Eleven” shopping festival. Data was collected from six different districts i.e.
Haidian, Chaoyang, Fengtai, Changping, Tongzhou, and Dongcheng of Beijing.
The survey was conducted on the Double Eleven shopping festival date on 11th
November 2016. Convenience sampling method was utilized to collect the large
data.
4.1 Questionnaire
Paper questionnaire and online survey techniques was used for data collection.
Initially, questionnaire was developed/adapt in English and subsequently trans-
lated into Chinese and back translation into English for accuracy. Services of
Chinese language experts and translators were employed to translate the ques-
tionnaire into Chinese and then it was also translated back into English to check
accuracy. Translation and back translation was validated by International Chi-
nese Training Center, Beijing. The first part of questionnaire was designed to
ensure the respondents have online buying experience on “Double Eleven” shop-
ping festival and in order to get information related with respondents’ character-
istics. The second section was designed to examine the relationship among online
impulse buying behavior, hedonic and utilitarian motivations. Pilot testing were
utilized (n = 65). Additionally, total 470 questionnaires were distributed among
online shoppers and (n = 426) was valid for the data analysis with response rate
90.63%. Note books were provided to the respondents in order to motivate and
seduce their patience, which cost about 5 RMB (.7 $).
4.2 Instrumentations
Five items were adapted from [36] to measure the online impulse buying behavior
e.g., My purchases were spontaneous. In order to measure the utilitarian web
browsing five items were adopted from [9] e.g., I browse to buy better items
in price and quality and to measure the hedonic web browsing four items were
adopted from [9] e.g., while web browsing, I am able to forget my problems and
to feel relaxed. Rezaei [32] also used these three instruments in his study.
5 Results and Hypothesis Testing

Table 2 reports the demographic characteristics of the respondents. Total 470 ques-
tionnaires were floated, out of 470, 426 respondents were found who had online
shopping experience. Demographics show that there are higher number of female
respondents as compared to males i.e. 63% are females and 37% males. Maximum
respondents were belonging to 25–31 years of age (i.e. 38.6% of total respondents).
Final collected respondents came out to be 51.8% students, 20.6% employed, 11.2%
individuals having managerial job and 9.8% having their own business.
686 U. Akram et al.
Table 2. Respondents Demographic Information (n = 426)
Age Gender Education

18–24 32.1% Male 37% Doctorate 11.5%
25–31 38.6% Female 63% Master 23.2%
32–40 25.3% Bachelor 35.3%
above 40 4.0% Diploma 25.0%
Other 5.0%
Occupation Years using the internet Online Impulse Experience
Business Owner 9.8% 1–4 7.5% Yes 426
Managerial level 11.2% 4–9 53.1% No 470
Employee level 20.6% 10–14 35.6%
Student 51.8% More than 14 3.8%
Other 6.6%
Kaiser-Meyer-Olkin (KMO) test formula

2
i=j rij
KMOj = 2
2 ,
i=j rij + i=j uij
k 2 ⎡ k 2 k
⎤

AVE = λi ⎣ λi + 1 − λ2i ⎦ .
i=1 i=1 i=1
k k
k
Measured with the Joreskog Rho = i=1 λ2i i=1 1 − λi
2 2
i=1 λi + .
Before examining the structure model, multi-collinearity issue was assessed
through variance inflation factor (VIF) produced by SPSS 21 version. The results
of VIFs indicated the value of 1.34 for online impulse buying, 2.70 for utilitarian
web browsing and 1.76 for hedonic web browsing variable, all stated values are
lower than the threshold value 10 [10]. Therefore, the results show that no serious
problem of multicollinearity in our study. Missing values in data set is a big
challenge for social science scholars like information system, Human resource
management and marketing areas. Many techniques can be performed for the
treatment of missing values but multiple imputation considered effectively [32].
In this study, to handle missing values and impute missing values effectively by
using SPSS, expectation maximization algorithm (EMA) was utilized. For the
missing value treatment, little’s missing completely at random MCAR which is
chi-square χ2 test for missing completely at random Little’s MCAR test: χ2 =
334.270, df = 332, significance level = .444. Furthermore, before factor analysis
Kaiser-Meyer-Olkin (KMO) test was assessed for the sampling adequacy which
shows KMO value .867. KMO value .867 is less then benchmark value .10 [16].
To examine the structure model confirmatory factor analysis (CFA) was
used to test the hypothesis and measurement model by using AMOS 21 ver-
sion. After removing problematic items, we reran the CFA analysis, the results
indicated that satisfactory fit (CMIN/DF: 1.82; p < .000; NFI = .98; IFI = 0.96;
Table 3. Convergent Validity and Reliability
Construct Item Loadings Cronbach’s alpha Composite reliabilitya AVEb

Online Impulse Buying (OIB) OIB1 0.875 0.894 0.98 0.81
OIB2 0.863
OIB3 0.924
OIB4 0.911
OIB5 0.889
Utilitarian web browsing UWB1 0.945 0.877 0.912 0.9
UWB2 0.865
UWB3 0.876
UWB4 0.878
Hedonic web browsing HWB1 0.888 0.845 0.92 0.79
HWB2 0.899
HWB3 0.924
HWB4 0.789
Notes: a Composite reliability (CR) = (Square of the summation of the factor loading)/(square of
the summation of the factor loadings) + (square of the summation of the error variance); b Average
variance extracted (AVE) = (summation of the square of the factor loadings)/(summation of the
square of the factor loadings) + (summation of the error variances).
Table 4. Results of structural relationships and hypothesis testing
Construct Mean Std. Deviation OIB UWB HWB

OIB 4.03 1.56 0.81 - -
UWB 3.98 1.77 0.54 0.9 -
HWB 4.16 1.55 0.452 0.64 0.79
Note: Acronyms: Online impulse buying (OIB); Utili-
tarian web browsing and hedonic web browsing; Diago-
nal bold values represent AVEs and off-diagonal denotes
square correlation among constructs. [∗p < 0.05, ∗∗p < 0.01,
∗∗∗p < 0.001].
GFI = .887; CFI = .91; RMSEA = 0.05). All values of CFA analysis were sat-
isfactory and meet the threshold value. Reliability and Validity test performed
in the model. Convergent validity was examined by using three parameters i.e.,
(1) factor loading grater then .70 with statistical significance (2) composite reli-
ability (CR) larger then .80. (3) Average variance extraction (AVE) higher then
.50 [16]. Table 3 indicated that all values of factor loading is greater than .70, in
addition, all constructs show the high level of reliability internal consistency CR
and Cronbach’s alpha values ranges from .84 to .98. The value of AVE is greater
then the benchmark value .50. Thus, we attained the good level of convergent
validity and reliability. Table 3 reports all statistical values of mean, standard
deviation and discriminant validity. We used Hair’s criterion to evaluate the dis-
criminant validity which directs square root to average variance extraction AVE
for all constructs should be higher then it correlations [16]. Discriminant validity
among all variables were confirmed based upon Hair’s criterion (see Table 4).
688 U. Akram et al.
Table 5. Results of structural relationships and hypothesis testing
Hypotheses Path R2 Path Coefficient Standard error t-statistics Decision

H1 HWB → OIB 0.424 0.823 0.067 21.4∗∗∗ Supported
H2 HWB → OIB 0.541 0.764 0.089 32.1∗∗∗ Supported
Notes: t-values for two tailed test: ∗∗∗ 2.58 (significance level = 1%); ∗∗ 1.96 (significance
level = 5%); and ∗ t-value 1.56 (significance level 10%) [17]; Acronyms: online impulse buying
(OIB); Hedonic web browsing (HWB) and utilitarian web browsing (UWB).
Table 5 depicts the results of path coefficients (β) which indicates the hypoth-
esized association among the constructs. Figure 2 represents the results of struc-
ture model. H1 that proposes there is a positive relationship between hedonic
web browsing and online impulse buying was approved with path coefficient (β)
0.823, t-statistics 21.4 and standard error 0.067. Results demonstrated that hedo-
nic web browsing positive and significantly influences on online impulse buying
on “Double Eleven” shopping festival. H2 proposed that utilitarian web brows-
ing positively influence on online impulse buying (utilitarian web browsing →
online impulse buying) with coefficient (β) of 0.764, standard error of 0.089 and
the value of t-statistics 32.1. H2 was also supported. Furthermore, Table 4 shows
the values of R2 for the relationships of constructs. The R2 value for (UWB →
OIB) .541 and for (HWB → OIB) .424, which means 54.1% change in online
impulse buying due to utilitarian web browsing while 42.4% change in online
impulse buying due to hedonic web browsing. Hedonic web browsing strongly
influences on online impulse buying than utilitarian web browsing. The findings
confirm that UWB and HWB strongly influenced on online impulse buying.
6 Discussions and Conclusion

With a remarkable growth of Chinese internet users and tremendous penetra-
tion rate of internet (see Fig. 1), web browsing motivation is considered as an
essential part of online shoppers’ experience. The current study provides impor-
tant insights of online impulse behavior based on web browsing motivation i.e.,
Hedonic and Utilitarian for e-tailers to develop their e-business strategies spe-
cially on big event like “Double Eleven”. This study investigates online impulse
buying shoppers towards “Double Eleven” e-shopping festival and examines
whether hedonic and utilitarian web browsing have any impact on it. This man-
uscript contributes to existing studies regarding online impulse buying behavior
by using exogenous constructs i.e., hedonic and utilitarian web browsing. With
the remarkable growth of online business, “Double Eleven” has turned out not
only a festival to celebrate, but also most popular and highest sales volume fes-
tival in online shopping. This study provides a significant insight of utilitarian
and hedonic web browsing behavior which influence on online impulse buying
behavior. In China, e-commerce companies are facing high rivalry environment
and they need to make innovative e-business strategies in order to increase online
impulse buying, that makes the results of this paper useful.
The results of this study demonstrated that Hedonic and utilitarian web
browsing both influenced on online impulse buying behavior on “Double Eleven”
shopping festival. The findings also support broadened theory of impulse pur-
chase behavior [5], which recommends that web browsing motivation is a key to
enhance online impulse purchase for apparel purchase from both hedonic and
utilitarian perspective. Inclination with the findings of this study Kim and Eas-
tim argued that there is a huge difference between hedonic and web browsing
behavior in online shopping. With the perspective of utilitarian value, consumers
are focused towards completing consumption goals [3,4,7,29], while hedonic
value, consumers are more attentive on entertainment, emotional and fun when
dealing with online browsing [7,22]. Park [9] found that hedonic web browsing
positively influences on OIB whereas utilitarian web browsing negatively influ-
enced on online impulse buying. In an online shopping context, this evidence
supports the hedonic nature of online impulse behavior [6,30].
E-tailer should make successful strategies by focusing on hedonic and util-
itarian web browsing motives for e-shoppers by ensuring in professional man-
ner, for instance; easy purchase process, good selection of variety, elegance and
security of the web site. These all factors lead to online impulse shopper. This
manuscript has some important practical implication towards OIB on “Double
Eleven” shopping activity. This study’s finding will help the e-tail managers and
web hosts associated with online selling to exploit exogenous factors HEB and
UWB to enhance their impulse sales through web traffic. Utilitarian and hedonic
web browsing plays a significant role to motivate e-shoppers to buy impulsively
specially on “Double Eleven” event. In China, there is rapid growth towards neti-
zens and penetration rate of internet users, now every individual prefer buying
online. The outcomes of this study will useful for two e-tail big players in China
(i.e., www.jd.com and www.taobao.com) as well as small e-tail businesses. There-
fore, marketers should focus on these strategies in order to increase the impulse
buying behavior which is major contribution in the field of e-commerce retailing
and marketing (Table 6).
Despite all the strengths and potential contributions of the study, to general-
ize the findings, future research should consider the limitation of study. First, in
this study we collect data from Beijing and just focus on “Double Eleven” shop-
ping festival. Results can be enhanced by expanding the study sample and incor-
porating more festivals that are found to increase online sales. Moreover, this
manuscript has considered only two motivational factors i.e., UWB and HWB,
future scholars may use other factors related to web site personality, emotional
factors such as perceived usefulness and perceived ease of use and outcomes
could be used to improve other type of e-commerce web site. This study is based
on quantitative data analysis and cross sectional in nature. For online shopping
characteristic of Double Eleven festival longitudinal and qualitative method (i.e.,
interviews and focus group) may also be used for deep understanding of online
shopping environment. Pre-and post-purchase behavior with regard to “Dou-
ble Eleven” shopping festival may also be analyzed by future researchers. Lastly,
future studies are recommended to analyze the proposed model in different e-tail
cultural environment.
690 U. Akram et al.
Acknowledgements. The authors want to extend their gratitude towards the

Editor (Jiuping, Xu) and the anonymous Reviewers for their indispensable and valu-
able suggestions and comments that improved the quality of the paper significantly.
Furthermore, they thank Prof. Peng Hui, Professor of finance, School of economics
and management, Beijing University of Posts and Telecommunications for valuable
suggestions.
Appendix
Table 6. Measurement items
No Construct Scale Source

1 Utilitarian web browsing UWB1 I browse to buy better items in price or quality [30]
UWB2 I browse the shopping web sites to gather
information about products
UWB3 I look around the shopping web sites to
comparison shop
UWB4 I browse the shopping web sites to get
additional value as much as possible
UWB5 I browse for efficient shopping online
2 Hedonic web browsing HWB1 While web browsing, I am able to forget my [30]
problems and to feel relaxed
HWB2 During web browsing, I am very excited, like
playing
HWB3 I enjoy web browsing enough to forget a time
out
HWB4 I look around at items on the internet just for
fun
3 Online impulse buying OIB1 My purchase was spontaneous [36]
OIB2 My purchase was unplanned
OIB3 I did not intend to do this purchase before this
shopping trip
OIB4 Before visiting the site, I did not have the
intention to do this purchase
OIB5 I could not resist doing this purchase at the site
References
1. Akram U, Peng H et al (2017) Impulsive buying: a qualitative investigation of the
phenomenon. Springer, Singapore
2. Akram U, Peng H et al (2016) Impact of store atmosphere on impulse buying
behaviour. Moderating effect of demographic variables. Int J U-and E-Serv Sci
Technol 9(7):43–60
3. Babin BJ, Griffin M (1994) Work and/or fun: measuring hedonic and utilitarian
shopping value. J Consum Res 20(4):644–656
4. Babin LA, Babin BJ, Boles JS (1999) The effects of consumer perceptions of the
salesperson, product and dealer on purchase intentions. J Retail Consum Serv
6(6):91–97
5. Baumeister RF (2002) Yielding to temptation: self-control failure, impulsive pur-
chasing, and consumer behavior. J Consum Res 28:670–676
6. Beatty SE, Ferrell ME (1998) Impulse buying: modeling its precursors. J Retail
74(2):161–167
7. Choi J, Li YJ et al (2014) The odd-ending price justification effect: the influence of
price-endings on hedonic and utilitarian consumption. J Acad Mark Sci 42(5):545–
557
8. CNNIC (2016) Statistic report on internet development in China
9. Dholakia UM (2000) Temptation and resistance: an integrated model of consump-
tion impulse formation and enactment. Psychol Mark 17(11):955–982
10. Diamantopoulos A, Siguaw JA (2006) Formative versus reflective indicators in
organizational measure development: a comparison and empirical illustration. Br
J Manage 17(4):263–282
11. Dittmar H, Long K, Meek R (2004) Buying on the internet: gender differences in
on-line and conventional buying motivations. Sex Roles 50(5):423–444
12. Eastin SK, Matthew S (2011) Hedonic tendencies and the online consumer: an
investigation of the online shopping process. J Internet Commer 10(1):68–90
13. Eroglu SA, Machleit KA, Davis LM (2001) Atmospheric qualities of online retail-
ing: a conceptual model and implications. J Bus Res 54(2):177–184
14. Gohary A, Hanzaee KH (2014) Personality traits as predictors of shopping motiva-
tions and behaviors: a canonical correlation analysis. Arab Econ Bus J 10(9):166–
174
15. Ha S, Stoel L (2012) Online apparel retailing: roles of e-shopping quality and
experiential e-shopping motives. J Serv Manage 23(2):197–215
16. Hair JF, Black WC et al (2010) Multivariate data analysis: a global perspective
17. Hair JF, Sarstedt M (2011) PLS-SEM: indeed a silver bullet. J Mark Theor Pract
19(2):139–151
18. Hausman A (2000) A multi-method investigation of consumer motivations in
impulse buying behavior. J Consum Mark 17(5):403–426
19. Inman JJ, Ferraro R, Winer RS (2004) Where the rubber meets the road: a model
of in-store consumer decision making
20. Jones MA, Reynolds KE et al (2003) The product-specific nature of impulse buying
tendency. J Bus Res 56(7):505–511
21. Kacen JJ, Hess JD, Walker D (2012) Spontaneous selection: the influence of prod-
uct and retailing factors on consumer impulse purchases. J Retail Consum Serv
19(6):578–588
22. Kaltcheva VD, Weitz BA (2013) When should a retailer create an exciting store
environment. J Mark 70(1):107–118
23. Kervenoael RD, Aykac DSO, Palmer M (2009) Online social capital: understanding
e-impulse buying in practice. J Retail Consum Serv 16(4):320–328
24. Liu Y, Li H, Hu F (2013) Website attributes in urging online impulse purchase: an
empirical investigation on consumer perceptions. Decis Support Syst 55(3):829–837
25. Madhavaram SR, Laverie DA (2004) Exploring impulse purchasing on the internet.
Adv Consum Res 31:59–66
692 U. Akram et al.
26. Martinez-Lopez FJ (2006) Motivations for consumption behaviours on the web: a

conceptual model based on a holistic approach. Int J Electron Mark Retail 1(1):3–
20
27. Mattila AS, Wirtz J (2008) The role of store environmental stimulation and social
factors on impulse purchasing. J Serv Mark 22(7):562–567
28. Novak TP, Hoffman DL, Duhachek A (2003) The influence of goal-directed and
experiential activities on online flow experiences. J Consum Psychol 13(1):3–16
29. Odekerken-Schröder G, Wulf KD, Schumacher P (2003) Strengthening outcomes of
retailer-consumer relationships: the dual impact of relationship marketing tactics
and consumer personality. J Bus Res 56(3):177–190
30. Park EJ, Kim EY, Forney JC (2006) A structural model of fashion-oriented impulse
buying behavior. J Fashion Mark Manage 10(4):433–446
31. Piron F (1991) Defining impulse purchasing. Adv Consum Res 18(1):509–514
32. Sajad R, Faizan A et al (2015) Online impulse buying of tourism products: the
role of web site personality, utilitarian and hedonic web browsing. J Hospitality
Tourism Technol 7(1):60–83
33. Schiffman K (2005) Behavior del consumer, 8th edn. Prentice Hall, México
34. Sharma P, Sivakumaran B, Marshall R (2010) Impulse buying and variety seeking:
a trait-correlates perspective. J Bus Res 63(3):276–283
35. Stern H (1962) The significance of impulse buying today. J Mark 26(2):59–62
36. Verhagen T, Dolen WV (2011) The influence of online store beliefs on consumer
online impulse buying: a model and empirical application. Inf Manage 48(8):320–
327
37. Wells JD, Parboteeah DV, Valacich JS (2011) Online impulse buying: understand-
ing the interplay between consumer impulsiveness and website quality. J Assoc Inf
Syst 12(1):32–56
38. Wu K, Vassileva J et al (2016) Complexity or simplicity designing product pictures
for advertising in online marketplaces. J Retail Consum Serv 28:17–27
Decision Support System
Fuzzy Multi-attribute Grey Relational Analysis
Using DEA and AHP
Mohammad Sadegh Pakkar(B)
Faculty of Management, Laurentian University, Sudbury, ON P3E 2C6, Canada

ms pakkar@laurentian.ca
Abstract. This research applies the method of grey relational analy-

sis (GRA) for multiple attribute decision making (MADM) problems in
which the attribute weights are completely unknown and attribute val-
ues take the form of fuzzy numbers. In order to obtain the attribute
weights, this research proposes an integrated data envelopment analysis
(DEA) and analytic hierarchy process (AHP) approach. According to
this, we define two sets of weights in a domain of grey relational loss,
i.e., a reduction in grey relational grade, between each alternative and
the ideal alternative. The first set represents the weights of attributes
with the minimal grey relational loss in DEA. The second set represents
the priority weights of attributes, bounded by AHP, with the maximal
grey relational loss. Using a parametric goal programming model, we
explore the various sets of weights in a defined domain of grey relational
loss. This may result in various ranking positions for each alternative in
comparison to the other alternatives. An illustrated example of a nuclear
waste dump site selection is used to highlight the usefulness of the pro-
posed approach.
Keywords: Grey relational analysis · Data envelopment analysis · Ana-

lytic hierarchy process · Multiple attribute decision making · Goal pro-
gramming · Fuzzy numbers
1 Introduction
Multiple attribute decision making (MADM) aims to find the ranking position
of alternatives in the presence of multiple incommensurate attributes. Many
MADM problems take place in an environment in which the information about
attribute weights are incompletely known and attribute values take the form of
intervals and fuzzy numbers [16,24,29].
Grey relational analysis (GRA) is part of grey system theory [3], which is
suitable for solving a variety of MADM problems with both crisp and fuzzy data.
The application of GRA with fuzzy data has recently attracted the attention of
many scholars [5,8,25].
GRA solves MADM problems by aggregating incommensurate attributes for
each alternative into a single composite value while the weight of each attribute is
696 M.S. Pakkar
subject to the decision maker’s judgment. When such information is unavailable

equal weights seem to be a norm. However, this is often the source of controver-
sies for the final ranking results. Therefore, how to properly select the attribute
weights is a main source of difficulty in the application of this technique. For-
tunately, the development of modern operational research has provided us two
excellent tools called data envelopment analysis (DEA) and analytic hierarchy
process (AHP), which can be used to derive attribute weights in GRA.
DEA is an objective data-oriented approach to assess the relative perfor-
mance of a group of decision making units (DMUs) with multiple inputs and
outputs [2]. Traditional DEA models require crisp input and output data. How-
ever, in recent years, fuzzy set theory has been proposed as a way to quantify
imprecise and vague data in DEA models [7,14,26]. In the field of GRA, DEA
models without explicit inputs are applied, i.e., the models in which only pure
outputs or index data are taken into account [11,27,28]. In these models, each
DMU or alternative can freely choose its own favorable system of weights to max-
imize its performance. However, this freedom of choosing weights is equivalent
to keeping the preferences of a decision maker out of the decision process.
Alternatively, AHP is a subjective data-oriented procedure which can reflect
the relative importance of a set of attributes and alternatives based on the formal
expression of the decision maker’s preferences. AHP usually involves three basic
functions: structuring complexities, measuring on a ratio-scale and synthesizing
[23]. Some researchers incorporate fuzzy set theory in the conventional AHP to
express the uncertain comparison judgments as fuzzy numbers [9,12,17].
However, AHP has been a target of criticism because of the subjective nature
of the ranking process [4]. The application of AHP with GRA can be seen in
[1,10,30].
In order to overcome the problematic issue of confronting the contradic-
tion between the objective weights in DEA and subjective weights in AHP,
this research proposes an integrated DEA and AHP approach in deriving the
attribute weights in a fuzzy GRA methodology. This can be implemented by
incorporating weight bounds using AHP in DEA-based GRA models. It is worth
pointing out that the models proposed in this article are not brand-new models
in the DEA-AHP literature. Conceptually, they are parallel to the application
of DEA and AHP in GRA using crisp data as discussed in [20]. Nevertheless,
it is the first time that these models are applied to a fuzzy GRA methodology.
Further research on the integration of DEA and AHP approach in deriving the
attribute weights with fuzzy data can be seen in [19].
2 Methodology
2.1 Fuzzy Multiple Attribute Grey Relational Analysis
Let A = {A1 , A2 , · · · , Am } be a discrete set of alternatives and C = {C1 ,

C2 , · · · , Cn } be a set of attributes. Let ỹij = (y1ij , y2ij , y3ij , y4ij ) be a trape-
zoidal fuzzy number representing the value of attribute Cj (j = 1, 2, · · · , n) for
Fuzzy Multi-Attribute Grey Relational Analysis 697
alternative Ai (i = 1, 2, · · · , m). Using α-cut technique, a trapezoidal fuzzy num-

ber can be transformed into an interval number as follows:
− +
yij = [yij , yij ] = [αy2ij + (1 − α)y1ij , αy3ij + (1 − α)y4ij )], (1)
− + −
where yij = [yij , yij ], yij ≤ yij +
, is an interval number representing the value of
attribute Cj (j = 1, 2, · · · , n) for alternativeAi (i = 1, 2, · · · , m). Then alterna-
− + − + −
tive Ai is characterized by a vector Yi = [yi1 , yi1 ], [yi2 , yi2 ], · · · , [yin +
, yin ] of
attribute values. The term Yi can be translated into the comparability sequence
− + − + −
Ri = ([ri1 , ri1 ], [ri2 , ri2 ], · · · , [rin +
, rin ]) by using the following equations [31]:
− +

− +
yij yij
[rij , rij ] = + , + ∀j, yj(max)
+
= max{y1j + +
, y2j , · · · , ymj
+
} (2)
yj(max) yj(max)
for desirable attributes,

− −

− +
yi(min) yi(min) − − − −
[rij , rij ] = − , + ∀j, yj(min) = min{y1j , y2j , · · · , ymj } (3)
yij yij
for undesirable attributes.

Now, let A0 be
a virtual ideal alternative which is characterized by a reference
sequence U0 = [u− 01 , u +
01 ], [u −
02 , u +
02 ], · · · , [u −
0n , u +
0n ] of the maximum attribute
values as follows:
u− − − −
0j = max{r1j , r2j , · · · , rmj } ∀j, (4)
u+
0j = +
max{r1j +
, r2j ,··· +
, rmj } ∀j. (5)
− +
To measure the degree of similarity between rij = [rij , rij ] and u0j =
− +
[u0j , u0j ] for each attribute, the grey relational coefficient, ξij , can be calculated
as follows:
− +
mini minj [u− − + − +
0j , u0j ] − [rij , rij ] + ρ maxi maxj [u0j , u0j ] − [rij , rij ]
+
ξij = − + , (6)
[u , u ] − [r− , r+ ] + ρ maxi maxj [u− , u+ ] − [r− , r+ ]
0j 0j ij ij 0j 0j ij ij
while the distance between u0j = [u− , u+ ] and rij = [rij − +

, rij ] is measured by
− + 0j+ 0j
|u0j − rij | = max u0j − rij , u0j − rij . ρ ∈ [0, 1] is the distinguishing coef-
−
ficient, generally ρ = 0.5. It should be noted that the final results of GRA
for MADM problems are very robust to changes in the values of ρ. Therefore,
selecting the different values of ρ would only slightly change the rank order of
attributes [13]. To find an aggregated measure of similarity between alternative
Ai , characterized by the comparability sequence Ri , and the ideal alternative
A0 , characterized by the reference sequence U0 , over all the attributes, the grey
relational grade, Γi , can be computed as follows:
n

Γi = wj ξij , (7)
j=1
698 M.S. Pakkar
n
where wj is the weight of attribute Cj and j=1 wj = 1. In practice, expert
judgments are often used to obtain the weights of attributes. When such infor-
mation is unavailable equal weights seem to be a norm. Nonetheless, the use
of equal weights does not place an alternative in the best ranking position in
comparison to the other alternatives. In the next section, we show how DEA can
be used to obtain the optimal weights of attributes for each alternative in GRA.
2.2 DEA-Based GRA Models
Since all the grey relational coefficients are benefit (output) data, a DEA-based
GRA model can be formulated similar to a classical DEA model without explicit
inputs [15]:
n

Γk = max wj ξkj , (8)
j=1
n

wj ξij ≤ 1 ∀i, (9)
j=1
wj > 0 ∀j, (10)
where Γk is the grey relational grade for alternative under assessment Ak (known
as a decision making unit in the DEA terminology). k is the index for the
alternative under assessment where k ranges over 1, 2, · · · , m. wj is the weight
of attribute Cj . The first set of constraints (9) assures that if the computed
weights are applied to a group of m alternatives, (i = 1, 2, · · · , m), they do not
attain a grade of larger than 1. The process of solving the model is repeated
to obtain the optimal grey relational grade and the optimal weights required
to attain such a grade for each alternative. The objective function (8) in this
model maximizes the ratio of the grey relational grade of alternative Ak to
the maximum grey relational grade across all alternatives for the same set of
weights (max Γk /maxi=1,··· ,m Γi ). Hence, an optimal set of weights in the DEA
based-GRA model represents Ak in the best light in comparison to all the other
alternatives. It should be noted that the grey relational coefficients are normal-
ized data. Consequently, the weights
n attached to them are also normalized. In
addition, adding the constraint j=1 wj = 1 to the DEA-based GRA model is
not recommended here. In fact, the sum-to-one constraint is a non-homogeneous
constraint (i.e., its right-hand side is a non-zero free constant) which can lead
to underestimation of the grey relational grades of alternatives or infeasibility in
the DEA-based GRA model (see [22])
2.3 Minimax DEA-Based GRA Model Using AHP
We develop our formulation based on a simplified version of the generalized dis-

tance model (see for example [6]). Let Γk∗ (k = 1, 2, · · · , m) be the best attainable
grey relational grade for the alternative under assessment, calculated from the
DEA-based GRA model. We want the grey relational grade, Γk (w), calculated
from the vector of weights w = (w1 , · · · , wn ) to be closest to Γk∗ . Our definition
of “closest” is that the largest distance is at its minimum. Hence we choose
the form of the minimax model: minw maxk {Γk∗ − Γk (w)} to minimize a single
deviation which is equivalent to the following linear model:
min θ (11)
n

s. t. Γk∗ − wj ξkj ≤ θ, (12)
j=1
n

wj ξij ≤ Γi∗ ∀i, (13)
j=1
θ ≤ 1, (14)
θ, wj ≥ 0 ∀j. (15)
The combination of (11)–(15) forms a minimax DEA based-GRA model that
identifies the minimum grey relational loss θmin needed to arrive at an optimal
set of weights. The first constraint ensures that each alternative loses no more
than θ of its best attainable relational grade, Γk∗ . The second set of constraints
satisfies that the relational grades of all alternatives are less than or equal to their
upper bound of Γk∗ . It should be noted that for each alternative, the minimum
grey relational loss θ = 0. Therefore, the optimal set of weights obtained from
the minimax DEA based-GRA model is exactly similar to that obtained from
the DEA-based GRA model.
On the other hand, the priority weights of attributes are defined out of the
internal mechanism of DEA by AHP. In order to more clearly demonstrate how
AHP is integrated into the newly proposed minimax DEA-based GRA model,
this research presents an analytical process in which attributes’ weights are
bounded by the AHP method. The AHP procedure for imposing weight bounds
may be broken down into the following steps:
Step 1. A decision maker makes a pairwise comparison matrix of different
attributes, denoted by B with the entries of bhq (h = q = 1, 2, · · · , n).
The comparative importance of attributes is provided by the decision
maker using a rating scale. Saaty [23] recommends using a 1–9 scale.
Step 2. The AHP method obtains the priority weights of attributes by comput-
ing the eigenvector of matrix B (Eq. (16)), e = (e1 , e2 , · · · , ej )T , which
is related to the largest eigenvalue, λmax .
Be = λmax e. (16)
To determine whether or not the inconsistency in a comparison matrix
is reasonable the random consistency ratio, C.R. can be computed by
the following equation:
λmax − N
C.R. = , (17)
(N − 1)R.I.
700 M.S. Pakkar
where R.I. is the average random consistency index and N is the size
of a comparison matrix.
In order to estimate the maximum relational loss θmax necessary
to achieve the priority weights of attributes for each alternative, the
following set of constraints is added to the minimax DEA-based GRA
model:
wj = αej ∀j. (18)
The set of constraints (18) changes the priority weights of attributes to

weights for the new system by means of a scaling factor α. The scaling
factor α is added to avoid the possibility of contradicting constraints
leading to infeasibility or underestimating the grey relational grade of
alternatives (see [22]).
2.4 A Parametric Goal Programming Model

In this stage we develop a parametric goal programming model that can be solved
repeatedly to generate the various sets of weights for the discrete values of the
parameter θ, such that 0 ≤ θ ≤ θmax . Let w(θ) be a vector of attribute weights
for a given value of parameter θ. Let w∗ (θmax ) be the vector of priority weights
of attributes obtained from the minimax DEA-based GRA model after adding
the set of constraints (18). Our objective is to minimize the total deviations
between w(θ) and w∗ (θmax ) with a city block distance measure. Choosing such
a distance measure, each deviation is being equally weighted subject to the
following constraints:
n

−
min Zk (θ) = (d+
j + dj ) (19)
j=1
−
s. t. wj − d+
j + dj = αej ∀j, (20)
−
d+
j , dj ≥ 0. (21)
−
and constraints (12)–(15), where d+ j and dj are the positive and negative devi-
ations from the priority weight of attribute Cj (j = 1, 2, · · · , n) for alternative
Ak (k = 1, 2, · · · , m). The set of Eq. (20) indicates the goal equations whose right-
hand sides are the priority weights of attributes adjusted by a scaling variable.
Because the range of deviations computed by the objective function is dif-
ferent for each alternative, it is necessary to normalize it by using relative
deviations rather than absolute ones. Hence, the normalized deviations can be
computed by:
Zk∗ (0) − Zk∗ (θ)
Δk (θ) = , (22)
Zk∗ (0)
where Zk∗ (θ) is the optimal value of the objective function for 0 ≤ θ ≤ θmax . We
define Δk (θ) as a measure of closeness which represents the relative closeness
of each alternative to the weights obtained from the minimax DEA-based GRA
model in the range [0, 1] after adding the set of constraint (18) to it. Increasing
the parameter (θ), we improve the deviations between the two systems of weights
obtained from the minimax DEA-based GRA model before and after adding the
set of constraints (18). This may lead to different ranking positions for each
alternative in comparison to the other alternatives. It should be noted that in a
special case where the parameter θ = θmax = 0, we assume Δk (θ) = 1.
3 Numerical Example: Nuclear Waste Dump Site

Selection
In this section we present the application of the proposed approach for nuclear
waste dump site selection. The multiple attribute data, adopted from Wu and
Olson [27], are presented in Table 1. There are twelve alternative sites and 4
performance attributes. Cost, Lives lost, and Risk are undesirable attributes and
Civic improvement is a desirable attribute. Cost is in billions of dollars. Lives lost
reflects expected lives lost from all exposures. Risk shows the risk of catastrophe
(earthquake, flood, etc.) and Civic improvement is the improvement of the local
community due to the construction and operation of each site. Cost and Lives
lost are crisp values as outlined in Table 1, but Risk and Civic improvement have
fuzzy data for each nuclear dump site.
We use the processed data as reported in [27]. First the trapezoidal fuzzy
data are used to express linguistic data in Table 1. Using the α-cut technique,
the raw data are expressed in fuzzy intervals as shown in Table 2. These data
are turned into the comparability sequence by using the Eqs. (2) and (3). Each
attribute is now on a common 0–1 scale where 0 represents the worst imaginable
attainment on an attribute, and 1.00 the best possible attainment.
Table 3 shows the results of a pairwise comparison matrix in the AHP model
as constructed by the author in Expert Choice software. The priority weight for
each attribute would be the average of the elements in the corresponding row
of the normalized matrix of pairwise comparison, shown in the last column of
Table 3. One can argue that the priority weights of attributes must be judged
by nuclear safety experts. However, since the aim of this section is just to show
the application of the proposed approach on numerical data, we see no problem
to use our judgment alone.
Using Eq. (6), all grey relational coefficients are computed to provide the
required (output) data for the DEA-based GRA model as shown in Table 4.
Note that grey relational coefficients depend on the distinguishing coefficient ρ,
which here is 0.80.
Solving the minimax DEA-based GRA model for the site under assessment,
we obtain an optimal set of weights with minimum grey relational loss (θmin ).
It should be noted that the value of the grey relational grade of all waste dump
sites calculated from the minimax DEA-based GRA model is identical to that
calculated from the DEA-based GRA model. Therefore, the minimum grey rela-
tional loss for the site under assessment is θmin = 0 (Table 5). This implies that
702 M.S. Pakkar
Table 1. Data for nuclear waste dump site selection
Site Cost Lives Risk Civic

Nome 40 60 Very high Low
Newark 100 140 Very low Very high
Rock Sprgs 60 40 Low High
Duquesne 60 40 Medium Medium
Gary 70 80 Low Very high
Yakima 70 80 High Medium
Turkey 60 70 High High
Wells 50 30 Medium Medium
Anaheim 90 130 Very high Very low
Epcot 80 120 Very low Very low
Duckwater 80 70 Medium Low
Santa Cruz 90 100 Very high Very low
Table 2. Fuzzy interval nuclear waste dump site data
Site Cost Lives lost Risk Civic

Nome [0.80–1.00] [0.40–0.70] [0.00–0.10] [0.10–0.30]
Newark [0.00–0.05] [0.00–0.05] [0.90–1.00] [0.90–1.00]
Rock Sprgs [0.70–0.95] [0.70–0.90] [0.70–0.90] [0.70–0.90]
Duquesne [0.50–0.85] [0.70–0.90] [0.40–0.60] [0.40–0.60]
Gary [0.40–0.60] [0.10–0.30] [0.70–0.90] [0.90–1.00]
Yakima [0.50–0.70] [0.10–0.30] [0.10–0.30] [0.40–0.60]
Turkey [0.75–0.90] [0.20–0.40] [0.10–0.30] [0.70–0.90]
Wells [0.85–0.95] [0.85–1.00] [0.40–0.60] [0.40–0.60]
Anaheim [0.00–0.30] [0.00–0.10] [0.00–0.10] [0.00–0.10]
Epcot [0.10–0.40] [0.00–0.20] [0.90–1.00] [0.00–0.10]
Duckwater [0.30–0.50] [0.20–0.40] [0.40–0.60] [0.10–0.30]
Santa Cruz [0.10–0.40] [0.10–0.30] [0.00–0.10] [0.00–0.10]
Table 3. Pairwise comparison matrix of 4 attributes
Attribute Cost Lives Risk Civic Priority

Cost 1 1/5 1/2 3 0.131
Lives 5 1 2 9 0.545
Risk 2 1/2 1 6 0.275
Civic 1/3 1/9 1/6 1 0.05
C.R. = 0.01
Table 4. Results of grey relational coefficient for nuclear waste dump site selection
Site Cost Lives lost Risk Civic

Nome 0.9383 0.6281 0.4578 0.4872
Newark 0.4444 0.4444 1 1
Rock Sprgs 0.8352 0.8352 0.7917 0.7917
Duquesne 0.6847 0.8352 0.6032 0.6032
Gary 0.6281 0.5033 0.7917 1
Yakima 0.6847 0.5033 0.4872 0.6032
Turkey 0.8837 0.539 0.4872 0.7917
Wells 0.9383 1 0.6032 0.6032
Anaheim 0.472 0.4578 0.4578 0.4578
Epcot 0.5033 0.472 1 0.4578
Duckwater 0.5802 0.539 0.6032 0.4872
Santa Cruz 0.5033 0.5033 0.4578 0.4578
the measure of relative closeness to the AHP weights for the site under assess-
ment is Δk (θmin ) = 0. On the other hand, solving the minimax DEA-based GRA
model for the site under assessment after adding the set of constraints (18), we
adjust the priority weights of attributes (outputs) obtained from AHP in such
a way that they become compatible with the weights’ structure in the minimax
DEA-based GRA model. This results in the maximum grey relational loss, θmax ,
for the site under assessment (Table 5). In addition, this implies that the mea-
sure of relative closeness to the AHP weights for the site under assessment is
Δk (θmax ) = 1.
Table 6 presents the optimal weights of attributes as well as its scaling factor
for all nuclear waste dump sites. It should be noted that the priority weights of
AHP (Table 3) used for incorporating weight bounds on the attribute weights
w
are obtained as ej = αj .
Going one step further to the solution process of the parametric goal pro-
gramming model, we proceed to the estimation of total deviations from the AHP
weights for each site while the parameter θ is 0 ≤ θ ≤ θmax . Table 7 represents
the ranking position of each site based on the minimum deviation from the pri-
ority weights of attributes for θ = 0. It should be noted that in a special case
where the parameter θ = θmax = 0, we assume Δk (θ) = 0. Table 7 shows that
Wells is the best alternative in terms of the grey relational grade and its relative
closeness to the priority weights of attributes.
Nevertheless, increasing the value of θ from 0 to θmax has two main effects
on the performance of the other sites: improving the degree of deviations and
reducing the value of the grey relational grade. This, of course, is a phenom-
enon, one expects to observe frequently. The graph of Δ(θ) versus θ, as shown
in Fig. 1, is used to describe the relation between the relative closeness to the
priority weights of attributes, versus the grey relational loss for each site. This
704 M.S. Pakkar
Table 5. Minimum and maximum grey relational losses for each nuclear waste dump
site
Site Γk∗ θmin θmax

Nome 1 0 0.2876
Newark 1 0 0.2761
Rock Sprgs 1 0 0.0487
Duquesne 0.8921 0 0.0345
Gary 1 0 0.2774
Yakima 0.7855 0 0.1742
Turkey 1 0 0.3251
Wells 1 0 0.0000
Anaheim 0.5735 0 0.0409
Epcot 1 0 0.2811
Duckwater 0.7351 0 0.0869
Santa Cruz 0.5943 0 0.0283
Table 6. Optimal weights of minimax DEA-based GRA model for all nuclear waste
dump sites bounded by AHP
w1 w2 w3 w4 α
0.1516 0.6308 0.3183 0.0579 1.1575
Table 7. The ranking position of each site based on the minimum distance to priority
weights of attributes
Site Z ∗ (θmin = 0) Rank Site Z ∗ (θmin = 0) Rank

Nome 1.1568 12 Turkey 0.8884 9
Newark 0.3955 4 Wells 0.0000 1
Rock Sprgs 0.1479 2 Anaheim 0.7521 7
Duquesne 0.2069 3 Epcot 0.5274 5
Gary 0.7049 6 Duckwater 0.8371 8
Yakima 0.9496 11 Santa Cruz 0.8972 10
may result in different ranking positions for each site in comparison to the other
sites. In order to clearly discover the effect of grey relational loss on the rank-
ing position of each nuclear dump site, as shown in Table 8 in Appendix, we
performed a Kruskal-Wallis test. The Kruskal-Wallis test compares the medi-
ans of rankings to determine whether there is a significant difference between
them. The result of the test reveals that its p-value is quite smaller than 0.01.
Therefore, we conclude that increasing grey relational loss in the whole range
[0.0, 0.33] changes the ranking position of each site significantly. Note that at
Wells 1
Rock 0.9
Anaheim
Duck 0.8
Nome
Duquesne 0.7
Yakima Epcot
0.6
Newark
Gary 0.5
Turkey
Santa 0.4
0.3
0.2
Δ (θ)
0.1
0
0 0.05 0.1 0.15 θ 0.2 0.25 0.3 0.35
Fig. 1. The relative closeness to the priority weights of attributes [Δ(θ)], versus grey
relational loss (θ) for each site
θ = 0 sites can be ranked based on Zk∗ (0) from the closest to the furthest from
the priority weights of attributes. For instance, at θ = 0, Nome, Newark and
Rock Sprgs with grey relational grades of one, are ranked in 12th , 4th and 2nd
places, respectively (Tables 5 and 7). However, with a small grey relational loss
at θ = 0.01, Nome, Newark and Rock Sprgs take 9th , 10th and 5th places in the
rankings, respectively. Using this example, as a guideline, it is relatively easy
to rank the sites in terms of distance to the priority weights of attributes. At
θ = 0.02, Newark moves up into 9th place while Nome and Rock Sprgs drop
in 10th and 6th places, respectively. It is clear that both measures, Zk∗ (0) and
Δk (θ), are necessary to explain the ranking position of each nuclear dump site.
4 Conclusion
We develop an integrated approach based on DEA and AHP methodologies for

deriving the attribute weights in GRA with fuzzy data. We define two sets of
attribute weights in a minimax DEA-based GRA framework. The first set rep-
resents the weights of attributes with minimum grey relational loss. The second
set represents the corresponding priority weights of attributes, using AHP, with
maximum grey relational loss. We assess the performance of each alternative (or
DMU) in comparison to the other alternatives based on the relative closeness of
the first set of weights to the second set of weights. Improving the measure of
relative closeness in a defined range of grey relational loss, we explore the vari-
ous ranking positions for the alternative under assessment in comparison to the
other alternatives. To demonstrate the effectiveness of the proposed approach,
an illustrative example of a nuclear waste dump site using twelve alternative
sites and 4 attributes is carried out. Further studies can apply the simultaneous
application of DEA and AHP to the field of GRA by considering the hierarchical
structures of attributes in the ranking positions of alternatives [18,21].
706 M.S. Pakkar
Appendix
Table 8. The measure of relative closeness to the priority weights of attributes [Δk (θ)]
verses grey relational loss[θ] for each nuclear waste dump site
θ Nome Newark Rock Sprgs Duquesne Gary Yakima Turkey Wells Anaheim Epcot Duckwater Santa Cruz
0 0 0 0 0 0 0 0 0 0 0 0 0
Rank N/A N/A N/A N/A N/A N/A N/A 1 N/A N/A N/A N/A
0.01 0.0451 0.0445 0.2054 0.2051 0.1409 0.0921 0.0369 1.0000 0.7774 0.0418 0.2693 0.8839
Rank 9 10 5 6 7 8 12 1 3 11 4 2
0.02 0.0870 0.0882 0.4108 0.5901 0.2019 0.1596 0.0725 1.0000 0.8644 0.0835 0.4711 0.9474
Rank 10 9 6 4 7 8 12 1 3 11 5 2
0.03 0.1263 0.1310 0.6162 0.8726 0.2586 0.2270 0.1068 1.0000 0.9292 0.1251 0.6439 1.0000
Rank 10 9 6 4 7 8 12 1 3 11 5 2
0.04 0.1627 0.1734 0.8217 1.0000 0.3100 0.2941 0.1405 1.0000 0.9940 0.1667 0.7656 1.0000
Rank 11 9 5 3 7 8 12 1 4 10 6 1
0.05 0.1967 0.2158 1.0000 1.0000 0.3550 0.3610 0.1743 1.0000 1.0000 0.2082 0.8284 1.0000
Rank 11 9 5 1 8 7 12 1 4 10 6 1
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
0.29 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9070 1.0000 1.0000 1.0000 1.0000 1.0000
Rank 1 1 1 1 1 1 12 1 1 1 1 1
0.3 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9335 1.0000 1.0000 1.0000 1.0000 1.0000
Rank 1 1 1 1 1 1 12 1 1 1 1 1
0.31 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9600 1.0000 1.0000 1.0000 1.0000 1.0000
Rank 1 1 1 1 1 1 12 1 1 1 1 1
0.32 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9865 1.0000 1.0000 1.0000 1.0000 1.0000
Rank 1 1 1 1 1 1 12 1 1 1 1 1
0.33 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
Rank 1 1 1 1 1 1 1 1 1 1 1 1
References
1. Birgün S, Güngör C (2014) A multi-criteria call center site selection by hierarchy
grey relational analysis. J Aeronaut Space Technol 7(1):1–8
2. Cooper WW, Seiford LM, Zhu J (2011) Handbook on data envelopment analysis.
Springer, US
3. Deng JL (1982) Control problems of grey systems. Syst Control Lett 1(5):288–294
4. Dyer JS (1990) Remarks on the analytic hierarchy process. Manage Sci 36(3):249–
258
5. Goyal S, Grover S (2012) Applying fuzzy grey relational analysis for ranking the
advanced manufacturing systems. Grey Syst 2(2):284–298
6. Hashimoto A, Wu DA (2004) A DEA-compromise programming model for com-
prehensive ranking. J Oper Res Soc Jpn 2(2):73–81
7. Hatami-Marbini A, Saati S, Tavana M (2010) Data envelopment analysis with
fuzzy parameters: an interactive approach. Core Discuss Pap Rp 2(3):39–53
8. Hou J (2010) Grey relational analysis method for multiple attribute decision mak-
ing in intuitionistic fuzzy setting. J Converg Inf Technol 5(10):194–199
9. Javanbarg MB, Scawthorn C et al (2012) Fuzzy AHP-based multicriteria decision
making systems using particle swarm optimization. Expert Syst Appl 39(1):960–
966
10. Jia W, Li C, Wu X (2011) Application of multi-hierarchy grey relational analysis

to evaluating natural gas pipeline operation schemes. Springer, Heidelberg
11. Jun G, Xiaofei C (2013) A coordination research on urban ecosystem in Beijing
with weighted grey correlation analysis based on DEA. J Appl Sci 13(24):5749–
5752
12. Kahraman C, Cebeci U, Ulukan Z (2003) Multi-criteria supplier selection using
fuzzy AHP. Logist Inf Manage 16(6):382–394
13. Kuo Y, Yang T, Huang GW (2008) The use of grey relational analysis in solving
multiple attribute decision-making problems. Comput Ind Eng 55(1):80–93
14. Lertworasirikul S, Fang SC et al (2003) Fuzzy data envelopment analysis (DEA):
a possibility approach. Fuzzy Sets Syst 139(2):379–394
15. Liu WB, Zhang DQ et al (2011) A study of DEA models without explicit inputs.
Omega 39(5):472–480
16. Olson DL, Wu D (2006) Simulation of fuzzy multiattribute models for grey rela-
tionships. Eur J Oper Res 175(1):111–120
17. Osman MS, Gadalla MH et al (2013) Fuzzy analytic hierarchical process to deter-
mine the relative weights in multi-level programming problems. Int J Math Arch
4(7):282–295
18. Pakkar MS (2015) An integrated approach based on DEA and AHP. CMS
12(1):153–169
19. Pakkar MS (2016) An integrated approach to grey relational analysis, analytic
hierarchy process and data envelopment analysis. J Cent Cathedra 9(1):71–86
20. Pakkar MS (2016) Multiple attribute grey relational analysis using DEA and AHP.
Complex Intell Syst 2(4):243–250
21. Pakkar MS (2016) Using DEA and AHP for hierarchical structures of data. Ind
Eng Manage Syst 15(1):49–62
22. Podinovski VV (2004) Suitability and redundancy of non-homogeneous weight
restrictions for measuring the relative efficiency in DEA. Eur J Oper Res
154(2):380–395
23. Saaty RW (1987) The analytic hierarchy process - what it is and how it is used.
Math Modell 9(3–5):161–176
24. Wei G, Wang H et al (2011) Grey relational analysis method for intuitionistic fuzzy
multiple attribute decision making with preference information on alternatives. Int
J Comput Intell Syst 4(2):164–173
25. Wei GW (2010) GRA method for multiple attribute decision making with
incomplete weight information in intuitionistic fuzzy setting. Knowl-Based Syst
23(3):243–247
26. Wen M, Li H (2009) Fuzzy data envelopment analysis (DEA): model and ranking
method. J Comput Appl Math 223(2):872–878
27. Wu DD, Olson DL (2010) Fuzzy multiattribute grey related analysis using DEA.
Comput Math Appl 60(1):166–174
28. Xie Z, Mo L (2013) Analysis method and its application of weighted grey relevance
based on super efficient DEA. Res J Appl Sci Eng Technol 5(2):470–474
29. Xu Z (2005) On method for uncertain multiple attribute decision making problems
with uncertain multiplicative preference information on alternatives. Fuzzy Optim
Decis Making 4(2):131–139
30. Zeng G, Jiang R et al (2007) Optimization of wastewater treatment alternative
selection by hierarchy grey relational analysis. J Environ Manage 82(2):250–9
31. Zhang J, Wu D, Olson DL (2005) The method of grey related analysis to multiple
attribute decision making problems with interval numbers. Math Comput Modell
42(9):991–998
A Fuzzy Multi-criteria Decision Making
Approach for Supplier Selection Under
Fuzzy Environment
Adnan Sarwar1 , Ziqiang Zeng1,2(B) , Richard AduAgyapong1 ,

Nguyen ThiHoaiThuong1,3 , and Talat Qadeer1
1
Uncertainty Decision-Making Laboratory, Sichuan University, Chengdu 610064,
zengziqiang@scu.edu.cn
2
Department of Civil and Environmental Engineering, University of Washington,
Seattle, WA 98195, USA
3
Faculty of Financial Accounting and Business Management,
Vietnam National University of Agriculture, Hanoi, Vietnam
Abstract. Supplier selection is crucial and multi-criteria decision mak-

ing problem in supply chain. Nowadays in competitive business environ-
ment, supplier selection is a critical task for purchasing department in
every organization. Appropriate supplier helps manufacturer to reduce
cost, consistent quality product, and enhance competitiveness in market.
In order to select potential supplier, it is essential to tradeoff between
tangible and intangible factors. Uncertainty and vagueness of decision
makers opinion is an important characteristic of this problem. Fuzzy ana-
lytical hierarchical process (AHP) and extent analysis method is used to
choose appropriate supplier under fuzzy environment. The application
procedure of FAHP and extent analysis method is elaborated through
numerical example.
Keywords: Supplier selection · Multi-criteria decision making · Fuzzy

set theory · Fuzzy AHP · Extent analysis
1 Introduction
Supplier selection is the most challenging issue in supply chain. In order to
develop sustainable supply chain companies need to select right supplier. Appro-
priate selection of supplier helps companies to provide high quality product at
right price and at right time. The term “Supply chain management” (SCM) can
be defined as it is a process that includes: sourcing raw material, convert raw
material to finish goods, and delivers product to final customer. The purpose of
SCM is to reduce inventory and production cycle time, to increase production as
well as achieve long terms goal of firms with respect to customer stratifications.
Appropriate selection of supplier is important task of purchasing department in
A Fuzzy Multi-criteria Decision Making Approach 709
any organization. In order to select supplier, it is necessary to evaluate supplier

based on different criteria. The supplier selection problem is considered as a
multi-criteria decision making problem. The process of supplier evaluation and
selection consists both quantitative and qualitative criteria [3]. Selection of sup-
plier against single factor is not sufficient. Supplier should be evaluated against
multi-criterion [11]. In supplier selection problem, it is necessary to be tradeoff
between tangible and intangible factors which may conflict each other. Optimal
selection of supplier can enable companies to manufacture innovative quality
product and gain competitive advantage in market. An effective supplier selec-
tion process plays important role to success of any organization. Whether right
or wrong selection of supplier directly effect on price, quality, and delivery time.
The purpose of supplier selection is to minimize purchase risk and build strong
relationship between supplier and buyer in competitive scenario [17]. Selection
of potential supplier is strategic decision that helps firms to achieve goal for
a longer period of time. Nowadays companies are changing reactive buying to
proactive buying to select wealthy supplier [12].
Supplier selection problem can be categorized into two types. Single sourc-
ing and multiple sourcing. Single sourcing means one supplier can provide all
required material to the buyer. In multiple sourcing, single supplier unable to sat-
isfy all needs of manufacturer. Hence companies need to split their required mate-
rial among various supplier [7]. Dickson [8] ranked 23 different criteria includes
quality, delivery, warranties, performance history, price, supplier profile, finan-
cial position, and technical capability for the purpose of supplier evaluation and
selection. The author mentioned that the most important criteria for supplier
selection is quality. Weber, et al. [26] reviewed 74 articles from the period 1966
to 1991 about supplier selection criteria and techniques. The author classified
supplier selection approaches into three different types; linear weighting meth-
ods, mathematical programming models, and statistical techniques. De Boer, et
al. [4] defined various phases of supplier selection process. In the initial phase,
define problem, formulation of criteria, qualification of supplier, and selection of
potential supplier among suppliers.
The process of supplier selection can be simple if one criteria is used in the
process of decision making. However, in case of more than one criteria, buyers
need to consider a range of different criteria in decision making process. If man-
ufactures are used multiple criteria in supplier selection process, manufacture
needs to consider influence of each criteria in decision making process, whether
weights are equally given to each criteria or according to type and importance
of different criteria [28]. In supplier selection process, two issues are very signif-
icant. Formulation of criteria that affect performance of supplier and determine
methodology that can be used to select the best supplier. Therefore, selection of
appropriate supplier is a critical task for any organization that affects organiza-
tion efficiency and profitability. Every supplier has a number of strengths and
weakness that should be assessed by buyer before ranking them.
710 A. Sarwar et al.
2 Literature Review
Recently, supply chain management and supplier selection process have gained
great attention in the literature. Appropriate supplier selection is a multi-criteria
problem but traditionally supplier selection based on only one criteria price. The
supplier selection process is one of the most important task for every firm to
establish effective supply chain. In competitive environment, it is critical task
for firms to select right supplier because potential supplier helps organizations
to produce high quality product at reasonable price. In past, various methodolo-
gies have been used to select supplier [1]. Proposed integrated methodology data
envelopment analysis (DEA) and group analytical hierarchy process (GAHP) to
evaluate and select most efficient supplier. Chatterjee, et al. [12] proposed case
based reasoning and decision support system including multi-attribute analy-
sis to measure supplier management performance. Usually, supplier evaluation
consists price, quality, flexibility, but in this paper some environment factors
were considered in supplier selection process using knowledge-based system.
Humphreys et al. [18] presented two methodologies for supplier selection prob-
lem and compare relatives performance of organization. Vikor and outranking
methods were used to select supplier and compare the relative performance of
supplier. Multi-criteria decision making model was applied in a construction
industry for supplier evaluation and selection [21]. Multi-criteria decision mak-
ing enabled construction industry to build good relationships among its supplier,
managers, and partners.
The balance and ranking method were used to select supplier based on multi-
criteria; profitability of supplier, technological capability, conformance quality,
and relationship closeness. The application of model consists three steps; con-
struction of out-ranking matrix, determine relative frequency of each supplier,
then triangularised out-ranking matrix to obtain implicit order of each sup-
plier with the help of balancing method [24]. The activity based costing (ABC)
technique was proposed for evaluation and selection of vendor. The ABC tech-
nique helped to calculate total cost caused by vendor in a company’s production
process, to judge performance of supplier [19]. The multi-criteria group decision
model for supplier selection was applied. This model based on ELECTRE IV
and VIP analysis approach. The proposed model consists two stages. In first
stage, ELECTRE IV utilized to determine ranking of different criteria. In sec-
ond stage, VIP analysis method was used to select alternatives [2]. Dimensional
analysis technique proposed to measure performance of supplier and to get index
called vendor performance index (VPI). In the process of supplier performance
evaluation both qualitative and quantitative criterion were used. To overcome
blindness of human and vagueness fuzzy approach applied [16]. Supplier selec-
tion is crucial task for every organization. The analytical network process (ANP)
method was proposed to select the best supplier based on three criteria, business
structure of supplier, manufacture capability of supplier, and quality system of
supplier. These main three criteria further classified to 45 sub-criteria [10,27].
Neural network (NN) approach used to select potential supplier. The determi-
nant factors quality, performance history, geographical and price were selected
in the process of supplier selection.
In Telecommunications Company AHP model was proposed for selection of
vendor. Vendor selection is a complex and multi-criteria decision making prob-
lem. In application process of AHP, two strategic factors (criteria) cost and
quality selected for evaluation of vendor, further both criteria categorized into
26 sub-criteria [23]. Shyura & Hsu-Shih [22] developed integrated model for ven-
dor selection. The hybrid model was included analytical network process (ANP),
TOPSIS, and nominal group.
A fuzzy multi-objective integer programming method was developed for sup-
plier selection problem which incorporate three objectives cost minimization,
maximize on-time delivery, and quality maximization [14]. Proposed strategy-
aligned fuzzy simple multi-attribute rating technique (SMART) to solve supplier
selection problem in supply chain. Strategy-aligned fuzzy SMART technique
involves three stages. In first stage, define objective and strategy, second step,
develop supply chain strategy, third stage recognize criteria for supplier selec-
tion [6,15]. In this study, fuzzy analytical hierarchy process method was proposed
with the concept of benefits, opportunities, cost, and risks (BOCR). With the
application of this approach supplier can be evaluated various aspects based on
quantitative and qualitative criteria. Fuzzy set theory was applied to overcome
human ambiguity in decision making process.
In this paper, fuzzy AHP method is applied to select best supplier. In fuzzy
AHP method, linguistic variables and triangular fuzzy numbers are used as a
pairwise comparison scale to determine priority of main criteria and sub-criteria.
After that, extent analysis method is applied to calculate final weight priority
of main decision criteria, sub-criteria and alternatives.
3 Fuzzy AHP
The analytical hierarchical process (AHP) introduced by [20]. AHP is very useful
method to solve multi-criteria decision making problem. It offers hierarchical
procedure to solve problem using both subjective and objective measures. The
hierarchical model of supplier selection problem consists four stages; objective,
criteria, sub-criteria, and last stage is decision of alternatives.
The advantage of AHP is simple and ease of use but the most important
disadvantage of this approach is that it uses nine-point scale which is unable to
handle uncertainty in the process of comparisons of decision variables. Tradition
AHP is unable to solve uncertain decision making problem, to overcome this
problem, linguistic variables and triangular fuzzy numbers are used to decide
priority of different decision variables. The extent analysis method is applied to
calculate priority of weight using triangular fuzzy numbers. The fuzzy analytical
hierarchical process is an extension of AHP which is used to handle fuzziness in
supplier selection process based on both quantitative and qualitative criteria.
Fuzzy set theory and fuzzy number
Fuzzy set theory proposed by [29]. Fuzzy set theory helps to overcome vague-
ness of human, imprecise data, and uncertainty in decision making. Fuzzy set
theory is used to represent mathematically human vagueness and uncertainty
in decision making problem. A fuzzy set is a class of elements with continuum
grades of membership. The membership value of each object lies between 0 and
1. Triangular fuzzy number is a part of fuzzy number whose membership func-
tions are defined by three numbers [9]. Where parameter m donates the most
promising value, l and u indicates lower and upper bond of fuzzy event.
The membership function µÃ is defined as
⎧ ⎫
⎪
⎪ 0, x<1 ⎪
⎪
⎨ ⎬
(x − l/m − l) , l ≤ x ≤ m,
µÃ = . (1)
⎪
⎪ (u − x/u − m) , m ≤ x ≤ u ⎪ ⎪
⎩ ⎭
0, x>u
3.1 Extent Analysis for Fuzzy AHP
The extent analysis method introduced by [25]. In this paper extent analysis
method is applied for supplier selection problem. In traditional AHP, nine point
scale was used to determine priority of one criteria over another. In fuzzy AHP,
linguistics variables and triangular fuzzy numbers are used. The fuzzy AHP
based on extent analysis is applied for supplier selection problem. Let X =
{x1 , x2 , · · · , xn } represent object set of alternatives and P = {p1 , p2 , p3 , · · · , pn }
represent goal set of supplier selection. Each object is taken and extent analysis
for each goal, p1 is performed respectively [5]. Then m extent analysis value of
each object can be obtained, using following signs:
m
1
Mgi 2
, Mgi , · · · , Mgi , i = 1, 2, · · · , n, (2)
j
where Mgi (j = 1, 2, 3, · · · , m) are Triangular fuzzy numbers. The steps of extent
analysis method proposed by [25];
Step 1. The value of synthetic extent analysis with respect to ith object is
defined as
⎡ ⎤−1
m n m
Si = j
Mgi ⊗⎣ j ⎦
Mgi . (3)
j=1 i=1 j=1
m j
To obtain j=i Mgi , perform the fuzzy addition operation of m extent analy-
sis values for a particular matrix such that,
⎛ ⎞
m m m m
j
Mgi =⎝ lj , mj , uj ⎠ , (4)
j=i j=1 j=1 j=1
m −1
n j
and to obtain i=1 j=1 Mgi , perform the fuzzy addition operation of
j
Mgi (j = 1, 2, · · · , m) values such that,
n m
n n n

j
Mgi = li , mi , ui , (5)
i=1 j=1 i=1 i=1 i=1
and then the inverse of the vector Eq. (5) is calculated such that,
⎡ ⎤−1
n m
⎣ j ⎦ 1 1 1
Mgi = n , n , n . (6)
i=1 j=1 i=1 ui i=1 mi i=1 li
Step 2. The degree of possibility of M2 = (l2 , m2 , u2 ) ≥ M1 = (l1 , m1 , u1 ) is

defined as
V (M2 ≥ M1 ) = sup min (µM1 (x), µM2 (y)) (7)
y≥x
and can be equivalently expressed as follows;

⎧ ⎫
⎨ 1, if m2 ≥ m1 ⎬
V (M2 ≥ M1 ) = hgt (M1 ∩ M2 ) = µM2 (d) = 0, if l1 ≥ u2 ,
⎩ l1 −u2 ⎭
(m2 −u2 )−(m1 −l1 )
, otherwise
(8)
where d is the ordinate of the highest intersection point D between µM1 and
µM2 . In Fig. 1, the intersection between M1 and M2 has been shown.
To compare M1 and M2 , we need both the values of V (M1 ≥ M2 ) and
V (M2 ≥ M1 ).
Step 3. The degree of possibility for a convex fuzzy number to be greater than
k convex fuzzy numbers Mi (1, 2, · · · , k) can be defined by
V (M ≥ M1 , M2 , · · · , Mk ) = V [(M ≥ M1 ) and (M ≥ M2 ) and · · · and (M ≥ Mk )]
= min V (M ≥ Mi ) ,
(9)
Assume that
d (Ai ) = min V (Si ≥ Sk ) . (10)
For k = 1, 2, · · · , n; k = 1.
T
W = (d (A1 ) , d (A2 ) , · · · , d (An )) , (11)
where Ai (i = 1, 2, · · · , n) are n elements.
Step 4. Via normalization, the normalized weight vectors are
T
W = (d (A1 ) , d (A2 ) , · · · , d (An )) , (12)
where W is a non-fuzzy numbers [25]. It gives the weight priority one alter-
native to another.
Fig. 1. The intersection between M2 and M1
4 Numerical Example
Suppose a manufacturing company wants to select supplier. Supplier selection is

critical task for every company. The fuzzy AHP method is applied to select poten-
tial supplier. In supplier selection process, three suppliers have been selected
for evaluation against three decision criteria. In evaluation process of supplier,
each criteria is defined using linguistic judgments provided by decision makers
(Table 1). The criteria is given below:
• Price: relates to product price and logistics management cost.

• Quality: relates to product durability and supporting service of quality.
• Flexibility: relates to production capacity flexibility and customization capa-
bility.
In Table 2, fuzzy pairwise comparison matrix of main decision criteria is made

with respect to triangular fuzzy numbers. Fuzzy synthetic extent values Eq. (3)
is used to calculate priority weight of main criteria (Fig. 2).
Table 1. Linguistic scale for supplier evaluation [13]
Linguistic variable Fuzzy triangular number

Just equal (1,1,1)
Equally preferred (1/2,1,3/2)
Weakly more preferred (1,3/2,2)
Strongly preferred (3/2,2,5/2)
Very strongly preferred (2,5/2,3)
Absolutely preferred (5/2,3,7/2)
Table 2. Fuzzy pairwise comparison matrix of main attribute
Price Quality Flexibility

Price (1,1,1) (1/2,1,3/2) (1,3/2,2)
Quality (2/3,1,2) (1,1,1) (3/2,2,5/2)
Flexibility (1/2,2/3,1) (2/5,1/2,2/3) (1,1,1)
Selection of
supplier
Price Quality Flexibility
Production
Product price Product Durability
capacity flexibity
Logistic Supporting service Customization

management cost of quality capability
Supplier 1 Supplier 2 Supplier 3
Fig. 2. Hierarchical structure of supplier selection problem

1 1 1
Price = (2.5, 3.5, 4.5) ⊗ , , = (0.20, 0.36, 0.66), (13)
12.67 9.67 6.86

1 1 1
Quality = (3.17, 4, 5.5) ⊗ , , = (0.25, 0.41, 0.80), (14)
12.67 9.67 6.86

1 1 1
Flexibility = (1.9, 2.17, 2.67) ⊗ , , = (0.15, 0.22, 0.39). (15)
12.67 9.67 6.86
Next step is to calculate degree of possibility using Eqs. (7) and (8).
(0.25 − 0.66)
V (Price ≥ Quality) = = 0.89, (16)
(0.36 − 0.66) − (0.41 − 0.25)
V (Price ≥ Flexibility) = 1, (17)
V (Quality ≥ Flexibility) = 1, (18)
V (Quality ≥ Price) = 1, (19)
(0.20 − 0.39)
V (Flexibility ≥ Price) = = 0.57, (20)
(0.22 − 0.39) − (0.36 − 0.20)
(0.25 − 0.39)
V (Flexibility ≥ Quality) = = 0.42. (21)
(0.22 − 0.39) − (0.41 − 0.25)
Table 3. Fuzzy pairwise comparison matrix with respect to price
Product price Logistics management cost

Product price (1,1,1) (3/2,2,5/2)
Logistics management cost (2/5,1/2,2/3) (1,1,1)
Normalized weight vector is WP = (1, 0).
Table 4. Fuzzy pairwise comparison matrix with respect to quality
Product durability Supporting service quality

Product durability (1,1,1) (1,3/2,2)
Supporting service quality (1/2,2/3,1) (1,1,1)
Normalized weight vector is WQ = (0.68, 0.32).
Table 5. Fuzzy pairwise comparison matrix with respect to flexibility
Production capacity flexibility Customization flexibility

Production capacity flexibility (1,1,1) (2,5/2,3)
Customization flexibility (1/3,2/5,1/2) (1,1,1)
Normalized weight vector is WF = (1, 0).
Table 6. Fuzzy pairwise comparison matrix of alternative with respect to product

price

Supplier 1 (1,1,1) (3/2,2,5/2) (1,3/2,2)
Suuplier 2 (2/5,1/2,2/3) (1,1,1) (1/2,1,3/2)
Supplier 3 (1/2,2/3,1) (2/3,1,2) (1,1,1)
Normalized weight vector is WP P = (0.51, 0.20, 0.29).
Table 7. Fuzzy pairwise comparison matrix of alternative with respect to logistics

management cost

Supplier 1 (1,1,1) (2/3,1,2) (1,3/2,2)
Supplier 2 (1/2,1,3/2) (1,1,1) (3/2,2,5/2)
Supplier 3 (1/2,2/3,1) (2/5,1/2,2/3) (1,1,1)
Normalized weight vector is WLM C = (0.39, 0.44, 0.16).
Applying Eq. (10) to calculate minimum degree of possibility:
d (Price) = min(0.89, 1) = 0.89, (22)

d (Quality) = min(1, 1) = 1, (23)
d (Flexibility) = min(0.57, 0.42) = 0.42. (24)
Table 8. Fuzzy pairwise comparison matrix of alternatives with respect to product

durability

Supplier 1 (1,1,1) (1/2,1,3/2) (2/7,1/3,2/5)
Supplier 2 (2/3,1,2) (1,1,1) (1,3/2,2)
Supplier 3 (5/2,3,7/2) (1/2,2/3,1) (1,1,1)
Normalized weight vector is WP D = (0.08, 0.39, 0.53).
Table 9. Fuzzy pairwise comparison matrix of alternatives with respect to supporting

service quality

Supplier 1 (1,1,1) (1,3/2,2) (5/2,3,7/2)
Supplier 2 (1/2,2/3,1) (1,1,1) (2,5/2,3)
Supplier 3 (2/7,1/3,2/5) (1/3,2/5,1/2) (1,1,1)
Normalized weight vector is WSSQ = (0.62, 0.38, 0.00).
Table 10. Fuzzy pairwise comparison matrix of alternatives with respect to production
capacity flexibility

Supplier 1 (1,1,1) (1/2,2/3,1) (2,5/2,3)
Supplier 2 (1,3/2,2) (1,1,1) (2/7,1/3,2/5)
Supplier 3 (1/3,2/5,1/2) (5/2,3,7/2) (1,1,1)
Normalized weight vector is WP CF = (0.41, 0.14, 0.45).
Table 11. Fuzzy pairwise comparison matrix of alternatives with respect to customiza-
tion flexibility

Supplier 1 (1,1,1) (2,5/2,3) (1/2,1,3/2)
Supplier 2 (1/3,2/5,1/2) (1,1,1) (5/2,3,7/2)
Supplier 3 (2/3,1,2) (2/7,1/3,2/5) (1,1,1)
Normalized weight vector is WCC = (0.43, 0.42, 0.15).
The weight vector calculated as W = (0.89, 1, 0.42). After the normalization

process, weight vector of main criteria price, quality, and flexibility was cal-
culated WG = (0.39, 0.43, 0.18). The same procedure of calculation is used to
determine priority weight of other pairwise comparison matrices (Tables 3, 4
and 5).
Therefore, supplier 1 is the best supplier for company among them because of
high priority weight. The second best supplier for company is supplier 3 (Tables 6,
7, 8, 9, 10, 11 and 12).
Table 12. Summary of combined priority weights
5 Conclusion
Suppler selection process is a broad and complex task for firms. In this paper,
multi-criteria decision making approach fuzzy AHP has been used which includes
both quantitative and qualitative criteria, to measure supplier performance. In
this model, linguistic variables and triangular fuzzy numbers are used to over-
come vagueness and uncertainty of decision makers. Multiple criteria helps deci-
sion makers to measure overall performance of supplier more efficiently. The
advantage of this approach is that it takes less time and more accurate to solve
supplier selection problem. Furthermore, fuzzy AHP model can be applied to any
manufacturing company to choose best supplier. The application of this model
is significantly effective and easily implement to choose best supplier.
Acknowledgements. The authors would like to give our great appreciation to editors
and anonymous referees for their helpful and constructive comments and suggestions,
which helped to improve this paper. This research is supported by the Youth Program
of National Natural Science Foundation of China (Grant No. 71501137), the Gen-
eral Program of China Postdoctoral Science Foundation (Grant No. 2015M572480),
the International Postdoctoral Exchange Fellowship Program of China Postdoctoral
Council (Grant No. 20150028), the Project of Research Center for System Sciences
and Enterprise Development (Grant No. XQ16B05), and Sichuan University (Grant
No. SKQY201647).
References
1. Ahadian B, Abadi AGM, Chaboki RM (2012) A new DEA-GAHP method for
supplier selection problem. Manage Sci Lett 2(7):2485–2492
2. Alencar LH, Almeida AT (2008) Multicriteria decision group model for the selection
of suppliers. Pesquisa Operacional 28(2):321–337
3. Ayhan MB (2013) A fuzzy AHP approach for supplier selection problem: a case
study in a gear motor company. Int J Manag Value Supply Chains 4(3):11–23
4. Boer LD, Labro E, Morlacchi P (2001) A review of methods supporting supplier
selection. Eur J Purchasing Supply Manage 7(2):75–89
5. Chang DY (1992) Extent analysis and synthetic decision. Optim Tech Appl 1:352–
355
6. Chou SY, Chang YH (2008) A decision support system for supplier selection based
on a strategy-aligned fuzzy smart approach. Expert Syst Appl 34(4):2241–2253
7. Demirtas EA, Ustun O (2009) Analytic network process and multi-period goal
programming integration in purchasing decisions. Comput Ind Eng 56(2):677–690
8. Dickson GW (1966) An analysis of vendor selection systems and decision. In: Mate-
rials science forum, pp 1377–1382
9. Dubois D, Henri P (1980) Systems of linear fuzzy constraints. Fuzzy Sets Syst
3(1):37–48
10. Gencer C, Grpinar D (2007) Analytic network process in supplier selection: a case
study in an electronic firm. Appl Math Model 31(11):2475–2486
11. Ho W, Xu X, Dey PK (2010) Multi-criteria decision making approaches for supplier
evaluation and selection: a literature review. Eur J Oper Res 202(1):16–24
12. Humphreys P, Mcivor R, Chan F (2003) Using case-based reasoning to evaluate
supplier environmental management performance. Expert Syst Appl 25(2):141–153
13. Kahraman C, Tijen E, Buyukozkan G (2006) A fuzzy optimization model for QFD
planning process using analytic network approach. Eur J Oper Res 171(2):390–411
14. Kumar M, Vrat P, Shankar R (2006) A fuzzy programming approach for vendor
selection problem in a supply chain. Int J Prod Econ 101(2):273–285
15. Lee AHI (2009) A fuzzy supplier selection model with the consideration of benefits,
opportunities, costs and risks. Expert Syst Appl 36(2):2879–2893
16. Li CC, Fun YP, Hung JS (1997) A new measure for supplier performance evalua-
tion. IIE Trans 29(9):753–758
17. Pal O, Gupta AK, Garg R (2013) Supplier selection criteria and methods in supply
chains: a review. Int J Soc Behav Educ Econ Bus Ind Eng 7(10):2667–2673
18. Prasenjit C, Poulomi M, Shankar C (2011) Supplier selection using compromise
ranking and outranking methods. J Ind Eng Int 7(14):61–73
19. Roodhooft F, Konings J (1995) Vendor selection and evaluation an activity based
costing approach. Eur J Oper Res 96(1):97–102
20. Saaty TL (1980) The analytic hierarchy process: planning, priority setting, resource
allocation. McGraw-Hill, New York
21. Schramm F, Morais DC (2012) Decision support model for selecting and evaluating
suppliers in the construction industry. Pesquisa Operacional 32(32):643–662
22. Shyur HJ, Shih HS (2006) A hybrid mcdm model for strategic vendor selection.
Math Comput Model 44(7–8):749–761
23. Tam MCY, Tummala VMR (2001) An application of the AHP in vendor selection
of a telecommunications system. Omega 29(2):171–182
24. Vahdani B, Zandieh M, Alemtabriz A (2008) Supplier selection by balancing and
ranking method. J Appl Sci 8(19):3467–3472
25. Veerabathiran R, Srinath KA (2012) Application of the extent analysis method on
fuzzy AHP. Int J Eng Sci Technol 4(7):649–655
26. Weber CA, Current JR, Benton WC (1991) Vendor selection criteria and methods.
Eur J Oper Res 50(1):2–18
27. Wei S, Zhang J, Li Z (1997) A supplier-selecting system using a neural network. In:
IEEE international conference on intelligent processing systems, vol 1, pp 468–471
28. Yahya S, Kingsman B (1999) Vendor rating for an entrepreneur development pro-
gramme: a case study using the analytic hierarchy process method. J Oper Res
Soc 50(9):916–930
29. Zadeh LA (1965) Fuzzy sets*. Inf Control 8(3):338–353
Cement Plant Site Selection Problem
with Carbon Emission Trading Mechanism
Lurong Fan1 , Zongmin Li2(B) , and Pingwen Wu3

1
Low Carbon Technology and Economy Research Center, Sichuan University,
2
lizongmin@scu.edu.cn
3
School of Economics, Sichuan University, Chengdu 610064,
Abstract. Location problem is a fundamental task for the cement plant

enterprises. This paper mainly investigates this problem with carbon
emission trading mechanism. In practice, most of cement plant location
problems are multi-objective and uncertain in nature. To enhance the
general practice of cement plant location problem, this paper proposes a
fuzzy random bi-level multi-objective model with carbon emission trad-
ing mechanism. Genetic algorithm is used to seek the approximate best
solution of the model. A numerical example is also given to show the
effectiveness of the proposed model.
Keywords: Cement plant location problem · Carbon emission trading

mechanism · Genetic algorithm
1 Introduction
Cement is the most widely used essential construction material which is the
main component of concrete, in addition, the cement industry is one of the most
intensive energy consumptions [4,5]. Cement industry has been always among
the largest CO2 emission sources. Almost 5–7% of global CO2 emissions are
caused by cement plants, while 900 kg CO2 is emitted to the atmosphere for
producing one ton of cement [3].
Along with the phenomenon that carbon emission trading is accepted by
many countries, the cement industry how to make product plan and select site to
get the maximum profit and customer satisfaction under carbon dioxide trading
mechanism, which are one of the most significant decision-making challenges and
practical significance problem. As a consequence, the cement industry worldwide
is facing growing challenges in conserving material and energy resources, as well
as reducing its CO2 emissions [7]. Because of the high carbon emission by cement
industry, a series of researches have been constructed to attempt to reduce carbon
emissions. Benhelal [2] presented new design of pyro-processing unit in a cement
722 L. Fan et al.
factory and proved such novelties new process can significantly reduce 66% of
CO2 emission compared to the existing process. Madlool [6] found that carbon
emission reduced by using energy efficiency measures in the cement industry.
However, there are few researches about cement plant site selection problem
in the traditional article. Ataei [1] used a multi criteria decision-making method
to rank alternative plant locations and found the best site. The problem of
location of cement plant mainly has considered the cement industry. In fact,
the customers purchase-choosing behavior should been taken into account in
the cement plant location problem. It is often difficult to estimate cement sale
revenues because these are decided by the customers. In turn, cement industry
can also have an impact on the customers purchase-choosing behavior. Therefore,
the cement plant location problem is formulated as a bi-level model with two
decision-makers: cement industry and customers. Because the purchase-choosing
behavior among the customers is uncertain, it is more reasonable to regard the
customer’s demands as a fuzzy-random variable. Therefore, the objective of this
study aims to develop a fuzzy-random bi-level model for cement plant location
problem with carbon emission trading mechanism.
The structure of the paper is as follows. Section 2 is the problem statement
part. A fuzzy-random bi-level multi-objective model for location of cement plant
with carbon emission trading mechanism is introduced in Sect. 3. And then a
solution algorithm for the problem based on the KKT condition and GA (Genetic
Algorithm) in Sect. 4. Section 5 gives a numerical example to illustrate the appli-
cation of the model. Finally, some conclusions are presented in Sect. 6.
2 Problem Statement
The problem considered in this paper is a decision making problem between the
cement industry and its consumers with carbon emission trading mechanism.
For cement plant location problem, because each consumer is independent indi-
vidual, cement plant site selection plan is not only determined by the cement
industry, but also influenced by consumers. The cement industry choose some
points to establish cement plants, decide the output and sale price, and after
obtaining the cement price, the consumers make their decision in an consumer
optimal manner to meet their demand and achieve the goal of minimal each
cost at the same time. Therefore, the cement industry influences the consumers’
purchase choosing behavior by adjusting the production plan and cement sale
price, whereas each consumer tries to meet their demand to make the minimal
cost based on decision made by the cement industry.
This paper considers the cement plant site selection problem as a leader-
follower game where the cement industry is leader and the consumers who can
freely chose the cement plant to purchase cement are the followers. In this sit-
uation, the cement plant site selection problem can be represented as a bi-level
problem where the cement industry is the upper-level and the consumer is the
lower-level (see Fig. 1).
Cement Plant Site Selection Problem 723
The upper level Carbon

Market
Carbon
Carbon allowance
allowance price
quantity
Cement Industry
Cement Cement Cement

plant 1 plant i plant I
Sale price Output

decision
Maximize profits
Material flow Capital flow
The lower level Consumers
Meet cement Minimize

Purchasing decision
demand cost
Fig. 1. The bi-level relationship between the cement plant and customers
3 Modeling Formulation
In this section, the relevant assumptions are first outlined, and then the bi-level
model for the cement plant site selection problem under the carbon emission
trading mechanism and considering fuzzy random is constructed. The mathe-
matical description of the problem is given as follows.
3.1 Assumptions
To construct a model for the location problem of cement plant, the following
assumptions are adopted:
(1) If customers buy cement, the cement plant must help the customer transport
the cement without fee.
(2) The total alternative points is fixed.
(3) There are not any old cement plants before the new ones to be built.
3.2 Notations
The following mathematical notations are used to describe the cement plant
location problem.
Index:
i: index of cement plants, where i = 1, 2, · · · , I;

j: index of demand nodes, where j = 1, 2, · · · , J;
n: index of cement species, where k = 1, 2, · · · , N ;
724 L. Fan et al.
Certain parameters:
I: number of cement plants;

J: number of demand nodes;
N : number of cement species;
tmax
j : the maximum tolerant transportation time of demand node j;
tj : the minimum transportation time of demand node j;
min
βn1 : the percentage of Ca in the cement n;

βn2 : the percentage of Mg in the cement n;
αi1 : the emission coefficient in product process at cement plant i;
αi2 : the emission coefficient in transportation process at cement plant i;
R: the initial allocated carbon allowance for cement company
Uncertain parameters:
f˜¯i : the annual fixed costs for cement plant i;

v̄˜i1 : the unit variable costs for cement plant i in product process;
v̄˜i2 : the unit variable costs for cement plant i in transportation process;
η̄˜: the carbon emission quota price;
˜ijn : the unit transportation time of cement n between cement plant i and
ā
demand node j;
D˜ ¯jn : the consumer j demand of cement n
Decision variables:
Oin : output decision of cement plant i about cement n;

ui : the binary variable, if Oin > 0, ui = 1, otherwise ui = 0;
Pn : the sale price of cement n;
xijn : purchasing decision of demand node j
3.3 Upper-Level Model for Cement Plant Location Problem
It is not easy for the cement industry to make decision to get maximize profits
and service satisfaction at the same time. Moreover, cement plant which is one of
the major contributors to carbon emission, location problem is more meaningful
to study with carbon emission mechanism.
(1) Objective Functions
In this paper, we consider cement plant which own two objectives, as the
upper-level. In order to maximize economic profit, taking into account both
improving the sales and saving the cost.
1 Sale revenues
Cement industry revenues come from cement sales which related to the
sale
N quantity
I andJcement price. So the expected value of economic benefits is
n=1 Pn i=1 j=1 xijn .
2 Carbon emission cost

The cement industry carbon emission comes from energy consumption for
complete process and the raw material calcination process. The carbon emission
by calcination process mainly from the reaction [3,8–10]:
CaCO3 → CaO + CO2 , (1)

MgCO3 → MgO + CO2 . (2)
From the reaction,

N we can calculateN the carbon emission from the calcination
11
process is: 10 n=1 O in βn1 + 11
6 n=1 Oin βn2 . Approximately 50% of the total
CO2 emissions from the cement manufacturing process are released by the chem-
ical processes and approximately 40% are released from burning fuel [10]. A
minor contribution, 10%, is derived from electricity use or transport processes
[11]. So the carbon emission from energy consumption includes produce processes
N N J
and transport processes, i.e. αi1 n=1 Oin + αi2 n=1 j=1 xijn . The carbon
emission is limited by the government, the amount of cement industry’s car-
bon emission would better not more than the limited standard. If initial carbon
allowances allocated are sufficient, the carbon related cost is negative value, oth-
erwise it indicates needing to buy carbon allowances. The carbon emission cost
is as follows:
I
11
N
11
N N N J
Oik βn1 + Oin βn2 + αi1 Oin + αi2 xijn − R Ed (η̄˜).
i=1
10 n=1
6 n=1 n=1 n=1 j=1
(3)
3 Construction cost
The construction costs take into account fixed cost Ed (f˜ ¯ ) and vari-
N i
n=1 Oin Ed (v̄i1 ) +
able costs which are related to the cement quantity, ˜
N J
j=1 xijn Ed (v̄i2 ).
n=1
˜
According to the aforementioned information, the equivalent for the total
economic profits of cement industry is established as:

N
I
J
I
11
N
11
N
max F = Pn ui − Ed (η̄˜) Oin βn1 + Oin βn2
n=1 i=1 j=1 i=1
10 n=1 6 n=1

N
N
J
+ αi1 Oin + αi2 xijn − R
n=1 n=1 j=1

I
I
N
N
J
− Ed (f˜
¯i )− Oin Ed (v̄˜i1 ) xijn Ed (v̄˜i2 ) .
i=1 i=1 n=1 n=1 j=1
726 L. Fan et al.
(2) Constraints
The cement industry will choose no more than M points from the total
alternative points to built cement plant, so we have:

I
ui ≤ M, (4)
i=1
where ui is decided by the cement plant output xijn ,

1, xijn > 0
ui = (5)
0, xijn = 0.
It acquire the variable costs and fixed costs which will quickly be sank cost
not more than G,
⎛ ⎞
I N
N
J
I
⎝ Oin Ed (v̄˜i1 ) + xijn Ed (v̄˜i2 )⎠ + Ed (f˜
¯i ) ≤ G. (6)
i=1 n=1 n=1 j=1 i=1
3.4 Lower-Level Model for Cement Plant Location Problem
The lower-level problem represents the customers’ choice behaviors and the
demand of distributed among cement plants, that is to say each customer assigns
his demand among the cement plants to minimize the total cost.
(1) Minimum Cost Objective
The lower-level objective is the minimum cost, it is decided by the cement
price and purchase quantity, the customer’s total cost is as follows:

J
N
I
min L = Pn xijn . (7)
j=1 n=1 i=1
(2) Constraints
The purchase plan must satisfy the cement demand:

I
xijn ≥ Ed (D˜¯jn ). (8)
i=1
The total purchase quantity of cement plant i is not more than the capacity
of it:

J
xijn ≤ Oin . (9)
j=1
3.5 Global Model
The lower-level customer choose cement plant and the amount of cement buying
from it. The up-level cement industry choose some points from the fixed points
to get the maximum profit and improve the service satisfactory; The customer
try to influence the cement industry’s decision to make their cost minimum. On
the one hand, the lower cement price, the lower cost to the customer; On the
other hand, the higher cement price, the higher profit to the cement industry.
The location of cement plant can be represented as an game, they both can
influence but can not control the other part.
The complete location problem involves the cement industry output decisions
and sale price and the consumers purchase decision. Therefore, the global model
includes bi-level objectives and constraints. In Noncooperative environments, we
consider bi-level programming problems and it can be expressed as:
⎧
⎪ max F = N P I J x − E (η̄˜) I
⎪ 11
N
O β 11
N
O β
⎪
⎪ n ijn d 10 in n1 + 6 in n2
⎪
⎪ Oin ,Pn i=1 n=1
⎪
⎪
n=1 i=1 j=1 n=1
⎪
⎪ N N J I
⎪
⎪ +αi1 Oin + αi2 xijn − R − ˜
ui Ed (f¯i )
⎪
⎪
⎪
⎪
n=1 n=1 j=1 i=1

⎪
⎪
⎪
⎪ I N N J
⎪
⎪ − Oin Ed (v̄˜i1 ) + xijn Ed (v̄˜i2 )
⎪
⎪
⎪
⎪ ⎧ i=1 n=1 n=1 j=1
⎪
⎪
⎪
⎪ ⎪ I
⎪
⎪
⎪ ⎪
⎪ ui ≤ M i = 1, 2, · · · , I
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪
i=1
⎨ ⎪ ui = i = 1, 2, · · · , I
⎪
⎪
⎪ 0 or 1,
⎪ ⎪
⎪ I N N J I (10)
⎪
⎪ ⎪
⎪ Oin Ed (v̄˜i1 ) + xijn Ed (v̄˜i2 ) + Ed (f˜¯i ) ≤ G
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪ i=1 n=1 n=1 j=1 i=1
⎪
⎪ ⎪
⎪
⎪
⎪ ⎨ where x ijn solve
⎪
⎪
⎪ s.t. ⎪
⎪
J N I
⎪
⎪ ⎪
⎪ min L = Pn xijn
⎪
⎪ ⎪ ijn⎧ j=1 n=1
⎪ x
⎪
⎪ ⎪
⎪
i=1
⎪
⎪ ⎪
⎪ ⎪ I x
⎪
⎪ ⎪
⎪ ⎪
⎪ ijn ≥ Ed (Djn )
˜
¯ j = 1, 2, · · · , J, n = 1, 2, · · · , N
⎪
⎪ ⎪
⎪ ⎪
⎨ i=1
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪ s.t. J
⎪
⎪ ⎪
⎪ ⎪ xijn ≤ Oin i = 1, 2, · · · , I, n = 1, 2, · · · , N
⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎩ ⎪
⎩ ⎪
⎩ j=1
xijn ≥ 0 i = 1, 2, · · · , I, j = 1, 2, · · · , J, n = 1, 2, · · · , N
4 Solution Approach
The proposed model (9) is a bi-level decision making problem, which reflects the
interactive relationship between the cement industry and consumers. There are
many algorithms available to solve bi-level programming. Therefore, the bi-level
model (9) can be converted into a single level model with additional constrains
by using KKT optimal conditions. So, the single level transformed from (10) by
728 L. Fan et al.
KKT conditions is as follows:

⎧ N I J I N N

⎪
⎪ F P x − E η̄
˜) 11
O β 11
O β
⎪
⎪ max = n ijn d ( 10 in n1 + 6 in n2
⎪
⎪ Oin ,Pn n=1 i=1 n=1 n=1
⎪
⎪
i=1 j=1
⎪
⎪ N N J I
⎪
⎪ +αi1 Oin + αi2 xijn − R − ui Ed (f˜ ¯i )
⎪
⎪
⎪
⎪
n=1 n=1 j=1 i=1

⎪
⎪
⎪
⎪ I N N J
⎪
⎪ − Oin Ed (v̄˜i1 ) + xijn Ed (v̄˜i2 )
⎪
⎪
⎪
⎪ ⎧ i=1 n=1 n=1 j=1
⎪
⎪
⎪
⎪ ⎪
⎪
I
⎪
⎪ ⎪
⎪ ui ≤ M i = 1, 2, · · · , I
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪ i=1
⎪ ui = 0 or 1,
⎪
⎪ ⎪
⎪ i = 1, 2, · · · , I
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪ I N N J I
⎪
⎪ ⎪
⎪ Oin Ed (v̄˜i1 ) + xijn Ed (v̄˜i2 ) + Ed (f˜¯i ) ≤ G
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪ i=1 n=1 n=1 j=1 i=1
⎪
⎪ ⎪
⎪
⎪
⎨ ⎪
⎪
N
⎪
⎪ pn − λjn − ϕin + δijn = 0, i = 1, 2, · · · , I, j = 1, 2, · · · , J
⎪
⎪ (11)
⎪ ⎪
⎪
n=1
⎪
⎪ ⎪
⎪ I
⎪
⎪ ⎪ λjn Ed (Djn ) −
⎪
˜
¯ xijn = 0, j = 1, 2, · · · , J, n = 1, 2, · · · , N
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎨ i=1
⎪
⎪
⎪
⎪ s.t. ϕin
J
⎪
⎪ ⎪ xijn − Oin = 0, i = 1, 2, · · · , I, n = 1, 2, · · · , N
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪
j=1
⎪
⎪ ⎪
⎪ δijn (xijn ) = 0, i = 1, 2, · · · , I, j = 1, 2, · · · , J, n = 1, 2, · · · , N
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪ I N
⎪
⎪ ⎪
⎪ xijn ≥ Ed (D˜ ¯jn ), j = 1, 2, · · · , J
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪ J
i=1 n=1
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪ xijn ≤ Oin , i = 1, 2, · · · , I, n = 1, 2, · · · , N
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪
j=1
⎪
⎪ ⎪
⎪ xijn ≥ 0, i = 1, 2, · · · , I, j = 1, 2, · · · , J, n = 1, 2, · · · , N
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪ ϕin ≥ 0, i = 1, 2, · · · , I, n = 1, 2, · · · , N
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪ δijn ≥ 0, i = 1, 2, · · · , I, j = 1, 2, · · · , J, n = 1, 2, · · · , N
⎪
⎩ ⎩
λjn ≥ 0, j = 1, 2, · · · , J, n = 1, 2, · · · , N
The KKT conditions switched the cement sales from a game between cement
industry and consumers to single decision making problem only by cement indus-
try. The cement industry must try to control the sale price to get more market
share. The consumers make purchase decisions by the cement price that reflects
the competition.
5 Practical Application and Results

In this section, a case study of a practical cement plant selection problem in
Conch cement company in China is introduced to demonstrate the effectiveness
of the proposed model and optimization method.
5.1 Description of Case Problem

In this part, Conch cement company is taken as an application to explore the
problem. Conch cement is a listed company with more than 80 subsidiaries in
17 province in China, and the parent company is located at Anhui province.
The average sales increased rate of Conch cement is 31.01% in last three years,
ranking at 471 in total 1710 listed company and ranking at (9/29) in the con-
struction industry. At the Copenhagen climate conference in 2009, the Chinese
government pledged that China would reduce carbon emissions intensity (i.e.
carbon emissions per unit GDP) by 40% ∼ 45% based on 2005 levels by 2020.
In order to achieve this target, the government has proposed policy to control
carbon emission. As one of the largest company in the construction industry,
Conch cement must be more cautious when deals with decision problem like
cement plant selection problem with limited carbon emission allowance.
5.2 Data Collection

Conch cement industry choose some points to build new cement plants which
not more than 4 points in the possible 6 points. Every cement plants produce 2
different type cement. The Ca and Mg’s percentage are different in the different
cement which will influence the carbon emission from natural release, is shown
in Table 1. The basic information for theses possible points, such as unit vari-
able cost in production and transportation process, maximum output, emission
coefficients in the production and transportation process are show in Tables 2
and 3. There are 3 different companies need buy some cements to build road,
Table 1. The percentage of Ca and M g in the cement n
Cement n Ca Mg
n=1 β11 = 12.19% β12 = 20.31%
n=2 β21 = 23.16% β22 = 38.59%
Table 2. The basic information of the cement plants
Cement plant Unit variable cost in Unit variable cost in Maximum

index i produce process v̄˜i1 transportation process v̄˜i2 output li
(RMB/Tone) (RMB/Tone) (104 Tone)
i=1 (165, ξ1 , 179) (55, θ1 , 72) 195
ξ1 ∼ N (172, 0.0258) θ1 ∼ N (57.28, 0.0127)
i=2 (190, ξ2 203) (69, θ2 , 81) 168
ξ2 ∼ N (195, 0.0081) θ2 ∼ N (70.34, 0.0215)
i=3 (230, ξ3 , 258) (96.2, θ3 , 101.7) 175
ξ3 ∼ N (242, 0.0195) θ3 ∼ N (98.5, 0.0097)
i=4 (262, ξ4 , 288.75) (79, θ4 , 105.2) 210
ξ4 ∼ N (273, 0.0347) θ4 ∼ N (83.2, 0.0438)
i=5 (181, ξ5 , 293) (73.9, θ5 , 87.15) 151
ξ5 ∼ N (224.5, 0.0491) θ5 ∼ N (76.5, 0.0278)
i=6 (176, ξ6 , 227.6) (82.05, θ6 , 97.63) 220
ξ6 ∼ N (187, 0.0392) θ6 ∼ N (86.42, 0.0365)
730 L. Fan et al.
Table 3. The basic information of the cement plants
Cement Emission coefficient in Emission coefficient in Annual fixed cost ū

˜i
plant produce process αi1 transportation (108 RBM)
index i (kg/Tone) process αi2 (kg/Tone)
i=1 302.12 90.24 (1.0542, 1 , 1.4518)
1 ∼ N (1.0589, 0.0653)
i=2 247.95 68.98 (2.0786, 2 , 2.4776)

2 ∼ N (2.2318, 0.1142)
i=3 444.44 103.65 (0.8269, 3 , 0.9287)

3 ∼ N (0.8675, 0.0625)
i=4 362.58 78.29 (1.8841, 4 , 2.0012)

4 ∼ N (1.9857, 0.0929)
i=5 233.92 57.48 (2.5987, 5 , 2.7014)

5 ∼ N (2.6013, 0.0471)
i=6 471.12 114.78 (1.6638, 6 , 1.8247)

6 ∼ N (1.6959, 0.0516)
Table 4. The information of the consumers
Consumer indexj The consumers’ demand The consumers’ demand of

˜ (Tone)
of cement 1 D̄ ˜ (Tone)
cement 2 D̄
1j 2j
j=1 (48.4, ζ11 , 75.12) (11.7, ζ21 , 24)

ζ11 ∼ N (50.6, 0.0621) ζ21 ∼ N (12.4, 0.0315)
j=2 (95.8, ζ21 , 119.9) (26.6, ζ22 , 42.1)
ζ21 ∼ N (105.37, 0.0323) ζ22 ∼ N (31.2, 0.0519)
j=3 (146, ζ31 , 170.05) (54.3, ζ32 , 89.7)
ζ31 ∼ N (156, 0.0282) ζ32 ∼ N (76.7, 0.0724)
airport, house, the relative information for these consumers, such as demand
for the cement are show in Table 4. The data about ca and mg percentage in
Table 1 are from the Conch cement industry. Most data in the Tables 2 and 3
were estimated based on expert consultations as well as the data in Table 4. The
“allocation cap” was calculated to be 849.6 × 104 Tonnes, according to the total
maximum output 944 × 104 Tonnes. The allowance price is (20, γ, 60), where
γ ∼ N (27.5, 10.36).
5.3 Results and Analysis
Almost 5–7% of global CO2 emissions are caused by cement plants, therefore, the
government allocated the allowance to the cement industry by the proportion.
But to ensure the efficiency of the proved model, the carbon emission allowance
R considering the following 3 scenarios.
Table 5. The purchase decision of the consumers
Consumer Cement Purchase Consumer Cement Purchase Consumer Cement Purchase

j n decision j n decision j n decision
xijn (104 xijn (104 xijn (104
Tones) Tones) Tones)
j=1 n=1 x111 = 0 j=2 n=1 x211 = 9 j=3 n=1 x311 =
156
x121 = x221 = 6 x321 = 0
51
x131 = 0 x231 = 0 x331 = 0
x141 = 0 x241 = 0 x341 = 0
x151 = 0 x251 = 0 x351 = 0
x161 = 0 x261 = x361 = 0
90
n=2 x112 = n=2 x212 = n=2 x312 = 0
14 30
x122 = 0 x222 = 1 x322 =
77
x132 = 0 x232 = 0 x332 = 0
x142 = 0 x242 = 0 x342 = 0
x152 = 0 x252 = 0 x352 = 0
x162 = x262 = x362 = 0
12.4 0.6

j=1 n=1 x111 = j=2 n=1 x211 = 0 j=3 n=1 x311 =
53 25
x121 = 0 x221 = 0 x321 = 0
x131 = 0 x231 = 0 x331 = 0
x141 = 0 x241 = x341 = 0
110
x151 = 0 x251 = 0 x351 = 0
x161 = 0 x261 = 0 x361 =
130
n=2 x112 = n=2 x212 = 0 n=2 x312 =
14 22
x122 = 0 x222 = 0 x322 = 0
x132 = 0 x232 = 0 x332 = 0
x142 = 0 x242 = x342 = 0
33
x152 = 0 x252 = 0 x352 = 0
x162 = 0 x262 = 0 x362 =
57
732 L. Fan et al.
(1) Scenario1: R = 1 × 106 Tones

In this scenario, the allowance is controlled by the government very strict.
Conch cement company should take action to decrease the carbon emission.
The optimal cement price are 507.2 RBM/Tone and 700 RBM/Tone, and other
optimal results are in Tables 5 and 7.
In this scenario, the government allocated enough allowance to the cement
company. Conch cement company can make full use of the resource. The optimal
cement price are 447.29 RBM/Tone and 580.21 RBM/Tone, and other optimal
results are in Tables 6 and 7.
In this scenario, the government allocated more allowance to the cement
company. Conch cement company can make full use of the resource and improve
the cement production at the same time. The optimal cement price are 370.9
RBM/Tone and 550.21 RBM/Tone, and other optimal results are in Tables 7
and 8.
6 Policy Implications
The above results indicates that carbon emission allowance paly a important
role in the cement site selection problem. When the government allocated low
level allowance (i,e. 1 × 106 Tones) to cement company, it will control the output
which just meet the consumers’ demand, decrease the cost by limit the carbon
emissions and make higher sale price; When the cement company gets medium
level (i,e. 4×106 Tones), it will expand the scale suitably, increase the output and
take lower sale price policy to improve the sale quantity; When the government
allocated high level allowance (i,e. 8×106 Tones) to cement company, it will built
more cement plants, take low price policy to arouse consumers extra demand to
sale more cement.
As one of the most contributor to the carbon emission, the cement industry
has different decisions with different allowance. On the one hand, building new
cement plants and exploiting new market with high level allowance. The demand
Table 7. The output decision of cement plant
Cement R = 1 ×106 Tonnes R = 4 ×106 Tonnes R = 8 ×106 Tonnes

Plant i
Cement 1 Cement 2 Cement 1 Cement 2 Cement 1 Cement 2
(104 Tones) (104 Tones) (104 Tones) (104 Tones) (104 Tones) (104 Tones)
i=1 O11 = 165 O12 = 30 O11 = 145 O12 = 20 O11 = 131 O12 = 37
i=2 O21 = 57 O22 = 78 O21 = 0 O22 = 0 O21 = 123 O22 = 28
i=3 O31 = 0 O32 = 0 O31 = 0 O32 = 0 O31 = 118 O32 = 50
i=4 O41 = 0 O42 = 0 O41 = 141 O42 = 40 O41 = 0 O42 = 0
i=5 O51 = 0 O52 = 0 O51 = 0 O52 = 0 O51 = 0 O52 = 0
i=6 O61 = 90 O62 = 13 O61 = 148 O62 = 62 O61 = 80 O62 = 100

j=1 n=1 x111 = j=2 n=1 x211 = 0 j=3 n=1 x311 =
100 45
x121 = 0 x221 = x321 = 0
123
x131 = 0 x231 = 0 x331 =
118
x141 = 0 x241 = 0 x341 = 0
x151 = 0 x251 = 0 x351 = 0
x161 = 0 x261 = 0 x361 =
50
n=2 x112 = n=2 x212 = 0 n=2 x312 = 0
20
x122 = 0 x222 = x322 = 0
28
x132 = 0 x232 = x332 =
20 30
x142 = 0 x242 = 0 x342 = 0
x152 = 0 x252 = 0 x352 = 0
x162 = 0 x262 = 0 x362 =
80
of cement will increased rapidly with the expanded of population scale, the
cement company need to produce more cement to meet consumers’ demand; On
the other hand, the cement industry will control the output or do more research
to exploit new technology to decrease carbon emission in the cement produce
and transportation processes. Otherwise, if the cement company increase the
output without new technology, it will face huge emission cost.
7 Conclusion
In this paper, a bi-level programming model with fuzzy random variables is pro-
posed to deal with the cement plants site selection problem with carbon emission
trading mechanism. Getting total profits maximum and making the total cost
minimum are the cement company (i.e. the upper) and the consumers (i.e. the
lower) targets, respectively. In contrast to the previous studies, the research takes
the natural release and energy consumption which includes produce and trans-
portation process as the total carbon emissions from the cement company. To
solve the complex programming model, a KKT convert technique and a multi-
cut points-based genetic algorithm are taken as a solution method. Finally, the
results of the case study in Conch cement plants site selection problem are proved
the efficiency of the optimization model and method.
734 L. Fan et al.
Acknowledgements. We are thankful for financial support from the Research Center
for Systems Science & Enterprise Development, Key Research of Social Sciences Base
of Sichuan Province (Grant No. Xq15C01), the National Natural Science Foundation
of China for Younger Scholars of China (Grant No. 71601134) and Project funded by
China Postdoctoral Science Foundation.
References
1. Ataei M (2005) Multicriteria selection for an alumina-cement plant location in East
Azerbaijan province of Iran. J S Afr Inst Min Metall 105(7):507–513
2. Benhelal E, Zahedi G, Hashim H (2012) A novel design for green and economical
cement manufacturing. J Cleaner Prod 22(1):60–66
3. Benhelal E, Zahedi G et al (2013) Global strategies and potentials to curb co2
emissions in cement industry. J Cleaner Prod 51(1):142–161
4. Kim SH (2015) Ventilation impairment of residents around a cement plant. Ann
Occup Environ Med 27(1):3
5. Lamas WDQ, Palau JCF, Camargo JRD (2013) Waste materials co-processing in
cement industry: ecological efficiency of waste reuse. Renew Sustain Energy Rev
19(1):200–207
6. Madlool NA, Saidur R et al (2013) An overview of energy savings measures for
cement industries. Renew Sustain Energy Rev 19(1):18–29
7. Schneider M, Romer M et al (2011) Sustainable cement production present and
future. Cem Concr Res 41(7):642–650
8. Stefanovi GM, Vuc̆kovi G et al (2010) Co2 reduction options in cement industry -
The Novi Popovac case. Therm Sci 14(3):671–679
9. Strazza C, Borghi AD et al (2011) Resource productivity enhancement as means for
promoting cleaner production: analysis of co-incineration in cement plants through
a life cycle approach. J Cleaner Prod 19(14):1615–1621
10. Usón AA, López-Sabirón AM et al (2013) Uses of alternative fuels and raw mate-
rials in the cement industry as sustainable waste management options. Renew
Sustain Energy Rev 23(4):242–260
11. WBCSD (2002) The Cement Sustainability Initiative: Our Agenda for Action.
World Business Council for Sustainable Development (WBCSD), Switzerland
Representation and Analysis
of Multi-dimensional Data to Support Decision
Making in the Management
Shavkat Ayupov1 and Abdulla Arifjanov2(B)

1
Institute of Mathematics, Academy of Sciences of the Republic of Uzbekistan,
Tashkent, Uzbekistan
2
Center for the Development of Software Products and Hardware-Software
Complexes at the Tashkent University of Information Technologies,
Tashkent, Uzbekistan
arifjanov@yandex.ru
Abstract. The paper considers questions related to the formal descrip-

tion of concepts used in the multivariate analysis of data, as well as to
the set of operations on multidimensional cubes structuring the data. Dis-
cussed are questions in regard of the effectiveness of a class of multidimen-
sional data models, and searching ways to solve the problem of “explosive”
growth in the volume of data in a multidimensional OLAP-analysis.
Keywords: Data aggregation · Data hypercube · Data mining ·

Data selection · Data warehouse · Multidimensional representation of
data · OLAP (On-Line Analytical Processing) · Operations in data
manipulation
1 Introduction
An effective control of any production process, technical maintenance, or social

configuration is not possible without taking timely, well-weighted adequate deci-
sions. The complete work on such decisions requires the responsible teams and
leaders to analysis large volumes of information, and as a rule, in conditions of
rush and time limitations.
In the data analysis and decision support three basic concepts are used:
• The coding and structure of data base to ensure flexible navigation through
the database, targeting generation of ad hoc queries and presentation of the
results in the form of various reports. Their display should be based on the
topological realization of algorithms for spatial analysis;
• The multi dimensional data manipulations to allow an easy organization of
aggregated information from the storage in the self-saved form of a hyper-
cubic model. This model would provide comfortable visualization and analysis

736 S. Ayupov and A. Arifjanov
of data. Simultaneously, in the cells of a hypercube where numeric values are

stored as aggregated indicators, the measurements to allow the organization
of data in accordance to its chronology, geography and to other classifications,
based on the reference storage;
• The search of dependences within the accumulated information to allow the
use of algorithms for intellectual data analysis.
The concept of interactive multidimensional data analysis OLAP (On-Line

Analytical Processing) was proposed in 1993, by Codd et al. [2], one of the
ideologists of the relational databases. It was in the form of 12 basic rules that
must be satisfied by the OLAP-systems, as well as the products providing the
ability to perform online analytical processing. These rules can be considered as
a theoretical basis for online analytical processing. In 1995, based on the Codd’s
et al. set of requirements, Nigel Pendse formulated the so-called Fast Analysis of
Shared Multidimensional Information (FASMI) test, defining requirements for
the applications of multidimensional data analysis.
OLAP-technology is an actual and essential topic of research. Its practical
results are widely applicable. In this regard, the research analysis and presenta-
tion of multidimensional data have lately become a priority in the development of
methods for designing information systems and systems for information analysis.
The relevance of operational support systems for multidimensional data analysis
has been around, and there are already many publications in this field. Despite
a long history of research, materials on the formalized description of OLAP-
systems are very few. In general, the available literature concerning this subject
is devoted to the concepts, technology of OLAP-system, as well as to some
practical issues. There can be found a lot of information on OLAP-systems,
but virtually nowhere is written and said about how it works internally. Due
to this reason, the solution of most problems in the construction of OLAP-
systems is mainly carried out by the method of trials and errors. Currently,
there is virtually no rigorous, unified mathematical foundation for systems of
multidimensional data storage and analysis, such as that which exists for rela-
tional database management systems (DBMS). Little work is available on the
formal description of the mechanism of aggregation in OLAP-cubes. Until now,
there is no single terminology standard about data transmission, data query
language, and about the formation of OLAP-cubes. In the existing publications
it is mostly proposed multi-dimensional models describing the subject studied
domains. Typically, these models were built for specific tasks. Moreover, their
authors are using different terminology in description of various aspects of the
subject area at different stages of design and data analysis. In order to be able
to compare and analyze these works, we must have a unified system for classifi-
cation of multidimensional models.
A rigorous mathematical description of the concept of multidimensional data
and methods for constructing multivariate models provide several advantages.
First and foremost, is to simplify the design process of multidimensional models,
just as the normal form simplifies the design of relational models. Mathematical
description of multidimensional operations can help to build simple and clear
Representation and Analysis of Multi-dimensional Data 737
requests at the data analysis phase. Therefore, the development of a mathemati-

cal model for multi-dimensional representation of data and methods of analysis,
is currently very topical problem. This article attempts to fill this gap.
Proceeding from the aforesaid, the purpose of this paper is: Firstly, to provide
a formal description of concepts used in the multivariate analysis of data, as
well as many operations on the multidimensional cube, by structuring the data;
Secondary, to offer a study of the efficacy of a class of multidimensional data
models, and to help in finding ways to solve the problem of “explosive” growth
in the volume of data in multidimensional OLAP-analysis [1,2].
2 Basic Concepts of Multidimensional Data Model

Before proceeding to a formal description of a multidimensional data model, we
adopt the basic concepts of multidimensional data model. Until now, there is no
generally accepted point of view for the multidimensional models of data. Despite
the absence of fundamental differences in existing views, however, there are some
differences in terminology. We follow the terminology described in [1,3,5–7].
Basic concepts of multidimensional data models are:
• Data Hypercube is the entity of stored information;
• Dimension: a set of objects of one or several types organized in a hierarchical
structure, which provides information context for a numeric indicator. Dimen-
sions are usually visualized as the edges of a multidimensional cube;
• Facts: components of the measurements, tags (sequence of characteristics).
These are objects, the totality of which forms the measurement. Facts are
visualized as points or areas, located on the axes of the hypercube. These
objects should be organized in a hierarchical structure so that the objects of
the same type belong to only one level of hierarchy;
• Cells: atomic structures of the cube corresponding to particular values of some
indicators. Cells in the visualization are inside the cube and are used to display
the corresponding values of the indices;
• Indicator (Measure): a characteristic (usually numeric), which itself can be
subject of analysis. An OLAP cube can have one or more indicators.
Data hypercube may contain one or more dimensions and is an ordered
set of cells. Each cell is determined by the one and only one set of values of
measurement–the facts. A cell can contain data - some measure or can be empty
(Fig. 1).
In typical applications of OLAP-systems, it is usually used a “basic” attitude,
which is called as the table of actual values (in short a fact table). Attributes of
this attitude can be perceived as separate measurements of the multidimensional
space in the form of a measurements table, or cube the original data (further
called the hypercube), and each train - as a point in this space. At the intersection
of the axes of measurement are located the data, quantitatively characterizing
the analyzed facts C the measures. Data cube (or OLAP-cube) is different from
the original data cube in the following aspects:
Fig. 1. An example of a data hypercube
• Along with the original data it contains information, aggregated over all sub-
sets of measurements;
• Points in the data cube can be aggregated versions of the starting points of the
original data. For example, the dimension of a “quarter” of the cube of initial
data that determines the information about the progress for each quarter of
a year may be reduced to the dimension of “half of a year”.
3 A Mathematical Model of Multi-dimensional

Representation and Data Analysis
OLAP systems represent software whose construction ideology is open towards
analysis of large volume of structured information. Technology and application of
class OLAP is focused on online analytical processing of multidimensional data
sets, used to analyze the current and predict future activities of object man-
agement to managerial decision-making. OLAP technology simplifies the work
with multi-purpose accumulated data, on the activities in decision making from
the past with their extremely high volume. It also transfers sets of quantita-
tive indicators into qualitative ones which are easy to understand. The system
allows analysts, managers, and responsible personnel to form their own view on
the data, using a fast, consistent, rapid access to diverse forms of information.
These forms, based on primary data, allow the user to get a full picture of the
activities which are subject of the control.
OLAP technology is an alternative to traditional methods of data analysis
based on various systems implementation of SQL-queries to a relational data-
base. OLAP systems play a crucial role in analysis and planning activities in the
subjects of control. The analysis of data is mainly based on the requirements
of decision makers to the information provided, traditionally existing individual

characteristics of doing business in the objects of control, and decision-making
mechanisms. From a user perspective the main difference between the OLAP
systems and the Data Warehouse (DWH) systems is not technical, but in the
subjected structure of the information.
The hallmark of OLAP systems is the representation of data in the form
of multidimensional cubes (see e.g. [1,4]. All information about the business
process is systematized, divided into categories, which, in turn, are converted to
axes (measurements) of the cubes. Such an approach allows us to provide all the
data in the form of two, three, or more dimensional sections of these cubes. The
user can choose the sections on those or other categories, which allow combining,
comparing and analyzing information in many possible directions. Moreover,
each marginal direction can in turn be divided into sub-directions depending on
the type of information.
Cube, this is a multidimensional database that stores two kinds of fields:
measurement and evidence. Cube has a hierarchical structure which is optimized
to be instantly used for immediate excerption of aggregated facts at the inter-
section of values (or components) of the measurements. The cube can contain
pre-computed aggregates or provide special sampling data to calculate desirable
units in real time. Its design provides instant performance and manipulations:
data filtering, searching, changing the rules of aggregation, and more. In the
sequel we consider OLAP features in detail.
Formal descriptions of terms used in the multivariate analysis of data, as well
as many operations on multidimensional cubes structuring the data, we consider
on the example of the data hyper-cube shown in Fig. 1.
Let consider first the algorithms for constructing a multidimensional data
model to create a multidimensional database and data analysis using the mech-
anisms described in [1,3,5–7].
On Fig. 1 are depicted: H(D, M ) is the hypercube itself, i.e. plurality of cells;
V (H) is the set of measures; D = {d1, d2, · · · , dn } is the set of measures; Mdi =
{m1i , m2i , · · · , mki } is the set of measurements; - the set of labels pertinent to
measurement di ; M = Md1 ∪ Md1 ∪ · · · ∪ Mdn is the whole set of labels; D ⊆ D
is a set of selected fixed dimensions; M ⊆ M is a set of selected fixed labels;
H (D , M )|H ⊆ H is a slice selected from the hypercube by the use of the sets
D and M .
In order to gain access to the data, the user must specify one or more cells by
selecting the measurement values (indicators), which correspond to the desired
cells (the labels). This process of choosing some wished values of measurement
is called fixing the facts, and the set of selected values of measurements - a set of
fixed facts. The set of cells corresponding to the fixed labels and measurements
is denoted as slice of the hypercube.
Hence, let D = {d1, d2, · · · , dn } be the set of dimensions of the hypercube;
Mdi = {m1i , m2i , · · · , mki } is the set of labels for the measurement;
M = Md1 ∪ Md1 ∪ · · · ∪ Mdn is the set of facts contained in the hypercube;
D ⊆ D is a set of fixed dimensions;
M ⊆ M is a fixed array of facts.
The data hypercube consists of the array of all cells H(D, M ), corresponding
to the sets D and M .
A subset of H(D, M ), corresponding to some subsets of fixed values D , M .
We denote as H (D , M ).
Each cell h of the data hypercube H(D, M ) corresponds to only one possible
set of facts measurement Mh ⊂ M . The cell can be empty (contains no data),
or contains a value index - a measure.
The set of all measures included in the hypercube H(D, M ) is denoted by
V (H).
In a data hypercube one can perform either of the following operations (or
manipulations) with data:
• operation “slicing”;
• operation “rotation”;
• operation “convolution and detail”;
• operation “Data Aggregation”;
• operation “sampling”.
Consider these operations.

(1) The operation “slicing”
Any subset H (D , M ) of the hypercube, obtained by fixing the labels from one
or more dimensions, is called a slice. The operation of constructing the slice
is carried out to obtain the desired subset of cells H ⊂ H, and cutting the
“unnecessary” values by successively fixing the needed tags. The slice can be
interpreted as a two-dimensional array (a two-way table).
The label mji ∈ M specifies the hyper-plane section of a data hypercube,
corresponding to an appropriate measurement di ∈ D. The set of fixed labels
M ⊆ M , thus defines a set of hyperplane intercepts (dissections) of the data
hypercube corresponding to the set of fixed dimensions D ⊆ D. The common
intersection of these hyper-planes defines a set of cells H (D , M ) is the slice of
the data hypercube which is of interest to the user. The essence of the process
of retrieving data from a data hypercube, thus consists in constructing a slice-a
sub-hypercube of data H (D , M ), by specifying the sets D , M . Sometime, the
result of this operation is called dice, and the operation itself is called dicing.
(2) The operation “rotation”
Any change in the order of presentation (imaging and visualization) of the mea-
surements is called rotation. The rotations allow visualization of the data in
forms, most comfortable for their perception.
(3) The operation “Data Aggregation” (Drill Up, consolidation, roll-up)
Data hypercube generally has hierarchies or formula-based relationships of data
within each dimension. Aggregation involves computing all of these data rela-
tionships for one or more dimensions. While such relationships are normally sum-
mations, any type of computational relationship or formula might be defined.
This means to get the values corresponding to the labels from a certain higher
level l, l1 of the hierarchical dimension D based on the values from the previous
lower hierarchical level l − 1.
Consider the hierarchical dimension D with L levels (Fig. 2). The primary
data (events, facts) correspond to the lower level of the hierarchy (l = 0). Com-
putation of aggregates is made in accordance with the applicable method of
aggregation. In the case of summation, the values Xj1 of the units at the level of
Mj 0
the hierarchy l = 1 can be calculated by the formula Xj1 = i=1 xi , where Mj
is the number of values corresponding to the facts appearing subsidiaries to the
label j.
Fig. 2. Aggregation of data hypercube. One-dimensional representation
Summarizing, we obtain formulas for calculating the aggregated units accord-

ing to the method of summation at other levels of hierarchy

Mj
Xjl = i , l = 1, · · · , L, j = 1, · · · , Nl .
xl−1
i=1
The number
L of aggregated units in a data hypercube for a single measurement
is NA = i=1 Ni .
Generalizing, for the case of an arbitrary number of dimensions D, for the
number of aggregated units we obtain

D
Lj

D
NA = Nij − N0j ,
j=1 i=1 j=1
where Nij is number of labels the ith level of the hierarchy of measurements j,
l = 1, · · · , D, and Lj is the number of hierarchical levels of measurement j.
(4) The operation “convolution and detail” (Drill Down)
Drilling down is a specific analytical technique, where the user navigates among
levels of data ranging from summarized data at higher level to more detailed
data, located at lower level. This operation is carried out thanks to the hierar-
chical structure of the measurements. Readings (events, measurements) may be
combined in a hierarchy, consisting of one or more levels. For example:
Day ⇒ Month ⇒ Quarter ⇒ Year;
Manager ⇒ unit ⇒ Region ⇒ Company ⇒ Country;
Model Car ⇒ The manufacturing plant ⇒ Country.

When going down to more and more detailed information the user is actually
doing drill down.
(5) The operation “sampling”
This is a process of constructing a query to retrieve data from a data hypercube.
It can be presented as a model of the network graph.
Consider the network graph G(S, V ) on Fig. 3, whose vertices correspond
to the facts in the data hypercube H(D, M ). The set S of vertices consists
of n subsets, or layers Si ⊂ S. Among these are the “steps” of the user (a
pattern) with which will be restricted the facts needed for the sample. The layer
Si , i = 1, · · · , n consists of specific vertices Sji , ji ∈ {1, · · · , |M |}, corresponding
to facts that will be selected by the user on the ith step.
Fig. 3. Query construction
i
The set of edges of the graph, V , is a set of pairs of vertices vjk = {sji , ski+1 }.
The edge vjk ∈ V of the graph, thus connecting the j − th vertex Sji from the
i
layer Si with the k − th vertex Ski+1 from the layer Si+1 .

The procedure for constructing a user request to retrieve data from a data
hypercube H(D, M ) can be represented as a sequence of steps, at each of which
the user captures a fact mdi ∈ M from the measurement di ∈ D. By fixing
facts in desired dimensions, the user gets access to the cell corresponding to
recorded and requested facts. The sequence of fixing a series of facts can be
presented on the network graph G(S, V ) as a sequence of edges (paths) w =
{vjk
0 1
; vjk , · · · , vjk
n
}, reflecting the process of formation of the user’s query.
The initial vertex on the graph S00 corresponds to the beginning of the
procedure of forming a user’s query and to the absence of any fixed facts. In
the first step (n = 1), the user fixes a fact mdi ∈ M from the measurement
di ∈ D of the data hypercube H(D, M ), and this corresponds to the vertex vji
on the layer S1 of the graph G(S, V ). In the second step, the user captures a fact
mdi ∈ M \Mdi from the measurement dj ∈ D\di , since the measurement di has
already been fixed. This procedure is performed consecutively n times, where n
is the number of measurements needed to be fixed to obtain the required data
slice (in general n = |D|).
The final vertex S0n+1 corresponds to the end of the procedure of forming a
request and obtain access to the cell h(w) of the data hypercube. Here w is the
path selected by user’s query on the graph G(S, V ).
In a diluted (scarce) data hypercube the requested cell h(w), generally speak-
ing, may be empty. Thus, time spent on the formation of a user query will be
“wasted”, because this particular query result is empty. The problem of optimiz-
ing the formation of user’s queries by using step-by-step fixing facts measurement
of the hypercube H(D, M ), is thus reduced to finding the set of paths Wtrue on
a network graph G(S, V ), leading mainly to non-empty cells [7].
Unfortunately, the use of the described multidimensional data models and
operations in the hypercube has shown poor performance on large volumes of
data. For example, if the hypercube contains information about the sales in one
year, and if it has only 3 measurements–Customers (250) Products (500) and
Dates (365), we obtain a matrix of facts of the size 250 × 500 × 365 = 45625000.
The total number of non-empty cells in the matrix can be practically only a
few thousand. Moreover, the more the number of dimensions, the more tenuous
will be the matrix. Therefore, to work with such matrix it is preferable to create
and use special methods of processing data in sparse matrices. To solve this
problem, it is possible by preliminary “cleaning” the data before to use them to
build the cubes. But this approach is not always applicable. Another drawback is
that the choice of a higher level of minuteness in the creation of a hypercube can
greatly increase the size of the multidimensional database. Because of these, and
some other additional reasons, commercially available multidimensional database
management systems are not able to handle large amounts of data. It is advisable
to use this multivariate model, when the size of the database is small and has a
homogeneous in time set of measurements.
In contrast to the multidimensional databases, the relational management
systems of databases are capable to store huge amounts of data. But they lose
their advantages when compared in the speed of execution of analytical queries.
When using relational database management systems, the principles for data
storage are organized in a special way. The most commonly used so-called radial
(or “star”) pattern. In this scheme, two types of tables are used: (1) Fact table
and (2) Several reference tables (dimension tables). The fact table usually con-
tains data, the most intensively used for analysis of data in the cube. If we draw
an analogy with the multivariate model, a record in the factual table corresponds
to a cell in the hypercube. In the reference tables are listed all the possible values
of each one dimension of the hypercube. Each dimension is described by its own
lookup table. Fact table is indexed on a complex key that is built from individual
keys help reference tables. It provides a link-up table with the factual by suitable
sets of key attributes.
To shorten the response time in an analytical system one can use some special
tools. In the composition of powerful relational database management systems
there are typically included query optimizers. When creating a data warehouse
based on relational database management systems, their presence is of particular
importance. Optimizers analyze the request and determine the best position, in
regard to some criterion, of the sequence of operations to access the database for
this specific request implementation. For example, in this way one can minimize
the number of physical disk accesses for the query. The query optimizers use
sophisticated algorithms for statistical processing, which operate on the number
of entries in the tables, ranges of keys, etc.
Each of the models described above has both, advantages and disadvantages.
Multivariate model allows implementation of fast analysis of data, but cannot
store large amounts of information. Relational model, by contrast, has virtually
no limit on the amount of accumulated data, but such database does not provide
the desirable speed in the run of analytical queries, as the multidimensional
database.
4 Ways to Solve the “Explosive Volume Growth”

Problem in Multidimensional OLAP
Data warehouses and OLAP analytical system are designed for analysis and
synthesis of detailed data. So, they can include a full range of detailed data
contained in database sources, and may increase the amount of data as a result
of denormalization and duplication of detailed data. In addition, the amount of
data can grow when aggregated information is added to the detailed data. This
explains the rapid growth of data in the OLAP structures, as compared with
the rate of increase of the volume in source databases. Often to describe this
phenomenon, the experts are using the term “explosive” nature of increase of
data in hypercube.
Accelerated growth of data volumes dramatically reduces the performance
of multivariate analysis and limits its scalability. We have investigated various
strategies to improve query performance in relational data warehouses and ana-
lytical OLAP-systems. We tried to achieve the rapid growth of the processed
data and typical compact solutions without built-in mechanisms to improve sys-
tem’s performance and scalability. Considered were the following strategies:
• Distributing the data according to its relevance of use;
• Partitioning the storage fact table and the hypercube structure into sections
and into sub-cubes of smaller size;
• Organization of parallel query processing.
The results show that the largest integrated effect on improved performance
and scalability provides a combined strategy. It combines the strategies of parti-
tioning and parallel processing of sections. While the available storage capacity
increases in proportion to the number of sections in the fact table, the query
performance increases by several times. It was found that the specificity of the
multivariate analysis cannot directly transfer the methods for partitioning fact
tables on the hypercube. This is due to the difference in the processes of indexing
data in hypercube, and these in relational tables.
As noted above, OLAP-technology has emerged as an alternative to the tradi-
tional methods of data analysis based on limited and SQL-queries to a relational
database. The main advantage of OLAP lies in the wide possibility of forming
ad-hoc queries for analytical databases. However, our studies show that from a
theoretical point of view, unlimited opportunity to form ad hoc queries to OLAP
hypercube when working with the above model, under certain conditions can lead
to complex problems, associated with ensuring the integrity of the multidimen-
sional data. A similar problem arises, for example, when using the methods of the
additive decomposition of a hyper-cube’s structure into sub-cubes of smaller size.
In such cases an ad hoc query, being addressed to the complete hypercube (as a
whole), first processes the sub-cube, and the results are then consolidated into
the overall result. With ad-hoc queries indexing hypercube becomes arbitrary,
varies from request to request, and may contribute to errors in the procedure of
consolidation in the settlement of the results from sub-cubes. We have investi-
gated the issue and presented a new method for its solution. It eliminates data
corruption, despite the non-additive indexing of additively decomposed hyper-
cubes. The conditions for indexing data in sub-cubes were analyzed. An assertion
was stated, that allows to identify any subset of undistorted data sub-set Xsec
and the potentially faulty data subsets X’sec at the intersections of indices in
the following form
n−1

Xsec ≡ X\Xsec = X\ U ( U (Xi IXj )),
i=1 i<j≤n
here n is the number of decomposed sub-cubes, and X is a notation for the

original hypercube:

X = Xsec ∪ Xsec ; (Xsec |Xsec = ∅).
Fig. 4. The structure of the algorithm of multidimensional data aggregation

In accordance with the obtained by us conditions for the algorithm of multi-

dimensional data aggregation, the overall structure is shown in Fig. 4.
The developed algorithm is tested in the solution of specific applications. It
allows the use of a convenient way for additive decomposition of hyper-cube’s
structure in cases of ad hoc requests from users, resulting in non-additive index-
ing of the data in the hypercube.
5 Example
As an example, consider the formation of OLAP-cube based on existing relational
fact table sales from managers, M1, M2 and M3 cars Nexia, Tico and Damas in
the years of 2007, 2008 and 2009. To do this, it is needed to select the values of
all measures and carry out the aggregation of all the obtained values, and record
it in the cube. Consider the application of the obtained models on the example
of a cube with 3 dimensions (Fig. 5).
Here measurements are: the model made of cars (products, a categorical
variable), the sale person (manager, another structural attribute), and the year
of production (time dimension), i.e. this is the triplet C, M, T ;
Facts are: Model car (Nexia, Tico, Damas), labeled as (N, T, D); Managers
(Manager 1, Manager 2, Manager 3) labeled as (M 1, M 2, M 3); The time (year)
recorded (2007, 2008, 2009) labeled as (Y 4, Y 5, Y 6);
Measures are: the volume of sales (7, 7, 4, etc.).
Fig. 5. A cube with three dimensions
In our example we have:

The set of dimensions of the hypercube is D = C, M, T ;
The set of labels for the measurement di : MC = {N, T, D}, MM =
{M 1, M 2, M 3} and MT = {Y 4, Y 5, Y 6}.
In this example each value in the field of sales is uniquely determined by a

combination of fields: Model car, Manager, Year of sale. The indicator is the
sales volume, located in the cell of the intersection of specified dimension of each
kind.
Consider the application of operations to manipulate dimensions:
(1) Formation of a “Slice”
If we limit the value “dimensions of a model car” = “Nexia GL”, then we obtain
a subset of the hypercube (in our case - two-dimensional table) containing infor-
mation about the history of sales of this model by different managers in different
years. This is the content on the left wall of the cube on Fig. 5.
(2) Operation “Rotation”
This operation provides the ability to visualize data in a manner most comfort-
able for its perception. Rotating consists of swapping the rows and columns. For
example, by changing the Measures of Products down on front X-axis by the
Time measure across the side Y -axis into a report, an entirely new vision of the
data hypercube on Fig. 4 will be obtained. This will be equivalent to a rotation
in 90◦ counter clock wise along the Z-axis (the managers).
(3) Operation “Aggregation”
To get the sales for each car model and for each year, we need to accumulate the
numbers in the cell of each column for all the managers. To get the total sales
of each manager throughout the years we need to add his/her results for each
car model for all years. This is aggregation.
From user’s perspective, in a larger dimension hierarchical model Division,
Region, Company, Country is exactly the same kind of measurements as the
manager is. But each of these added dimensions would correspond to a new,
higher level of aggregated values of the index “volume of sales”. In the process
of analysis the user works not only with a variety of data slices and performs
their rotation, but also moves from detailed data to aggregated, i.e. applies the
operation Aggregation. For example, after noticing how successfully Manager 1
sold the models of “Nexia” and “Tico” in 2007, his supervisor may want to see
how looks its relationship with the sales of these models at the unit level, where
Manager 1 works. And after that the supervisor may want to obtain a similar
certificate for the Region or firm.
(4) Operation “Detailization”
The transition from aggregated to more detailed data is the operation “Detailiza-
tion”. For example, beginning the analysis at the level of total sales through all
the years, the supervisor may wish to obtain more precise information about
specific local sale in a year and by a M anager.
(5) The “Sampling”
To construct a query to retrieve data from the data hypercube presented on
Fig. 4 a user need to select a sequence of addresses (vertices of an associated
graph) of cells whose content are required in the sample. For instance, the choice
(N, M 2, Y 5) ⇒ (T, M3, Y4) ⇒ (T, M 1, Y 6) will retrieve the content of the
corresponding cells on the display in this order.
6 Conclusions
It is shown that the use of the conventional multidimensional data models and
operations in the hypercube does not provide the desired speed performance
on large volumes of data. This problem can be solved by pre-treatment of data
before to use them to build the data cubes. Another drawback is that the choice
of a high level of detail in the creation of a hypercube can significantly increase
the size of a multidimensional database. Because of these and some other rea-
sons, commercially available multidimensional database management systems
are not able to handle large amounts of data. It is advisable to use a standard
(conventional) multivariate model, if the size of the database is small and has a
homogeneous set of measurements. Largest integrated effect of improved perfor-
mance and scalability provides a combined strategy that combines partitioning
and parallel processing sections. Then the available storage capacity increases in
proportion to the number of sections in the fact table, and query performance
increases by several times.
It was found that the specificity of the multivariate analysis cannot imme-
diately transfer the methods for partitioning the fact table on the hypercube.
This is due to the difference in the processes of indexing data for hyper-cubes,
and for relational tables.
References
1. Arifjanov AS (2008) Algorithms for multidimensional reporting and analysis of data.
Uzbek J Probl Inf Energ 8(1):65–72
2. Arifjanov AS (2014) Some questions in mathematical modeling of multidimensional
data reporting and analysis. In: Book of abstracts. Flint international statistical
conference “FLINT:1 City Hundred Years under Variability”
3. Ayupov S, Arifjanov A (2017) Information–analytical technologies of decision sup-
port in management of power systems. Springer, Singapore
4. Codd EF, Codd SB, Salley CT (1993) Providing OLAP (on-line analytical process-
ing) to user-analysts: an it mandate, vol 32. E.F. Codd & Associates, pp 12–20
5. Hrustalev EM (2013) Aggregation of data in olap-cubes. http://www.olap.ru
6. Sakharov AA (2017) The concept of building and implementing information sys-
tems, with a focus on data analysis. http://olap.ru/basic/saharov.asp (in Russian)
7. Zabotnev MS (2017) Methods of presenting information in a sparse data hypercube.
http://www.olap.ru
Fuzzy Logic Applied to SCADA Systems
Tahar Benmessaoud1,2 , Alberto Pliego Marugán2 , Kamal Mohammedi1 ,

and Fausto Pedro Garcı́a Márquez2(B)
1
Laboratory of Energetic-Mechanic and Engineering, FSI, M’hamed Bougara
University of Boumerdes, Boumerdès, Algeria
2
Ingenium Research Group, Universidad Castilla-La Mancha,
13071 Ciudad Real, Spain
FaustoPedro.Garcia@uclm.es
Abstract. This article focuses on the monitoring of a wind farm in

real time based on big data collected by Supervisory Control and Data
Acquisition (SCADA) system. The decision-making of the type of main-
tenance to be applied can be insured by SCADA system. This system
generates alarms based on the collected data. False alarms cause false
interventions by the maintenance team resulting in loss of production
and costs. The reduction of these false alarms makes it possible to con-
tribute better to the management of the maintenance of the wind farm.
In this paper, we propose a new approach for the identification of alarms
by Fuzzy Logic based on the data collected by the SCADA system. The
alarms generated in this case can be divided into two categories: orange
alarms corresponding to faults requiring the intervention of preventive
maintenance and red alarms corresponding to critical states that can
cause system failures.
Keywords: Wind farm · Monitoring · SCADA · Alarms · Fuzzy logic

approach
1 Introduction
The generation of electricity through wind energy systems is demanded by the
industry in order to increase the competitiveness of the business. The main-
tenance of these systems becomes very complex because the stochastic city of
the random loads can cause catastrophic failures [3,12]. The optimization of the
maintenance planning is a crucial factor for the efficiency of the wind farms [15].
With this purpose, wind turbines (WT) can be monitored by SCADA or con-
dition monitoring systems to detect failures [13,14], risks and to take necessary
actions online [8]. Several studies have proved the efficiency of this systems in
the WTs [9] and other types of industries [7,13,19].
False alarms generated by SCADA system are a important problem because
they cause unnecessary stops, false interventions by the maintenance team, loss
of production and, consequently, extra costs [11,17,18]. Some studies aim to
750 T. Benmessaoud et al.
eliminate false alarms of WTs [5]. Chen et al. [5] present a study on the treatment
of SCADA using an artificial neural network ‘ANN’. Qiu et al. proposes two
methods (sequential and probabilistic) for the analysis of alarms for two large
onshore wind farms considering the defects of pitch and converter systems [20].
This paper proposes a new approach based on fuzzy logic for controlling
wind farms considering the big data collected by SCADA system, and give a
complementary response to strengthen the response of the SCADA. Figure 1
shows the general structure of a fuzzy system [1].
Rule Base
Input
Output
Fuzzification Defuzzification
Interface Interface
Decision-making unit
(fuzzy) (fuzzy)
Fig. 1. Basic configuration of a fuzzy logic system
In the fuzzification part, any numeric value of the input data is converted
into a linguistic value between 0 and 1. The fuzzy system requires functions for
each input data:
• Input (1): Variable (1). Subsets: low, average and maximum.
• Input (n): Variable (n). Subsets: low, average and maximum.
Fuzzy inference is a method that interprets the values in the input vector and,
based on fuzzy rules, assigns values to the output vector [16]. With this purpose,
it is necessary to stablish the rules that will perform the defuzzification process
to provide the output. The Fuzzy rules (IF antecedent THEN consequent) in
expert system are usually is following [10]:
IF Var(1) is A11 and/or Var(2) is A21, · · · THEN y is B1
else
IF Var(1) is A12 and/or Var(2) is A22, · · · THEN y is B2
else
IF Var(1) is A1n and/or Var(2) is A2n, · · · THEN y is Bn
where Var(1), Var(2), · · · , Var(n) are the fuzzy input (antecedent) variables, y is
a single output (consequent) variable, and A11, · · · , A1n are the fuzzy sets [5].
Generally, there are n input variables consisting of 3 fuzzy linguistic variables.
Therefore, a total of 3n rules will provide all possible combinations of the input
variables.
Fuzzy Logic Applied to SCADA Systems 751
Defuzzification is the process of producing a quantifiable result given fuzzy

sets and corresponding membership degrees. There are many types of defuzzi-
fication methods, usually maximum membership and centroid techniques are
used [22].
2 Methodology Proposed
In this paper, a novel methodology is presented to make an original treatment

of the SCADA dataset. The flowchart of the methodology proposed is shown by
Fig. 2.
Fig. 2. Alarm identification flowchart
In the diagram adopted for the identification of alarms, it is integrated the

real data collected by SCADA system with the fuzzy system. With this purpose,
the SCADA data will be preprocessed before of being inputted in the fuzzy
system. This preprocess starts with the evaluation of the simple moving average
(SMA) of the SCADA signals. Then, the difference between the signals and the
SMA is calculated. These differences will be used as input vectors for the fuzzy
system.
Once the new variables have been obtained, the next step is to establish some
control laws for them. The new variables will be analyzed statistically to obtain
two thresholds. These thresholds will separate the graph into three different
regions. Figure 3 shows the definition of the thresholds and the regions (good,
acceptable, unacceptable) considered
This classification allows for determining the member functions of the inputs
of the fuzzy system and generate the fuzzy system. The control of each physical
parameter collected by SCADA can be represented by the difference between
the values of all the peaks and the mean value. The fuzzy inference system is
based on different rules to generate the occurrence probabilities of the alarms in
the output.
35
30
25
Unacceptable
Tbear (ºC)
20
15
Tu
10
Acceptable
Ta
5
Good
0
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Time(10 min intervals)
Fig. 3. Classification for member functions
3 Real Case Study
The data considered for this real case study is obtained from the European
Project entitled OPTIMUS [4]. The database used in this paper come from a
SCADA system that provides different measures every 10 min in the period from
01/01/2015 to 28/03/2015. The system measure 37 physical variables, but in this
paper only 4 parameters have been considered to simplify the example. These
parameters correspond to the speed of main shaft (Vel), vibration at the main
shaft (Vibr), oil temperature (Toil) and bearing temperature (Tbear). Figure 4
shows these parameters.
Fig. 4. Evolution of real data in time
Figure 5 shows the preprocessing of the variable Tbear before of being inputed
in the fuzzy system. The blue line corresponds to the “Temperature of bearing
(Tbear)”. The red line evaluates the moving average (MA) calculated for a period
of 2 h. The black line represents the absolute value of the difference blue and
the red line. This new variable will be named “Difference of Tbear” and noted
(DTbear).
100
Tbear
Average
80 Difference
60
T (ºC)
40
20
0
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
samples
Fig. 5. Preprocessing of SCADA signals
The same procedure will be applied to the rest of variables. The new variables
will be obtained by the following equations:
Dvel =| Vel − SMA(Vel) |,

DVibr =| Vibr − SMA(Vibr) |,
DToil =| Toil − SMA(Toil) |,
DTbear =| Tbear − SMA(Tbear) | .
Consequently, Fig. 4 will be transformed to Fig. 6, considering the new vari-

ables.
Fig. 6. Evolution of real data in time
For stablishing the different thresholds pointed in Fig. 3, a maximum value

have been assigned to each input considering several norms and standards [2,6,
12,21]. These values correspond to:
• max(Rotor velocity) = 20.2 rpm.

• max(Vibration) = 115 µm, max(Temperature of oil) = 95 ◦ C.
• max(Temperature of bearing) = 110 ◦ C.
Thus, some of the rules defined in this example are:

• IF DVel is good and DVibr is good and DToil is good and DBear is good
THEN the output is green.
• IF DVel is good and DVibr is acceptable and DToil is unacceptable and DBear
is Good THEN the output is red.
• IF DVel is unacceptable and DVibr is good and DToil is unacceptable and
DBear is unacceptable THEN the output is red.
• If DVel or DVibr or DToil or DBear is unacceptable then the output is red.
More rules of the Fuzzy system are presented graphically in the Fig. 7:
Fig. 7. Rule based Fuzzy Inference System for generation of probabilistic alarms
The red line of each variable shows the value of the variable. The displace-
ments to the left or the right of these lines generates a new position of the
output, and therefore, a new probability of alarm. Once, the fuzzy system is
defined, it is possible to represent surfaces that explain the behavior of the sys-
tem under different conditions. For example, Fig. 8 represents the probability of
alarm depending on DTbear and DVibr.
Fig. 8. Surface view of alarms probability with respect to DTbear and DVibr
It is noted that the probability varies proportionally with the temperature

of the bearings and the vibration. For values of DTbear and DVibr respectively
less than certain value, the probability has very minimal values, in opposite, the
probability of alarm increase exponentially when these values are exceeded.
The system built has been proved by performing a simulation with the
SCADA data. The results of the simulations are shown in Fig. 9. The system
provides a certain probability of alarm regarding to the inputs.
Fig. 9. Probability of alarms generation
A total of 10768 inputs have been analyzed through the created fuzzy system.
The outcomes are:
• 10407 normal measures. This represent the 96.7% of the total data.
• 361 orange alarms. This represent a 3.3% of the total data.
• No red alarms. In the period studied there is not any value exceeding the
limits of critical alarm.
This method allows to transform the data collected by the SCADA system
into probability of alarms. It can be a useful information for complementing the
decision making. The methodology can aid to reduce false alarms because when
a critical alarm arises from the SCADA system, the response of the fuzzy system
can reinforce that alarm.
4 Conclusions
In this paper, a new methodology based on fuzzy logic is proposed in order to
analyze the SCADA data and provide an alternative decision support. The main
purpose is to process the signals collected by the SCADA system from a different
perspective. These signals are converted into new variables to be inputted in the
fuzzy system. A fuzzy system has been created using several standards and
considering the signals of the SCADA system from a statistical point of view.
The creation of the fuzzy systems implies the definition of a set of fuzzy rules.
In this case, the variables considered correspond to the distance of the value
measured by the SCADA to the simple moving average. The more distance to
the average, the more probability of being an abnormal measure, and therefore,
the more probability to generate an alarm.
Three different outcomes of the fuzzy system have been considered. Firstly,
the values are in range of normal behavior and no actions are required. Sec-
ondly, orange alarms where the probability of alarm reaches exceeds a defined
threshold but it is not a critical point. Finally, red alarm when the probability
is unacceptable and the system need an urgent action.
The results of this methodology can become a statistical support for the
generation of alarms. The methodology can be also used as a complementary
information for evaluate the priority of each alarm.
Acknowledgements. The work reported herewith has been financially supported

by the Spanish Ministerio de Economáa y Competitivid ad, under Research
GrantsDPI2015-67264-P and RTC-2016-5694-3.
References
1. Abreu GD, Ribeiro FJ (2003) On-line control of a flexible beam using adaptive
fuzzy controller and piezoelectric actuators. Sba Controle & Automação Sociedade
Brasileira De Automatica 14(4):319–338
2. Acciona (2017) Wind power evolved aw1500 technical specifications. http://
www.acciona-windpower.es/media/1390114/11052015-aw11051500 inusa
abril-11052012.pdf
3. Benmessaoud T, Mohammedi K, Smaili Y (2013) Influence of maintenance on the
performance of a wind farm. Przeglad Elektrotechniczny 89(3):174–178
4. Camacho E, Requena V et al (2014) Demonstration of methods and tools for the
optimisation of operational reliability of large-scale industrial wind turbines. In:
International conference on renewable energies offshore renew
5. Chen B, Qiu YN et al (2011) Wind turbine scada alarm pattern recognition. In:
Renewable power generation, pp 1–6
6. Errichello R, Muller J (1994) Application requirements for wind turbine gearboxes.
NASA STI/recon Technical report N 95
7. Garcı́aárquez FP, Chacón Muñoz JM (2012) A pattern recognition and data analy-
sis method for maintenance management. Int J Syst Sci 43(6):1014–1028
8. González-Carrato RRDLH, Márquez FPG, Dimlaye V, Ruiz-Hernández D (2014)
Pattern recognition by wavelet transforms using macro fibre composites transduc-
ers. Mech Syst Sig Process 48(1–2):339–350
9. González-Carrato RRDLH, Márquez FPG, Dimlaye V (2015) Maintenance man-
agement of wind turbines structures via MFCS and wavelet transforms. Renew
Sustain Energy Rev 48:472–482
10. Khanna V, Cheema RK et al (2013) Fire detection mechanism using fuzzy logic.
Int J Comput Appl 65(12):82–97
11. Marquez FPG (2006) An approach to remote condition monitoring systems man-
agement. In: The institution of engineering and technology international conference
on railway condition monitoring, pp 156–160
12. Márquez FPG, Garcı́a-Pardo IP (2010) Principal component analysis applied to

filtered signals for maintenance management. Qual Reliab Eng 26(6):523–527
13. Márquez FPG, Pedregal DJ, Roberts C (2013) New methods for the condition
monitoring of level crossings. Int J Syst Sci 46(5):878–884
14. Marugan AP, Marquez FPG (2016) A novel approach to diagnostic and prognostic
evaluations applied to railways: a real case study. Proc Inst Mech Eng Part F J
Rail Rapid Transit 230:271–284
15. Marugán AP, Márquez FPG, Pérez JMP (2016) Optimal maintenance management
of offshore wind farms. Energies 9(1):46
16. MathWorks (2007) Fuzzy logic toolbox. https://es.mathworks.com/products/
fuzzy-logic.html
17. Muñoz CG, Márquez FG (2016) A new fault location approach for acoustic emis-
sion techniques in wind turbines. Energies 9(1):40
18. Muñoz CQG, Márquez FPG, Tomás JMS (2016) Ice detection using thermal
infrared radiometry on wind turbine blades. Measurement 93:157–163
19. Muñoz JMC, Márquez FPG, Papaelias M (2013) Railroad inspection based
on ACFM employing a non-uniform b-spline approach. Mech Syst Sig Process
40(2):605–617
20. Qiu Y, Feng Y et al (2012) Wind turbine scada alarm analysis for improving
reliability. Wind Energy 15(8):951–966
21. SKF (2013) SKF spherical roller bearings, setting the standard for performance
and reliability. SKF. http://www.skf.com/binary/143-148465/146100 EN.pdf
22. Zhang P (2010) Advanced Industrial Control Technology
On Causal Analysis of Accident and Design
of Risk-Proof Procedure for Nuclear Materials
Operation: The Case of JCO Accident
Sachiko Oshima(B)
Faculty of Science and Engineering, Waseda University, Tokyo, Japan

oshima.sachiko@aoni.waseda.jp
Abstract. This paper aims to examine a risk-proof procedure for

nuclear materials operation through the causal analysis of JCO acci-
dent in 1999. Nuclear chain reactions are, by now, commonly used in
the nuclear reactors, and thus it seems that there is no basic problem
in fission processes from the scientific point of view. However, the criti-
cality accident that occurred in JCO suggests that one should carefully
examine this accident from the nuclear physics point of view. Indeed the
nuclear chain reactions should have taken place in the small area of sedi-
mentation tank. In fact, when people carry the uranium nitrate solution
into the tank, then this solution with uranium should get into the critical
state at given uranium nitrate solution. The root cause of the accident
should not be very simple from the nuclear physics point of view, and
it should be quite important to examine why the uranium nitrate solu-
tion could have become critical. Based on the analysis result, this paper
also proposes the conditions for the operation without criticality accident
which leads to serious damage of workers’ life.
Keywords: Risk management · Operation system design · Nuclear

fission · Criticality accidents
1 Introduction
In this paper the contents and/or expressions of the previous paper published
by the first author [8] is reused for readers’ convenience and understanding. The
criticality accident that occurred in JCO company in 1999 (called Tokai-mura
JCO-rinkai-jiko) in Japanses must be most serious, and it should not be very
easy to understand why the nuclear chain reactions could proceed in a small
area of sedimentation tank for a finite period of time. In this sense, it should
be quite important to carry out the careful examination of criticality accidents
from the nuclear physics point of view. It should be, of course, difficult to claim
that the JCO accident can be a target of the scientific study as it is impossible
to make the experimental study of the JCO type accidents. However, we believe

On Causal Analysis of Accident and Design of Risk-Proof Procedure 759
that the basic mechanism of the criticality accident should be clarified why it
could naturally occur in the small area.
Here, we explain briefly the theoretical examination of the criticality accident,
which are carried out in terms of the multiple scattering theory, and we show
why the nuclear chain reactions can proceed in the small area [8]. In particu-
lar, the nuclear fission reactions (nucleon-nucleon collision together with nuclear
fission) is traced each by each, and the microscopic processes why and how the
criticality accident occurred is clarified. As a result, this paper provides some
specific reasons why the chain reactions can proceed, and this can be done by
making use of the mean free path which is the result of the nuclear multiple
scattering theory.
When the causal event of this criticality accident is identified, another ques-
tion as to why the criticality stop must be analyzed. This study tries to find
an answer for this question, though not necessarily sufficient. This mechanism
of stopping criticality may be related to the quick settle of the uranium com-
pound. The calculation of this paper tells a possible dangerous situation which
was thought to be due to the eighth batch, if it were carried into the sedimen-
tation tank. The estimated energy release after the virtual eighth batch should
become the same order of magnitude as the Chernobyl nuclear accident.
Finally, a procedure is designed to prevent in advance criticality accidents
from the standpoint of three manufacturing resources that is Human, Machine
and Material [11], grounded in the above physical theoretical discussion.
2.1 Nuclear Chain Reactions
Nuclear fission reaction by incident neutrons can be written as Bohr et al. [1].
n + 235 U → A1 + A2 + (2 ∼ 3)n, (1)
where A1 , A2 are new nuclei which are produced in the reactions. In this reaction,
there are two important points. The first one is concerned with two or three
neutrons which are produced in the reactions. The second point is that the
probability of this nuclear reactions is strongly based on the incident neutron
energy, and the biggest cross section is for the incident neutron with almost zero
energy (thermal energy).
The chain reactions indicate that the produced neutrons should be absorbed
by another 235 U such that the nuclear fission can proceed further on. In addition,
if the chain reactions continue to proceed without the aid of other external
neutron sources, then this situation is called a criticality stage. In reactors, this
criticality must be kept by controlling the number of neutrons involved in the
chain reactions.
In normal reactors, a few percent enriched uranium (235 U exists a few percent
of total amount) should be commonly used, but in this JCO accident, 18.8%
enriched uranium were used, and this high enrichment should be one of the
strong reasons why the nuclear reactions run wild.
760 S. Oshima
2.2 JCO Criticality Accident

This criticality accident occurred when workers in JCO were carrying the ura-
nium nitrate solution (18.8% enriched uranium) into sedimentation tank [5,9,10].
They executed the following procedure. First, they made the uranium nitrate
solution which was composed of 2.4 kg U3 O8 with the nitric acid of 1.7 in the
stainless vessel. In addition, they add water to the uranium nitrate solution until
the total volume became 6.5. Then, they carries the 6.5 solution into the sedi-
mentation tank, and this working procedure was called one batch. The criticality
accident should have occurred in the middle of the seventh batch since the work-
ers noticed blue lights that should be due to the Cherenkov radiation. In fact,
two of the workers suffered from the neutron radiation.
A question should arise as to how the nuclear chain reactions could proceed
within the small sedimentation tank (45 cm diameter, 60 cm high). There are,
of course, some analyses of this criticality accident [10]. However, these studies
are mainly carried out by the computer simulation such that the total energy
emitted via radiations can be reproduced in some way or the other. These investi-
gations are, of course, very important in order to understand the accident cause.
However, it is also important to carry out the study of the criticality accident
from the nuclear physics point of view.
3 Causal Analysis of JCO Case

3.1 Why Criticality?
Now, a question is as to why the criticality is realized in the small area of
the sedimentation tank with 50 of the uranium nitrate solution. That is, why
nuclear chain reactions continue to occur in this small area. Here this section
clarifies the basic mechanism of the criticality accident.
(1) Neutron source
The nuclear chain reactions should require thermal neutrons to start for
the initial fission reactions. Since neutrons should decay within 15 min, they
do not exist as a natural source. Neutrons should be produced in some way
or the other. Here in this accident, the neutron source should be the decay of
238
U spontaneous fissions. The life time of 238 U is about 4.5 billion years and,
in addition, the rate of the spontaneous fission to the total width is around
5.45 × 10−7 . Therefore, 1 gram weight of 238 U make the spontaneous fission of
0.01 times per second. Since one batch contains 1.6 kg of 238 U , we should find
about 20 neutrons per second in the one batch solution.
(2) Mean free path of n − 235 U fission (fast neutrons)
The probability of nuclear fission of 235 U induced by neutrons should be
evaluated in terms of mean free path of λ inside the uranium nitrate solution.
This mean free path of nuclear reactions can be obtained from the multiple
scattering theory as
1
λf = . (2)
ρσf
This derivation of the mean free path (2) is based on the Glauber theory [4], and
this theoretical frame work is well examined in atomic and nuclear reactions [2,3].
Here, ρ denotes the number density of 235 U in solution and σf corresponds to
the nuclear fission cross section of 235 U induced by neutrons. In fact, the number
density of 235 U in one batch solution is ρ ≈ 1.5 × 1020 numbers/cm3 which is a
constant. On the other hand, the nuclear fission cross section σf of 235 U induced
by neutrons crucially depends on the incident energy of neutrons. The incident
energy dependence of the observed cross sections σf can be written as [7].

585b : En ≈ 0.025eV
σf ≈ (3)
1b : En ≈ 1MeV,
where 1b = 10−24 cm2 .

(3) Mean free path of prompt neutrons in nuclear fission
In fission process, the average energy of prompt neutrons is around 1 MeV,
and therefore the average mean free path of the prompt neutrons after fissions
becomes
1
λf = ≈ 67 m. (4)
ρσf
This is quite long in comparison with the scale of the tank, and therefore this
prompt neutrons by themselves cannot induce subsequent fissions in correspond-
ing solution in the tank. In this respect, we ask a question as to why the criticality
should take place within the small sedimentation tank.
3.2 Collision Between Neutrons and Water Molecule

In reality, the prompt neutrons may collide with protons in water molecule, and
they should lose their energy by nucleon-nucleon collisions. Since the nuclear fis-
sion cross sections become largest for the thermal neutrons, the fission processes
should start in case the prompt neutrons lose most of their energy inside the
uranium nitrate solution.
(1) Energy Loss after the Collision of Prompt Neutrons with Protons in Water
When the prompt neutron scatters with protons in water, this neutron should
lose a half of its energy. This can be easily understood in the following way.
First, we denote the incident momentum and energy of the neutron by p, En (=

p2 /2M ), and the final momentum and energy by k, En with En (= k 2 /2M ).
In this case, we find an equation from the conservation law of momentum and
energy as
2
P2 k2 (p − k)
= + , (5)
2M 2M 2M
which can be solved and its solution becomes
k = p cos θ. (6)
762 S. Oshima
Since the observed scattering cross section does not depend on the scattering
angles, we can make an average over the angles, and we obtain the average
energy after the scattering

1 π k2 1 π p2 1
En = dθ = cos2 θdθ = En . (7)
π 0 2M π 0 2M 2
This means that a neutron should lose a half of its energy in each scattering
process.
(2) The Mean Free Path of Neutrons inside Water
Now we calculate the mean free path of neutrons after the scattering with
protons in one batch solution. The number density of protons in one batch
solution is ρp ≈ 4.9 × 1022 numbers/cm3 . The neutron-proton cross section at
low energy is observed as σnp ≈ 20b [6] and thus the mean free path of neutron
in one batch solution becomes
1
λp = ≈ 1cm. (8)
ρp σnp
Therefore, a prompt neutron with 1 MeV energy should have its energy after it
travels around 25 cm,
25
1
En = 1MeV × ≈ 0.03eV. (9)
2
This neutron does not have to travel linearly, but in any case, it should become
a thermal neutron.
(3) Mean Free Path of Thermal Neutron in the n − 235 U Fission Process
We can easily calculate the mean free path of the thermal neutron before the
nuclear fission in one batch solution. Since σf = 585b, we find
1
λf = ≈ 11 cm. (10)
ρσf
From these considerations, we see that prompt neutrons with 1 MeV should
travel around 25 cm, and then they become thermal neutrons. Further, after
they travel 11 cm, they can induce nuclear fissions. Thus, if one carries 50 of
the uranium nitrate solution into the sedimentation tank with 45 cm diameter
and 25 cm height, then nuclear chain reactions may well start quickly and proceed
further on.
(4) Reaction Time of Neutrons
Now we see that when prompt neutrons travel 36 cm, then they can induce
nuclear fissions. Therefore, we should estimate the duration time that is neces-
sary to travel this 36 cm. Since the nuclear reaction time must be smaller than
10−15 second, we can ignore this time duration. Since the prompt neutron with 1
MeV should spend τ 0 ≈ 7.6 × 10−10 second to proceed 1 cm, its energy becomes
a half of the previous energy after 1 cm walk.

√ Therefore, the time to proceed the
next 1 cm becomes larger by a factor of 2. In this way, if the prompt neutron
proceed 25 cm, then the total time to spend must be
√ 25

T0 = 1 + 2 + · · · + 2 2 τo ≈ 15 µs. (11)
After that, this neutron becomes thermal, and it should proceed 11 cm before
the nuclear fission. Since the thermal neutron may have the energy of 0.03 MeV,
it should take τth ≈ 46 µs. Thus, the total time that is necessary for the prompt
neutron to induce a fission reaction should be Ttotal ≈ 61 µs.
3.3 Total Energy of Fission with Criticality

Here, we should estimate the total amount of energy which is released from this
accident. This evaluation must be very difficult, but we want to calculate it in an
approximate way and obtain an order of magnitude of the total energy. First, the
number of neutrons which is required for the criticality reactions should be taken
as nr = 1.001, which is assumed to be consistent with the total energy released as
calculated from the computer simulation. In nuclear reactors, one should make
use of all the possible techniques to keep the number as nr = 1. In addition, we
assume that the number of nuclear fissions should be N = 40000. This number
is chosen so that the total nuclear energy release should be consistent with the
computer simulation which can reproduce all the observed radiation energies.
In this case, the total reaction time of fission becomes Tf ≈ 2.4s, and the total
number of fissions becomes
Ntotal = 1.00140000 ≈ 2.3 × 1017 . (12)
Further, we evaluate the neutron number at the beginning, and this neutron
should come from the spontaneous fission of 238 U . The number of neutrons
in one batch solution must be around 20, and a half of them are assumed to
contribute followed reaction. The energy release from the nuclear fission must
be around 200 MeV in each reaction, and therefore the total energy becomes
Etotal ≈ 4.6 × 1026 eV, (13)
which is just similar to the result of the computer simulation.
3.4 Why Does the Criticality Stop?

It is true that the criticality accident produced a huge amount of energy by the
nuclear chain reactions, and the accident is indeed quite serious. In this sense, we
here clarify as to how the chain reactions started and continued by reaching the
critical stage. However, we face to the more serious problem at this point. That
is, why the criticality accident could stop? We should understand any reason why
the criticality could stop, namely there were only one burst and not any more
burst, but why?
764 S. Oshima
(1) Nuclear Fission in the Seventh Batch

Here, we try to answer for this question, though it should be extremely
difficult. In order to find a possible mechanism for the stopping of the critical-
ity, we assume that the uranium compound should settle faster than any other
compounds in the solution. Further, we assume that uranium should be settled
within 20% height from the bottom of the sedimentation tank.
In this case, after the sixth batch, the uranium should be settled up to the
4.9 cm from the bottom. Thus, water should be found for 19.7 cm long in the
sedimentation tank. By taking into account this fact, we can calculate the total
energy release by nuclear fission as
Etotal ≈ 4.6 × 1026 eV ≈ 7.4 × 107 J. (14)
The duration time of this nuclear reactions can be estimated and should be
around Tf ≈ 2.4 s, which should correspond to the time that the uranium com-
pound is coming down to the bottom.
(2) Nuclear Fission in the Sixth Batch
The same calculation can be carried out for the sixth batch case. In this
case, we see that the total energy must be 1000 times smaller than that of the
seventh batch case. This is not very large, but at the sixth batch, the nuclear
chain reactions already started, and indeed there were a small burst.
From this calculation, we now understand the reason why the criticality
stopped. In case the uranium were settled at the bottom of the tank, then the
nuclear chain reaction cannot proceed further since the prompt neutrons cannot
lose their energy because of the lack of water.
(3) Nuclear Fission in the Eighth Batch
From now on, we only present a possible scenario of nuclear accident, if the
eighth batch were carried into the tank. In this case, the number of uranium
involved in the nuclear fission must be proportional to the height of water, and
thus it should be 22.9/19.7 more than the seventh batch. Thus, the number
becomes
22.9
N = 40000 × ≈ 46500. (15)
19.7
This means that the number of nuclear fissions should be also increased and the
total number becomes
Ntotal = 1.00146500 ≈ 1.5 × 1020 . (16)
Therefore, the total energy becomes
Etotal = 3 × 1029 eV ≈ 4.8 × 1010 J. (17)

This energy 4.8 × 1010 J corresponds to 11 ton of TNT powder which is quite a
serious explosion. The accident of Chernobyl nuclear power plant is believed to
correspond to around 100 ton of TNT powder, and therefore, if the eighth batch
were thrown away, then the accident would have been more than serious.
4 Design of Risk-Proof Procedure
Table 1 illustrates a procedure, including control factors of the manufacturing

system of uranium compound and objectives, designed to prevent criticality
accidents in advance from the standpoint of three manufacturing resources that
is Human, Machine and Material, grounded in the above theoretical discussion.
Proposed procedure mainly aims to reduce the opportunity of accelerated spon-
taneous occurrence of fission.
For the viewpoint of human, a control factor is the number of times work-
ers carry uranium nitrate into sedimentation tank. Whenever this operation is
activated, the neutrons settled in the tank should not contact with water mole-
cule contained in uranium nitrate solution. The opportunity between the two
materials must be minimized if possible.
For the viewpoint of a machine, a control factor is the space of sedimentation
tank. The shape of the tank id better to be shallow with thinly fill of water
because neutrons will be severely scattered in this case. Owing to it, the moving
distance of neutrons in the water of tank can be minimized.
For the viewpoint of a material, two control factors are identified which can
regarding to two components of uranium nitrate solution i.e. neutrons and water
molecule. The former is the core of fission. The more the number of neutrons
are, the more the opportunity of fission of neurons increases. It is necessary
to control suitable number density of 235 U . The latter has the role to reduce
Table 1. Risk-proof procedure
Manufacturing Control factors Objectives

resources
Human The number of times Minimization of the opportunity of
workers carry into neutrons to make contact with water
sedimentation tank molecule
Machine Space of sedimentation Minimization of moving distance of
tank neutrons into the water in
sedimentation tank
Material The number density of Reduction of the probability of
235
U induced fission by neutrons
The amount of water Prevention of over- slowdown material
766 S. Oshima
energy of neutrons which lead to accelerated fission. Over-input of water may

cause diversification of neutron occurrence and losing of their energy. Suitable
amount of water should be fixed to prohibit any accident.
The above conditions were derived from consideration on operational
resources. In the future, the comprehensive design of effective countermeasures
are needed. The accident focused in this paper is old case, however, the pro-
posal approach that intends to develop the procedure based on the principle of
physical phenomenon will be useful for experienced IEer.
The purpose of this study aims to examine a risk-proof procedure for nuclear
materials operation through the causal analysis of JCO accident in 1999. In the
analysis, the basic mechanism of the accident is discussed in terms of the nuclear
multiple scattering theory. As the result, some specific reasons why the chain
reactions can proceed and why the criticality could stop. The control factors of
operation resources are extracted referred to physical phenomena regarding to
fissions.
This paper aims to assist managers and academic researchers in the area of
atomic power generation. As atomic power plants rely on many human opera-
tions, mangers have to cope with operational effectiveness and safety. Therefore,
they have to concern to provide systematic procedure based on the proposed
methodologies.
Especially, this paper proposed control factors to prevent accident. Subse-
quently, because of its early stage, we hope related studies to this approach is
advanced by researchers interested in various areas.
Acknowledgements. We, the authors, sincerely acknowledge Dr T. Fujita, the super-

visor of Doctor of Science Degrea of the 1st author, for his kind suggestions and support
on preparation of this paper.
References
1. Bohr A, Mottelson BR (1998) Nuclear Deformations. Nuclear Structure
2. Fujita T, Hüfner J (1979) Inclusive hadron-nucleus scattering at high energy. Phys
Lett B 87(4):327–330
3. Fujita T, Hüfner J (1980) Momentum distributions after fragmentation in nucleus-
nucleus collisions at high energy. Nucl Phys A 343(3):493–510
4. Glauber RJ (1959) Lectures in Theoretical Physics. Interscience
5. JCO accident investigation committee (2005) Criticality accident in JCO; its solu-
tion, fact, cause and study (jco rinkaijika; sonozenbou no kai jijitsu youin taiou).
In: Iinkai JJC, Gakkai NG (eds) Atomic Energy Society of Japan. Tokai University
Press, Kanauawa
6. Nuclear Data Center (2016). http://wwwndc.jaea.go.jp/j40fig/jpeg/h001 f1.jpg
7. Okajima S, Kugo T, Mori T (2012) Reactor Physics (Genshiro Butsurigaku).
Ohmsha, Tokyo
8. Oshima S, Hisada M et al (2016) Reexamination of criticality accident in JCO. J

Mod Phys 07(12):1500–1505
9. Tateno J, Noguchi K, Aoyagi N (2000) Criticality Accident in Tokaimura; Complete
Analysis (Tetteikaimei Tokaimura Rinkaijiko). Shin-nihon Shuppan-sha, Tokyo
10. Tonoike K, Nakamura T et al (2003) Power profile evaluation of the jco pre-
cipitation vessel based on the record of the gamma-ray monitor. Nucl Technol
143(3):364–372
11. Yamauchi K, Murata K, Katayama H (2010) Reinforcement of performance com-
petitiveness by concurrent strategy: a case of chemical product manufacturing line.
In: Proceedings of the 23rd International Congress on Condition Monitoring and
Diagnostic Engineering Management (COMADEM 2010), pp 913–920
Multi-stage Logistics Inventory for Automobile
Manufacturing by Random Key-Based GA
Hisaki Inoue1(B) , Jung Bok Jo2 , and Mitsuo Gen3

1
Miyazaki Sangyo Keiei University, Miyazaki 880-0931, Japan
inoue-lab@sankei.info
2
Division of Computer Engineering, Dongseo University, Busan, Korea
3
Abstract. When evaluating a logistics system, automobile companies

commonly search for the minimum transportation cost, which is sig-
nificantly influenced by inventory problems. These inventory problems
are extensive and varied. In many actual logistics systems, there are
the three-stage network models take inventory values into consideration.
Safety inventories are kept in distribution centers (DCs). In this study,
we adapted a model to set-up a number of plants and DCs. We then per-
formed numerical experiments by using demand data that we created on
the basis of data disclosed by an automobile company. In this study, we
propose a random key-based genetic algorithm (rk-GA) with the distrib-
uted environment scheme, we compared it with random key and spanning
tree-based GAs, and report the advantages of the proposed method, ran-
dom key-based genetic algorithm with distributed environment scheme
(des-rkGA).
Keywords: Automobile manufacturing · Multi-stage logistics · Inven-

tory control · Random key-based genetic algorithm
1 Introduction
Many difficult inventory problems involve production control and asset man-
agement, but those that involve logistics systems are important and have gar-
nered strong interest in recent years. Particular focus has been paid to inventory
problems. Doboshas presented many papers on inventory problems. In 2001, he
presented the problem of reverse logistics, adjusting the relationship of holding,
production, and disposal costs [4]. In 2005, he investigated production inventory
adjustment [5]. In 2007, he presented a paper on the total production cost of two
companies for adjustment [6]. Minner et al. [18] have also presented papers on
inventory problems. In 2003, they evaluated reverse logistics, where the inven-
tory has several supply methods. In 2005, they presented a paper on the problem
of adjusting shipping, replenishment, and lost sales opportunities for two inven-
tories [19]. In 2008, Thangam et al. [28] presented a paper on how to determine

Multi-stage Logistics Inventory for Automobile Manufacturing 769
replenishment for Poisson demand. In 2003, Mahadevan et al. [16] treated a facil-
ity as an inventory problem where the returned products are remanufactured.
In 2004, Miranda et al. [20] analyzed the inventory decision problem using the
Lagrangian relaxation method and subgradient methods; their ordering point
method was based on the economic order quantity. In 2009, Rieksts et al. [24]
analyzed the inventory problem with ordering intervals using power- of- two
policies.
There are many other inventory problems, including the reduction of the
total safety inventory quantity, or on-hand inventory [8], and the calculation of
the inventory value at each step, or echelon inventory [12].
In this study, we propose a new model where inventories are managed at
distribution centers (DCs), taking actual conditions into account. Holding costs
only occur in DCs, but additional inventory costs are incurred for the product
value because a product near dealers has more value. For inventory costs other
than the holding cost, we can get an interest charge if the product is exchanged
with cash. Another factor is the lost product value due to age depreciation. We
calculated the annual supply and demand value, and created demand data that
were based on the Poisson demand of time-series fluctuations. Logistics models
such as these are known as the NP-hard problem [3]. Soft computing methods
such as simulated annealing, neural networks, and genetic algorithms (GA) are
well-suited to solve this problem [1,17,26,29].
In this study, we adopted random key-based genetic algorithm (rk-GA) with
distributed environment scheme (des-rkGA) as the proposed method, which is
an improved version of rk-GA; we compared the proposed method with rk-GA
and spanning tree-based GA (st-GA) to confirm its suitability [9].
We propose a model that addresses many of the different inventory problems
studied earlier; we demonstrate des-rkGA algorithm to solve the multi-logistics
inventory problem, and present the computational results with the effectiveness
by the proposed algorithm.
2 Inventory Problem in Logistics System

In this section, we detail the contents for the inventory problem treated in this
study. Inventory costs result from holding costs and other factors. For the model
used in this study, the product is held in a DC for general managing. All products
have needed costs as their product value, and the production control cost is
needed to calculate the safety inventory in a DC. The safety inventory quantity
is decided by the service level but is affected by the production control and
holding costs. The safety inventory cost is based on the service level and has a
trade-off relation with lost costs based on lost sales opportunities.
Figure 1 shows many different kinds of inventories; they are difficult to cat-
egorize clearly, but it is important for inventory management to categorize and
manage them. The inventory cost without the holding cost includes all inven-
tory, such as the pipeline, production process, lot size and DC inventories. The
inventory costs determine the product value, inventory value, inventory time,
and inventory holding ratio.
770 H. Inoue et al.
Therefore, the inventory cost without holding costs and the product value,
inventory value, inventory time, and inventory holding ratio have a trade-off
relation. In this section, we describe the inventory holding ratio, safety inven-
tory, production adjustment, pipeline inventory, and inventory on a production
process, lot size inventory, and DC inventory as well as a summary of the inven-
tory problems evaluated in this study. Although there are many kinds of inven-
tories, we adopted the idea of a value chain for the inventories; we used the
pipeline, lot size, production process, and DC inventories.
Fig. 1. Inventory category
The general outline of the logistics model used for the inventory in this study
is shown in Fig. 2.
Fig. 2. General outline of logistics model with inventory
Below, we define the indices, constants, parameters, decision variables, and

units used in this paper. Units of measurement are in square brackets.
Indices:
i = 1, 2, · · · , I : index of suppliers;
j = 1, 2, · · · , J : index of plants;
k = 1, 2, · · · , K : index of DCs;
l = 1, 2, · · · , L : index of customers;
t = 1, 2, · · · , T : index of cycles in the logistics system. Tcycle intervals are
described later.
Parameters:
A : number of unit parts to constitute;
c1ij : shipping cost of unit parts or material from supplier i to plant j
[yen];
c2jk : shipping cost of unit production from plant j to DC k [yen];
c3kl : shipping cost of unit production from DC k to customer l [yen];
gij
1
: shipping time of unit parts or material from supplier i to plant j [h];
gjk
2
: shipping time of unit production from plant j to DC k [h];
gkl
3
: shipping time of unit production from DC k to customer l [h];
Tplant : producing time of plants [h];
α : safety inventory coefficient;
σ : standard deviation of demand;
Tcycle : cycle time in the logistics system. It shows how much time it takes
for a load to be moved once [h];
h : inventory holding ratio (2.1 value chain.);
H : holding cost of unit production in a DC [yen];
ek : inventory cost or DC k [yen];
Mcost : material cost [yen];
rj1 : fixed cost for operating plant j [yen];
rk2 : fixed cost for operating DC k [yen];
Ui1 : upper limit of supply for parts and materials in supplier i;
Uj2 : upper limit of supply for production in plant j;
Uj3 : upper limit of supply for production in DC k;
pconst
j : unit cost production at steady state [yen];
pexceed
j : unit cost production at excess state, [yen];
shortage
pj : unit cost production at shortage state [yen];
Itotal : total inventory cost without holding cost;
Ndelay : number of delays for plant production;
W1 : number of plants that can be operated;
W2 : number of DCs that can be operated;
Raverage : production quantity for one period of cycle time as calculated
from the annual average production quantity;
Si1 : supplier production ratio; ratio of supplier i to gross supplier product
capability;
Sj2 : plant production ratio; ratio of plant j to gross plant product
capability;
zklevel : base inventory level in DC k.
772 H. Inoue et al.
Decision Variables:
u1i (t) : supply amount of parts and materials in supplier i at cycle t;
u2j (t) : supply amont of production in plant j at cycle t;
u3k (t) : supply amont of production in DC k at cycle t;
u4l (t) : demand amount of production in customer l;
zk (t) : inventory volume for DC k at cycle t [yen];
Sk3 (t) : DC demand ratio; demand ratio of DC k to the gross DC demand
quantity;
Rj2 (t) : shipping quantity in plant j;
Rk3 (t) : receive cargo quantity in DC k;
Bl (t) : back order quantity in customer l;
Dl(t) : order quantity at this period in customer l;
DlDC (t) : request quantity to DC at this period in customer l;
Δzk (t) : difference of inventory quantity and base inventory level;
Δzmax (t) : maximum of difference of inventory quantity and base inventory
level;
zkreq (t) : request quantity in DC k;
p1val : product value when delivery is completed to plants [yen];
p2val : product value when delivery is completed to DCs [yen];
p3val : product value when delivery is completed to customers [yen];
Ujexceed : threshold for determination when exceeding production;
Ujshortage : threshold for determination when reducing production;
Decision Variables:
pj (uj 2 ) : producing cost of unit production in plant j, [yen]
x1ij (t) : amount supplied of unit parts or material from supplier i to plant j
at cycle t;
x2jk (t) : amount supplied of production from plant j to DC k at cycle t;
x3kl (t) : amount supplied of production from DC k to customer l at cycle t;
p1j (t) : operating flag for plant j at cycle t (= 1 when plant j is used,
= 0 otherwise);
p2k (t) : operating flag for DC k at cycle t (= 1 when DC k is used,
= 0 otherwise).
2.1 Value Chain
When considering the product value and inventory holding ratio in addition to
the holding cost, the inventory cost becomes high near the customer (dealer),
as shown in Fig. 3 [23]. The inventory holding ratio is determined by the con-
stant number by interest rate when the product is cashed, the decrease in prod-
uct value, etc. In this study, we treated the inventory holding ratio as uniform
because we were dealing with engineered products such as automobiles that do
not experience degradation. We adopted the value chain concept for all products.
The value chain can be defined as:

Icost : Inventory Cost;
Pvalue : Product Value;
Ivolume : Inventory Amount.
Icost = h × Pvalue × Ivolume (1)
Fig. 3. Value chain
2.2 Safety Inventory
The safety inventory is the inventory required to prevent stock out. The safety
inventory and service level have a mutual trade-off relation. The formula for the
safety inventory is shown below.

Isafety = ασ Tcycle + max(gjk ) ∀j, k, (2)
where Isafety is safety inventory.
2.3 Production Adjustment
Order fluctuations mean that factories have to adjust production. Extra charges
include extra pay, etc. if production is over the steady state, and extra fixed
costs for the equipments if production is under steady state. The formula for
production adjustment is shown below.
Pj (u2j (t)) =Pconst + max{Pexceed (u2j (t) − Ujexceed ), 0} (3)

+ max{Pshortage (Ujshortage − u2j (t)), 0}.
2.4 Pipeline Inventory
The pipeline inventory denotes the inventory during shipping. There is a trade-off
between the pipeline inventory and shipping cost. In this study, we constructed a
model based on the idea of an inventory holding ratio. We treated each product
during shipping between a supplier and plant and between a DC and customer
774 H. Inoue et al.
as pipeline inventory. The formula of the pipeline inventory cost Ipileline is shown
below.
⎧ ⎛
⎨T I
J J
K
Ipipeline =h ⎝p1val gij
1 1
xij (t) + p2val gjk
2 2
xjk (t) (4)
⎩
t=1 i=1 j=1 j=1 k=1
K
L
+p3val gkl
3 3
xkl (t) .
k=1 l=1
The cost is approximated as follows because we presumed that the product

has the same value when received by any factory, DC, or dealer. The value of
completed product is approximated by material and average shipping costs.
T I J
j=1 cij xij (t)
1 1
t=1 i=1
p1val = Mcost + T I J
. (5)
j=1 xij (t)
1
t=1 i=1
The product value when shipped to DCs is approximated by its value when
shipped to plants and the average shipping cost.
T J K
k=1 cjk xjk (t)
2 2
t=1 j=1
p2val = p1val + T J K
. (6)
k=1 xjk (t)
2
t=1 j=1
The product value when shipped to dealers is approximated by its value when
shipped to DCs and average shipping cost.
T K L
l=1 ckl xkl (t)
3 3
t=1 k=1
p3val = p2val + T K L
. (7)
l=1 xkl (t)
3
t=1 k=1
2.5 Inventory of a Production Process
For the inventory of a production process, we constructed a model based on the

idea of an inventory holding ratio. We treated an unfinished product as inventory
on a production process.
The formula is shown below. The product value on a production process
changes with the progress conditions of the manufacturing processes. However,
in this study, the value was approximated as the material cost and half of the
production cost in the plant.
J

Iproduction = hTplant u2j (t)(p1val + Pj (u2j (t))/2). (8)
j=1
2.6 Lot-Size Inventory
The lot size inventory is the inventory of the completed products during shipping.
We treated completed products as lot size inventory. All products are shipped
from plants to DCs in the same interval. The lot size inventory cost was cal-
culated as half the product of the shipping values, production values in plants,
and inventory holding ratio, as shown in Fig. 4; we used this formula because
all products were shipped from a plant to a DC in the same time intervals. The
formula is shown below.
J
T K
gjk
2 2
xjk (t)(p1val + Pj (u2j (t)))
Ilotsize = h . (9)
t=1 j=1 k=1
2
Fig. 4. Lot size inventory value
2.7 DC Inventory
We treated products in a DC as DC inventory. In this study, the DC inventory

plays the central role of the inventory of the logistics system; we need extra
holding costs for each car in DCs. The DC inventory cost is calculated as the
product of the DC inventory amount, cycle time, and production value when
shipped to DCs multiplied by the inventory holding ratio. The formula is shown
below.
K
T T
K
IDC = hTcycle p2val zk (t) + HTcycle zk (t). (10)
t=1 k=1 t=1 k=1
2.8 Total Inventory Cost Without Holding Cost
As mentioned above, the formula of the total inventory cost without the holding
cost used in this study consists of the pipeline, production process, lot size, and
DC inventory costs; it is shown as follows.
776 H. Inoue et al.
Itotal= Ipipeline + Iproduction + Ilotsize + IDC (11)

T
I
J J
K
Itotal =h (p1val gij
1 1
xij (t) + p2val gjk
2 2
xjk (t)
t=1 i=1 j=1 j=1 k=1
L
K J

+ p3val gkl
3 3
xkl (t)) + Tplant u2j (t)(p1val + Pj (u2j (t))/2) (12)
k=1 l=1 j=1
T
J
K K
T
gjk
2 2
+ + Tcycle p2val zk (t)
t=1 j=1 k=1
2 t=1 k=1
T
K
+ HTcycle zk (t).
t=1 k=1
2.9 Demand Data

In this study, we constructed the logistics system in terms of inventories with
Poisson demand [25]. To confirm the availability, we created demand data from
an automobile company’s public data. Many products flow in the supply chain
network (SCN) by dealer’s demand, and these demands are approximated with
Poisson demand time fluctuations. We created new demand data using Poisson
randomization. The Poisson equation is shown below.
e−λ λk
p(k) = , (13)
k!
here, p(k) is the demand, λ is the average order values, and k is the number
of order times. Figure 5 shows the outline of the Poisson randomization. In this
study, we created daily demand data each experiment by applying a Poisson
random number to the annual demand.
Fig. 5. Relationship between amount and times of orders

3 Mathematical Model of Multi-stage Logistics System

with Inventory
3.1 Assumptions
In this study, we constructed the logistics model with the following assumptions.
A1. The transit times are known between suppliers and plants, plants and DCs,
and DCs and customers.
A2. The shipping costs are known between suppliers and plants, plants and
DCs, and DCs and customers.
A3. The supplies are delivered without delay to a factory in population to the
planned production.
A4. In this model, suppliers provide multipurpose parts for effective optimiza-
tion. The parts are examined and classified for easy assembly work. This
process is carried out assuming that A package parts are used in the assem-
bly of one car. The conditions for assembling one car require the package
parts of A units. Other parts are not targeted in this model because the
supply route is decided from the first time the supplier side is entrusted
with delivery, few of the detailed parts have management value in the
logistics system, etc.
A5. The products made at a plant are shipped to a DC by lot size.
A6. The DCs have space to accept products from plants.
A7. We consider only the inventory costs in DCs.
A8. The customer addresses are known.
A9. The existence of inventory is known in advance through an inventory check;
orders for inventory that is out of stock are treated as reservations.
A10. A product has the same value when received by any factory, DC, or dealer.
A11. All products are delivered within the limits time of a Tcycle .
3.2 Multi-stage Logistics System
Generally, when a logistics system is seen from the functional constitutive prop-
erty, it is modeled by a three-stage production network and distribution system,
which is called a supply chain network. The first stage is the supplier phase,
which involves parts and a supplier. The second stage is the plant phase, which
consists of a production plant or outsourcing. The third stage is the DC phase
and consists of a distribution center or storehouse. A sample of an actual auto-
mobile company’s logistics system is shown in Fig. 6. In many cases, the actual
logistics system is comprised of three stages. A three stage logistics system model
is shown in Fig. 7. The logistics model used in this study had 14 suppliers, 2 fac-
tories, 4 DC, and 22 customers. The mathematical model used in this study is
shown below.
778 H. Inoue et al.
Fig. 6. Sample of actual logistics system
T
I
J J
K
min Z1 = A c1ij x1ij (t) + c2jk x2jk (t)
t=1 i=1 j=1 j=1 k=1
K
L T
I
J
+ c3kl x3kl (t) + h (p1val gij
1 1
xij (t)
k=1 l=1 t=1 i=1 j=1
J
K K
L
+ p2val gjk
2 2
xjk (t) + p3val gkl
3 3
xkl (t))
j=1 k=1 k=1 l=1
J

+ Tplant u2j (p1val + Pj (u2j (t))/2) (14)
j=1
J
T K
gjk
2 2
+
t=1 j=1 k=1
2
K
T T
K
+ Tcycle p2val zk (t) + HTcycle zk (t)
t=1 k=1 t=1 k=1
J
T J
K

+ u2j (t)P (u2j (t)) + rj1 p1j (t) + rk2 p2k (t)
t=1 j=1 j=1 k=1
J

s. t. x1ij (t) ≤ u1i (t), ∀i, t (15)
j=1
I

x1ij (t) ≤ u2j (t)p1j (t), ∀j, t (16)
i=1
J

x2jk (t) ≤ u3k (t)p2k (t), ∀k, t (17)
j=1
K

x3kl (t) ≥ u4l (t), ∀l, t (18)
k=1
J
J

x1ij (t) = x2jk (t), ∀j, t (19)
j=1 j=1
K
K

x2jk (t) = zk (t − 1) − zk (t) + x3kl (t), ∀j, t (20)
k=1 k=1
J

p1j ≤ W1 (21)
j=1
K

p2k ≤ W2 (22)
k=1
x1ij (t), x2jk (t), x3kl (t), zk (t) ≥ 0, ∀i, j, k, l, t (23)
p1j (t), p2k (t) = {0, 1}, ∀j, k, t (24)
Pj (u2j ) = Pjcont + max(Pjexceed (u2j (t) − Ujexceed ), 0)
+ max(Pjshortage (Ujshortage − u2j (t)), 0) (25)
T I J
j=1 cij xij (t)
1 1
t=1 i=1
p1val = Mcost + T I J
(26)
j=1 xij (t)
1
t=1 i=1
T J K
k=1 cjk xjk (t)
2 2
t=1 j=1
pval = pval +
2 1
T J K
(27)
k=1 xjk (t)
2
t=1 j=1
T K L
l=1 ckl xkl (t)
3 3
p3val = p2val + t=1 T
k=1
K L
. (28)
l=1 xkl (t)
3
t=1 k=1
4 Advanced Genetic Algorithm
4.1 Random-Key Based Genetic Algorithm
Gen and Lin [7] surveyed genetic algorithms in Wiley Encyclopedia of Computer
Science and Engineering and recently many researchers applied GA to various
areas in logistics systems. Inoue and Gen [10] reported multistage logistics sys-
tem with inventory considering demand by hybrid GA, Neungnatcha et al. [22]
reported adaptive genetic algorithm (AGA) for solving sugarcane loading sta-
tions with multi-facility services problem, Jamrus et al. [11] reported discrete
particle swarm optimization (PSO) approaches and extended priority based-
HGA for solving multistage production distribution under uncertainty demands
780 H. Inoue et al.
Fig. 7. 3-stages logistics model
and Lee et al. [13] reported multi-objective hybrid genetic algorithm (MoGA) to
minimize the total cost and delivery tardiness in a reverse logistics.
Lin and Gen [15] proposed a random key-based genetic algorithm (rk-GA) for
solving AGV (automatic guided vehicle) dispatching problem in flexible man-
ufacturing system (FMS). Now we are going to use it for multistage logistics
system with inventory. Now we define the following example of the cost matrix:
Fig. 8. Sample of cost matrix
The algorithm created using the rk-GA technique has three logistics stages.
Figure 7 shows the third stage process. Figure 8 shows a sample cost matrix.
Figure 9 shows a sample rk-GA chromosome.
Gen and Cheng successfully applied rk-GA encoding to the shortest path
and project scheduling problems in 2000 [2]. For transportation problems, a
chromosome consists of the priorities of sources and depots, which make up a
transportation tree; its length is equal to the total number of sources m and
depots n, i.e., m + n. The transportation tree corresponding to a given chro-
mosome is generated by the sequential arc between sources and depots. At each
Fig. 9. Sample of rk-GA chromosome
Fig. 10. Sample of one point crossover
Fig. 11. Sample of insertion mutation
step, a single arc is added to a tree that selects a source (depot) with the highest
priority and connects it to a depot (source) to minimize cost.
Figure 13 shows the brief decoding at each stage. Figure 14 shows the process
of the m-logistics problem. In this study, we used the one-point crossover, which is
the simplest method when using rk-GA. We used insertion and swap mutations.
We used the roulette wheel approach, which selects the chromosome in ascending
order of fitness. Examples of one-point crossover, insertion mutation, and swap
mutation are shown in Figs. 10, 11 and 12, respectively.
The new generated chromosome is evaluated. It is selected in ascending order
of fitness based on the number of popSize in the parent and newly generated
chromosomes. The order of fitness then helps determine the next generation of
chromosomes.
782 H. Inoue et al.
Fig. 12. Sample of swap mutation
Fig. 13. Brief decoding of each stage
Inversion and displacement mutations are used in st-GA. The inversion muta-
tion select two positions within a chromosome at random and then inverts the
substring between these two positions. The displacement mutation selects a sub-
string at random and inserts it in a random position.
4.2 Total Logistics Process
The total logistics system process can be explained as follows. Figure 15 shows
the whole logistics system process in logistics cycle periods. The safety inventory
is given first each cycle period. The inventory quantity is renewed last in each
cycle period. The plant product shipping quantity is based on past demand,
because of the production time in plants. In this study, this is called the number
of delays for plant production Ndelay where the plant production time is the
number of order times. This is shown below.
Fig. 14. m-Logistics Process
Tplant
Ndalay = . (29)
Tcycle
Here, we use an example for the shipping products when the customer’s
total demand quantity changes from 600 before delay cycle times (Ndelay ) to
1200. A pull-type demand quantity is applied to DC-customer and supplier-plant
product distributes based on the demand quantity at the time. The shipping
quantity is 1200. A push-type demand quantity is applied to plant-DC product
distributions based on the demand quantity before the number of delay times
for plant production (Ndelay ). The shipping quantity is 600.
The load of the customer’s demand is shared by the total inventory quantity
in DCs and total production in plants and suppliers. An example process is
shown in Fig. 16. Step 1 shows a renewal of the demand quantity (u4l ) and back
order quantity (Bl ). Step 2 calculates the planned order quantity (u2j (t)). Step 3
calculates the order quantity in suppliers (u1i (t)). Step 4 calculates the planned
shipping quantity (Rj2 (t)). Step 5 calculates the quantity of cargo received in
DCs (Rk3 (t)).
Fig. 15. Total m-logistics process in a cycles period

784 H. Inoue et al.
Fig. 16. Initial process
4.3 rk-GA with Distributed Environment

The immigration scheme is an independent genetic operation where each island
is made up of several divided populations for the same generation. The basic con-
cept of immigration is shown in Fig. 17. Immigration is an information exchange
that is performed continuously for a group of chromosomes. The ratio to the
number of immigration populations is called the immigration ratio, and the
generation interval during which immigration occurs is called the immigration
interval. In 2008, Lin et al. [14] solved the shortest route problem for cost and
time using a priority-based GA with immigration, and reported their result.
Nominated Same
Group of Genes Immigation Generation
Island-A
Island-B
Island-D
Island-C
Fig. 17. Basic Concept of Immigration
In this study, we propose rk-GA with a distributed environment scheme (des-

rkGA) that changes the crossover and mutation rates for each island. Miki et al.
[21] proposed multi-objective genetic algorithms with a distributed environment
scheme (MoGA-DES). We adopted an immigration rate of 0.1; we selected other
islands randomly and changed the chromosome asynchronously. There are two
methods for parallel processing [27]. The multithread method performed by one
programming using time sharing. The multi-task method performs several tasks
at the same time [21]. In this study, we adopted the multi-task method because
it is easy to disperse processing to several PCs. There are also two process-
ing methods for GA. The synchronous method synchronized the time for each
generation. The asynchronous method does not do so. We adopted the asyn-
chronous method because we would need a wait time for the slowest island if we
adopted the synchronous method. We created nine islands with crossover prob-
ability (PC) and lower mutation probability (PM) than the center island. Two
of nine islands were chosen at random with the generation timing of the center
island. The direction of immigration was decided at random, and the popula-
tions immigrated at a 0.1 immigration rate. Next, 10% of each island’s worst
chromosomes were destroyed, and 10% of other islands’ best chromosomes were
adopted as immigrations. The concrete values of PC and PM used by this study
are discussed in detail in the next section. We performed the experiment using
parallel processing with three PCs.
We performed prior experiments to determine PC and PM for the center island
and obtained the solution is shown in Table 1. We adopted PC = 0.4 and PM =
0.6 as the center island values for st-GA and PC = 0.6 and PM = 0.4 as the
center island values for rk-GA because they provided the best solutions.
We adopted PC = 0.6 and PM = 0.4 as the center island values for des-rkGA,
because it is based on rk-GA. We created 9 islands with PC = (0.4, 0.6, 0.8) and
PM = (0.2, 0.4, 0.6) for the experiments.
popSize denotes the population size. maxGen is the maximum generation
size used as the terminating condition for the experiments. We performed exper-
iments repeated 20 times using maxGen = 5000 and popSize = (20, 50, 100).
However, we reported maxGen = 1000 as sufficient for the evolutive process
because no more improvement was detected after 1000 generations. Table 2 shows
the evolutive processes for the best value by each method. The best value Z1
was at gen = (300, 500, 1000).
Fig. 18. Evolutive process for each method
Figure 18 shows the evolutive processes for each method when popSize was
100. The proposed Des-rkGA produced the best final result because its evolutive
786 H. Inoue et al.
Table 1. Solution at each PC and PM (unit: yen)
PM PC
0.2 0.4 0.6 0.8
st-GA 0.2 451, 461, 818 442, 315, 078 447, 088, 045 454, 269, 696
0.4 445, 044, 804 440, 918, 667 444, 501, 564 444, 782, 298
0.6 442, 653, 511 440, 602, 712 445, 820, 848 452, 372, 968
0.8 445, 865, 777 443, 723, 937 451, 882, 261 454, 711, 777
rk-GA 0.2 412, 925, 351 406, 037, 962 408, 355, 139 413, 251, 110
0.4 406, 280, 823 405, 783, 627 405, 476, 817 405, 988, 745
0.6 408, 604, 249 406, 269, 725 405, 763, 570 413, 679, 127
0.8 412, 120, 741 406, 777, 058 411, 116, 273 413, 073, 894
Table 2. Evolutive process for each popSize (unit: yen)
PopSize Gen
300 500 1000
st-GA 20 551, 429, 175 547, 095, 287 541, 769, 426
50 499, 464, 718 485, 094, 520 482, 185, 269
100 447, 629, 126 423, 064, 081 402, 412, 188
rk-GA 20 467, 164, 557 463, 804, 677 433, 653, 257
50 452, 496, 499 434, 436, 525 394, 435, 691
100 366, 335, 225 365, 780, 289 364, 869, 285
des-rkGA 20 418, 809, 097 396, 053, 541 389, 092, 924
50 353, 999, 460 346, 130, 700 346, 030, 174
100 348, 536, 438 345, 863, 047 345, 695, 925
Table 3. Evolution of each island by des-rkGA (unit: yen)
Island Gen
10 20 50 70
A 590, 914, 974 505, 586, 248 442, 219, 078 407, 869, 638
B 545, 289, 874 519, 883, 420 431, 958, 888 411, 047, 134
C 596, 086, 519 530, 215, 824 413, 252, 989 399, 855, 446
D 589, 080, 440 496, 206, 544 433, 337, 814 386, 073, 858
E 569, 117, 032 535, 673, 932 419, 654, 636 412, 582, 256
F 556, 063, 137 497, 222, 107 454, 229, 946 381, 231, 373
G 582, 061, 836 535, 413, 007 412, 034, 067 417, 752, 697
H 548, 377, 594 520, 688, 854 446, 742, 473 378, 950, 050
I 572, 220, 316 493, 015, 494 415, 259, 267 382, 084, 372
des-rkGA 545, 289, 874 493, 015, 494 412, 034, 067 378, 950, 050
speed was faster than the other compared methods. As shown in Table 2, Z1 was
(402, 412, 188) when st-GA was used. It was (364, 869, 285) with rk-GA and (345,
695, 925) with des-rkGA. des-rkGA showed 16.41% and 5.55% improvements
compared to st-GA and rk-GA, respectively. des-rkGA was confirmed to provide
stable results because the standard deviation was only 3210 compared to 20,
338 with st-GA and 7780 with rk-GA. Table 3 shows the evolutive process of
each island when popSize was 100 as in Table 2. The highlight shows the best
solution for each generation, of which the best was with des-rkGA. In a prior
experiment, the PC and PM for island E were the best combination. However, in
the evolutive process for each island, many of the best solutions were produced
at other islands. Table 4 shows the solution at each immigration rate, and the
best solution by rk-GA is shown as reference. The experimental results show
that if the immigration rate surpassed 50%, the results become bad. Figure 19
was created from Table 4; the best value was produced at 10% migration rate.
Table 4. Immigration Rate of des-rkGA
Immigration rate (%) Z1 (yen)

0 352, 543, 849
5 349, 209, 452
10 346, 035, 987
15 349, 468, 738
20 353, 121, 891
50 373, 634, 700
rk-GA 364, 769, 003
Fig. 19. Immigration Rate of des-rkGA
Table 5 summarizes the experimental results in terms of the des-rkGA

improvement rate and standard deviation for Z1 of maxGen when popSize
788 H. Inoue et al.
Table 5. Experimental results
st-GA rk-GA des-rkGA

Z1 (yen) 402, 412, 188 364, 869, 285 345, 695, 925
Improvement rate 16.41% 5.55% -
standard deviation 20, 338 7, 780 3, 210
st-GA rk-GA des-rkGA

Each PC Total PCs
maxGen 55.56 29.96 32.47 97.41
For 1d 35.78 19.98 20.45 61.35
Time to arrive at the maxGen value 55.56 11.07 5.74 17.22
Time to arrive at the rk-GA value - - 3.15 9.45
was 100. As shown in Table 2, there were differences when the number of gener-
ations was small.
In this experiment, we used three PCs of the same kind dual core AMD1212
2.0 GHz/2 MB; the memory size was 2 GB, and the development language was
C#.
The computational time is shown in Table 6. We experimented with the test
data for 90 days. When converted to 1 day, the computation time was 35.78,
19.98, and 20.45 s using st-GA, rk-GA, and des-rkGA, respectively. The average
generation number when the solution arrived at the maxGen value was 685,
467, and 78 using st-GA, rk-GA, and des-rkGA, respectively. The generation
time is shown in Table 6 as time to arrive at the maxGen value. The time was
55.56, 11.07, and 5.74 s using st-GA, rk-GA, and des-rkGA, respectively. There
was no drastic improvement because the total CPU processing time was 9.45 s,
but the CPU processing time could be distributed. Moreover, the average of the
generation numbers when the des-rkGA value became better than maxGen of
rk-GA was 165, and it took 3.15 s. We also confirmed the advantage of des-rkGA
in terms of computer time. This was due to using parallel processing with 3 PCs
of the same kind. We believe that the PC environment affects the solutions.
6 Conclusions
Using date of an actual automobile company in this study, we proposed random
key-based genetic algorithm with distributed environment scheme (des-rkGA),
for a multi-stage logistics system that calculates inventory values for many differ-
ent cases. We proposed a logistics system that keeps the safety inventories only
in DCs and that can cope with the location allocation problem. We performed
numerical experiments with test data that were based on the disclosed data of a
certain automobile company. The proposed des-rkGA showed improvements of

16.25% and 5.43% compared to the performances of st-GA and rk-GA, respec-
tively. The des-rkGA also showed a small dispersion of solutions compared with
the other methods. For future work, we intend to investigate the effectiveness
of the logistics system in actual conditions while continuing to advance research
and analysis of case studies.
Acknowledgements. This work is partly supported by the Grant-in-Aid for Scientific

Research (C) of Japan Society of Promotion of Science (JSPS): No. 15K00357.
References
1. Altiparmak F, Gen M, Lin L (2004) A priority-based genetic algorithm for supply
chain design. Comput Ind Eng 13:22–25
2. Cauty R (2000) Genetic algorithms and engineering optimization. Wiley, Hoboken
3. Gen M, Cheng R (1997) Genetic algorithms and engineering design. Wiley, New
York
4. Dobos I (2001) Production strategies under environmental constraints: continuous-
time model with concave costs. Int J Prod Econ 71(1–3):323–330
5. Dobos I (2005) The effects of emission trading on production and inventories in
the Arrow-Karlin model. Int J Prod Econ 93–94(1):301–308
6. Dobos I (2007) Tradable permits and production-inventory strategies of the firm.
Int J Prod Econ 108(1–2):329–333
7. Gen M, Lin L (2009) Genetic algorithms. In: Wah B (ed) Wiley encyclopedia of
computer science and engineering. Wiley, pp 1367–1381
8. Graves SC, Willems SP (2003) Supply chain design: safety stock placement and
supply chain configuration. Handbooks Oper Res Manage Sci 11:95–132
9. Inoue H (2008) Utilization of genetic algorithm for transportation planning
improvement in multi-objective logistics system. J Soc Plant Eng Jpn 19:252–259
(in Japanese)
10. Inoue H, Gen M (2012) A multistage logistics system design problem with inventory
considering demand change by hybrid genetic algorithm. IEEJ Trans Electron Inf
Syst 95(5):56–65
11. Jamrus T, Chien CF et al (2015) Multistage production distribution under uncer-
tain demands with integrated discrete particle swarm optimization and extended
priority-based hybrid genetic algorithm. Fuzzy Optim Decis Making 14(3):265–287
12. Lagodimos AG, Koukoumialos S (2008) Service performance of two-echelon supply
chains under linear rationing. Int J Prod Econ 112(2):869–884
13. Lee JE, Chung KY et al (2015) A multi-objective hybrid genetic algorithm to
minimize the total cost and delivery tardiness in a reverse logistics. Multimedia
Tools Appl 74(20):9067–9085
14. Lin L, Gen M (2008) An effective evolutionary approach for bicriteria shortest path
routing problems. IEEJ Trans Electron Inf Syst 128(3):416–423
15. Lin L, Gen M (2009) A random key-based genetic algorithm for agv dispatching
in FMS. Int J Manufact Technol Manage 16(1):58–75
16. Mahadevan B, Pyke DF, Fleischmann M (2003) Periodic review, push inventory
policies for remanufacturing. Soc Sci Electron Publ 151(3):536–551
17. Min H, Zhou G (2002) Supply chain modeling: past, present and future. Comput
Ind Eng 43:231–249
790 H. Inoue et al.
18. Minner S (2003) Multiple-supplier inventory models in supply chain management:

a review. Int J Prod Econ 81–82(02):265–279
19. Minner S, Silver EA (2005) Evaluation of two simple extreme transshipment strate-
gies. Int J Prod Econ 93–94(1):1–11
20. Miranda PA, Garrido RA (2004) Incorporating inventory control decisions into a
strategic distribution network design model with stochastic demand. Transp Res
Part E Logistics Transp Rev 40(3):183–207
21. Miyata S, Kudo K et al (2003) Mass simulations based design approach and its
environment: multi-objective optimization of diesel engine with distributed genetic
algorithm using iSIGHT MOGADES and HIDECS. Parallel computational fluid
dynamics, pp 499–506
22. Neungmatcha W, Sethanan K et al (2013) Adaptive genetic algorithm for solving
sugarcane loading stations with multi-facility services problem. Comput Electron
Agric 98(7):85–99
23. Porter ME (1985) Competitive advantage: creating and sustaining superior perfor-
mance, pp 33–69
24. Rieksts BQ, Ventura JA, Herer YT (2009) Power-of-two policies for single-
warehouse multi-retailer inventory systems with order frequency discounts. Com-
put Oper Res 36(7):2286–2294
25. Saul AT et al (1993) Numerical recipes in C. Cambridge University Press, Cam-
ridge, pp 293–295
26. Syarif A, Gen M (2003) Double spanning tree-based genetic algorithm for two stage
transportation problem. Knowl Based Intell Eng Syst 7(4):214–221
27. Taniar D, Leung C et al (2008) High performance parallel database processing and
grid databases. Wiley, New York
28. Thangam A, Uthayakumar R (2008) A two-level supply chain with partial backo-
rdering and approximated poisson demand. Eur J Oper Res 187(1):228–242
29. Zhou G, Min H, Gen M (2002) The balanced allocation of customers to multiple
distribution centers in the supply chain network: a genetic algorithm approach.
Comput Ind Eng 43:251–261
The Impact of Temporary Employment
on Employees’ Organizational Citizenship
Behavior and Turnover Intention:
The Moderating Effect of Organizational
Identification
Xiaoye Qian1 , Qian Li2(B) , Qiong Wu2 , and Yilun Wu3

1
Business School of Sichuan University, Chengdu 610064, People’s Republic of China
2
International Business School, Beijing Foreign Studies University, Beijing 100089,
liq@bfsu.edu.cn
3
School of Mechanical, Aerospace and Civil Engineering,
The University of Manchester, Manchester M13 9PL, UK
Abstract. Temporary employment, compared with permanent employ-

ment, impacts on employees’ attitudes and behaviors. This study pro-
vides empirical evidence on the relationships between temporary employ-
ment and employee’s turnover intention and organizational citizenship
behavior. Using a two-wave panel design, we collect data from a sample of
355 employees from 66 teams in four Chinese organizations. Our results
show that, temporary employees have higher turnover intention than
permanent employees, but in terms of employee’s OCB, the difference
between temporary and permanent employees is not significant. Further,
we also explore the mechanism how employee’ss organizational identi-
fication as a moderator-for temporary employees high in organizational
identification, the negative relationships between temporary employment
and OCBs and the positive relationships between temporary employment
and turnover intention would be both attenuated.
Keywords: Temporary employment · Organizational citizenship behav-

ior · Turnover intention · Organizational identification
1 Introduction
To be competitive in a globalized economy, many organizations have been using

nonstandard work arrangements to reconstruct their workforce to improve flex-
ibility, to reduce employment-related costs, therefore improve competitiveness
in product market [2]. One wide-adopted flexible employment practice con-
ducted by Chinese firm is temporary employment. Kalleberg and his colleagues
define standard work as “work done on a fixed schedule-usually full-time-at the
792 X. Qian et al.
employer’s place of business, under the employer’s control, and with the mutual
expectation of continued employment”; any arrangement that lacks one or more
of these attributes is nonstandard [7]. There are many different forms of non-
standard work arrangements, such as temporary, contract, part-time and tempo-
rary agency work, to enhance organizational flexibility and reduce employment-
related cost [7]. The temporary employment, as a most widely adopted form of
nonpermanent employment, refers to employment featured with shorter length
of contracts and lower expectations of continued employment.
According to the latest report by United Nations International Lab our Orga-
nization, in the United States, one in four employees worked part time in 2014,
up from 19.6% in 2009. In 33 European countries, 12.3% employees were aver-
agely on temporary contracts in 2014. In Asia, the portion of temporary workers
are even higher, ranging from 24% in Philippines to 67% in Viet Nam. The
percentage is sizeable in China, India, Indonesia and Malaysia.
In response to the growing use of nonstandard workers, more scholars are
conducting empirical and theoretical research on this phenomenon. Much of the
studies focus on the differences between standard and nonstandard employees in
work-related attitudes, including satisfaction, commitment, loyalty, and in-role
& extra-role behaviors, like organizational citizenship behavior, turnover and
work performance [2,5,12,14].
Other than permanent employment, Contrast with the wide-acceptance in
management practices, scholars have argued that there are quite a few draw-
backs of temporary employment. Many studies reported that temporary jobs
are connected with lower work status, characterized with low-wage, low-welfare
and high-stress [4]. Scholars also observed that temporary employees often put
in less effort than standard employees, hence alternative work arrangements
increase the difficulty in human resource management [12].
Previous studies in U.S. and Euro examined the relationship between tempo-
rary employment and employee’s work attitudes and performance. Christin and
Linn [14] recruited a sample of 350 entry-level employees from 6 restaurants in
United States, and found that part-time workers did show some organizational
citizenship behavior (OCB) differences compared to their full time colleagues.
Thorsteinson [5] did a meta-analysis to demonstrate that, full-time employees
were more involved in their jobs than part-time employees and there was little
difference between full-time and part-time employees in job satisfaction, organi-
zational commitment, and turnover intention.
Compared with the abundant empirical studies in western countries, research
on temporary employment has not attracted enough attention in China. To
bridge the gap in the literature, we use a two-wave panel design study to exam-
ine whether the different work arrangements would influence employees’ OCB
and turn over intention. We have two major contributions to the literature.
Firstly, this study is to investigate the relationship between temporary employ-
ment status and OCB and turnover intention in Chinese cultural context. Our
research fills this gap in the literature and advances current understanding on
blended workforce arrangement and employees’ OCB and turnover intention.
The Impact of Temporary Employment on Employees’ Organizational 793
Secondly, we test the hypothesized moderating effects of individuals’ organi-

zational identification on the relationships between work arrangements and
employee’s OCB and turnover intention. We propose that, the negative rela-
tionship between temporary employment and OCB will be attenuated, while
the positive relationship between temporary employment status and turnover
intention will be we propose that organizational identification play the role of
moderating the relationship between temporary employment, OCB and turnover
intention. The moderator is of great importance for us to understand how the
effect works.
This paper is organized as follows. Section 2 presents concepts and hypothe-
ses. Section 3 introduces the methodology. Section 4 summarizes the results and
Sect. 5 concludes the paper.
2 Concepts and Hypotheses

2.1 Temporary Employment
During the twentieth century, the “normal” or “standard” work arrangement
for most workers was that they were hired full-time and would work for the
organization until they retired [14]. However, to deal with the changing inter-
nal and external environment, an increasing number of enterprises realize that
they should not only boost productivity but also maintain flexibility. There-
fore, many of them started to adopt a “non-standard” employment, hiring both
“contingent” or “fixed-term contract” workers and standard employees [5].
The differences between “standard” and “contingent” employment arise from
two fundamental criteria: the duration of their employment relationship and the
level of HR investments [6]. While standard employees have sustained employ-
ment relationships and receive higher level of HR investments (e.g., economic
incentives, training, empowerment opportunities), contingent ones have employ-
ment relationships of limited duration and generally receive lower level of HR
investments [13].
2.2 Temporary Employment and Organizational Citizenship

Behavior
There are some studies on the relationships between temporary employment and
OCB. Van Dyne and Ang [3] tested their hypotheses based on a sample of 155
professional workers from 2 large service organizations, a bank and a hospital in
Singapore. They found that the relationship between two kinds of commitment
and psychological contracts and OCBs were stronger for temporary workers than
standard employees. In contrast, Conway and Briner’s analyses [1] showed that
the relationships between psychological contract fulfillment and outcomes were
rarely moderated by work status, suggesting that part-time employees would
respond in a similar way as full-time employees to adjust their psychological
contract. It seems that the relationship between temporary employment and
OCB remains in dispute.
794 X. Qian et al.
Thus, we propose the following hypothesis:

Hypothesis 1 : Temporary employees are less likely to engage in organizational
citizenship behavior, as compared to permanent employees.
2.3 Temporary Employment and Turnover Intention
Turnover intention can be measured by ideas of quitting, expectation of find-

ing alternative employment, job search behavior and intention to quit. It is an
important indicator to measure employees’ emotional condition. If an employee
has a high turnover intention, it is likely that he will not perform his job with
all the effort.
Regarding existing studies on turnover intention, Peter, Jackofsky & Salter
[11] found that there was differential predictability of turnover across full-time
and part-time employees, suggesting that the two groups may have a different
“psychology of work”. Accordingly, we come up with the second hypotheses:
Hypothesis 2 : Temporary employees are more likely to have turnover intention,
as compared to permanent employees.
2.4 The Moderating Effect of Organizational Identification
Organizational identification is defined as a perceived oneness with an organiza-

tion and the experience of the organization’s successes and failures as one’s own.
The more people identify with an organization, the more the organization’s val-
ues, norms, and interests are incorporated in the self-concept [8]. Knippenberg
and Sleebos [8] found that organizational identification has a significant impact
on turnover intention. Following this direction of this analysis, we propose the
hypothesis below:
Hypothesis 3 : Organizational identification will moderate the relationships
between temporary employees and employee’s OCB. Specifically, for employees
high in organizational identification, the negative relationships between tempo-
rary employees and OCB will be attenuated.
Hypothesis 4 : Organizational identification will moderate the relationships
between temporary employees and employee’s turnover intention. Specifically,
for employees high in organizational identification, the positive relationships
between temporary employees and turnover intention will be attenuated.
The framework of this research is shown in Fig. 1.
3 Methodology
3.1 Data
This study was conducted in 4 hospitals in Shaoxing City, Zhejiang Province of

People’s Republic of China (PRC), which are Lingzhi Town Hospital, Mashan
Fig. 1. Theoretical model of this study
Town Hospital, Sunduan Town Hospital and Fushan Community Hospital. All
the hospitals are public community hospitals with more than 150 employees,
providing service to people living in the town or the community.
The sample consists of both standard employees and temporary employees,
including doctors, nurses and other employees like drivers and accountants who
provide assistant service. Using information obtained from the leaders of 4 hos-
pitals, we distributed questionnaires to each individuals separately. We finally
got 209 valid questionnaires with a response rate of 80.2% (shown in Table 1).
Table 1. Summary of the sample
N %
Sample Valid sample 209 57.89
Invalid sample 152 42.11
SUM 361 100
3.2 Measurement
Unless otherwise noted, we use a 5-point Likert scale (“strongly disagree” = 1 to
“strongly agree” = 5) for all scales. Each scale’s coefficient α is noted below in
Table 5. Because the questionnaires are distributed in Chinese organizations, we
follow the strict procedure of translating, back-translating and cultural adjust-
ment of the original scales to generate a Chinese version.
(1) Employment Status
We collect the information of subordinates’ employment status with a three-
step procedure. Firstly, the supervisors are asked to report the subordinates’
employment status two weeks before the survey. Then in the first wave sur-
vey, the subordinates report their own employment status. The survey ques-
tion is “You are currently 1 ‘permanent employee’, 2 ‘contract employee’, or 3
796 X. Qian et al.
Table 2. Organizational citizenship behavior used in this study
Strongly Disagree Neither Agree Strongly

disagree agree nor agree
disagree
m56. I will attend functions 1 2 3 4 5
that are not required but that
help the hospitals image
m57. I will keep up with 1 2 3 4 5
developments in the hospital
m58. I will defend the hospital 1 2 3 4 5
when others criticize it
m59. I will show pride when 1 2 3 4 5
representing the hospital in
public
m60. I will offer ideas to 1 2 3 4 5
improve the functioning of the
hospital
m61. I will express loyalty 1 2 3 4 5
toward the hospital
m62. I will take action to 1 2 3 4 5
protect the hospital from
potential problems
m63. I will demonstrate 1 2 3 4 5
concern about the image of
the hospital
‘dispatched employee”’. Lastly, we identify the respondents’ employment status

through crosschecking the employment status information obtained from both
parties and asking the HR department to verify inconsistent information. In the
empirical analysis, the three different employment types are categorized into two:
1 is the standard employee, 0 is the temporary employee (contract employee and
dispatched worker).
(2) Organizational Citizenship Behavior
Organizational citizenship behavior was measured by using an 8-item OCB scale
developed by Organ [10]. Here we only used 8 items which are beneficial to
the organization. Examples are: “I will attend functions that are not required
but that help the hospital’s image”, “I will keep up with developments in the
hospital” (Cronbach’s α = .90). All the items of the OCB scale are listed in
Table 2.
(3) Turnover Intention
Four items were used to assess each employee’s turnover intention. Examples
are: “I have frequent thoughts of leaving this hospital”, “I often seek information
Table 3. Turnover intention items used in this study

disagree
m25. I have frequent 1 2 3 4 5
thoughts of leaving this
hospital
m26. I intend to find a 1 2 3 4 5
new job
m27. I often seek 1 2 3 4 5
information about other
job possibilities
m28. I will quit my job 1 2 3 4 5
soon
about other job opportunities” (Cronbach’s α = .94). All of items of the turnover
intention scale are listed in Table 3.
(4) Organizational identification
Organizational identification was measured by using a 6-item scale developed
by Mael and Ashforth [9]. Examples are: “When someone criticizes my team, it
Table 4. Organizational identification items used in this study

disagree
e24. When someone criticizes 1 2 3 4 5
my team, it feels like a
personal insult
e25. When someone praises 1 2 3 4 5
my team, it feels like a
personal compliment
e26. My teams successes are 1 2 3 4 5
my success
e27. I am very interested in 1 2 3 4 5
what others think about my
team
e28. When I talk about my 1 2 3 4 5
team, I usually say “we”
rather than “they”
e29. If others criticized my 1 2 3 4 5
team, I would feel
embarrassed
798 X. Qian et al.
feels like a personal insult”, “I am very interested in what others think about
my team” (Cronbach’s α = .91). All the items of the turnover intention scale
are listed in Table 4.
(5) Control Variables
We controlled five demographic variables, which are age, gender, education back-
ground, organizational tenure, and employees’ difference. Age and tenure were
measured in years. Gender was measured as a dummy variable, 1 for male and
0 for female. As for education background, we used 1 to represent bachelors
and above, 2 to represent technical school graduates, 3 to represent senior high
school graduates and below. Treatment difference is employees’ cognition on the
organization’s treatment differences between temporary workers and permanent
workers on income level, welfare, uniform, training opportunities, etc.
4 Results
4.1 Descriptive Analysis
Tables 5 and 6 provide the descriptive statistics and correlations of the variables
in this study. In the Table 5, the mean, standard deviation, minimum and max-
imum value of each variable are reported. As the figures in Table 5 shows, in a
final sample (N = 209), 26.3% are male. The average age of the participants
are 37.6 years old (SD = 8.76). The average organizational tenure is 11.7 years
(SD = 8.83). As for education background, we used 1 to represent bachelors
and above, 2 to represent technical school graduates, 3 to represent senior high
school graduates and below. The average education background of participants
is 1.41 (SD = 0.69).
Table 6 reports the correlation matrix of the research variables. As expected,
independent variables of employment status are significantly correlated with
organizational identification (r = 0.17), turnover intention (r = −0.12) and
OCB (r = 0.12). The significance of correlation between employment status
and the organizational identification, turnover intention and OCB indicate that
permanent employees show higher organizational identification, less turnover
intention and more OCB compared, as compared to temporary employees.
The Cronbach’s alphas of major variables are listed in Table 6. The Cron-
bach’s alphas of organizational identification, OCB and turnover intention’s
Cronbach’s alphas are 0.91, 0.91 and 0.94. All the figures exceed the accept-
able criteria, which is 0.70, which suggests that the validity of the scales chosen
in this study is acceptable.
4.2 Confirmatory Factor Analysis
To further determine the factor structure of your data set, we conduct the con-
firmatory factor analysis of OCB, turnover intention and organizational identi-
fication. For OCB, the factor loadings are range from 0.60 to 0.85. For turnover
Table 5. Descriptive statistics of research variables
Variable Number Mean Std. Dev. Min Max

Age 209 37.56938 8.755011 20 67
Edu 209 1.416268 0.689149 1 3
Gender 209 0.2631579 0.4414046 0 1
Tenure 209 11.70793 8.831362 0 46
Treatment difference 209 1.516373 0.358681 1 2.888889
ES 209 0.8373206 0.3699591 0 1
OI 209 3.844169 0.6415724 2 5
TUR 209 1.745135 0.9175642 1 5
OCB 209 3.551993 0.6366862 1 5
Note: Edu = education background; Gender is coded 0 = female, 1 = male;
Tenure = working years; Treatment difference = the differences between
permanent and temporary employees of how the organization treat them
in uniform, wage, benefits, training and promotion, 5-likert scale; ES =
Employment status, it is coded as 0 = contingent employee, 1 = regular
employee; OI = organizational identification; TUR = turnover intention;
OCB = organizational citizenship behavior.
Table 6. Correlations and cronbach’s alpha of research variables
Variable 1 2 3 4 5 6 7 8 9
1. Age 7.1 - - - - - - - -
2. Edu 0.47∗∗∗ 1 - - - - - - -
3. Gender 0.27∗∗∗ 0.128∗ 1 - - - - - -
4. Tenure 0.67∗∗∗ 0.30∗∗∗ 0.09 1 - - - - -
5. Treatment 0.03 0.14∗∗ 0 −0.06 1 - - - -
difference
6. ES −0.13∗∗ −0.53∗∗∗ −0.03 0.03 −0.32∗∗∗ 1 - - -
∗∗
7. OI −0.06 −0.11 −0.15 −0.02 −0.08 0.17∗∗ −0.91 - -
8. TUR −0.20∗∗∗ −0.03 0.08 −0.23∗∗∗ −0.01 −0.12∗∗ −0.25∗∗∗ −0.94 -
9. OCB 0.02 −0.09 0.03 −0.05 −0.01 0.12∗ 0.57∗∗∗ −0.35∗∗∗ −0.91
Note: N = 209; ∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01.
intention, the factor loadings are from 0.82 to 0.95. For organizational identi-
fication, the factor loadings are from 0.67 to 0.90. The detailed factor loading
of each item of OCB are listed in Table 7 as an illustration. Since component
reliability of each items are higher than the criteria, i.e., 0.45. The measurement
in this study is acceptable.
We then test the Hypotheses 1 and 2 first. Then after we mean-center the
variables, we use moderated regression analyses to test Hypotheses 3 and 4.
800 X. Qian et al.
Table 7. CFA of organizational citizenship behavior
Standardized Coef. OIM Std. Err. z P > |z| 95% Conf. Interval
m56←OCB 0.598 0.049 12.28 0 0.502 0.693
cons 4.19 0.216 19.37 0 3.766 4.614
m57←OCB 0.732 0.036 20.27 0 0.661 0.803
cons 4.265 0.22 19.41 0 3.834 4.695
m58←OCB 0.695 0.04 17.52 0 0.617 0.773
cons 4.222 0.218 19.39 0 3.796 4.649
m59←OCB 0.758 0.034 22.54 0 0.692 0.824
cons 4.471 0.229 19.49 0 4.022 4.921
m60←OCB 0.685 0.04 16.98 0 0.606 0.764
cons 3.715 0.194 19.11 0 3.334 4.096
m61←OCB 0.846 0.024 34.88 0 0.799 0.894
cons 4.375 0.225 19.45 0 3.934 4.816
m62←OCB 0.762 0.033 22.81 0 0.696 0.827
cons 4.834 0.246 19.62 0 4.351 5.317
m63←OCB 0.812 0.028 28.53 0 0.756 0.868
cons 4.697 0.24 19.58 0 4.227 5.165
Note: LR test of model vs. saturated: chi2(20) = 107.71, Prob > chi2 =
0.0000
4.3 Employment Status and OCB, Turnover Intention
The results of OLS regression are presented in Table 8. In Model 1, we put in 5

control variables, including age, education background, gender, tenure and par-
ticipants’ difference. None of these five control variables has significant impact
on employees’ OCB. When the independent variable, “temporary employment”,
is added into the regression in Model 2, the results show that permanent employ-
ment has a positive impact on OCB but the impact is insignificant (β = 0.202,
p > 0.1). In Model 3, we remove the control variable of education background,
then we find that permanent employment has a significant positive impact of
OCB (β = 0.251, p < 0.1). These results suggest that if education background is
not considered, standard employees show more organizational citizenship behav-
ior than temporary employees.
In Model 4, the dependent variable is turnover intention. We also put in
5 control variables, including age, education background, gender, tenure and
participants’difference. Results show that age (β = −0.018, p < 0.1) has a neg-
ative impact on turnover intention while gender (β = 0.272, p < 0.1) has a
positive impact on turnover intention. When the independent variable, “tem-
porary employment”, is added into the Model 5, results show that permanent
employment has a significant negative impact on turnover intention (β = −0.402,
p < 0.1). Therefore, Hypothesis 2 is supported.
Table 8. Regressions results
Variable Dependent variable

OCB Turnover intention
M1 M2 M3 M4 M5
Age 0.108 0.010 0.009 −0.018 −0.018
Edu −0.115 −0.055 0.108 −0.116
Gender 0.028 0.025 0.024 0.273 0.279
Tenure −0.008 −0.010 −0.010 −0.156 −0.012
Difference −0.100 0.038 0.040 −0.638 −0.160
∗∗
ES 0.201 0.251 −0.402∗∗
2
R 0.020 0.029 0.027 0.077 0.090
ΔR2 −0.004 0.000 0.003 0.054 0.070
F 0.830 1.000 0.636 3.380 3.470
Note: N = 209; ∗∗ p < 0.1, ∗∗ p < 0.05, ∗∗ p < 0.01.
4.4 The Moderating Effect of Organizational Identification

We also conduct OLS regression to examine Hypothesis 3. The results are
reported in Model 6 of Table 9. In this model, besides the control variables,
we introduce the independent variable Employment status (ES), the modera-
tor organizational identification (OI) and the interactions of ES × OI (both
variables are mean-centralized). Results show that the interactions of ES × OI
has a positive impact on employees’ OCB, but it is not significant (β = 0.178,
p > 0.1). Therefore, the Hypothesis 3 is not supported.
We follow the same steps to test Hypothesis 4. The results are demonstrated
in Model 7 of Table 9. In this model, besides the control variables, we introduce
the independent variable Employment status (ES), the moderator organizational
identification (OI) and the interactions of ES × OI (both variables are mean-
centralized). We find that the interactions of ES × OI has a significant positive
impact on employees’ turnover intention (β = 0.638, p < 0.1). Therefore, our
Hypothesis 4 is supported.
5 Conclusion
In this study, we investigate the relationships between temporary employment,
employees’ OCB and turnover intention. We also propose that employee’s orga-
nizational identification is a moderator in the above relationships. Results show
that if education is not controlled, standard employees show more OCB than
temporary employees. Compared with temporary employees, standard employees
show less turnover intention. Organizational identification positively moderates
the relationship between temporary employees and turnover intention, indicat-
ing that if a standard employee has a higher organizational identification, the
impact of employment form on turnover intention would be much stronger.
802 X. Qian et al.
Table 9. Moderating Effect
Variable Dependent variable

OCB M6 Turnover intention M7
Age 0.009 −0.019∗∗
Edu −0.048 0.001
∗∗
Gender 0.144 0.202
Tenure −0.008 −0.012
Difference −0.074 −0.149
OI 0.561 −0.385
EF × OI 0.178 0.638∗∗
2
R 0.35 0.159
2
ΔR 0.324 0.126
F 13.45 4.74
∗∗ ∗∗
Note: N = 209; p < 0.1, p < 0.05,
∗∗
p < 0.01.
Our research makes two contributions to current literature. First, we have

examined a conceptual model in cross cultural context. Although several studies
in west have investigated the relationships between temporary employment and
in-role/out-role performance, we do not know whether their results are consistent
in Chinese context. Therefore, it is theoretically important for us to investigate
these relationships in China. Second, our research demonstrates the important
role of organizational identification. More specifically, we contribute to the liter-
atures by examining and confirming organizational identification as a moderator
in the above relationships.
Our paper also provides some managerial implications. To maximize tem-
porary employees’ work efforts, it’s vital to increase their job satisfaction by
improving their salary and welfare. Meanwhile, human resource department need
to take effective actions to promote enterprise culture among temporary employ-
ees, thus raising their organizational identification and loyalty to work.
This paper also has some limitations. First, all data were collected within
a single industry, which decreases external validity. Second, we use the self-
reported data for the analysis, which may raise the possible common method
bias. We recommend that future research could collect the data from different
resources. Whether the findings of our study apply to workers in other settings
or to other types of workers, should be examined by future studies.
Acknowledgements. The authors thank for the support by the National Science
Foundation of China (Grant No. 71402108), the Humanities and Social Sciences Foun-
dation of the Ministry of Education (Grant No. 14YJC630103), the Fundamental
Research Funds for the Central Universities (No. 2016JJ019) and Young Faculty
Research Fund of Beijing Foreign Studies University (2016JT003).
References
1. Conway N, Briner RB (2002) Full-time versus part-time employees: understanding
the links between work status, the psychological contract, and attitudes. J Vocat
Behav 61(2):279–301
2. Davisblake A, Broschak JP, George E (2003) Happy together? How using non-
standard workers affects exit, voice, and loyalty among standard employees. Acad
Manage J 46(4):475–485
3. Dyne LV, Ang S (1998) Organizational citizenship behavior of contingent workers
in singapore. Acad Manage J 41(6):692–703
4. Johnson CD, Messe LA, Crano WD (1984) Predicting job performance of low
income workers: the work opinion questionnaire. Pers Psychol 37(2):291–299
5. Kalleberg AL (2000) Nonstandard employment relations: part-time, temporary and
contract work. Ann Rev Sociol 26:341–365
6. Kalleberg AL (2001) Organizing flexibility: the flexible firm in a new century. Br
J Ind Relat 39(4):479–504
7. Kalleberg AL, Reskin BF, Hudson K (2000) Bad jobs in America: standard and
nonstandard employment relations and job quality in the United States. Am Sociol
Rev 65(2):256–278
8. Knippenberg DV, Sleebos E (2006) Organizational identification versus organi-
zational commitment: self-definition, social exchange, and job attitudes. J Organ
Behav 27(5):571–584
9. Mael F, Ashforth BE (1992) Alumni and their alma mater: a partial test of the
reformulated model of organizational identification. J Organ Behav 13(2):103–123
10. Organ DW (1988) Organizational citizenship behavior: the good soldier syndrome.
Adm Sci Q 41(6):692–703
11. Peters LH, Jackofsky EF, Salter JR (1981) Predicting turnover: a comparison of
part-time and full-time employees. J Organ Behav 2(2):89–98
12. Stamper CL, Dyne LV (2003) Organizational citizenship: a comparison between
part-time and full-time service employees. Cornell Hotel Restaurant Adm Q
44(1):33–42
13. Stirpe L, Trullen J, Bonache J (2013) Factors helping the hr function gain greater
acceptance for its proposals and innovations: evidence from spain. Int J Hum Res
Manage 24(20):3794–3811
14. Thorsteinson TJ (2003) Job attitudes of part-time vs. full-time workers: a meta-
analytic review. J Occup Organ Psychol 76(2):151–177
Integration of Sound and Image Data
for Detection of Sleep Apnea
Takehiro Kasahara1,2(B) , Kota Nomura3 , Yoshihiro Ueda2 , Yuji Yonezawa2 ,

Masatoshi Saito4 , Hirohisa Toga4 , Yuki Fujimoto4 , Koji Kojima4 ,
Haruhiko Kimura1 , and Hidetaka Nambo1
1
Kanazawa University, Kakuma, Kanazawa, Ishikawa 920-1192, Japan
kasaharat@irii.jp
2
Industrial Research Institute of Ishikawa,
2-1 Kuratsuki, Kanazawa, Ishikawa 920-8203, Japan
3
Akatsuki Inc., 8F Oak Meguro, 2-13-30 Kamiosaki, Shinagawa-ku,
Tokyo 141-0021, Japan
4
Kanazawa Medical University, 1-1 Daigaku, Uchinada,
Kahoku, Ishikawa 920-0293, Japan
Abstract. To detect sleep apnea syndrome (SAS), we propose a method

that integrates sound and image processing results using overnight
sound and video data. In sound processing, snoring sounds are initially
extracted using an unsupervised method. The zero-cross ratio (ZCR) and
power (PWR) are calculated for the snoring sounds, and the total val-
idations (TV) for ZCR and PWR are calculated using the sound data.
Then, a support vector machine (SVM) is used to classify SAS and non-
SAS subjects using TV values. In image processing, inter-frame absolute
difference values are calculated every 1 s from video data, and multiscale
entropy (MSE) is calculated using the time series data. Then, an SVM
is employed to classify SAS and non-SAS subjects using two MSE values
selected based on the Bhattacharyya distance. To integrate sound and
image data, we focused on the number of detected snores. When two
results do not match and the number of extracted snores is greater than
five per hour, the sound processing result is adopted. When the num-
ber of snores is less than five per hour, the image processing result is
adopted. The proposed integrated method demonstrates better accuracy
than that obtained using only sound or image processing results.
Keywords: Sleep apnea · Hypopnea · Integration of image and sound

processing · Healthcare services
1 Introduction
Sleep apnea syndrome (SAS) is a disorder in which breathing stops during sleep.
SAS can cause sleepiness during the day, which may lead to traffic accidents.
In addition, SAS induces a serious circulatory organ disease [6]. SAS symptoms
Integration of Sound and Image Data for Detection of Sleep Apnea 805
occur in approximately 4% of men and 2% of women. If people without subjective

symptoms are included, this increases to 24% of men and 9% of woman [11],
which is a significant portion of the population. In Japan, the prevalence of
SAS is thought to be the same or more than in the United States [9]. Apnea
is a temporary cessation of breathing for more than 10 s while sleeping, and
hypopnea is an abnormally low ventilation, i.e., less than 50% of normal, lasting
more than 10s. The apnea hypopnea index (AHI) represents the occurrence
frequency per hour. Apnea is diagnosed when the AHI is five or more. Apnea and
hypopnea can be classified as obstructive, central, or mixed. Simple noninvasive
SAS diagnostic tools include pulse oximeters and apnomonitors. More reliable
diagnoses can be obtained by polysomnography (PSG). PSG is performed in a
hospital and requires multiple sensors. To collect data, the subject is required to
stay in hospital overnight. Various techniques to detect SAS have been proposed.
These techniques use sound data [2], video data [4], and a combination of depth
video and sound [10]. Previously, we proposed a method that uses microphones
and cameras, which are widely used noncontact sensors [7].
2 Methods
In this study, we have improved the previously proposed method. We used PSG
data and evaluated the accuracy of the improved method to classify SAS and
non-SAS more precisely by integrating the results of sound and image processing.
2.1 Data Acquisition
Overnight PSG data were collected from the sleep studies of 55 patients (nine
females, 61.2 ± 12.4 years) who were suspected to suffer from SAS and referred
to the Kanazawa Medical University (Ishikawa, Japan). The subjects provided
written consent to participate in the sleep studies, which were conducted in 2014.
Four subjects were diagnosed as normal and 51 were diagnosed with different
levels of SAS. To obtain more non-SAS subject data, overnight pulse oximeter
data were collected from 24 employees of the Industrial Research Institute of
Ishikawa or CosmoSummit Co., Ltd. in 2014–2015. These employees, none of
whom complained about sleep apnea symptoms, also provided written consent,
took data according to a data acquisition manual themselves. From the analysis
of the pulse oximeter data, three subject datasets were considered to indicate
SAS, and 24 subject datasets were considered to indicate non-SAS. Subject
information, such as age, gender, height, and weight, was also collected (Table 1).
The overnight sounds of the participants were recorded using a microphone
(SONY ECM-360) placed near the subject’s head. The sampling rate was
11025 Hz at 16-bit resolution. Overnight video of the participants was recorded
using a camera (IDS UI-1220LE-M-GL) with a lens (TAMRON12VM412ASIR)
that is sensitive to infrared light. The camera was placed at the side of the bed.
An example video frame is shown in Fig. 1:
806 T. Kasahara et al.
Table 1. Overview of subject data
Parameter SAS subjects: 44 male, 10 female Non-SAS subjects: 26 male, 2 female

Mean ± σ Range Mean ± σ Range
Age 59.4 ± 14.2 26–88 38.5 ± 9.9 23–66
BMI 25.8 ± 3.6 17.2–33.4 22.8 ± 3.2 17.8–30.2
AHI 38.8 ± 18.8 5.5–76.0 2.6 ± 1.1 0.6–4.7
Fig. 1. Example video frame
2.2 Sound Processing
Since various sounds, e.g., subject and bedding movements, and coughs, are
included in the recorded sound data, we used an unsupervised method proposed
by Azarbarzin et al. [1] to extract snore sounds. The unsupervised snoring sound
extraction was performed as follows. After applying a band pass filter (BPF, 150–
5000 Hz) to the recorded data, a section in which the volume is greater than
a predefined threshold continues for 0.4 to 2.0 s was extracted as an episode.
Extracted episodes include snoring episodes and episodes due to noise other
than snoring. For one episode, 50% overlap is multiplied by a 50-ms window,
and short-time Fourier transform (STFT) was performed. The STFT results for
an episode were averaged, and the spectral intensities in the range 0–5000 Hz (in
practice, BPF in the range 150 to 5000 Hz) were summed every 500 Hz to create a
10-dimensional vector. Principal component analysis was performed using the 10-
dimensional vector of all episodes. We reduced the dimensionality of the feature
space to a two-dimensional vector, and classification was performed using the
Fuzzy c-Means clustering method. In this study, we adopted the number three
classes employed in Azarbarzin’s method. Among the three classes, clusters with
a greater number of episodes were considered snoring clusters, and episodes in
this cluster were considered snoring episodes.
For snoring episodes extracted by unsupervised methods, SAS and non-SAS

were classified using the method proposed by Azarbarzin et al. [2]. The zero-
crossing rate (ZCR) and power value (PWR) were calculated for the extracted
snoring episodes. Snoring sounds are highly variable relative to the severity of
SAS. As severity increases, characteristics, such as frequency, will vary; thus,
ZCR and PWR were used to calculate the amount of change (total valida-
tion norm, TV), i.e., TVzcr and TVpwr. The calculation of TVi is shown in
Eq. (1). These two values are output as a two-dimensional feature amount for
the recorded data from a single night. However, TVzcr and TVpwr output 0
when no snoring sounds were detected. The two-dimensional features are used
for support vector machine (SVM) classification. A radial basis function kernel
(RBF) was used with the SVM, and parameters C and were determined by grid
search.
N −1
1
T Vi = |fi (k + 1)−fi (k)|. (1)
N
k=1
2.3 Image Processing

The video recordings had a duration of approximately 6 hours and were recorded
in MJPEG format at a resolution of 752 × 480 and an average frame rate of 5
frames per second (fps). Considering that the fastest physiologically important
body movements associated with SAS during sleep are respiratory movements
and the respiratory rate for adults at rest is on average 12–18 breaths per minute,
the frame rate was reduced to 1 fps. The overall motion signal was derived by
calculating the mean absolute pixel difference between successive frames.
Multiscale entropy (MSE) analysis was performed using the obtained signal.
MSE analysis is an analytical method proposed by Costa et al. [3], which is
highly useful for physiological indices, especially heart rate variability. Generally,
for a signal with high periodicity, the values obtained are lower, and the values
obtained for waveforms with lower periodicity are higher. Therefore, by adapting
the MSE analysis to the respiratory motion signal, the disorder degree of the
periodicity of the respiratory signal was evaluated and used as a classification
parameter.
MSE analysis is primarily performed in two steps. First, a coarse grain
process averages the original data x = {x1 ,x2 , · · · ,xn } by t to obtain data
y (t) = {y1 ,y2 , · · · , yN (=n/t) }. Next, entropy is measured for each coarse-grained
time series corresponding to t. Different measures of entropy can be employed
for this step. Here, sample entropy proposed by Richman and Moorman [8]
was used as the entropy metric. This metric takes two parameters, i.e., pattern
length m and similarity criterion r, which is the tolerance for accepting pattern
matches. r is a positive real value and is typically chosen between 10% to 25%
of the standard deviation of the time series. Two patterns of length m match
if each point in the first pattern is within distance r from the corresponding
point in the second pattern. The distance between two vectors is defined as the
maximum absolute difference between the components of those two vectors. A
vector xm (i) = {xi , · · · , x(i+m)−1 } where 1 ≤ i ≤ N − m is defined as a pat-

tern of length m, Note that i can only go up to N − m to ensure that vector
xm+1 of length m + 1 is also defined. nm i (r) is defined as the number of vectors
xm (j) that have distance less than r with xm (i), (i = j) to exclude self matches.
Uim (r) = nmi (r)/(N − m) is the probability that the distance between vector
xm (i) and any other vector xm (j) is less than r. U m (r) is the probability of any
two vectors of length m being within distance r of each other.
Therefore, the probability U m (r) is expressed as
N
−m
U m (r) = 1/(N − m) Uim (r). (2)
i=1
The sample entropy is defined as

U m+1 (r)
HSE (m, r) = lim − ln . (3)
N →∞ U m (r)
This is estimated by the following statistic
U m+1 (r)
HSE (m, r, N ) = − ln . (4)
U m (r)
The sample entropy was calculated for each subject’s motion signal over
scale factors t ∈ T = {1, · · · , 60}, pattern lengths m ∈ M = {2, 3, 4, 5, 6}, and
similarity criteria r ∈ R = {0.10, 0.15, 0.20, 0.25}. Therefore, a total of 1200 sets
of entropy values were calculated for all subjects.
The Bhattacharyya distance [5] is a measure of similarity of two continuous
or discrete probability distributions. In classification, it is used as a measure
of separability between two classes. The distance is calculated from the Bhat-
tacharyya coefficient ρ, which is a measure of the amount of overlap between
two probability distributions. For discrete distributions p and q over a domain
X, the Bhattacharyya coefficient is defined as follows

ρ= p(x)q(x), (5)
x∈X
where 0 ≤ ρ ≤ 1. The Bhattacharyya distance ΔB is defined as follows

ΔB = − ln(ρ), (6)
such that 0 ≤ ΔB ≤ ∞. Low ρ values or high ΔB values indicate little overlap
between p and q. The 1200 combinations of MSE parameters were analyzed and
ranked based on the distance between their corresponding sample entropy distri-
butions of SAS and non-SAS subjects. Larger Bhattacharyya distance between
the two distributions indicates that the sample entropy values are better feature
candidates for classification.
An SVM with an RBF kernel was used for classification of SAS and non-SAS
subjects. The features used in the classification were the sample entropy values,
and parameters C and σ were determined by grid search.
2.4 Integration of Sound and Image Processing Results
We performed final classification from the results obtained by image processing

and sound processing. If both results match, the result can be the final clas-
sification result. On the other hand, when each result differs, either result is
adopted by considering the reliability of the acquired data. In this study, we
considered the snoring number extracted from the sound data. When the num-
ber of extracted snores was greater than five per hour, the sound processing
result was adopted. When the number of extracted snores was less than five
per hour, the image processing result was adopted. In the data acquisition envi-
ronment, the image data could not be acquired successfully because the degree
of installation difficulty of the camera is greater than that of the microphone,
and the distance is too close or the angle does not match. Therefore, when the
extracted snore number is detected as 5 per hour or more, the sound processing
result was adopted. In addition, several cases were confirmed where snoring did
not occur even before and after the occurrence of apnea, and such cases could
not be distinguished from cases where recording failed; thus, if snoring was not
detected, the image processing result was adopted.
Fig. 2. SVM result in sound processing using all data

3 Results
In sound processing, non-SAS data with a small snore count were greatly devi-
ated from the origin when plotted on 2D feature coordinate graph, and learning
in the SVM did not provide proper results. The SVM result obtained using all
data is shown in Fig. 2. Since snoring was not detected sufficiently, data with
than five snores per hour were assumed to be non-SAS, and data classified by
the SVM for the remaining data were taken as the sound processing result. The
SVM result using data with less than five snores per hour is shown in Fig. 3:
In image processing, MSE analysis was performed to calculate the Bhat-
tacharya distance, and the parameters sorted in descending order are shown in
Table 2. The top of 2 parameter sets of (m, r, t) = (4, 0.10, 30), (2, 0.25, 34) were
adopted for SVM. The sample entropy when (m, r) = (4, 0.10) is shown in Fig. 4.
The result of SVM is shown in Fig. 5.
Accuracy was calculated by leave-one-out cross validation. With only image
processing, the accuracy was 81.7%, sensitivity was 81.5%, and specificity was
82.1%. With only sound processing, the accuracy was 85.4%, sensitivity was
85.2%, and specificity was 85.7%. These results show that sound processing
Fig. 3. SVM result in sound processing with more than five snores per hour
Table 2. Sample entropy parameters of the highest bhattacharyya dist
Order Bhattacharyya dist m r t

1 0.7140 4 0.10 30
2 0.7135 2 0.25 34
3 0.7119 2 0.25 44
4 0.7100 3 0.10 38
5 0.7074 2 0.10 49
Fig. 4. Sample entropy when (m, r) = (4, 0.10)
outperformed image processing. In the integrated result, the accuracy was 89.0%,
sensitivity was 92.6%, and specificity was 75.0%, which shows a decrease in
specificity compared to the sound processing results; however, improvements to
accuracy and sensitivity were observed. Table 3 compares classification results
and the final results for image and sound processing.
Overnight PSG was used to obtain the number of apnea events and the
number of hypopnea events. Subjects were classified as apnea or hypopnea type
by comparing these events. In the sound processing results, there were eight
false negatives out of 54 positive cases; however, in the integrated results, false
negatives were reduced to four. In this case, among the four subjects that changed
from false negatives to true positives, three subjects were hypopnea type. Of the
19 apnea type subjects, only one was mistaken as a false negative with sound
processing. Thus, sound processing is effective for detecting apnea type SAS but
less effective for hypopnea type SAS. However, by integrating the sound and
image processing results, it is possible to reduce errors by half.
Fig. 5. SVM result in image processing using all data
Table 3. Individual results
Process Image Sound Integrated

Accuracy (%) 81.7 85.4 89.0
Sensitivity (%) 81.5 85.2 92.6
Specificity (%) 82.1 85.7 75.0
Positive predictive value (%) 89.8 92.0 87.7
Negative predictive value (%) 69.7 75.0 84.0
Screening methods that can easily detect SAS are expected. Therefore, we
analyzed sound data and image data obtained using microphones and cam-
eras, which are inexpensive and familiar sensors. We have proposed a screening
method to classify SAS and non-SAS subjects. In this study, we focused on the
number of snores per hour. When the number of extracted snores was large, we
used sound processing due to good accuracy. However, low accuracy was obtained
with sound processing when the number of snores was small; thus, we proposed a
method using image processing. The experimental results demonstrate that the
proposed integrated method (sound and image data) can screen at high accu-
racy compared to screening with only sound or image data. Furthermore, the
proposed method is effective for screening hypopnea SAS.
Microphones and cameras are standard sensors on smartphones; therefore, it

can be expected that SAS screening can be performed easily using smartphones
by applying the results of this study.
References
1. Azarbarzin A, Moussavi Z (2011) Automatic and unsupervised snore sound extrac-
tion from respiratory sound signals. IEEE Trans Biomed Eng 58:1156–1162
2. Azarbarzin A, Moussavi Z (2012) Snoring sounds variability as a signature of
obstructive sleep apnea. Med Eng Phys 27:479–485
3. Costa M, Goldberger A, Peng C (2002) Multiscale entropy analysis of complex
physiologic time series. Phys Rev Lett 89:068102
4. Gederi E, Clifford G (2012) Fusion of image and signal processing for the detec-
tion of obstructive sleep apnea. In: 2012 IEEE-EMBS International Conference on
Biomedical and Health Informatics, pp 890–893
5. Kailath T (1967) The divergence and bhattacharyya distance measures in signal
selection, communication technology. IEEE Trans Commun Technol 15:52–60
6. Kario K (2009) Obstructive sleep apnea syndrome and hypertension: ambulatory
blood pressure. Hypertens Res 32:428–432
7. Nomuara K, Kasahara T et al (2015) Development of a system for detecting
obstructive sleep apnea by combining image and sound processing. J Nanjing Univ
Aeronaut Astronaut 27:12–20
8. Moorman Richman J JS (2000) Physiological time-series analysis using approx-
imate entropy and sample entropy. Am J Physiol Heart Circulatory Physiol
278:H2039–H2049
9. Tanigawa T, Tachibana N et al (2004) Relationship between sleep-disordered
breathing and blood pressure levels in community-based samples of japanese men.
Hypertens Res 17:479–484
10. Yang C, Cheung G et al (2016) Sleep apnea detection via depth video & audio
feature learning. IEEE Trans Multimedia 1–5
11. Young T, Palta M et al (1993) The occurrence of sleep-disordered breathing among
middle-aged adults. New Engl J Med 17:1230–1235
Pricing Strategy Study
on Product Crowdfunding
Min Luo1,2,3 , Yong Zhang4(B) , Huanmei Zhang4 , and Xia Li4

1
Guanghua School of Management, Peking University,
2
Mathematics Department, Chengdu University of Information Technology,
3
4
School of Mathematical Sciences, Beijing Normal University,
zhangyong@bnu.edu.cn
Abstract. The pricing strategy of crowdfunding has always been its

highlights and difficulties. A pricing strategy formula concerning with
the price discrimination is obtained through the in-depth studies of the
price discrimination theory and the mechanism of crowdfunding pricing,
considering the crowdfunding financial limits which constitute constraint
conditions. This article carries out a study on the effect of advertising
profit of crowdfunding and founds the profit function. Furthermore, it
studies the model of crowdfunding based on the 1P theory to solve this
issue. Finally, the article provides a corresponding pricing strategy on
product crowdfunding based on the above studies. It is expected to con-
tribute to the field of theoretical research on crowdfunding and the pric-
ing strategy through this article, and provide inspiration and reference
for enterprises and entrepreneurs when they make the pricing strategy
decisions.
Keywords: Crowdfunding · Pricing strategy · Price discriminate ·

Profit
1 Introduction
Product crowdfunding refers that the investors put their money into develop-
ing a product or service in accordance with the fundraisers; when the product
or service begins to be pre-sold or has the conditions for external sales, the
fundraisers will provide the developed product or service for the investors. Prod-
uct crowdfunding needs a third party to be supported by the public. The relevant
information of the products, services and producers can be obtained through the
platforms, which would save transaction cost while realizing a certain advertising
effect.
Pricing Strategy Study on Product Crowdfunding 815
The consumers participating in the crowdfunding have strong interests in

the project products when these are in the creating or initial stage, willing to
provide funds to make the products available as soon as possible, and the price
paid for each unit may not be lower than that of the product later on. It is an
external direct financing that a project can be conducted smoothly when the
crowdfunding reaches the planned fixed amount, which can reduce the difficulty
of financing for the companies and can predict the market’s reflection and selling
through probing the participating action.
It is necessary to establish an adequate pricing and feedback module for a
successful project. In general, in order to attract the public to support the par-
ticipation of supporters, the project launchers will announce the prices of the
products and a combination of preferential prices on the platform pre-sale in
advance to the market. In addition, the use of the price discrimination strategy
can tell the people who have different needs apart to sell. Grasping the psycho-
logical expectations of consumers and taking differentiated pricing not only can
promote the products and return the funds ahead of time, but also can maximize
the corporate profits.
2 Literature Review
For the platforms of the product crowd funding, pre-sale is generally the main
way, which functions to find prices well and can help the project launchers or
entrepreneurs to create some favorable conditions for price discrimination. Pigou
[8] defines that the price discrimination refers to different prices charged by firms
for different markets, consumers or different numbers of purchases, and he also
puts forward a three-level price discrimination theory. As a kind of theory, price
discrimination belongs to the category of the pricing strategy without judging.
In the industries with more competition, price discrimination is widely used in
a variety of flexible forms. It is an effective pricing strategy, not only helping to
enhance the competitiveness of enterprises to achieve its business objectives, but
also adapting to the psychological differences of consumers to meet the needs
at multiple levels. Rochet and Thanassoulis [10] declares that the price discrim-
ination is an economic phenomenon and a trading strategy, which is popularly
used by the sellers; and according to economic thinking, the implementation of
the price discrimination strategy for the sellers is positive in terms of the welfare
and can thus expand their own profits.
Nocke et al. [7] studied the relations between the product crowd funding and
the price discrimination in the context of asymmetric information. In an intertem-
poral setting in which the individual uncertainty is resolved over time, advance-
purchase discounts can serve to the price discrimination between consumers with
different expected valuations for the product. Consumers with a high expected val-
uation will purchase the product before learning the actual valuation at the offered
advance-purchase discount; consumers with a low expected valuation will wait and
purchase the goods at the regular price only in the occasion where their realized
valuation is high. The consumers with higher expectations of participation in the
816 M. Luo et al.
product quality in the case of unknown product quality should enjoy discounts and
have the lower prices than those for the regular customers. Enterprises have to tar-
get at the customers with different expectations. Belleflamme et al. [11] isolated
some important features of crowdfunding on the basis of a unique, hand-collected
dataset and proposed a model of crowdfunding that encompasses several of these
key features. By constructing the profit function, they obtained the pricing expres-
sion of the product and the ownership of the product. On this basis, the enterprise
financing limit is given, and the uncertainty of the product quality and informa-
tion are elaborated. Hu et al. [2] and Ping [9] maximized the profits of producers as
the objective function optimization model to analyze the pricing mechanism of the
crowdfunding products, drawing the conclusion that the model will help to open up
the initial public offering of the product market, and the profits of manufacturers
can be expanded with the same total sales.
Lawton and Marom [5] pointed out that crowdfunding is not only a new
source of funding for projects, but also a way to quickly attract the attention of
users, form early user communities; and because of the large number of media
reports and discussions, pre-release has become a common corporate marketing
strategy, so crowdfunding is still a very effective marketing tool. Based on the
choices of consumers, Bayus [1] applied the model of maximizing their own prof-
its to jointly determine the time choice and pricing decision from the consumer
utility function. Zhang et al. [14] and Luo et al. [6] constructed a two-stage pric-
ing model based on the impact of IWOM on customers’ perceived value to the
product. IWOM can change the cognitive value of customers to the products,
the customers who have profound insight seek the overall effectiveness of maxi-
mizing products purchasing and brands spreading, and the dominant sellers can
maximize the two-stage total profits through using reasonable pricing strategies.
Huang [3] and Krishnan et al. [4] elaborated the pricing strategy and analyzed
four strategies, such as new product pricing strategy and discount strategy, which
can help enterprises make full use of pricing strategy to provide reasonable mar-
ket prices to obtain the most competence. But all the above literatures just study
the traditional price strategy and cannot explain the popular free phenomenon
nowadays. 1P business model proposed by Wang [12,13] breaks the pricing space
of the business strategy by integrating the third party to obtain a new competi-
tive advantage, with putting the pricing into the center position and innovatively
stressing positioning and locking the third party to pay, which can make profit
with the price that is lower than the average cost and can promote with the
price that is higher than the expectation of customers.
In this paper, through the deep research on price discrimination theory and
crowdfunding pricing mechanism, we not only consider the influence of crowd-
funding price on needs, but also further study the effect of network brand spread-
ing and advertising, constructing the maximum profit function of the product
crowdfunding in the two occasions to obtain the pricing expression under the
constraint for financing amount. This paper also studies the pricing theory of
business model proposed by Wang [13] and applies it into the crowdfunding
business model with drawing the corresponding pricing strategy.
3 Crowdfunding Pricing Model

3.1 Crowdfunding Pricing Model for Price Discrimination
According to the level of discrimination, the price discrimination can be divided
into three levels. The first-degree price discrimination, also known as complete
price discrimination, regulates the price according to the possible maximum
amount of money paid by each buyer to obtain the full consumer surplus. Based
on different consumptions or “segments” to obtain a different price, the second-
degree price discrimination only regulates different prices for different segments
and thus obtains part of the consumer surplus. The third-degree price discrimi-
nation classifies the customers according to a particular standard to make clear
which category each customer belongs to, with taking different prices in different
markets or to different customer groups for one kind of product.
The first and the second degrees of price discrimination enable the manu-
facturers to grab the consumer surplus wholly or partly, and this part of the
consumer surplus will be transferred into profits, achieving the principle of effec-
tive allocation of resources P = M C (equilibrium price equals to marginal cost.
That is the requirement of maximizing profits). However, the first-degree price
discrimination cannot be fully realized in reality. The pricing of crowdfunding
products will usually take the pricing strategies of the first and second degrees
of price discrimination, using direct information on needs to choose the con-
sumers indirectly through the choices done by them about different packages.
The inverse elasticity rule is taken for different consumer groups, meaning that
the consumers who have a high price elasticity will be charged low and the con-
sumers who have a low price elasticity will be charged high. The maximization
of profits can be obtained through the price combinations.
By the function of demand and price
Qd = d(P ).
The demand function of the consumers belonging to different levels in the

price combinations is
Qn = (An − Pn )/Ep ,
here An is the price that the consumers are willing to pay and Ep refers to the
price elasticity of demand.
Assuming that β is the variable cost component of producing a unit of this
product and the initial funding amount K is the fixed cost amount put by the
firm to produce the required product smoothly. The total cost of product is

n
C =K +β Qi .
i=1
The total profit of the project producer is

n
n
R= Pi Qi − K − β Qi .
i=1 i=1
818 M. Luo et al.
Only if the total payment of the participant is not less than the total amount
of financing K can the raisers get the money from the platforms smoothly. It
requires:
n
Pi Qi ≥ K.
i=1
To maximize the profits of crowdfunding, it is required to consider how to
design the relations between various levels of prices and quantities, that is, to
figure out the maximum profit objective function with some constraints

n
n
max R = Pi Qi − K − β Qi
i=1 i=1
⎧
⎪
⎪ Pn = An − Ep Qn ,
⎨ n
s. t. Pi Qi ≥ K,
⎪
⎪
⎩ i=1
Pi ≥ 0.
3.2 The Crowdfunding Price Model Concerning with the

Advertising Effect
As a result of extensive and effective brand and spreading effects, the product
crowdfunding not only can sell the products but also reach the goal of advertising
through the release on the Internet platform. Compared with the common adver-
tising tunnels, the influences of crowdfunding include a lower price and sound
effects. However, considering that the effect is related much to the demand and
price, it is still necessary to research the pricing strategy prudently to maximize
the profits.
Assuming the same amount of customer traffic, denote the conventional unit
of product advertising budget by Gr and the cost of crowdfunding platform by
Kc , the function relation between effect profits and the demands of the rather
common advertisements is

n
Gc = Gr Qi − K,
i=1
where the function relation between the price and the demand is
Qn = (An − Pn )/Ep .
An refers to the price which the crowdfunding consumers are willing to pay and
EP is the price elasticity of demands.
Assuming that β is the variable cost component of producing a unit of this
product and the initial funding amount K is the fixed cost amount that the firm
can produce the required product smoothly. The total cost of product is

n
C = K + Kc + β Qi .
i=1
The total profit of the project producer is

n
n
R= Pi Qi − K + Gc − β Qi .
i=1 i=1
Only if the total payment of the participant is not less than the total amount
of financing K can the raiser get the money from the platform to produce
smoothly. It needs to meet with:

n
Pi Qi ≥ K.
i=1
Taking the advertising effect into account, it is needed to figure out the
maximum of profit objective function with constraints to obtain the maximum
profits of the crowdfunding

n
n
max R = Pi Qi − K − Kc + (Gr − β) Qi .
i=1 i=1
3.3 The Crowdfunding Pricing Model Using the 1P Theory

1P theory refers that the enterprises can expand the profit space through inte-
grating the third party to benefit mutually, to participate in income payments
and cost sharing and to break the upper and lower limits of strategic pricing.
The third party represents the value of all associated networks, and is the value
of the associated network of the enterprise and its customers and partners.
Crowdfunding as a business pattern can also be explained by the 1P theory
which wields the third party to pay, and all the crowdfunding patterns can be
categorized in the 1P theoretical business model. The product crowdfunding
views that the investor as a third party to share some or all of its fixed cost,
channels and promotional cost, thereby reducing the cost of the product and
improving the total profit of the product. Therefore, the product is also the
value logic and trading structure in which the companies, customers and other
third-party partners can jointly create and trade value to make mutual benefits.
In the 1P theory, the 4Ps are separated according to the nature of the income
and cost, where the price P is the income per unit of product; the product P ,
the channel P and the sales promotion P are referred to as 3P, which is the unit
cost (that is, the average cost AC and 3P = AC). π refers to the unit profit, and
Q means the production, and π = P − 3P = P − AC, therefore the total profit
is π = Q(P − AC). Here, P − AC is both the pricing space and profit margin.
According to the linear value hypothesis of strategic marketing, the product
price P is paid by the target customer C, so P = P C. The unit product cost
AC is burdened by the enterprise E itself, so AC = ACE, and the product price
must be greater than or equal to the average cost P ≥ AC, that is P C ≥ ACE,
or there is money loss. When the profit is zero, π = P − AC = 0, i.e. π =
P C − ACE = 0, P C = ACE.
820 M. Luo et al.
The introduction of the third-party investors in the product pool will make
the unit product price P no longer equals to the target customer’s payment
price P C, for it should be added with the price of PB paid by the third-party
investors B, that is, P = P C + P B, P C = P − P B; when the profit is zero,
π = (P C + P B) − AC = 0, P C = AC − P B < AC. This proves that even if
the price of the product sold to the target customer is less than the price of the
product and the average cost, it is also profitable, and how much the profit will
be depends on the price paid by the third party and the number of the shared
cost.
The theoretical business model of 1P whose pricing space breaks the upper
and lower limits of pricing is shown in Fig. 1.
Fig. 1. The business pricing space model of 1P theory
4 Crowdfunding Pricing Strategies

4.1 The Crowdfunding Pricing Strategy Using the Price
Discrimination
The crowdfunding products are faced up with different consumer groups who
have different demand elasticity to products, creating some conditions for the
implementation of price discrimination. The consumers who have higher elastic-
ity can be charged with lower prices to increase sales to gain profits while the
consumers who have lower elasticity can be charged with higher prices to raise
the profit without decreasing the sales obviously. Considering from elasticity, the
supporters of crowdfunding are willing to pay higher prices. When considering
uncertainty about the quality of products, supporters of the crowdfunding will
face a greater degree of uncertainty, then they need to be given with discounts
and the price should be lower. Hence, it is not clear which part of the consumers
will face higher prices in the crowdfunding projects. In addition, because the
consumers in the crowdfunding projects can choose whether to participate in at
their own accordance and the products are marked with personalized features,
it is assumed that consumer groups do not trade with each other, and producers
can implement price discrimination.
Obviously, price discrimination can make the sellers benefit as much as pos-
sible, for the consumer surplus belonging to the buyers will also be transferred
to the sellers through price discrimination. Price discrimination is economically
efficient. The maximum benefit of sellers equals to the value of social welfare
maximization through price discrimination. Of course, the sellers have to be able
to distinguish different characteristics of the buyers to make price discrimination
work. This difference may exist in the buyers’ demand intensity, the quantity
of purchase or the price elasticity of demand. The basic principle of implement-
ing price discrimination is that the marginal returns in different markets are
equal and equal to marginal cost. The crowdfunding management can regulate
a higher price for the market with less price elasticity of demand, implementing
the strategy of “less sales but more profit”, while a lower price can be regulated
for the market with greater price elasticity of demand, and the maximum profit
can be obtained through the implementation of “less profit but more sales”.
4.2 The Crowdfunding Pricing Strategy Considering the

Advertising Effects
Plenty of crowdfunding products are marked as innovative products. If the prod-
ucts are new to the market, it can be considered in the price combination about
the three pricing strategies including skim-milk pricing, penetrating pricing and
satisfying pricing. Also, as a new product to come out, to select the crowdfund-
ing is also a working advertising way which can lead to a sound effect. Thus, the
advertising effect needs to be considered in the crowdfunding pricing strategy,
and the ways mentioned above should be implemented flexibly according to the
actual situation to achieve profit maximization.
Skimming pricing is a high-priced pricing strategy in the early stage of a new
product, provided that the product’s quality and image must support the high
price of it. Also, consumers have a strong desire for consumption and there is no
potential competitor in the market. In this case, consumers are willing to pay
for high prices; also, the elasticity of demand, sale and the overall profit income
are rather high.
Penetrating pricing strategy is a low-priced strategy when a new product
first comes into the market, in which many consumers will be attracted by the
low price to win a sound reputation and a comparatively large market. The
prerequisite of this pricing is: the market is highly price-sensitive and the cost of
producing and selling will be reduced with increasing sales. Also, the competitors
can be expelled through a low price. Such a disruptive and innovative pricing
method tends to stimulate consumers’ willingness to buy and bring about sound
advertising effects, resulting in more profits.
Satisfying pricing strategy is a strategy between skimming pricing and pen-
etrating pricing. It is an intermediate pricing which can satisfy both producers
and consumers. The prerequisite of such a pricing is that this product is com-
paratively mature and the consumers are familiar with the product. Also, the
information is symmetry. The consumers can buy this product according to their
own preferences. If the consumers are satisfied with the price, more attention of
822 M. Luo et al.
them will be placed on the product experience, then the high-quality product
will win a good name, and the manufacturers will thus obtain rich profits.
In short, it needs to take the actual situation of crowdfunding products into
account to choose skimming pricing, penetrating pricing or satisfying pricing.
Considering the cost of advertising and error-trying, the maximum product profit
can be gained with using a crowdfunding way to sell and make a sound pricing
strategy because of the novelty and advertising needs of these products.
4.3 The Crowdfunding Pricing Strategy Using the 1P Theory

The 1P theory claims that any business model is a revenue-cost transaction
structure in which customers, third-party associated customers, and third-party
production partners participate in co-creating and trading network values. In the
business model of the product crowdfunding, the consumer may be a producer
or the producer may be a consumer at the same time. The third-party associated
customers who pay the prices can be either consumers or enterprises; the third-
party production partners who share costs can be either a business or a consumer.
Transaction cost is the most common network value in the network economy, and
the emergence of the third parties makes transaction cost lower. The reduction in
the number of transaction equals to the result of the number of brands multiplied
by the number of customers minus the number of brands then adds the number of
customers, and the difference will be multiplied by the cost of each transaction.
The result is the transaction cost saved by the third-party integrated brands
and the customers. Therefore, the number of the third parties can be expanded
through the pricing strategy to lower the transaction costs in a greater degree
to maximize profits.
The business model innovation of crowdfunding should break the lower and
upper limits of the commercial competitive pricing to access to new competi-
tive advantages and expand profit space through discovering, creating, trading
network values and integrating the third-parties’ paying. In the formulation of
pricing strategies, it should be focused on discovering, creating and capturing the
network values, paying attention to the cooperation among businesses, customers
and partners to jointly benefit from expanding the total values. Also, it needs to
be considered to use the ecological network thinking to re-examine the economic
nature of the businesses, customers, competitors, suppliers and even the whole
commercial environment to assess the values of the associated network within
or outside the enterprise and find out the high-volume crowdfunding-targeted
customers to jointly create and trade network values, achieving the purpose of
win-win cooperation.
5 Conclusion
5.1 Theoretical Contribution
In this paper, we studied profoundly the price discrimination theory and the
pricing mechanism of the crowdfunding pricing. By considering the effect of the
crowdfunding pricing on demand to compose the profit function and the con-
straint conditions of the financing, it was figured out the crowdfunding pricing
expression concerning with the price discrimination. Compared with the tradi-
tional sales modes, the total sales of the crowdfunding products remain the same,
while the proportions of crowdfunding consumers and conventional consumers
change oppositely. The crowdfunding mode can distinguish the different pref-
erences of different consumer groups, according to which the market’s attitude
towards the products is likely to be judged. The unit price given by crowdfund-
ing consumers is higher than that of conventional consumers. The crowdfunding
raisers are faced up with a funding ceiling, and if the ceiling is touched, the
crowdfunding will be a failure. Not only can the new medium and small enter-
prises finance without much difficulty by crowdfunding, but also more profits
can be achieved for the businesses compared with the traditional sale modes.
Also, we further studied the maximum profit function of the product crowd-
funding with the advertising effects. Finally, we studied the pricing theory of
the business model in which the third parties pay in the theory of 1P, and then
applied it to the business model of crowdfunding to get the corresponding crowd-
funding pricing strategies, aiming to obtain the maximum profit. The paper is
expected to contribute to the theoretical research in the crowdfunding financing
and pricing field, and also provide references for enterprises and entrepreneurs
to make decisions on pricing strategies for their products.
5.2 Marketing Significance

Pricing strategy is an important part of the marketing combination for a busi-
ness. Enterprises need to prudently formulate the pricing strategies of the prod-
ucts when making the crowdfunding plans. Only by innovating and conforming
to the objective laws of the market economy can the enterprise win more profits
in the target market and achieve a better reputation and advertising effects.
Making the right crowdfunding pricing strategy is not only an important factor
in maintaining the interests of manufacturers and customers, but also the key
to overcome competitors, develop and consolidate the market.
5.3 Limitations and the Future Research Prospects
Due to the short rising period of the Internet crowdfunding, the domestic and
foreign literature about its product pricing is correspondingly scarce. Crowd-
funding product pricing research also belongs to the emerging field, which has
great research value and profound potential. Moreover, with the rapid iteration
of technologies and business models, the industry of crowdfunding is changing
incessantly and the pricing strategy research should follow the development of
the situation. This paper gives the theoretical expression of the crowdfunding
pricing using the price discrimination as well as the maximum profit function
considering the effects of advertising, and there may be more factors which should
be taken into account when planning the actual cases.
824 M. Luo et al.
With the support of the policies, the developing potential of crowdfunding

model is not to be limited, especially in the items of culture innovation and tech-
nologies. Young people are more likely to express themselves and communicate
on the Internet, thus, the consumer groups of the products of culture innovation
and technologies match well with the main active users on the Internet. With
the help of the crowdfunding platforms and the social network, the crowdfund-
ing projects can be focused as soon as possible. Not only can the money be
obtained but also the sound advertising effects can be achieved. If we introduce
the analyzing technologies of large data into pricing strategies, it may be more
accurate in terms of dimensions and data processing to make the strategies more
effective.
Acknowledgements. This project was supported by the National Natural Science

Foundation of China (11501035, 70831005) and the Key Project of China Petroleum
and Chemical Corporation (GJ-73-0706).
References
1. Bayus BL (1997) Speed-to-market and new product performance trade-offs. J Prod
Innov Manag 14(6):485–497
2. Hu M, Li X, Shi M (2014) Product and pricing decisions in crowdfunding. Soc Sci
Electron Publishing 34(3):331–345
3. Huang ZY (2012) Discussion on product pricing strategies. J Nanchang Coll Educ
27(11):12–18 (in Chinese)
4. Krishnan TV, Bass FM, Jain DC (1999) Optimal pricing strategy for new products.
Manag Sci 45(12):1650–1663
5. Lawton K, Marom D (2013) The crowdfunding revolution: how to raise venture
capital using social media (ebook). McGraw Hill, New York
6. Luo Y, Gao L, Huang J (2015) Price and inventory competition in oligopoly tv
white space markets. IEEE J Sel Areas Commun 33(5):1002–1013
7. Nocke V, Peitz M, Rosar F (2011) Advance-purchase discounts as a price discrim-
ination device. J Econ Theor 146(1):141–162
8. Pigou AC, Todd AJ (1934) The economic of welfare. Am J Sociol 19(1):23–30
9. Ping YY (2015) The pricing mechanism of commodity crowdfunding. Zhongnan
Univ Econ Law 40(3):23–30 (in Chinese)
10. Rochet JC, Thanassoulis JE (2016) Stochastic delivery and intertemporal price
discrimination with multiple products. Social Science Electronic Publishing
11. Sahm M, Belleflamme P, Lambert T et al (2014) Corrigendum to “crowdfunding:
tapping the right crowd”. J Bus Ventur 29(5):610–611
12. Wang JG (2015) An investigation into modeling marketing. J Peking Univ
52(4):95–110 (in Chinese)
13. Wang JG (2016) The classification, innovation and design of business model in
business ecosystem. Peking Univ 6(1):1–17 (in Chinese)
14. Zhang MX, Lei M, Zheng XN (2013) The impact of online word-of-mouth dissem-
ination on the oligarch seller’s pricing strategy. Mark Sci 9(2):71–89 (in Chinese)
The Support System of Innovation-Driven
Strategy in Private Enterprises: A Theoretical
Model
Tao Xie1 , Guichuan Zhou1(B) , and Shuangyi Zheng2

1
zhouguichuan@vip.163.com
2
School of Management of South-Central, University for Nationalities, Wuhan
430074, People’s Republic of China
Abstract. The theory of Innovation System has been used to under-

stand the complexity and instability of innovation. Recently Strategy
management researchers have found Innovation-Driven Strategy is trans-
ferred from national level to enterprise level. The dualistic Innovation
System is dominated by government agencies and enterprises in parallel,
which is less conducive to innovation of private enterprises. The dualistic
main-body Innovation System is converted into the multi-nested Innova-
tion Support System by building a Support System of Innovation-Driven
Strategy (SSIDS) in private enterprises. Through analyzing the connota-
tion, structure and characteristic, a new theoretical model is pioneered
that serves as a guide for innovation strategy in private enterprises.
Keywords: Private enterprises · Innovation-driven strategy · Support

system
1 Introduction
Managers should put additional importance on innovation that plays a significant
role as a bridge between competitive strategies and firm performance in a devel-
oping economy environment [1]. As stepping into the “new normal”, the devel-
opment of economy is not investment-driven, but innovation-driven in China,
which reflects the reconstruction of economic engine and switching of develop-
mental strategy [12]. Much of research has examined that private enterprises
(PE) are superior to state-owned enterprises (SOE) about innovative ability
[8]. According to the statistics, private enterprises have more R&D projects,
R&D investment and patents than state-owned enterprises. Therefore, private
enterprises with innovative potential play a leading role in innovation-driven
strategy. However, market mechanism with inherent weaknesses has blocked the
development of innovation in China. “Perfect market” also breeds Opportunism
which jeopardizes innovative performance of PEs [4]. Although with huge inno-
vative potential, PE faces systemic risk. When in the immature market, entre-
preneurs improve transformational leadership with the expectation of favorable
826 T. Xie et al.
business environment that the government creates. The dualistic Innovation Sys-
tem dominated by government agencies and enterprises in parallel is extensively
implemented in China [16]. There are some divergences concerning relationship
between the government and enterprises, which roots in complexity and instabil-
ity of innovation. The theory of Innovation System cannot solve new problems
of PEs.
How does PE build the internal support system to keep innovative perfor-
mance? How does the government improve the business ecosystem without inter-
ference to PE This article will solve these problems and develops a theoretical
model of Innovation-driven Support System, which can change traditional inno-
vation paradigm and offer a framework to further empirical studies. To do this
we build our discussion on extensive relevant literatures and enterprise practices.
Using methods and theories of the system, the Support System of Innovation-
Driven Strategy (SSIDS) is built.
In the following sections, we first review current researches and state defi-
ciencies. We then go on to analyze connotation, structure and characteristic of
SSIDS. Finally, we discuss the implications of our theoretical model for managers
and provide directions for further research.
2 Literature Review
The inflexion point theory demonstrates that according to the decreasing law of
marginal utility, contribution degree of basic elements, such as natural resources
and capital, present descending trend. Therefore, the long-term development of
economy depends on intellectual and technological factors. The inflexion point
theory promotes innovation-driven theory, which manifests investment-driven
strategy is replaced by innovation-driven strategy. The concept of innovation-
driven was initially used by Porter [13]. Porter argued that economic develop-
ment could be divided into four periods: factor-driven stage, investment-driven
stage, and innovation-driven, wealth-driven stage. In innovation-driven stage,
the effect of basic factors and investment is led by innovation and a system of
innovation-driven strategy taking technological innovation as core is built.
After that innovation-driven development became national strategy, Chinese
academics conduct quiet a few examinations regarding innovation-driven strat-
egy. Liu [11] contended that the innovation-driven development means the
change of economic engine in China. And he further demonstrated that the
growth of economy has transformed from depending on learning and imitation
into relying on self-designing, R&D, knowledge-creating. In addition, Zhang [17]
analyzed the main features of innovation-driven development, i.e. people ori-
ented, acquiring first-mover advantage and entrepreneur-driven. Li and Wang
[10] proposed internal innovation-driven elements including enterprise culture,
leaders full of entrepreneurship, expected return; and external innovation-driven
elements including technical progress and institutional environment. Meanwhile,
WANG Tao and Qiu [16] emphasized that the innovative environment domi-
nated by entrepreneurs is immature and constructed innovation-driven strategy
The Support System of Innovation-Driven Strategy 827
of “two-way drive” effect model that dominated by government and enterprises

in parallel.
Innovation origins in Economics, gradually progresses to Management. Mean-
while, innovation-driven development transfers from national strategy to enter-
prise strategy and entrepreneurs play a major role in innovation. Implementing
the innovation-driven strategy is complicated and nonlinear systems engineering.
Therefore, many studies have advanced the combination between innovation and
system and developed the theory of Innovation System, such as National Inno-
vation System (NIS) [15], Regional Innovation System (RIS) [5], Technological
Innovation System (TIS) and Innovation Ecosystem (IE) [2,6]. However, more
recent studies argued that Innovation Support System (ISS) is different Inno-
vation System and suggested the performance of innovation actors is restricted
by environment, in China. He and Li [7] proposed how the Innovation Support
System is established. In the process of innovation, various innovative elements
generate and spread new knowledge and technology by interaction to form a
network system.
In summary, innovation-driven strategy not only depend on the core compe-
tency of enterprises, but need a favorable business environment. Previous studies
built the Innovation System that is a dualistic model, which cannot clarify the
role of the government and enterprises. The Innovation Support System provides
a feasible theoretical framework for innovation-driven strategy in private enter-
prises. It is how private enterprises effetely combine their core competency and
environmental support from the government mostly, which is the deficiency of
present literature.
3 The Structure of SSIDS Model
It is complex system engineering for private enterprises to implement the

innovation-driven strategy, which is influenced by various factors inside and out-
side the system. Elements form inherent complex relationships by interaction,
such as resources, capabilities and environment, which is regarded as a system.
Therefore, it is necessary to build a system to support the innovation-driven
strategy in private enterprises.
3.1 Strategic Support System
Strategic support system is a collection of subsystems composed by a set of

key managerial issues, which support the realization of strategic objectives. This
system can be divided into internal and external support subsystems. Internal
strategic support subsystem is constituted by internal key managerial issues
within the process of strategic implementation and usually includes three parts:
impetus subsystem, behavior subsystem and capability subsystem. Specifically,
the focus of impetus subsystem is to form power of strategic implementation;
the focuses of behavior subsystem include deciding, organizing, motivating and
controlling behaviors during strategic implementation; the focuses of capability
828 T. Xie et al.
subsystem is mainly about the optimization of enterprise capabilities including

strategic management, R&D, capital operation and marketing. External strate-
gic support subsystem is the collection composed of various external factors
and their interactions in the process of implementation of strategies. It mainly
refers to external supportive environment, for example, market competition envi-
ronment, resource environment, policy & laws environment and public service
environment.
Support System of Innovation-Driven Strategy (SSIDS) in private enterprises
refers to the organic integration and coupling operation of enterprises’ internal
and external supporting elements. In this system, various elements promote and
restrict mutually to seek the balance point supporting the implementation of
private enterprises’ innovation-driven strategy. Therefore, this system will help
to enhance the willingness of private enterprises to implement innovation-driven
strategy, facilitate the exchange of internal and external factors such as material,
energy and information during strategy implementation process, and further
improve their innovative performance.
SSIDS in private enterprises includes two basic parts: internal and external
support system. It covers every supportive element inside and outside enterprises
during the process of the implementation of innovation-driven strategy, centering
on the improvement of innovative performance.
3.2 Internal Support System of SSIDS
Internal support system is a relatively stable network system formed endoge-

nous in private enterprises, consisting of innovation support elements and their
mutual relationships. The process of innovation-driven development in private
enterprise can be described as: generate innovation impetus, implement inno-
vation behavior, form innovation capability, which leads to the formation of
innovation impetus subsystem, innovation behavior subsystem and innovation
capability subsystem, as illustrated in Fig. 1.
Fig. 1. Internal support system
(1) Impetus Subsystem

The degree of innovation impetus directly influences the speed and scale
of innovation in private enterprise. According to the different sources, innova-
tion impetus can be divided into internal impetus supply and external impetus
supply. As showed in Fig. 2, internal impetus supply includes resource, entre-

preneurship and corporate culture; external impetus supply includes technologi-
cal advancements, market demands, competitive pressure and the support from
government. Impetus of resource comes from continuously innovation resources’
injection to internal innovation system by enterprises. These innovation resources
include technical personnel, innovative capital and technical equipment etc.
Abundant resource can improve the success rate of innovation, increase inno-
vative returns, and thereby strengthen the willing of implementing innovation
strategy.
Fig. 2. Impetus subsystem
Entrepreneurial spirit becomes the impetus of continuous innovation in pri-

vate enterprises through entrepreneurs’ self-driven and self-motivation. The inno-
vative spirit of entrepreneurs can motivate more members to participate the
practice of innovation-driven strategy and further improve the market perfor-
mance.
Enterprise innovative culture can influence staff’s values, thoughts and behav-
iors, which infect and motivate innovation of organization, intrinsically and
intangibly. Innovative culture performs its function as guiding, constraint, coher-
ence and motivation [14].
When converted to differentiated products or services, achievements in sci-
entific and technological progress can bring higher monopoly profits. Also that
motivates enterprises consistently to absorb new results of scientific and techno-
logical progress and to re-innovate. The impetus and influence of technological
progress performs as: the induction of new technology ideas, technical track and
technical expectations.
Market demand is the basic starting point of implementing innovation-driven
strategy for private enterprises, which provide new market opportunities and
innovative ideas for enterprises. In fact, the market demand has the function of
guiding the direction of technological invention, testing the results of technolog-
ical innovation and providing technical promotion space.
The pressures of competition enhance the crisis awareness of private enter-
prises. Meanwhile, the incentive of innovation can also be triggered by more
market share and monopoly profit. Under fierce competition, the private enter-
prise will have more intense willing of innovation, more emphasis on market
intelligence, more advanced concept and shorter product update cycle.
830 T. Xie et al.
The government support can make up inherent weaknesses of market. Inno-

vation activities characterized of high risks and externalities, lead to the decline
of innovative impetus of private enterprises. To reduce this negative effect, the
government should give private enterprises appropriate support such as making
up for the overflowed loss of innovation, building innovative infrastructure and
sharing risk in innovation.
(2) Behavior Subsystem
When facing various stimulations from external environment, private enter-
prises make responses by implementing innovation-driven strategy. The behav-
iors of innovation-driven strategy integrate the interaction of macro and micro,
vertical and horizontal, external environment and internal factors, and finally,
form a closely related dynamic open system. Figure 3 illustrates the composition
of innovation behavior subsystem of private enterprises, the subsystem includes
four aspects: policy-making, organization, motivation and control. According
to the theory of organizational behavior, the evolution direction of private
enterprises’ innovation behavior is: policy-making behavior organization behav-
ior motivation behavior, moreover, the control behavior is penetrating in each
behavior stage. As the elements of the system, different behaviors also have
reverse exchange of information and energy. The detailed path is shown in Fig. 3.
Fig. 3. Behavior subsystem
Behaviors of policy-making imply that managers revise innovation utility

function and the risk function by collecting information and decide whether to
carry out the innovative activities, which intends to reduce the uncertainty dur-
ing innovation process. Managers with strong policy-making ability can not only
grasp the market information accurately and enhance the willingness of imple-
menting innovative activities, but also can form effective methods and means to
improve the success rate of innovative activities.
Behaviors of organization refer to that private enterprises conduct optimized
recombination of innovation elements such as technical knowledge, innovation
capital and intelligent team in the process of strategic implementation [3]. The
quantity, quality and synergy of these factors determine the output and risk
of innovative activities. Technical knowledge is the foundation of innovation,
which determines the feasibility and risk of innovation. Capital is the blood
of innovation, which ensures stability and continuity of the innovative activities.
Intelligent team is the brain of innovation; high quality and compound innovative
team provides innovation activities the insurance of science and sustainability.
Behaviors of motivation include a series of specific regulations and reward-

punish system, by which behaviors of staff are in line with the goal of inno-
vative activities, and their potential are exerted fully. The implementation of
innovation-driven strategy is catalyzed by innovation incentives such as spiritual
incentives and material incentives. The intensity of incentive will affect the desire
and responsibility of innovative activities. Meanwhile, the way of incentive will
influence the relationship between innovation behavior and expected returns.
Behaviors of control imply that managers diagnose, adjust and feed back all
other behaviors, which can reduce the loss of error behavior, and improve the
incomes of correct behavior. Specifically, behaviors of control includes perfor-
mance evaluation, risk monitoring, modification, feedback and so on. Remark-
ably, control behavior should run through the whole innovation process because
innovation activities are characterized of extreme complexity and uncertainty.
(3) Capability Subsystem
Innovation capability refers to the probability, efficiency and continuity that
enterprises carry out innovative activities and obtain innovative income. As
shown in Fig. 4, the capability is mainly embodied in knowledge stock and
innovative output. With strong innovation capability, enterprise can constantly
develop new products, expand the development of the enterprise space, and
improve the level of its profitability to maintain strong innovative motivation [9].
Fig. 4. Capability subsystem
Knowledge stock consisting of explicit knowledge and tacit knowledge is the

key element of enterprise growth, resource allocation and innovation. The flow
and upgrading of knowledge can nurture flexible thoughts and enhance the recog-
nition of innovation strategy. It is the basis of implementing innovation-driven
strategy to improve quantity, quality and structure of the knowledge stock.
Innovative outputs, generally speaking, include new products and intellectual
property rights. Once the innovation results are converted into new products
with competitiveness, enterprises can attain revenue from innovation. The higher
the market value of new products is, the stronger the willingness of enterprise
to innovate will become. As the means to ensuring the benefits of innovation,
intellectual property rights have an important impact on the sustainability and
stability of the strategy. The more intellectual property rights you have, the
more security your innovative performance will gain.
832 T. Xie et al.
3.3 External Support System of SSIDS
There are lots of external elements of enterprise that are closely related to
innovation-driven strategy. External support system is exactly consisted of these
elements and their mutual relationship together. It is necessary to implement the
innovation-driven strategy and can be generalized as the environmental subsys-
tem of innovation. As shown in Fig. 5, the system commonly includes market
environment, policy environment and service environment.
Fig. 5. External support system
(1) Market Environment

Influenced by historical reasons and institutional factors, Chinese private
enterprises are facing more and special market environment. Market environ-
ment mainly includes the market rule, the fairness of market competition and
the degree of government’s intervention to market. With the help of adequate
market rules, essential market ethics and credit, less intervention from govern-
ment, private enterprises can compete in a fair market environment and further
improve their willingness and performance of innovation.
(2) Legal Environment
Laws & regulations play an important role of ensuring and adjusting inno-
vative activities of enterprises. Therefore, government should focus on creating
suitable environment and help to promote innovative activities in enterprise.
For example, the authority can exert its influence on launching, organizing
and implementing of enterprises’ innovative programs by constantly adjusting
and perfecting the laws and regulations on intellectual property rights and the
standards of environment & energy. These laws, especially intellectual property
protection regulations, can help private enterprises to enhance technical barri-
ers, protect the property rights of innovation, and improve performance from
innovation.
(3) Service Environment
The insurance of technology, personnel and capital is necessary for private
enterprises to implement innovation-driven strategy. Therefore, service environ-
ment is vital for innovative activities. First, it can serve as bridge connecting
technology and economy by providing private enterprises mature intermediary
institutions technology and market. Second, it helps to introduce and select all
Fig. 6. Support system of innovation-driven system
kinds of innovative talents for private enterprises. Finally, it promotes the forma-
tion of a sound financial market so that private enterprises can raise innovation
funds from this important capital source.
Based on above analysis of internal and external support system, we build a
new support system of innovation-driven strategy in private enterprises with the
purpose of improving performance of this strategy (as illustrated in Fig. 6). The
support system of innovation-driven strategy totally differs from the traditional
innovation systems that define the role of the government and private enterprise.
The SSIDS is an open, dynamics and complex system as it involves a wide range.
4 Conclusion
This article analyzes internal support system form innovative impetus, behav-
ior and capability, discusses external support system from market mechanism,
laws and regulations, service. Based on internal support system and external
support system, a new theoretical model, SSIDS, is built, which is conductive to
innovative practices of private enterprises.
Innovation is unpredictable system engineering. According to Schumpeter’s
theory of innovation, entrepreneurs are situated in primary status. However, at
the initial stage of the market, Chinese private enterprises lack of mature entre-
preneurs. Therefore, traditional theory of Innovation System insists that both
the government and enterprises are situated in primary status, which is harmful
to development of Chinese entrepreneurs. Contrary to the dualistic main-body
Innovation System, the SSDIS with a multi-nested structure is instrumental and
assistant. In the SSDIS, the government provides service and supervision; private
enterprises carry out innovative practice. This article builds a theoretical model
and change innovation paradigm. Nevertheless, many empirical studies are nec-
essary to the validation and optimization of SSIDS model. How the SSIDS model
operates is a valuable question for further research.
Acknowledgements. We thank the support of the Projects of the National Social

Science Foundation of China (16AGL003), Social Science Planning Annual Project of
Sichuan Province (SC16C011), Soft science research plan project of Sichuan province
834 T. Xie et al.
(2017ZR0098, 2015ZR0091), Soft science research project of Chengdu city (2015-RK00-

00156-ZF), The Central University Basic Scientific Research Business Expenses Project
of Sichuan University (skqy201623, skqy201209), Projects of the Center for systems
science and enterprise development (xq15B11).
References
1. Bayraktar CA, Hancerliogullari G et al (2016) Competitive strategies, innovation,
and firm performance: an empirical study in a developing economy environment.
Technol Anal Strateg Manage
2. Carlsson B (1991) On the nature, function and composition of technological sys-
tems. J Evol Econ 1(2):93–118
3. Chakraborty S, Thompson JC, Yehoue EB (2016) The culture of entrepreneurship.
J Econ Theory 163(3):288–317
4. Chen B (2014) On the connotation and characteristics of innovation driven and the
condition of realizing innovation driven: from the perspective of realizing “China
dream”. J Theor Ref 7:123–134 (in Chinese)
5. Cooke PN, Heidenreich M, Braczyk HJ (2004) Regional innovation systems: The
role of governances in a globalized world. Eur Urban Reg Stud 6(2):187–188
6. Edquist C (1997) Systems of innovation: technologies, institutions and organiza-
tions. Soc Sci Electron Publ 41(1):135–146
7. He M, Li F (2009) Innovation supporting system of Zhongguancun science park. J
Beijing Inst Technol (Soc Sci Ed) 2:46–55 (in Chinese)
8. Jiang S, Gong L, Wei J (2011) The path to the catch-up of the innovative ability
of the late comers of enterprises in the transitional economy: comparing SOE with
POE. Manage World 12:43–72 (in Chinese)
9. Lancker JV, Mondelaers K et al (2016) The organizational innovation system: a
systemic framework for radical innovation at the organizational level. Technovation
52–53:40–50
10. Li Y, Wang J (2014) The research on the relationship between driving factors and
performance of innovation in the context of vertical integrated industry chain. Res
Finan Econ Issues 7:21–29 (in Chinese)
11. Liu Z (2011) From late-development advantage to first-mover advantage: theoretic
thoughts on implementing the innovation-driven strategy. Ind Econ Res 4:65–72
(in Chinese)
12. Liu Z (2015) Innovation-driven strategy in new high-level opening-up. J Nanjing
Univ (Philos Humanit Soc Sci) 2:37–47 (in Chinese)
13. Michael P (1990) The competitive advantage of nations. Ashgate, Brookfield
14. Mir M, Casadesüs M, Petnji LH (2016) The impact of standardized innovation
management systems on innovation capability and business performance: an empir-
ical study. J Eng Technol Manage 41:26–44
15. Trevor M (1989) Technology policy and economic performance. lessons from Japan.
R&D Manage 19(3):278–279
16. Wang T, Qiu G (2014) A research on “bidirectional-driven” effects of innovation-
driven strategy. Technoeconomics Manage Res 6:44–57 (in Chinese)
17. Zhang L (2013) Innovation-driven development. China Soft Sci 1:80–88 (in Chi-
nese)
The Emission Reduction and Recycling of Coal
Mine Water Under Emission Allowance
Allocation of Government
Ning Ma and Shuangyi Zheng(B)

zhengshuangyi@scu.edu.cn
Abstract. Water shortage and water pollution problems have become

increasingly severe, especially in these areas with the coal mining indus-
try. In this paper, a bi-level multi-objective optimization model based on
the trade-and-cap mechanism with fuzzy coefficients is developed to deal
with environmental water problems in large scale coal fields, in which
both the emission right and mine water recycling are considered. This
model fully considers the relationship between the government and the
coal mines and also the equilibrium between economic development and
environmental protection. To deal with the inherent uncertainties, the
model is defuzzified using a possibility measure, and a solution app-
roach based on the Karush-Kuhn-Tucker condition and interactive fuzzy
programming is designed to search for the solutions. A case study is
presented to demonstrate the practicality and efficiency of the model.
The results showed practicality and efficiency in the proposed model of
reducing the coal-water conflict. Finally, a comprehensive discussion is
provided and some propositions is given as a foundation for the proposed
management recommendations.
Keywords: Water environment · Coal mining · Mine water recycling ·

bi-level multi-objective programming
1 Introduction
Coal is still the most abundant fossil fuel in 2015 and it provided around 30.0%
of the global primary energy need, though oil and natural gas reserves have
increased over time. As such an important fossil fuel contributing to worldwide
energy generation, coal will still takes a huge market share in world energy and
coal mining is still a major industry for developing economy. But coal mining can
cause a lot of pollution for water, soil and air environment, and these ecological
environment issues will deepen the contradiction between the land and humans,
and constrains economic and social development. However, for safety, signifi-
cant quantities of groundwater are discharged in underground mining [17,19]. If
836 N. Ma and S. Zheng
untreated, the water combining with heavy metal ions such as iron and mercury
ions would definitely pollute the groundwater [7], which has many pollutants
and waste the cherish water resource [16,18]. Thus, it is of importance for the
coal mine to control the mine drainage.
In order to mitigate the environmental damage, several policy instruments
have been developed to attempt to reduce pollution emissions, such as emis-
sion taxes, command-and-control, and cap-and-trade [5,6,9]. The cap-and-trade
mechanism, also known as the emission trading scheme (ETS), is an applica-
tion of Coase [3], and has showed its effectiveness of controlling emissions and
has been successfully applied in practice [2]. Therefore, this paper adopts this
mechanism to control the mine drainage. In this mechanism, the initial carbon
emission allowances are defined and allocated by local government for free or at
auction or a combination of both [4,20] and coal mines decide their coal pro-
duction under the emission allowance. Based on this mechanism, there are two
kinds of decision makes in this problem. However, government only concerns the
whole benefit and coal mines only consider their individual benefits. There is a
conflict between government and coal mimes. Therefore, in order to solve this
conflict, this paper adopts bi-level programming to balance the benefit between
the whole and individuals. Meanwhile, this paper also considers the effect of
mine water recycling on reducing waste water emission. Coal mines can have
more coal production by recycling mine water to reduce waste water emission
under the limited emission right.
Based on the above discussion, this paper proposes an optimization model
that integrates a trade-and-cap mechanism and waste water recycling to mitigate
mine water pollution. In Sect. 2, the government use to control pollution through
emission allowance allocation, and the coal mines treat and reuse mine water to
guarantee production under limited emission right. Then, as an abstraction of
the real problem, a bi-level multi-objective model is built based on the discus-
sion and a algorithm is applied in Sect. 3. In Sect. 4, a case study is presented
to demonstrate the significance of the proposed model and solution method.
Conclusions and future research directions are given in Sect. 5.
2 Problem Statement
In this paper, we consider a trade-and-cap mechanism to control the total waste
water emission. Government decides the number of permits to discharge specific
quantities of a specific mine water per time period. Coal mines need to decide
their production plan according to permits allocated by government, because
the quantity of mine water is related to the coal production. Meanwhile, coal
mines can reduce their actual waste water emission though mine water recycling.
In this problem, government hope to minimize the total waste water emission
in the terms of ecology. But coal production is an important finance source
for government. It’s unreasonable to only consider ecology or economy singly.
Therefore, government have to make a balance between economic and ecological
benefits. And coal mines as profitable organizations only consider their economic
benefits. The relationship between government and coal mines is shown as Fig. 1.
The Emission Reduction and Recycling of Coal Mine Water 837
Fig. 1. The logic diagram of emission right allocation and mine water recycling
3 Modelling
Based on the analysis above, the hierarchical interests equilibrium between gov-
ernment and coal mines, and the economic and ecological conflicts are respec-
tively the key points of solving the mine water emission reduction. Therefore,
we adopt the bi-level programming to achieve the interests equilibrium which
DMs in a hierarchical relationship to solve the emission right allocation, and the
multi-objective programming to balance the economic and ecological benefits.
3.1 Notation
Index:
i: Coal mine index, i = 1, 2, · · · , I;

j: Coal mine water treatment method index, j = 1, 2, · · · , J, and the quality of
treated water is from high to low when j from 1 to J, and method J can just
make the mine water reach the emission standard.
Parameters:
ω: Tax rate of coal sale;

H: Maximum environment carrying capacity;
P C: Unit coal price;
CiC : Unit coal mining cost of coal mine i;
Bij
R
: Unit economic benefit of reusing the waste water treated by method j in
coal mine i;
Cij
R
: Unit economic cost of treating the waste water by method j in coal mine
i;
θi : Coal drainage coefficient when coal mine i produces unit coal.
Decision variables:
Qi : Waste water emission allowance of government allocating to coal mine i;
xi : Coal production amount of coal mine i;
yij : Waste water amount of coal mine i treating by method j.
3.2 Objective for Government

(1) Economic Objective
For the region with abundant coal resource, coal industry is generally the main
industry influencing the local economic development and human salary level.
Therefore, local government must consider the economic objective, especially
in developing country. Generally speaking, the financial revenue of the local
government is determined by local tax revenue, central tax returns and transfer
payments. The coal industry’s contribution to regional financial revenue comes
mainly from taxes. Therefore, the economic objective of government can be
expressed as:

I
C
max F1 = (P − C C ) × xi × ω, (1)
i=1
(2) Ecological Objective

Environmental protection is a major responsibility for local government, and this
is especially important in large scale coal mining regions. In terms of environmen-
tal protection, government hopes to minimize the actual waste water emission.
Therefore, it is the optimal for ecological objective when total emission allowance
is minimum.

I
min F2 = Qi . (2)
i=1
3.3 Constraints for Government

(1) Maximum Environment Carrying capacity
Emission right allocation is to control the environmental damage effectively.
Therefore, the total waste water should be under the maximum environment
carrying capacity H. However, it is very difficult to estimate the exact value of
the constraint due to the inherent uncertainty and complexity. Therefore, we
adopt fuzzy number to express this kind of uncertainty.

I
Qi ≤ H. (3)
i=1
3.4 Objective for Coal Mine
For the market-based coal mines, pursuit of the highest profit is the absolute
priority when the managers make decisions. Generally, coal mines’ profit come
mainly from coal production and sales. In this paper, we also consider the recy-
cling and utilization of coal mine water. So the economic influence coming from
waste water reuse and selling or buying emission right should be considered in
the economic objective.

J
J
max fi = (P C − C C ) × xi + r
(Bij − Cij
r
) × yij − Cid × (xi × θi − yij ).
j=1 j
(4)
3.5 Constraints for Coal Mine
(1) Mining Ability Restrictions

Coal mines’ mining ability is limited and can not be quickly improved. Therefore,
when coal mine i makes a decision about the quantity that should mined, this
quantity must not exceed its mining ability:
xi ≤ Cimax , i = 1, 2, · · · , I. (5)
(2) Available Waste Water Restrictions

The amount of mine water drained by coal mine is relevant to the coal exacting
amount, and the total amount of treated mine water cannot exceed the total
mine water amount. Therefore, we can get

J
yij = xi × θi , i = 1, 2, · · · , I. (6)
j
(3) Recycling Technology Restriction

To clean mine water to different quality, coal mine needs to use different meth-
ods. Obviously, the treated mine water amount by method j cannot exceed the
maximum treatment capacity of this kine of technology the coal mine having.
Therefore, the relationship can be described as:
0 ≤ yij ≤ Tij , i = 1, 2, · · · , I, j = 1, 2, · · · , J. (7)
(4) Recycling Water Demand Restriction

The treated mine water will be used in different method according to their qual-
ities respectively, but users’ demands for water are limited. Therefore, treated
mine water cannot exceed the user’s demand, otherwise, the excessive water will
be wasted.
0 ≤ yij ≤ Dij , i = 2, · · · , I, j = 1, 2, · · · , J − 1. (8)

(5) Environmental Protection Restriction

To protect water environment, the actual waste water emission cannot exceed
the allowance allocated by government. And the actual waste water emission
should be the waste water without recycling. So the relationship is as following:

J
xi × θi − yij ≤ Qi , i = 1, 2, · · · , I. (9)
j
3.6 Global Model

Based on the previous discussion, by integrating Eqs. (1)–(9), the following global
expected model of bilevel programming can now be formulated for the problem:

I

max F1 = (P C − C C ) × xi × ω
i=1

I
min F2 = Qi
i=1
⎧ I
⎪
⎪ i=1 Qi ≤ H
⎪
⎪
⎪
⎪ Qi ≥ 0, i = 1, 2, · · · , I
⎪
⎪
⎪
⎪ max fi = (P C − C C ) × xi + j=1 (Bij
J r
− Cij
r
) × yij (10)
⎪
⎪
⎪
⎪ − Ci × (xi × θi − j yij )
d J
⎪
⎪ ⎧
⎨ x ≤ Cimax , i = 1, 2, · · · , I
⎪
⎪
s.t. ⎪
⎪
i
J
⎪
⎪ ⎪
⎪ xi × θi ≤ j=1 Tijr + T d i = 1, 2, · · · , I
⎪
⎪ ⎪
⎨ J
⎪
⎪
⎪ j yij ≤ xi × θi , i = 1, 2, · · · , I
⎪
⎪
⎪
⎪ ⎪
⎪ 0 ≤ yij ≤ Tij , i = 1, 2, · · · , I, j = 1, 2, · · · , J
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪ 0 ≤ yij ≤ Dij , i = 2, · · · , I, j = 1, 2, · · · , J − 1
⎪
⎩ ⎪
⎩ J
xi × θi − j yij ≤ Qi , i = 1, 2, · · · , I.
4 Solution Approach
In this section, we adopt a KKT to solve the bi-level problem and fuzzy goals
programming to solve multi-objective problem.
4.1 Fuzzy Goal Programming for Bi-Level Multi-objective Model

The basic concept of fuzzy goal programming approaches is that each lower level
DM considers the upper level DM’s goal when seeking to optimize their own
objective function. In the decision process, the fuzzy goal membership functions
for all the DMs’ variables are considered and the fuzzy programming problem is
solved using a constraint on the overall satisfactory degree of any upper levels.
[13,14]. As membership levels are in the same dimension, this method can also
solve the weighted sum scalarization with different dimensions in model (10)
Meanwhile, this method has been previously adopted to solve multi-objective

model problems in different areas [1,10] and has proved to be highly effective.
Therefore, this paper adopts fuzzy goal programming to transform the bi-level
multi-objective model into a single level and single objective model.
First, the optimal solution for each objective of each decision maker fξB is
calculated and the DMs decided on the worst value fξW they can accept (ξ is the
objective index), from which the fuzzy goal f1 (x, y, z) f1B and f2 (x, y, z) f2B
can be determined, and their membership functions are as Eq. (11)
⎧
⎪ 1, if fξ (q, x, y) ≤ fξB ,
⎨
fξ (q,x,y)−fξW
μξ (fξ (q, x, y)) = fξB −fξW
, if fξB < fξ (q, x, y) ≤ fξW , (11)
⎪
⎩
0, if fξ (q, x, y) ≥ fξW .
Using Eq. (11), the multi-objective model can be transformed into a single
objective model by weighting and summing the negative deviations between the
actual satisfactory level and the aspiration satisfactory level for each objective,
as shown in model (12):

max F = ωξ × μξ (Fξ (q, x, y))
ξ∈Ξ
μ̄ij − μij (fij (q, x, y)) ≤ σ̂

s.t. (12)
(q, x, y) ∈ S,
where the auxiliary variable ωξ is the importance degree for objective ξ

and ξ∈Ξ ωξ = 1; μ̄ij is the individual aspiration levels of each coal mine;
μij (fij (q, x, y) is the actual the practical satisfaction level of each decision maker;
σ̂ is the maximum deviation between the practical satisfaction level and the aspi-
ration satisfaction level of each lower decision maker; S is the feasible region of
Eq. (10). By minimizing the weighted sum of these deviation, decision makers
can determine an overall satisfactory solution close to the optimal solution.
However, based on these characteristics, however, there is a limitation to the
use of interactive fuzzy programming as it can only be used to solve decision
making problems in organizations with cooperative behavior. And the solution
of model (12) is not optimal for lower decision makers. To solve this problem,
we adopt KKT to transform the lower model to guarantee the optimal solution
of lower decision makers.
4.2 KKT
Karush-Kuhn-Tucker (KKT) approach has been proven to be a valuable analy-
sis tool with a wide range of successful applications for bi-level programming
[11]. The fundamental strategy for the KKT approach is that it replaces the
follower’s problem with its KKT conditions and appends the resultant system
to the leader’s problem [8,12]. Based on previous work by Shi et al. [15], we
transformed the bi-level model into a single-level model. However, KKT cannot
guarantee the equality between lower decision makers when there are multiple
lower decision makers in a bi-level problem. But in this paper, we have used
fuzzy goal programming to transform the bi-level model into single model and
can guarantee the benefit balance between lower decision makers by the satisfac-
tory deviation. Therefore, in this paper, we just adopt KKT to seek the optimal
solution of lower model under upper decision maker’s strategy. And the model
(10) is finally transformed as Eq. (13).
max F =ω1 × μξ (F1 (q, x, y)) + ω2 × μξ (F2 (q, x, y))
⎧
⎪
⎪
I
F1 = i=1 (P C − C C ) × xi × ω
⎪
⎪
⎪
⎪ F2 = i=1 Qi
I
⎪
⎪ ⎧
⎪
⎪ ⎪ 1, if fξ (q, x, y) ≤ fξB ,
⎪
⎪ ⎨
⎪
⎪ f (q,x,y)−f W
⎪ μξ (fξ (q, x, y)) =
⎪
ξ ξ
, if fξB < fξ (q, x, y) ≤ fξW ,
⎪
⎪ ⎪
⎩
fξB −fξW
⎪
⎪ if fξ (q, x, y) ≥ fξW ,
⎪
⎪ 0,
⎪
⎪
⎪
⎪ μ̄ij − μij (fij (q, x, y)) ≤ σ̂
⎪
⎪
⎪
⎪
I
i=1 Qi ≤ H
⎪
⎪
⎪
⎪ Q i ≥ 0, i = 1, 2, · · · , I
⎪
⎪
⎪
⎪ x i ≤ Ci
max
, i = 1, 2, · · · , I
⎪
⎪
⎪
⎪
J
xi × θi ≤ j=1 Tijr + T d i = 1, 2, · · · , I
⎪
⎪
⎪
⎪ J
⎪
⎪ j yij ≤ xi × θi , i = 1, 2, · · · , I
⎪
⎪ ≤ yij ≤ Tij , i = 1, 2, · · · , I, j = 1, 2, · · · , J
⎪
⎨ 0
s.t. 0 ≤ yij ≤ Dij , i = 2, · · · , I, j = 1, 2, · · · , J
⎪
⎪ J
xi × θi − j yij ≤ Qi , i = 1, 2, · · · , I
⎪
⎪
⎪
⎪ (P − C ) − Cid × θi + v1 + (v2 − v3 + v11 ) × θi = 0
C C
⎪
⎪
⎪
⎪ J
⎪
⎪ j=1 (Bij − Cij ) + Ci + v3 + v4 − v5 + v6 − v11 = 0
d
⎪
⎪
⎪
⎪ v1 g1 (Qi , xi , yij ) + v2 g2 (Qi , xi , yij ) + v3 g3 (Qi , xi , yij )
⎪
⎪
⎪
⎪ +v4 g4 (Qi , xi , yij ) + v5 g5 (Qi , xi , yij ) + v6 g6 (Qi , xi , yij )
⎪
⎪
⎪
⎪ +v7 g7 (Qi , xi , yij ) =
⎪
⎪
0
⎪
⎪ g (Q , x , y ) = C max
− xi ≥ 0, i = 1, 2, · · · , I
⎪
⎪
1 i i ij i
⎪
⎪
J
g2 (Qi , xi , yij ) = j=1 Tijr + T d − xi × θi ≥ 0, i = 1, 2, · · · , I
⎪
⎪
⎪
⎪ g3 (Qi , xi , yij ) = xi × θi − j yij ≥ 0,
J
⎪
⎪
⎪
⎪
⎪
⎪ g4 (Qi , xi , yij ) = Tij − yij , i = 1, 2, · · · , I, j = 1, 2, · · · , J
⎪
⎪
⎪ g5 (Qi , xi , yij ) = yij ≥ 0, i = 1, 2, · · · , I, j = 1, 2, · · · , J
⎪
⎪
⎪
⎪ g6 (Qi , xi , yij ) = Dij − yij ≥ 0, i = 2, · · · , I, j = 1, 2, · · · , J
⎪
⎪
⎩ g (Q , x , y ) = Q − x × θ + J y ≥ 0.
7 i i ij i i i j ij
(13)
5 Case Study
5.1 Case Presentation
Lu’an coal field is located on the southeastern edge of the middle eastern Qin-
shui coal field, which is in the center of Shanxi Province. There are nine major
collieries (i.e., i = 4) in this area. These are the Shiyijie mine, the Changcun
mine, the Zhangcun mine and the Wangzhuang mine.
In Lu’an coal field, there is average annual rainfall of 584 mm, and an aver-
age annual evaporation of 1732 mm. With the capacity expansion of the coal
mines, coal mining quantity has significantly increased, and this caused mine
drainage increases simultaneously. Moreover, untreated mine water damages the
surrounding water environment. Meanwhile, due to the water from the mine
drainage, the groundwater supplement is insufficient, which affects plant growth
and adds to environmental degradation. Therefore, this paper adopt Lu’an coal
field as a case to show the practicality of the model.
5.2 Data Collection
Detailed data for the research region were obtained from the Statistical Yearbook
of Chinese coal industry (National Bureau of Statistics of the Peoples Republic
of China (2013)), the Statistical Yearbook of Chinese energy (National Bureau
of Statistics of the Peoples Republic of China (2013)) and field research. Certain
parameters taken from the Statistical Yearbooks and data published by the
companies are shown in Tables 1, 2 and 3.
5.3 Results and Discussions
First, the upper and lower bounds of the government and coal mines objective
functions were calculated to identify the membership function for the fuzzy goals,
Then, we can build the fuzzy goals though the optimal solutions and worst
solutions of each objective of government and coal mines. And we set ω1 =
ω2 = 0.5, μ̄1 = μ̄2 = μ̄3 = μ̄4 = 1. Finally, through adapting the maximum
deviation σ̂, we can get some groups optimal solutions under different attitude
of government, as Table 5.
Table 1. Parameters for government
Government H (104 t) ω
1000 0.17
Table 2. Parameters of coal mine (1)
r (RMB/m3 )
Bij r (RMB/m3 )
Cij r (104 m3 )
Tij r (104 m3 )
Dij
j=1 j=2 j=3 j=1 j=2 j=3 j=1 j=2 j=3 j=1 j=2 j=3
i = 1 2.8 2.8 4.1 2 1.7 1.5 5 20 120 26 38.5 359
i = 2 2.8 2.8 4.1 0 1.66 1.49 0 15 90 20 28.1 240
i = 3 2.8 2.8 4.1 0 1.72 1.52 0 10 75 17 30 305
i = 4 2.8 2.8 4.1 1.98 1.59 1.4 3 34 100 24.9 35.7 367
Table 3. Parameters of coal mine (2)
Cimax (104 t) Pic (yuan/t) Cic (yuan/t) θi (m3 /t) Tid (104 m3 ) Cid (RMB/m3 )
i = 1 230 780 219 2.15 500 0.63
i = 2 170 780 230 2.12 320 0.6
i = 3 136 780 227 2.14 283 0.57
i = 4 250 780 220 2.12 490 0.61
Table 4. Pay-off table with 5 individual single-objective problems
max F1 min F2 max f1 max f2 max f3 max f4

F1 68914.81 29885.33 44858.20 30641.72 28683.95 45683.80
F2 1000 0 1000 1000 1000 1000
f1 128173.37 60852.47 169949.24 12156.43 12156.43 12156.43
f2 94745.76 27240.28 62580.98 77884.18 62580.98 62580.98
f3 75799.63 32076.54 55084.51 55084.51 77157.82 55084.51
f4 140199.70 57563.49 18477.52 18477.52 18477.52 132121.31
Table 5. Optimal solutions under different attitude of government
Qi xi (104 yij (104 m3 ) fi (104 μ(fi )i F1 F2 μ(F )

(104 t) RMB) (104 (104
m3 ) RMB) m3 )
j =1 j =2 j =3
σ̂ = 1 i = 1 0.00 107.84 5.00 20.00 250.00 60852.47 0.00 61846.81 810.00 0.5044
i = 2 320.00 182.77 0.00 15.00 100.00 102073.49 1.00
i = 3 0.00 56.82 0.00 10.00 140.00 32076.55 0.00
i = 4 490.00 304.55 3.00 34.00 210.00 170248.06 1.00
σ̂ = 0.8 i = 1 100.00 147.06 5.00 20.00 250.00 82671.82 0.20 67600.55 966.60 0.4998
i = 2 320.00 182.77 0.00 15.00 100.00 102073.49 1.00
i = 3 56.60 78.26 0.00 10.00 140.00 44007.47 0.20
i = 4 490.00 304.55 3.00 34.00 210.00 170248.06 1.00
σ̂ = 0.6 i = 1 200.00 186.27 5.00 20.00 250.00 104491.18 0.40 68525.06 1000.00 0.495
i = 2 320.00 182.77 0.00 15.00 100.00 102073.49 1.00
i = 3 113.20 99.70 0.00 10.00 140.00 55938.39 0.40
i = 4 366.80 253.64 3.00 34.00 210.00 141915.94 0.75
σ̂ = 0.4 i = 1 300.00 225.49 5.00 20.00 250.00 126310.53 0.60 68085.13 1000.00 0.4894
i = 2 236.20 147.56 0.00 15.00 100.00 82476.55 0.74
i = 3 169.80 121.14 0.00 10.00 140.00 67869.30 0.60
i = 4 294.00 223.55 3.00 34.00 210.00 125174.24 0.60
σ̂ = 0.38 i = 1 310.00 229.41 5.00 20.00 250.00 128492.46 0.62 68029.88 1000.00 0.4887
i = 2 210.74 136.87 0.00 15.00 100.00 76522.63 0.66
i = 3 175.46 123.28 0.00 10.00 140.00 69062.40 0.62
i = 4 303.80 227.60 3.00 34.00 210.00 127427.93 0.62
From Table 5, when the maximum deviation σ̂ = 1, which means govern-

ment doesn’t consider the equality between coal mines fully, government can get
its optimal solution and the satisfactory is 0.5044. When government begin to
decrease the maximum deviation from 1 to 0.38 to consider the equality between
coal mines, its satisfactory level drop continuously. Because coal mines only
concern about economic benefit, the ecological benefit decreases faster than eco-
nomic benefit with σ̂ draping. σ̂ from 1 to 0.6, the ecological objective value drop
to the worst value. And σ̂ from 0.6 to 0.38, the whole economic benefit decreases
and the benefit balance between coal mines is improved. When σ̂ < 0.38, there
is no feasible solution for model (13).
Comparing these results, we can find government prefer allocate emission
right to these coal mines with better waste water recycling capacity because
which can guarantee the economic and ecological benefit simultaneously in the
biggest extend. And when government consider the whole satisfactory level, the
upper benefit would be damaged. Therefore, government should push these coal
mines faster the reformation step to achieve the economy and ecology develop-
ment together.
6 Conclusion
This paper proposed a bi-level multi-objective programming model to deal with
economic and ecological conflicts in large scale coal fields. An environmental pro-
tection based mining quotas competition mechanism was established using the
proposed model which considers not only the relationship between the govern-
ment and coal mines, but also the equilibrium between economic development
and environmental protection. To solve the complex bi-level multi-objective pro-
gramming model and an extended KKT and fuzzy goal programming approach
were combined as a solution method. The proposed method was then applied
to the Lu’an coal field which includes four major coal mines. By inputting the
data into the model and computing it using the proposed solution approach, the
effectiveness of the model was demonstrated.
References
1. Baky IA (2010) Solving multi-level multi-objective linear programming problems
through fuzzy goal programming approach. Appl Math Model 34(9):2377–2387
2. Clo S (2009) The effectiveness of the EU emissions trading scheme. Clim Policy
9(3):227–241
3. Coase RH (1960) The problem of social cost. J Law Econ 3:1–44
4. Cong RG, Wei YM (2012) Experimental comparison of impact of auction format
on carbon allowance market. Renew Sustain Energ Rev 16(6):4148–4156
5. Gersbach H, Requate T (2004) Emission taxes and optimal refunding schemes. J
Public Econ 88(3):713–725
6. Goulder LH, Hafstead MA, Dworsky M (2010) Impacts of alternative emissions
allowance allocation methods under a federal cap-and-trade program. J Environ
Econ Manag 60(3):161–181
7. Haibin L, Zhenling L (2010) Recycling utilization patterns of coal mining waste in
China. Resour Conserv Recycl 54(12):1331–1340
8. Hanson MA (1981) On sufficiency of the Kuhn-Tucker conditions. J Math Anal
Appl 80(2):545–550
9. Harrington W, Morgenstern RD (2007) Economic incentives versus command and

control: whats the best approach for solving environmental problems? In: Acid in
the Environment. Springer, pp 233–240
10. Hu CF, Teng CJ, Li SY (2007) A fuzzy goal programming approach to multi-
objective optimization problem with priorities. Eur J Oper Res 176(3):1319–1333
11. Lu J, Shi C, Zhang G (2006) On bilevel multi-follower decision making: general
framework and solutions. Inf Sci 176(11):1607–1627
12. Lu J, Shi C, Zhang G, Dillon T (2007) Model and extended Kuhn-Tucker approach
for bilevel multi-follower decision making in a referential-uncooperative situation.
J Glob Optim 38(4):597–608
13. Moitra Bhola Nath, Pal Bijay Baran (2002) A fuzzy goal programming approach
for solving bilevel programming problems. In: Advances in Soft Computing, AFSS
2002. Springer, Heidelberg, pp 91–98
14. Pal BB, Moitra BN (2003) A fuzzy goal programming procedure for solving
quadratic bilevel programming problems. Int J Intell Syst 18(5):529–540
15. Shi C, Lu J, Zhang G (2005) An extended Kuhn-Tucker approach for linear bilevel
programming. Appl Math Comput 162(1):51–63
16. Silva LF, de Vallejuelo SFO, Martinez-Arkarazo I et al (2013) Study of environmen-
tal pollution and mineralogical characterization of sediment rivers from Brazilian
coal mining acid drainage. Sci Total Environ 447:169–178
17. Sun W, Wu Q, Dong D, Jiao J (2012) Avoiding coal-water conflicts during the
development of Chinas large coal-producing regions. Mine Water Environ 31(1):74–
78
18. Wolkersdorfer C, Bowell R (2005) Contemporary reviews of mine water studies in
Europe, Part 2. Mine Water Environ 24(1):2–37
19. Younger PL, Wolkersdorfer C (2004) Mining impacts on the fresh water environ-
ment: technical and managerial guidelines for catchment scale management. Mine
Water Environ 23:s2–s80
20. Zhixin Z, Ya L (2011) The impact of carbon tax on economic growth in China.
Energ Procedia 5:1757–1761
Modeling and Solving the Vehicle Routing
Problem with Multiple Fuzzy Time Windows
Fang Yan(B) and Yuanyuan Wang
School of Economic and Management, Chongqing Jiaotong University,

yanfang@cqjtu.edu.cn
Abstract. Considering that the customers could accept service in mul-

tiple time periods and the fuzziness of service time periods, this paper
deals with the multiple time windows as fuzzy variables and quantifies
the customer satisfaction level according to the membership function of
the beginning time to be served, on the basis of the given acceptable
satisfaction level, the vehicle routing model with multiple fuzzy time
windows is constructed in order to minimize the transportation cost and
number of vehicles and maximize the satisfaction level. Then accord-
ing to the model characteristics, we use the punishment factors to deal
with the constraints and apply particle swarm operations to solve the
proposed problems. The experimental results show the effectiveness of
proposed algorithm in solving the vehicle routing problems with multi-
ple fuzzy time windows. Comparing the calculated results with the hard
time window model results, it is found that our proposed model is more
effective to reduce the cost of distribution.
Keywords: Multiple fuzzy time windows · Particle swarm operations ·

Vehicle routing · Customer satisfaction
1 Introduction
Vehicle routing problem was first proposed by Dantzing and Ramser [5]. After
more than 50 years of research, it has attracted wide attention of scholars in the
field of operations research and combinatorial optimization. The vehicle routing
problem with time window (VRPTW) is add the time window constraint on the
basis of the capacity constraint model. At present, domestic and foreign schol-
ars have done a lot of research on the VRPTW problem, Desrochers et al. [6]
solved VRPTW problem by combined branch-and-bound algorithm and column
generation algorithm. Bent and Hentenryck [3] proposed a two-stage hybrid algo-
rithm for VRPTW problem. In addition, the algorithm for the VRPTW problem
is also includes genetic algorithm [15], particle swarm optimization algorithm
[1], ant colony system [17], etc. However, considering the customer service time
requirements are not completely rigid, Lin [12] considered customer satisfaction

848 F. Yan and Y. Wang
level and proposed vehicle routing problems based on the fuzzy time window.
Ghannadpour et al. [9] presented a multi-objective dynamic vehicle routing prob-
lem with fuzzy time windows considered customer-specific time windows and pro-
posed the genetic algorithm (GA) and three basic modules to solve the problem.
While these studies have contributed significantly to solving the vehicle routing
problem with single time window, customers usually have more than one time
period to receive service, they ignored the vehicle routing problem with multiple
time windows (VRPMTW).
The vehicle routing problem with multiple time windows (VRPMTW) is
an extension problem of the traditional vehicle routing problem, refers to each
customer has a number of non overlapping time windows to receive services,
delivery vehicles must choose one time window for service, and compared with
the single time window of vehicle routing problem, VRPMTW is much closer to
the reality. The research on VRPMTW is mainly focused on two aspects: model
and algorithm. In the aspect of model, Doerner et al. [7] presented the model of
vehicle routing problem with multiple interdependent time windows and solved
the problem by several variants of a heuristic constructive procedure as well as
a branch-and-bound based algorithm. Ma et al. [14] considered delivery can be
splited and established the split delivery vehicle routing problem with multiple
time windows (SDVRPMTW) model. Yan et al. [18] applied VRPMTW model
in military oil transport path optimization and solved problem with particle
swarm optimization algorithm. And the research of algorithms mainly focus on
the ant colony system algorithm [8], simulated annealing algorithm [13], hybrid
intelligent algorithm [4], hybrid variable neighborhood-tabu search heuristic [2]
and intelligent water drop algorithm [11], etc. However, customer usually have
multiple fuzzy time windows and service time requirements are not completely
rigid in reality. This kind of problem is more complicated and more realistic than
VRPMTW, so it is necessary to carry out relevant research.
In this paper, VRPMTW in a fuzzy environment is considered and cus-
tomer satisfaction is quantified by the time to start service. We aim to minimize
transportation cost, minimize vehicles, and maximum customer satisfaction, thus
establish the model of vehicle routing problem with multiple fuzzy time windows.
The particle swarm optimization algorithm is used to solve the problem and the
experimental results are analyzed and discussed.
2 Problem Statement and Modeling
2.1 Problem Statement
The vehicle routing problem with multiple fuzzy time windows can be stated as:
A distribution center has m cars to serve n customers, customer i has Wi non
overlapping fuzzy time windows. Each customer’s needs, the maximum loading
capacity of each vehicle and the distance between any two customers is known.
Vehicles starting from the distribution center, select one time window of cus-
tomer to serve and return distribution center until completion of distribution.
Modeling and Solving the Vehicle Routing Problem 849
On the basis of satisfying customer demand, vehicle loading capacity and cus-
tomer time window, the objective function is optimized through the reasonable
path planning. The model considers the following constraints: (1) Vehicle loading
capacity constraint: the actual loading of each vehicle must not exceed the max-
imum loading capacity. (2) Multi fuzzy time windows constraint: each customer
has multiple fuzzy time windows, but vehicle can only choose a time window ser-
vice. (3) Customer satisfaction constraint: customer satisfaction is greater than
the value of the decision maker. (4) Visit uniqueness constraint: each customer
can only served by a car and can only serve once. (5) Central constraint: vehicle
start from the distribution center and after implement the customer service to
return. (6) Loop elimination constrain: the path of each vehicle can only exist
the loop between the starting point and point, there can be no other circuit.
2.2 Notations
The variables and parameters used in the model are:
L = {1, 2, · · · , n} : customer set, with 0 and n + 1 said distribution center
K = {1, 2, · · · , m} : vehicle set;
Qk : the loading capacity of vehicle k;
C : vehicle start-up costs;
qi : the demand of customer i;
Dk : the longest running distance of vehicle k;
cij : the cost of traveling from nodes i to j;
ti : the service start time of customer i ;
si : the service duration time of customer i ;
tij : the travel time from customer to customer i to j ;
Wi : the number of customer time windows;
[Eiα , aα
i , bα
i , Lα
i ] : the fuzzy time window α of customer i;
Eiα : the endurable earliness time of time window α of
customer i;
aαi : the earliest start service time which customer i expect
to be served in time window α;
bα
i : the latest start service time which customer i expect to
be served in time. window α;
xijk : a binary variable, xijk = 1, if vehicle travels from
customer i to customer j, otherwise, xijk = 0;
yiα : a binary variable, yiα = 1, if vehicle serve the time
window α of customer i, otherwise, yiα = 0.
2.3 Fuzzy Time Window

Trapezoidal fuzzy time windows is used in this paper, as shown in Fig. 1. Each
customer can own multiple trapezoidal fuzzy time windows, customer is satis-
fied when be served in the expected time period and the satisfaction level is 1,
while customer satisfaction will be reduced when the service time is not met or
exceeded the customer expected service time.
Fig. 1. Trapezoidal fuzzy time window
When the customer has multiple trapezoidal fuzzy time window can accept
the service, customer satisfaction can be defined by the membership function of
the service start time:
⎧
⎪
⎪ 0, ti < Eiα
⎪
⎪
⎨ (ti − Eiα )/(aα i − Ei ) , Ei < ti < ai
α α α
μi (ti ) = 1, ai < ti < bi (1)

⎪
⎪
⎪
⎪ (L α
− t i )/(Lα
− b α
) , bα
< t i < Lα
⎩ i i i i i
0, ti > Lα i .
2.4 Modelling
The vehicle routing problem with multiple fuzzy time windows is as follows:

n n+1 m
min Z1 = cij xijk (2)
i=0 j=0 k=1
n
min Z2 = C x0jk , (3)
j=0
1 n
max Z3 = μi (ti ), (4)
n i=1
s. t. μi (ti ) ≥ ηi , ∀i ∈ L, (5)
⎛ ⎞
n n
⎝ qi xijk ⎠ ≤ Qk , ∀k ∈ K, (6)
i=1 j=0

n n+1
dij xijk ≤ Dk , ∀k ∈ K, (7)
i=0 j=1

n+1
x0jk = 1, ∀k ∈ K, (8)
j=1

n+1
xin+1k = 1, ∀k ∈ K, (9)
i=1
n m
n
m
xijk = xjik = 1, ∀j ∈ L, (10)
i=1 k=1 i=1 k=1

xijk ≤ |S| − 1, S ⊆ L, ∀k ∈ K, (11)
i,j∈S×S
i ≤ Ei , ∀i ∈ L, α ∈ {1, 2, · · · Wi − 1} ,
α+1
Lα (12)

Wi
max yiα Eiα , (ti + si + tij ) xijk ≤ tj , ∀i, j ∈ L, ∀k ∈ K, (13)
α=1

Wi
tj ≤ j , ∀j ∈ L,
yjα Lα (14)
α=1

Wi
yiα = 1, ∀i ∈ L, (15)
α=1
xijk = 0 or 1∀i, j, k, (16)
yiα = 0 or 1, ∀i ∈ L, α ∈ {1, 2, · · · , Wi } . (17)
Objective (2) minimizes the total travel cost of all vehicles which is the most
important objective to the decision makers. Objective (3) minimums the number
of vehicles. Objective (4) maximizes the average customer satisfaction. Equa-
tion (5) ensures that each customer satisfaction is higher than ηi , ηi is given by
the decision maker based on experience. Equation (6) ensures that the loading
of each vehicle is not more than the maximum loading capacity. Equation (7)
makes sure the distance of each vehicle is not more than the longest running
distance. Equations (8) to (10) ensure that each customer is served by a car and
the vehicle flow keep balance. Equation (11) is to eliminate the sub loop. Equa-
tion (12) is to sort time windows by time. Equations (13) and (14) makes sure
that customer is served in time window. Equation (15) ensures each customer
has a time window to be served. Equations (16) and (17) is the variable range.
3 Particle Swarm Optimization

Particle swarm optimization algorithm was first proposed by Kennedy and
Eberhart [10], and it is a kind of stochastic optimization technique based on
population. Compared with other evolution algorithm based on population, their
initialization is a set of random solutions and through the iterative to search the
optimal solution. The difference is that evolutionary computation follows the
principle of survival of the fittest and PSO simulates the society. PSO express
each possible solution as a particle of group, each particle has its own position
vector and velocity vector, as well as a fitness which is determined by the objec-
tive function. All particles are flying at a certain speed in the search space, by
following the optimal value of the current search to find the global optimal value.
3.1 Structure of Solution
Construct the expression of the solution and make the solution correspond to the
particles in the PSO algorithm is the key link of the algorithm. Salman et al. [16]
described the expression of solution in detail, we draw lessons from this method.
A 3L-dimensional matrix is used to represent the solution of the proposed model
with L customers, each customer corresponds to three dimensional: vehicle num-
ber vector, vehicle ranking vector and time windows number vector. In order to
express clearly and calculate conveniently, the 3L-dimensional matrix of each
particle is divided into three L-dimensional vectors, they are Xv (vehicle num-
ber vector), Xr (vehicle ranking vector) and Xw (time windows number vector).
Corresponding velocity vectors V are Xv , Xr and Xw .
For example, there are three vehicles to complete the distribution tasks of
eight customers, the position vector of a particle is: Due to the original assump-
tion that the vehicle starting from the distribution center and finally return to
Consumers 1 2 3 4 5 6 7 8
Xv 1 1 1 2 3 3 3 3
Xr 1 3 2 1 4 1 3 2
Xw 1 2 1 1 1 1 2 1
the distribution center, the corresponding vehicle routing is:
Vehicle 1 0 → 1 → 3 → 2 → 0
Vehicle 1 0 → 4 → 0
Vehicle 1 0 → 6 → 8 → 7 → 5 → 0
In the coding, the 0 omitted in order to simplify the calculation. The third
line represents the time window number of each customer to be served, customer
1 choose the first time window service, for the customer 2 choose second time
window service and for customers 3 choose the first time window service, etc.
3.2 The Fitness Value Function
According to the objective function (2), (3), (4) and the constraint conditions
design the fit value function, as shown in the following
⎧⎡ ⎛ ⎞ ⎤ ⎫

n
n
m
m ⎨ n
n ⎬
min Z = Cm + cij xijk +β1 max ⎣ ⎝di xijk ⎠ − Qk ⎦ , 0
⎩ ⎭
i=0 j=0 k=1 k=0 i=1 j=0

n
n
+ β2 max(Eiα − ti , 0) + β3 max(ti − Lα
i , 0) (18)
i=1 i=1
n
+ β4 max [ηi − μi (ti ) , 0],
i=1
where β1 , β2 , β3 and are punishment coefficients, β1 is the penalty factor for

overloading, β2 and β3 penalty coefficient for violation of the time windows, β4 is
the penalty coefficient for violation of the customer satisfaction. Decision makers
can set the punishment coefficient base on their predilection, thus eliminating
the infeasible solution.
3.3 Particle Updating

Suppose there are n particles in the D-dimensional search space, the position
vector of the particle i is Xi = (xi1 , xi2 , · · · , xiD ), the optimal position of the
particle i during the flight is, global best position is Pibest = (pi1 , pi2 , · · · , piD ),
Vi = (vi1 , vi2 , · · · , viD ) is the flight speed of particle. Using the update method
of Li et al. [11] to update, the update formula is as follows:
vid (k + 1) = ω × vid (k) + c1 × rand () × [pid (k) − xid (k)] + c2 × rand ()
× [pgd (k) − xid (k)] , (19)
xid (k + 1) = xid (k) + vid (k + 1) 1 ≤ i ≤ n; 1 ≤ d ≤ D. (20)
w is the inertia factor which is calculated by the following formula:
ω(t) = ω(gen) + (t − gen)/(1 − gen) × [ω(1)−ω(gen)], (21)
where c1 and c2 are accelerating factor, rand () generate random number between
0 and 1.
3.4 The Procedure of PSO

Step 1. Randomly generated the initial position and velocity of the particles
within the range of the values.
Step 2. Based on formula (18) to calculate the fitness value of each particle.
Step 3. Calculate the fitness value of each particle and compared with the fitness
value of the best position it has experienced, if the fitness value of the
current location is better, set the current position as its personal best.
Step 4. Compare the personal best fitness of each particle with global best fit-
ness, if the fitness value of the current location is better, set the current
position as its global best.
Step 5. Update the particle velocities and positions according to the formu-
las (19) and (20).
Step 6. If the stopping criterion is met, stop; otherwise, return to Step 2.
4 A Case Study
In this section, we use an example to illustrate the model of vehicle routing
problem with multiple fuzzy time windows. Example is as follows: a distribution
center with six vehicles to provide services for 15 customers and each customer
has two fuzzy time windows, the maximum loading capacity of each vehicle is 50
tons and the furthest distance is 120 km, the average vehicle speed is 35 km/h
with the driving cost is 5 yuan/km and the start-up cost of each vehicle is 100
yuan, the average customer satisfaction level is set to 0.7 (Table 1).
Table 1. Customer information
Customer di si Coordinate Ei1 a1i b1i L1i Ei2 a2i b2i L2i
1 4 0.3 (1,8) 6.5 8 9 10.5 12.5 14 15.6 17.1
2 9 0.4 (9,−8) 7 8.5 9.5 11 11.5 13 15 16.5
3 6 0.4 (−16,−20) 7 8.5 9.8 11.3 12 13.5 14.6 16.1
4 2 0.2 (19,12) 7 8.5 9.5 11 11.5 13 14.8 16.3
5 9 0.3 (−21,−15) 7 8.5 10 11.5 12 13.5 14 15.5
6 3 0.5 (3,25) 7 8.5 10.4 11.4 11.5 13 14 15.5
7 5 0.3 (12,1) 6.5 8 9 10.5 11.5 13 14 15.5
8 7 0.2 (14,−19) 7 8.5 9.5 11 11.3 12.8 14.5 16
9 8 0.3 (5,16) 6.5 8 9.3 10.4 10.5 12 14 15.5
10 7 0.2 (−20,18) 7 8 10.2 10.9 11 12.2 13.4 14.9
11 5 0.3 (−25,9) 6 7.5 9 10.4 10.5 12 14 15.5
12 8 0.2 (20,−18) 7.5 9 10 11.5 12 13.5 15 16.5
13 8 0.4 (−11,−7) 7 8.5 9.5 11 11.5 13 14 15.5
14 5 0.2 (−3,−17) 6.5 8 9.4 10.5 11 12.5 14 15.5
15 10 0.5 (2,−30) 7.5 9 11 12.5 13 14.5 16 17.5
Particle swarm optimization parameters set as: population size is 200, itera-
tions gen = 3000, c1 = c2 = 2. There are eighteen times to achieve the average
level of satisfaction in the twenty experiments, the eligible results of the experi-
ment as shown in Table 2.
On the basis of satisfying the satisfaction level, the optimal experimental
results are: two cars driving 232.35 km and total distribution cost is 1361.75
yuan, average customer satisfaction was 0.85. The optimal experimental results
are shown in Table 3.
In order to compare the optimization results of the proposed model and the
VRPMTW model, using the PSO algorithm to solve the case in accordance with
the VRPMTW model ignored the customer fuzzy time windows, the experimen-
tal results are compared shown in Table 4. According to the comparison of the
experimental results can be found that on the basis of vehicle routing problem
Number of Vehicle Total travel Average Total

experiments number distance (km) customer distribution
satisfaction cost (Yuan)
1 2 240.83 0.87 1404.15
2 2 232.35 0.85 1361.75
3 2 232.35 0.85 1361.75
4 2 240.83 0.87 1404.15
5 2 240.83 0.87 1404.15
6 2 240.83 0.87 1404.15
7 2 232.35 0.85 1361.75
8 2 232.35 0.85 1361.75
9 2 232.35 0.85 1361.75
10 2 232.35 0.85 1361.75
11 2 239.67 0.8 1398.35
12 2 232.35 0.85 1361.75
13 2 232.35 0.85 1361.75
14 2 232.35 0.85 1361.75
15 2 239.67 0.8 1398.35
16 2 239.67 0.8 1398.35
17 2 240.83 0.87 1404.15
18 2 239.67 0.8 1398.35
Average result 2 236.33 0.84 1381.66
Table 3. Optimal experimental results
Vehicle Optimal routings Travel distance (km) Cost (Yuan)

1 0-4-7-2-12-8-15-14-0 113.42 667.09
2 0-1-9-6-10-11-5-3-13-0 118.93 694.66
Table 4. Comparison of experimental results
The optimal solution of VRPMTW The optimal solution of proposed model

Total Total dis- Vehicles Total Total dis- Vehicles Average
travel tribution number travel tribution number customer
distance cost distance cost satisfac-
(km) (Yuan) (km) (Yuan) tion
247.95 1539.75 3 232.35 1461.75 2 0.85
with multiple fuzzy time windows, keeping a higher average customer satisfac-
tion degree can effectively reduce vehicle travel distances and lower distribution
costs. Secondly, decision maker can set satisfaction degree parameters accord-
ing to his own different emphasis on distribution costs and satisfaction degree
and change distribution strategies. Compared to VRPMTW, this model is more
flexible and realistic.
5 Conclusions
The vehicle routing problem with multiple fuzzy time windows has been widely
used in the practical life, but the study of such problems are relatively few
recently. Based on the multiple fuzzy time windows provided by customers, this
paper adopts the membership function of start service time to quantify customer
satisfaction degree, and establish universal vehicle routing problem with multiple
fuzzy time windows as well as solve mathematical model by means of the particle
swarm algorithm to minimize the total distribution costs and maximize customer
satisfaction degree. The results of numerical example show that compared with
VRPMTW, this model can reduce distribution costs more effectively. In addition,
the flexible setting of customer satisfaction parameters and the expansion of
the time window can further reduce the distribution cost, it has more realistic
meaning and reference value for decision makers.
However, there are more complex and uncertain factors in practical appli-
cations such as vehicle travel time, the demand of customers and so on. The
exploration of such extensive problems will be the main research focus on the
next step.
Acknowledgements. This research was supported by NSFC (Grant No. 71401020,

Grant No. 71401093) and Human Social Science for Universities of Hebei (Grant No.
BJ2016057).
References
1. Ai TJ, Kachitvichyanukul V (2009) A particle swarm optimisation for vehicle rout-
ing problem with time windows. Int J Oper Res 6(19):519–537
2. Belhaiza S, Hansen P, Laporte G (2014) A hybrid variable neighborhood tabu
search heuristic for the vehicle routing problem with multiple time windows. Com-
put Oper Res 52:269–281
3. Bent R, Van Hentenryck P (2004) A two-stage hybrid local search for the vehicle
routing problem with time windows. Transp Sci 38(4):515–530
4. Bitao P, Fei W (2010) Hybrid intelligent algorithm for vehicle routing problem
with multiple time windows. Int Forum Inf Technol Appl 1:181–184
5. Dantzig GB, Ramser JH (1959) The truck dispatching problem. Manage Sci
6(1):80–91
6. Desrochers M, Desrosiers J, Solomon M (1992) A new optimization algorithm for
the vehicle routing problem with time windows. Oper Res 40(40):342–354
7. Doerner KF, Gronalt M et al (2008) Exact and heuristic algorithms for the vehicle
routing problem with multiple interdependent time windows. Comput Oper Res
35(9):3034–3048
8. Favaretto D, Moretti E, Pellegrini P (2013) Ant colony system for a VRP with
multiple time windows and multiple visits. J Interdiscip Math 10(2):263–284
9. Ghannadpour SF, Noori S et al (2014) A multi-objective dynamic vehicle rout-
ing problem with fuzzy time windows: model, solution and application. Appl Soft
Comput 14(1):504–527
10. Kennedy J, Eberhart R (1995) Particle swarm optimization. In: 1995 Proceedings
of the IEEE International Conference on Neural Networks, vol 4, pp 1942–1948
11. Li Z, Zhao F, Liu H (2014) Intelligent water drops algorithm for vehicle routing
problem with time windows. In: 11th International Conference on Service Systems
and Service Management, vol 24, pp 1–10
12. Lin JJ (2006) Multi-objective decision making for vehicle routing problem with
fuzzy due time. IEEE Int Conf Syst Man Cybern 4:2903–2908
13. Ma H, Zuo C, Yang S (2009) Modeling and solving for vehicle routing problem
with multiple time windows. J Syst Eng 24(5):607–613 (in Chines)
14. Ma HW, Ye HR, Xia W (2012) Improved ant colony algorithm for solving split
delivery vehicle routing problem with multiple time windows. Chin J Manage Sci
15. Ombuki B, Ross BJ, Hanshar F (2006) Multi-objective genetic algorithms for vehi-
cle routing problem with time windows. Appl Intell 24(1):17–30
16. Salman A, Ahmad I, Al-Madani S (2002) Particle swarm optimization for task
assignment problem. Microprocess Microsyst 26(8):363–371
17. Toklu NE, Gambardella LM, Montemanni R (2014) A multiple ant colony system
for a vehicle routing problem with time windows and uncertain travel times. J
Traffic Logistics Eng 2(1):52–58
18. Yan H, Gao L et al (2015) Petrol-oil and lubricants support model based on mul-
tiple time windows. J Comput Appl 35(7):2096–2100
Optimization of Operating of the Systems
with Recurrent Service by Delays
Asaf Hajiyev(B) and Narmina Abdullayeva
Institute of Control Systems, Azerbaijan National Academy of Sciences,

Baku, Azerbaijan
asaf.hajiyev@gmail.com
Abstract. For systems with recurrent service delays of the beginning

service are introduced. It is shown that for some systems delays reduce a
customer average waiting systems. The class of so called stochastic sys-
tems for which it is advisable to introduce delays is described. Numerical
and graphical examples demonstrating results are given. There are given
also examples, when introducing of delays can increase a customer wait-
ing time before service.
Keywords: Systems with recurrent service · Delays of beginning ser-

vice · Customer average waiting time · Stochastic systems
1 Introduction
Consider a queuing system, where: t1 , t2 , · · · , tn is random sequence when ser-
vice starts in the system, x is customer arrival instant. At each instant ti all
customers who arrived at the interval [ti−1 , ti ) immediately will get service (see,
Fig. 1). Such models are typical for applications and can be used in traffic, com-
munication systems, network of computers and others. Denote t∗1 , t∗2 , · · · , t∗n is
(random sequence) service starts in the system with delays of beginning service
(see, Fig. 1).
The problem is: Is it possible by introducing delays to reduce a customer
average waiting time? May be it is a Paradoxical Idea to introduce delay of
the beginning service for reducing a customer average waiting time? Idea to
introduce delays of beginning service belong to various authors [1–4]. Below we
will give some examples of queueing systems with delays of beginning service.
Example 1. The Main Building of M.V. Lomonosov Moscow State University.
There are several lifts in the hall (see, Fig. 2).
If at least two lifts are coming to the first floor almost at the same time, then
one of them must be delayed!. Mathematical model has the following form (see,
Fig. 3).

Optimization of Operating of the Systems with Recurrent Service 859
customer waiting time
t1 t2 ... tn
t1* t2*... t n*
Fig. 1. General mathematical model
distance
Fig. 2. Model with several elevators
Initial system
0 t1 t2 ti-1 …… ti tk
t1* t 2* ti*1 ti*

T
controlled (delayed) system
Fig. 3. Mathematical model with delays
If in the system, during the time-interval T there are k services in an initial

system, but in a controlled system there are i services and i < k, but w∗ < w?
PARADOX? This mathematical model can be used in communication systems
(Fig. 4).
860 A. Hajiyev and N. Abdullayeva
2 Communication System
How long time we can keep communications?
III
II
satelites
I II I
0 t1 t2 t3 ……
Fig. 4. Mathematical model of communication system
3 Mathematical Models
Denote η1 = t1 η2 = t2 − t1 , · · · , ηn = tn − tn−1 ; η1 , η2 , · · · , ηn are independent
and identically distributed random variables with distribution function F (x),
Eη1 = μ, V arη1 = δ2.
Stationary flow of customers arrives to service. At the each instant ti all
customers, who arrived at the interval [ti−1 , ti ) immediately will get service.
Denote w is customer average waiting time before service (Fig. 5).
From the sequence t1 , t2 , · · · , tn we pass to the new sequence for which we
have
η1∗ = t∗i − t∗i−1 = ηi + g(ηi ), (ηi = ti − ti−1 ),
where g ∈ G is a class of measurable and nonnegative functions. Denote w(g) is
customer average waiting time before service in a system with
Eη12
MF (g) = w(g) − w, c= ,
2Eη1
control function g(.).
t1 t2 tn
t1* t 2*
Fig. 5. Simple mathematical model with delays

Definition 1. Service can be improved if there exists g ∈ G such that

MF (g) < 0.
Below will be formulated some results without proof but an explanations of
these theorem will be given.
Theorem 1. Service can be improved if and only if ∃x0 < c such that F (x0 ) > 0
(Figs. 6 and 7).
F(x)
Fig. 6. Service can be improved
Fig. 7. Service can not be improved
Definition 2. Function g ∗ (x) ∈ G is called an optimal function if
min MF (g) = MF (g ∗ ).
g∈G
Theorem 2. Under the conditions of Theorem 1 an optimal function has the

form g ∗ (x) = max(0, c1 − x) = (c1 − x)+ , where c1 is a unique solution of the
equation ∞
c21 = (c1 − x)2 dF (x).
c1
Example 2. F (x) = 1 − e − x, x ≥ 0. Poisson flow of customers with bounded

intensity arrives to service. Simple calculations yield w = 1, g ∗ (x) = (0, 9 − x)+,
w∗ = 0.9; The gain in customer average waiting time is 10%.
Example 3. Let t1 , t2 , · · · , tn , · · · be times when service starts in some system.
We assume ti = i+ d xi , i = 1, 2, · · · , where d is a constant x1 , x2 , · · · , xn is a
sequence of independent and identically distributed random variables (i.i.d.r.v.)
with d.f .
Such a scheme describes, for example, the behavior of a public transport
system where id(i = 1, 2, · · · ) is the ideal timetable (schedule) for the arriving
transport and id + xi (i = 1, 2, · · · ) is the real timetable (schedule). We denote
ηI = ti − ti−1 = id + xi − (i − 1)d − xi − 1 = d + xi − xi−1 . Let w and σ 2 be
the expectation and the variance of the customer waiting time until service. If
t1 , t2 , · · · , tn , · · · is a stationary renewal process, tn = η1 + η2 +, · · · + ηn , then
we have
2
Eη 2 Eη 3 Eη 2
w= , σ2 = − ,
2Eη 3Eη 2Eη
where η is a random variable with d.f. F (x). It is clear that
2
Eη∗2 Eη∗3 Eη∗2
w∗ = , σ∗2 = − ,
2Eη∗ 3Eη∗ 2Eη∗
In our case η1 , η2 , · · · , ηn , · · · are identically distributed random variables,
but they are dependent. The flow of customers is stationary with intensity μ < ∞
and independent of t1 , t2 , · · · , tn . At the instant ti all customers, which arrived
during the time interval [ti−1 , ti ) are served immediately, in particular we assume
that service time equals zero. For this case also it is succeded to get similar
formula for w and σ, but we need to use for that wellknown Campell’s formula.
Theorem 3. Service in the model A can be improved by delays if and only if
there exists x0 < c such, that F (x0 ) > 0 where
Eη 2
c= .
2Eη
Definition 3. We call g̃(x) an optimal function if
min MF (g) = MF (g̃).

g∈G
Theorem 4. Under the conditions of Theorem 1, the optimal function has the
following form
g̃(x) = max {0, (c1 − x)} = (c1 − x)+ ,
where c1 is the unique solution of the equation
∞
c21 = (c1 − x)2 dF (x).
c1
Theorem 5. Under the conditions of Theorem 1, we have σ 2 (c1 ) ≤ 2(c2 ) ≤

σ 2 , c2 ≤ c1 .
Example 4 (numerical). Let t1 , t2 , · · · , tn , · · · be times when service starts in
system. We assume ti = id+xi , i = 1, 2, · · · , where d is a constant x1 , x2 , · · · , xn
is a sequence of independent and uniformly distributed random variables on
[−1/2, 1/2].
Denote ξi = t−ti−1 = (i+1)d+xi −id−xi −1 = d+xi −xi −1 = 1+xi −xi −1,
put d = 1, Eξi = 1, Eξi 2 = 13/12, w = Eξi 2/2Eξi = 0.54.
Simple calculations yield g ∗ (x) = max(0, 0.53 − x), w∗ = 0.53; Gain in cus-
tomer average waiting time is 1.8%.
3.1 Realization at the Practice 1

(1) Bus Stop
It is necessary to calculate c1 , for instance, in the Example 1, c1 = 0.9; d is
an interval between departure previous bus and coming to the stop next bus
If d < c1 , then bus will be delayed for (c1 − d), otherwise nothing is changed
(Fig. 8).
c1 =0.9;
d =0.8;
0 t1 t2 tn
Fig. 8. Electronic plate at the bus stop
Consider two different service systems: w1 (g) and w2 (g) are customer average
waiting time with delay function g(x). Problem is: when
w1 (0) w2 (0)
> ?
w1 (g1∗ ) w2 (g2∗ )
i.e. where (in which system) service can be improved more strong? Introduce k,
which means variation coefficient, i.e.
Eη 2 k w1 (0) w2 (0)
k= 2; k1 > √2 ⇒ ∗ > .
(Eη) 1 + k2 − 1 w1 (g1 ) w2 (g2∗ )
Let us consider Examples 1 and 2.

Consider the random variable η has d.f. F1 (x) = 1 − exp(−λx), x ≥ 0, λ = 1.
Eη = (1/λ) = 1, Eη 2 = 1/λ2 + 1/λ = 2, k1 = 2. Gain in customer average
waiting time was 10%. Random variable ξi = 1 + xi − xi − 1, where xi has

uniform distribution at the [−1/2, 1/2]. Eξi = 1, Eξi2 = 7/6, k2 = 13/12. Gain
in a customer average waiting time was 1.5%. Our formula gives: if (k1 )1/2 =
1.4 > k2 /[1 + (k2 − 1)1/2 ) = 0.83; then gain is greater (Fig. 9).
0 t1 t2 t3 …… tn
t 1* t2
t1 t 1* t2 t 2*
Fig. 9. Complicated mathematical model with delays
Model B:
η1∗ = η1 + g1 (η1 ), η2∗ = (η1 + η2 − η1∗ )+ + g2 (η1 + η2 − η1∗ ),
ηn∗ = (η1 + η2 + · · · + ηn − η1∗ − η2∗ − · · · ηn−1

∗
)+ + gn ( , · · · ).
From construction it is clear that the model has complicated form and to
apply directly formula for calculation of a customer average waiting time not
so easy. We will try to get new formula for calculation of a customer average
waiting time before service.
Lemma 1. N
1 2
N i=1 Eηi
w = lim N
;
N →∞ 1
N2 i=1 Eηi
Generalization of Khinchin’s formula
Suppose 1, 2, · · · , n is i.i.d.r.v. Eη1 =, V arη1 = σ 2 , F (x) is d.f.
Theorem 6. Service can be improve ⇔ σ 2 > 0.
Definition 4. System is called stochastic if σ 2 > 0
Remark 1. In any stochastic system service always can be improved.
Theorem 7. For stochastic system optimal function has the form
g̃(x) = (μ − x)+ = max (0, μ − x).
Example 5. F (x) = 1 − e − x, x ≥ 0.
∗
w = 1; σ 2 = 1; g ∗ (x) = (1 − x)+ , w∗ = 0.5; σ 2 = 0.
The gain in customer average waiting time is 50% and variance equals zero
(Fig. 10).
<1
0 t1 t2 t3 …… tn
=1 >1 =1
* *
t
1 t 2 t3*
Fig. 10. Realization at the practice
(2) Shuttle Systems with Finite Volume

Denote η1 = t1 , η2 = t2 − t1 , · · · , ηn = tn − tn−1 . η1 , η2 , · · · , ηn is independent
and identically distributed random variables with distribution function F (x),
Eη1 = μ, V arη1 = δ 2 .
At the each instant ti all customers, who arrived at the interval [ti−1 , ti )
immediately will get service, w is customer average waiting time before service .
Our aim: reduce w by introducing delay of beginning service.
Introduce delay, hence from the sequence t1 , t2 , · · · , tn we pass to the new
sequence for which we have t∗1 , t∗2 , · · · , t∗n ,
η1∗ = t∗i − t∗i−1 = ηi + g(ηi ), (ηi = ti − ti−1 )
Denote m volume of Shuttle.
Theorem 8. Service can be improved if the conditions: λEη1 < m x0 < c such
that F (x0 ) > 0 are hold. Denote ζ(t) number of customers in a system at the
instant t.
Theorem 9. Optimal function has the form
a(x) = inf {t : ξ(x + t) ≥ m} where c1 is a unique solution of the equation
∞
2
2
c1 = (x − c1 ) dF (x).
c1

It is necessary to calculate c1 , for instance, in the Example 1, c1 = 0.9; d is
an interval between previous bus departure and coming to the stop next bus If
d < c1 , then bus will be delayed either until the instant when number of the
passengers at the interval becomes m or for (c1 − d) (Figs. 11 and 12).
Theorem 10. Service can be improved if λEη1 < m, σ 2 > 0.
Theorem 11. Optimal function has the form
g̃(x) = max {0, min(a(x), μ − x)} .
Denote ξ(t) number of customers in the system at the Instant
a(x) = inf {t : ξ(x + t) ≥ m} .

c1 =0.9;
d =0.8;
0 t1 t2 tn
0 t1 t2 t3 …… tn
t1* t2
t1 t1* t2 t2*
Fig. 12. Model B’
Theorem 12. Optimal function has the form g̃(x) = max{0, min(a(x), c1 − x)},
∞
2
where c1 is a unique solution of the equation c21 = (x − c1 ) dF (x).
c1

Eη = μ is given or can be calculated from statistics. d is an interval between
previous bus departure and coming to the stop next bus. If d < μ, then bus will
be delayed either until the instant when number of the passenger at the interval
becomes m or for (μ − d) (Fig. 13).
µ =1;
d =0.8;
0 t1 t2 tn
bus arrival to stop
Remark 2. The proof of some results can be found in [3,4].

References
1. Carter GM, Chaiken JM, Ignall E (1972) Response areas for two emergency units.
Oper Res 20(3):571–594
2. Hajiyev AH, Jafarova H (2010) Mathematical models of complex queues and their
applications. Proc Inst Math Mech ANAS 33:57–66
3. Hajiyev AH, Djafarova HA, Mamedov TS (2010) Mathematical models of moving
particles without overtaking. Dokl Math 81(3):395–398
4. Newell GF (1974) Control of pairing of vehicles on a public transportation route,
two vehicles, one control point. Transp Sci 8(3):248–264
An Epidemic Spreading Model Based
on Dynamical Network
Yi Zhang(B)
Mathematics School, Yibin University,

Yibin 644007, People’s Republic of China
245806513@qq.com
Abstract. This paper proposes a epidemic spreading model consider-

ing dynamic change of networks. A function p(t) reflecting the dynamic
change of network is established, which is a probability that the infected
from the susceptible at time t. This function replaces the constant rate
in the traditional model and a variant of the SIR model was built. This
model can reflect the dynamic change of network. At the same time, a
dynamic analysis of the model was conducted. Then, epidemic spreading
model simulations are conducted with different parameters in ER ran-
dom networks. The results show that epidemic spreads faster and more
broadly in network when the smaller the parameter c, the larger parame-
ter q. That is, epidemic often spread faster and more broadly when people
take protection measures slowly and moving speed of people is slow.
Keywords: Epidemic spreading · Dynamical change · Moving speed
1 Introduction
Human beings are still at the risk of the outbreak of epidemic diseases
now even after the development of modern medicine and the appearance of
antibiotics [1,2]. The epidemic disease cause big disaster to human in history.
For example, the SARS outbreak in 2003 and the H1N1 outbreak in 2009. These
epidemic disease swept the globe in a short time, and caused a great loss, even
more change the life way of human. With the worsening environment, epidemic
disease is break out more frequently, and affect the humanity’s survival and devel-
opment. Therefore, the epidemic rule, transmission mechanism and the strategy
of preventing disease are major problem which need to be solved. These problems
have attracted many researchers attention increasingly.
The mathematical modeling of epidemic disease spreading has been exten-
sively studied for a long time [13]. The main mathematical approach is the so-
called compartmental models which are composed of ordinary differential equa-
tions [6–8,11]. In this kind of approach, the entire population are divided into
different compartments and each compartment corresponds to an epidemiolog-
ical state which depends on the characteristics of the particular disease being
An Epidemic Spreading Model Based on Dynamical Network 869
modeled. The commonly used compartment; the susceptible (S), is not infected
but likely infected individuals, the exposed (E) is already infection but don’t
have infectious individuals, the infected (I) consists of individuals who have
been infected and had infectious, the removed (R) consisting of individuals who
have been restored or death. A classical compartmental model is the SIR model,
which was constructed by Kermack and Mckendric in 1927, and they also pro-
posed SIS model in 1932. Based on these research, a lot of compartmental model
were proposed. According to the different propagation process of disease, com-
mon epidemic model are: SI, SIS, SIR, SIRS, SEIR, SEIRS, SEI and SEIS.
However, most of research of epidemic on complex networks assume that the
network structure remains the same, it does not fit the reality. Some papers have
realized the dynamic nurture in epidemic spreading. A model was built to inves-
tigate the disease spreading dynamics and synchronization behavior on complex
networks [12]. A SIRS epidemic model with feedback mechanism on adaptive
scale-free networks is presented [9]. The real social network is the evolution of
dynamic interaction between nodes and links [5]. In fact, people cut back on
frequency of going out once the outbreak of epidemic, and the government take
some measures to reduce mobility of people. Because of action of self-protection
and government protection, the topology of the network is dynamic change. That
is, the mobility of people change topology of networks, thus change the topo-
logical feature of network and the way of epidemic spreading. Therefore, it is
significant that dynamic network affected by action of protection is considered
in epidemic model. And it is goes some way to providing an in-depth under-
standing of epidemic spreading and how to control such epidemic spreading on
dynamic social networks.
In this paper, We establish a function to quantify the dynamic change of
the network, and propose a epidemic spreading model which consider both the
change of network and the speed of the change. In this model, the function can
dynamically measures the probability that the susceptible becomes the infected.
A variant of the SIR model is then built, in which the traditional fixed parameters
replaced with our probability function, and conduct the stability analysis of
model. Then investigate how different values for parameters can effect epidemic
spreading using MATLAB in ER random networks. Finally, we conclude this
paper and provide several avenues for further research.
2 Epidemic Spreading Model
In this section, we construct a function of time t which considers the dynamic

change of networks. Then, with this function added as a parameter to a differ-
ential dynamic spread model, a new epidemic propagation model is proposed.
2.1 Function
Consider a population of N individuals and E links respectively representing

individuals and their interactions, each of individuals is in one of three states
870 Y. Zhang
S (susceptible), I (infected), or R (removed). At each time step t, each indi-

vidual comes into contact with one other individual, drawn randomly from the
population.
In epidemic spreading process, the behaviors of protection of the individuals
are ignored. In reality, people will reduce the probability of going out because
they are afraid of being infected when the epidemic is outbreak, and they will take
some protection measures, for example, doing more exercise. With the moving
of the people, the topology of network is changed, and the object they contact
is also changed. Generally speaking, as time passes, because of the behavior of
protection, the probability that the susceptible contact the infected is decrease
gradually, and go to zero finally. Moreover, we believe that the speed at taking
the measure of protection should be considered. If people do protection behavior
quickly, the epidemic effect is small, but in a contrary situation, the epidemic
effect is far greater.
This function is a probability p(t) that an susceptible become a infected at
time t, which is affected by the topology of networks. Therefore, p(t) reflects the
dynamical behaviors of networks at time t. In light of the above analysis, the
probability at time t can be denoted as
p(t) = e−ct , 0 < c ≤ 1, t ≥ 1,
where c is the parameter to reflect the speed of taking the measure of protection.
A larger c indicates that the individual take the measure of protection in a shorter
time. Figure 1 shows the probability as a function of t, given a different c. When c
tends to zero, the speed of taking the measure of protection slows down. We have
carefully checked that p(t) decreases sharply when c ∈ (0.1, 1), it is suggested
that parameter c < 0.1.
In this case, the parameter for c was (0.1, 1). p(t) is the probability, and in
(0, 1), p(t) reflects the probability that an individual become a infected from an
susceptible. Therefore, p(t) tends to zero as times passes.
2.2 Model
Denote S(t), I(t), R(t) as the density of the susceptible, infected, removed at
time t, S(t)+I(t)+R(t)=1. As shown in Fig. 2, the epidemic spreading rules can
be summarized as follows:
(1) When a susceptible individual contacts an infected individual, the susceptible

individual becomes a infected individual at the rate of p(t), which is affected
by the topology of networks. Therefore, the reduced speed of the susceptible
dS(t)/dt is proportional to the number of the susceptible S(t) and infected
I(t), so we get the differential equation:
dS(t)
= −p(t)S(t)I(t). (1)
dt
Fig. 1. Probability as a function of t
Fig. 2. Structure of epidemic spreading process
(2) At each time step, some infected individuals become removed individuals
because of death or immunity. We suppose that the above case happens
at probability q. Therefore, the reduced speed of the infected individuals
dI(t)/dt is proportional to the number of the infected individuals I(t), Addi-
tionally, an susceptible individual becomes a infected individual at the rate
of p(t), so we have
dI(t)
= p(t)S(t)I(t) − qI(t). (2)
dt
(3) The increasing speed of the removed individuals dR(t)/dt is proportional to
the number of existing I(t) from (2), so we get
dR(t)
= qI(t). (3)
dt
Based on the previous discussion, by integrating Eqs. (1)–(3), the follow-
ing global expected model can now be formulated for the epidemic spreading
process:
⎧
⎪
⎪
dS(t)
dt = −p(t)S(t)I(t)
⎪
⎨ dI(t)
dt = p(t)S(t)I(t) − qI(t) (4)
⎪
⎪
dR(t)
⎪ dt = qI(t)
⎩ S(0) = S0 , I(0) = 1 − S0 , R(0) = 0.
872 Y. Zhang
2.3 Steady-State Analysis of Model

In the epidemic spreading process, initially the system has only the susceptible
and the infected. Finally, the system has only the susceptible and the removed
as it reaches an equilibrium state. So here, we analyze the system’s steady-state
in this period.
Because S(t) + I(t) + R(t) = 1, and the differential Eqs. (1) and (2) are not
related to the R(t), we only consider the differential Eqs. (1) and (2). Then the
reduced limiting dynamic system is given by
⎧ dS(t)
⎨ dt = −p(t)S(t)I(t)
dI(t)
⎩ dt = p(t)S(t)I(t) − qI(t)
(5)
S(0) = S0 , I(0) = 1 − S0 > 0.
From differential equations theory, system (4) and (5) are homogeneous,
which means that analyzing the properties of system (5), is equal to analyz-
ing the properties of system (4). First, a required theorem is introduced.
Our model is a non-autonomous differential dynamic system, the general
form of which is
dx
dt = f(t, x) (6)
f(t, 0) = 0, x ∈ Rn .
Suppose that I = [t0 , +∞], U = {x | x ≤ h}, and V (t, x) is a continu-

ous differentiable function defined in I × U.W (x) is a continuous differentiable
function defined in U .
Theorem 1. [10] Suppose that V (t, x) is a positive definite function in I ×Rn ,

which has an infinitesimally small upper bound and an infinite lower bound, and
V̇ (t, x) is a negative semi-definite, so the zero solutions for system (6) are global
uniformly asymptotically stable.
According to the differential equations theory, we can get two theorem as

follows.
Theorem 2. P ∗ = (S ∗ , 0)(0≤ S ∗ < 1)is the equilibrium of the system (5).
Proof. Let the right side of each of the differential Eq. (8) be equal to zero in
the system which gives the equation
−p(t)S(t)I(t) = 0, (7)
p(t)I(t) − qI(t) = 0. (8)
The feasible region for the equations is R2 , and so we study the equations in
a closed set A = {(S, E, I) ∈ R2 | S + I ≤ 1, S, I ≥ 0}
From Eq. (8), we can get I(t) = 0(p(t) = q). Substituting I(t) = 0 into
Eq. (7), we can get the system has the equilibrium P ∗ = (S, I) = (S ∗ , 0)
(0 ≤ S ∗ < 1).
Therefore, we can obtain the equilibrium for the system (4) Q∗ = (S ∗ , 0, R∗ ),

where S ∗ + R∗ = 1; that is, the epidemic must disappear with time, and all S ∗
and R∗ that satisfy S ∗ + R∗ = 1, represent stable situations. Next, we look at
the stability of the equilibrium P ∗ = (S ∗ , 0, 0).
Theorem 3. The equilibrium P ∗ is globally uniformly asymptotically stable.
Proof. Let the Liapunov function be V (S, I) = F (S) + G(I), and taking the
derivative of V versus t along the solution for the equations, we have
V (S, I) = − F (S)p(t)S(t)I(t) + G (I)[p(t)S(t)I(t) − qI(t)].
In order to ascertain V (S, I) and V (S, I), and taking the value F (S) = 1,

G (I) = 1, we get V (S, I) = S + I, and
V (S, I) = − p(t)S(t)I(t) + p(t)S(t)I(t) − qI(t) = −qI(t) ≤ 0.
So we get V (S, I), which is positive definite, 0 ≤ V (S, I) ≤ 1 + S, that

is, it has an infinitesimally small upper bound and an infinite lower bound, and
V (S, I) is a negative semi-definite in the feasible region A = {(S, I) ∈ R2 |S+I ≤
1, S, I ≥ 0}.
From Theorem 1, the equilibrium P ∗ is globally uniformly asymptotically
stable.

The probability that a susceptible individual becomes a infected individual is
given a constant over the whole spreading process in traditional epidemic prop-
agation models, but dynamical network needs to be considered in the model as
Fig. 3. Density of the infected with different c on a random network

874 Y. Zhang
can be seen in the above analysis. Therefore, the probability is a variable which
changes over time in our model. In the following sections, we examine the results
with a different values for parameter c to demonstrate how the act of isolated
affects epidemic propagation. Simulations were carried out using MATLAB.
To compare the results, the proposed model with different value for c was
tested on a artificial networks–ER random networks with the network size at
N = 10000 and the average degree <k> = 16 [4]. Two different nodes are
connected with a probability of p = 0.0016, so we can get random networks with
−1)
N nodes, and pN (N 2 edges. ER random networks degree-distributions are an
approximate Poisson distribution <k> = pN ≈ p(N − 1). This distribution
reaches its peak value at the average degree <k> [3].
Given that the other parameters are fixed, we compared the epidemic spread-
ing processes on the random networks with different values for c. So here, we
set c = 0.005, c = 0.03, c = 0.7 when q = 0.05 respectively. Figure 3 illustrates
how the density of the infected changes over time for different values of c in a
random network. From a macroscopic perspective, we found that as parameter
c decreased, the number of the infected peaked increased, because a smaller c
indicates a faster movement speed. The green line indicates a scenario in which
the individuals move quickly, that is, the topology of network change quickly.
The blue line represents a scenario in which the individuals move speed is slow
relatively, and the red line represents a scenario in which the individuals hardly
move. It can be seen that the higher parameter c is, the smaller the infected
peak value. Figure 4 describes how the density of the removed changes with
changes in the parameter c. We found that as parameter c decreased, the num-
ber of the end the size of the removed increased, because a smaller c indicates a
faster movement speed. The final value for the removed density R(t) is greater,
which indicates the number of people affected by an epidemic when c is smaller.
Fig. 4. Density of the removed with different c on a random network

Fig. 5. Density of the infected with different q on a random network
Fig. 6. Density of the removed with different q on a random network
Clearly, the smaller the value of c when other parameters are fixed, the broader
the epidemic’s influence. A smaller c indicates that if the individuals have a faster
movement speed, less susceptible individuals change into removed individuals,
and therefore the influence of the epidemic decreases. Generally speaking, it can
be seen that epidemic spread more broadly and last longer when c is smaller.
However, there is a more significant impact from protection measures in random
networks.
Given that the other parameters are fixed, we compared the epidemic spread-
ing processes with different values for q. we set q = 0.05, q = 0.1, and q = 0.2.
Figures 5 and 6 illustrates how the density of the infected and removed changes
over time for the different values of q. From the Fig. 5, we can see that the bigger
876 Y. Zhang
the value of q, when other parameters are fixed, the smaller peak value of the
infected. It can be seen that the higher parameter q is, the larger the infected
peak value, and the slower the epidemic terminates. From the Fig. 6, it is easy to
find that the final value for the removed density R(t) is greater, which indicates
the number of people is infected when q is smaller. Clearly, q is the transforma-
tion probability that infected individuals become removed individuals, in reality,
along with the increase of q, there is more infected individuals become removed
individuals. As a result, the number of removed increases, the influence of the
epidemic decreases.
4 Conclusions
In this paper, we proposed a variant of the SIR epidemic spreading model by
considering the dynamic change of networks because of the behavior of protec-
tion of individuals. In our model, we established a function p(t) to describe the
dynamic change of networks. p(t) also was a probability that the infected from
the susceptible at time t. The parameter c of p(t) reflecting the change speed of
taking measure of protection. If the individuals take protection measures quickly,
c is greater, with the epidemic effect is small. On the contrary, if c is smaller,
the epidemic effect is greater. Generally speaking, as time passes, because of the
behavior of protection, the probability that the susceptible contact the infected
is decrease gradually, and go to zero finally. This function as a parameter was
added to dynamic differential equations, and a variant of the SIR model was
built. At the same time, a dynamic analysis of the model was conducted.
We then do the simulation results from our epidemic spreading model using a
different values for c, q in ER random networks. The simulations showed that epi-
demic spread faster and more broadly when c is smaller. That is, epidemic often
spread faster and more broadly when people take protection measures slowly. At
the same time, the simulation results suggested that the spread of epidemics is
influenced significantly by parameter c in a random network. q is the transforma-
tion probability that infected individuals become removed individuals, in reality,
along with the increase of q, the infected peak value is large, and the epidemic
terminates slowly. These results indicated that epidemic spreads faster and more
broadly in network when the smaller the parameter c, the larger parameter q.
The dynamic networks plays an important role in epidemic spreading. With
this in mind, in the future we plan to study feature of dynamical networks.
In addition, further studies will be conducted on epidemic control strategies
according to different network topologies.
Acknowledgements. This research was supported by project of Education Depart-

ment in Sichan province (Grant No. 15ZB0295). We appreciated all support in finance
and in spirit.
References
1. Albert R, Barabási AL (2002) Statistical mechanics of complex networks. Rev Mod
Phys 74(1):47–97
2. Boccaletti S, Latora V, Moreno Y et al (2006) Complex networks: structure and

dynamics. Phys Rep 424(4–5):175–308
3. Bollobas B (2001) Random Graphs. Academic Press, New York
4. Erdös P, Rényi A (1959) On random graphs I. Publicationes Math 6:290–297
5. Gao L, Liu JM, Zhang SW et al (2010) Dynamic changes of social network structure
caused by memory characteristics. Comput Eng 36(10):11–13
6. Hethcote HW (2000) The mathematics of infectious diseases. Siam Rev 42(4):599–
653
7. Korobeinikov A, Wake GC (2002) Lyapunov functions and global stability for sir,
sirs, and sis epidemiological models. Appl Math Lett 15(8):955–960
8. Li CH, Tsai CC, Yang SY (2014) Analysis of epidemic spreading of an sirs model in
complex heterogeneous networks. Commun Nonlinear Sci Numer Simul 19(4):1042–
1054
9. Li T, Liu X, Wu J, Wan C et al (2016) An epidemic spreading model on adaptive
scale-free networks with feedback mechanism. Phys A Stat Mech Appl 450:649–656
10. Ma ZE, Zhou YC (2001) The Qualitative and Stability Theory of Ordinary Differ-
ential Equations. Science press in Chinese
11. Moneim IA (2007) Seasonally varying epidemics with and without latent period:
a comparative simulation study. Math Med Biol 24(1):1–15 A Journal of the IMA
12. Xu D, Xu X, Yang C et al (2016) Spreading dynamics and synchronization behavior
of periodic diseases on complex networks. Phys A Stat Mech Appl 466:544–551
13. Zhu G, Fu X, Chen G (2012) Spreading dynamics and global stability of a gen-
eralized epidemic model on complex heterogeneous networks. Appl Math Model
36(12):5808–5817
Analysis of Enterprise Microblog Marketing
in Different Industries Based on DEA Model
Dan Zhang(B) , Yufeng Ma, Aixin Wang, Yue He, and Changzheng He

253510764@qq.com
Abstract. In this paper, we propose a evaluation system for the enter-

prise microblog marketing effectiveness, combine DEA method to estab-
lish the evaluation model, and applying the quantitative evaluation
method of input and output to evaluate the effect of microblog market-
ing among different enterprises. In the empirical part, in order to verify
the feasibility of the model, we compare the effectiveness of the relative
marketing effect of 50 official microblog in 5 different industries based
on Sina microblog. The result shows that microblog marketing effect
among different enterprises is different. In the five major industries, the
integrated marketing effect of mobile phones industry is best, followed
by Home Furnishing, skin care, snacks, the worst is the women’s cloth-
ing industry. In the end, by combining the development characteristics
and situation of each industry, the paper puts forward some measures to
improve the marketing effect of enterprise microblog.
Keywords: Microblog marketing · Enterprise microblog · Marketing

effectiveness · DEA
1 Introduction
CNNIC [10] has pointed out that as of June 2016, Chinese netizens has reached
710 million, the size of microblog users to 242 million. The huge user resources of
microblog provide a rich fertile soil of marketing activities for enterprises. Many
companies have realized the importance of microblog. Therefore, they launch
the online marketing on microblog. However, how to assess the effectiveness of
microblog marketing becomes the practical problems faced by enterprises.
With the Matthew effect of microblog is gradually prominent, the number of
the domestic and foreign scholars who study on the effect of microblog marketing
are on the rise. Scholars on microblog marketing research mainly focused on three
aspects: marketing effect, the impact of microblog marketing and marketing
strategy. In the study of the marketing effect of microblog, Leung [13] attempted
to explore the marketing effectiveness of two different social media sites (Face-
book and Twitter) in the hotel industry, Integrated the attitude-toward-the-ad
Analysis of Enterprise Microblog Marketing 879
(Aad) model with the concepts of attitude-toward-social-media-page, and finally

proposed a theoretical model of hotel social media marketing effectiveness. For
different groups of people, microblog marketing may have different influence,
Barry [2] assessed whether alcohol companies restrict youth/adolescent access,
interaction, and exposured to their marketing on Twitter and Instagram. Clark
[9] found that due to the youth presence on Twitter and the clinical uncertainty
of the long term health complications of electronic cigarette consumption, the
protection of public health warrants scrutiny and potential regulation of social
media marketing. Chen [8] explored young consumers’ interpretation of Twitter
and marketing information on this particular social media platform from the
perspective of consumers. Good marketing strategy is very important for enter-
prises, Liu [14] developed a guideline for advertisers to make the best marketing
practices by visualizing and studying the evolution of popular Twitter hash-
tags. In order to understand how social media affect the hospitality and tourism
field has increased, Park [15] discussed and demonstrated social media analytics
using Twitter data referring to cruise travel, finally provided feasible marketing
strategies. Kafeza [12] introduced a novel methodology to achieve information
diffusion within a social graph that activates a realistic number of users, the
methodology is useful to marketers who are interested to use social influence
and run effective marketing campaigns.
To sum up, the innovation of this paper is to put forward the microblog mar-
keting activity is input and output activities, through the introduction of DEA
model, we establishes the comprehensive evaluation index system using quanti-
tative method to calculate the input-output efficiency of enterprises microblog
marketing, and then evaluate the effect of enterprise microblog marketing.
2 Related Work
2.1 Enterprise Microblog Marketing Effect Evaluation System
There are relatively abundant indicators to assess the effect of microblog mar-
keting. Through quantities of case studies and quantitative research methods,
some scholars found that there are many factors which have influence on the
enterprise microblog marketing effectiveness. Xue [18] found that sales promo-
tion is easy to promote consumer to make information share by forwarding to
a friend or to make comments. In addition, Chang [5] thought liking or sharing
social media messages can increase the effects of popular cohesion and message
diffusion. What’s more, Zhu [20] argued that social media marketing efforts need
to be congruent and aligned with the different needs of social media users.
Bi [3] thought that the comments and reposts can reflect the microblog mar-
keting effects. Saulles [16] considered not only the simply counting Twitter fol-
lowers and volumes of tweets as indicators of effectiveness but also utilises social
authority scoring from the digital marketing analysts. Based on the above find-
ings, this paper selects the number of fans, the number of microblog, the fre-
quency of releasing microblog, and prizes value as the input evaluation indexes.
The selected output evaluation indexes are the number of fans forwarding, the
880 D. Zhang et al.
number of fans comments, the number of fans point of praise, and non negative
sentiment index.
From the input index perspective, this paper puts forward active comments
and passive comments. Active comments means that the numbers of blogger’s
active replies to followers. Here the follower’s comments do not express the ten-
dency of doubt or the desire to get any responds from the blogger. However, the
passive comments refer to the number of the replies. The replies are the blogger
who make to answer the doubts or the questions which fans put forward to. In
fact, there are obvious marks in these comments. For example, these comments
consist of the symbol@, or the interrogative and obvious punctuation, microblog
expressions.
From the output index perspective, the original microblog forwarded and
the original microblog comments were proposed. The reason is that the official
microblog generally contain original and non-original microblog(namely forward-
ing other’s microblog). The total numbers of the original microblog forwarded
also contain the original microblog forwarded and the non-original forwarded.
Therefore, in order to more clearly measure the spread of the original microblog,
the article puts forward the two indicators.
Based on the above analysis, the evaluation index system of enterprise
microblog marketing effectiveness is established, as is shown in Fig. 1.
Fig. 1. Enterprise microblog marketing effect evaluation index system
2.2 Analytic Approach

Data Envelopment Analysis which can be used for evaluating the relative effi-
ciency of decision-making units in organizations was first proposed by Charnes
et al. [6]. While DEA prototype can be traced back to 1957, in the analysis of
British agricultural productivity, Farrell [4] put forward the thought of envelope.
Later, the related DEA methods are mostly building on the idea of Farrell to
develop a mathematical programming model. Other than comparing efficiency
across DMUs within an organization, DEA has also been used to compare effi-
ciency across firms.
DEA model has many forms. This paper selects the BCC model which can
evaluate the effectiveness of scale and technical. Sun [17] proposed that we can
suppose there are n decision-making units, each unit DMUj (j = 1, 2, · · · , n) has
m inputs and s outputs. Respectively, they are showed by input X and output Y .
Xj = (x1j , x2j , · · · , xnj ), Yj = (y1j , y2j , · · · , ynj )T , j = 1, 2, · · · , n.
To evaluate DM Uj , Cooper [11] thought that we can use the linear program-
ming model (P) and dual programming model (D), as shown below:
⎧
⎪
⎪ max μT y0 = Vp
⎨
ω xj − μTi ≥ 0(j = 1, 2, ..., n)
T
(p) = (1)
⎪
⎪ ω T x0 = 1
⎩
ω ≥ 0, μ ≥ 0,
⎧
⎪
⎪ min VD = θ
⎪
⎪ n
⎪
⎪ xj λj + s− = θx0
⎪
⎪
⎨ j=1
(D) = n
(2)
⎪
⎪ yj λj − s+ = y0
⎪
⎪ j=1
⎪
⎪
⎪
⎪ λj ≥ 0, (1 ≤ j ≤ n)
⎩
s+ ≥ 0, s− ≥ 0.
In the above formula, s− and s+ are slack variable. λj is the weights coefficient
of input and output indicators. θ represents the ratio of input reduction. If
θ∗ = 1, s−∗ = s+∗ = 0, the unit is considered to be valid. If θ∗ = 1, the s−∗ ,
s+∗ are not all 0, the unit is considered weak effective. If θ∗ < 1, the unit is
considered to be invalid. When using the model to measure technical efficiency,
technical efficiency can be decomposed into pure technical efficiency and scale
efficiency.
3 Data Collection
3.1 Sample Selection
The research time was setting on March 5, 2014-March 11, 2014. When choosing
the brands, this paper discretionarily chooses five different industries, which is
Women’s Clothing, Mobile Phones, Skin Care, Home Furnishings, and Snacks.
Then the microblog of enterprises who have the sina official certification will
be determined to as the study sample. The final choice of enterprises list is as
shown in Table 1.
882 D. Zhang et al.
Table 1. The chosen enterprises in this research
Category Women’s clothing Mobile phones Skin care Home furnishings Snacks
1 Ochirly Apple EsteeLauder Lin shi mu ye Three squirrel
2 HSTYLE Samsung Pechoin Bell Land Be & Cheery
3 Vero Moda MIUI Lancome Gudxon Haoxiangni
4 ELF SACK Huawei Laneige Coleshome FERRERO ROCHER
5 Artka VIVO L’Oreal QuanU ZHOUHEI YA
6 INMAN Coolpad Kiehl’s Flisa Lou lan mi yu
7 GIRDEAR Nokia MEIFUBAO Hegou BESTORE
8 ONLY HTC Marykay KUKa Xinnongge
9 LIEBO Sony AFu LUHU Xiyumeinong
10 PeaceBird Lenovo CHCEDO YIMILOVE Houstage
3.2 The Experimental Data Preprocessing
This paper collects 1363 microblog and 128483 comments on the 50 sina official
microblog. In the all comments, there are 123782 fans comments, 2584 active
comments and 2117 passive comments. The definition of main input and output
indicators are as follows:
(1) The input indicators
Followers and microblog are all gotten from the pameng crawler.
The frequency of releasing microblog (FRW). This paper collects microblog
of enterprise during 7 days. The frequency of post microblog = the numbers of
microblog/7.
The number of active comments (NAC). This paper invites two researchers
to judge the emotion of the comments. When judge the comments, the researcher
according to the emotion and tone in the sentence pretend to be a follower to
determine whether the comment has the tendency to question.
The number of passive comments (NPC). Use the similar collecting method
with the active comments.
Prizes value (PV). Through sina advanced search, screening the microblog
which contain prizes information. Then, recording the number of prizes and
price. At last, this paper calculates the price of the total prize according to the
market price.
(2) The output indicators
The number of microblog forward (NWF), the number of fans praise (NWL), the
original microblog forwarded (OWF), the original microblog comment (OWC).
They are all gotten from the pameng crawler.
The number of follower comments (NWC). Follower comments = microblog
comments − active comments − passive comments.
Non negative sentiment rates (NNS). Non negative sentiment rates = positive
emotion + neutral emotion)/all kinds of emotion. This represents the ratio of
non-negative sentiment occupying the total emotional. This rate can also reflect
the ratio of negative emotion from the side. The enterprise collects negative
emotions to effectively understand the user, and then help enterprises to improve
the service.
In order to improve the accuracy of the emotional data, this paper use arti-
ficial method to judge the emotion in the comments. When collecting the com-
ments, this paper regards the spam comments as the neutral emotion to avoid
its interference.
After the comments collection is to classify the emotion of the comments.
This paper use the content analysis method. Then, these two researchers will
classify the comments which belong to one enterprise that is chosen randomly
among 50 enterprises. These comments will be used to practice the researcher
repeatedly to ensure the accuracy. Then, the Kappa Statistic in reliability studies
will be used to test the researcher. As a result, the researcher’s kappa value is
0.87. That shows a good consistency. Therefore, these two researchers can be
used in the phase of emotional classification. Finally, we can get the number
of positive comments, negative comments and neutral comments. Average these
three comments are the brand’s corresponding comments. We can use these data
in our study.
Then, each brand’s microblog indexes in 7 days (except for the frequency to
send microblog, the prize value, non-negative sentiment index) are also on aver-
age. Due to the numerous tables, this paper only selects the ladies enterprises’
data, as shown in Tables 2 and 3.
From the table we can see, there are many data whose value is 0 in the
column of active comments, passive comments and prize value. That is because
that different enterprise has different marketing strategy. Thus, some input data
value of enterprises is 0. Besides, this paper uses the infinitesimal (0.0001) to
replace the 0 due to the forbiddance that the input DEA data cannot be 0.
3.3 Experimental Result

This paper uses the transverse, longitudinal comparison to do the comprehensive
analysis in the part of results analysis.
Table 2. The input indicators of Women’s Clothing enterprises
Followers (Ten Microblog FRW NAC NPC PV (Hundred)

thousand) (Thousand)
Ochirly 11.96 2.05 0.71 1.6 0.2 0
HSTYLE 14.67 0.26 9.29 0.42 0.11 3.65
1Vero Moda 15.61 2.52 1.71 0.08 0
ELF SACK 5.55 9.82 8.86 2.89 0.27 2.5
Artka 3.7 2.11 7.14 0.34 0.04 14.46
INMAN 22.26 6.33 8.71 1.92 0.08 5
GIRDEAR 1.05 3.91 0.29 0 0 0
ONLY 47.26 4.56 5 0.11 0 0
LIEBO 27.52 6.72 2.71 0.37 0.21 9
PeaceBird 6.86 4.01 2 0 0 0
884 D. Zhang et al.
Transverse analysis is a method of comparative analysis of different enter-

prises’ microblog marketing effect between input and output efficiency within
the same industry. Longitudinal analysis refers to the comparative analysis of
different industries. First, arranging all the 50 brands DEA results data in a
table. Then regarding the comprehensive technical efficiency as the key word,
according to the descending order, to compare the input and output efficiency
of different industries.
(1) The Transverse Comparison
This paper is based on BCC model of DEA which assumes that scale reward is
variable. DEAP2.1 software is applied to this experiment. Through calculating,
we can obtain the evaluation results. Because of the 50 brands, the number of
relate tables are also very large. In this thesis, the representative list is only
shown in Table 4.
Table 3. The output indicators of Women’s Clothing enterprises
NWL NWF NWC OWF OWC NNS

Ochirly 9.2 6.4 11.2 7.75 13 0.93
HSTYLE 3.22 10.38 4.14 14.5 5.64 0.99
Vero Moda 4.42 3.58 4.33 3.58 4.33 0.63
ELF SACK 3.73 3.95 13.9 4.68 15.3 0.99
Artka 3.5 3.64 3.52 3.36 3.06 0.98
INMAN 7.36 125.56 35.69 84.4 44.09 0.92
GIRDEAR 0.5 0.5 2.5 0.5 2.5 1
ONLY 6.97 10.2 3.63 10.47 3.74 0.91
LIEBO 4 17.16 15.68 43.57 39.86 0.99
PeaceBird 0.29 0.29 0.43 0.29 0.43 0.5
Table 4. The marketing efficiency of Women’s Clothing
Enterprises Crste Vrste Scale

Ochirly 1 1 1 -
HSTYLE 0.895 0.898 0.997 irs
Vero Moda 1 1 1 -
ELF SACK 0.522 0.522 1 -
Artka 0.543 0.545 0.997 irs
INMAN 0.45 0.453 0.994 irs
GIRDEAR 1 1 1 -
ONLY 1 1 1 -
LIEBO 0.31 0.31 0.998 irs
PeaceBird 0.5 1 0.5 irs
If the value of the comprehensive efficiency of an enterprise is 1, that shows

that the input-output efficiency of the enterprise marketing activities is ideal,
namely the enterprise marketing effect is good. The irs in the table represents
the enterprises marketing activity is in the state of increasing, drs represents the
decreasing state.
The comprehensive technical efficiency is the measurement and evaluation of
the ability of resource allocation and resource utilization of the decision-making
unit. Pure technical efficiency is the production efficiency of enterprises due to
the factors such as management and technology, and the specific meaning is the
ability to get the maximum output under the given input. Scale efficiency is the
production efficiency which is influenced by the scale of enterprise. It reflects the
gap between the actual size and the optimal production scale. According to the
production status of activities in the part of returns to scale, the scale efficiency
can be divided into three states, increasing returns to scale (irs) and decreasing
returns to scale (drs) and constant returns to scale (crs).
It can be seen from Tables 3 and 4, in the 10 brands of Women’s Clothing
category, only 4 enterprises reach the ideal input-output ratio. They are Ochirly,
VEROMODA, GIRDEAR and ONLY. HSTYLE, Artka, INMAN, LIEBO’s mar-
keting effect are not ideal, because the pure technical and scale efficiency are
invalid. In addition, they are in a state of increasing return of scale. This shows
that its allocation of resources and the scale of investment are not good enough.
The low comprehensive efficiency of ELF SACK is due to the invalidation of
scale. The best method to improve the condition is to increases the scale of
activities and expands its exposure rate.
After the comprehensive comparison of the data table of other industry, the
article finds that in five industries, the integrate marketing effect of mobile phone
is best. 80% enterprises achieve the satisfied efficiency. Then the following is
Home Furnishing 70%, Skin Care 60%, and Snacks 50%. The worst result is
Women’s Clothing industry, only 40%. The low satisfaction of Women’s clothing
industry may be due to the fans’ activity participation is not high because of
its boring and unattractive contents. Its microblog mostly show new products,
seldom interacting with fans. So the official microblog is similar to enterprise’s
advertising platform, rather than a communication platform with fans. Lacking
of interaction and humanization, the marketing effect of the product is naturely
not very good.
(2) The Longitudinal Comparison
In order to better evaluate the microblog marketing effect between the five major
industries, this paper proposes to use the overall ranking of comprehensive tech-
nical efficiency in DEA results in the table to identify the success or failure
of enterprise marketing effect. Ranking the Comprehensive technical efficiency
value of the 50 brands from high to low sequence, if the technical efficiency is
equal, the brands are in same ranking. After that, the ranking will be as the
enterprise scores. Higher ranking namely that the ranking order is more back-
ward explains that the enterprise marketing effect is poorer. Calculate every
886 D. Zhang et al.
Table 5. The ranking status of comprehensive technical efficiency which is not 1
Ranking Enterprises Crste Vrste Scale

31 Pechoin 0.99 1 0.99 irs
32 Lin shi mu ye 0.968 1 0.968 irs
33 Haoxiangni 0.956 1 0.956 irs
34 Lenovo 0.955 1 0.955 irs
35 HSTYLE 0.895 0.898 0.997 irs
36 Nokia 0.673 0.677 0.994 drs
37 Laneige 0.655 0.704 0.93 drs
38 EsteeLauder 0.589 0.591 0.998 drs
39 Artka 0.543 0.545 0.997 irs
40 ELF SACK 0.522 0.522 1 -
41 PeaceBird 0.5 1 0.5 irs
42 INMAN 0.45 0.453 0.994 irs
43 MEIFUBAO 0.444 0.445 0.999 -
44 Three squirrel 0.414 0.414 0.999 irs
45 LIEBO 0.31 0.31 0.998 irs
46 BESTORE 0.243 0.244 0.996 drs
47 Be & Cheery 0.142 0.143 0.999 -
48 Xinnongge 0.127 0.128 0.997 -
49 Bell Land 0.021 1 0.021 irs
50 Gudxon 0.001 1 0.001 irs
enterprise’s score, and then classified calculate the score of each category. The
higher score shows that the category’s marketing effect is poorer.
Experimental data shows that the top 30 enterprises’ marketing effects are
very ideal, because the comprehensive technical efficiency, pure technical effi-
ciency and scale efficiency value are 1. The comprehensive technical efficiency of
last 20 enterprises are different, as shown in Table 5.
The data shows that in the top 30 enterprises, mobile phones companies have
8, accounting for 26.7%; Home Furnishing class has 7, accounting for 23.3%;
skin care companies have 6, occupying 20%; snacks has 5, accounting for 16.7%;
women’s clothing enterprises have 4, occupying 13.3%.
In whole 50 enterprises, achieving the satisfactory efficiency of enterprises in
the mobile phones industry accounted for 16%, followed by home furnishing for
14%, skin care for 12%, snacks for 10%. Women’s clothing class accounts for
only 8%. The above data shows that in five industries, the microblog marketing
effect of mobile phones companies are doing relatively well, women’s relatively
weak.
Then, classifying the ranking scores of all brands, mobile phone has the lowest
score 78 points. Home furnishing gets 138 points. Skin care is 155 points. Snack is
223 points. Women’s clothing gets the highest score 246 points. Therefore, from
the perspective of input and output, the input-output ratio of mobile phones is
the most ideal, which show that there are good microblog marketing effect in
this category. The home furnishing and skin care industries are also relatively
well. The input of snacks, women’s clothing is redundant, while the output is
insufficient. They should properly adjust the resources reasonably, display the
resources value to maximize to increase enterprise microblog marketing effect.
In summary, this paper finds the marketing effect of different goods of differ-
ent types of enterprises on sina microblog is different. The best effect is mobile
phones industry, followed by Home Furnishing, skin care, snacks. The worst is
the women’s clothing industry.
3.4 Results
(1) Microblog Marketing Input-Output Analysis
As can be seen from the previous section, the invalid marketing effects of enter-
prises have 3 kinds:
• The technical efficiency, pure technical efficiency, scale efficiency are all not 1;
• The technical efficiency, pure technical efficiency is not 1, scale efficiency is 1;
• The pure technical efficiency is 1, the technical efficiency, scale efficiency is
not 1.
The paper selects the first representative case to analyze. There are some
enterprises satisfying the condition, randomly select one company. As a result,
this paper chooses Be & Cheery as the sample. Table 6 is Be & Cheery’s unit
adjusted value: Table 6 The adjustment values of Be & Cheery.
The comprehensive efficiency value of Be & Cheery is 0.142. The pure tech-
nical efficiency value is 0.143. The scale efficiency value is 0.999. The compre-
hensive efficiency and pure technical efficiency value are very low, that shows
there is a big problem. From the table we can see, there are input redundancy
in 6 input indexes, especially on the number of followers. The original value of
followers is 80.650, but the actual target value is 5.508. Be & Cheery spends too
much energy on the increasing number of followers, and this part of input do
not actually convert into the corresponding output, which cause a great waste of
resources. Be & Cheery releases 42 microblog within 7 days, with high number
of active and passive comments. Relatively speaking, official microblog is active,
but the active behavior does not bring good interaction. The average number of
forwarding and comments are all only 5, the number of praise is 2. The output
is significantly inadequate. Besides, the prizes of Be & Cheery are attractive,
the price reaching 1489 yuan. However, they do not get corresponding fans par-
ticipation. In addition, Be & Cheery is good at guiding the followers’ emotion
and controlling the scale of marketing. The future improvement is mainly on the
allocation of resources. To sum up, the above adjustment values can provide an
important basis for the evaluation of enterprise marketing effects.
888 D. Zhang et al.
Table 6. The adjustment values of Be & Cheery
Enterprises Indicators Original Redundant Insufficient Effective

value input value output value target value
Be & Cheery NWL 1.881 0 0.965 2.846
NWF 5.929 0 0.005 5.934
NWC 5.714 0 0.953 6.668
OWF 6.639 0 0.015 6.654
OWC 6.25 0 0 6.625
NNS 0.992 0 0 0.992
Followers 80.65 −69.15 −5.992 5.508
microblog 6.13 −5.256 0 0.874
FRW 6 −5.144 0 0.856
NAC 0.667 −0.572 0 0.095
NPC 0.262 −0.225 0 0.037
PV 14.89 −12.767 −1.563 0.56
(2) The Improve Measurement for Enterprises

Through calculation of the DEA model, this paper finds that the microblog
marketing effect of firms in different industries is different. According to the dif-
ferent situation of the enterprise, this paper puts forward the following measures
to help enterprises improve.
1 Increase the enterprise’s official microblog exposure, propagandize enterprise
microblog through many ways.
According to the “2015 electric business microblog development report” released
by sina microblog, the current number of active electric business enterprises in
sina microblog account for a total of more than 1600. If you want to highlight
your own business account in so many official accounts, you must understand
the main way for users to pay attention to enterprise. Seize the basic demands
for user to follow the enterprise.
The sina “2015 microblog user development report” shows, users follow the
enterprise’s microblog primarily through three ways: advertising, others forward-
ing and corporate propaganda. Besides, Baird [1] thought the establishment of
social customer relationship, seized the customer psychology, and the use of inter-
active communication strategy were also important. In addition to the above
measures, the enterprise can also increase the promotion in the major holidays
and special time point.
2 Pay attention to negative emotions, encourage the positive emotions, and
guide the neutral mood.
Chatterjee [7] said that consumers always believe the negative information have
more critical value than the positive information when they make purchase deci-
sions or scan the comments. Therefore, they always have greater degree of depen-
dence on the negative information. Through microblog to guide consumers is the
main reason for some enterprise to do the microblog marketing. The negative
comments may cause certain extent damage to the product of enterprise. There-
fore, some enterprises must pay much attention to the negative emotion and
know the origin of it. That can help the enterprise to eliminate the negative
emotion.
For the positive emotions, the enterprise should take the initiative to interact
with the followers and encourage users to maintain this positive emotion, then
to increase the reputation of enterprise. For the neutral emotions, the enterprise
should show its advantage to consumers with the objective data.
3 Both active and passive comments should be concerned
Most followers want to get reply or attention when they give comments to the
blogger. If bloggers can timely find this kind of behaviors of followers, and then
give timely response, that can largely reduce the negative emotions and increase
the positive interaction with followers.
4 Diversify the types of microblog
Zhang [19] believed that the types of microblog can to a certain extent influence
the browsing and forwarding of fans. Generally the type of blog is divided into:
pictures, short chain, video, text, or a mixture of the four types. In order to
maintain the freshness of microblog, the enterprise should cross use multiple
types of posts to give user a new experience.
4 Conclusions
This paper through the literature research, put forward the evaluation indicators
of the microblog marketing effect in enterprises, and set up a set of comprehensive
evaluation system.
Combining with the method of data envelopment analysis (DEA), this paper
proposes the evaluation model, with input and output of the quantitative evalu-
ation method to evaluate the microblog marketing effect between different enter-
prises. In the empirical part, this paper selects 50 different enterprises to verify
the feasibility of the model. According to the objective data, this paper makes
a comparison of different enterprise marketing effect, and provides a feasible
improvement measures for it.
In fact, this paper also has certain limitation in the selection of indicators.
The article does not use the enterprise actual sales as the evaluation index.
Besides, this paper only studies the different enterprises in the short period. The
future research can increase the time and carry on deeper research on microblog
marketing effects.
Acknowledgements. The authors acknowledge the financial support National Nat-

ural Science Foundation Project of China, project (No. 71372189), and Sichuan
University Fundamental Research Funds for the Central Universities Project (No.
skqy201406). The author is grateful to the anonymous referee for a careful checking of
the details and for helpful comments that improved this paper.
890 D. Zhang et al.
References
1. Baird CH, Parasnis G (2011) From social media to social customer relationship
management. Strategy Leadersh 39(5):30–37
2. Barry AE, Bates AM et al (2016) Alcohol marketing on twitter and instagram:
evidence of directly advertising to youth/adolescents. Alcohol Alcohol 51(4):487–
492
3. Bi L (2013) Enterprise micro-blogging marketing effect evaluation model and
empirical research based on micro-blogging dissemination of information flow. J
Intell 7:67–71
4. Bressler BRG (2010) The measurement of productive efficiency. In: Western Farm
Economic Association, Proceedings, vol 3, pp 253–290
5. Chang YT, Yu H, Lu HP (2015) Persuasive messages, popularity cohesion, and
message diffusion in social media marketing. J Bus Res 68(4):777–782
6. Charnes A, Cooper WW, Rhodes E (1978) Measuring the efficiency of decision
making units. Eur J Oper Res 2(6):429–444
7. Chatterjee P (2001) Online reviews: do consumers use them? Adv Consum Res
28:133–139
8. Chen H (2015) College-aged young consumers interpretation of twitter and mar-
keting information on twitter. Young Consum 16(2):208–221
9. Clark EM, Jones CA et al (2016) Vaporous marketing: uncovering pervasive elec-
tronic cigarette advertisements on twitter. Plos One 11(7):e0157304
10. CNNIC (2016) Statistical report on internet development 38th China internet net-
work
11. Cooper WW, Seiford LM, Tone K (2001) Data envelopment analysis: a comprehen-
sive text with models, applications, references and dea-solver software. Springer,
Heidelberg
12. Kafeza E, Makris C, Vikatos P (2016) Marketing campaigns in twitter using a
pattern based diffusion policy. In: 2016 IEEE International Congress on Big Data
(BigData Congress), pp 125–132
13. Leung XY (2015) The marketing effectiveness of social media in the hotel industry:
a comparison of facebook and twitter. J Hospitality Tourism Res 39(2):147–169
14. Liu G, Amini MH et al (2016) Best practices for online marketing in twitter:
an experimental study. In: IEEE international conference on electro information
technology, pp 504–509
15. Park SB, Ok CM, Chae BK (2016) Using twitter data for cruise tourism marketing
and research. Travel Tourism Mark 33(6):885–898
16. Saulles MD (2015) Push and pull approaches to using twitter as a marketing tool.
In: Proceedings of the European conference on social media, pp 105–111
17. Sun YL (2008) Evaluation on the capacity of agricultural sustainable development
in sichuan province based on dea method. Soft Sci 6:100–103 (in Chinese)
18. Xue JP, Yu WP, Niu YG (2013) Research on brand communication effect of e-
commerce enterprises’ microblogging - a case study of 51buy. com’s microblogging.
Soft Sci 12:67–71 (in Chinese)
19. Zhang L, Peng TQ et al (2014) Content or context: which matters more in infor-
mation processing on microblogging sites. Comput Hum Behav 31(1):242–249
20. Zhu YQ, Chen HG (2015) Social media and human need satisfaction: implications
for social media marketing. Bus Horiz 58(3):335–345
Accounting for Clustering and Non-ignorable
Missingness in Binomial Responses:
Application to the Canadian National Survey
on Child Safety Seat
Syed Ejaz Ahmed1(B) , Abdulkadir A. Hussein2 , Anne Snowdon2 ,

and Yiwen Tang3
1
Brock University, 500 Glenridge Avenue, St. Catharines, ON L2S 3A1, Canada
babak rezaee@aut.ac.ir
2
University of Windsor, 401 Sunset Avenue, Windsor, ON N9B 3P4, Canada
3
Cancer Care Ontario, 505 University Avenue, Toronto, ON M5G 1X3, Canada
Abstract. Motivated by a Canadian national survey on child safety seat

use, we propose a procedure for incorporating non-ignorable missingness
in a binary response using logistic regression model with random effects.
The proposed method applies an expectation maximization (EM) algo-
rithm to the Penalized quasi-likelihood of the artificially completed data.
We provide closed form formulae for estimating the covariance of the
regression coefficients as well as the odds ratios of interest and probabil-
ities of missingness. The proposed algorithm can be easily implemented
by using familiar statistical packages that fit generalized linear mixed
models. The method is illustrated by applying it to data from Canadian
national survey on child safety seat use.
Keywords: Generalized linear mixed model · EM algorithm · Non-

ignorable missingness · Data augmentation · Logistic regression · Child
safety
1 Introduction
Subjects are likely to decline participation in a survey if the variables measured
are socially sensitive. This, raises the issue of non-ignorable missingness whereby,
the missignness mechanism depends on the value of the response itself. The
current research was motivated by data from a large Canadian survey in which
200 retail parking lots across Canada were random chosen and vehicles entering
to the parking lot were asked consent to participate in a survey which measured
the correct use of child safety seats in vehicles. Age, height, weight and type of
car safety seat used were recorded for up to three children in the participating
vehicles. By using these variables, a binary correct use variable was then defined.
892 S.E. Ahmed et al.
Unfortunately, drivers of more than 50% of the vehicles refused to participate in

the survey. Since correct uses of children in the same participating vehicle are
correlated, in addition to the missing responses, the data presented also a cluster
correlation feature.
Ibrahim et al. [3] have developed an EM algorithm for the generalized linear
mixed models (GLMM) with normal random effects and non-ignorable missing-
ness in the responses by using MCMC sampling techniques. Ibrahim et al. [3]
modeled the respondent data as a function of covariates by using GLMM with
logit link and likewise they modeled the non-response data as a function of the
same set of covariates as well as the response itself. Thus, they captured the non-
ignorable nature of the non-response through a logit link as well. The authors
then, proposed an EM (Expectation Maximization) algorithm along with MCMC
sampling for estimating the various parameters of the model.
Ibrahim and Molenberghs [8] have recently given an exhaustive review of
methods for accommodating non-ignorable missingness in longitudinal data.
Also, Huang and Carriere [2] discussed accommodation of incomplete responses
data in small sample longitudinal data.
In this paper, we propose a direct method which does not require Monte
Carlo sampling. The method proposed here is based on the penalized quasi-
likelihood of Breslow and Clayton [1] (PQL) which utilizes Laplacian method
of integral approximation of integrals. We use GLMM to model the respondent
data, thus accommodating the clustering. On the other hand, we model the
dependence of the non-response on covariates and on the true values of the
response itself through generalized linear model GLM as was done in Ibrahim
et al. [3]. However, we use the PQL method of Breslow and Clayton via Laplace
transform and Taylor series approximation [1,6] in order to eliminate the random
effects. We then apply an EM algorithm to estimate the regression coefficients
and Louis’s method [4] to obtain formulae for the their covariances. Following
Zeger et al. [9], and by using delta method, we compute marginal probabilities
of the response as well as odd ratios and probabilities of non-response and their
respective confidence intervals. The proposed method can be easily implemented
through familiar statistical software that fit GLMM. In Sect. 2 we give technical
details of the model and the algorithm used for fitting it as well as formulae for
variances of the various parameters of interest. In Sect. 3 we apply the method
to the Canadian national survey on child safety seat use. Finally, in Sect. 4, we
give some discussions.
2 The Proposed Method
2.1 Model Details
Let yi be a binomial variable with size mi and success probability pi . Suppose

that the complete data consists of a sample y = (y1 , · · · , yn ) , from such distri-
bution. Given an n-dimensional vector b = (b1 , · · · , bn ) of random effects, we
assume that yi | b follows generalized linear mixed model (GLMM), that is
Accounting for Clustering and Non-ignorable Missingness 893
yi | b ∼ Binomial (mi , pi ) ,
b ∼ M V N (0, D) ,
where M V N stands for multivariate normal distribution with covariance matrix

D, n × n diagonal matrix with diagonal elements σ 2 .
Let β = (β0 , β1 , · · · , βp ) be a vector of fixed effects regression coefficients
associated with the vector of covariates xi = (1, xi1 , · · · , xip ) . Let also, ti be an
1 × n vector of explanatory variables for the random effects associated with the
ith observation. The ith component of ti is 1 and the rest are all 0. Such vectors
can be compacted for all i = 1, · · · , n in an n×n identity matrix T = (t1 , · · · , tn ),
which is simply the design matrix associated with the random effects. By stan-
dard theory of GLMM, yi | b has the density

2 mi
f (yi | xi , β, σ , b) = exp yi log(pi /(1 − pi )) + mi log(1 − pi ) + log , (1)
yi
with the first two moments given by
E(Yi | b) = μi (b) = mi pi (2)

Var(Yi | b) = mi pi (1 − pi ). (3)
We use the logit link function to relate covariates (both fixed and random)
and the pi as follows
log(pi /(1 − pi )) = ηi (b) = xi β + ti b = xi β + bi . (4)
We introduce a missing data indicator variable ri for each yi observation in

the complete data,
1, if yi is missing;
ri =
0, if yi is observed.
Now, as in [3], we assume that the missingness indicators, r = (r1 , · · · , rn ) ,
are independent of the random effects, bi , but depend on the covariates xi
as well as on the value of the response, yi , through a logistic regression
model. Thus, by setting zi = (xi , yi ) to be the vector of explanatory vari-
ables for the non-response model with corresponding regression coefficients,
α = (α0 , α1 , · · · , αp+1 ) , we can write the density of the ri as
f (ri |zi , α) = exp {ri zi α − log (1 + exp {zi α})} ,
where the logit link function is again used to connect covariates and probability
of non-response πi = Pr(ri = 1 | xi , yi ) as follows
log(πi /(1 − πi )) = zi α.
The log-marginal likelihood of y can now be written as

logf (y|x, β, σ 2 ) = log f (y | x, β, σ 2 , b)f (b)db = log eh(b) db, (5)
where h(b) = logf (y | x, β, σ 2 , b) + logf (b).

Since the random effects are assumed to be normal, we have

1 n 1
logf (b) = − b D−1 b − log2π − log |D|
2 2 2
and hence,
h(b) = logf (y | x, β, σ 2 , b) + logf (b)
1 n 1
= logf (y | x, β, σ 2 , b) − b D−1 b − log2π − log |D| .
2 2 2
The log-likelihood in Eq. (5) is not easy to deal with as it involves a high
dimensional integral. To simplify it, we apply the Laplace approximation of
[1,6,7], based on a second-order Taylor series expansion,
2
. n 1 ∂ h(b)
h(b)
log e db = h(b̂) + log2π − log − , (6)
2 2 ∂b∂b
where b̂ is solution to

∂h(b)

−1
T (y − μ(b)) − D b = 0.
∂b b=b̂ b=b̂
It can be shown that [6],

∂ 2 h(b)
− = (T W T D + I)D−1 , (7)
∂b∂b
where W is the n × n diagonal matrix with diagonal element wi = mi pi (1 − pi ).
By using Eqs. (6) and (7), the log-likelihood in Eq. (5) can be
approximated by
. 1 1
logf (y|x, β, σ 2 ) = logf (y | x, β, σ 2 , b̂) − b̂ D−1 b̂ − log |T W T D + I| ,
2 2
or equivalently,
n

2 . mi 1 b̂2i
logf (y|x, β, σ ) = yi ηi (b̂) − mi log 1 + exp ηi (b̂) + log −
i=1
yi 2 σ̂ 2

1 2
− log mi exp ηi (b̂) 1 + exp ηi (b̂) σ̂ 2 + 1 . (8)
2
The log-likelihood so obtained is the penalized quasi-log-likelihood and most

of the available statistical software can fit such model.
Finally, the joint log-likelihood of the complete data and the missingness
indicator can be written as
l(α, β, σ 2 |r, y, x) = logf (r, y|x, α, β, σ 2 )
n

= logf (ri | zi , α) + logf (yi | xi , β, σ 2 ) .
i=1
This implies that the ith observation’s contribution to the log-likelihood is
li (α, β, σ 2 |ri , yi , xi ) = logf (ri | zi , α) + logf (yi | xi , β, σ 2 )
and so, the ith observation’s contribution to the expected log-likelihood is

⎧ mi
⎨ w (y )l (α, β, σ 2 |r , y , x ), if original y is missing;
i i i i i i i
E[li (α, β, σ 2 |ri , yi , xi )] = yi =0
⎩ 2
li (α, β, σ |ri , yi , xi ), if original yi is observed,
where the weights wi (yi ) are
f (ri | zi , α)f (yi | xi , β, σ 2 )

wi (yi ) = , (9)

mi
f (ri | zi , α)f (yi | xi , β, σ 2 )
yi =0
and f (yi | xi , β, σ 2 ) is the marginal density of yi obtained from Eq. (8). Note that
if the original yi is observed, then wi (yi ) = 1. Now, the expected log-likelihood
for all n observations in the complete binomial sample is
mi
n mi
n

L(α, β, σ 2 ) = wi (yi )logf (ri | zi , α) + wi (yi )logf (yi | xi , β, σ 2 ).
i=1 yi =0 i=1 yi =0
(10)
This last equality indicates that the model for the complete data and that of
the missingness indicator can be separated as in Ibrahim et al. [3], and therefore,
estimation of the regression parameters can be carried out separately.
2.2 Estimating the Regression Coefficients

The regression coefficients of the model can be estimated by using Eq. (10). First,
we create an augmented data set (yi , ri , xi ), in which each missing observation yi
is replaced by its mi + 1 possible values, then we apply the following algorithm
on the augmented data set.
(1) Set the initial weights to 1 and fit ordinary weighted GLMM to the aug-
mented (complete) data (y, x) by using any statistical software that fits
GLMM via penalized quasi-likelihood methods, and hence, obtain the start-
ing values of β, b, and σ 2 .
(2) Set the initial weights to 1 and fit weighted GLM to the missing indicator
data (r, z) and obtain the starting values of α.
(3) By using (9), update the weights wi (yi ) based on the current values of β, α,
b, and σ 2 .
(4) Update β, b, and σ 2 by fitting weighted GLMM to the augmented (complete)
data (y, x) with the current values of wi (yi ).
(5) Update α by fitting weighted GLM to (r, z) with the current values of wi (yi ).
(6) Update the weights wi (yi ) based on the new β, α, b, and σ 2 .
(7) Repeat steps 4–6 until β and α converge.
The final estimates of β, α, σ 2 are obtained upon convergence of the above

algorithm. Following Louis [4], the estimated observed Fisher information matrix
of the parameters (β, α, σ 2 ) can be computed as
I (α̂, β̂, σ̂ 2 ) = Q(α̂, β̂, σ̂ 2 ) − H(α̂, β̂, σ̂ 2 ),
where
mi
n

H(α̂, β̂, σ̂ 2 ) = ŵi (yi )Si (α̂, β̂, σ̂ 2 |ri , yi , xi )Si (α̂, β̂, σ̂ 2 |ri , yi , xi )
i=1 yi =0
n
− Ui (α̂, β̂, σ̂ 2 |ri , yi , xi )Ui (α̂, β̂, σ̂ 2 |ri , yi , xi ) ,
i=1
f (ri | zi , α̂)f (yi | xi , β̂, σ̂ 2 )

ŵi (yi ) = ,

mi
f (ri | zi , α̂)f (yi | xi , β̂, σ̂ 2 )
yi =0
mi

Ui (α̂, β̂|ri , yi , xi ) = ŵi (yi )Si (α̂, β̂, σ̂ 2 |ri , yi , xi ).
yi =0
The score vector Si (α̂, β̂, σ̂ 2 |ri , yi , xi ) is given by

∂li (α, β, σ 2 |ri , yi , xi )
Si (α̂, β̂, σ̂ 2 |ri , yi , xi ) =
∂(α, β, σ 2 ) (α,β,σ 2 )=(α̂,β̂,σ̂ 2 )
= ((ri − π̂i ), (ri − π̂i )xi1 , · · · , (ri − π̂i )zip , (ri − π̂i )yi , di , di xi1 ,

· · · , di xip , ∂li /∂σ 2 ,
where
mi σ̂ 2 p̂i (1 − p̂i )(1 − 2p̂i )
di = yi − μ̂i (b̂) − ,
2(mi σ̂ 2 p̂i (1 − p̂i ) + 1)
∂li b̂2i mi p̂i (1 − p̂i )
= − .
∂σ 2
2 (σ̂ 2 )
2 2(mi σ̂ 2 p̂i (1 − p̂i ) + 1)
The matrix
Q1 (α̂) 0
Q(α̂, β̂, σ̂ 2 ) =
0 Q2 (β̂, σ̂ 2 )
is such that the (j, k)th element of Q1 (α̂) is given by

∂ 2 L(α, β, σ 2 )
Q1,jk (α̂) = − ,
∂αj ∂αk α=α̂
and the (j, k)th element of Q2 (β̂, σ̂ 2 ) is

∂ 2 L(α, β, σ 2 )
Q2,jk ((β̂, σ̂ ) = −2
,
∂βj ∂βk β=β̂,σ2 =σ̂2

2
L(α,β,σ 2 )
for j, k = 0, 1, · · · , p and Q2,p+1,p+1 (β̂, σ̂ ) =
2
− ∂ ∂σ 2 ∂σ 2 . These
β=β̂,σ 2 =σ̂ 2
matrices are, respectively, the usual Fisher and penalized Fisher information
matrices associated with the GLM and GLMM models and they can be eas-
ily extracted from existing software upon convergence of the EM algorithm
described above.
2.3 Computing the Marginal Response Probabilities and

Probabilities Of Missingness
In binary response studies, investigators are often interested in the marginal

probabilities of the response of interest, p̄i = P (Yi = 1). This is, in fact, not
easy in a random effects model in which a logit link function has been used, due
to the non-linearity of the logit function. A way about this difficulty is to use
a Taylor approximation of the type proposed by Zeger et al. [9] in which p̄i are
approximately estimated by

exp k̂xi β̂
p̄î = Ê(yi ) ≈ ,
1 + exp k̂xi β̂
√
where k̂ = (1 + c2 σ̂ 2 )−1/2 and c = 16 3/(15π). By applying delta method to
η̂i = g(β̂, σ̂ 2 ) = k̂xi β̂, we obtain estimators of p̄î ,
V̂ (η̂i ) ≈ Ĝ Σ̂ Ĝ,

where Σ̂ is the estimated asymptotic covariance matrix of β̂, σ̂ 2 , that is the lower
right block of I −1 (α̂, β̂, σ̂ 2 ), and

∂g(β,σ 2 )
∂β |β̂,σ̂ 2
G = ∂g(β,σ 2
)
.
∂σ 2 |β̂,σ̂ 2
An approximate 95% confidence interval for ηi is then given by

L U
η̂i , η̂i ≡ η̂i − 1.96 Var(η̂i ), η̂i + 1.96 Var(η̂i ) ,
whereas an approximate 95% confidence interval for p̄i = exp {ηi } /

(1 + exp {ηi }) is
exp η̂iL exp η̂iU
, .
1 + exp η̂iL 1 + exp η̂iU
Finally, V̂ (p̄î ) can be computed by a second delta method as V̂ ar(p̄î ) ≈ [p̄î

(1 − p̄î )]2 V̂ ar(η̂i ). The probabilities of non-response, π̂i , can be easily obtained
from the output of the GLM model fit as these probabilities depend only on the
regression parameters αi .
3 Application
The data analyzed in this example are based on a Canadian child safety seats
survey described in Snowdon et al. [5]. The main objective of the survey was
to measure whether children traveling in vehicles on Canadian roads are cor-
rectly restrained in safety devices that are appropriate for their ages, weights
and heights. Drivers of vehicles entering in 200 randomly selected retail park-
ing lots across Canada were asked to participate in the survey. For the vehicles
whose drivers agreed to participate in the survey, heights, weights, ages and type
of safety restraints used were recorded for up to three child occupants in the vehi-
cle. For vehicles whose drivers refused to participate in the survey, the number
of children in the vehicle were recorded. In general, there is no consensus on the
definition of correct seat for a child. There are several criteria of correct use based
on one or combination of age, weight and height. In this article, we employed
correct use definition given in Table 1, based only on child’s age. We defined
the number of children who are correctly restrained in a vehicle as yi , assumed
Table 1. Definition of correct use of CSS based on the child’s age groups
Child safety seats Age groups

Rear-facing infant seat (Age 1 year)
Forward-facing infant seat (1 years < Age 4 years)
Booster seat (4 years < Age 9 years)
Seat belt (Age > 9 years)
Table 2. Sample proportions of correct use of CSS and Marginal probabilities of correct
use adjusted for missingnes
Category Proportion Estimate SE 95% CI

Lower Upper
AB 79.4% 65.87% 0.69% 64.50% 67.20%
BC 79.9% 68.01% 0.77% 66.49% 69.49%
MB 74.6% 60.55% 1.39% 57.79% 63.24%
NB 76.0% 68.35% 0.84% 66.68% 69.97%
NFL 72.9% 64.31% 1.12% 62.09% 66.48%
NS 79.3% 72.64% 1.00% 70.64% 74.56%
NT 70.4% 61.47% 2.62% 56.24% 66.46%
ON 84.1% 69.92% 0.54% 68.85% 70.97%
PEI 66.7% 62.20% 1.85% 58.51% 65.76%
QC 71.0% 60.81% 1.07% 58.69% 62.89%
SK 75.7% 65.10% 1.55% 62.01% 68.07%
Nation 78.6% 67.42% 0.34% 66.75% 68.09%
Table 3. Estimated regression parameters β.
Parameter Estimate SE t Value Pr >| t |

Intercept 0.9202 0.0315 29.2553 0.0000
AB −0.2032 0.0562 −3.6134 0.0003
BC −0.0977 0.0627 −1.5582 0.1192
MB −0.4530 0.0938 −4.8296 0.0000
NB −0.0803 0.0698 −1.1498 0.2503
NFL −0.2777 0.0841 −3.3010 0.0010
NS 0.1449 0.0889 1.6304 0.1031
NT −0.4105 0.1756 −2.3378 0.0194
PEI −0.3768 0.1357 −2.7761 0.0055
QC −0.4410 0.0774 −5.6954 0.0000
SK −0.2400 0.1107 −2.1669 0.0303
Table 4. Odds ratios of correct use of CSS and their 95% confidence intervals for all
provinces as compared to Ontario
Province Odds ratio 95% CI

Lower Upper
AB 0.8161 0.7124 0.9349
BC 0.9069 0.7806 1.0537
MB 0.6357 0.5062 0.7985
NB 0.9228 0.7852 1.0845
NFL 0.7575 0.6231 0.9209
NS 1.1559 0.9461 1.4123
NT 0.6633 0.4377 1.0052
PEI 0.6860 0.5077 0.9271
QC 0.6434 0.5363 0.7720
SK 0.7866 0.6048 1.0231
to be Binomial(mi , pi ), where mi = 1, 2, 3. For vehicles whose drivers refused

to participate in the survey, yi is not observed but only mi is known. There-
fore, our objective is to incorporate this missingness into the estimation of the
probabilities of correct use, pi , and probabilities of missingness, πi , for the vari-
ous provinces of Canada. The covariates of interest in this study are the dummy
codings for 10 Canadian provinces and territories, x1i , x2i , · · · , x10i . We assumed
the province of Ontario to be the reference province. We then applied the EM
algorithm proposed in this article and summarized results in Tables 2, 3, 4 and 5.
In Table 2, we reported the estimated adjusted marginal probabilities of correct
use for the various provinces as well as for the entire Canada. It is clear that the
adjusted rates exhibit downward bias as compared to the raw proportions.
Table 5. Estimated regression parameters α.
Parameter Estimate SE Z Value Pr > | Z |

Intercept 1.2630 0.0722 17.505 0.0000
AB 0.1908 0.1001 1.907 0.0566
BC −0.0396 0.1061 −0.374 0.7087
MB 0.2648 0.1820 1.456 0.1455
NB −0.8039 0.1068 −7.524 0.0000
NFL −0.7253 0.1310 −5.535 0.0000
NS −1.2439 0.1306 −9.522 0.0000
NT −0.1025 0.3148 −0.325 0.7448
PEI −1.4363 0.2204 −6.515 0.0000
QC −0.4741 0.1268 −3.739 0.0002
SK −0.3529 0.1830 −1.929 0.0538
yi −0.2964 0.0428 −6.927 0.0000
Table 6. Probabilities of non-response, π̂0 , π̂1 , π̂2 , π̂3 , respectively, when correct use
within vehicle is zero, one, two and three
Category Raw non response rate π̂0 π̂1 π̂2 π̂3

AB 75.95% 81.06% 76.09% 70.29% 63.75%
BC 70.74% 77.27% 71.65% 65.26% 58.28%
MB 79.14% 82.17% 77.41% 71.81% 65.44%
NB 55.10% 61.28% 54.06% 46.66% 39.41%
NFL 59.81% 63.13% 56.00% 48.62% 41.30%
NS 45.27% 50.48% 43.11% 36.04% 29.52%
NT 66.67% 76.14% 70.35% 63.82% 56.74%
ON 74.18% 77.95% 72.44% 66.15% 59.24%
PEI 43.62% 45.68% 38.47% 31.73% 25.68%
QC 65.53% 68.76% 62.07% 54.89% 47.49%
SK 69.83% 71.30% 64.88% 57.86% 50.52%
Nation 68.74% 72.73% 67.14% 61.02% 54.54%
In Tables 3 and 4, we reported the estimated regression coefficients, β̂, and

odds ratios based on these parameters. The provinces of Alberta, Manitoba,
Newfoundland, Prince Edward Island and Quebec have significantly lower odds
of correct use as compared to Ontario. In Table 5, the missing indicator regression
parameters, α̂, are reported. We clearly see that the coefficient of yi , α11 , is
statistically significant, which indicated that the missingness is non-ignorable.
In Table 6, we reported the non-response proportions as well as non-response
proportions estimated from the EM model. The probabilities of non-response
decrease as the number of correctly restrained children in a vehicle increase.

These probabilities are denoted by π̂i , where i = 0, 1, 2, 3 is the number of
correctly restrained children in the vehicle.
4 Conclusion
In this manuscript, we proposed a method for handling non-ignorable missing
responses in a clustered binary data. We employed GLMM along with penalized
quasi-likelihood estimation method for fitting the parameters of the complete
data and GLM for fitting those of the missingness mechanism and combined
them via the EM algorithm. Procedures for estimating the marginal probabilities
of the response are derived via delta method. The proposed methods are then
applied to a Canadian national survey for estimating rates of correct use of child
safety seats in vehicles. The proposed methods are attractive in the sense that
they only require existing statistical software that have capabilities for fitting
GLMM and GLM.
References
1. Breslow NE, Clayton DG (1993) Approximate inference in generalized linear mixed
models. J Am Stat Assoc 88(421):9–25
2. Huang R, Carriere KC (2006) Comparison of methods for incomplete repeated mea-
sures data analysis in small samples. J Stat Plann Infer 136(1):235–247
3. Liechty JC, Roberts GO (2001) Missing responses in generalised linear mixed models
when the missing data mechanism is nonignorable. Biometrika 88(2):551–564
4. Louis TA (1982) Finding the observed information matrix when using the EM algo-
rithm. J Roy Stat Soc 44(2):226–233
5. Snowdon AW, Hussein A et al (2009) Are we there yet? Canada’s progress towards
achieving road safety vision 2010 for children travelling in vehicles. Int J Inj Control
Saf Promot 16(4):231–237
6. Song XK (2007) Correlated data analysis: modeling, analytics, and applications.
Springer, New York
7. Tierney L, Kadane JB (1986) Accurate approximations for posterior moments and
marginal densities. J Am Stat Assoc 81(393):82–86
8. Ugarte MD (2009) Missing data methods in longitudinal studies: a review. Test
18(1):1–43
9. Zeger SL, Albert PS (1988) Models for longitudinal data: a generalized estimating
equation approach. Biometrics 44(4):1049–60
Equity and Sustainability Based Model
for Water Resources Planning
Yan Tu1 , Jun Gang2 , Benjamin Lev3 , and Xiaoyang Zhou4(B)

1
School of Management,
Wuhan University of Technology, Wuhan 430070, People’s Republic of China
tuyan1988@whut.edu.cn
2
Sichuan Institute of Building Research,
3
Decision Sciences Department, LeBow College of Business,
Drexel University, Philadelphia, PA 19104, USA
4
International Business School, Shaanxi Normal University,
283509632@qq.com
Abstract. This paper presents a decentralized multi-objective bi-level

model for water resources planning, in the upper level of the model, the
regional authority as the leader determines to allocate how much water
to each sub-area to maximize equity and minimize the environmental
pollution, and then the sub-areas in the lower level make decisions on
allocating water to water sectors based on different allocation principles
with the analysis of relationship between supply and demand to max-
imize the benefits. In the decision making process, fuzzy random envi-
ronment is considered to describe the internal uncertainty, an expected
value operator is used to deal with the fuzzy random variables. Finally,
a case study is used to verify the applicability of the proposed model.
Keywords: Water resources allocation · Equity · Sustanability ·

Bi-level programming · Fuzzy random variable
1 Introduction
In the water resources allocation system, there exists the interaction between
regional authority and sub-areas in a river basin [11,12]. Efficiency and sustain-
ability of water resources allocation often conflict with each other, for without
considering ecological water to destroy the sustainability of water resources allo-
cation, other water sectors that can generate economic benefits can obtain more
water, the efficiency of water resources allocation will improve [4,5,8]. In addi-
tion, to guarantee the efficiency and sustainability of water using will lead the
contradiction between two-level decision makers (i.e., regional authority and sub-
areas).

Equity and Sustainability Water Resources Planning 903
Rogers and Louis [9] pointed out that there were a variety of activities and
objectives in a water system as well as complicated supply and demand contra-
dictions, which brought pressures to regional authority and sub-areas managers,
because these contradictions and ineffectiveness management limit the economic
development and environment protection. Therefore, an equitable and efficient
water resources allocation is an important measure to deal with water crisis and
improve water management, especially when it is very lack of water.
Based on the above, this paper considers to establish a user-friendly water
resources allocation model, with two-level structure, in which the regional
authority and sub-areas are the upper level and lower level decision makers
respectively, as there are multiple decision makers in the lower level, it is a
decentralized bi-level model. In the model, the regional authority as the leader
aims to maximize the equity (i.e., minimize the Gini coefficient), the sub-areas
as the followers expect to maximize the economic benefits. In addition, prac-
tical water resources allocation process faces complex uncertain environments,
technology progress and equipment update and climatic change will bring uncer-
tainty to parameters in the water transport process, and it is hard to describe the
uncertainty using simple random variable or fuzzy variable, therefore, the fuzzy
random variable is considered. Based on the discussion above, a decentralized
multi-objective bi-level model with fuzzy random coefficients will be established
in this paper.
The reminder of this paper is structured as follows. In Sect. 2, the problem
statement is given, including the bi-level structure of the model and the motiva-
tion to consider fuzzy random variable. Then an expected decentralized bi-level
model is established in Sect. 3. In Sect. 4, a case study is conducted to verify the
applicability of the model. Finally a conclusion remark is given in Sect. 5.
2 Problem Statement
For a regional water resources allocation system, in the case of insufficient water
supply, the limited water should be allocated optimally to each sub-area, and
then be allocated by each sub-area to different water sectors (i.e., ecological
water, municipal water, industrial water, agricultural water) [10]. The fair allo-
cation of water can balance the contradiction between water supply and demand,
therefore, the regional authority the leader in the upper level expects to max-
imize the equity of water allocation, in the lower level, the sub-areas hope to
maximize the benefits through water allocation based on different allocation
principles (i.e., efficiency principle, stress principle, priority principle) with the
analysis of relationship between supply and demand. The bi-level structure of
the water allocation model is shown in Fig. 1.
The need to address uncertainty in water allocation system is widely recog-
nized, there is a strong motivation for considering the fuzziness and randomness
in water resources planning problems [1,3,6]. For example, the loss ratio of water
transfer is not fixed because of the effect of many uncertain elements such as
lack of historical data, technology progress and equipment update, dynamics of
904 Y. Tu et al.
Upper level: Allocation among subareas
Avaliable water in Objective:

the basin Regional Equity maximization
authority
Equity principle Allocating water
feedback
Lower level: Allocation among water sectors
Sub-area 1 Sub-area 2 Sub-area m
Allocating Allocating Allocating

water water water
Ecological water (meet first) Ecological water (meet first) Ecological water (meet first)
Minicipal water Municipal water Municipal water
Agricultural water Agricultural water Algricultural water
Industrial water Industrial water Industrial water
Supply and demand

Allocation bsed on no Avaliable water in determ inatio n
water demand sub-areas<water demand? Allocation principles
yes
Efficiency principle no Avaliable water in sub-areas yes

Priority principle
Stress principle <minimum water requirement?
O bjective :
Benefit maximization
Fig. 1. Bi-level model structure of water allocation
water resources systems, in this case, it is appropriate to use a fuzzy variable to

describe the loss ratio of water transfer, which is the fuzziness of the loss ratio of
water transfer. The decision making is happen in the future, some experts will
be asked to describe the variable using different ranges, which is the random-
ness of the loss ratio of water transfer. In this study, triangle fuzzy number is
used to describe the fuzziness, normal distribution random variable is used to
describe the most possible value in the triangle fuzzy number. Therefore, fuzzy
random variable is used to describe the two-hold uncertainty of the loss ratio of
water transfer, the flowchart of the loss ratio of water transfer as a fuzzy random
variable is shown in Fig. 2.
Fig. 2. The flowchart of the loss ratio of water transfer as a fuzzy random variable
3 Modelling
3.1 Assumptions
The following assumptions are considered for constructing the model.
(1) All available water is from a river basin.

906 Y. Tu et al.
(2) The leader (i.e., regional authority) and the followers (i.e., sub-areas) will
make rational decision makings for avoiding unsatisfying solutions.
(3) Loss ratio of water transfer is considered as a fuzzy random variable, in which
the parameters are determined using data analysis based on historical data
and experience.
3.2 Notations
Indices
i: index of sub-area;
j: index of water sector;
k: index of special water sector that requirements should be met;
Parameters
Q: available water of basin;
ei,j : economic benefit for using water of water sector in sub-area;
ei : economic benefit for using water in sub-area;
oi,j : amount of the pollutant BOD in unit wastewater discharge of water sector
in sub-area;
wi,j : wastewater discharge coefficient of water sector in sub-area;
vi : economic benefit of sub-area;
qi : water source of sub-area;
Zimin : minimum capacity of the physical connection between water source and
sub-area;
Zimax : maximum capacity of the physical connection between water source and
sub-area;
Simax : maximum capacity of sub-area;
dmin
i,j : minimum water requirement of water sector in sub-area;
dmax
i,j : maximum water requirement of water sector in sub-area;
edmax
i,j : minimum ecological water requirement of sub-area;
Fuzzy random variables

˜ iloss : loss ratio of water transfer from water source to sub-area;
ᾱ
Decision variables
xi : water allocated to sub-area;
yi,j : water allocated to water sector in sub-area
3.3 Model Formulation
Based on the problem statement and the above preparatory work, the decen-
tralized bi-level model for water resources allocation with analysis of supply and
demand is established as follows.
m
m

1 xk xl

max G =
x
m
xk
ek − el (1)
2m ek
k=1 l=1
k=1
m
n
min P = 0.01oi,j wi,j yi,j (2)
x
i=1 j=1
m

s.t. xi ≤ Q (3)
i=1
t

(1 − E[ᾱ
˜ iloss ])xi + qi ≥ dmin
i,k , ∀i (4)
k=1
zimin ≤ xi ≤ zimax , ∀i (5)
(1 − E[ᾱ
˜ iloss ])xi + qi ≤ smax
i , ∀i (6)
For given x, y solves
n
max vi = ei,j yi,j , ∀i (7)
y
j=1
i,j ≤ yi,j ≤ di,j ,

s.t. dmin ∀i, j
max
(8)
n
(1 − E[ᾱ
˜ iloss ])xi + qi − yi,j ≥ edmin
i , ∀i (9)
j=1
Equation (1) represents the first objective of the leader (i.e., regional
authority), which aims to maximize the equity of the water allocation based on
Gini coefficient [13]. Equation (2) represents the second objective of the leader
(i.e., regional authority), which aims to minimize the environmental pollution to
achieve the sustainability in water using at most. Constraint in Eq. (3) requires
the water allocated to all sub-areas should be less than the available water in
the river basin. Constraints in Eq. (4) require the water allocated to sub-area
should meet the water using of special waters in sub-area. E[•] is the expected
value operator proposed in Liu [7]. Constraints in Eq. (5) describe the range of
water allocated to sub-area. Constraints in Eq. (6) restrict all water sub-area
can get should be less than its maximum capacity. Equations in Eq. (7) repre-
sent the objectives of the followers (i.e., sub-areas), which aim to maximize the
economic benefit of all the sub-areas. Constraints in Eq. (8) define the range of
water allocated to water sector of sub-area. Constraints in Eq. (9) require that
the ecological water should be met in each sub-area.
4 Model Analysis
From the decentralized bi-level model, the following results can be obtained.
3
(1) When (1 − E[ᾱ ˜ loss
i ])xi + qi − edi
min
< j=1 dmini,j , allocation principle: equity
principle in the upper level, priority principle in the lower level.
908 Y. Tu et al.
3
(2) When (1 − E[ᾱ ˜ iloss ])xi + qi − edmin ≥ j=1 dmin i,j ,
i
3
xi ≥ min
j=1 di,j +edi
min
− qi /(1 − E[ᾱ
˜ iloss ]), and ei1 = ei2 = ei3 , allocation
principle: equity principle in the upper level, stress principle in the lower
level. 3
(3) When (1 − E[ᾱ ˜ iloss ])xi + qi − edmin
i ≥ j=1 dmin i,j ,

3
xi ≥ min
j=1 di,j +edi
min
− qi /(1 − E[ᾱ
˜ i ]), and ei1 = ei2 = ei3 , allocation
loss
principle: equity principle in the upper level, benefit-based principle in the

lower level.
5 Case Study
Qujiang basin is taken as a case example to illustrate the proposed model.
Qujiang river is the largest tributary of Jialing river basin in the upper reaches
of Yangtze, it is across three provinces in southwest China, in this paper, the
part of Qujiang basin in Sichuan Province is studied (see Fig. 3). From Fig. 3,
the available water is supplied to five sub-areas: Bazhong (BZ), Nanchong (NC),
Guang’an (GA), Guangyuan (GY), Dazhou (DZ). The data for the related para-
meters can be seen in Tables 1, 2 and 3.
Sichuan Province
China
Qujiang
Guangyuan
Bazhong
Dazhou Qujiang Basin

Nanchong
Qujiang
Guang an
Fig. 3. The map of studied Qujiang river basin

Table 1. Related parameters of 3 water sectors in 5 sub-areas
˜ loss
Sub-area ᾱ qi (104 m3 ) smax (104 m3 ) zimax (104 m3 ) zimin (104 m3 ) ei,j (104 m3 )
i i
IND AGR MUN
BZ (0.42, αloss
1 , 0.47) 42764 50000 53157 47905 60.61 42.55 43.86
αloss
1 ∼ N (0.45, 0.01)
NZ (0.36, αloss
2 , 0.40) 23870 40000 45320 39514 54.35 34.48 32.26
αloss
2 ∼ N (0.38, 0.04)
GA (0.28, αloss
3 , 0.32) 33997 40000 46889 18622 67.57 47.62 47.17
αloss
3 ∼ N (0.30, 0.01)
GY (0.45, αloss
4 , 0.55) 2874 30000 38543 12448 74.07 31.25 37.88
αloss
4 ∼ N (0.50, 0.09)
DZ (0.29, αloss
5 , 0.35) 115382 60000 67832 26308 86.96 43.11 45.45
αloss
5 ∼ N (0.32, 0.01)
Note: IND—Industrial water; AGR—Agricultural water; MUN—Municipal water.
Table 2. Water demand of 3 water sectors in 5 sub-areas (104 m3 )
Sub-area Industrial water Agricultural water Municipal water Ecological water

dmax
i,1 dmin
i,1 dmax
i,2 dmin
i,2 dmax
i,3 dmin
i,3
BZ 48848 11715 76465 51917 15510 11410 285

NC 47202 5643 42032 36349 7845 5675 163
GA 45149 16356 44277 33187 9616 6697 149
GY 35988 653 4224 3422 812 659 17
DZ 78174 52444 141104 104172 28092 21702 598
Before solving the proposed model, the expected values of loss ratios of water
transfer should be given. For the loss ratios of water transfer are triangle fuzzy
random variables, in which the random parameters follow normal distributions.
The expected value operator for triangle fuzzy random variables proposed in Xu
et al. [12] can be applied to deal with fuzzy random loss ratios of water transfer.
Thus, the model can be subsequently processed with the expected value of loss
ratios of water transfer.
For dealing with the multiple objectives in the upper level, the weighted sum
method based on satisfactory degree proposed by Gang et al. [2] is applied, the
Table 3. Water using pollution parameters of 3 water sectors in 5 sub-areas
Sub-area Industrial water Agricultural water Municipal water

oi,1 wi,1 oi,2 wi,2 oi,3 wi,3
BZ 250 0.75 650 0.45 120 0.9
NC 300 0.85 850 0.55 160 0.8
GA 275 0.8 750 0.5 140 0.95
GY 325 0.7 1000 0.45 150 0.85
DZ 200 0.9 800 0.6 170 0.75
910 Y. Tu et al.
Table 4. Water allocation results (104 m3 )
Sub-area Industrial water Agricultural water Municipal water Water allocated

by authority
BZ 5550 51917 11410 47955
NC 6050 36349 5675 39364
GA 21858 33187 6697 39792
GY 4850 3322 659 12298
DZ 23289 104172 21702 50490
weights are set as: for the first objective (i.e., the equity objective), for the second
objective (i.e., the pollution control objective). Based on the data in Tables 1, 2
and 3, the water allocation planning of this case can be shown in Table 4.
From the results in Table 4, we know that the total available water in Qujiang
basin (i.e., 4089 × 106 m3 ) is less than the total minimum demand in all sub-
areas, but is more than the total minimum demand for agricultural sectors and
municipal sectors, which can be known that the water shortage situation is seri-
ous currently in the basin. In this paper, the regional authority makes decisions
on allocating water based on equity (mainly) and pollution control. Through
the comparison of the results with the practical allocation, it can be found that
there is scarce water, the minimum demand for agricultural and municipal water
sectors of all sub-areas can be satisfied using the proposed model. However, for
more economic benefits, more water is allocated for industrial sectors, and the
minimum demand for agricultural and minimal using cannot be satisfied in prac-
tice, which is not reasonable for the steady and development of the river basin.
What’s more, the decision making of the proposed model can effectively control
the pollution under the premise of guaranteeing equity.
6 Conclusion
This paper proposed a decentralized multi-objective bi-level model considering

equity, economic benefit and ecological environment for water resources alloca-
tion. In the upper level of the model, the regional authority aimed to achieve
two objectives of equity and environmental pollution control by allocating water
to sub-areas, in the lower level of the model, the sub-areas expected to get the
maximal economic benefit by allocating water to different water sectors. In the
model, fuzzy random loss ratio of water transfer was considered. Finally, a case
study was conducted to verify the applicability of the proposed model. The com-
prehensive research on equity, sustainability and effectiveness of water resources
allocation under different uncertain environments will be the future research
direction.
Acknowledgement. This research was supported by the National Natural Science

Foundation of China (Grant No. 71401093), the Fundamental Research Funds for
the Central Universities (WUT: 2016VI002), the Fundamental Research Funds for the
Central Universities (Grant No. 14SZYB08), Soft Science Research Project of Shaanxi
Province (Grant No. 2016KRM089) and Research Center for Systems Science & Enter-
prise Development (Grant No. Xq16B01).
References
1. Ajami NK, Hornberger GM, Sunding DL (2008) Sustainable water resource man-
agement under hydrological uncertainty. Water Resourc Res 44(11):2276–2283
2. Gang J, Tu Y et al (2014) A multi-objective bi-level location planning problem for
stone industrial parks. Comput Oper Res 56:8–21
3. Guo P, Huang GH et al (2010) A two-stage programming approach for water
resources management under randomness and fuzziness. Environ Model Softw
25:1573–1581
4. Hu Z, Chen Y et al (2016) Optimal allocation of regional water resources: from a
perspective of equity–efficiency tradeoff. Resour Conserv Recycl 109:102–113
5. Lartigue C (2015) Efficient use of water resources for sustainability. Springer,
Switzerland
6. Li YP, Liu J, Huang GH (2014) A hybrid fuzzy-stochastic programming method
for water trading within an agricultural system. Agric Syst 123(2):71–83
7. Liu B (2004) Uncertainty theory: an introduction to its axiomatic foundations.
Springer, Berlin
8. Ni J, Xu J, Zhang M (2016) Constructed wetland planning-based bi-level opti-
mization to balance the watershed ecosystem and economic development: a case
study at the Chaohu lake watershed, China. Ecol Eng 97:106–121
9. Rogers JW, Louis GE (2007) A financial resource allocation model for regional
water systems. Int Trans Oper Res 14(1):25–37
10. Song WZ, Yuan Y et al (2016) Rule-based water resource allocation in the Central
Guizhou Province, China. Ecol Eng 87:194–202
11. Tu Y, Zhou X et al (2015) Administrative and market-based allocation mechanism
for regional water resources planning. Resour Conserv Recycl 95:156–173
12. Xu J, Tu Y, Zeng Z (2012) Bilevel optimization of regional water resources alloca-
tion problem under fuzzy random environment. J Water Resour Planning Manage
139(3):246–264
13. Yitzhaki S (1979) Relative deprivation and the Gini coefficient: reply. Q J Econ
93(2):321–324
SCADA and Artificial Neural Networks
for Maintenance Management
Alberto Pliego Marugán(B) and Fausto Pedro Garcı́a Márquez
Ingenium Research Group, Universidad Castilla-La Mancha, 13071 Ciudad Real, Spain
alberto.pliego@uclm.es
Abstract. Nowadays, the reliability of the wind turbines is essential to ensure

the efficiency and the benefits of the wind energy. The SCADA system installed
in a wind turbine generates lot of data that need to be processed. The informa-
tion obtained from these data can be used for improving the operation and man-
agement, obtaining more reliable systems. The SCADA systems operate through
different control rules that are predefined. However, a static control of the wind
turbine can generate a miscorrelation between the control and the real conditions
of the wind turbine. For example, two wind turbines can be separated several
kilometers in the same wind farm, therefore, the operation conditions must be dif-
ferent and the control strategy should not be unique. This research work presents
a method based on neural networks for a dynamic generation of the control strat-
egy. The method suggests that the thresholds used for generating alarms can vary
and, therefore, the control of the wind turbine will be adapted to each specific
wind turbine.
Keywords: Wind turbine · Reliability · SCADA systems · Advanced control

analytics · Artificial neural networks
1 Introduction
The wind energy is currently the most important renewable energy, the capacity
installed currently is more than 420 GW and it is estimated to be more than 1000GW
in 2030 [3]. The maintenance and operation costs of conventional wind turbines are
12% of the total costs, but the wind energy is evolving towards the offshore location. It
causes an important increase of these costs, being for offshore wind farms around 23%
of the total costs [12].
SCADA systems are widely introduced in wind turbines (WTs) due to their effec-
tiveness has been proved in other industries for detection and diagnostics of failures
[9, 11, 16]. They are presented as an inexpensive and optimal solution [20] to control
feedback for the health monitoring while reducing the operation and management costs
[19]. Nevertheless, they also present some minor disadvantages due to the operational or
reliability conditions [14, 21]. These systems consider a large amount of measurements
such as temperatures or wind and energy conversion parameters [18]. Data have raised
considerable interest in different areas, e.g. wind power forecasting [17], production
assessment [22] or fault detection [4, 6, 8, 10].
SCADA and Artificial Neural Networks for Maintenance Management 913
In the case of WTs, the introduction of SCADA systems verifies the efficiency when
their components are deteriorated. This degradation can indicate problems of different
nature such as misalignments in the drive-train, friction caused by bearing or gear faults.
The basic elements of the performance monitoring consist on a first collection of raw
values by the sensors. After the application of the appropriate filters, anomalies are
detected. Finally, a diagnosis will be provided. The anomaly detection includes a series
of techniques that range from simple threshold checks to statistical analyses [5, 7, 15].
It has been demonstrated that the WT suffers a gradual loss of production. Figure 1
shows that the power has a decrease year by year (this study has been developed by
Sheffield university in the OPTIMUS project [2]). The SCADA data has been consid-
ered over 5 years and the wind speed-power curve has been estimated.
A103
750
700
650
2010
2011
600 2012
2013
2014
550
500
8.5 9 9.5 10 10.5
Fig. 1. Loss of power of a WT over the time
This paper present a signal processing considers the minor changes in the behavior
of the WT. Some alarms will be affected by the decrease of the power shown in Fig. 1.
The following alarms will be activated considering the power as a cause.
• Activation of the ice safe mode: One of the statements considered for this alarm is
that the power is low for the measured wind.
• High aerodynamic deterioration: It is considered for low power with high wind.
If the power presented a minor reduction, then some false alarms would be activated
because the SCADA was prepared for a higher power. However, if the control laws are
dynamic, this problem will be solved.
2 Proposed Method
The method proposed aims to improve the control of the system by following a super-
vised iterative process. The objective is to determine if the SCADA control is coherent
914 A.P. Marugán and F.P.G. Márquez
Fig. 2. Flowchart of the method proposed
with the historical data and, therefore, if the system has changed. The main capability
of the method is to find differences in the behavior of the system over the time. Figure 2
shows a flowchart of the method proposed.
The iterative process starts with a database of the historical SCADA data. In this
paper, two different types of data are considered: the value of the measured parameters
and an alarm report. This data is employed to build an artificial neural network (NN)
with a supervised training. The values of the parameters are employed as the inputs and
the alarms will represent the outputs of the NN.
NN are used in problems that cannot be formulated as an exact method or an analyti-
cal solution. NN learns by itself and provides a good solution for the problem simulating
the biological neurons in a reasonable time. An artificial NN consists of neurons that
are simple processing units and weighted connections between those neurons [1].
It can be used for improving the control of the WT when the NN that defines the
logic of the system has been generated. The SCADA system will provide online data
of the condition of the WT. These data will be processed following two different eval-
uations: the SCADA system processing and the NN processing. Both outcomes will be
compared. If the outcomes are equal, then the data will be added to the database, there
for the NN processing will became an adaptive process.
3 Case Study
The European Optimus project [2] has provided the SCADA data used in this work.
It measures a lot of parameters, but after a filtering process, only 34 parameters (see
Table 1) have been considered for the analysis presented hereby. The values of this
parameters will be considered as inputs for the NN.
Basically, the NN receives a dataset and make a training process to recognize sev-
eral patterns. The training process fits the different weights to provide the output. If
the output is known, then the training is defined as supervised, otherwise, it is called
unsupervised training. The condition of the WT corresponds to the desired outputs of
the NN. The data used to design the NN is divided in following groups:
Table 1. Inputs and outputs for neural network
No Signal No Signal
1 General accumulator blade 1 pressure 18 Environmental temperature
2 General accumulator blade 2 pressure 19 Drive end side generator bearing temperature
3 General accumulator blade 3 pressure 20 Non-drive end side generator bearing temperature
4 Phi cosine 21 Generator winding temperature
5 Turbulence level 22 Nacelle temperature
6 Oscillation level 23 Lower gearbox radiator
7 Vibration level 24 Upper gearbox radiator
8 Pitch 1 angle 25 Gearbox bearing temperature
9 Pitch 2 angle 26 Transformer 1 temperature
10 Pitch 3 angle 27 Transformer 2 temperature
11 Active power 28 Transformer 3 temperature
12 General accumulator pressure 29 Grid voltage
13 Brake pressure 30 Total reactive power
14 Hydraulic group pressure 31 Generator speed
15 SP pitch angle 32 Rotor speed
16 Hydraulic group oil temperature 33 Wind speed
17 Gearbox oil temperature 34 Yaw
• Training set: Around 75% of the total amount of data.

• Validation set: Around 15% of the total amount of data.
• Testing set: Around 15% of the total amount of data.
The NN is expected to be able to learn from data and predict the output when a
generic input is considered.
The output of the NN will be specified by a Report of Alarms generated by the
SCADA [2]. The data available consists of a serial of alarms registered from 01/10/2012
to 15/05/2015, where there are different types of alarms identified by a specific code.
The codes are not indicated for reasons of confidentiality. In this work, only 5 different
scenarios have been considered to simplify the example. These scenarios correspond to
4 different alarms and the absence of alarms.
Therefore, the NN will be created following the structure shown in Table 2.
The design of the NN has been carried out using the geometrical pyramid rule [13].
Figure 3 shows the structure of the NN built. This rule express an approximation of the
number of neurons h at the hidden layer. It is defined by:
√
h = m × n,
where m is the number of elements of each input and n is the number of possible outputs,
being m = 34, n = 5, and therefore, h = 13.03.
Figure 4 shows the statistics of the NN generated by using the SCADA data. This
is a global confusion matrix and corresponds to the sum of the training, validation and
testing sets. The rows of the matrix show the outcomes of the NN for the dataset used.
Table 2. Inputs and outputs for NN
Input data Output data

1 · · · 34 Alarm 1 Alarm 2 Alarm 3 Alarm 4 None Alarm
Date 1 a11 a··· 34
1 a1 j 1 0 0 0 0
Date 2 a12 a···
2 a34
2j 0 1 0 0 0
Date 3 a13 a···
3 a34
3j 0 0 1 0 0
Date 4 a14 a···
4 a34
4j 0 0 0 1 0
Date 5 a15 a···
5 a34
5 0 0 0 0 0
m n
(hidden layer) Alarm 1
1
2 Alarm 2
Number of neurons
3 Alarm 3
h=13 Alarm 4
…
Fig. 3. Structure of the neural network
1 28 0 0 0 1 96.6%
28.6% 0.0% 0.0% 0.0% 1.0% 3.4%
2 0 9 1 0 1 81.8%
0.0% 9.2% 1.0% 0.0% 1.0% 18.2%
3 0 0 9 0 1 90%
Output
0.0% 0.0% 9.2% 0.0% 1.0% 10%
4 0 0 0 9 0 100%
0.0% 0.0% 0.0% 9.2% 0.0% 0.0%
5 0 2 1 0 36 92.3%
0.0% 2.0% 1.0% 0.0% 36.7% 7.7%
6 100% 81.8% 81.8% 100% 92.3% 92.9
0.0% 18.2% 18.2% 0.0% 7.7% 7.1%
1 2 3 4 5 6
Target
Fig. 4. Results of the neural network: confusion matrix
The columns show the real output, i.e. the output established by the alarms. The diago-
nal of this matrix (grey cells) contains the desired solutions. The sixth row and the sixth
column show a summary of the rights (green percentage) and wrongs (red percentage).
The outcomes of the NN agree more than 90% with the alarms that the SCADA
system generates. The method proposed establishes that the control of the SCADA
needs to be uploaded when the alarms cannot be predicted by the NN with enough
accuracy. If the SCADA control and this NN have the same response, then the data
processed will be added to the database. Therefore, a “healthy” dataset will be created
and the control will be adapted to the real conditions over the time.
4 Alarm Prediction
The method proposed here by also allows for predicting a possible alarm. In the previ-
ous Sect. 3, a NN was generated by inputting the data that the SCADA system employs
for generating the alarms. In this section, the input will be the data obtained before the
alarm is activated, i.e. the objective is to predict an alarm before the SCADA system
generates it. The NN is performed using the same techniques that in previous section. In
this case the NN will be designed to distinguish whether alarm will be activated or not.
Figure 5 shows the scheme used for the prediction. The inputs employed for performing
the NN are the dataset collected before the alarms are activated.
Fig. 5. Structure of the second neural network
Following the structure of Fig. 5, a NN has been performed obtaining the results
shown in Fig. 6. This confusion matrix shows that the NN can predict a 20.1% of the
alarms. Despite it is a low percentage when the NN suggests an alarm, it has a success
of 62.1%.
5580 143 97.5%

1
96.5% 2.5% 2.5%
22 36 62.1%
2
Output 0.4% 0.6% 37.9%
99.6% 20.1% 97.1%
0.4% 79.9% 2.9%
1 2
Target
Fig. 6. Results of the second neural network: confusion matrix
This method can be used for determining the predictability of some alarms. This
can be a useful tool to identify possible alarms before the WT can be damaged.
5 Conclusions
In this work, a new methodology is presented for extracting information from SCADA
dataset. The methodology is based on the generation of neural network from the quanti-
tative (value of parameters) and qualitative dataset (alarms) of the SCADA system. The
methodology has two different purposes. In first place, the adaptation of the SCADA
control rules to the variable condition of the wind turbines. A neural network has been
generated to determine which data should be added to the database. The creation of a
“healthy” database allows for adapting the SCADA control rules to the real condition
of the wind turbine over the time. Secondly, an additional neural network has been cre-
ated for making predictions of the activation of alarms. This can be used to identify an
abnormal state of the wind turbine earlier than the SCADA.
Acknowledgement. The work reported herewith has been financially by the Spanish Ministerio
de Economı́a y Competitividad, under Research Grant Ref.: DPI2015-67264-P.
References
1. Arbib MA (2003) The handbook of brain theory and neural networks. MIT Press
2. Camacho E, Requena V et al (2014) Demonstration of methods and tools for the optimisation
of operational reliability of large-scale industrial wind turbines. In: International conference
on renewable energies offshore renew
3. Council GWE (2015) Global wind report annual market update 2014. Technical report, http://
www.gwec.net/wp-content/uploads/2015/03/GWEC Global Wind 2014 Report LR.pdf
4. Gómez Muñoz CQ, Márquez FPG (2016) A new fault location approach for acoustic emis-
5. González-Carrato RRDLH, Dimlaye V, Ruiz-Hernández D (2014) Pattern recognition by
wavelet transforms using macro fibre composites transducers. Mech Syst Signal Process
48(1):339–350
6. González-Carrato RRDLH, Márquez FPG, Dimlaye V (2015) Maintenance management
of wind turbines structures via mfcs and wavelet transforms. Renew Sustain Energy Rev
48:472–482
7. Kim K (2011) Use of scada data for failure detection in wind turbines. In: ASME 2011 5th
International conference on energy sustainability, pp 2071–2079
8. Kusiak A, Li W (2011) The prediction and diagnosis of wind turbine faults. Renew Energy
36(1):16–23
9. Márquez FPG, Garcı́a-Pardo IP (2010) Principal component analysis applied to filtered sig-
nals for maintenance management. Qual Reliab Eng Int 26(6):523–527
10. Márquez FPG, Muñoz JMC (2012) A pattern recognition and data analysis method for main-
tenance management. Int J Syst Sci 43(6):1014–1028
11. Márquez FPG, Pérez JMP, et al. (2016) Identification of critical components of wind turbines
using fta over the time. Renew Energy 87
12. Marugán AP, Márquez FPG, Pérez JMP (2016) Optimal maintenance management of off-
shore wind farms. Energies 9(1):46
13. Masters T (1993) Practical neural network recipes in C++. Academic Press
14. Muñoz CQG, Márquez FPG, Tomás JMS (2016) Ice detection using thermal infrared radiom-
etry on wind turbine blades. Measurement 93:157–163
15. Munoz CQG, Arenas JRT, Marquez FPG (2014) A novel approach to fault detection and
diagnosis on wind turbines. Global Int J 16(6):1029–1037
16. Mylaraswamy D, Olson L, Nwadiogbu E (2007) Engine performance trending. In: AIAC12-
HUMS conference
17. Song Z, Jiang Y, Zhang Z (2014) Short-term wind speed forecasting with markov-switching
model. Appl Energy 130(5):103–112
18. Sun P, Li J et al (2016) A generalized model for wind turbine anomaly identification based
on scada data. Appl Energy 168:550–567
19. Wymore ML, Dam JEV et al (2015) A survey of health monitoring systems for wind turbines.
20. Yang W, Tavner PJ et al (2010) Cost-effective condition monitoring for wind turbines. IEEE
Trans Industr Electron 57(1):263–271
21. Yang W, Court R, Jiang J (2013) Wind turbine condition monitoring by the approach of scada
data analysis. Renew Energy 53(9):365–376
22. Zang X, Sun H, Trivedi KS (1998) A bdd-based algorithm for reliability graph analysis.
Department of Electrical Engineering
Advances in Engineering Management
of the Eleventh ICMSEM
Advances in Green Supply Chain, Resource
Optimization Management, Risk Control and Integrated
Project Management Based on the Eleventh
ICMSEM Proceedings
Jiuping Xu(B)

xujiuping@scu.edu.cn
Abstract. Management Science and Engineering Management (MSEM) have

significantly contributed to developments in society and economy, especially in
management and control processes. In this paper, we first describe the basic con-
cepts covered in the eleventh ICMSEM proceedings Volume II. We then conduct
a review of engineering management (EM) research to identify the key areas,
from which green supply chain, resource optimization management, risk control
and integrated project management were found to be the most widely discussed
research areas. After a summary of the key research achievements in the four
areas, the high frequency EM keywords in the proceedings volume II are identi-
fied using NodeXL. The research trends both from MSEM journals and the ICM-
SEM are then summarized using the CiteSpace tool. As always, we are committed
to providing an international forum for academic exchange and communication
through the ICMSEM and plan to continue our innovative MSEM progress in the
future.
Keywords: Green supply chain · Resource optimization management · Risk

control · Integrated project management
1 Introduction
Management Science and Engineering Management, which emphasizes the theories,

methods and engineering practice of complex management decision making, has been
playing a vital role in both scientific endeavor and societal development. The increase
in MSEM research in the past few decades has brought new vitality to management and
engineering practice. At the same time, through the continuing and expanding develop-
ment of innovative managerial tools, MSEM has significantly contributed to advance-
ments in domestic and international economic development and increased general sci-
entific management consciousness. MSEM is a complex synthesis with a broad research
focus that combines complex management theories and practical engineering solutions
to successfully solve management practice problems. This kind of cross-functional,
924 J. Xu
multidisciplinary research supports real world management execution, improves man-

agement efficiency, and contributes to energy conservation.
MSEM research is widely applied to management problems that involve extensive
engineering concepts. In proceedings volume I, the focus is on management science
(MS) and its future development trends. Engineering management (EM) is therefore the
main focus of proceedings volume II. Kocaolgu’s defined EM as a field that examines
engineering management using comprehensive scientific research to develop innovative
engineering solutions that can improve technical organization, technical resources, and
technical systems [5], in areas such as manufacturing, construction, design engineering,
and industrial engineering. The ICMSEM proceedings Volume II focuses on four key
EM areas; the green supply chain, resource optimization management, risk control, and
integrated project management.
2 Literature Review
To better analyze the pertinent research fields and possible research directions, we
reviewed the most popular research areas in the most recent EM research. What
emerged was that the green supply chain, resource optimization management, risk con-
trol, and integrated project management have been the most widely studied in recent
years. In this section, we review the related literature to analyze the developmental
tracks in these four areas.
2.1 Green Supply Chain

The supply chain management concept emerged in the early 1980s, with the green
supply chain (GSC) being first proposed by Michigan State University Manufactur-
ing Research Association in 1996. Lambert et al. defined supply chain management as
the integration of key business processes from the end user to the original suppliers to
provide products, services, and information to add value for customers and other stake-
holders [8]. Brazil Sarkis examined the social and cultural dimensions in green supply
chain management and found that cultural boundaries govern the type of management
skills and control processes used [7]. To empirically investigate the construction and
scale of green supply chain management (GSCM) implementation by manufacturers,
Zhu et al. tested two GSCM implementation measurement models and compared them
using confirmatory factor analysis [19]. As a supply chain grows in scale and operation,
its structure becomes more complicated. Some scholars have taken a fresh look at inte-
grated green supply chain management [4, 13, 14] to ensure continuing sustainability.
2.2 Resource Optimization Management

Resource optimization management (ROM) is the efficient and effective deploy-
ment and allocation of an organization’s resources when and where they are needed.
Such resources may include financial resources, inventory, human skills, production
resources, or information technology. Resources and energy shortages have become
the main factors restricting the sustainable development of national economies, as
Advancement of GSC, ROM, RC and IPM Based on the 11th ICMSEM Proceedings 925
low resource efficiency and high energy consumption can result in a serious waste
of resources putting significant pressure on resources and environmental governance.
Vadenbo presented a general multi-objective mixed-integer linear programming (MILP)
optimization model aimed at providing decision support for waste and resource man-
agement in industrial networks [15]. Further, energy resource shortage problems associ-
ated with rapid social and economic development have been of critical concern to both
national and local governments worldwide for many decades [9]. Water, as one of the
most important resources on earth has therefore received a great deal of research atten-
tion, with water optimization management being a major focus in the past few years
[1, 12, 18]. ROM involves all aspects of social life, so research strives to further develop
this area using modern computer technology.
2.3 Risk Control
Risk control (RC) is the coordinated and economic application of resources to mini-
mize, monitor and control the impact probability of unfortunate events, or to maximize
the realization of opportunities after the identification, assessment and prioritization of
risks. Muriana and Vizzinib presented a deterministic technique for assessing and pre-
venting project risks by determining the risk of the Work Progress Status [10]. Valtonen
et al. studied public risk management related to the use of public land development
by analyzing case studies in Finland and the Netherlands, both of which have strong
public land development traditions [16]. By identifying risks, specific state support and
special project management measures have been developed to limit the negative influ-
ence of the possible project risks [17]. Therefore, developing a general framework to
analyze corporate risk management policies and ensure risk control is vital [6]. In brief,
RC includes transferring risk to another party by avoiding risk, reducing the negative
effects of risk, and accepting some or all of the consequences of a particular risk.
2.4 Integrated Project Management
Integrated project management (IPM) is a philosophy that recognizes the different ele-
ments involved in projects to apply strong team leadership and encourage a collabora-
tive ethos with clear purposes and strategies to ensure success. Planning, organizing,
securing, and managing resources to successfully complete specific project goals and
objectives are the main IPM research areas, the applications for which have been used
in manufacturing, construction, design engineering, industrial engineering, technology,
production and many other areas. Atkinson provided some thoughts about the success
criteria for project management in which cost, time, and quality have become inextri-
cably linked to project management success over the last 60 years [2]. An integrated
methodology was developed for planning construction projects under uncertainty that
relied on a computer supported risk management system to identify the risk factors in
the integrated project [11]. IPM has also been applied to integrated waste management
systems to identify the optimal breakdown between materials and energy recovery from
municipal solid waste [3]. In all, integrated project management is a complex subject
and needs to be examined from several perspectives.
926 J. Xu
3 Central Issues in the Proceedings Volume II

Keywords are an important index for the identification of essential content in research
articles. The keywords in the 75 articles in Volume II were extracted to a database to
allow for a tendency analysis. The keywords were sorted using the Keywords Match
software that matched the keywords that were uniform or similar. Finally, the matched
keywords were imported into NodeXL software to obtain the initial visual image shown
in Fig. 1.
Fig. 1. DAS literature mining
With EM as the key component, the open source software tool NodeXL was used to
facilitate the learning of the concepts and methods. To begin with, some key words were
processed to unify expressions and keywords with the same meaning. This preliminary
process reduced the number of keywords, making it possible to develop an efficient
network. At the same time, the vertices’ shapes were set to determine the betweenness
and closeness centrality. When the degree of a vertex’s betweenness and closeness cen-
trality was more than five, the shape of this vertex was a red diamond in the resulting
figure. The aim of this analysis was to determine the key EM concepts that connected
with other research topics through the primary nodes. After calculating, clustering and
filtrating, the results in Fig. 2 illustrate the important research relationships.
The above analysis highlighted the key areas focused on in the proceedings vol-
ume II; green supply chain management, resource optimization management, risk con-
trol and integrated project management problems related to green and pro-environment
concepts. Tao focuses on a simplified manufacturing and distribution supply chain plan-
ning network focused on carbon emission constraints, for which a bilevel programming
model was mathematically formulated. Through the study of related emissions poli-
cies, Lu finds that under different carbon policies, revenue sharing contracts are able to
coordinate the supply chain with the manufacturer’s revenue sharing proportion under
Fig. 2. Research topics in EM for the Eleventh ICMSEM
a cap-and-trade being always less than a situation in which there is no carbon emis-
sions constraint. Machado and Duarte present an overview of Industry 4.0, a new devel-
oping concept in the field of industrial engineering, and lean and green supply chain
management, for which a conceptual model is developed. Wang et al. concludes that
when big data suppliers are part of the supply chain competition and one party gains a
dominant supply chain position, collaborative decision-making is the key to enhancing
overall supply chain profitability. These studies highlight the many new characteristics
involved in green supply chains, such as big data and carbon emissions reductions.
Resource optimization management primarily covers resource processing, exploita-
tion, production, and consumption. Arifjanov and Zakhidov consider an issue for the
development of methods for the estimation of electrical loads for various electricity cus-
tomers in residential and public buildings, and address issues related to the integration
of distributed generators to the UES to ensure the optimal management of technologi-
cal modes. Gómez Munoz et al. develops an optimal maintenance strategy that includes
the NDT system to reduce costs and increase the competitiveness of renewable energy
sources. For water resource optimization, Tu et al. proposes an equity and sustainability
based model that considers equity, economic benefit, and ecological environment and
Liu et al. uses a DEA method to establish an evaluation model to empirically study the
urban agglomeration of Chengdu city from 2008–2014. Because of the wide use and
serious shortage of some resources around the world, scientific research into resource
optimization management has increased to assist governments and enterprises around
the world.
Risk control is important in all organizations and enterprises. In this section, Zhou
et al. presents a model to mitigate the risk of budget overruns during raw materials pro-
curement, and Chen et al. provides a risk allocation model using a fuzzy comprehen-
sive evaluation method aimed at optimizing risk allocations, reducing transaction costs,
and enhancing comprehensive benefits from the perspective of state-owned enterprises.
Beside these studies, to better control risks occurring in financial markets, Shibli et
al. investigates the link between the foreign exchange markets and the stock market in
928 J. Xu
Bangladesh by examining the volatility spillover between the markets, volatility persis-
tence, and the asymmetric effect of information on the volatility of these two financial
markets. In addition, Zhang et al. studies a projection pursuit risk assessment model
using a combined method to model PPP risk under the background of Big Data. Effec-
tive risk control can reduce potential risk factors and assist managers gain increased
benefits.
The last section in Volume II focuses on the developments in integrated project
management. Elchan defines the scientific problems associated with management at the
beginning stages of an innovation project in the departments of a technology park at a
higher education school in Azerbaijan. Based on principal-agent theory and game the-
ory, a theoretical framework for a knowledge network conflict coordination mechanism
is constructed by Wei, which divides the conflict coordination mechanism into three
levels; contract mechanism, self-implementation mechanism, and a third-party conflict
coordination mechanism. Sheng et al. proposes suggestions and countermeasures for
the equalization of public services in urban and rural areas from three integrated project
aspects and Tu et al. develops two staged fuzzy DEA models with undesirable outputs to
evaluate the banking system. The overall process of leading, organizing, staffing, plan-
ning, and controlling activities requires managers to use systems viewpoints, methods,
and theories to optimize the work involved when seeking effective integrated project
management under limited resource constraints.
4 Evaluation of EM and ICMSEM Development Trends

In this section, we evaluate the ICMSEM to validate that the research trends are in line
with the trends emerging in current MSEM journals. At first, CiteSpace, an information
visualization technology developed by Chen, was utilized to draw the a scientific knowl-
edge map that display the development trends in a discipline or knowledge domain over
a certain period, and identifies the possible evolution through an analysis of the research
frontiers.
“Advanced search” with “engineering management” was used as the identifier in
the “web of science” database with a timespan from 2000 to 2017. Initially, 42499
published articles were identified and after the final screening using TI = (engineering
management) 2963 articles were reviewed.
4.1 The Development Trends of EM
The statistics from the 2963 articles recorded output were saved and converted into
CiteSpace which transformed the data into a format that could be identified by the soft-
ware to allow for parameter selection. In this operation, the time span was from 2000
to 2017 with the time slice set at one year and the theme selection based on the titles,
abstract subject words, identifiers, and keywords to allow for node selection. Then, each
zone with the 30 highest keyword records were clustered and analyzed, from which a
map was drawn for the minimum spanning tree. As shown in Fig. 3, by setting the
“Threshold = 30”, a total of 348 nodes were obtained, with the overall network density
being 0.0098. Therefore, the system frequency identified green supply chain manage-
ment, environmental management, project management, risk management, and mod-
els and systems as the highest ranked areas. This not only displayed the most popular
research fields in engineering management but also implied the future EM development
trends.
Fig. 3. The results of co-occurring keywords in EM
With a reference frequency running from high to low, the top thirty were analyzed
from the 348 keywords. As shown in Table 1, the keywords such as green supply chain
management, risk management, environmental performance and sustainability had a
relatively high centrality.
Using the keyword with the label title clustering, 40 categories were identified; how-
ever, only the top 10 categories are shown in Fig. 4. The other topics were relatively
dispersed so are not displayed, which indicated that scientific engineering management
research was relatively loose. In particular, the new keywords for 2015 and 2016 have
not resulted in significant research attention.
930 J. Xu
Table 1. The top thirty hot keywords of EM

84 0.11 2010 Performance
77 0.03 2011 Green supply chain management
54 0.07 2002 Model
54 0.08 2002 Risk management
52 0.27 2007 System
48 0.21 2011 Environmental management
41 0.15 2007 Implementation
40 0.1 2012 Sustainability
39 0.05 2007 Project management
38 0.12 2011 Impact
36 0.11 2011 Industry
32 0.05 2011 Strategy
29 0.08 2010 Perspective
29 0.12 2010 China
25 0.04 2011 Green supply chain
24 0.05 2010 Framework
21 0.06 2011 Innovation
21 0.04 2009 Integration
20 0.05 2008 Firm
20 0.05 2011 Reverse logistics
20 0.05 2011 Design
17 0.01 2012 Capability
17 0.02 2013 Initiative
16 0.21 2004 Risk
16 0.05 2011 Pressure
16 0.01 2010 Environmental performance
15 0.1 2008 Quality
15 0.01 2013 Selection
15 0.06 2007 Uncertainty
15 0.05 2008 Optimization
Fig. 4. The results of research fields clustering of EM
4.2 Future Development Predictions
To further understand the focused engineering management research areas, we exam-

ined each category to determine the common themes. At “Threshold = 30”, the high fre-
Fig. 5. The category clustering of EM timeline view

932 J. Xu
Table 2. The results of top thirty categories

307 0.21 2002 Engineering
185 0.01 2004 Business & Economics
176 0.01 2002 Computer Science
104 0.05 2002 Operations Research & Management Science
102 0.02 2002 Environmental Sciences & Ecology
95 0.22 2002 Environmental Sciences
73 0.12 2002 Computer Science, Information Systems
71 0.07 2002 Engineering, Electrical & Electronic
70 0.02 2005 Engineering, Industrial
68 0.13 2004 Business
60 0.02 2007 Engineering, Civil
51 0.17 2006 Economics
50 0.02 2007 Computer Science, Theory & Methods
50 0.1 2005 Computer Science, Interdisciplinary Applications
50 0.09 2002 Engineering, Environmental
47 0.23 2004 Water Resources
42 0.05 2006 Business, Finance
41 0.12 2010 Science & Technology - Other Topics
39 0 2002 Cardiovascular System & Cardiology
39 0.04 2007 Engineering, Manufacturing
38 0.06 2006 Computer Science, Artificial Intelligence
32 0 2013 Green & Sustainable Science & Technology
28 0.11 2005 Construction & Building Technology
28 0.04 2002 Telecommunications
27 0 2004 General & Internal Medicine
26 0.08 2004 Medicine, General & Internal
25 0 2002 Cardiac & Cardiovascular Systems
24 0.04 2010 Materials Science
23 0 2003 Peripheral Vascular Disease
quency research areas and timezone view diagram were identified, as shown in Fig. 5.
Most of these high frequency words appeared in the early days, indicating that engineer-
ing management research has been focused around these topics for quite a long time.
The analysis of the top 30 most frequent words (Table 2), found that engineering, busi-
ness and economics, computer science, environmental science and operations research
are currently the most popular research areas.
From Tables 1 and 2, the papers presented in this year’s ICMSEM proceedings
volume II closely reflect the most pertinent engineering management research areas;
green supply chain (systems management, industry, green and environmental protec-
tion), resource optimization management (environmental sciences, water resources,
engineering education), integrated project management (project management, indus-
try, integrated project) and risk control (risk management, risk, design, information).
In addition, the computer science and environmental science research highlights how
high-tech can spur social progress and environmental protection.
We believe that EM should focus on the study of specific EM problems as well
as popularizing MS knowledge. Excellent academic research can effect developments
across the world, but a further focus on regional areas is also needed. To ensure a bright
EM future, there is a need for inspiration, practical theories, effective methods, and
extensive applications in future developments. To achieve this, EM knowledge needs to
be popularized, which is the duty of all MSEM academics. In the future, more focus on
low carbon emissions, environmental protection, big data, energy utilization, and other
popular EM issues are needed.
5 Conclusion
Engineering management is a complex area that involves all engineering aspects. The
open source software tool NodeXL identified the four areas covered in the eleventh
ICMSEM proceedings Volume II, from which we summarized the key research in this
year. To analyze the EM and ICMSEM development trends, we identified the main
search terms and keywords using CiteSpace. Our key objective is to continue to improve
the quality of papers in the proceedings and to ensure the ICMSEM organization is
dynamic and appealing to EM researchers worldwide. EM research is continuously
developing and new trends are appearing every year; however, more research is neces-
sary so as to popularize EM developments and provide a more active research forum.
Acknowledgements. The author gratefully acknowledges Jingqi Dai and Lin Zhong’s efforts on
the paper collection and classification, Zongmin Li and Lurong Fan’s efforts on data collation
and analysis, and Ning Ma and Yan Wang’s efforts on the chart drawing.
References
1. Afshar A, Massoumi F et al (2015) State of the art review of ant colony optimization appli-
cations in water resource management. Water Resour Manage 29(11):3891–3904
2. Atkinson R (1999) Project management: cost, time and quality, two best guesses and a phe-
nomenon, its time to accept other success criteria. Int J Project Manage 17(6):337–342
3. Consonni S, Giugliano M et al (2011) Material and energy recovery in integrated waste
management systems: project overview and main results. Waste Manage 31(9–10):2057–
2065
4. Fahimnia B, Sarkis J, Davarzani H (2015) Green supply chain management: a review and
bibliometric analysis. Int J Prod Econ 162:101–114
5. Farr JV, Buede DM (2003) Systems engineering and engineering management: keys to the
efficient development of products and services. Eng Manage J 15(3):3–9
934 J. Xu
6. Froot KA, Scharfstein DS, Stein JC (1993) Risk management: coordinating corporate invest-
ment and financing policies. J Financ 48(5):1629–1658
7. Govindan K, Sarkis J et al (2014) Eco-efficiency based green supply chain management:
current status and opportunities. Eur J Oper Res 233(2):293–298
8. Lambert DM, Cooper MC (2000) Issues in supply chain management. Ind Mark Manage
29(1):65–83
9. Ming Z, Song X et al (2013) New energy bases and sustainable development in China: a
review. Renew Sustain Energ Rev 20(4):169–185
10. Muriana C, Vizzini G (2017) Project risk management: a deterministic quantitative technique
for assessment and mitigation. Int J Project Manage 35:320–340
11. Schatteman D, Herroelen W et al (2008) Methodology for integrated risk management and
proactive scheduling of construction projects. J Constr Eng Manage 134(11):885–893
12. Singh A (2014) Simulation and optimization modeling for the management of groundwater
resources. I: distinct applications. J Irrig Drainage Eng 140(4):04013,021
13. Singh A, Trivedi A (2016) Sustainable green supply chain management: trends and current
practices. Competitiveness Rev 26(3):265–288
14. Srivastava SK (2007) Green supply-chain management: a state-of-the-art literature review.
Int J Manage Rev 9(1):53–80
15. Vadenbo C, Guilléngosálbez G et al (2014) Multi-objective optimization of waste and
resource management in industrial networks - part ii: model application to the treatment
of sewage sludge. Resour Conserv Recycl 89(4):52–63
16. Valtonen E, Falkenbach H, Krabben EVD (2017) Risk management in public land devel-
opment projects: comparative case study in Finland, and The Netherlands. Int J Covering
Aspects Land Use, Land Use Policy 62:246–257
17. Wang HB (2007) Application of risk management to wind power project. Sci Technol Man-
age 43:48–51
18. Wang R, Li Y, Tan Q (2015) A review of inexact optimization modeling and its application
to integrated water resources management. Front Earth Sci 9(1):51–64
19. Zhu Q, Sarkis J, Lai KH (2008) Confirmation of a measurement model for green supply
chain management practices implementation. Int J Prod Econ 111(2):261–273
Green Supply Chain
Modelling a Supply Chain Network of Processed
Seafood to Meet Diverse Demands
by Multi-branch Production System
Koichi Murata1(B) , Seiichiro Isobe1 , and Hiroshi Katayama2

1
College of Industrial Technology, Nihon University, Tokyo, Japan
murata.kouichi30@nihon-u.ac.jp
2
Abstract. The demand for processed seafood has recently increased

owing to diversity of taste and the crises of food supplies. Processed
seafood companies have the potential to manufacture various products
by combining fish varieties and processing methods. They also associate
with many customers such as other food processing companies, food ser-
vice industries, FMCG companies, and end users. This study aims to
reveal the supply chain model of processed seafood to multi-customers
by the multi-branch production system. A model is constructed to under-
stand the practical system from a theoretical perspective. The main
finding of the study is that four paths of the proposed network model
between a seafood processing company and customers exist: (1) defrost-
ing to processing, (2) cleaning to service, (3) processing to service, and
(4) seasoning to retail. These findings clarify the supply chain network
in the industry and help consider business opportunities.
Keywords: Food supply chain management · Seafood processing indus-

try · Model approach
1 Introduction
Fish and seafood have recently become important materials to meet diversified
customer needs resulting from the food crisis caused by population explosion and
the requirement of seafood which is essentially healthy (i.e. low calorie count
and nutrient intakes) and which provides convenient meals for a two-income
family, a child-rearing family, and aging society in advanced counties. To match
the resultant explosive demand, the globalization of the seafood supply chain
has increasingly progressed with technological innovation. For example, the cold
chain has expanded to maintain the freshness of materials through improvement
of freezing technologies for transport worldwide [11]. A traceability system has
been developed to obtain the trust of customers by informing them of the safety
of materials by a physical sensor system and information technologies [6].

938 K. Murata et al.
In the Japanese context, to aim for the creation of a tourism nation and for
the next summer Olympic Games in 2020, an enrichment of the food supply
chain is required. The Japanese dietary culture called ‘Washoku’, registered as
a 2013 Intangible Cultural Heritage by UNESCO, must be a powerful weapon
for executing the strategy. Processed seafood, one core ingredient of ‘Washoku’,
is manufactured in small and medium enterprises (SMEs) mostly located in
the area that experienced the Great East Japan earthquake (Higashi Nihon
Daishinsai) on 11 March 2015. Furthermore, the occurrence of climate change
including global warming and abnormal weather and over-fishing by neighbour-
ing countries causes a decrease of marine resources in the Pacific Ocean. These
SMEs operate on a small scale and sell processed seafood to identified customers
according to their requirements. For the future, they must overcome difficul-
ties and reinvent themselves to obtain new markets and customers worldwide
through business innovation.
Many studies related to the seafood supply chain have presented new per-
spectives, including open innovation of seafood value chain [12], an interna-
tional distribution system [1,2], firm structure [4], quality assurance with labelled
seafood products [7], a sustainable system [3], marketing and economic innova-
tion [5,9,14], seafood supply chain management [10], and an inventory system
[8]. However, there are not enough studies related to innovation of the business
model which give an overall perspective to change the present business structure.
Based on the recognition of current practical and academic conditions, the
present study aims to explore the business model of the processed seafood indus-
try with the concept level as the starting point of the study. Specifically, the
network between a food processing company and its customers is focused on
understanding the exciting business style of the food processing company and
grasping the possibility of cultivating routes to new customers.
2 Methodology
This study tries to draw the network structure between one food processing com-
pany and its candidate customers because a clarification of the relation among
participants firstly needs to retrieve investment points for business expansion.
The description is performed based on an observation and an interview of sev-
eral seafood processing companies in the rebuilding project after the Great East
Japan earthquake for three years. The authors of the present study have grad-
ually understood these companies business structure through investigation and
a trial-and-error method. This study presents the results. The method to build
the described model is business modelling which reveals the following things
pursuant to the aim of the study [15]:
The business model is emerging as a new unit of analysis.
• Business models emphasize a system-level, holistic approach to explain how

firms ‘do business’.
• Firm activities play an important role in the various conceptualizations.
Modelling a Supply Chain Network of Processed Seafood 939
• Business models seek to explain how value is created, not just how it is cap-
tured.
This research also reference typical models related to a production system,

such as make-to-order (MTO) and make-to-stock (MTS) models [13].
Four steps are adopted in this study. The first is based on the seafood cook-
ing process. Seafood processing companies cook materials at the location of mass
customers. The step will be useful to identify fundamental processes in a produc-
tion system for seafood processing. The second is the deployment of candidates
of a customer. They are other food processing companies, food service compa-
nies, fast-moving consumer goods (FMCG) companies, and consumers. This step
describes the details of the four customers because the purpose of the step is to
become aware of new customers that differ from existing customers. In the third
step, a network model is constructed to visualize a possibility of new business
through a combination of outputs in the first and second steps. The final step
is a confirmation of the capability of the constructed network through an ini-
tial discussion of business expansion in the future. The flowchart of all steps is
provided in Fig. 1.
Start
Step1: Analysis of seafood cooking process
Step 2: Identification of customer of seafood

processing companies
Step3: Construction of network between one

seafood processing company and seafood
service industry
Step 4: Discussion of future strategy of

seafood processing industry based on the
constructed network
End
Fig. 1. Research procedure.
3 Construction of Seafood Supply Chain Model
This section describes the result of each step illustrated in Fig. 1.

3.1 Seafood Cooking Process (Step 1)
The journey to develop a processed seafood business model in this study starts
with the understanding of daily eating habits as follows. The main processes to
eat seafood have not changed from ages past, as shown in Fig. 2: fishing, cleaning,
processing, seasoning, and eating. The middle three processes exist mainly to add
value to fresh fish. These can be omitted and exchanged in the cooking process.
Many techniques and skills are included to make the materials taste even better.
The detailed explanation of each type is as follows:
(1) Fishing process
This process is considered a preparation part for processed seafood produc-
tion involving obtaining the materials. The quality level of the materials depends
on size, weight, appearance, freshness, variety, and so forth. Over-fishing and cli-
mate change seriously affect a haul of fish. A fishing quota by multiple nations
and enclosed aquaculture with the newest biotechnologies are countermeasures
for sustainable marine resources.
(2) Cleaning process
The complicated body of fish and seafood causes a decline in the productivity
of seafood processing. It is basic and important to protect customers’ safety from
dangerous parts which they cannot eat (i.e. hard fish, thick bone, and the internal
organs of a fish including poisonous substances). The variety of processes includes
carving, boning, scaling, and filleting. Accurate fileting of large fish like salmon
requires immense skill with yield rate and waste disposal.
(3) Processing process
This process is the main portion of seafood processing. Representative meth-
ods of the process are grilling, reducing, frying, steaming, and drying. The selec-
tion of each method relates to a core business of a seafood processing company
with a large investment in production facilities. The design of the process pro-
file of production by the selected method is a company secret relating to the
organoleptic feel of processed seafood because many companies adopt the com-
mon categories of methods noted above.
(4) Seasoning process
This process is the portion that adds tastiness and flavour-value to a
processed item. In Japan, there are five traditional popular seasonings, such
as sugar (Sato), salt (Shio), vinegar (Su), soy sauce (Shoyu), and soybean paste
(Miso). They characterize the taste of Japanese seafood called ‘Washoku’, which
was registered as a 2013 Intangible Cultural Heritage by UNESCO. Chemical
seasoning widens the possibility of new ways to enjoy processed seafood.
(5) Eating process
This is the final process to synthetically evaluate the quality of the four
processes by consumers. In this stage, not only the taste of processed seafood,
but also a dishing, including a selected plate and a side dish, influences the total
impression. The family styles of advanced countries, such as a two-income family,
a child-rearing family, and the aging society, also require the convenience of a
microwave oven and reheat pouches.
Fishing Cleaning Processing Seasoning Eating
Fig. 2. Basic process to eat seafood
In the present age, the division of the five processes between fisheries and
consumers depends on consumers’ requirements. It indicates the following five
models proposed in Fig. 3.
Model 1 Consumers
Model 2 Fisheries Consumers
Fishing Cleaning Processing Seasoning Eating
Fig. 3. Division of roles between fisheries and consumers
Consumers do not need support from fisheries in Model 1, which constitutes

a hobby. In the case of Model 2, the fishing industry will exist in a market near
a fishing port with consumers. In the three models consisting of two processes
and more by fisheries (i.e. Models 3 to 5), the division of roles occurs among
three industries: (I) fishing industry, (II) seafood processing industry, and (III)
seafood service industry including restaurant and retail businesses. The division
is considered based on the two following assumptions:
• The fishing industry is necessarily in charge of the fishing process.

• The orders of the three industries do not become reversed.
Nineteen sub-models of the three models are summarized in Table 1. Mod-

els 3, 4, and 5 contain 3, 6, and 10 sub-models. Processed seafood is sold directly
to consumers or through a mail-order system and internet service when the sub-
model consists of (I) the fishing industry and (III) the seafood service industry.
(II) The seafood processing industry asks (III) the seafood service industry to
sell their products in addition to the above-mentioned selling methods when
these two entities exist in the sub-model. This table shows that (II) the seafood
processing industry can choose, in theory, appropriate sub-models from the alter-
natives. The flexibility seems to suggest the possibility of expanding the business
of each seafood processing company.
Table 1. Sub-models of Models 3 to 5
Model Fishing Cleaning Processing Seasoning Note

3 I I - - Selling directly to customers
I II - - Selling by a service industry
or directly to customers
4 I I I - Selling directly to customers
I I II - Selling by a service industry
I I III - Selling directly to customers
I II II - Selling by a service industry
I II III - Selling directly to customers
I III III - Selling directly to customers
5 I I I I Selling directly to customers
I I I II Selling by a service industry
I I I III Selling directly to customers
I I II II Selling by a service industry
I I II III Selling directly to customers
I I III III Selling directly to customers
I II II II Selling by a service industry
I II II III Selling directly to customers
I II III III Selling directly to customers
I III III III Selling directly to customers
(I) fishing industry, (II) seafood processing industry, and (III) seafood service industry.
3.2 Customers of Processed Seafood Industry (Step 2)
Customer candidates for (II) the processed seafood industry are rich in vari-
ety. The main candidates are four types of seafood service companies: other
food processing companies, food service companies, fast-moving consumer goods
(FMCG) companies, and consumers. The explanations of each type are as
follows:
(1) Other food processing companies
These customers use the product supplied from seafood processing companies
as one material of their final products. For example, raw materials are supplied
from a company which manages a large-capacity frozen warehouse to store them
cheaply and in huge quantities. Cleaned materials (i.e. seafood paste and fish cut
into small cubes) also have high value for customers who do not have technical
know-how.
(2) Food service companies

These customers are food service companies including restaurant businesses;
home-delivery services; services for the provision of meals by a school, hospital, or
welfare facilities. They basically have a kitchen which has the capacity for making
and serving meals with materials supplied from seafood processing companies
to their customers. Moreover, if their business model is a franchise chain, they
construct a supply network with a central factory which executes the middle
processing of materials for speedy and effective servicing of meals at all their
shops.
(3) FMCG companies
These customers are types of FMCG companies (e.g. a convenience store,
supermarket, shopping mall, or department store). They mainly offer two ser-
vices. One is the direct sale of goods supplied from seafood processing companies.
The other is the sale of a daily dish including simple processed/cooked materials.
Their customers are end users who live near these small shops. FMCG compa-
nies obtain large sales accumulated across their many branches, but the number
of end users is limited for one store.
(4) Consumers
Seafood processing companies usually do not directly access end users except
in face-to-face sales like at a regional festival and special sales once a year. They
contact end users to sell their products through a mail-order system with mass-
communication and a system of online electronic commerce. This opportunity
will increase in the era of the internet.
3.3 Network Between Seafood Processing Company and Seafood

Service Industry (Step 3)
Figure 4 illustrates the relationship between one seafood processing company
and seafood service companies. The left side of the figure is a typical produc-
tion system in a seafood processing company. This system mainly consists of
four processes: defrosting, cleaning, processing, and seasoning. The first one is a
process to prepare frozen materials and deliver them to maintain their freshness
while restraining freezer burn if possible, by a fishing company, and to maintain
a sufficient inventory. The other three processes add appropriate values to the
defrosted materials described in the previous section. The notable characteristic
of the system is a multi-branch production system with the potential to supply
semi-processed products as final products to customers.
Owing to the capability of the industry, candidate customers are rich in
variety, as shown in the right side of the figure. Main candidates are the four
types of seafood service companies mentioned above (i.e. other food processing
companies, food service companies, FMCG companies, and consumers). While
centring on seafood processing companies, they should consider all combinations
between processes of seafood processing and four types of seafood service compa-
nies. In particular, the following four paths are useful for food service companies
to realize their service while reducing initial workloads.
A seafood processing company Seafood service companies
Other food
processing
Defrosting companies
Food
Cleaning
service
companies
Processing
FMCG
companies
Seasoning
Consumers
Fig. 4. Relationship between a seafood processing company and seafood service com-
panies.
(1) A path from defrosting to processing (DTP)

In this path, seafood processing companies supply frozen/defrosted materials
to other food processing companies. The added-value is a supply system of mate-
rials sufficient both in quality and quantity. Seafood processing companies must
improve the refrigeration capacities of inventory and transportation. Owing to
these functions, other food processing companies outsource procurement opera-
tions.
(2) A path from cleaning to service (CTS)
In this path, seafood processing companies supply frozen materials to seafood
service companies. The added-value is a system to clean masses of fish and
seafood. The seafood processing company must improve its capability through
training of skills related to the operation and its automation. Owing to this
function, food service companies will concentrate on creative cooking and close
servicing, which are considered to be their core functions.
(3) A path of processing to service (PTS)
In this path, seafood processing companies supply processed materials to
seafood service companies. The added-value is a mass production system of
processed seafood. Seafood processing companies should improve the main
process in their business. Owing to this function, food service companies will
provide effective and uniform services to their customers with simple cooking.
(4) A path of seasoning to retail model (STR)
In this path, seafood processing companies supply seasoned materials to
FMCG companies. The added-value is a mass supply system of processed
seafood. Seafood processing companies should improve the main processes in
their business. Owing to this function, retail companies can sell various moder-
ately priced and tasty items to end users every day.
3.4 Future Strategy for Seafood Processing Company (Step 4)

To confirm the utility of the constructed network model in the previous sections,
a future strategy of seafood processing companies will be considered in this
section.
Most SMEs continue to do business along one path of the network model
described in the proposed network in Fig. 4. They may consider the path as the
only possibility for their business with limited management resources. However,
they should understand two characteristics of their industry.
One is the flexibility of their production system. It is possible to supply
products from any process in their production system. For example, if the current
path of the network model is STR, semi-products of the path become fully final
products for the other three paths (i.e. DTP, CTS, and PTS) without large
facility investment. The other is that they have many candidate customers. If the
current path of the network model is CTS, the same product could be supplied by
PTS. The development of a new path is naturally difficult, even if the potential
of the path is recognized. However, it will be a valuable trial to expand business
in the future.
The utility of the proposed network is confirmed because it is useful to con-
sider a subject to realize new business possibilities with the characteristic pro-
duction system based on the above discussion.
This study aims to reveal the supply chain model of processed seafood to mul-
tiple customers via a multi-branch production system. The findings consist of
four paths: (1) DTP, (2) CTS, (3) PTS, and (4) STR. They clarify the supply
chain network in the industry and consider the opportunity of the business. A
future study could provide a mathematical formulation of the network model to
quantitatively simulate and evaluate the impact of each path.
References
1. Abrahamsen MH, Håkansson H, Naudè P (2007) Perceptions on change in business
networks: Norwegian salmon exporters and japanese importers. In: 23rd Annual
IMP Conference, Manchester
2. Abrahamsen MH, Naudé P, Henneberg SC (2011) Network change as a battle of
ideas? analysing the interplay between idea structures and activated structures.
Imp Group 5(2):122–139
3. Alden R (2011) Building a sustainable seafood system for maine. Maine Policy Rev
20(1):87–95
4. Brydon K, Dana LP (2011) Globalisation and firm structure: comparing a family-
business and a corporate block holder in the new zealand seafood industry. Int J
Globalisation Small Bus 4(2):206–220
5. Chang KS, Waters EC (2006) The role of the alaska seafood industry: a social
accounting matrix (SAM) model approach to economic base analysis. Ann Reg Sci
40(2):335–350
6. Farooq U, Tao W et al (2016) epedigree traceability system for the agricultural

food supply chain to ensure consumer health. Sustainability 8(9):839
7. Jaffry S, Pickering H et al (2004) Consumer choices for quality and sustainability
labelled seafood products in the UK. Food Policy 29(3):215–228
8. Murata K, Watanabe N et al (2015) Diagnosis of production system of marine
frozen products by inventory management theory - a case of blue fins. Int J Intell
Inf Syst 4(2–1):18–24
9. Quagrainie KK (2006) Analysis of US catfish fillet market share using a flexible
logistic model. Marine Res Econ 21(1):33–45
10. Roheim CA (2008) Seafood supply chain management: Methods to prevent
illegally-caught product entry into the marketplace. IUCN World Conservation
Union
11. Salin V, Nayga RM (2003) A cold chain network for food exports to developing
countries. Int J Phys Distrib Logistics Manag 33(10):918–933
12. Silva DD, Bjondal T (2013) An open innovation and its role in global fish and
seafood value chains: Beyond the conventional wisdom
13. Vollmann B, Berry W, Claywhybark D (1997) Manufacturing Planning and Control
Systems. Irwin, New York
14. Wessells CR, Anderson JL (1992) Innovations and progress in seafood demand and
market analysis. Mar Resour Econ 07(4):209–228
15. Zott C, Amit R, Massa L (2011) The business model: recent developments and
future research. J Manag 37(4):1019–1042 Official Journal of the Southern Man-
agement Association
Hybrid Global Supply Chain Design Considering
Product Families and Layered Cellular Design
Jue Jiang and Gürsel A. Süer(B)
Department of Industrial and Systems Engineering,

Ohio University, Athens, OH 45701, USA
suer@ohio.edu
Abstract. This paper proposes a new hybrid global supply chain strat-
egy for a global blood sugar manufacturing company. In this approach,
products are analyzed based on product families. Product families can
share cells across the globe if needed. The results are compared with the
host market strategy where three manufacturing plants located in three
regions to meet world demand. Manufacturing plants are designed con-
sidering layered cellular design approach under stochastic demand. This
approach allows three cell types; (1) dedicated cells where each dedicated
cell can be used only by one family, (2) shared cells where a cell can be
shared by two product families, and (3) remainder cells where a cell can
be used by three or more families. The main focus of this paper is to com-
pare both alternatives with respect to number of manufacturing cells and
number of machines needed. In this case, number of machines remained
the same for both approaches. However, the proposed new approach led
to more dedicated cells.
Keywords: Global supply chain · Layered cellular manufacturing · Sto-

chastic demand
1 Introduction
This research aims designing a manufacturing system for a global blood sugar
strip manufacturer. Three manufacturing facilities located in different continents
are assumed to meet the demand of these regions. By using layered cellular man-
ufacturing concepts, number and type of manufacturing cells are determined for
each manufacturing facility considering stochastic demand data and production
rates. Later, this supply chain strategy is compared with the proposed one in
this paper where some cells are shared globally. A probabilistic method is used
to design the cells.
The type of a manufacturing system mostly depends on the layout of the
manufacturing system. As mentioned in Süer, Huang and Maddisetty [19], man-
ufacturing systems can be classified into four categories based on the layout:
process layout, fixed layout, cellular layout and product layout. Fixed Layout

948 J. Jiang and G.A. Süer
is used when products are heavy and large and therefore are mostly stationary.
Therefore, products typically stay in the same position and workers, machines
and equipment are brought to the products to perform the tasks [19]. Product
Layout is appropriate when production volume is high and product variety is
low. Product layout is usually very efficient since it can be configured to perfec-
tion for a single product but this makes it an inflexible system. Process Layout
is the best system for low production volume systems with high product variety
[19]. These systems are very flexible since they can handle many different prod-
ucts but unfortunately they are not very efficient since they requires extensive
material handling and thus lead to long leadtimes. On the other hand, Cellular
Layout is more flexible than Product Layout and more efficient than Process
Layout. It suits well for high product variety with low to moderate demand [19].
Figure 1 shows the classification of four layout types. Obviously, a manufactur-
ing system can indeed consist of any number of layout types as well such both
Cellular Layout and Process Layout etc.
Cellular Manufacturing is based on the grouping of similar products with
respect to similar processes into one or more similar cells as required. In the real
manufacturing world, many uncertainties exist in the system such as demand
uncertainty, supply uncertainty and processing uncertainty. These uncertain-
ties have been discussed extensively in many research works. Süer, Huang and
Maddisetty [19] proposed to consider the uncertainties of product demand and
processing times. By probabilistic market demand calculation, the part-family
assignment is achieved [19]. Then, low utilized cells are grouped to increase the
utilization of the system i.e., to reduce the number of machines and thus invest-
ment, even though some of these cells may no longer be pure cells. These different
cell types, namely, dedicated (DC), shared (SC) and remainder (RC) cells are
also shown in Fig. 1.
Supply chain is the network connecting suppliers, manufacturers, distribu-
tion centers and customers [16]. Dicken [7] discussed many supply chain models.
Fig. 1. Four types of manufacturing layout [19]

Hybrid Global Supply Chain Design 949
Fig. 2. Host market production model [7]
Among them, host market production model is illustrated in Fig. 2. In this strat-
egy, each of the geographic regions covers its own demand. In this study, this
strategy will be compared with the newly proposed hybrid global supply chain
strategy that considers product family definitions and layered cellular approach.
2 Literature Review
Literature regarding global supply chain and cellular manufacturing systems is

summarized in this section.
Chopra and Meindl [6] introduced that supply chain model includes three
logistical drivers and three cross functional drivers. They defined logical drivers
as facilities, inventory and transportation and three cross functional drivers as
information, sourcing and pricing. Brush, Maritan and Karnani [2] mentioned
that many factors affected a network of multinational manufacturing facilities.
Besides considering facility selection, Haug [9] also considered production shift-
ing. A facility location model was developed to study the location decision of high
technology firms. The model identified the international manufacturing facility
location based on domestic and potential international production markets. This
allowed production to be transferred from domestic manufacturing facilities to
foreign facilities. This allowed production to be transferred from domestic man-
ufacturing facilities to foreign facilities. Lowe, Wendell and Hu [13] included
production capacity to develop a two-phase multi-screening approach to incor-
porate uncertainty about exchange rates and exchange rate risk. Canel and Das
[4] improved national market into global supply chain market by considering
connection among global markets. Canel and Das [4] developed an integrative
mathematical model to connect global manufacturing and marketing. Jiang and
Suer [11] compared two supply chain strategies in their work. The paper is based
on their work and hybrid supply chain strategy is developed as an alternative
approach. Ates [1] designed supply chain for the same product category and later
simulated competition among four global manufacturers.
Group Technology (GT) was introduced to improve productivity in the Cel-
lular Manufacturing System (CMS). Hyer and Wemmerlov [10] discussed Group
Technology in cellular manufacturing. Rajamani, Singh and Aneja [17] also men-
tioned that GT played an important role in cellular manufacturing. Singh and
Mohanty [18] observed that not many works in the literature used fuzzy con-
cepts to deal with multi-objective framework in the process planning. Chen and
Cheng [5] proposed a supplementary procedure to solve the limitation of Adap-
tive Resonance Theory (ART). They mentioned that the performance of ART
depended on the initial matrix of bottleneck process. The proposed supplemen-
tary procedures could improve the reliability of results. Moreover, a new math-
ematical model based on cell utilization was conducted by Mahdavi, Javadi,
Fallah-Alipour and Slomp [15]. A comparison of part-machine grouping from
this proposed method with the mathematical model from Chen and Cheng [5]
was tested. The model from Mahdavi, Javadi, Fallah-Alipour and Slomp [15]
tended to produce better results. A mixed integer non-linear model was ana-
lyzed by Bulgak and Bektas [3] for CMS. In this paper, the proposed model was
an integrated approach to combine production planning and system reconfigu-
ration. This CMS model was a new model, which includes sequence, duplicate
machines, capacity of machines and lot splitting.
The literature reviews discussed so far included the deterministic CMS prob-
lem. However, cellular manufacturing is difficult to design in the real world due
to uncertainty of manufacturing process. In order to deal with the uncertainty
of product demand along with processing time, another research is proposed
by Süer, Huang and Maddisetty [19]. A heuristic methodology was conducted
to distinguish cell types in the CMS - Dedicated Cell (DC), Shared Cell (SC)
and Remainder Cell (RC). The product family configuration and cell allocation
are accomplished by mathematical analysis. The designed manufacturing system
turned to successfully solve the uncertainty of product demand and processing
time through simulation method. The methodology conducted by Süer, Huang
and Maddisetty [19] is implemented in the current research for the purpose of
designing the manufacturing system given the market demand, part-family for-
mations, and the operations required to process the products. Erenay, Suer, Jing
[8] also developed a mathematical model for designing alternative layered cellular
system and they compared both works.
3 Problem Definition
In this research, a blood glucose test strip manufacturer is studied and alter-
native global supply chain strategies are compared. The strategies compared
are, (1) independent facilities per region (host market strategy) and (2) newly
proposed hybrid approach that allows family-based cell sharing across multi-
ple facilities. Customers from three continents are considered in this study and
they are Europe, Asia and North America. Three manufacturing facilities are
assumed to produce these products and they are located in Ireland, China and
Puerto Rico. The production data and manufacturing processes are obtained
from Lobo [12]. All of the data are converted into common units by considering
market share, revenue, and product price from Ates [1].
In most manufacturing systems, different products require to be processed

on different machines. Due to high product variety, products are grouped into
several families based on their similarity. Table 1 shows an example of product-
machine incidence matrix. In this table, “1” in row i and column j indicates
that product i needs to be produced on machine j. For example, Product 2 (P 2)
is processed on Machine 1 (M 1) and Machine 3 (M 3). One can observe that
products with similar manufacturing processes are grouped together. Table 2
shows families and cells they are assigned to for this example.
However, in real life manufacturing systems, some product families may have
quite high demand, which means they cannot be produced in one cell. Table 3
shows this multiple cell production system. For example, due to high demand,
product families 1, 2 and 3 may need 2, 2 and 3 cells, respectively.
In many cases, demand values for product families follow a probabilistic
distribution. In these situations, expected utilization for some cells of families
may be low and as a result it may be desirable to share one cell by multiple
families. A Dedicated Cell (DC) is assigned to one product family. A Shared
Cell (SC) can run two product families, which have relatively similar operations.
A Remainder Cell (RC) handles more than two product families. Both Shared
Cells and Remainder Cells usually handle product families that have medium or
low expected utilization values for some of its cells. Table 3 shows the cell sharing
between three product families. For example, Cell 1 (C1) is Dedicated Cell for
Product Family 1 (F 1). C2 is also Dedicated Cell for F 2. C4 is a Remainder
Cell to be shared by F 1, F 2 and F 3. Finally, C3 is a Shared Cell between F 1
and F 2 (Table 4).
Table 1. An example of product-machine incidence matrix
Product Machines
M1 M2 M3 M4 M5 M6
P1 1 1 1 - - -
P2 1 - 1 - - -
P3 1 1 1 - - -
P4 - - - 1 1 -
P5 - - - 1 1 -
P6 1 - 1 - - 1
P7 1 - 1 - - 1
Table 2. Product families and cells
Family Products Cell Machines in the Cell

F1 P 1, P 2, P 3 Cell1 M 1, M 2, M 3
F2 P 4, P 5 Cell2 M 4, M 5
F3 P 6, P 7 Cell3 M 1, M 3, M 6
Table 3. Family vs. multiple cells due to high demand
Family Cells
C1 C2 C3 C4 C5 C6 C7
F1 1 1 - - - - -
F2 - - 1 1 - - -
F3 - - - - 1 1 1
Table 4. Layered cellular design due to stochastic demand
Family Cells
C1 C2 C3 C4
F1 1 - 1 1
F2 - 1 1 1
F3 - - - 1
(DC) (DC) (SC) (RC)
4 Alternative Supply Chain Designs
In this section, two alternative supply chain design strategies are discussed. Strat-
egy 1 is the host market strategy where each facility in a continent covers the
demand of the same continent. Strategy 2 discusses the proposed hybrid global
supply chain strategy where cell sharing opportunities are explored across vari-
ous facilities with product families in mind.
4.1 Strategy 1: Independent Supply Chain Design
In this strategy, each region produces many types of products to meet the
demand of its own demand. Products are produced independently in different
facilities, which lead to no transportation and information sharing between dif-
ferent regions. Figure 3 shows that the blood sugar strips are produced in three
manufacturing facilities-China, Ireland and Puerto Rico [11].
4.2 Strategy 2: Hybrid Global Supply Chain Design
In this strategy, each region produces some of their own products. For example,
products in Family 1 and Family 2 are produced in their own facilities in their
continents. However, products in family 3 are produced in a single location and
distributed to the remaining regions. Obviously, manufacturing cell(s) producing
products in F 3 are shared by various regions while this requires transportation
of some products (Fig. 4).
Fig. 3. Independent manufacturing systems
Fig. 4. The proposed hybrid global supply strategy
5 The Methodology Used

In each manufacturing facility, there are both fabrication and packaging cells.
Products have been already divided into five families and this is out of scope
of the paper. Two alternative strategies are discussed. Most of the computa-
tions are identical for both strategies. The only difference is how shared and
remainder cells are created. To be more specific, in the host market strategy,
DC, SC and RC are defined for each facility independently while, in the newly
proposed strategy these opportunities are sought across multiple facilities poten-
tially located in different continents. The general methodology is illustrated in
Fig. 5. Cell utilizations are calculated by using cell capacity, product demand,
etc. Each cell capacity is assumed 2000 h annually.
6 Mean Capacity Requirements and Standard Deviation
Historical demand values of four companies - Roche, LifeScan, Bayer and Abbott
from 2002 to 2010 are used to forecast the 2011 demand [1]. In this research,
demand values for families are assumed to follow normally distribution. Standard
Deviation (σ) values for each product family are generated as a percentage of
the corresponding mean demand (20%–25%). From [1], mean demand by family
in all markets is shown in Table 5 along with Standard Deviations (σ).
Mean Demand and

Standard Deviation
Calculate Capacity
Requirements
Determine Probability
of Demand Coverage
Compute Expected
Cell Utilization Values
Determine DC, SC and Determine DC, SC and

RC for individual RC considering all
facilities facilities
Strategy 1 simultaneously
Strategy 2
Compare Results
Fig. 5. The general methodology followed in this paper
Table 5. Mean demand and standard deviation by family in three regions
Family China (C) Ireland (I) Puerto Rico (PR)

Mean % STDEV Mean % STDEV Mean % STDEV
1 1,422,286 25 355,571 1,422,925 25 355,731 1,337,087 25 334,271
2 7,098,188 24 1,703,565 7,101,377 24 1,704,330 6,672,986 24 1,601,516
3 6,711,423 24 1,610,741 6,714,438 24 1,611,465 6,309,389 24 1,514,253
4 24,313,261 21 5,105,784 24,324,183 21 5,108,078 22,856,826 21 4,799,933
5 3,137,454 25 784,363 3,138,863 25 784,715 2,949,511 25 737,377
Once mean demand and standard deviation values are known, the mean
capacity requirements by product family are calculated by using Eq. (1) in [14].
Bottleneck Processing Time is defined as the longest processing time among
operations in the cell.
Bottleneck Processing Time
MCRFamilyNo = MeanDemand × (hr). (1)
60
For example, the mean capacity requirements for Product Family 2 in the
manufacturing system of China region is decided by Mean Demand of Prod-
uct Family 2 in China region which is 7,098,188. BPT (Bottleneck Processing
Time) is 1/80 = 0.0125 min in China region. The results of Mean Capacity
Requirements and Standard deviation for different regions are shown in Table 6.
0.0125
MCRF 2 = 7, 098, 188 × = 1479 h,
60
0.01252
Standard DeviationCapacityF 2 = 1, 703, 5652 × = 335.
3600
Table 6. Mean demand and standard deviation by family in three regions

MCR STDEV MCR STDEV MCR STDEVpity
1 296 74 296 74 279 70
2 1479 355 1479 355 1390 334
3 1868 448 1869 449 1756 421
4 5065 1064 5068 1064 4762 1000
5 654 163 654 163 614 154
7 Calculating Demand Coverage Probabilities
The demand coverage probability is calculated for various cell levels. It shows
the probability that a given number of cells will meet the demand. Capacity
Requirements are assumed to follow the normal distribution. Each cell is avail-
able 2000 h per year. Demand Coverage Probability (DCP) for a family and cell
combination is calculated by Eq. (2).

2000 × CellNo. − MCRFamilyNo
DCPFamilyCell = Normsdist . (2)
STDEVCapacity
Demand Coverage Probability for all cells and families are calculated based
on Mean Capacity Requirement and Standard Deviation. The Mean Capacity
Requirement for Product Family 2 for China region is 1479 and Standard Devi-
ation is 355. Based on these values, the Demand Coverage Probability for the
first cell for family 2 is 93%. In other words, one cell covers demand at a very
high level.

2000 × 1 − 1479
DCPF 1C1 = Normsdist = 0.93.
335
All of the results of Demand Coverage Probability for different regions are
shown in Table 7. For family 2 in China facility, by adding a second cell, the
Demand Coverage Probability jumps to 99.99%.
7.1 Expected Cell Utilization Calculation
Expected Cell Utilization is determined by using Demand Coverage Probability,

Mean Capacity Requirements and Standard Deviation as shown in Eq. (3)
E(C = X) =P (CR > X) × P U1 + P X − 1 ≤ CR ≤ X × P U2

(3)
+ P (CR < X − 1)P U3 ,
Table 7. Demand coverage probabilities

1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
1 1.00 - - - - 1.00 - - - - 1.00 - - - -
2 0.93 1.00 - - - 0.93 1.00 - - - 0.97 1.00 - - -
3 0.62 1.00 - - - 0.61 1.00 - - - 0.72 1.00 - - -
4 0.00 0.16 0.81 0.99 1.00 0.00 0.16 0.81 0.99 1.00 0.00 0.22 0.89 0.99 1.00
5 1.00 - - - - 1.00 - - - - 1.00 - - - -
where
E(C = X): Expected cell utilization for the X th cell in a product

family;
P (CR > X): Probability that Capacity Requirements (CR)exceed > X;
P U1 : Percentage utilization of the X th cell when CR > X,
P U1 = 1.0;
P (X −1 ≤ CR ≤ X): Probability that CR between X − 1 and X;
P U2 : Percentage utilization of X th cell when CR between X−1
and X;
P (CR < X −1): Probability that CR < X − 1;
P P U3 : Percentage utilization of X th cell when CR < X − 1,
P U3 = 0.0
P U2 is solved by Eq. (4).

2000X
y × f (y)
P U2 = dy − (X − 1), (4)
2000(X−1) 2000 × A
where
y: Variable represents CR;

f (y): Probability density formation for CR;
A: Probability that CR between X − 1 and X
f (y) and A are calculated by Eqs. (5) and (6)

1 2 1
f (y) = √ e−(y−µ) 2σ2 , (5)
σ 2π
A = P (X − 1 ≤ CR ≤ X). (6)
For example, Expected Cell Utilization of Product Family 1 for China region
is decided by Probability that the Cell Required number larger than 1, Percent-
age utilization of 1st cell when CR > 1, Probability that CR between 0 and 1
and Percentage utilization of 1st cell when CR between 0 and 1.
E(C = 1) = P (CR > 2000) × P U1 + P (0 ≤ CR ≤ 2000) × P U2 + P (CR < 0) × P U3

= 0 × 1 + 1 × P U2 ,
2000
y 1 −(y−296)2 × 1 dy
P U2 = × √ e 2×742 = 0.1480.
0 2000 × 1 74 × 2π
All of the results of Expected Cell Utilization calculations for different regions
are given in Table 8.
8 The Heuristic Algorithm for Layered Cellular Design

Having determined Expected Cell Utilization values for all possible cell segments,
Dedicated Cells (DC), Shared Cells (SC), and Remainder Cells (RC) are identi-
fied using the heuristic algorithm described briefly here. The heuristic algorithm
by Süer, Huang and Maddisetty [19] is used for identifying cell types. Expected
Cell Utilization values are sorted in the decreasing order first. The assignment
starts with the highest Expected Cell Utilization. If the expected cell utilization
is 100%, this cell is considered to be a Dedicated Cell (DC). If the expected cell
utilization is larger than 50%, a cell will be allocated to a product family. Then
other similar product families are allocated to the cell to make the cell coverage
close to 100% by considering similarities among families. These cells are named
Shared Cells (SC) if they process only two product families and remainder cells
if they process three or more product families.
Threshold value is the lowest acceptable similarity coefficient that allows
two families to be grouped in a cell. The Similarity Threshold is set to 77%
Table 8. Demand coverage probabilities

1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
1 0.15 0.00 - - - 0.15 0.00 - - - 0.14 0.00 - - -
2 0.73 0.01 0.00 - - 0.73 0.01 0.00 - - 0.69 0.00 - - -
3 0.87 0.06 0.00 - - 0.87 0.06 0.00 - - 0.84 0.04 0.00 - -
4 1.00 0.96 0.52 0.06 0.00 1.00 0.96 0.52 0.06 0.00 1.00 0.94 0.42 0.03 0.00
5 0.33 0.00 - - - 0.33 0.00 - - - 0.31 0.00 - - -
Table 9. Similarity coefficients between product families
Family 1 2 3 4 5
1 - 1.00 0.89 0.78 0.70
2 1.00 - 0.89 0.78 0.70
3 0.89 0.89 - 0.70 0.80
4 0.78 0.78 0.70 - 0.89
5 0.70 0.70 0.80 0.89 -
Table 10. Cell Types for independent regions

1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
1 - - - 0.15 - - - - 0.15 - - - - 0.14 -
2 - 0.01 - 0.73 - - 0.01 - 0.73 - - - - 0.69 -
3 - - 0.87 0.06 - - 0.87 0.06 - - - 0.84 0.04 -
4 1.00 0.96 - 0.06 0.52 1.00 0.92 - 0.02 0.52 1.00 0.94 - 0.03 0.42
5 - - - - 0.33 - - - - 0.33 - - - - 0.31
DC SC DC RC SC DC SC DC RC SC DC DC DC RC SC
Tabulate the expected cell

utilization (ECU) and
similarity coefficient values
Sort ECU in the

decreasing order
Consider a coverage
segment with highest ECU
Is ECU = Allocate a cell to the

Yes
100%? family
No
No Decrease ST & consider existing

Is ECU
No cells for possible assignment (if
>=50%
capacity is avaliable)
Yes
Allocate a cell to the

No Are all coverage Remaining segments from
family No
segments assigned? a remainder cell
Can the cell be utilized by other Yes

coverage segments?
No
Allocation of families to cells is
Yes complete
Set similarity threshold
(ST) & compare similarity
between original family
allocated to cell and other
families
Is the similarity coefficient

between the family/families >=
ST?
Yes
Yes
Assign the family to the

cell similar to the original
family
Are there any other

segments for possible
assignment?
Fig. 6. The heuristic procedure for the layered cellular design

Table 11. Cell types for China based on new approach
Family China
1 2 3 4 5(C + PR) 6(C + I + PR)
1 - - - 0.15 - -
2 - 0.01 - 0.73 - -
3 - - 0.87 0.06 - -
4 1.00 0.96 - 0.06 0.52 + 0.42 -
5 - - - - - 0.33 + 0.33 + 0.31
DC SC DC RC DC DC
Table 12. Cell types for Puerto Rico and Ireland based on new approach
Family Puerto Rico (PR) Ireland (I)

1 2 3 4 1 2 3 4 5
1 - - - 0.14 - - - 0.15 -
2 - - - 0.69 - 0.01 - 0.73 -
3 - - 0.84 0.04 - - 0.87 0.06 -
4 1.00 0.94 - 0.03 1.00 0.92 - 0.02 0.52
5 - - - - - - - - -
DC DC DC DC DC DC DC DC DC
in this paper. For example, when sorting the ECU for Ireland region, the highest
ECU is 100% for Family 4 in Cell 1. Then Product family 4 is allocated to cell
1. When the second cell with 92% utilization is considered, it is allocated to a
new cell − Cell2. Table 10 is used to search the similar Families with Product
Family 4. From Table 9, Families 1, 2 and 5 can be considered to share a cell
with Product Family 4 (similarity coefficient > 0.77). However, merging this cell
with Family 1 or Family 5 will exceed 10% utilization. The only option is to
merge Cell 2 (1% utilization) of Family 2 with Family 4. In Ireland case, there
are two Dedicated Cells, two Shared Cells and one Remainder Cell. The results
are summarized in Table 10. The heuristic procedure is given in Fig. 6.
9 The Proposed Hybrid Global Supply Approach

The proposed approach looks into global supply chain and explores the oppor-
tunities for forming shared cells and remainder cells across global facilities. In
this paper, a new cell (cell 6) is opened in China where the entire Family 5 is
produced to meet the global demand. In the same facility, demand for product
family 4 original assigned to Puerto Rico facility is re-assigned to China facility.
The Puerto Rico facility is thus reduced to four cells with three DCs and one
RC. Ireland facility is also changed to 3 DCs, one SC and one RC as shown in
Table 13. Comparison of results
Parameters Independent supply chain New hybrid supply chain

China Ireland P.R. Total
# of DCs 2 2 2 6 9
# of SCs 2 2 2 6 3
# of RCs 1 1 1 3 3
# of total cells 5 5 5 15 15
# of Cell Util > 0.9 3 3 3 9 11
# of workers 82 82 82 246 234
# of machines 66 66 66 198 198
Table 12. In short, China facility becomes a global supplier for Product Family 5
and partial global supplier for Product Family 4. All three facilities supply their
own markets with respect to product Families 1, 2 and 3 (Table 11).
10 Conclusion
The comparison between two designs is presented in Table 13. It shows that total
number remains the same for both designs. However, the number of workers
is reduced in the new design. This is due to balancing issues in the cells. In
other words, each cell is equipped with all the machines needed to run all the
families (maximum number of machines) while some of them remained idle while
running different families. Another benefit was that there were more dedicated
cells under the new design. Furthermore, number of cells with a utilization value
greater than 0.90 was higher. It is expected that flowtimes will be shorter in
dedicated cells. This in turn reduces the work-in-process inventory. For a more
complete analysis, transportation costs, labor costs as well as investments costs
have to be included. These results could not be included due to space concerns.
References
1. Ates O (2013) Tglobal supply chain and competitive business strategies: a case
study of blood sugar monitoring industry. https://etd.ohiolink.edu/
2. Brush TH, Marutan CA, Karnani A (1999) The plant location decision in multi-
national manufacturing firms: an empirical analysis of international business and
manufacturing strategy perspectives. Prod Oper Manag 8:109–132
3. Bulgak AA, Bektas T (2009) Integrated cellular manufacturing systems design
with production planning and dynamic system reconfiguration. Eur J Oper Res
192:414–428
4. Canel C, Das SR (2002) Modeling global facility location decisions: integrating
marketing and manufacturing decisions. Ind Manag Data Syst 102:110–118
5. Chen SJ, Cheng CS (1995) A neural network-based cell formation algorithm in
cellular manufacturing. Int J Prod Res 33:20–542
6. Chopra S, Meindl P (2013) Supply chain management: strategy, planning & oper-
ations, 2nd edn. Prentice Hall, Upper Saddle River
7. Dicken P (1992) Global shift: the internationalization of economic activity, 2nd
edn. Paul Chapman Publishing, London
8. Erenay B, Suer GA et al (2015) Comparison of layered cellular manufacturing
system design approaches. Comput Ind Eng 85:346–358
9. Haug P (1992) An international location and production transfer model for high
technology multinational enterprises. Int J Prod Res 30:559–572
10. Hyer NL (1992) Education: Mrp/gt: a framework for production planning and
control of cellular manufacturing. Decis Sci 13:681–701
11. Jiang J, Suer GA (2016) Alternative global supply chain design strategies for a
blood sugar strip manufacturer considering layered cellular design. In: Proceedings
of the global joint conference on industrial engineering and its application areas
2016, Istanbul, Turkey, 14–15 July, pp 14–15
12. Lobo R (2006) Comparison of connected vs disconnected cellular systems using
simulation. Ph.D. thesis, Ohio University
13. Lowe TJ, Wendell RE, Hu G (2002) Screening location strategies to reduce
exchange rate risk. Eur J Oper Res 136(3):573–590
14. Maddisetty S (2005) Design of shared cells in a probabilistic demand environment.
Ph.D. thesis, Ohio University
15. Mahdavi I, Javadi B et al (2007) Designing a new mathematical model for cellular
manufacturing system based on cell utilization. Appl Math Comput 190:662–670
16. New SJ, Payne P (1995) Research frameworks in logistics: three models, seven
dinners and a survey. Int J Phys Distrib Logistics Manag 25:60–77
17. Rajamani D, Singh N, Aneja YP (1990) Integrated design of cellular manufacturing
systems in the presence of alternative process plans. Int J Prod Res 28:1541–1554
(in Chinese)
18. Singh N, Mohanty BK (1991) A fuzzy approach to multi-objective routing problem
with applications to process planning in manufacturing systems. Int J Prod Res
29:1161–1170
19. Süer GA, Huang J, Maddisetty S (2010) Design of dedicated, shared and remainder
cells in a probabilistic demand environment. Int J Prod Res 48:92–114
Design of Closed-Loop Supply Chain Model
with Efficient Operation Strategy
Xing Chen1 , Anudari Chuluunsukh1 , YoungSu Yun1(B) , and Mitsuo Gen2

1
Department of Business Administration, Chosun University, Gwangju, Korea
ysyun@chosun.ac.kr
2
Fuzzy Logic Systems Institute, Tokyo University of Science, Katsushika, Japan
Abstract. In this paper, a closed-loop supply chain model with effi-

cient operation strategy (CLSC-OS) is proposed. In the CLSC-OS, var-
ious facilities which can be used in forward logistics (FL) and reverse
logistics (RL) are taken into consideration. A mathematical formulation
is proposed to design the CLSC-OS and it is implemented using genetic
algorithm (GA) approach. In numerical experiment, for efficient opera-
tion strategy, several scenarios using various factors such as remanufac-
turing rate, profit premium rate, discount rate, and return rate based on
revenue-cost rate is considered and analyzed. Experimental results shows
that discount rate and profit premium rate have a significant influence
on revenue-cost rate.
Keywords: Closed-loop supply chain · Efficient operation strategy ·

Forward logistics · Reverse logistics · Genetic algorithm
1 Introduction
In general, designing closed loop supply chain (CLSC) model is to maximize the
utilization of resources. Especially, return management such as reuse and resale
of the recovered parts and products has been focused on many companies. For
example, by 1997, approximately 73,000 U.S. companies had earned $53 billion
though resale activities of remanufactured products [1]. This means that return
management using used product becomes more and more important.
Return management using used product is usually consisted of three aspects.
The first aspect is the operational management of returned product, the sec-
ond one is that of remanufacturing activity, and the third one is that of sec-
ondary market [6]. Ozkir and Basligil considered the operational management of
returned product [8]. They found that regulating the rate of returned product
can increase the total profit and revenue. Subramoniam found that remanufac-
turing activities can increase profit, save resource and energy, create new market
[9]. Van Daniel et al. suggested that con-sidering both new market with new
product activity and secondary market with resale activity can get more profit
than the new market alone [7].
Design of Closed-Loop Supply Chain Model 963
However, unfortunately, little attention to the relationship among the three

aspects of return management has been done in conventional studies. There-
fore, in this paper, we consider the relationship based on various scenarios using
remanufacturing rate, profit premium rate, discount rate, and return rate and
analyze which factor can be an important one to increase total profit.
For achieving this objective, a CLSC model with efficient operation strategy
(CLSC-OS) is proposed in Sect. 2. The CLSC-OS is consisted of various facilities
such as supplier, product manufacturer, retailer, recovery center, and secondary
market at each stage. In Sect. 3, a mathematical formulation is proposed for
implementing the CLSC-OS. The mathematical formulation is represented as
a nonlinear mixed integer programming (NIMIP). The NIMIP is implemented
using a genetic algorithm (GA) approach in Sect. 4. In numerical experiment
of Sect. 5, the CLSC-OS is presented and various scenarios using the CLSC-OS
is considered. For various scenarios, remanufacturing rate, profit premium rate,
discount rate, and return rate are considered and they all are used for analyzing
the revenue-cost rate resulting from the implementation of the CLSC-OS. Some
conclusions and a room for improvement to future study are followed in Sect. 6.
2 Proposed CLSC-OS
The proposed CLSC-OS is an integrated model combining forward logistics (FL)
and reverse logistics (RL), which is consisted of suppliers in areas 1, 2, 3, and 4,
product manufacturer, part inventory distribution center, product distribution
center and retailer in FL and customer, collection center, recovery center, sec-
ondary market and waste disposal center in RL. For effective use of the recovered
parts from recovery center, part inventory distribution center is also taken into
consideration. The conceptual network structure of the proposed CLSC-OS is
shown in Fig. 1.
In Fig. 1, the production and recovery flows are as follows. New part types
1, 2, 3, and 4 (NP1, NP2, NP3, and NP4) are produced at the suppliers of
areas 1, 2, 3, and 4, respectively. NP1, NP2, NP3, and NP4 are then sent to
part inventory distribution center. Recovered parts (RP1, RP2, RP3, and RP4)
from recovery center are also sent to part inventory distribution center. Product
manufacturer produces product using NP1, NP2, NP3, NP4, RP1, RP2, RP3,
and RP4 from part inventory distribution center. The product is sent to retailer
through product distribution center and sold to customer. The returned product
from customer is collected at collection center and then sent to recovery center.
At recovery center, the returned product is checked and classified into recover-
able and unrecoverable products. The quality of the recoverable product with
a1 % is recovered at recovery center and then resold at secondary market. The
unrecoverable product is disassembled into recoverable and unrecoverable parts.
The quality of the recoverable part with a2 % is recovered at recovery center and
then sent to part inventory distribution center. The unrecoverable part with a3 %
is sent to waste disposal center to be burned or landfilled.
Especially, in the proposed CLSC-OS, the part inventory distribution cen-
ter can regulate the transportation amount of NP1, NP2, NP3, NP4, RP1, RP2,
964 X. Chen et al.
Fig. 1. The conceptual network structure of the CLSC-OS
RP3, and RP4. For instance, if product manufacturer want to produce 100 prod-
ucts and a2 is 10% (10 recovered parts = 100 returned product ×10%), then part
inventory distribution center sends 90 NP1, 90 NP2, 90 NP3, and 90 NP4 as new
parts and 10 RP1, 10 RP2, 10 RP3, and RP4 as recovered parts to product man-
ufacturer.
Beofre suggesting the mathematical formulation for the CLSC-OS, the following
assumptions should be considered.
• Single product is produced within a fixed period.

• New market (retailer) and resale market (secondary market) are separated.
• The qualities of the new parts (NP1, NP2, NP3, and NP4) from supplier and
those of the recovered parts (RP1, RP2, RP3, and RP4) from recovery center
are identical.
• All products (new and recovered products) at retailer and secondary market
are sold to customer.
• The number of facilities at each stage is known.
• All facilities at retail/customer, secondary market, and waste disposal center
should be opened. However, only one facility at supplier, product manufac-
turer, product distribution center, part inventory distribution center, collec-
tion center, and recovery center should be opened, which means that the other
facilities at each stage should be closed.
• Fixed costs of all facilities are different and known.
• Unit handling costs of all facilities at same stage are identical and known.
• Unit transportation costs of all facilities at each stageare different and known.
• The handling capacity of the facilities at a stage is the same or greater than
that of previous ones.
• The return rate from customer is fixed.
• The discount rates of recovered product and part are calculated according to
the quality of returned product and part.
The index sets, parameters, and decision variables for the mathematical for-
mulation are as follows.
Index set
a : index of area at supplier; a ∈ A;
b : index of new product; b ∈ B;
h : index of supplier; h ∈ H;
i : index of part inventory distribution center; i ∈ I;
j : index of product manufacturer; j ∈ J;
k : index of product distribution center; k ∈ K;
l : index of retailer/customer; l ∈ L;
m: index of collection center; m ∈ M ;
n : index of recovery center; n ∈ N ;
o : index of secondary market; o ∈ O;
p : index of waste disposal center; p ∈ P
Parameter
FDSha : fixed cost at supplier h of area a;
FDIi : fixed cost at part inventory distribution center i;
FDOj : fixed cost at product manufacturer j;
FDDk : fixed cost at distribution center k;
FDCm : fixed cost at collection center m;
FDRn : fixed cost at recovery center n;
UHSha : unit handling cost at supplier h of area a;
UHOj : unit handling cost at product manufacturer j;
UHDk : unit handling cost at product distribution center k;
UHCm : unit handling cost at collection center m;
UHRn : unit handling cost at recovery center n;
UHEe : unit handling cost at secondary market e;
UHWp : unit handling cost at waste disposal center p;
UTShai : unit transportation cost from supplier h of area a to part inventory
distribution center i;
UTIij : unit transportation cost from part inventory center i to product
manufacturer j;
UTOjk : unit transportation cost from product manufacturer j to product
distribution center k;
UTDkl : unit transportation cost from product distribution center k to
retailer/customer l;
966 X. Chen et al.
UTAlm : unit transportation cost from retailer/customer l to collection

center m;
UTCmn : unit transportation cost from collection center m to recovery
center n;
UTEno : unit transportation cost from recovery center n to secondary
market o;
UTIni : unit transportation cost from recovery center n to part inventory
distribution center i;
UTLnp : unit transportation cost from recovery center n to waste disposal
center p;
KCIi : holding cost at part inventory distribution center i;
UPNb : unit price of new product b;
PRNb : profit premium rate for new product b at retailer l
Decision variables
TCsha : treatment capacity at supplier h of area a
TCii : treatment capacity at part inventory distribution center i
TCoj : treatment capacity at product manufacturer j
TCdk : treatment capacity at product distribution center k
TCul : treatment capacity at retailer/customer l
TCcm : treatment capacity at collection center m
TCen : treatment capacity for recovered product at recovery center n
TCin : treatment capacity for recovered part at recovery center n
TCln : treatment capacity for unrecovered part at recovery center n
TCee : treatment capacity at secondary market e
TCwp : treatment capacity at waste disposal center p
Q : quantity of new product
Rt : return rate
DSo : discount rate for recovered product
DSi : discount rate for recovered part
ysha : takes the value of 1 if suppler h at area a is opened and 0 otherwise
yii : takes the value of 1 if part inventory distribution center i is opened
and 0 otherwise
yoj : takes the value of 1 if product manufacturer j is opened and 0
otherwise
ydk : takes the value of 1 if product distribution center k is opened
and 0 otherwise
ycm : takes the value of 1 if collection center m is opened and 0 otherwise
yrn : takes the value of 1 if recovery center n is opened and 0 otherwise
Using the above mentioned index sets, parameters, and decision variables,
the mathematical formulation is suggested as follows:
Maximize Total Revenue (TR)
TR =revenue of new product + revenue of recovered product

(1)
+ revenue of recovered part

= Q × PRNb + Q × UPNb × α1 ×DSo + Q×UPNb ×α2 × DSi , (2)
i o o
where
UPNb = (TCF × (1 + PRNb )) /Q. (3)
Minimize Total Cost (TC)
TC = total cost of FL (TCF) + total cost of RL (TCR). (4)
TCF = total FL fixed cost + total FL handling cost

+ total FL transportation cost

= (FDSha × ysha ) + (FDIi × yii ) + (FDOj × yoj )
a h i i

+ (UHSha × T Csha × ysha ) + (UHOj × T Coj × yoj )
a h j

+ (UHDk ×T Cdk ×ydk )+ (U T Shai × T Csha × ysha × yii )
j a h i

+ (UTIij × T Cii × yii × yoj ) + (UTOjk × T Coj × yoj × ydk )
ij jk

+ (UTDkl × T Cdk × ydk ) ,
kl
(5)
TCR = total RL fixed cost + total RL handling cost
+ total RL transportation cost + holding cost

= (FDCm × ycm ) + (FDRn × yrn ) + (UHCm × T Ccm × ycm )
m i a h

+ (UHRn × T Cen × yrn ) + (U HRn × T Cin × yrn )
n n

+ (UHRn ×T Cln ×yrn )+ (UHEe ×T Cee )+ (UHWp ×T Cwp )
n e p

+ (UTAlm × T Cul × ycm ) + (UTCmn × T Ccm × ycm × yrn )
lm mn

+ (UTEno × T Cen × yrn ) + (UTIni × T Cin × yrn )
no ni

+ (UTLnp × T Cln × yrn ) + (KCIi × yii )
ni i
(6)
968 X. Chen et al.

s.t. ysha = 1, ∀a ∈ A (7)
h

yii = 1 (8)
i

yoj = 1 (9)
j

ydk = 1 (10)
k

ycm = 1 (11)
m

yrn = 1 (12)
n

(U HShai × ysha ) − (T Cii × yii ) = 0, ∀a ∈ A (13)
h i

(T Cii × yii ) − (U HOj × yoj ) = 0 (14)
i j

(UHOj × yoj ) − (UHDk × ydk ) = 0 (15)
i k

(UHDk × ydk ) − T Cul = 0 (16)
k l

T Cul − (UHCm × ycm ) = 0 (17)
l m

(T Cii × yii ) − α2 (UHRn × yrn ) ≥ 0 (18)
i n

(HCrcn × x rcn ) − α3 (U HWp × ywp ) ≥ 0 (19)
n p

(UHRn × yrn ) − (UHCm × ycm ) = 0 (20)
n o

(UHRn × yrn ) − α1 UHEe ≥ 0 (21)
n e
ysha = {0, 1}, ∀h ∈ H, a ∈ A (22)
yii = {0, 1}, ∀i ∈ I (23)
yoj = {0, 1}, ∀j ∈ J (24)
ydk = {0, 1}, ∀k ∈ K (25)
ycm = {0, 1}, ∀m ∈ M (26)
yrn = {0, 1}, ∀n ∈ N (27)
UHShai , UHOj , UHDk , UHCm , UHRn , UHEe , UHWp ≥ 0
∀h ∈ H, a ∈ A, ∀i ∈ I, j ∈ J, ∀k ∈ K, (28)
∀m ∈ M, ∀n ∈ N, ∀e ∈ E, ∀p ∈ P
The objective function is divided into two parts. First part is to maximize
total revenue and second one is to minimize total cost. Maximizing the total
revenue is represented in Eq. (2) and minimizing the total cost is in Eq. (4).
Each objective function should be maximized and minimized under satisfying
all constraints which represented from Eqs. (7)–(28). Equations (7)–(12) mean
that only one facility at each stage should be opened. Equations (13)–(21) show
that the capacity of each opened facility is the same or greater than that of the
previous one. Equations (22)–(27) means that each decision variables can take
the value 0 or 1. Equation (27) refers to non-negativity.
4 Genetic Algorithm (GA) Approach

Genetic Algorithm (GA) approach was proposed by Goldberg [5] and has been
used to solve many optimization problems [3,4,10]. GA approach starts with
producing a random initial population and each individual in the population
represents a potential solution. The next step is that each individual is selected
by evaluating the measure of fitness and producing a next new population. All
individuals in the new population are undergoing crossover and mutation oper-
ators. For crossover and mutation opera-tors, various methodologies such as
one-point crossover operator (1X), two-point crossover operator (2X), and ran-
dom mutation operator have been developed. After applying the crossover and
mutation operators to each individual, an offspring is produced and the fitness is
evaluated. These procedure is repeated until a pre-defined stop condition is sat-
isfied or optimal solution already known is located. Finally, the optimal solution
is produced.
The GA approach [2,3] for implementing the mathematical formulation in
Sect. 3 is as Fig. 2:
Fig. 2. The GA approach
In numerical experiment, various scenarios are considered for evaluating the rate
of the total revenue in Eq. (1) and the total cost in Eq. (4). First, a scale for
970 X. Chen et al.
the proposed CLSC-OS is presented in Table 1. As shown in Table 1, numbers of

the suppliers (S) at each area, part inventory distribution center (PI), product
manufacturer (PM), product distribution center (PD), retailer/customer (RT),
collection center (CC), recovery center (RC), secondary market (SM), waste
disposal center (WD) are displayed.
Table 1. Scale of the CLSC-OS
No. of S No. of PI No. of No. of No. of No. of No. of No. of No. of

Area 1, PM PD RT CC RC SM WD
2, 3, 4
4 2 4 2 10 4 2 10 2
Table 2 shows various cases of remanufacturing for returned product at recov-

ery center. a1 , a2 , and a3 refer to the rates of recovered product, recovered part,
and unre-covered part, respectively. For instance, Case 1 means that (i) 60% of
all returned products are recovered and sent to secondary market, (ii) 30% of
them are recovered and sent to part inventory distribution center and (iii) 10%
of them are unrecovered and sent to waste disposal center.
Table 2. Various cases of remanufacturing
Case a1 a2 a3
1 0.6 0.3 0.1
2 0.6 0.2 0.2
3 0.6 0.1 0.3
4 0.7 0.2 0.1
5 0.7 0.1 0.2
6 0.8 0.1 0.1
Table 3 shows various scenarios. Each scenario has four factors using the
case of remanufacturing, return rate, discount rate and profit premium rate.
By the each scenario, various relationships between total revenue and total cost
are compared. For instance, Scenario 1 indicates that various cases in Table 2
and various profit premium rates (Pm = 0.4, 0.3, 0.2, 0.1, and 0.0) under the
fixed values at return rate and discount rate are considered to solve the two
mathematical formulations in Eqs. (1) and (4). The computation results are
used for comparing the relationships between total revenue and total cost.
Each scenario in Tables 2 and 3 under the scale of the proposed CLSC-OS is
implemented using GA approach. The parameters used in GA approach is that
population size is 20, crossover rate 0.5, and mutation rate 0.3. The GA approach
is executed under the following computation environment: Matlab R2015 under
Table 3. Various scenarios
Scenario Case Return rate Discount rate Profit premium rate

1 1, 2, 3, 4, 5, 6 1.0 0.5 0.4, 0.3, 0.2, 0.1, 0.0
2 1 1.0 0.5 0.4, 0.3, 0.2, 0.1, 0.0
3 1 1.0 0.5, 0.4, 0.3, 0.2 0.4, 0.3, 0.2, 0.1, 0.0
4 1, 2, 3, 4, 5, 6 0.9 0.5 0.4, 0.3, 0.2, 0.1, 0.0
IBM compatible PC 3.40GHZ processor (Inter Core i7-3770 CPU), 8GB and
Window 7.
Figure 3 shows the computation results of total revenue-cost rates in Scenario
1. Each curve with different colors shows the changes of the total revenue-cost
rates under the changes of profit premium rate. For instance, the top blue-colored
curve shows a fixed values of 1.4 in the total revenue-cost rates of all cases under
the fixed values of 0.4 profit premium rate, 100% return rate and 0.5 discount
rate. Similar situations are also shown in the other colored-curves. Therefore,
the results shown in Fig. 3 indicates that the changes of remanufacturing rate
has no significant influence on the increase of total profit.
Fig. 3. Computation results of total revenue-cost rates in Scenario 1
Figure 4 shows the changes of the total revenue-cost rates under various profit
premium rates, when return and discount rates are fixed at 100% and 0.5, respec-
tively. As the changes from profit premium rate 0.4 to 0.0, the total revenue-cost
rate is continuously decreasing. Therefore, the results shown in Fig. 4 indicate
that the profit premium rate has significant influence on the increase of total
profit.
Figure 5 shows the changes of the total revenue-cost rates for each profit
premium rate under various discount rates, when return rate is fixed at 100%.
For instance, the top blue-colored curve shows that total revenue-cost rates is
972 X. Chen et al.
increasing according to the changes of discount rate from 0.5 to 0.2. Similar sit-
uations are also shown in the other colored-curves. Therefore, the results shown
in Fig. 5 indicate that the change of discount rates has significantly influence on
the increase of total profit.
Similar to Figs. 3 and 6 shows the computation results of total revenue-cost

rates in Scenario 4. Each curve with different colors shows the changes of the
total revenue-cost rates under the changes of profit premium rate. For instance,
the top blue-colored curve shows a little change around 1.55 for each case when
the discount rate and profit premium rate are fixed at 90% and 0.4, respectively.
Similar curves are also shown in the profit premium rates 0.3 and 0.0. However,
the curves of profit premium rates 0.2 and 0.1 do not show any changes for each
case. Therefore, the results shown in Figs. 3 and 6 indicate that the changes of
remanufacturing rate and return rates has no significant influence on the increase
of total profit, though there is a little changes according to each case.
6 Conclusion
The first objective of this study is to design a CLSC-OS with various facilities at
each stage and the second one is to find a good alternative for increasing total
profit under the changes of remanufacturing rate, discount rate, profit premium
rate, and return rate.
For achieving the two objectives, the CLSC-OS has been proposed. The pro-
posed CLSC-OS has supplies at areas 1, 2, 3, and 4, product manufacturer,
product distribution center, and retailer in FL and customer, collection cen-
ter, recovery center, secondary market and waste disposal center in RL. This
is represented in a mathematical formulation which has two objectives of the
maximization of total revenue and the minimization of total cost under various
constraints. The mathematical formulation is implemented using GA approach.
In numerical experiment, four types of scenarios with the changes of remanufac-
turing rate, discount rate, profit premium rate, and return rate have been con-
sidered and executed using GA approach. The experimental results have shown
that the changes of the profit premium rates and the discount rates have no
significant influence on the increase of total profit, but the changes of the profit
premium rates and the discount rates have significant influence on the increase
of total profit. Therefore, decreasing the profit premium rates and the discount
rates can be a good alternative for increasing the total profit (= total revenue
− total cost) in the proposed CLSC-OS.
However, in this paper, since only four scenarios have been considered, more
various scenarios with larger-scaled CLSC-OS will be considered to compare the
changes of remanufacturing rate, discount rate, profit premium rate, and return
rate. This will be left to future study.
Acknowledgements. This work is partly supported by the Grant-in-Aid for Scientific

974 X. Chen et al.
References
1. Atasu A, Sarvary M, Wassenhove LNV (2008) Remanufacturing as a marketing
strategy. Manage Sci 54(10):1731–1746
York
3. Gen M, Cheng R (2000) Genetic algorithms and engineering optimization. Wiley,
New York
4. Gen M, Zhang W et al (2016) Recent advances in hybrid evolutionary algorithms
for multiobjective manufacturing scheduling. Comput Ind Eng doi:10.1016/j.cie.
2016.12.045
5. Goldberg DE (1989) Genetic algorithms in search, optimization and machine learn-
ing, vol 13. Addsion-Wesley, Reading
6. Guide VDR, Wassenhove LNV (2009) The evolution of closed-loop supply chain
research. Oper Res 57(1):10–18
7. Li J (2010) The potential for cannibalization of new products sales by remanufac-
tured products. Decis Sci 41(3):547–572
8. Ozkır V, Başlıgıl H (2012) Modelling product-recovery processes in closed-loop
supply-chain network design. Int J Prod Res 50(8):2218–2233
9. Subramoniam R, Huisingh D, Chinnam RB (2009) Remanufacturing for the auto-
motive after market-strategic factors: literature review and future research needs.
J Cleaner Prod 17(13):1163–1174
10. Yun YS, Chung HS, Moon C (2013) Hybrid genetic algorithm approach for
precedence-constrained sequencing problem. Comput Ind Eng 65(1):137–147
Optimization of Supply Chain
Based on Internet Plus
Wen Hua(B)
School of Tourism and Economic Management,

Chengdu University, Chengdu 610000, People’s Republic of China
846937934@qq.com
Abstract. “Internet plus supply chain” one of the internet plus strate-
gic action, basing on information technology and platform, promotes the
high fusion of elements of supply chain and modern science & technology,
and upgrades efficiency of supply chain. The basic factors-information
technology guarantee and information sharing, which keep the supply
chain operation efficiently, protect the “three flow”-material flow, infor-
mation flow and fund flow efficient and order movement on the supply
chain. This paper analyses the situation and the problem about “three
flow”, and discusses how to improve the efficiency of supply chain sys-
tem, from client, to manufacture, and to logistics and financial service
with the help of internet technology.
Keywords: Internet plus · Logistics management · Supply chain man-

agement
1 Introduction
In 2015, China Premier Li Keqiang puts forwards the “Internet plus” strategic
action plan in the government work report to integrate the Internet with tradi-
tional industries by means of the Internet platform and information technology,
to encourage industrial networks and Internet banking, and to guide Internet-
based companies to increase their presence in the international market.
The internet provides information and communication technology and plat-
form for enterprises, promotes the effective communication and coordination
between enterprises, business partners, customers and inner-enterprise, and
makes it possible to implement supply chain effectively. The enterprises in sup-
ply chain can make the synergistic responds quickly, reduce costs of supply chain
effectively and improve enterprise efficiency. Even The ratio of logistics cost to
GDP keeps declining from 18.4% in 2007 to 16.0% in 2015, and is 0.6% lower
than the year earlier [8], it is still higher. The logistics cost in our country is 5%
higher than developed countries, which shows the obvious gap. Global Logistics
Performance Index (LPI) issued by the World Bank every two years is a interna-
tional index that demonstrates the domestic logistics abilities of one country or
976 W. Hua
one region to participate global supply chain. The 2016 assessment result shows
China ranks at 27th [9] and is in the second echelon. As a whole, the status
quo of the weak soft power in circulation filed, such as high cost, low efficiency,
low intensive level, insufficient industrial support and issues of trust, standard,
talents, safe and environmental protection, constraints the efficient operation of
supply chain.
The supply chain is a sequential and interconnected organization structure
including the design, production and sales and a value chain involving suppliers,
manufacturers, transportation and all other components. Meanwhile, the supply
chain links business activities to achieve accurate forecast, less inventory and
meet real-time needs based on information system. This paper starts with the
situation and the problem about “three flow”, and discusses how to improve the
efficiency of supply chain system, from client, to manufacture, and to logistics
and financial service with the help of internet technology.
2 Current Situation of Logistics, Information Flow and

Fund Flow in the Supply Chain in China
In the Internet plus era, the supply chain is showing a trend of flat model to
ensure that logistics, information flow and fund flow perform effectively and
orderly in the system. As globalization lengthens and broadens the supply chain,
the competition is more and more intense and the competition between enter-
prises becomes the competition between supply chains.
2.1 Logistics in Supply Chain
Modern logistics links every supply chain activity from raw material procurement
to retails, achieving external integration and reflecting integration advantage.
With the speeding up of globalization, raw material procurement and product
sales expand in scale in China’s manufacturing enterprises. Procurement and
sales are attached with the same importance. In terms of transportation, owing
to China’s road especially expressway construction, consummation and construc-
tion of highway networks and transformation of consumption patterns, the road
transportation scope was widen and highway logistics accounted for 75.52% of
the total freight volume (2015) as shown in Table 1.
Table 1. Annual traffic volume of transportation (10000 ton)
Year Railway Highway Water transportation Civil aviation Total

1990 150681 724040 80094 37.0 970602
2000 178581 1038813 122391 196.7 1358682
2015 336000 3150000 614000 625.3 4171000
Source: National Bureau of Statistics et al., China Statistics Yearbook
Optimization of Supply Chain Based on Internet Plus 977
The data has shown [3] that: In China, there exist over 7 million logistics
companies with more than 16 million vehicles and an average of 2 cars for each
company. The trucks empty driving rate is 50% and the cargo are transferred
5 to 6 times before transport drivers. As the individual drivers owe more than
90% of road transport capacity, the road logistics has low degree in scale and
intensification and road transport is in small, scattered, chaotic, disorder state.
All in all, long-distance road transportation has no advantage in economy and
environmental protection.
Revenue composition of logistics enterprises: transport and warehousing busi-
ness make up for 60%, freight forwarding business for 13.9%, distribution and
distribution processing for 5% [2], indicating that logistics enterprises focus on
transport and warehousing that are the base business with low profit margins.
And for profit margin: integrated logistics enterprises accounts for 8.9%, ware-
housing enterprises for 6.9% and transportation enterprises for 3.3%.
In general, for the majority of small and medium-sized logistics enterprises,
they are lack of fund, labor force, capacity and application of information tech-
nology; the labor-intensive operation relies on a large number of labors; they have
insufficient capacity to provide value-added logistics service since their business
focus are basic logistics services such as transportation, warehousing. What’s
more, the general low profitability can be attributed to two reasons. One is that
demand of logistics services reduces as China’s economy slows down. At the
same time, logistics enterprises has weak ability to provide quality service and
low efficiency, which cannot meet the current market demand or meet in the way
the market needs.
2.2 The Information Flow in Supply Chain

The value of information technology application and information sharing among
upstream and downstream enterprises in supply chain is widely concerned and
recognized by scholars and enterprises. From 2013, the National Bureau of Sta-
tistics released information technology application and e-commerce situation
according to enterprise types, indicating that the application of information tech-
nology has increasing impact on China’s economy. Table 2 suggests the enterprise
information and e-commerce situation in 2014. The proportion of enterprises
with websites in manufacturing and transportation industry as well as ware-
housing industry were separately 70% and 45%, but enterprises dealing with
e-commerce accounted for only 8.6% and 4%. The website is mainly used for
non-transactional corporate propaganda and product display.
Logistics information technology is divided into four categories: data acqui-
sition technology (such as bar code, RFID technology), information technology,
warehousing management technology and transportation management technol-
ogy, and respectively the proportion of technology used by Chinese logistics
enterprises is 56%, 94.9%, 73.8%, 88.5% [7]. According to the market research
report on China’s logistics industry in 2015, the utilization rate of logistics
applications are as follows: storage management system accounts for 17.21%,
transportation management system for 17.21%, and vehicle tracking system for
978 W. Hua
Table 2. Information technology application and E-commerce of enterprises in 2014
Project Number of Number of Number of Number of Dealing with E-commerce

enterprises computers enterprises websites
used per with owned by
100 websites per 100
workers enterprises
Number of Percentage (%)
enterprises
Manufacturing 347376 18 244822 70 29938 8.6
Transportation, 35187 25 15908 45 1402 4
warehousing,
postal service
Source: National Bureau of Statistics et al., China Statistics Yearbook (http://www.stats.gov.cn/tjsj/ndsj/
2015/indexch.htm)
19.35% [3]. Application of logistics information technology and the software does
not mean that the supply chain has achieved information sharing, collaborative
response and quick response. Only by breaking the blocked business model and
interest chain established by the system and institution can we cross the bound-
aries of organizations, regions and industries.
2.3 Fund Flow in Supply Chain
On the whole, the financing difficulties are the bottleneck of enterprise devel-
opment for many SMEs. With the concept of supply chain management is
strengthen, the overall foundation of the supply chain is increasingly appar-
ent. This development has gone through three stages. The first stage, logistics’
factors combine with financial activities. The second stage, the development of
supply chain finance moves from factors to integrating management of logistics
operation, business operation and financial management. The third stage, inter-
net + supply chain finance relieve information asymmetry and reduce the costs
of information access and information processing, and the industry supply chains
are optimized through financial resources [10]. The Supply chain finance takes
the core enterprise as the starting point, and the systematically arranges financ-
ing for all member enterprises in the supply chain. Supply chain finance inputs
the funds into the small and medium-sized upstream and downstream enterprises
in the supply chain, which provides the financial support for the supply chain.
The application of Internet achieves financial support for supply and demand,
that is, the third-party payment platform reduces financing costs for enterprises
in supply chain, and consequently the bottleneck of funding is alleviated.
3 The Optimization of Supply Chain
The essence of “Internet plus” is to online and digitalize the traditional indus-
try. When the commodity display and transaction behaviors shift to Internet
and online, information and data can flow, exchange, mine, transfer and share.
Besides, supply and demand can be connected at the lowest cost. However,
the enormous information and data are featured with scattered property, low
value intense, low connection. In this case, big data, cloud storage, data mining
and other technologies are adopted to provide valuable information for decision-
making by data integration, processing and analysis.
The integration of Internet and industry moves forwards generally from the
downstream to the upstream in the supply chain and changes consumers in the
first place. According to the Internet Statistical Report of 2015, the popularity
rate of Chinese netizens is 50.3%, among which there are 413 million, that 60.03%
of the netizens [4] making up for 30.20% of the total population conduct online
shopping and transactions (26.68% in 2014). Reports [1] have shown that China’s
online retail market has 3.8 trillion yuan volume, which accounts for 12.7% of
total retail sales(10.6% in 2014), represent the proportion of retails that applying
Internet. The application rate of E-commerce, from the consumer → retail →
production → transport services along the supply chain, are respectively 26.68%,
10.6%, 8.6%, 4.0% in 2014 as shown in Fig. 1.
Fig. 1. Application rate of E-commerce in different links of supply chain (Year 2014)
3.1 Demand Side Management-Define and Expand Effective

Demands
In the open network environment, the demand of the increasing amount of end
users participating in innovation and revolution must be given prominent atten-
tion by the supply side and placed in the logic starting point and attention
center of enterprise decision making. Otherwise, once the consumers’ upgrading
needs are not responded, then the low-cost, convenient external resources are
their alternatives, which results in decrease of internal demand.
(1) Demand side management-Define and expand effective demands
Diversification and uncertainty of demand reduce the versatility of the product
and increase the inventory of supply side and make production and management
more difficult. Thanks to the application of Internet information technology, it
reduces the cost of collecting, integrating and utilizing consumer information,
thus the enterprise is sensitive to marketplace and explores and meets the end-
user demand in more convenient, fast, efficient and low cost way. What’s more,
980 W. Hua
business to customer model achieves disintermediation effect, which greatly

weakened the profits divided by the distributors, retailers and other traditional
business roles in the supply chain and is conducive to get access to first hand
market information because supply chain is directed to end users. Data analysis,
cloud computing and other tools utilize the multi-dimensional information con-
nect by the internal and external network platform to dock, integrate, analysis
trend. Enterprises are enabled to increase supply chain forecasting and decision-
making capabilities and conduct scientific regulation of production, inventory,
logistics and other key activities in supply chain.
(2) Exploit and foster consumer market by innovative activities
Efficient supply chain must be the interaction, feedback and balance between
efficient supply and demand. The supply side should strengthen understanding
and awareness of the market demand and positive response, maintaining this
market demand is true, reliable and sustainable. Supply chain finance provides
credit support for a large number of SMEs in the supply chain, and consequently
the consumer finance of supply chain has emerged. The typical representative
such as Alibaba, Jingdong Mall and Suning use their own funds to provide online
supply chain financial services, among which installment and other shopping
financing services increase the consumption by 30%–50%.
By fully utilizing these trading platforms, supply side sells in domestic and
overseas markets as well as expands sales channels at a lower cost. Owing to
the strategic cooperation with the e-commerce, large-scale production enter-
prises combine the market development, capital operation and brand upgrading
together, improve the ability to meet consumption experience based on Internet
and actively respond to market demand for quality products and services. These
efforts made by supply side improve brand awareness and acceptance and con-
sequently improve the company’s position in the value chain. What’s more, the
global business opportunities in supply chain are added.
(3) Guide the consumers to deeply participate in supply chain activities
The emergence of E-commerce has changed the objects of orders and logis-
tics delivery from retailers to consumers. Since business opportunities are found
from the user’s selection and behavior changes, production and service enter-
prises should attach importance to user experience and customer voice to ensure
consumers have more opportunities to participate in the supply chain activi-
ties. Consumers become a crucial factor in supply chain management. On one
hand, the demand side influences the product and service design of supply chain
through order, feedback, reviews and so on. On the other hand, the supply side
can effectively analyze preference trends, behavior characteristics and potential
demands of consumers through large data analysis, and accordingly make cus-
tomized designs based on personalized needs. It means that by application of
Internet technology, the ideas of individual end consumer become reality prod-
ucts and services. The customized design innovation featuring “User-Generated
Content” is becoming the big trends [6].
The visual distribution of E-commerce Suning, a kind of offline services
assisted with the “Internet plus”, brings great fun for users and enhance the
user experience by giving a sense of involvement and control. Now a lot of out-
standing enterprises are marching in deepening their involvement in the supply
chain management. They not only collect customer information, but also focus
on guiding consumers to participate in collaborative intelligent development of
product development, customized products, and experience processes. Enter-
prises are expected to improve the ability to satisfy customer needs based on the
Internet and enhance the quality of supply capacity.
3.2 Supply Side Management-Improve Quality Supply
Supply and demand coexist. The supply side reform means that we shouldn’t
only focus on demand stimulation and it should be emphasized that the economic
growth can’t simply depend on stimulating the demand. In the long run, the main
supporting factor lies in effective supply’s response to high-quality demand and
guidance. (Jia Kang, 2015) [5] While, the supply side reform should utilize the
production factors including the labor, land, resources, capital, technology to
achieve the intensive production model characterized by economic innovation
and efficiency, and to achieve the upgrading from low-level supply to high-level
supply through knowledge, technology, products and institutional innovation.
(1) Data Analysis Transformation-Big Data Analysis
With the speeding up of convenient Internet and wide application of new tech-
nologies, the consumption patterns of end-users have changed and demand
expectations are becoming diversified and personalized. In this case, manufactur-
ing enterprises should quickly respond to external changes and effectively mine
the data potential in order to solve supply chain management issues such as
demand forecasting, replenishment forecasting, inventory optimization, foundry
collaboration, production arrangements and supplier management. Besides, the
manufacturing enterprises should monitor and manage the process and results
including process tracking, exception reminder, order tracking, transaction his-
tory data and so on. The solution is the big data analysis which is characterized
by dynamic real-time, multi-dimension, public sharing and low cost. Based on
the connection of data, big data can help all the node enterprises in the supply
chain to watch and control every activity including product design, purchas-
ing, manufacturing, order processing, logistics, financial services and consumers.
Besides, it can display inventory, production, orders, circulation, distribution
and other activities in real time, and assist the enterprises to optimize, forecast
and adjust the balance of supply and demand.
The production systems can cope with the uncertainty brought by the rapidly
changing environment. By application of information technology and Internet,
the production system can do as follows: 1 Grasp the true needs of the market.
Process the fragmented information through big data, cloud computing and other
information technology, thus the demand trends will assist decision-making; 2
Improve production flexibility. It means to promote the production management
that quickly responds to multiple products, small batch and rapidity. The “rigid”
production mode of large-scale production transforms to the adaptable and
“flexible” production mode; 3 Integrate external advantage resources timely.
982 W. Hua
The enterprises in supply chain can share information at a lower cost through
the information platform.
(2) Manufacturing and production transformation-Quick response
O2O model, characterized by small batch, multiple species and fast response, has
become a popular market demand. The “Internet plus” mode accelerates this
consumption and further improve the consumer expectations. The changes are so
fast that manufacturing enterprises are confronted with the following situations:
1 Due to poor versatility of personalized products, inventory risk is greatly
increased. 2 As the technology and product updates speed up, the marketabil-
ity value of the inventory and performance are reduced, bringing huge economic
losses. Therefore, it is necessary to adjust the established production manage-
ment, system arrangement, equipment selection, logistics and warehousing busi-
ness pattern that targeted for large-scale production, instead, to transform to
flexible manufacturing by full application of Internet technology.
From perspective of production management, it is necessary to adjust the
production arrangement timely, quickly and frequently and adopt a flexible and
diversified combination of production factors. Besides make full use of internal
and external resources advantages to prevent excessive production and meet the
market demand at the same time. From the perspective of production equip-
ment, various artificial intelligence tools, manufacturing equipment and com-
puting methods are widely used in the manufacturing process. Manufacturing
intelligence is fully utilized in the scheduling, design, processing, control, plan-
ning and other aspects of the production. Manufacturing enterprises must com-
prehensively utilize manufacturing technology, management science, computer
science, information technology, and conduct comprehensive processing of inter-
nal and external information and reasonably schedule the procurement, logistics
and production. The flexible production system based on information technology
can respond to the changing production and sales, and reduce response time to
changes, besides, production costs can fluctuate with the amount fluctuation.
(3) Structural transformation of supply side-Provide quality supply
On one hand, the overcapacity is prominent when shortage economy is shifting
towards surplus economy in China. On the other hand, as Chinese per capital
income continues to grow, the basic needs upgrade to development demands and
the demands are diversified and personalized. In the Internet era, consumers
have more choices in products, service, consumption channels and manners. The
popularity of overseas purchase heat demonstrates that people welcome contin-
uous protected products of high quality and fair, just and honest service. When
consumers are dissatisfied with the products and services, they turn to lower cost
channels as alternatives. Therefore, the effective way to resolve overcapacity and
overstocked products is to change the export-oriented supply system to improve
the effective domestic demand, provide favorable environment for domestic high-
quality supply enterprises to ensure supply keep ups with the pace of demand.
Expanding domestic demand has always been an important means of China’s
economic development. In practice, domestic demand growth comes mainly from
investment as the contribution of terminal consumers is limited. Supply side
structural reform, in essence, is the effective supply and utilization of production

factors. In particular, it gives full play to the role of capital and technology
factors. It means we should introduce advanced management ideas, improve
management capacity, reduce costs, eliminate backward production capacity,
get rid of low-cost competition, improve supply quality, and create space for the
upgrading of industry and supply chain.
To start with the supply side, it is necessary to resolve the contradiction
between consumer demands for quality products and services and the existing
low-end products provided by production and service enterprises. “Internet plus”
information technology reallocates the resources and provide a strong platform
to eliminate the information asymmetry caused by mismatched contradictions.
3.3 Enhance Logistics and Financial Services Capabilities

The modern logistics industry provides material flow service for collaborative
activities of enterprise in supply chain node, and improves the interconnection
and interdependence among the supply chain nodes. The financial innovation
in supply chain on this ground provides the financial support for supply chain
development. However, the high financing and logistics costs have restricted
the competitiveness of China’s industries, affecting the efficiency of the entire
industrial structure. Therefore, China’s supply side reform not only involves the
production field, but also includes the reform on the transaction behaviors, sys-
tems and circulation and service areas among the manufacturers, agents, dealers,
retailers in the supply chain.
(1) Integrate the fragmented logistics services and improve logistics supply capac-
ity and level
The development direction for the logistics industry to enter into network and
organized service is to integrate resources, expand logistics enterprises scale,
improve the introduction and application of information technology to provide
one-stop service for consumers, expand the logistics services scope and provide
value-added logistics services. The logistics platform gathering large sums of fund
and technology will become the breakthrough of logistics industry integration.
Various forms of logistics network platform coordinate the goods and capac-
ity, integrate scattered transport resources and sources of goods and improve the
efficiency of material circulation. When the supply chain efficiency is low, the
supply of raw materials cannot make timely production. Then most of the manu-
facturers will prepare front-end raw material storage and back-end finished prod-
ucts storage, it might cause high inventory costs and overstock inventory. With
the development of the logistics platform and popularization of smart phones
and networks, the highway logistics APPs have emerged. These APPs help
obtain real-time logistics information conveniently, breaks information asym-
metry, matches freight resources effectively and reduce resources wasting. Many
small micro-logistics enterprises link the supply side and demand side through
the supply chain. With the supply chain information platform, theses compa-
nies can rapidly accumulate massive data, accurately research into demand and
participate in supply chain services in real time.
984 W. Hua
(2) Overall penetration of financial innovation on “Internet plus”

To provide financial guarantee for the supply side reform, it is necessary to stim-
ulate innovation through policies, improve the total factor productivity, rational-
ize the imbalance of supply and demand structure, make the capital flow to the
real economy and optimize the allocation of financial resources. Financial inno-
vation should have all-round penetration on “Internet+” in the economy and
living. Supply chain finance means that regarding with the real business context
and the credit level of the core enterprises of the supply chain, the financial
institution provides credit support to small and medium-sized enterprises in the
upstream supply chain and consumer finance for downstream consumers. Since
core enterprises are the backbones of the supply chain system, they must estab-
lish an efficient resource sharing system to provide a sharing platform, maintain
membership cooperation, find value creation methods and finally improve the
competitiveness of the whole supply chain system.
4 Conclusion
The wide coverage and availability of the Internet has transformed the sup-
ply chain from a single, linear, discontinuous information flow into a complex
information network, which increases the possibility of choosing excellent sup-
ply chain partners and expands the access of quality products and service for
consumers. The impact of “Internet plus” on the supply chain, on one hand, is
reflected in the application of information technology and big data, by which
companies achieve real-time information, data collection and analysis, result-
ing that business decision-making is no longer simply based on structured data;
on the other hand, the impact is reflected in the changes of production mode
and the circulation ways. The Internet plus promotes the development of multi-
industry convergence and transformation and upgrading of traditional industries,
and fundamentally reduces the contradictions between high inventory and low
consumption satisfaction as well as between overcapacity and sluggish manufac-
turing.
The transformation process from the traditional supply chain management
to the “Internet plus supply chain” is not the simple application of informa-
tion technology to supply chain management, but the conversion and innovation
of information technology from subordinate roles to the dominant rules, which
promote customers’ true demands capturing, supply chain visualization process,
agile manufacturing, real-time data acquisition and sharing, and partner collabo-
ration, in the environment of continuous technology development of internet, big
data and cloud computing. A reviewed way of thinking, about market, custom,
production and enterprise value even the business ecosystem, is revealed.
Acknowledgements. Sichuan Province Key Research Base of Social Science—

Sichuan Province Cycle Economy Research Center, “Ability Evaluation Research of
the County Economy Development from the Perspective of Cycle Economy” (XHJJ-
1610); Sichuan Province Key Research Base of Humanities and Social Sciences–Sichuan
Province Education Development Research Center of China West Normal University,
“Adaptability Research between Higher Education Subject Distribution and Industry

Cluster in Sichuan Province” (CJF16007).
References
1. CEBRC (2016) Online retail market data monitoring report in China in 2014
[DB/OL]. China E-Business Research Centre. http://www.100ec.cn
2. CFLP (2016) statistical investigation report on logistics of key enterprises in China
in 2015 [DB/OL]. China federation of logistics & purchasing, China logistics infor-
mation center. http://www.clic.org.cn/wlxxwlyx/262043.jhtml
3. CFLP (2015) China logistics development report (2014–2015). China Federation of
Logistics & Purchasing. China Logistics Association. China Fortune Press, Beijing
4. CINIC (2016) Statistics on the development of the 37th China internet network
[DB/OL]. China Internet Network Information Centre. http://www.clic.org.cn/
wlxxwlyx/262043.jhtml
5. Jia K, Shu J (2015) Economic ‘new framwork’ and ‘new supply’: the pursuit of
important link and epitomize in innovation. Sci Technol Prog Policy 11:8–14
6. Li J, Liu C, Zhang N (2016) Anlyzing of design process and operational mode of
different crowd-sourcing supply chains with internet plus. Sci Technol Prog Policy
11:24–31
7. Lin CY (2008) Determinants of the adoption of technological innovations by logis-
tics service providers in China. Int J Technol Manage Sustain Dev 7(1):19–38
8. NDRC (2016) The national logistics operation status briefing in 2015 [DB/OL].
National development and reform commission. http://www.sdpc.gov.cn/jjxsfx/
201605/t20160531 806056.html
9. NDRC (2016) The world bank logistics performance index in 2016 [DB/OL]. China
Logistics Information Center. http://www.clic.org.cn/kjwl/270223.jhtml
10. Song H, Chen SJ (2016) Development of supply chain finance and internet supply
chain finance: a theoretical framework. J Renmin Univ China 35:10–16
Assessing the Recycling Efficiency of Resource
in E-Waste Based on MFA of Reverse Logistics
System
Minxi Wang, Xiaoling You, and Xin Li(B)
College of Management Science, Chengdu University of Technology,

Chengdu, The People’s Republic of China
Lixin2012@mail.cdut.edu.cn
Abstract. The waste electrical and electronic equipment (WEEE, e-

waste) has caused many issues, such as waste of resource and environmen-
tal pollution. This paper applied material flow analysis (MFA) method to
analyze the flows of recycled resource from WEEE and process of reverse
logistics from recycling to reuse. And analyzed the discarded TV data in
K company (a leading television producer in China), and calculated the
recycling efficiency of metals. The research showed that recycling effi-
ciency of the metals from discarded TV in K company is 12.3%, which
is far below the criterion (In Japan, the “Household appliance recycling
and reuse” recycling efficiency of useful resource from discarded televi-
sion by the television manufacturers must be above 55%). At last, we
gave some suggestions for developing recycling efficiency of manufactur-
ing enterprises and recovery enterprises in China: (1) Improving the recy-
cling efficiency of manufacturer, retailer and consumer can contribute to
improve the recycling efficiency in the whole reverse logistics system; (2)
Improving the technological level of decomposition and dismantling of
e-waste in the recycling process can increase the recycled material; (3)
Manufacturing enterprises should have the green production concept and
the participants in the supply chain should follow the concept of EPR
(Extended Product Responsibility).
Keywords: Reverse logistics · Recycling efficiency · MFA · WEEE
1 Introduction
WEEE is the waste of electrical or electronic products, it covers a wide spectrum
ranging from consumer goods such as discarded refrigerators, air conditioners,
washing machines, televisions and computers, mobile phones to capital goods
such as some unqualified products, parts and scraps in the production and repair
processes [6].
Because of containing precious metals (such as gold, copper, aluminum and
silver) and plastic, the reutilization of WEEE has magnificent recovery value and
Assessing the Recycling Efficiency 987
environmental protection advantages. Recycling resource from the WEEE has

long-term strategic significance, not only can reduce the waste of resource, pro-
mote social development and regional economic growth, but also can reduce the
serious threat from toxic and harmful substances in e-waste to the environment
and human health.
Many researchers have made a large amount of researches on the recycling in
reverse logistics system and e-waste recycling. In the aspect of reverse logistics
system mainly focused on architecture design for the network, model establish-
ing and waste management in reverse logistics system. Reverse logistics network
can be divided into different branches according to the type of goods in reverse
logistics system, applying the mixed-integer linear program (MILP) or integrat-
ing the analytical hierarchical process (AHP) can address the complex network
configuration of a reverse logistics system [2,9,11]. In order to design the network
of reverse logistics system, the capacities of each nodes and the number of recy-
cling facilities should be considered. By applying some models can minimize the
costs, and simultaneously can consider green and social issues to design a reverse
logistics network [10]. In addition, establishing the model of reverse logistics
also can solve some practical problems, for instance Ehab Bazan established the
model based on the economic order/production quantity (EOQ/EPQ) and the
joint economic lot size (JELS) settings for reverse logistics inventory systems [5].
Waste management is an important part in reverse logistics system, and waste
recycling is a multi-disciplinary problem, a holistic view and viewpoints from
different decision levels should be considered when modelling a reverse logistics
system for waste management [7]. Setting up the waste management strategies
can help firms to reduce waste and have a healthy reverse logistics system [15].
In the aspect of e-waste recycling, developing model and applying method can
analyze the WEEE based on a real world case. For instance, Ayvaz, B created the
two stage stochastic programming model provides acceptable solutions to make
efficient decisions for reverse logistics network design of e-waste [4]. By applying
the material flow analysis method can make a detailed analysis for recycling and
stock of e-waste, as well as the recycling efficiency assessment and optimization
in the closed-loop supply chain [1,14].
This paper combine recovery system with material flow analysis method in
reverse logistics system and analyze the process of resource recycling of e-waste.
Take recovery of discarded televisions of K company (a leading television pro-
ducer in China) as an example in this thesis, its recycling efficiency of metal in
the discarded television is calculated and assessed, based on that we gave some
suggestions for the manufacturing enterprises and recovery enterprises in China.
2 Theory and Methodology

2.1 Reverse Logistics and Reverse Logistics Network
A complete definition of reverse logistics is put forward by the European Working
Group on Reverse Logistics, and is formulated as follows: “The process of plan-
ning, implementing and controlling backward flows of raw materials, in-process
988 M. Wang et al.
inventory, packaging and finished goods, from a manufacturing, distribution or

use node, to a node of recovery or node of proper disposal” [3].
Reverse logistics network is the nodes distribution and the arrangement of
transportation routes between each node in the reverse logistics system. Typical
reverse logistics network is shown in Fig. 1.
Fig. 1. Typical reverse logistics network
2.2 Material Flow Analysis

Material flow analysis (MFA) is “a systematic assessment of the flows and stocks
of materials within a system predefined in space and time”. The method mainly
Fig. 2. The basic framework of material flow analysis method

involved the material inputs, outputs, and inventories [8]. According to the law
of conservation of mass, MFA could get an ultimate conclusion that indicates the
status of material input, stock and output in each stage of life cycle. It is based
on the material balance theory (Total material input = Total material output
+ Stock), by means of material flow analysis, we can estimate and evaluate the
condition of development, utilization and waste management of natural resources
and various substances. The basic framework of material flow analysis method
is shown in Fig. 2.
2.3 Material Flow Analysis of Reverse Logistics System

MFA of reverse logistics system is the main methodology in this paper which is
a way to analyze the material flow of e-waste in reverse logistics system.
Fig. 3. The material flow analysis frame of the reverse logistics system
As shown in Fig. 3, participants in the frame are recycler (which mainly

responsible for four recycling process: detection and maintenance, dismantling,
decomposition and innocent treatment), manufacturer, retailer and consumer.
3 Analysis and Discussion

3.1 Reverse Logistics Process for the WEEE
As a special kind of reverse logistics, the recovery process for e-waste includes
five aspects. Figure 4 shows the principal stages in the recovery process.
(1) Recovery
Recovery phase is the first and most important stage in the whole recovery
process. At present, the main sources and recovery channels of e-waste in China
990 M. Wang et al.
Fig. 4. Recovery process for the WEEE
Table 1. The main sources and recovery channels for e-waste
Category Comment
Main source The daily life of the residents;
Electronic products and household appliances manufacturers
and sellers;
Illegal imports of e-waste from abroad.
Recovery channel Door-to-door recovery;
For trading-in old one for new one in the Shop;
Flea Market of electronic products and household appliances;
Garbage recycling by government/The glean and collect scrap
man;
After-service station
(2) Detection and Classification

E-waste has a wide variety of categories, and each category can be divided
into small classes and the corresponding value of it is different. Therefore, the
work that must be done is rigorous detection and classification for the recycling of
e-waste according to their characteristics such as structure or components func-
tion to determine the subsequent process, such as after maintenance to resale,
reuse the parts after dismantling, material recycling after decomposition or waste
treatment.
(3) Reprocessing
In the reprocessing stage, e-waste can be reacquired the value that reuse,
remanufacture or recycle through reasonable reprocessing for recycled products
or parts. Reuse only aims at the recycled products can be used directly through
cleaning or simple maintenance, such as the toner cartridge which can be used
again through the simple work. Remanufacturing aims at the e-waste can enter
the manufacturing stage again after dismantling, replace or repair. And the
recycle aims at the parts, glass and plastic in the recycled e-waste.
(4) Redistribution
The recycled e-waste can be as commodity to reuse and enter the consumer
market or donate to the consumers in poor areas directly after the inspection
and reprocessing stage. In this stage, distribution is the most important work,
which can make the entire recovery process for the WEEE operate efficiently.
(5) Waste Treatment
For the e-waste that has no circular economic value or can bring large harm to
the environment, the available material in them can as the recycling resources
to be reused through the reasonable treatment, such as dismantling, melting,
refining and electrolysis. While for the e-waste cannot be reused, the disposing is
partitioned in two ways, one which is stored them in landfills; the other approach
is sent them to an incinerator.
3.2 MFA in Reverse Logistics System for E-Waste

(1) Date Acquisition
In this paper, according to the MFA of reverse logistics system, we built a
reverse logistics system for e-waste based on the self-operated recovery system of
the K television manufacturing company in China. Take the recycling efficiency
of the metal resources (copper, aluminum, iron) of the company as the concrete
calculation example. Through collecting television sales data of the company
over the years and applying the market supply method A [12] to forecast the
amount of discarded TV of the K television manufacturing company in 2016.
Through the existing research [13] can get the duration of use and the waste
rate of the television is shown in Table 2.
Table 2. The duration of use and the waste rate of the television
Duration of use(year) < 8 8–9 9–10 10–11 11–12 12–13 13–14 14–15 15–16 > 16
Waste rate 6% 6% 10% 13% 15% 15% 13% 10% 6% 6%
The sales date of K television manufacturing company over the years is shown
in Table 3.
The market supply method A formula as follows:

Q= S i × Pi , (1)
992 M. Wang et al.
Table 3. Sales of television over the years
Year Sales (Million)

2000 1.06
2001 1.09
2002 1.63
2003 2.41
2004 2.43
2005 3.99
2006 3.99
2007 3.54
2008 4.01
2009 5.8
2010 6.02
2011 6.41
2012 7.56
2013 8.05
2014 9.67
2015 9.92
where, Q represents the amount of discarded televisions a year, Si represents the

sales of televisions i years ago (assuming all the sales of televisions are used), Pi
is the waste rate represents the proportion of the discarded television in sales i
years ago.
By using this model, it can be concluded that the amount of waste television
in 2016 is 2.12million. Combining with the relevant information from the related
industry, 50% of discard television can be recovered, and the dismantling rate
(after dismantling can be recycled parts ratio) is set to 35%, and the decompo-
sition and utilization rate is set to 56%. The average weight of a television is
20 kg, the content of copper in a television is 5.4%, the content of aluminum in
a television is 5.4%, and the content of iron in a television is 5.3% [18]. Because
the resource loss of recycled metal to the environment in each stages, and in
order to facilitate the research, so the following assumptions are given:
(1) The structure of resource recycling in reverse logistics system is a closed
loop network, all the discarded television after disposing would return to
manufacturer;
(2) The return production from consumer must be sent to the manufacturer for
processing;
(3) The product could be reused after detection and maintenance would be
resold and enter the consumer market;
(4) All the discarded television (which could be recycled) must be recycled by
the recycler.
Therefore, the amount of recycled metal from discarded television is 3413.2

tons. The specific data are shown in Fig. 5.
Fig. 5. The specific data in reverse logistics system of discarded television. All units
measures in tons
(2) Empirical Analysis

According to the data in Fig. 5, the input-output table of reverse logistics
system of discarded television is shown in Table 4.
Table 4. The input-output table. All units measures in tons.
P1 P2 P3 P4 P5 P6 P7 P8 I0
P1 0 0 0 0 0 0 0 0 3413.2
P2 3413.2 0 0 0 0 0 0 0 308.4
P3 0 2639.8 0 0 0 0 0 0 0
P4 0 0 1028.1 0 0 0 0 0 823.5
P5 0 0 923.9 1036.9 411.8 274.2 0 0 754.5
P6 0 0 0 0 2302.3 0 203.2 0 0
P7 0 432.3 0 0 0 2231.3 0 0 0
P8 0 649.5 687.8 814.7 687.2 0 0 0 0
O0 0 0 0 0 0 0 2460.4 2839.2 0
994 M. Wang et al.
In order to quantify the relationship of various stages of metal cycle in reverse

logistics system, we give the mathematical expression of input-output relation-
ship based on material balance theory as Eqs. (1) and (2), xk is the total flow
in Pk , fkj is the flow of Pj to Pi .
n

xk = fkj + Ik k = 1, 2, · · · n, (2)
i=1
n
xk = fik + Ok k = 1, 2, · · · n. (3)
i=1
According to the Eqs. (2) and (3), the total flow of each node is shown in
Table 5. In order to assess the overall recycling efficiency of the system, we give
the mathematical expression Eq. (4)–(6).
Table 5. The total flow of each node
Node Amount (tons)

P1 3413.2
P2 3721.6
P3 2639.8
P4 1851.6
P5 2989.5
P6 2505.5
P7 2663.6
P8 2839.2
n
Ii + fij
ni = j=1
n , (4)
oi + j=1 fji
ni − 1
Ri = , (5)
ni
where, ni is the relation of the input and the final output of a single node. When
ni = 1 indicates that the recycled metal resources through the node i, but not
returned to the node i for recycling. When ni > 1 indicates that the recycled
metal resources through the node i, and returned to the node for recycling with
a direct or indirect way.
ni is recycling efficiency of a single node. When Ri = 0 indicates the flow of
the recycled metal resources through the node i is unidirectional. When Ri > 0
indicates the recycled metal resources through the node i can return the node
for circulating and using.
Table 6. The calculation results of ni and Ri
1 2 3 4 5 6 7 8
ni 1 1 1 1 1.14 1.09 1.19 1
Ri 0 0 0 0 12.30 % 8.30% 16% 0
The calculation results of ni and Ri are shown in the following Table 6.

According to weighted method, we use the recycling efficiency of each node
can get the overall recycling efficiency C in the system.
n
sc R × xi 1001.841
C= = i=1n i = ≈ 12.3%, (6)
s x
i=1 i 8158.6
where, Sc is the total recycling flow, S is the overall flow of nodes which the
recycling efficiency of a single node is not equal to zero [16].
Therefore, the overall recycling efficiency of the system is 12.3 %. That is the
recycled metal resources after the whole recovery process, 12.3% metal resources
in discarded television can be reused.
4 Conclusions
According to the analysis and discussion above, we concluded that the recycling
efficiency of metal resources in waste television of K television manufacturing
company which is representative in the television industry, so the recycling effi-
ciency has a certain reference value for home appliance manufacturing industry
in China.
There is not a sound assessment standard to evaluate the recycling efficiency
in China, so this paper take Japanese experience as reference. In Japan, the
“Household appliance recycling and reuse” started from 2001 specifies that the
recycling efficiency of useful resource (metal in the majority) from discarded
television by the television manufacturers must be above 55% [17]. Obviously,
the recycling efficiency of the television company was hard to reach the criterion
with the present resource recycle level, so it is feasible to make optimization for
reverse logistics recycling system of e-waste in China, and improve circulation
efficiency of the recycling system, so we give the following suggestions:
(1) From the above calculation process, the overall recycling efficiency lever
mainly depends on P5 , P6 and P7 (Fig. 5). Thus, we can increase the numeri-
cal value of P5 , P6 and P7 (Table 6) to increase the overall recycling efficiency,
that means improving the flow of manufacturers, retailers and consumers is
the key problem to improve the recycling efficiency of the system.
(2) Improving the secondary utilization ratio of the material in the recycling
system, such as improving the technological level of decomposition and dis-
mantling of e-waste in the recycling process, the recycled material in e-waste
can be increased, but the input of raw material can be reduced, and then
minimize the wastage of resource from the external environment.
996 M. Wang et al.
(3) Chinese electronic products and household appliances manufacturing enter-

prises should have the green production concept to decrease the amount of
waste that can’t be reused. In addition, the participants in the entire supply
chain should follow the concept of EPR (Extended Product Responsibility).
That means not just the manufacturer, the distributor, retailer and con-
sumer should also strengthen mutual cooperation and bear their respective
responsibilities, for the purpose of promoting the level of reuse and recycling
of resources.
Acknowledgements. We would like to thank all anonymous reviewers for their

constructive comments and suggestions. This research was supported by the fund
of the Excellent Innovation Research Team at Chengdu University of Technology
(KYTD201406).
References
1. Agamuthu P, Kasapo P, Nordin NAM (2015) E-waste flow among selected institu-
tions of higher learning using material flow analysis model. Resour Conserv Recycl
105:177–185
2. Alshamsi A, Diabat A (2015) A reverse logistics network design. J Manufact Syst
37:589–598
3. Alumur SA, Nickel S et al (2012) Multi-period reverse logistics network design.
Eur J Oper Res 220(1):67–78
4. Ayvaz B, Bolat B, Aydin N (2015) Stochastic reverse logistics network design for
waste of electrical and electronic equipment. Resour Conserv Recycl 104:391–404
5. Bazan E, Jaber MY, Zanoni S (2015) A review of mathematical inventory models
for reverse logistics and the future of its modeling: an environmental perspective.
Appl Math Model 40(5–6):4151–4178
6. Bertram M, Graedel TE et al (2002) The contemporary european copper cycle:
waste management subsystem. Ecol Econ 42(1–2):43–57
7. Bing X, Bloemhof JM et al (2015) Research challenges in municipal solid waste
logistics management. Waste Manag 48:584–592
8. Brunner PH, Rechberger H (2006) Practical handbook of material flow analysis.
Int J Life Cycle Assess 10(9):293–294
9. Galvez D, Rakotondranaivo A et al (2015) Reverse logistics network design for a
biogas plant: an approach based on MILP optimization and analytical hierarchical
process (AHP). J Manufact Syst 37:616–623
10. Govindan K, Paam P, Abtahi AR (2016) A fuzzy multi-objective optimization
model for sustainable reverse logistics network design. Ecol Ind 67:753–768
11. Mangla SK, Govindan K, Luthra S (2016) Critical success factors for reverse logis-
tics in indian industries: a structural model. J Clean Prod 129:608–621
12. Simon W, Noel D, Matt C (2001) Waste from electrical and electronic equipment.
US Environmental Protection Agency
13. Song X (2008) Establishment and economic analysis of the standardized recycling
system for waste and used home applications. Doctoral dissertation
14. Su LX, Wang ZH et al. (2012) MFA-based resource cycling efficiency analysis for
a closed-loop supply chain system. Ind Eng J
15. Van Der Wiel A, Bossink B, Masurel E (2012) Reverse logistics for waste reduction
in cradle-to-cradle-oriented firms: waste management strategies in the Dutch metal
industry. Int J Technol Manage 60(1/2):96–113
16. Wu Z (2013) Reverse logistics management l the WEEE as the research objects.
Science Press in Beijing
17. Yu P (2012) An analysis of Japan’s home appliance recycling. Contemporary Econ-
omy of Japan
18. Zeng X (2008) Mechanism and technology of nonferrous metals recycling from
typical e-waste parts. Doctoral dissertation
Analysis of Supply Chain Collaboration with Big
Data Suppliers Participating in Competition
Shuai Liu and Hongchun Wang(B)
School of Economics and Management Engineering,

Beijing University of Civil Engineering and Architecture,
wanghongchun@bucea.edu.cn
Abstract. The combination of big data theory and supply chain man-
agement theory has brought the subversive change for the enterprise coop-
eration mode, the paper considered that when the big data suppliers to
participate in the competition of the supply chain, the condition of over-
all profit of the supply chain in the two cases of cooperation and non-
cooperation decision making. Firstly, the profit model of each member
enterprise is established by the demand function theory, and the whole
profit model of the supply chain is obtained in the case of decentralized
decision-making, then, the profit model for common decision making of
supply chain based on the maximization of supply chain is analyzed. After
compared the two profit models, the conclusion was obtained that when
big data suppliers join supply chain competition and one party gains the
dominant position in the supply chain, collaborative decision-making is
the key to enhancing the overall profitability of the supply chain.
Keywords: Big data · Supply chain collaboration · Competition ·

Cooperation
1 Introduction
The method of supply chain management based on mutual cooperation between
enterprises is one of the most effective business strategies adopted by enterprises
in the fierce market competition, and the rise of big data industry is injected new
blood on Supply Chain Management. Data analysis methods are applied to all
aspects of supply chain management to effectively curb the procurement short-
age, lag of logistics, production bottlenecks. There are some significant achieve-
ments in the area of big data supply chain management. Firstly, supply chain
management has been intelligent, and a new generation of radio frequency iden-
tification technology can provide more reliable data for supply chains, these
data flow in the supply chain nodes, so that enterprises can exchange informa-
tion timely, the operating mechanism of the system can be optimized and the
decision making become automated. In addition, big data has also brought dis-
ruptive changes for the supply chain logistics. At present, enterprises can already
Analysis of Supply Chain Collaboration 999
achieve real-time monitoring of logistics and route optimization in the use of big
data technologies. Transport capacity and efficiency have been greatly improved
than before. The focus of future research is to predict supply chain behavior
rely on large data. Accurate forecasting is one of the effective measures to help
enterprises save costs, and the accuracy of large data projections is not reached
by other methods.
Current research on competition and cooperation model of the supply chain
is very mature, Yang and Bai, who build up on the Hotelling model to analyze
profit-sharing contracts between the supplier and the core firm, and study the
impact of upward decentralization on the core firm and the whole supply chain
under the assumption that the supplier does not have any advantage on produc-
tion costs [12]. Wu and Han, who researched the optimal production decisions
in closed-loop supply chains consisted of two competing manufactures and one
retailer with remanufacturing costs disruptions [9]. Xu et al. [11] gave a literature
review of supply chain contracts selection and coordination based on the supply
chain under competing multi-manufactures. Shi, who studied in non-cooperative
static game and Stackelberg game optimal advertising strategy in supply chain
under master-slave, proves that relevance of the manufacturer and the retailer’s
own ads level and cost and overall supply chain advertising level and costs [3].
Strategic of partner selection of supply chain model was constructed by Xin and
others based on game theory, and get optimal strategies for select partners in
the case of incomplete information [10]. Zhang, who thought the whole supply
chain benefits and members of associated enterprise level knowledge sharing is
very important, In light of this, he builds supply chain profit distribution model
based on knowledge-sharing cost [14]. Chen et al. [1] proposed the question of
how “big data analysis affects the value creation of the supply chain”, and it was
answered through the use of dynamic capability theory as a unique information
processing capability.
Research on supply chain based on the theory of large scale data, Wu [8]
based on dynamic differential game theory constructed the dynamic corporation
model of three-level supply chain under retailer paid contract, union pay con-
tracts, with big data service providers to participate in, and discusses contract
parameters on the profit impact of supply chain. Wang and Cong, [7] who stud-
ied the cloud computing service providers involved in two-stage supply chain
coordination problem and introduced in the competition model of supply chain
revenue-sharing contract and punished the contract so as to verify that the sup-
ply chain coordination and decision. Waller [5] established a matrix model to
solve when it should be applied or avoid data analysis, theory of how to use
predictive analytics improve issues such as promoting lower total logistics costs.
Tan [4] through impedimented to data acquisition analysis, set up a correspond-
ing data infrastructure framework, the framework of internal data, available
through collection of deep mining, collection capacity required data to build a
data network and helped provide supply chain optimal decision. A systematic
framework for the SAM (Roadmap, Strategic Consolidation, and Assessment)
roadmap was roadmap to help companies avoid getting caught in a mess of data
1000 S. Liu and H. Wang
by Sanders [2]. And the concept of supply chain analysis (SCA) driven by big
data was proposed and the function, process-based, collaborative and agile SCA
Maturity Mode was established by Wang and Angappa [6].
This paper analyzes the overall profitability of the supply chain under dif-
ferent decision-making situations when the big data suppliers participate in the
supply chain competition. The supplier and manufacturer’s demand curve is
taken as the input to construct the profit model of the supplier and the manu-
facturer when the manufacturer assumes the big data cost respectively. Through
the analysis of the model, the profit model of the supplier and the manufac-
turer is obtained respectively in the decision - The overall profitability of the
supply chain. After comparing the two profit models, we get the corresponding
conclusion.
The rest of the paper is organized as follows. In Sect. 2, the problem of is
described and some assumptions is proposed. In Sect. 3, the model of decen-
tralized decision-making profit and common decision making are built. In the
Sect. 4, two models are compared and analyzed. And there is a numerical analysis
in Sect. 5. In Sect. 6, conclusions are given.
2 Problem Description and Model Assumptions
The paper considers the supply chain system as a kind of supply and demand
relationship between a supplier and a manufacturer. The manufacturer produces
a product with a fixed life cycle and mature market demand. The supplier pro-
vides raw materials. The input-output ratio of the raw material to the final
product is 1: 1. In order to solve the problem of lagging information transmission
among member firms and to improve the accuracy of market demand forecast-
ing, the third-party big data enterprise is used to help the supply chain to build
a big data decision platform. The cost is borne entirely by the manufacturer.
Because of exists of the supply chain big data platform, the information on the
supply chain is completely symmetrical. Assuming that the profit-sharing prob-
lem in cooperative decision-making has already reached a corresponding contract
between supply and demand, the paper will not consider it.
Suppose that the manufacturer’s market demand function is D(pA ) = a −
bPA + bK(t), then we can get the manufacturer’s price curve PA = ab − 1b D, for
the convenience of calculation, let ab = m, 1b = n, then there is PA = m − nD +
K(t). It is worth noting that, where m is the limit price that the market can
afford [13], that is to say, at this price, the purchase rate of the commodity is 0;
Suppose that the market demand function of the suppliers is D(pB ) = α − βPB ,
we can get PB = α α
β − β D, let β = s, β = t, then PB = s − tD, empathy, s is
1 1
the limit price that the market can afford.

The symbols involved in the article are described as follows:
D : Market demand;
PA : The price of the product produced by the manufacturer;
PB : The price of the raw materials supplied by the supplier;
GB : The prices of the raw materials supplied by the sub-suppliers;
K(t) : Big data horizontal curve;
r : Big data cost coefficient;
ΦA : The profits of the manufacturers;
ΦB : The profits of the suppliers;
ΦAB : The total profit of the supply chain when making decisions together
Among them, the big data cost is F = rt 2 D, that is, the manufacturer’s
output is positively correlated with the big data costs, the higher the output,
the greater the cost of the big data.
3 Building of the Model

3.1 Decentralized Decision-Making Profit Model
In the case of decentralized decision making, the manufacturer’s profit depends
on the market price of its own products and the price of raw materials from the
suppliers, In addition, because of the cost of big data suppliers to join the supply
chain is entirely borne by the manufacturer, so the profits are:
rt
ΦA = (PA − PB ) × D − D. (1)
2
The profit of the suppliers depends on the price of the product to the manu-
facturer and the purchase price of the raw material. Therefore, the profit of the
suppliers is:
ΦB = (PB − GB ) × D. (2)
s − PB
It can be known from the supplier’s demand function that D = t . There-
fore, its profit is:
1
ΦB = s (PB − GB ) − PB 2 + PB GB .
t
Find the partial derivative of PB in the supplier profit model and let it be
equal to 0, and then the function of raw material price provided by the suppliers
and cost is obtained:
∂ΦB 1
= [s − 2PB + GB ] = 0,
∂PB t
s + GB
PB = . (3)
2
The manufacturer’s profit formula can be known from formulas (3) and (1):
s + GB rt
ΦA = [(m − nD + k(t) − )] × D − D.
2 2
Find the partial derivative of D and make it equal to 0, the manufacturer’s

optimal production decision can be obtained:
∂ΦA s + GB rt
= (m − 2n + k(t) − )− = 0,
∂D 2 2
m + k(t) − s +2GB − rt
D= 2
. (4)
2n
The manufacturer’s profit under the optimal production decision can be
known through formulas (3), (4) and (1):
s + GB m + k(t) − s +2GB − rt
2 rt m + k(t) − s +2GB − rt
2
ΦA = [(m − nD + k(t) − )] × − × .
2 2n 2 2n
It can be known after finishing:

2
(m + k(t) − s +2GB − rt
2)
ΦA = . (5)
4n
Because of the existence of the big data platform, the information on the
supply chain is completely symmetric. Therefore, the production decision of
the supplier is based on the optimal production decision determined by the
manufacturer. Then, the profit of the supplier under the manufacturer’s optimal
production decision can be obtained through formulas (3) and (4):

s + GB m + k(t) − s +2GB − rt
ΦB = − GB × 2
.
2 2n

s + GB rt
(s − GB ) × (m + k(t) − − 2)
ΦB = 2
. (6)
4n
According to the big data platform of the supply chain, in the case of decen-
tralized decision-making, the manufacturer preferentially determines its own pro-
duction decision and puts its own decision data on the big data platform. Based
on the information obtained through the big data platform, the supplier pro-
duces the corresponding number of parts, and then the total profit of the supply
chain at this time is:
rt 2
(m + k(t) − s +2GB − 2
) (s − GB ) × (m + k(t) − s + GB
2
− rt
2
)
Φ = ΦA + ΦB = + .
4n 4n

s + GB
(m + k(t) − − rt
2 )(m + k(t) − 3s−GB
− rt
2)
Φ= 2 2
. (7)
4n
3.2 Common Decision-Making Profit Model

In the model of collaborative decision-making, suppliers and manufacturers rely
on the big data platform for effective correlation, considering the overall inter-
ests of the supply chain to jointly determine the optimal production decision.
Therefore, the function of the collaborative decision-making profit is expressed
by the original profit function of the supplier and the manufacturer:
rt
Φ = ΦA + ΦB = (PA − PB ) × D − D + (PB − GB ) × D.
2

rt
Φ = m − nD + k(t) − GB − × D. (8)
2
Find the partial derivative of D and make it equal to 0, it can be known
that the optimal number of producers and suppliers at the time of collaborative
decision-making is:
∂Φ rt
= m − 2nD + k(t) − GB − = 0,
∂D 2
rt
m + k(t) − GB − 2
D= . (9)
2n
The total profit model of supply chain can be known when the optimal pro-
duction quantity has been determined under the condition of common decision
through formula (8) and formula (9):

m + k(t) − GB − rt rt m + k(t) − GB − rt
Φ = m − n × 2
+ k(t) − GB − × 2
.
2n 2 2n
rt 2
(m + k(t) − GB − 2)
Φ = . (10)
4n
4 Comparative Analysis
Through the above analysis, it obtains the profit model of the supplier and
the manufacturer in the case of decentralized decision-making and collaborative
decision-making when the manufacturer is fully committed to the big data cost.
In the case of decentralized decision-making, suppliers and manufacturers do
not take into account the overall interests of the supply chain to maximize, but
entirely from the perspective of their own interests to make decisions. Because
of the existence of big data platform, although the suppliers and manufacturers
make decisions discretely, the transmission of information flow is transparent.
The manufacturer makes the decision first, and the supplier makes its own pro-
duction decision according to manufacturer’s production plan.
In the case of collaborative decision-making, to achieve the overall profit

maximization of the supply chain is the premise of cooperation between suppliers
and manufacturers, manufacturers take into account the supplier’s production
capacity constraints, logistics conditions and other objective factors, and then
the two sides can achieve collaborative decision-making. The overall profit model
of the supply chain is determined according to the market demand function of
both parties, so that the production decision can be made with the maximum
profit of the supply chain.
The profit model of the two cases is compared and analyzed, and then:
2
(m + k(t) − GB − rt 2)
ΔΦ =Φ − Φ =
4n
(m + k(t) − s +2GB − rt
2 )(m + k(t) −
3s−GB
− rt
2)
− 2
,
4n
rt
for the convenience of calculation, let e = m + k(t) − G − 2, then there is
(e − GB )
2
(e − s +2GB )(e − 3s −2 GB )
ΔΦ = −
4n 4n
2
(e − GB ) (2e − s − GB )(2e − 3s + GB )
= −
4n 16n
2 2 2
(s − GB ) + 4[(e − GB ) − (e − s) ]
= .
16n
According to the above it can be known that s is the limit price that the
market can afford, so there must be GB s, so, (e − GB )2 − (e − s)2 0, and easy
to proof that ΔΦ 0, that is to say the profit in cooperative decision-making is
higher than that in decentralized decision-making.
It can be known that manufacturers dominate the entire supply chain and
completely control the supply chain’s big data platform by analyzes and contrasts
the two models, but it does not mean that the manufacturer can control the
information flow entirely. Because in the case of decentralized decision-making,
manufacturers do not take into account the conditions of suppliers when making
decisions, so the interests of suppliers are damaged and lead to supply chain
breaks. In the case of cooperative decision-making, manufacturers and suppliers,
under the condition of ensure their own interests, determine the optimal number
of production according to the conditions of both factors. After reaching the
corresponding contract, the cooperation tends to be stable, which is to make the
overall profit maximization of the supply chain as a prerequisite.
The above analysis can also prove that only cooperation can make the supply
chain system to achieve the maximum profit, the joining of big data suppliers
does not change the competition pattern of supply chain. Although manufactur-
ers dominate the supply chain, choose to cooperate with suppliers will achieve a
win-win situation.
5 Numerical Analysis
In this secondary supply chain, the overall profit of the supply chain with the
2
raw material prices and market demand changes, Suppose that K(t) = t2 and
large data cost coefficient r = 0.5, consider the profit situation of supply chain
when t = 10, raw material price and market limit price change, the concrete
numerical analysis is shown in Table 1.
Table 1. Profitability of raw material price and market time price change
m s GB Φ Φ ΔΦ Rate of profit increase

35 25 10 812.5 1314.1 501.6 61.80%
35 25 15 820.3 1139.1 318.8 72.00%
35 30 10 664.1 1314.1 650 50.50%
35 30 15 675 1139.1 464.1 59.20%
40 25 10 962.5 1501.6 539.1 64.10%
40 25 15 970.3 1314.1 343.8 73.8%
40 30 10 801.6 1501.6 700 53.30%
40 30 15 812.5 1314.1 501.6 61.80%
It can be seen from the numerical analysis that when the price of raw materi-
als is kept constant, the market limit price of the supplier is unchanged and the
increase of the market limit price of the manufacturer will increase the overall
profit of the two supply chains; The limit price of manufacture does not change
and the increase of the market limit price of supplier has no effect on the overall
profit of the supply chain. In addition, from the table can clearly see the com-
mon decision-making big data supply chain profits than traditional supply chain
profits increase of at least 50.5%.
6 Conclusion
Big data theory has spawned many new service and industry models, and it
is challenging for traditional industries to use big data to change their supply
chains operating. How big data can be used to implement change, what should
be targeted to gain competition advantage is the first issue enterprise should
consider.
The model constructed in the paper focuses on the competition problem
brought by big data suppliers join to the supply chain, and examines the supply
chain cooperation mode from the profit perspective. It is not sharing big data
cost by manufacturers and suppliers in the model, which is a common phenom-
enon in practice. In order to gain a dominant position in the supply chain, one
enterprise chooses to bear the cost of big data alone to achieve the control of big
data platform. But it does not means that the enterprise will be able to achieve
an increase in profits, partly because of the cost of containment, on the other
hand the stability of cooperation could be impacted after enterprise obtaining a
dominant position, and the overall profit of the supply chain maybe reduced in
that condition. Therefore, the supply chain member companies choose to coop-
erate with each other and make decisions collaboratively is the key to improve
the overall profit of supply chain through introducing the third-party big data
providers. The analysis of the article is also just to verify this conclusion.
Acknowledgements. This work was supported by the National Natural Science

Foundation of China under Grant No. 61472027.
References
1. Chen DQ, Preston DS, Swink M (2015) How the use of big data analytics affects
value creation in supply chain management. J Manage Inf Syst 32(4):4–39
2. Sanders NR (2016) How to use big data to drive your supply chain. Calif Manage
Rev 58(3):26–48
3. Shi KR, He P, Xiao TJ (2011) A game approach to cooperative advertising in a
two-stage supply chain. Ind Eng J 12:6–9
4. Tan KH, Zhan YZ et al (2015) Harvesting big data to enhance supply chain inno-
vation capabilities: an analytic infrastructure based on deduction graph. Int J Prod
Econ 165:223–233
5. Waller MA, Fawcett SE (2013) Click here for a data scientist: big data, predictive
analytics, and theory development in the era of a maker movement supply chain.
J Bus Logistics 34(4):249–252
6. Wang G, Gunasekaran A et al (2016) Big data analytics in logistics and supply
chain management: certain investigations for research and applications. Int J Prod
Econ 176:98–110
7. Wang HC, Cong JJ (2016) Study on coordination of cloud computation service
supply chain under joint contracts. Logistics Technol 5:157–160
8. Wu C, Zhao DZ, Pan XY (2016) Comparison on dynamic cooperation strategies
of a three-echelon supply chain involving big data service provider. Control Decis
7:1169–1171
9. Wu H, Han XH (2016) Production decisions in manufactures competing closed-loop
supply chains under remanufacturing costs disruptions. Comput Integr Manuf Syst
4:1130–1140
10. Xin L, Jia Y (2011) Supply chain strategic partner selection based on game theory.
Syst Eng 4:123–126
11. Xu L, Qi XU, Liu X (2015) A literature review on supply chain contracts under
competing multi manufacturers. Technoeconomics Manage Res 7:3–7
12. Yang DJ, Bai Y (2015) Supply chain core firms competition and decentralization
based on hotelling model. Proc Natl Acad Sci 102(4):1157–1162
13. Zhang A (2012) A comparative analysis of several demand function. J Jilin Province
Inst Educ 10:153–154
14. Zhang TF, Zhang YM (2011) Analysis of knowledge sharing in supply chain node
based on game theory. Stat Decis 17:184–186
Fuzzy GM(1,1) Model Based per Capital
Income Predicted of Farmers in the World
Natural and Cultural Heritage Area: Take
Leshan City for an Example
Zhiyi Meng1,2(B) , Caiyang Xu3 , and Hongxing Lan1,2

1
Business College,Sichuan Agricultural University,
zhiyimengscu@sina.com
2
Southwest Poverty Reduction Center,
Sichuan Agricultural University, Chengdu 610065, People’s Republic of China
3
Business College, Sichuan Agricultural University,
Abstract. The per-capita income of farmer is an important technology

index which is the measured level of agriculture development and the
status of farmers’ life. This paper aims at developing a new fuzzy grey
predicting model to predict the per-capita income of farmer. The fuzzy
possibility is used to check the error of the proposed model, and the test
results show that the accuracy of the model is quite high, providing a
scientific basis for policy makers.
Keywords: The per-capita income of farmer · Fuzzy grey prediction

model · Fuzzy possibility
1 Introduction
In the past few decades, China has experienced rapid economic growth and
significant reduce of poverty [3]. However, at the same time, a lot of contradic-
tions and problems have exposed. In recent years, the problem of slow growth
in farmers’ income has become the focus of the society from all walks of life.
The difficulties faced by the current farmers’ income growth have become a big
obstacle to the virtuous cycle of the whole national economy. Therefore, the
study of farmers’ income growth is of far-reaching significance.
Since the reform and opening, China has carried out a series of significant
reform in the countryside. In 1978, the rural reform which is market-orientation
is a historic turning point of the agricultural development of China: not only
break the bondage of the traditional system but also promoted the growth of
farmers’ income greatly. From 1978 to 2008, the per capita net income of farmers
increased from 134 to 4761 [1]. In particular, over the past 30 years, the changes
1008 Z. Meng et al.
of the growth of farmers’ income in China are as follows. With the inflation
factors deducted, from 1978 to 1985, it is growing at 15.2% yearly; from 1986
to 1991, at 2.7% yearly; from 1992 to 1996, at 5.6% yearly. Besides, after 1980,
1996 is the fastest growth year of farmers living income at the rate of 9% [7].
After 1997, enter the stage of slow growth while the growth rate of farmers’
income was 4.6% which was the half of 1996. In 1998, their income continued
going down while the speed of growth was only 4.3%. The year of 1999 and 2000
were the same and the speed of growth was 3.8% in 1999. The farmers’ income
was into a trough at 2.1% in 2000. In 2001, there was a recovery growth at 4.8%.
However, it reduced to 4.3% in 2003. It achieved a steady increase from 2004 to
2008 and 2008 was at 8% [6].
In conclusion, there is a large change of growth range of per farming capita
income and it will cause a negative influence to the development of rural economy
even the whole national economy.
As a result, the prediction of per farming capita income becomes the focus of
public concern. The key of long-term agriculture development is solving the prob-
lem of agriculture, rural areas and farmers (the three agriculture-related issues)
reasonably and effectively. And the problem of per capita income of farmers is
not only the core of this issue but also the key to solve the problem of the three
agriculture-related issues. According to the history of the economic development
of China, it’s clear that the per capita income of farmers has an obvious change
trend. In order to solve the problem of the three agriculture-related issues better,
it needs to forecast the per capita income of farmers and grasp the change trend
of it accurately as much as possible. However, some uncertain factors which are
in the statistical process have become the chief problem of the prediction. For
per capita income of farmers, there are many kinds of uncertain factors such as
quantitative and non-quantitative factor, known and unknown factor.
Neglecting those non-quantified and unknown factors, the uncertain factors
affecting the per capita income of farmers can be divided into two categories: (1)
The part of production grew and sold by farmers themselves; (2) The production
cost of agriculture.
There are many methods for forecasting the uncertain factors. In this paper,
the prediction is based on the fuzzy gray prediction model.
The gray forecasting method has been applied in the modeling process of
the dynamic system successfully in different fields such as agriculture, ecology,
economy, statistics, industry and environment. In the absence of long-term his-
torical data, it can use a system model to predict the incomplete or uncertain
information.
According to the statistical method of per farming capita income from the
statistical yearbook, the factors affecting per farming capita income (P) mainly
include: the total income of farmers (TI), the total population of agriculture
(IP). The relationship between them can be expressed by the following formula:
TI
P = IP .
Leshan city of Sichuan province, as an area with World double-heritages,
will be used as a case to study. The problem faced by the area with World
Fuzzy GM(1,1) Model Based Per Capital Income Predicted of Farmers 1009
double-heritages is how to exploit protectively. That is to say, the local natural

and cultural heritages should be protected while the economy is developed and
historical culture is promoted. Thus, there is a problem of degree. How to grasp
this degree plays a vital role of the development of local economy and cultural
protection which is also a model to the other areas with World double-heritages.
The purpose of this paper is to establish a fuzzy gray forecasting model to
predict the per capita income of local farmers, get its changed trend and approve
the effectiveness of the model.
2 Problem Statement
At present, per farming capita income is an important index which reflects the
development of rural economy [5]. Therefore, it is very significant to predict and
grasp accurately the changed trend of per farming capita income. This part will
introduce the condition of per farming capita income of Leshan city and use it as
an example to forecast per farming capita income with the fuzzy gray prediction
model.
Due to the continuous application of the agricultural science and technology
and the continuous improvement of the policies supporting agriculture by gov-
ernment, the total agricultural income of Leshan city has risen from 677.67 to
1992.15 million yuan in the past ten years. As shown in Fig. 1.
Fig. 1. The total agricultural output value of central district of Leshan city
From 2000 to 2009, the total agricultural output value of central district of
Leshan city increased at a stretch. In addition to the year of 2006 and 2007,
the data increased from 107.71 million yuan to 1.7 billion yuan. The overall
increase in agricultural output value is in a small range and a lower growth rate.
In order to a sustained and rapid growth of agricultural economy, the process of
agricultural production should use advanced agricultural science and technology
to increase the production and quality of agricultural products. At the same
time, the policies published by government also need to be further improved.
1010 Z. Meng et al.
Due to the implementation of the family control policy (a basic national

policy) and the continuous improvement of the comprehensive qualities of local
people, the agricultural population of central district had a tendency to reduce
in Leshan city. As shown in Fig. 2.
Fig. 2. The agriculture population of central district of Leshan city from 2000 to 2009
From 2000 to 2009, although there was a slight rebound from 2005 to 2007,
the agricultural population of central district, Leshan city is of the gradual
decline in the trend of change. The main reasons are as follows: On the one hand,
affected by the Asian financial turmoil, from 2000 to 2005, crop prices were gen-
erally low and the enthusiasm of farmers was severely affected. However, since
2006, with the effect of financial crisis declining, the price of agricultural products
rebounded and the enthusiasm of farmers was encouraged again. On the other
hand, along with the rapid development of Chinese economy, the construction
of urbanization is carried out in the vast rural areas. The acceleration of the
urbanization process makes the local farming population decrease continuously.
Thus, the agricultural population in central district, Leshan city also showed
a decreasing trend. In summary, the total value of agricultural production of
central district, Leshan city, has followed an increasing trend while the number
of agricultural population has continued to decrease. Affected by the total value
of agricultural production and the number of agricultural population, per capita
income of farmers should also show an upward trend, as shown in Fig. 3.
From 2000 to 2009, the per capital income of farmers was increasing continu-
ously with slow speed and low growth rate. The reasons are as follow: First, with
the progress of science and technology, many kinds of advanced science and tech-
nology have been widely applied to agricultural production; second, the Chinese
government has introduced various policies to support agricultural development
which reduces the burden of farmers. For a sustained, rapid and steady growth
of per capita income of farmers, it is necessary to improve the conversion rate of
agricultural science and technology and perfect the various supporting policies
of agriculture constantly.
Fig. 3. The per capital income of farmers in central district, Leshan city between 2000
and 2009
3 Modeling
The gray system theory was first proposed by Deng [2], who used the differential
equation as the forecasting model and the least square method to obtain the
coefficient of the equation which has a wide range of application. In the gray
forecasting system, the original data is defined as an accurate. However, in fact,
the original data is not accurate which unable deal with the fuzzy phenomenon
in reality effectively. In this paper, define the original data of per capita income
of farmers as triangular fuzzy number and forecast the fuzzy numbers with gray
forecasting system. Thus, this model consists of two parts: one is the fuzzy part;
the other is the gray forecasting part.
3.1 Fuzzification
As mentioned earlier, the per capita income of farmers is affected by two factors:
the total value of agricultural production (TI) and the agricultural population
TI
(IP). The relationship between them is: P = IP .
When calculating the per capita income of farmers, this method has some
shortcomings. Due to some indicators which unable get accurate enough data,
it can only be estimated by experience. As a result, it will reduce the accuracy
and scientificalness of the end result.
In the statistical process of the total income of farmers, some indicators
unable be calculated accurately such as the part of self-production and mar-
keting and the agricultural production cost of farmers. The self-production and
marketing is the process to cost their own agricultural products. These agricul-
tural products do not enter the market so that their value can not be measured
accurately. In fact, the agricultural production cost includes labor costs while
farmers own labor costs always unable be carried out. As a result, it unable be
measured accurately. In these two parts of the statistical process, some indexes
are often judged by experts’ experience which is lack of precision.
1012 Z. Meng et al.
Thus, the total income of the farmers in a region is only an estimate rather
than an exact value. In this situation, it can be obscured by triangular fuzzy
number to obtain a triangular fuzzy number of per capita incomes of farmers.
In order to solve the above problem, Zadeh proposed the fuzzy theory in
1965. Zadeh represent the fuzzy set by membership function following the idea
that general set is represented by feature function [8].
Definition 1. Any mapping from domain X to closed interval [0, 1]. μA : X →

[0, 1], x → μA : (X). Determine a fuzzy set A of X, μA is called the membership
function of A, μA : (X) is called the membership degree of X to A, fuzzy set A
is written as A = {(x, μA )|x ∈ X}.
Definition 2. Set A as a fuzzy set in domain X. If ∀α ∈ [0, 1], the α- cut sets
of A are all convex set so that the fuzzy set A is called as the convex fuzzy set.
Definition 3. If the fuzzy set M is a normal convex fuzzy set which is defined
in the real field R, it is satisfied: (1) There is a unique point x0 ∈ R, making
μM (x0 ) = 1(x0 is called as the mean value of M ); (2) μM (x) is continuous from
left to right, then M is called as a fuzzy number. The meaning of the fuzzy
number M is “the approximate real number of x0 ”.
From the definition of fuzzy number, the α- cut sets of Mα actually is a closed
interval in real number field R: Mα = {x ∈ R|μM (x) ≥ α} = [mlα , mrα ] mlα and
mrα represent severally the left and right end points of the closed interval Mα .
The general expression of the fuzzy number M is μM (x) = {L(x), l ≤ x ≤
m; R(x), m ≤ x ≤ r}; and L(x) is the right continuous increasing function,
R(x) is the left continuous decreasing function, 0 ≤ L(x), R(x) ≤ 1. If both
the function L(x) and R(x) are linear functions, M is called as triangular fuzzy
number, often denoted as M (l, m, r). The total income of farmers is converted to
triangular fuzzy number and expressed as M (T I − , T I, T I + ). The membership
function shown in Fig. 4.
Fig. 4. The membership function of triangular fuzzy number

The relationship between the per capita income of farmers and the total
income of farmers is:
TI
P = .
TP
We can get the triangular fuzzy number of the historical data of the per
capita income of farmers: M (P − , P, P + ).
3.2 GM(1.1) Model
The GM(1,1) forecasting model is described below: Regard the original data of
the per capita income of farmers as a set of original data series p(0) , we can get:
− + − +
p(0) = {(p(0) (1), (p(0) (1), p(0) (1)), · · · (p(0) (n), (p(0) (n), p(0) (n))}
and
− + − +
p(1) = {(p(1) (1), (p(1) (1), p(1) (1)), · · · (p(1) (n), (p(1) (n), p(1) (n))}.
The relationship between the two sequences is as follows:

i

− −
p(1) (i) = p(0) (m), i = 1, 2, 3 · · · , n,
m=1
i

p(1) (i) = p(0) (m), i = 1, 2, 3 · · · , n,
m=1
i

+ +
p(1) (i) = p(0) (m), i = 1, 2, 3 · · · , n,
m=1
y y
p̂ (1)
(i + 1) = (p0 (1) − )e−ai + .
z z
According to the least squares method, we can conclude:
p̂0 (1) = p̂1 (1).
Then, the final formula is obtained:
p̂0 (i + 1) = p̂1 (i + 1) − p̂1 (i).
4 Application
In order to get the triangular fuzzy number of per capita income of farmers in
central district, Leshan city, it’s necessary to fuzzify the historical data of per
capita income of farmers in this area in the past ten years. The historical data
are as shown in Table 1:
1014 Z. Meng et al.
Table 1. The original data of per capita income of farmers in central district, Leshan
city between 2000 to 2009
Year 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009
Value (Yuan) 2306 2408 2573 2748 3154 3753 3829 4497 5192 5608
We suppose that:
−
P0 (k) = (PK , Pk , Pk+ ),
Pk − Pk− = α,
Pk+ − Pk = β.
According to expert experience and related theories, suppose α = β = 100.

According to the above method, it can get the triangular fuzzy number of per
capita income of peasants in central district, Leshan city by fuzzifing the data
in the above table, as shown in Table 2:
Table 2. The triangular fuzzy number of original data of per capita income of peasants
in central district, Leshan city between 2000 to 2009
Year 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009
P− 2206 2308 2473 2648 3054 3653 3729 4397 5092 5508
P 2306 2408 2573 2748 3154 3753 3829 4497 5192 5608
P+ 2406 2508 2673 2848 3254 3853 3929 4497 5292 5708
By now, input the triangular fuzzy numbers of historical data of the per
capita income of farmers in the central district of Leshan city to the gray fore-
casting model and get the predicted values are as Tables 3 and 4 shown:
Table 3. The predicted values of per capita income of farmers in the central district
of Leshan city between 2001 to 2009
Year 2001 2002 2003 2004 2005 2006 2007 2008 2009
P− 2260 2532 2838 3180 3563 3993 4474 5014 5619
P 2166 2435 2737 3077 3459 3888 4371 4914 5524
P+ 2354 2630 2938 3283 3667 4097 4578 5114 5714
5 Result Analysis and Suggestion

Per capita income of farmers is an important indicator to reflect the develop-
ment of agricultural economy and the living conditions of farmers. Whether the
Table 4. The predicted values of per capita income of farmers in the central district
of Leshan city between 2001 to 2009
Year 2010 2011 2012 2013 2014 2015

P− 6209 6980 7847 8821 9916 11147
P 6454 7302 8262 9349 10578 11968
P+ 6384 7132 7968 8902 9945 11111
−
Fig. 5. The actual values and predicted values of PK
per capita income of farmers has a sustained and rapid growth is important
to the development of the entire national economy and the construction of a
harmonious socialist society. According to the above forecast results, per capita
income of farmers in central district, Leshan city is in the overall upward trend.
In addition, the higher degree of fit between the predicted value and the original
data indicates the validity of the fuzzy-gray forecasting model and also proves
the scientificity of the predicted value. (As shown in Figs. 5, 6 and 7).
In the past decade, due to the continuous progress of agricultural science and
technology and the continuous improvement of agricultural supporting policies,
to a certain extent, it improved the productive rate of agriculture and reduced
the burden on farmers. However, the per capita income of farmers is growing at
a low speed with a low growth rate. Although there is a rising tendency, the per
capita income of farmers is fluctuated because of some related economic factors.
In this view, the suggestions are as follows:
First of all, establish the work thinking of how to achieve stable increasing in
income of farmers. The overall work thinking: the aim is to improve the income
of farmers; the orientation is market; the support is resource; the basis is the
structural adjustment; the driving force is science and technology; the way is
1016 Z. Meng et al.
+
industrialization and agricultural processing; the key is reducing agricultural

population; develop green food, ecological agriculture and characteristic-brand
agriculture vigorously [4].
Second, deepen the reform of the system and overcome the external factors
that restrict farmers’ income. China’s history has proved repeatedly that when
the state land and its property policy are correct and secure, when the enthusi-
asm of farmers is high, the income level has also a corresponding increase.
Third, the Government should implement the policy and measures to ensure
that farmers’ income increase. (1) It is of great significance for agricultural
development and farmers’ incomes to increase financial investment reasonably
and carry out the agricultural subsidy policy. The funds for supporting agri-
culture should focus on poverty alleviation, agriculture foundational facilities,
the research and application of agricultural technology and the green ecolog-
ical agriculture. (2) Deepen the reform of rural financial. In the innovation of
rural financial system, developing small and medium-sized banks and rural finan-
cial guaranty companies will solve the problems of rural finance fundamentally
and provide a strong financial support for the development of three agriculture-
related issues. It has an important role in regulation and protection to improve
the production and income with the development of rural economy.
Finally, continue to promote the strategic adjustment of agricultural struc-
ture and improve the quality and efficiency of agriculture comprehensively. At
this stage, improving the efficiency of agriculture should be market-oriented
focusing on high quality and diversification of agricultural varieties. It means
a change from yield-oriented to quality and efficiency.
6 Conclusion
Through the analysis of the per capita income of the farmers during the 30 years
of reformation and opening, it is found that there are still problems in the growth
process such as low growth rate and large fluctuation. Steady growth of farmers’
income supports strongly for the economic with rapid growth of China and the
smooth progress of reformation and opening.
In this paper, a forecasting model of the per capita income of farmers is
established by fuzzy theory and gray forecasting system. As a case for our study,
the central district, Leshan city provides with the relevant data. First, use the
correlation theory of triangular fuzzy number to deal with the uncertain question;
then use the gray forecasting model to predict the triangular fuzzy number which
ensures the accuracy of the forecast results. At the same time, it provides a
scientific basis for the formulation and implementation of relevant policies.
Acknowledgements. This work was supported by the Humanities and Social Sci-
ences Foundation of the Ministry of Education (Grant No. 16YJC630089), and the
Sichuan Center for System science and Enterprise Development Research (Grant No.
Xq16C04).
References
1. Chen LQ (2013) Discussion on growth path about per capita annual net income of
rural households of guangxi farmers. J South Agric 44(8):1402–1407
2. Deng J (1990) Application of grey system theory in China. In: International Sym-
posium on Uncertainty Modeling and Analysis, 1990, Proceedings, pp 285–291
3. Lin JY (2012) Demystifying the Chinese Economy. Cambridge University Press,
New York
1018 Z. Meng et al.
4. Mao M, Chirwa EC (2006) Application of grey model GM(1, 1) to vehicle fatality

risk estimation. Technol Forecast Soc Change 73(5):588–605
5. Tang HJ (2008) Models and strategic countes measures for circular agriculture devel-
opment in China. J Agric Sci Technol 16:28–35
6. Xiang JH, Xiao L et al (2006) The situation and evaluation of China agriculture
information website development. Agric Netw Inf 2:4–7
7. Yu X, Zhao G (2009) Chinese agricultural development in 30 years: a literature
review. Front Econ China 4(4):633–648
8. Zadeh LA (1965) Fuzzy sets. Inf Control 8(3):338–353
Eco-Cycle Comprehensive Operation
Performance Evaluation–A Case Study
of Baotou Steel Group
Yuyan Luo, Zhong Wang(B) , Yao Chen, Yahong Wang, and Jiming Xie
College of Management Science, Chengdu University of Technology,

wangzhong2012@mail.cdut.edu.cn
Abstract. With the advent of circular economy, enterprises have no

longer given priority to the economic benefits, but also focus on social
responsibility and sustainable development. In order to accelerate the
ecological change of enterprise system, every enterprise vigorously car-
ries out green technology innovation and development. Therefore, how
to evaluate the performance of ecological cycle has become a problem for
management personnel. This paper takes Baotou steel group as an exam-
ple, which is one of the largest steel companies in China, and establishes
an ecological cycle performance evaluation index system from the aspects
of economic operation, resource impact control and ecological circulation
innovation. Based on that, the entropy-topsis model is used to evaluate
the performance of ecological cycle in Baotou steel group. Through the
evaluation, we find out that the ecological cycle performance of Bao-
tou steel rose before 2010 and still at the middle and primary stage of
its eco-cycle development. The results show the evaluation index sys-
tem and method can help management personnel to realize the status
of their company’s eco-cycle development. It also can help policy mak-
ers to make better green policies and strategies, promoting the eco-cycle
transformation in China.
Keywords: Eco-cycle · Operation performance evaluation ·

Entropy-topsis
1 Introduction
In recent years, with the transforming of the concept of ecological civilization
construction, to realize low carbon development and cycle development has been
the important goal of the international economic development, the development
and the transformation of the target. The development direction of enterprise
has changed from a single to rely on economic benefit as the guidance to the eco-
logical cycle of performance oriented transformation. The evaluation of ecologi-
cal cycle operating performance has become an important research topic, which
1020 Y. Luo et al.
specifically reflected the enterprise transformation and put forward applicable

suggestions for the future development of enterprise.
In the whole performance evaluation, building the evaluation index system
is the critical part, reviewing the related literature, a large number of scholars
have discussed the different content of the evaluation index system.
Such as the research using the three dynamic indicators including economic
growth, environmental improvement and technical features to differentiate the
enterprise ecological performance into 13 types including senior evolution to
severe degradation [6]. Besides, there are some researches focusing on the circular
economy developing standard of industrial park to evaluate, such as construct-
ing the circular economy evaluation index system of industrial park including
economic development, resource utilization, pollution control and material recy-
cling indicators and using the TOPSIS method to evaluate [13] and based on
AHP and FCE method, through economic development, the material reduction
and circulation, pollution control and park management to build comprehen-
sive evaluation model of circular economy of industrial parks [12], also, some
researchers are committed to have assessment for the performance of enterprises
and industrial parks, such as incorporating research and development (R&D)
and marketing policy to build a innovative supply chain performance measure-
ment system [15], to use discounted cash flow (DCF) and multi-attribute global
inference of quality (MAGIQ) methods to evaluate the economic, environmen-
tal and social performances of eco-industrial park (EIP) pilot project [1], and
based on principal component analysis and set pair analysis (SPA) and the fuzzy
mathematics method to construct the performance evaluation system [5,8]. In
addition, some use Delphi, LCA technique to select indicators to evaluate the
green public procurement [14], setting the economic indicator and three gener-
ally applicable simplified environmental indicators (raw material consumption,
energy consumption, and CO2 emission) as the eco-efficiency indicators to eval-
uate the ecology efficiency of EIP [14], using the non-radial DEA model to eval-
uate the eco-efficiency [10], taking advantage of R cluster and factor analysis to
have quantitative selection for indicators to build the green industry evaluation
system including three criterion layers of green production, green consumption,
and green environment [4], applying data envelopment analysis (DEA) to study
the resource efficiency of cities [2], and presenting a compliance-based multi-
dimensional trust evaluation system (CMTES) and also using improved TOPSIS
technique to derive trustworthiness from compliance values [9].
Through the above review, a large number of methods such as the entropy-
topsis, principal component analysis, gray correlation analysis, AHP and FCE,
factor analysis, DEA model, the fuzzy comprehensive evaluation method and
other evaluation methods have been widely used in the indicator system.
Based on the circular economy and industrial theory, this article builds the
system of ecological cycle operation performance evaluation index and then takes
the Baotou steel group as a case study, and analyzes the current eco-cycle oper-
ation performance and finally puts forward the corresponding countermeasures
and suggestions.
Eco-Cycle Comprehensive Operation Performance Evaluation 1021
2 Construction of Eco-Cycle Operation

Performance System
The evaluation index system should accurately reflect the actual situation of
enterprise’s production and operation management. The establishment of index
system is a key part of the whole ecological cycle operating performance’s evalu-
ation. It is essential to select the right index and guarantee the credibility of the
results, therefore, this paper combines with the construction principles includ-
ing system’s scientificity and accuracy, effectiveness and comparability, stability
and dynamic, operability, combination of qualitative and quantitative and so
on, and based on a large number of research results of scholars and the actual
work of ecological cycle transformation in enterprises, using analytic hierarchy
process and expert-consulting method to select the indicators, builds the eco-
cycle operational performance evaluation index system which includes economi-
cal operation, resource impact control, technology research and development and
eco-cycle innovation. However, based on the performance evaluation of ecological
recycling operation of Baotou steel, this research is aimed at finding that there
exists missing data of the related index data in development of the technology
and pollution control investment and resource compensation at resource impact
control layer by collecting concrete data of Baotou steel such as annual report,
social responsibility report and so on. Thus considering the actual operation of
Baotou steel, we delete the indicators including R & D investment proportion
of total amount, the number proportion of R & D in technology research and
development level and pollution control investment growth rate and resource
compensation fee increase rate in resource impact control layer. The final eco-
logical recycling performance indicators are as follows in Fig. 1.
3 Entropy-Topsis Mode
On the basis of system theory, entropy is an ordered metric reflecting degree of

disorder in the system [11]. Therefore, the smaller information entropy is, the
lower disorder of information is, and the greater the effect of the information
value and the weight information are, and vice versa [7]. Using entropy to reflect
the size of information can come to the more objective and fair conclusions.
Analytic hierarchy process (AHP) is usually used to determine the weighting
factor of evaluation index in TOPSIS method, Because of its strong subjective
factors, we use the improved entropy weight method to determine the weight of
the evaluation index and to some extent avoid the influence of subjective factors.
When the relative importance of evaluation index is difficult to determine, using
the weight of entropy calculation index is more reasonable [3].
The general idea for using entropy weight method: To construct the index
system and standard weighted matrix, and normalization of the relevant data
into the model in the moment. Specific steps are as follows:
1022 Y. Luo et al.
Target Criterion
Index layer
layer layer
A1: Reduction rate of operating cost, taxes and surcharge
A2: Reduction rate of three costs (management fees,

Economic operation(A)
financial cost and sales expense)
A3: Operating revenue increasing rate
A4: Net profit increasing rate

Eco-cycle comprehensive operation performance
B1: Pollutant emission controllable level
Resource impact
B2: Resource comprehensive recycling level
control(B)
system
B3: Environmental risk prevention level
C1; Green development concept
C2: Ecological circulation informatization degree
C3: Circulating technology level

Eco-cycle innovation(C)
C4: R&D of green technology innovation
C5: Technical transformation project implementation
Fig. 1. Eco-cycle comprehensive operation performance system
(1) Data standardization

The evaluation system has n objects to be evaluated, m evaluation indicators,
the evaluation matrix is X = (xab )n×m (a = 1, 2, · · · , n; b = 1, 2, · · · , m).
For benefit type index:
xab − min xb
yab = (1 ≤ a ≤ n, 1 ≤ b ≤ m). (1)
max xb − min xb
For cost type index:
max xb − xab
yab = (1 ≤ a ≤ n, 1 ≤ b ≤ m). (2)
max xb − min xb
Normalization:
yab
Qab = n (1 ≤ a ≤ n, 1 ≤ b ≤ m). (3)
a=1 yab
(2) Calculation of entropy index

n

Pb = −1/ ln n Qab ln Qab (1 ≤ b ≤ m). (4)
a=1
(3) Determine the index weight

m
ϕb = (1 − Pb )/ (1 − Pb ), (5)
b=1
m
where ϕb satisfies the two conditions 0 ≤ ϕb ≤ 1 and b=1 ϕb = 1:
(4) Standardized weighting matrix is obtained from normalized matrix and index
weight.
W = (wab )n×m = (ϕb yab )n×m . (6)
(5) Determine the positive and negative ideal solution

Positive ideal explanation imagines each index to achieve the most satisfactory
solution, with W + and negative ideal solution is to imagine each index to achieve
the most satisfactory solution, with W − .
W + = { max wab | b ∈ V + , min wab | b ∈ V − } = {w1+ , w2+ , · · · , wm

+
}, (7)
W −
= { min wab | b ∈ V , max wab | b ∈ V } =
+ −
{w1− , w2− , · · · −
, wm }, (8)
where V + = Benefit index set, V − = Cost index set.

(6) Calculated distance
Each evaluation object to the government ideal sister distance with Da+ , Da− ,
said, calculated by the following formula:

m 2
Da =
+
(wab − wb+ ) (a = 1, 2, · · · , n), (9)
b=1

m 2
Da − = (wab − wb− ) (a = 1, 2, · · · , n). (10)
b=1
(7) Calculate the relative closeness of each evaluation object
Ha = Da − /(Da − + Da + )(a = 1, 2, · · · , n), (11)
where Ha is the a object and how close the solution to, the value of Ha is
between 0 and 1, the larger the value of Ha , the smaller the distance between
the evaluation object and the positive ideal solution, the greater the distance
from the negative ideal solution, the closer to the optimal level.
1024 Y. Luo et al.
4 Case Study
Baotou steel group (hereinafter referred to as Baotou steel) is one of the earliest
construction of iron and steel industrial bases after the founding of the People’s
Republic of China, it was first constructed in 1954 and was put into production
in 1957. Turned into a corporate enterprise in 1998, it is the most important iron
and steel industrial base and the largest rare earth industrial base in China, and
one of the biggest rare earth scientific research and production bases as well.
Always committed to diversify development, except to the major business of
steel and rare earth, it also owns the mining and the steel industry. During the
“twelfth five-year”? “big steel”? will be built and commit to become the world’s
largest production base of rare earth steel and the most competitive rare earth
production and scientific research unit and its annual sales revenue will reach
more than 100 billion yuan. By 2015, the group’s total assets have amounted
to 144.9 billion yuan, 33459 people have been employed and 22.501 billion yuan
of sales income has been achieved. At present, Baotou steel has entered the
ranks of tens of millions of tons of iron and steel enterprises in China, and
has an important influence in the world. In recent ten years, Baotou steel has
been committing to adhere to lead innovation, promoting the ecological cycle
transformation and upgrading, keeping the environmental protection concept
and setting the image of green development. Therefore, based on the practical
situation of Baotou steel, evaluating its performance helps to future development
selection and strategic decision-making.
(1) Data selection
According to the availability of data, this paper selects Baotou steel annual
reports during the ten years from 2006 to 2015, part of the data are obtained
by calculations, and others are obtained by the expert scoring method.
(2) Empirical results
In the process of analyzing of the index data, in order to eliminate the
influence of dimension to the data, we use the formulas (1) and (2) to get
the standardized data shown in Table 1, then take use of formula (3) to
give normalized treatment to the data and finally get the weight P and the
entropy φ.
As is shown in Table 2, the weight of the indicators for economic opera-
tion ranges between 0.02 and 0.033, the weight of the indicators for resource
impact control is weighted between 0.06 and 0.12, the weight of the indicators
for ecological circulation innovation range between 0.04 and 0.132. Therefore,
the indicators of resource impact control have a great effect on the ecological
cycle performance of Baotou steel.
Most of the indicators are at a low level, comprehensive index of ten years in
a similar seven years posted schedule H below the 0.1 level. But it had a great
reversal in 2010, whose comprehensive value of H reached the level of 0.04607,
the value of H for economic operation even reached at 0.7 or so, and the value
of H for resource impact control and the ecological circulation innovation were
also at a high stage the ten years (Table 3).
Table 1. Standardized data
Year A1 A2 A3 A4 B1 B2
2006 0.1107 0 0.0397 0.0911 0.1048 0.1347
2007 0.1159 0.1197 0.0007 0.0902 0.2048 0.1347
2008 0.1071 0.1112 0.0552 0.089 0.1476 0.1796
2009 0.1117 0.1118 0.0322 0.0866 0.1667 0.1856
2010 0 0.0913 0.7332 0.121 0.119 0.2156
2011 0.1103 0.1126 0 0.1401 0.1571 0.0749
2012 0.1109 0.119 0.0383 0.0949 0 0.0449
2013 0.1108 0.1041 0.0389 0.0905 0.0524 0.015
2014 0.1116 0.1147 0.0351 0.1967 0.0476 0.015
2015 0.111 0.1156 0.0268 0 0 0
Year B3 C1 C2 C3 C4 C5
2006 0.0435 0.1019 0 0.1327 0.142 0.1393
2007 0 0.1019 0.1135 0.1467 0.1514 0.1333
2008 0 0 0.0054 0.1234 0.142 0.1274
2009 0.3913 0.051 0.0054 0.1292 0.0978 0.1298
2010 0.087 0.1911 0.1135 0.1467 0.142 0.1262
2011 0.0435 0.2229 0.2054 0.1257 0.1293 0.119
2012 0.0435 0.0127 0.1946 0.1118 0.142 0.1298
2013 0.1304 0.1274 0.1135 0.0012 0.041 0.0012
2014 0.1304 0.0637 0.1243 0 0.0126 0
2015 0.1304 0.1274 0.1243 0.0827 0 0.094
Table 2. Entropy weights of indicators
Item Indicator P ϕ
Economical operation A1 0.9542 0.0231
A2 0.953 0.0236
A3 0.4763 0.2634
A4 0.9355 0.0325
Resource impact control B1 0.8624 0.0692
B2 0.8474 0.0768
B3 0.7755 0.1129
Eco-cycle innovation C1 0.765 0.1182
C2 0.7384 0.1316
C3 0.9004 0.0501
C4 0.9 0.0503
C5 0.9036 0.0485
1026 Y. Luo et al.
Table 3. The closeness of the evaluation objects and the positive ideal solution
Year Comprehensive Economic Resource Eco-cycle

operation operation impact innovation
control
H Rank H Rank H Rank H Rank
2006 0.0658 10 0.0417 6 0.0876 8 0.0839 8
2007 0.0773 5 0.018 10 0.1099 3 0.116 5
2008 0.0686 9 0.057 2 0.1077 4 0.0557 10
2009 0.1295 2 0.0361 7 0.3043 1 0.0596 9
2010 0.4607 1 0.7017 1 0.135 2 0.1497 2
2011 0.1093 3 0.0215 9 0.0854 9 0.1988 1
2012 0.0819 4 0.0418 5 0.039 10 0.1379 3
2013 0.0769 6 0.042 4 0.0979 5 0.1073 6
2014 0.071 8 0.044 3 0.0973 6 0.0904 7
2015 0.0765 7 0.0298 8 0.0936 7 0.1167 4
0.8000
0.7000
0.6000
0.5000
0.4000
0.3000
0.2000
0.1000
0.0000
2006 2007 2008 2009 2010 2011 2012 2013 2014 2015
Comprehensive operaƟon Economic operaƟon

Resource impact control Eco-cycle innovaƟon
Fig. 2. The trend of closeness of each evaluation object and positive ideal solution
In order to reflect the current operating state of Baotou steel more intuitively
and clearly, we make a brief line chart, it shown in Fig. 2.
The chart can reflect the development status of the last few years clearly
and objectively, and combining with the specific content from annual report of
Baotou steel, the following conclusions can be reached.
Baotou steel’s comprehensive performance is at huge ups and downs. From
2006 to 2009, the overall ecological cycle operation of Baotou lagged behind
slightly, but still grew steadily. But in the year of 2010, the business of Baotou
steel had a great leap forward trend. The value of H rose from about 0.1 in
2009 to nearly 0.5. For Baotou steel, the year of 2010 is not only glorious, but
also a sharp peak of development. The ecological cycle is good, which is the
result of continuous accumulation. However, after 2010, Baotou steel business,

abruptly suffered a great depression. In 2011, the value of H straightly dropped
to about 0.05, and then was continuously declining during 2012, and was in a
state between 0.05 and 0.1 from 2012 to 2015. So it’s obvious to see that after
Baotou steel’s ecological cycle operates to a certain stage, the development of
the operation will slow down, the challenge of industrial upgrading will increase.
As the same as the composite index, there also existed a parabola change
trend for economic operation. The value of H reached 0.7 in 2010, but years
after that it stayed between 0.01 and 0.057, which is similar to the analysis in
(1). To be brief, Baotou steel business was in a condition of lower stage at first,
but later went through such a huge revolution in 2010 and showed a trend of
rapid increase. Thus it can be said that the year of 2010 is an important turning
point in Baotou. Meanwhile, at the annual report. In previous years, Baotou steel
has been committing to the construction of R & D center, and by the time of
2010, R & D center has been building more perfect. So it is not hard to explain
before 2010, Baotou economy gradually increased, to show the reason for the
substantial increase in 2010, and after a few years, Baotou research devoted to
technology, may it lead to the decline in corporate profits.
Ecological circulation doesn’t grow as such amazing speed as economic opera-
tion, but takes on a smooth curve of innovation. It presents a substantial increase
in 2009 to 2011, makes a breakthrough in 2011, and gradually decline after 2011,
finally is under the stable state. In general, compared with the 2000s, the ecolog-
ical circulation innovation level in 2010s goes up. However, in comparison with
the former two kinds of curve, the ecological circulation innovation reaches peak
in 2011, but the economic performance and the comprehensive ecological cycle
performance is relatively weak. Through the phenomenon and combining with
the annual report of Baotou steel, during 2009 to 2011, Baotou steel has been
promoting science and technology innovation and green development, introduc-
ing the resources recycling project to have an ecological cycle transformation,
which leads to decline in the economic benefits. And then shows a trend of
decline of ecological circulation innovation, it can draw that when ecological cir-
culation innovation to achieve in a certain state, the growing rate slows down
and breakthroughs are much more difficult to realize.
The trend of resource impact control is different from the former three.
Through the line chart, we can clearly see that the sharp change of resource
impact control is 2009 rather than 2010 or 2011, so, it appeared earlier than
the year of vast change, and it can be seen in the line changes, the resource
control increased in the previous year, then it would lead to eco-cycle compre-
hensive performance rising the next year, in the same way, when the resource
impact control had a downward trend, after a year it would lead to a decline in
the economic operation and eco-cycle comprehensive performance. Thus we can
speculate that resource impact control is the foundation of ecological cycle of
enterprise innovation and the daily basic implementation of enterprise resource
recycling also determines the enterprise’s operation and development.
Besides, usually, we can divide the H into five levels, lower stage (N < 0.2),
primary stage (0.2 ≤ N < 0.4), middle stage (0.4 ≤ N < 0.6), relatively
1028 Y. Luo et al.
advanced stage (0.6 ≤ N < 0.8), and advanced stage (0.8 ≤ N < 1.0). So it’s
obvious to come out that the highest H values 0.46, the eco-cycle development
of Baotou steel is in the middle and primary stage. Therefore, we can draw the
conclusion that before 2010, the performance of Baotou steel’s ecological cycling
operation increased slowly, until 2010, the ecological cycle achieved a good level,
but after 2010, industry innovation tended to be slow gradually, Baotou group
operating performance gradually tumbled and the eco-cycle development of Bao-
tou steel is in the middle and primary stage.
5 Conclusion
In view of the above empirical results, we conclude that the ecological perfor-
mance of Baotou Steel rose in 2010, reached a peak in 2010, then decreased
gradually, and the eco-cycle development of Baotou steel is in the middle and
primary stage. And in the case of data support analysis, we confirm the relia-
bility of the conclusion. So based on the conclusion of Baotou Steel Company,
some suggestions are purposed as follows:
(1) Increase resource recycling and sustainable development. In the above analy-
sis we can clearly observe the resource control impact indicator is the corner-
stone of enterprise ecological cycle operation. Benign circulation of resources
under the guidance of national policy is conducive to the development of
enterprises. In the indexes we studied we can also know that controlling pol-
lutant emissions, building chain impact prevention level, achieving effective
material use and effective cycle are the measures to promote the develop-
ment of enterprise resources. Of course, we can also through other ways such
as optimization of industry leading and the construction of new production
lines and other methods to improve the level of corporate resources impact
control.
(2) Increase investment in green technology research and development. Drawn
from the above analysis, eco-cycle innovation plays an important role in the
eco-cycle comprehensive development of enterprises, so the ecological cycle
innovation as a part of leading enterprise development is greatly important
and cannot be ignored, therefore, increasing investment in green technol-
ogy research and development is very important. The policy of the country
now is to encourage the development of green industry and environmental
protection enterprises, only to meet the mainstream trend is the real devel-
opment, and Baotou as a high energy consumption and high pollution steel
enterprise, designing and creating for the theme of environmental protection
industry chain is the priority among priorities of Technology in recent years.
(3) Improve the energy structure. Increasing the recycling of resources and
investment in green technology research and development are currently the
most urgent task and measures based on the current. And for the future
development of ten years, Baotou Steel Company should start in the long
run, improve the energy structure of enterprise, deepen the reform and build
a green enterprise whose development conforms to the trend of the times.

This is also the most useful measure in the future development.

Foundation of China (Grant no. 71501019, 71501138, 71601164 and 71371130), Youth
Program of Social Science Research of Sichuan Province for the Twelfth Five-year Plan
(Grant no. SC15C005 and SC15C030), General Program of Education Department in
Sichuan Province (Grant no. 16SB0071 and 16SB0049), General Program of Mineral
Resources Research Center in Sichuan Province (Grant no. SCKCZY2014-YB04).
References
1. Chan FTS, Nayak A et al (2014) An innovative supply chain performance mea-
surement system incorporating research and development (r & d) and marketing
policy. Comput Ind Eng 69(1):64–70
2. Deilmann C, Lehmann I et al (2016) Data envelopment analysis of cities-
investigation of the ecological and economic efficiency of cities using a benchmark-
ing concept from production management. Ecol Ind 67:798–806
3. Dong J (2010) Analysis of external environment of transmission project based on
improved entropy topsis method. Water Resour Power 3:152–154 (in Chinese)
4. Gongrong C, Yanghong D (2016) Performance evaluation of eco-industrial park
based on the fuzzy mathematics method. J Hunan Univ Sci Technol 19:82–89
5. Jung S, Dodbiba G et al (2013) A novel approach for evaluating the performance
of eco-industrial park pilot projects. J Cleaner Prod 39(5):50–59
6. Li P, Ren H, Zhao L (2005) Evaluation and analysis of enterprise eco-
industrialization based on dynamic indicators. Quant Technica Econ 22(12):16–24
7. Mavi RK, Goh M, Mavi NK (2016) Supplier selection with shannon entropy and
fuzzy topsis in the context of supply chain risk management. Procedia Soc Behav
Sci 235:216–225
8. Rashidi K, Saen RF (2015) Measuring eco-efficiency based on green indicators
and potentials in energy saving and undesirable output abatement. Energy Econ
50:18–26
9. Singh S, Sidhu J (2016) Compliance-based multi-dimensional trust evaluation sys-
tem for determining trustworthiness of cloud service providers. Future Gener Com-
put Syst 67:109–132
10. Song X, Shen J (2015) The ecological performance of eco-industrial parks in shan-
dong based on principal component analysis and set pair analysis. Resour Sci
37:546–554 (in Chinese)
11. Swendsen RH (2016) The definition of the thermodynamic entropy in statistical
mechanics. Phys A Stat Mech Appl 467:67–73
12. Tarantini M, Loprieno AD, Porta PL (2011) A life cycle approach to green public
procurement of building materials and elements: a case study on windows. Energy
36(5):2473–2482
13. Wu XQ, Wang Y et al (2008) Evaluation of circular economy development in
industrial park based on eco-efficiency theory and topsis approach. Chin J Ecol
27(12):2203–2208
14. Ying Z, Hongzhi W, Guotai C (2016) Construction and application of green indus-
try evaluation indicator system based on factor analysis. J Syst Manag 25:338–352
15. Yan L (2012) Assessment of circular economy in chlorine-alkali chemical industrial
parks based on ahp-fce. Ind Technol Econ 31:151–155
Optimization of Closed-Loop Supply Chain
Model with Part and Module Inventory Centers
YoungSu Yun1(B) , Anudari Chuluunsukh1 , Xing Chen1 , and Mitsuo Gen2

1
Department of Business Administration, Chosun University, Gwangju, Korea
ysyun@chosun.ac.kr
2
Fuzzy Logic Systems Institute, Tokyo University of Science, Katsushika, Japan
Abstract. In this paper, a closed-loop supply chain (CLSC) model with

part and module inventory centers is proposed and optimized. Part sup-
plier, module assembler, manufacturer, distribution center and retailer
in forward logistics (FL) and customer, collection center, recovery cen-
ter, secondary market and waste disposal center in reverse logistics (RL)
are taken into consideration for constructing the proposed CLSC model.
Especially, for the reuses of recovered part and module, part and module
inventory centers are also considered. The proposed CLSC model is rep-
resented by a mathematical formulation, which is to minimize the sum
of various costs resulting from each stage of FL and RL under satisfying
various constraints. The mathematical formulation is implemented by
an adaptive hybrid genetic algorithm (a-HGA) approach. In numerical
experiment, various scales of the proposed CLSC model are presented
and the performance of the a-HGA approach are compared with those
of several conventional approaches. In conclusion, the efficiency of the
proposed CLSC model and the a-HGA approach are proved.
Keywords: Closed-loop supply chain · Part and module inventory cen-

ters · Forward and reverse logistics · Adaptive hybrid genetic algorithm
1 Introduction
Closed-loop supply chain (CLSC) model is generally consisted of forward logis-

tics (FL) and reverse logistics (RL). For FL, part supplier, manufacturer, distri-
bution center and retailer are considered. For RL, customer, collection center,
recovery (remanufacturing or refurbishing) center, secondary (used) market and
waste disposal center are taken into consideration. In general, the objective for
implementing the CLSC model can be divided into two aspects as follows:
• Constructing various facilities which can be used in FL and RL stages and

considered in real-world situation.
• Optimizing the CLSC model.
Optimization of Closed-Loop Supply Chain Model 1031
There are many conventional studies considering the above two aspects in
the CLSC model [1–4,7,10,12].
For constructing various facilities, Fleischmann et al. [4] proposed the CLSC
model with three activities of product production, resale and waste disposal in
customer, reused market and disposer market, respectively. Amin and Zhang
[1] proposed the CLSC model with reuse activity. For the reuse activity, they
considered the new part from supplier as well as the reusable part from refur-
bishing center so that all parts are used for producing product at manufacturer
in FL. Wang and Hsu [12] suggested the CLSC model with reuse activity, that is,
recycler in RL classifies the returned product into reusable and unusable mate-
rials, respectively. The reusable materials are then reused at manufacturer in
FL and unusable materials are disposed in landfill area. Similar to Wang and
Hsu [12], Chen et al. [3] also suggested the CLSC model with various reuse
activities. In this CLSC model, recycling center collects the returned product
from customer and then classifies them into reusable and unusable products.
The reusable product is reused at retailer in FL and the unusable product is dis-
assembled into reusable and unusable materials. The reusable material is reused
at manufacturer and the unusable material is treated at waste disposal plant
in RL.
For optimizing the CLSC model, Amin and Zhang [1] and Wang and Hsu
[12] minimized the total costs resulting from each stage of FL and RL. However,
Chen et al. [3] maximized the total profit consisting of total revenue and total
cost.
Although, the conventional studies mentioned above considered various activ-
ities of each facilities in FL and RL, they do not explained exactly how to use
the reusable part (or material) for the facilities in FL. Therefore, in this paper,
we propose a new CLSC model with two inventory centers (part and module
inventory centers) so that the reusable part (or material) is exactly and effec-
tively used for the facilities in FL. In Sects. 2 and 3, the proposed CLSC model
is represented by a mathematical formulation, which is to minimize the sum
of various costs resulting from each stage of FL and RL under satisfying vari-
ous constraints. The mathematical formulation is implemented by an adaptive
hybrid genetic algorithm (a-HGA) approach in Sect. 4. In Sect. 5 for numerical
experiment, various scales of the proposed CLSC model are presented and the
performance of the a-HGA approach are compared with those of several con-
ventional approaches. Finally, in Sect. 6, as a conclusion, the efficiencies of the
proposed CLSC model and the a-HGA approach are proved.
2 Proposed CLSC Model with Two Inventory Centers

The proposed CLSC model is consisted of part suppliers in areas 1, 2, 3 and 4,
module assembler, product manufacturer, distribution center and retailer for FL
and customer, collection center, recovery center, secondary market and waste
disposal center in RL and two inventory centers (part and module inventory
centers). Its conceptual structure is shown in Fig. 1.
1032 Y. Yun et al.
Fig. 1. Conceptual structure of the proposed CLSC model
The difference between the proposed CLSC model and the conventional
CLSC models [1,3,12] is that the former considers two inventory centers and
the latter does not taken into account them. The detailed logistics are as fol-
lows. Each part supplier at areas 1, 2, 3, and 4 respectively produces new part
types 1, 2, 3, and 4 and then send them to part inventory center. Also, each
module assembler assembles new module types 1 and 2 and then send them to
module inventory center. Recovery center checks the returned product from col-
lection center and then classifies them into recovered modules (recovered module
types 1 and 2) and recovered parts (recovered part types 1, 2, 3, and 4). Each
recovered part and module are sent to part and module inventory centers, respec-
tively. Each inventory center has a function that new part and module in FL and
recovered part and module in RL can be used for respectively assembling and
producing module and product in module assembler and product manufacturer
in FL. The recovered product and unrecovered module at recovery center are
sent to secondary market and waste disposal center so that they are resold and
land filled, respectively.
First, some assumptions are presented.
• Only single product is produced.
• The numbers of facility at each stage are already known.
• Among all facilities at each stage, only one facility should be opened.
• Fixed cost for operating each facility are different and already known.
• Unit handling cost considered at same stage is the same and already known.
• Unit transportation cost considered between each facility are different and
already known.
• All products from customer are returned to collection center.
• The qualities of recovered part and module at recovery center are identical
with those of new part and module.
Under considering the above assumption, the mathematical formulation for

the CLSC model is proposed. Some indexes, parameters and decision variables
are defined.
Index Set:
a : index of part supplier; a ∈ A;
p : index of area of part supplier; p ∈ P ;
b : index of module assembler; b ∈ B;
m : index of area of module assembler; m ∈ M ;
v : index of part inventory center; v ∈ V ;
g : index of module inventory center; g ∈ G;
d : index of product manufacturer ; d ∈ D;
e : index of distribution center; e ∈ E;
y : index of retailer/customer; y ∈ Y ;
j : index of collection center; j ∈ J;
z : index of recovery center; z ∈ Z;
i : index of secondary market; i ∈ I;
k : index of waste disposal center; s ∈ S.
Parameters:
BSap : fixed cost at part supplier a of area p;
BIv : fixed cost at part inventory center v;
BMbm : fixed cost at module assembler b of area m;
BVg : fixed cost at module inventory center g;
BPd : fixed cost at product manufacturer d;
BDe : fixed cost at distribution center e;
BCj : fixed cost at collection center j;
BRz : fixed cost at recovery center z;
LSap : unit handling cost at part supplier a of area p;
LIv : unit handling cost at part inventory center v;
LMbm : unit handling cost at module assembler b of area m;
LVg : unit handling cost at module inventory center g;
LPd : unit handling cost at product manufacturer d;
LDe : unit handling cost at distribution center e;
LCj : unit handling cost at collection center j;
LRz : unit handling cost at recovery center z;
TSIapv : unit transportation cost from part supplier a of area p to part inventory
center v;
TIMvbm : unit transportation cost from part inventory center v to module assem-
bler b of area m;
1034 Y. Yun et al.
TMVbm : unit transportation cost from module assembler b of area m to module

inventory center g;
TVPgd : unit transportation cost from module inventory center g to product
manufacturer d;
TPDde : unit transportation cost from product manufacturer d to distribution
center e;
TDTey : unit transportation cost from distribution center e to retailer/customer
y;
TTCyj : unit transportation cost from retailer/customer y to collection center
j;
TCVjz : unit transportation cost from collection center j to recovery center z;
TVUzk : unit transportation cost from recovery center z to secondary market k;
TVIzv : unit transportation cost from recovery center z to part inventory center
v;
TVVzg : unit transportation cost from recovery center z to module inventory
center g;
TUWzk : unit transportation cost from recovery center z to waste disposal center
k;
Decision Variables:
iiv : handling capacity at part inventory center v;
imbm : handling capacity at module assembler o of area e;
ivg : handling capacity at module inventory center g;
ipd : handling capacity at product manufacturer d;
ide : handling capacity at distribution center e;
iry : handling capacity at retailer/customer y;
icj : handling capacity at collection center j;
ivz : handling capacity at recover center z;
iui : handling capacity at secondary market i;
iwk : handling capacity at waste disposal center k

1, if part supplier p of area a is opened
psap =
0, otherwise
1, if part inventory center v is opened
piv =
0, otherwise;
1, if module assembler b of area m is opened
pmbm =
0, otherwise;
1, if module inventory center g is opened
pvg =
0, otherwise;
1, if product manufacturer d is opened
ppd =
0, otherwise;
1, if distribution center e is opened
pde =
otherwise;
0,
1, if distribution center j is opened
pcj =
0, otherwise;

1, if distribution center z is opened
prz =
0, otherwise.
Objective function and constraints are as follows:
Minimize Total Cost (TC)

TC = (BSap psap ) + (BIv × piv ) + (BMbm × pmbm ) ,
p a v b m

+ (Bvg × pvg ) + (BPd × ppd ) + (BDe × pde ) + (BCj × pcj )
g d e j

+ (BRz × prz ) + (LSpa × ispa × pspa ) + (LIv × iiv × piv )
z a p v

+ (LMbm ×imbm ×pmbm )+ (LVv ×ivg ×pvg )+ (LPd ×ipd × ppd )
m b g d

+ (LDe ×ide ×pde )+ (LCj ×icj × pcj )+ (LRz ×irz ×prz ) (1)
e j z

+ (TSIapv ×ispa ×pspa × piv )+ (TIMvbm ×iiv ×piv ×pmbm )
p a v v m b

+ (TMVeoc ×imbm ×pmbm ×pvg )+ (TVPgd ×ivg ×pvg ×pdd)
m b g g d

+ (TPDde × pd × ppd × pde ) + (TDTey × ide × pde )
d e e y

+ (TTCj × icj × pcj ) + (TCVjz × irz × prz × pvj )
y j j z

+ (TVIzv × irz × prz × piv ) + (TVVzg × irz × prz × pvg )
z v z g

+ (TVUzi × irz × prz ) + (TVWzk × irz × prz ).
z i z k

s. t. psap = 1, ∀p ∈ P, (2)
p

piv , (3)
v

pmbm = 1, ∀m ∈ B, (4)
m

pvg = 1, (5)
g

ppd = 1, (6)
d

pde = 1, (7)
e

pcj = 1, (8)
j

rz = 1, (9)
z
1036 Y. Yun et al.

(isap × psap ) − (iiv × piv ) ≤ 0 ∀p ∈ P, (10)
ap v

(iiv × piv ) − (imbm × pmbm ) ≤ 0 ∀b ∈ B, (11)
v bm

(imbm × pmbm ) − (ivg × pvg ) ≤ 0 ∀b ∈ B, (12)
bm g

(ivg × pvg ) − (ipd × ppd ) = 0, (13)
g d

(ipd × ppd ) − (ide × pde ) = 0, (14)
d e

(ide × pde ) − qry = 0, (15)
e y

qry − (icj × pcj ) = 0, (16)
y j

(icj × pcj ) − (irz × prz ) ≥ 0, (17)
j z

(iiv × piv ) − a1 % (irz × prz ) ≥ 0, (18)
v z

(ivg × pvg ) − a2 % (irz × prz ) ≥ 0, (19)
g z

pui − a3 % (irz × prz ) ≥ 0, (20)
i z

pwk − a4 % (irz × prz ) ≥ 0, (21)
k z
psap = {0, 1} , a ∈ A, ∀p ∈ P, (22)

piv = {0, 1} , ∀i ∈ I, (23)
pmbm = {0, 1} , ∀n ∈ M, b ∈ B, (24)
pvg = {0, 1} , ∀g ∈ G, (25)
ppd = {0, 1} , ∀d ∈ D, (26)
pde = {0, 1} , ∀e ∈ E, (27)
pcj = {0, 1} , ∀j ∈ J, (28)
prz = {0, 1} , ∀z ∈ Z, (29)
psap , piv , pmbm , ppd , pde , pcy , prz pui , pwk ≥ 0, ∀p ∈ P, ∀a ∈ A, ∀v ∈ V,
∀m ∈ M, ∀b ∈ B, ∀d ∈ D, ∀e ∈ E, ∀y ∈ Y, ∀z ∈ Z, ∀i ∈ I, ∀k ∈ K, (30)
∀j ∈ J.
The objective function of Eq. (1) is to minimize the total sum of fixed costs,
handling costs and transportation costs. Equations (2)–(9) means that only one
facility is opened at each stage. Equation (10) implies that the sum of the han-
dling capacity at each supplier in areas 1, 2, 3 and 4 is the same or greater than
that of the module inventory center. The same meaning is considered in Eqs.
(11)–(17). Equation (18) implies that the sum of the recovered products at each
recovery center is the same or greater than that of the recoverable products with
a1 % at each collection center. Equation (19) restricts that the sum of the han-
dling capacity at all part inventory center is the same or greater than that of the
recoverable product with a2 % at all recovery centers. Equations (20)–(21) indi-
cate the same meanings of the Eqs. (18) and (19). Equations (22)–(29) restrict
the variables to integers 0 and 1. Equation (30) means non-negativity.
4 A-HGA Approach
The mathematical formulation is implemented using the a-HGA approach. The
a-HGA approach is a hybrid approach with adaptive scheme. For the hybrid app-
roach, conventional GA and Cuckoo search (CS) are used and for the adaptive
scheme, Srinivas and Patnaik’s approach [11] is used. Using the hybrid approach
can achieve a better improvement of solution rather than using a single app-
roach does. By using the adaptive scheme, the rates of crossover and mutation
operators used in GA are automatically regulated. The detailed implementation
procedure [5,6,8] is as follows.
Step 1. GA approach
Step 1.1. Representation 0–1 bit representation scheme is used for
effectively representing opening/closing decision of all facil-
ities at each stage.
Step 1.2. Selection Elitist selection strategy in enlarged sampling
space is used
Step 1.3. Crossover Two-point crossover operator (2X) is used.
Step 1.4. Mutation Random mutation operator is used.
Step 1.5. Reproduce offspring
Step 2. CS approach Apply Levy flight scheme [8] to the offspring of GA and
produce new solution.
Step 3. Adaptive scheme Apply the adaptive scheme used in Srinivas and Pat-
naik [11] to regulate crossover and mutation rates.
Step 4. Termination condition If pre-determined stop condition is satisfied,
then stop all Steps, otherwise go to Step. 1.2.
In numerical experiment, three scales of the proposed CLSC model are presented.
Each scale has various sizes of part suppliers in areas 1, 2, 3 and 4, par inventory
center, module assembler in area 1 and 2, module inventory center, product
manufacturer, distribution center and retailer in FL and customer, collection
center, recovery center, secondary market, and waste disposal center in RL.
The detailed sizes of each scale are showed in Table 1. For each scale, 1,500
products are produced in FL and handled in RL. The rates at recovery center for
handling the returned products from collection center are as follows: α1 = 60%,
1038 Y. Yun et al.
Table 1. Various scales of the proposed CLSC model
Scale Part Part Module Module Product Distri- Retalier/ Collect- Recovery Second- Waste
supplier inventory assembler inventory manufa- bution customer ion center ary disposal
center center cturer center center market center
1 2 3 4 1 2
1 4 4 4 41 3 3 1 4 3 10 4 3 10 1
2 12 12 12 12 2 8 8 2 12 8 15 12 8 15 2
3 23 23 23 23 3 20 20 3 23 20 25 20 20 25 3
Table 2. Each approach for comparison
Approach Description
GA Conventional GA [5]
HGA Conventional HGA by Kanagaraj et al. [6]
a-HGA Proposed approach in this paper
Lingo Conventional optimization solver by Lindo Systems [8]
Table 3. Measures of performance comparison
Measure Description
Best solution Best value of the objective functions under satisfying all
constraints
Average solution Averaged values of the objective functions under satisfying all
constraints
Average time Average value of the CPU time (Sec.) used for running each
approach
Percentage difference The difference of the best solutions of GA, HGA and a-HGA
when compared with that of Lingo
α2 = 20%, α3 = 15% and α4 = 5%, for recoverable products, parts, modules

and unrecoverable modules, respectively.
To prove the efficiency of the a-HGA approach, two conventional approaches
(GA, HGA) and Lingo [9] as a benchmark are used and they are summarized in
Table 2. The performances of each approach are compared using various measures
of performance as shown in Table 3.
All approaches ran on a same computation environment (IBM compatible
PC 1.3 Ghz processor-Intel core I5 − 1600 CPU, 4 GB RAM, OS-X EI) and pro-
grammed by MATLAB version R2014a. The parameter settings for GA, HGA
and a-HGA are as follows: For each scale, total number of generations is 1,000,
the population size is 20, the crossover and mutation rates are 0.5 and 0.3 respec-
tively. Total 30 trials were independently done for eliminating the randomness in
Table 4. Computation results of each approach
Best solution Average Average Average Percentage

iteration solution time diff.
Scale 1 GA 149,789 15 150,754 8.1 −7.41%
HGA 149,789 17 150,520 8.30 −7.41%
a-HGA 149,789 21 149,789 138.50 −7.41%
LINGO 161,777 - - - 0.00%
Scale 2 GA 149,868 35 152,381 8.33 −10.12%
HGA 149,852 42 152,679 8.83 −10.13%
a-HGA 149,765 54 151,616 193.70 −10.18%
LINGO 166,747 - - - 0.00%
Scale 3 GA 151,572 53 153,919 9.00 −9.31%
HGA 149,943 54 153,610 9.07 −10.27%
a-HGA 149,895 42 153,365 313.30 −10.30%
LINGO 167,103 - - - 0.00%
the search processes of the GA, HGA and a-HGA. Table 4 shows the computation
results by GA, HGA, a-HGA and Lingo.
In the scale 1 of Table 4, the a-HGA including the GA and HGA has the same
re-sult and their performances are greater than that of Lingo in terms of the
best solution and percentage difference. In terms of the average solution, the a-
HGA shows the best performance compared with the GA and HGA. However, in
terms of the average time, the a-HGA shows the worst performance and the GA
is the best performer. In scale 2, the performance of the a-HGA is more efficient
than the GA and HGA in terms of the best solution and average solution. The
difference in terms of the percentage difference means that the a-HGA is 0.05%
and 0.06% advantageous compared with the HGA and GA. However, in terms
of the average time, the a-HGA is the slowest and the GA is the quickest.
Similar results are also shown in the scale 3, that is, the a-HGA shows the
best performances in terms of the best solution, average solution and percentage
difference when compared with the GA, HGA and Lingo. However, the search
speed of the a-HGA is about thirty times slower than those of the GA and HGA.
Figure 2 shows the convergence behaviors of GA, HGA and a-HGA until the
generation number is reached to 200.
In Fig. 2, all approaches show rapid and various convergence behaviors during
initial generations. However, after about 50 generations, each approach does not
show any more convergence behaviors and the a-HGA shows to be more efficient
behaviors than the GA and HGA.
The result of the detailed material flows in the a-HGA for the scale 3 is shown
in Fig. 3. The opened facilities at each stage are displayed as white-coloured
boxes. The 975 new parts in each area are produced and then sent to part inven-
tory center. The recovery center recovers the quality of the returned products
1040 Y. Yun et al.
Fig. 2. Convergence behaviors of GA, HGA and a-HGA
Fig. 3. Detailed material flows and facility numbers opened at each stage
and then sent 225 recovered parts to part inventory center. Part inventory center
stores 975 new parts and 225 recovered parts. Total 1,200 parts (= 975+225) are
shipped to module assembler. Module assembler assembles 1,200 parts and pro-
duces 1,200 new modules. Recovery center recovers the quality of the returned
products and then sent 300 recovered modules to module inventory center. Mod-
ule inventory center stores 1,200 new modules and 300 recovered modules. Total
1,500 modules are sent to product manufacturer for producing 1,500 products.
1500 products are sent to each retailer via distribution center. In RL, 1,500 prod-
ucts from all customers are returned to recovery center through collection center.
900 recovered products (= 60% × 1, 500) of all returned products are resold at
secondary markets. 75 unrecovered parts (= 5%×1, 500) of all returned products
are sent to waste disposal center.
Based on the results of Table 4, Figs. 2 and 3, we can reach the following
conclusions.
• The proposed CLSC model can represent the detailed material flows at each
stage and effectively handle the recovered part and module by using two
inventory centers, when compared with the conventional models of Amin and
Zhang [1], Wang and Hsu [12], and Chen et al. [3].
• The a-HGA approach shows to be more efficient in many measures of per-
formance than the GA, HGA and Lingo, which implies that the former can
explore whole search space rather than the latter do.
• The search speed of the a-HGA is significantly slower than those of the GA
and HGA, since the former has an adaptive scheme to regulate crossover and
mutation operators and the search structure requires many times.
6 Conclusion
In this paper, we have proposed a new type of the CLSC model. The proposed
CLC model has part suppliers at areas 1, 2, 3, and 4, module assembler, product
manufacturer, distribution center and retailer for FL and customer, collection
center, recovery center, secondary market and waste disposal center for RL.
Especially, for effectively handling recovered part and module, two inventory
centers (part and module inventory centers) have been used.
The proposed CLSC model has been represented by a mathematical formula-
tion, which is to minimize the sum of handling cost, fixed cost and transportation
cost resulting from each stage of FL and RL under satisfying various constraints.
The mathematical formulation has been implemented by the a-HGA approach.
The a-HGA approach is a hybrid algorithm with GA and CS approach. Also,
using an adaptive scheme, the rates of crossover and mutation operators are
automatically regulated in a-HGA approach. The a-HGA approach has been
implemented in various scales of the proposed CLSC model to compare its per-
formance with those of GA, HGA and Lingo. The experimental results have
shown that the a-HGA approach is more efficient in terms of various measures
of performance than the GA, HGA and Lingo. However, since the search speed
of the a-HGA approach is significantly slower than those of the others, a room
for improvement is still left in the a-HGA approach.
Acknowledgments. This work is partly supported by the Grant-in-Aid for Scientific

1042 Y. Yun et al.
References
1. Amin SH, Zhang G (2012) An integrated model for closed-loop supply chain
configuration and supplier selection: multi-objective approach. Expert Syst Appl
39(8):6782–6791
2. Amin SH, Zhang G (2013) A multi-objective facility location model for closed-
loop supply chain network under uncertain demand and return. Appl Math Model
37(6):4165–4176
3. Chen YT, Chan FTS, Chung SH (2014) An integrated closed-loop supply chain
model with location allocation problem and product recycling decisions. Int J Prod
Res 53(10):3120–3140
4. Fleischmann M, Krikke HR et al (2000) A characterisation of logistics networks
for product recovery. Omega 28(6):653–666
York
6. Gen M, Cheng R (2000) Genetic algorithms and engineering optimization. Wiley,
New York
7. Georgiadis P, Besiou M (2008) Sustainability in electrical and electronic equip-
ment closed-loop supply chains: a system dynamics approach. J Cleaner Prod
16(15):1665–1678
8. Kanagaraj G, Ponnambalam SG, Jawahar N (2013) A hybrid cuckoo search and
genetic algorithm for reliability–redundancy allocation problems. Comput Ind Eng
66(4):1115–1124
9. Lingo (2015) Lindo Systems. www.lindo.com
10. Savaskan RC, Bhattacharya S, Van Wassenhove LN (2004) Closed-loop supply
chain models with product remanufacturing. Manage Sci 50(2):239–252
11. Srinivas M, Patnaik LM (1994) Adaptive probabilities of crossover and mutation
in genetic algorithms. IEEE Trans Syst Man Cybern 24(4):656–667
12. Wang HF, Hsu HW (2010) A closed-loop logistic model with a spanning-tree based
genetic algorithm. Comput Oper Res 37(2):376–389
The Innovation Research of College Students’
Academic Early-Warning Mechanism Under
the Background of Big Data
Yu Li and Ye Zhang(B)
University of Electronic Science and Technology,

Chengdu, People’s Republic of China
zhangye uestc@126.com
Abstract. With China’s higher education from “elite education” to

“mass education”, students’ academic problems have become increas-
ingly serious. In the big data era, building college students’ academic
early-warning mechanism is of great significance for the training of qual-
ified college students and improving the quality of higher education to
keep pace with the times. However, in recently, the college students’
research of academic early-warning mechanism based on the big data
is relatively rare. This paper introduces the current research status of
academic early-warning mechanism and the background of big data, and
summarizes the characteristics of big data. Then it expounds the impact
of big data thinking on the traditional academic early-warning mecha-
nism about the four aspects: decision-making, forecasting, service and
evaluation. This article focuses on putting forward an innovation model
of academic early-warning mechanism concerning the information col-
lection system, information processing and analysis system, program-
making system and information release system based on big data, which
will guide the practice of big data in the field of college students’ acad-
emic early-warning.
Keywords: Academic early-warning · Big data · Impact · The innova-

tion model
1 Introduction
After the reform and opening up, China’s higher education has changed from elite
education to mass education. The admission rate of college entrance examination
has risen from 7% to 80% Since the 1980s, but the quality of students has
been declining and the academic problems have been increasing year by year.
For instance, the number of students who demote and dropped out of school
increased obviously. Therefore, it is necessary to establish the academic early-
warning mechanism to solve these problems.

1044 Y. Li and Y. Zhang
Academic early-warning is an impact on individual students in serious acad-

emic situation, and promptly reminds the students themselves and informs the
parents. It is a academic early-warning system through the communication and
collaboration of school, parents and college students, targeting to take appro-
priate preventive measures between the undergraduate students in the learning
situation to give timely guidance to help students solve their academic prob-
lems and successfully complete their studies. In recent years, the research on
college students’ academic early-warning mechanism has lots of ideas. Chen [4]
put forward the design idea of the colleges’ early-warning mechanism from the
aspects of the connotation, goal, function, warning system and the early warning
level on account of the angle of the credit system of teaching management in
colleges and universities. Cao [3] showed various reasons of academic problems
behind the “learning difficulty students”, putting forward the establishment of
“learning difficulty students” assistance mechanism of academic pre-warning.
Ding [5] and his colleagues took Science and Technology University in Suzhou
province of China as an example, comparing and analyzing the data of school
enrollment of academic early-warning mechanism. Hua [7] introduced the regu-
lations and characteristics of academic early-warning system of Taiwan Universi-
ties in China, and then discuss the study of academic early-warning mechanism
of college students in Mainland China. According to the document of Ministry
of Education, which gives many opinions to improve the quality of teaching
on further deepening reform of undergraduate teaching in 2007, it proposed to
strengthen the management of personnel training process. “National Medium
and Long Term Education Reform and Development Plan (2010–2020)” clearly
published: establishing the assistance mechanism of students with learning dif-
ficulties, improving the excellent student training methods. In August 2007, the
Ministry of Education promulgated “Academic Early-warning” as one of 171
new Chinese Words. The above studies and documents indicate that the system
of academic early-warning is of great significance to the cultivation of talents in
colleges and universities.
With the advent of the big data era, data increases fast, and the date range
is expanded, which contribute to the formation of educational database. The
information plan of national education offered “two platforms” (the public ser-
vice platform of national education resources, the service platform of national
education management) based on the concept of big data in wisdom campus
construction and the large-scale opening online courses’ teaching practice. In
the whole sample data, mining the regulation of educational development is
conducive to the guidance of education in colleges and universities to improve
the accuracy of decision-making. To a certain degree, in the construction of edu-
cational information at different levels has taken a historic step, furthermore,
it also lays the foundation for the practice of big date in the field of academic
early-warning. However, although many scholars have explored the application
of big data in education field at present, the theoretical and applied research on
big data in the academic early-warning mechanism of college students in China
and abroad is still in a preliminary stage. Therefore, this article will analyze
The Innovation Research of College Students’ Academic Early-Warning 1045
the impact of big data on the traditional academic early-warning mechanism

and design the model of college students’ academic early-warning mechanism
under the background of big data, and expect to provide basis for the follow-up
research and application of big data in the field of college students’ early-warning
mechanism.
2 The Brief Introduction of Big Data
There is no uniform definition of big data at present, citing the McKinsey Global
Institute’s description of The Next Frontier in Innovation, Competition and Pro-
ductivity: big data refers to the date process that is different from traditional
software tools [8]. Although the definition of big data is not uniform, the char-
acteristics of big data are clear and recognized. (1) The huge data volume. A
variety of terminal equipment and sensors produce a lot of data, PB-scale data
sets can be described as normal. (2) The variety data types. In the era of big
data, more and more unstructured data, including network log, audio, video,
pictures, geographical information. These different types of data needs higher
requirements on data processing. (3) Processing speed. The most significant fea-
tures is to distinguish big data technology from traditional data technology. In
the face of massive amounts of complex data, big data can deal with real-time
data faster. (4) Low value density. The value density is inversely proportional to
the size of the data, for an hour video, the useful data may be only one or two
seconds in the continuous monitoring.
The first feature of big data is large-scale, with the development of infor-
mation technology, a variety of information systems, databases, cloud storage,
Internet, Web of Things and mobile intelligent terminals, especially in recent
years, increase rapidly in social system. The data sharing becomes very easy,
so that the data scale is expanding constantly. At the same time, big data also
has a complex and diverse data structures, owing to large, complex and volatile
date, it is becoming more and more difficult to obtain hidden, useful knowledge,
traditional data warehouse and data mining-related processing mode has been
unable to meet the requirements, big data has a stronger decision-making power,
insight and process optimization capabilities, which brings new changes to data
processing, storage and conversion.
The latent value of big data is the correlation between the data, big data
thinking is a conversion from the traditional causal analysis to the relevant analy-
sis. More and more countries, governments, industries, enterprises and other
institutions have realized that big data is becoming the organization’s the most
important asset, the capabilities of data analysis is becoming the organization’s
core competitiveness [11]. At present, the government has put big data into
application to promote people’s daily lives, the internet has penetrated into var-
ious industries. Big data has brought about an enormous effect on education. It
has been integrated with education and is promoting the reform of educational
model. Besides, it grounds for the application in college students’ academic early-
warning mechanism.
Big data is a prediction method based on objective data that has been
obtained rather than on subjective inferences. This analysis method is based
on the huge data, which is obviously different from the sample analysis method
which has been used in the past. It can correct the error of the sampling method
and improve the accuracy of the analysis result. Big data analysis can get the pre-
cise results, easier access to the user’s trust, and thus has a wide utilization. The
application of big data technology in college students’ academic early-warning
mechanism will have more timely and accurate early-warning.
3 The Impact of Big Data Thinking on the Traditional

Academic Early-Warning Mechanism
3.1 Rational Decision-Making
The main object of universities and colleges is college students who as the most
dynamic and most potential individuals, college students’ thinking, behavior and
personality are diverse. Due to the invisible and complex nature of thinking, it
is hard to understand one’s mind. In the past, we could only make judgments on
the basis of individual experience and simple sample statistics. This traditional
subjective decision-making method and the statistical analysis method of small
sample data will play a certain role in the traditional academic pre-warning, but
the cost of manpower, material and financial resources is huge, and the accuracy
of academic early-warning needs to be improved. In the era of big data, we can
effectively make more scientific judgments, more rational decision-making. Big
data can provide a data basis for us to make the rational decision through the
real-time monitoring of students learning and lives in and out of class. We can
also analyze the students’ thoughts and behavior through the internet, forming
regular recognizing of them by the analysis of vast amounts of date to achieve
scientific decision-making.
3.2 Precise Prediction

Forecast is the core of big data, the big data analysis method applied to the mas-
sive data, so as to predict the possibility of occurrence, to achieve the purpose
of prediction. A large number of data not only allows us to predict the devel-
opment of things, but also makes our prediction of human behavior possible. In
the era of big data, the behavior data that is interdependent and associated of
college students have been recorded and preserved. We can find the link between
their behaviors and predict trends through the depth analysis and integration of
behavioral data, so that we are able to react quickly to guide and intervene in
advance.
3.3 Personalized Service

The big data has opened the way of individualized education, it can reveal the
students’ learning pattern through the data tracking and analysis of the students’
learning process, formulating individualized educational programs for them, and

giving the practical direction for individualized education. College students are
highly individualized groups. They pay attention to strong character hoping to be
treated as a unique individual. In the past, students worked only from the overall
work programs, which ignore the personal differences and individual needs. Big
data concerns every person, and will make a big change to our students work. It
will make us re-examine the work of students not only from the overall picture,
but also pay more attention to individual work to carry out specific work to
promote individualized development of each college student. Through the whole,
timely and dynamic record of each student’s learning, life and social situation,
the big data contribute to a better understanding personality and growth needs
of students so that we can adopt ideological and political education measures,
career planning, psychological counseling, comprehensive quality education to
achieve personalized service of college students.
3.4 Scientific Evaluation

In the past, it is difficult to quantify the college students work whether it is
the ideological evaluation of students or students’ the economic situation of
the family. It is inevitable that there is a certain bias resulting in a highly
subjective evaluation that is because of the information only from counselors,
class teachers, students and other channels. But the big data can promote the
scientific evaluation of teaching to achieve more effective academic early-warning.
To assess the economic situation of students in their family, for example, we
can clearly grasp the specific period of time students’ income and expenditure
through their campus cards consumption records, shopping consumption records,
mobile payment list, personal account records, thus we can make an accurate
judgment of their personal economic situation as an important basis for judging
the economic status of their families to avoid the errors caused by the subjective
analysis. In the evaluation of the students’ ideological situation, the analysis of
the massive data can be more deterministic to test the dynamics of their thoughts
and behavior to reflect the information of their ideological characteristics by data
processing.
4 The Innovation Research: The Building of Academic

Early-Warning Mechanism Under the Big Data
Environment
In the era of big data, the academic early-warning mechanism consists of four
major systems: information collection system, information analysis and process-
ing system, program-making system and information release system.
4.1 Information Collection System

The collection of information is the primary work of the academic early-warning
system, and it is the key to the work of academic early-warning as well. In the
coming of big data, the basis of college students’ work is to master the huge
data which can really understand the characteristics of the behavior of college
students to effectively find out the education, management and service counter-
measure. First, colleges and universities should conduct the top-level design that
is an overall vision of an integrated data platform inside their own campus. Con-
siderable universities in the development of the wisdom campus’s construction,
different departments of the university only consider the work of the department
needs, so that it is difficult to achieve data sharing and integration. So the school
should set up a coordination department and academic pre-warning department
to build a data center to gather the information of the office, academic affairs,
school work, library and other related departments in order to form informa-
tion platform. Secondly, building a system of online data collection platform to
form a large database of students across the school in order to ensure the timely
and complete collection of all students of all data. At the same time, colleges
and universities should from the overall point of view do a good job of data
classification, hierarchical collection and planning work to ensure the diversify
of data sources and data types. Thirdly, colleges and universities should take
the initiative to share the social database contributing to the establishment of
big database of education. In this paper, the educational data can be divided
into three categories: explicit data, behavioral data and system data (Fig. 1).
Explicit data is the data input or output by the end user. Behavior data can
also be called control data, which is designed by the developer for the purpose
of recording the data of the operation process of the user, generally only seen
by the administrator. System data is automatically generated by the system,
behavioral data and system data are hidden data [9].
Fig. 1. The three categories educational data
4.2 Information Analysis and Processing System

The amount of information is large and have different structure, and a lot of
information have no effect on the academic early-warning, so we need informa-
tion processing and analysis system to sorting, deleting unnecessary information,
and find useful details to making the academic early-warning mechanism more
effective. The work of this system is to dynamically monitor the behavior of col-
lege students inside and outside the classroom, using the data mining technology
of big data to analyze the collected information, and discover the abnormal phe-
nomenon or behavior of college students. Under the big data environment, we
use the technology of Outlier Mining Algorithm to find out the data away from
the regular objects and launch the analysis of the outlier data. From the data
itself, outlier data are deviations from normal data, which may be produced
due to errors in the data formation process, and are often eliminated directly
in early data analysis. However, due to data from reality, a lot of outlier data
is not coming into being due to the errors, but the corresponding data source
itself does contain a special behavior that may also mean a very important
information that indicates the emergence of a new situation, which need us to
distinguish, so the mining and analysis of outlier data is of great significance.
The general idea of outlier mining is that: in a data set with n data objects,
a desired outlier objects k (k ≤ n) is given, and dig out a number of outliers
data that are significantly different from the rest of the objects in the data set
[1]. There are a number of outlier mining algorithms, including statistic-based,
distance-based and bias-based detection algorithms, and outlier-based data using
conventional data mining algorithms. Through scientific monitoring and analy-
sis, we can timely identify the adverse factors, issue academic early-warning to
avoid the depravation in academic.
4.3 Program-Making System

The Course Signals System divides students into three groups: green, yellow,
and red according to the current state [2]. The green signal indicates that the
students are likely to achieve the goal if he or she continues the current learning
state. The yellow signal indicates that the student has potential difficulties in
learning the courses, while the red signal indicates that students may fail in the
exams. Under the big data background, we are able to output the visualization
of the early-warning signals by using the technology of big data mining and
analysis, so that students can continue to maintain the current status or make
corresponding improvements.
(1) The early-warning program of the red signal: for red alert students, firstly,
we need to analyze the reasons of their learning difficulties one by one, and
arrange mutual help with each other. Counselor or class teacher would better
chant with each students as much as possible to help them find causes of
learning decline, and then seek out appropriate learning methods to improve
academic grade. Secondly, for red warning students, we can be appropriate
to reduce their extracurricular activities, and increase learning time, learning
content. At the same time, we should not encourage them to take too much
non-professional training curriculum.
(2) The early-warning program of the yellow signal: it is better to give encourage-
ment to the yellow alert students. We could not only affirmed their existing
achievements, but also put forward higher requirements. What’s more, the
yellow warning for students can also be divided into two types to be treated
in different. Some students study hard but not get good achievements. Their
academic performance is not good, which is not caused by laziness. They take
a lot of time to learn, but do not learn too much, we can arrange study-well
students to give them one-on-one help, and offer guidance on learning meth-
ods. Other students lack initiative in learning, it is better to carry out strict
management measures to require poor autonomy students. For example, we
can take certain treatment measures, such as the centralized management,
self-study at night, roll-call system.
(3) The early-warning program of the green signal: the green students are very
excellent. We should pay attention to control their complacency, and put
forward higher requirements to keep the existing achievements and continue
to improve their comprehensive ability. Besides, we should ask them to learn
well in the professional courses, doing well in the minor course at the same
time, and inspire them to reach high standards in its computer level and
English ability.
4.4 Information Release System of Academic Early-Warning
Release system is helpful for university students to understand their current

academic situation by setting up information release system of academic early-
warning. The school provides corresponding help for the students’ academic dif-
ficulties. The parents strengthen supervision and guidance, it will achieve better
warning effect through tripartite linkage among the school, family and students.
On one hand, when the students log on to their academic early-warning system,
the relevant early-warning information will automatically inform and highlight
prompt, students can timely understand their own learning status by this confi-
dential and practical warning, and they can form a certain psychological warning
to take some self-remedial measures according to their specific circumstances. So
we can set up release system by the method of abnormal affecting elements feed-
back [6].
Abnormal Affecting Elements Feedback
The key attribute subspace kas (p) of outlier object P is the main cause of
data outliers, and can be used to accurately feedback the students’ academic
abnormalities with kas (p). Release system from two aspects of the q test results
for output. Release system output test results of q times from tow aspects.
(1) Personalize abnormal elements analysis of outlier objects. Set the key
attribute subspaces of the same outlier object p from every test are kas(p),
i = 1, 2, · · · , p. If P in the i time is not outlier, then kas(p) is an empty
set. Set P is outlier in qo , cj is the occurring times of attribute aj in all
kas(p). Apparently, the bigger is cj , the more times of attribute aj affecting
P to outlier. Appraisal the affecting of attribute aj to P abnormal study by
cj /po , to draw personalize abnormal elements affection chart. For example,
one student is judged as abnormal 7 times in 10 tests data in one week, which
is more than 10/2, so he gets an orange warning and his personal affecting
elements and their degree of affecting are: Job performance is 5/7, times of
being late in class is 2/7, internet surfing duration is 4/7, downstream flow
is 3/7, community activity duration is 3/7, staying in dormitory duration
is 1/7, which shows the student’s constitute of study abnormal elements by
Fig. 2.
(2) Total degree of affecting analysis of element aj , set the abnormal degrees of
affecting of aj in every test are cj /k, i = 1, 2, · · · , q. Draw elements continu-
ously affecting chart to show the total situation of study abnormal affecting
elements.
Fig. 2. The individual influence about abnormal factors
5 The Process Model of Academic Early-Warning

Mechanism
In the big data environment, the main process of academic early-warning mecha-
nism is information collection, information analysis, and information prediction
and information release [10]. The process model of academic early-warning of
college students is constructed based on the technology of big data. As shown in
Fig. 3.
(1) The information collection system of academic early-warning: using the large
database of information collection in university and the large database of
education outside the university collects the internal and external informa-
tion in colleges and universities, and using big data technology to initialize
and classify them.
Fig. 3. The process model of academic early-warning of college students is constructed

based on the technology of big data.
(2) The information analysis and processing system of academic early-warning:

further cleaning and transforming to the collected information by using the
big data analysis and processing technology, and using the outlier data min-
ing and analysis technology of the big data to analyze the large number of
data, focusing on the abnormal information and identifying adverse factors
to make an early prevention.
(3) The program-making system of academic early-warning: according to Course
Signals System, the early-warning of academic is divided into green, yellow
and red pre-warning, and the big data is used to visualize the early-warning
signals.
(4) The information release system of academic early-warning: college students
get the feedback of results, and take measures to adjust learning state, which
will provide feedback information to begin the next round of information
collection.
6 Conclusion
Big data can collect data that is unable to obtain or acquire costly in the past,
and more easily utilize their teaching level and management ability in the process
of teaching analysis to fully enhance students’ academic performance, and it will
also set off a huge revolutionary by bringing influence in the field of education.
However, the goal of education is the growth of each person that driving people
to recognize things with a rational thinking, and gradually form a relatively
perfect thinking of self-consciousness, which is so complex and hard to make
perfect forecast. Therefore, on one hand, big data services for education in the
academic early-warning system to be required cautious implementation in the
practical application. On the other hand, it is so attractive and challenging
to create personalized education through big data, more confidence should be

given to design more detailed academic early-warning mechanisms to find more
targeted interventions.
Acknowledgements. We would like to thank International Society of Management

Science and Engineering Management and its workers. This article also benefit from the
proposals offered by seminar participants of the 416 office at University of Electronic
Science and Technology of China.
References
1. Annika W, Zdenek Z et al (2014) Improving retention: predicting at-risk students
by analyzing clicking behavior in a virtual learning environment
2. Arnold KE, Pistilli MD (2012) Course signals at purdue: using learning analytics
to increase student success. http://www.itap.purdue.edu/learning/does/research/
ArnolejPistilli-Purdue-University-Course-Signals-2012.pdf2015-12-31
3. Cao W (2014) Analysis of the mechanism of the early-warning and aid for the stu-
dents with learning disabilities. Higher Eng Educ Res 2014(5):93–96 (in Chinese)
4. Chen Q (2012) To explore the early warning mechanism of college students under
the credit system. J Guangxi Teachers Educ Univ 2012(28):60–65 (in Chinese)
5. Ding F (2015) An empirical study on the performance of academic early-warning
of college students. Frontiers Pract 2015(2):72–73 (in Chinese)
6. Ge D, Zhang S (2015) Educ Data Mining Methods Appl. Educational Science
Press, Beijing
7. Hua J (2016) Learning early-warning system in Taiwan University. Educ Sci
2016(3):41–44 (in Chinese)
8. Manyika J, Chui M, et al (2011) Big data: the next frontier for innovation, com-
petition, and productivity. Analytics pp 3–5
9. U.S. Department of Education Office of Educational Technology (2010) Trans-
forming American education: learning powered by technology. http://www.ed.gov/
sites/default/files/netp2010.pdf
10. Wei S (2016) Digging the value of education data in the era of big data. Mod Educ
Technol 23(2):5–11 (in Chinese)
11. Witten IH, Frank E (2000) Data mining: practical machine learning tools and tech-
niques with Java implementations. San Francisco: Morgan Kaufmann Publishers
Inc
Bottleneck Management of Multi-stage
Sorting-Packing Operations with Large-Scale
Warehouse System
Tetsuya Sato(B) and Hiroshi Katayama
Faculty of Science and Engineering, Waseda University, Tokyo 169-8555, Japan

tetsuya sato@aoni.waseda.jp
Abstract. Integration of plural local DCs (Distribution Centers) into

a large-scale DC with automatic warehouse system is considered as an
effective strategy among logistics industries for responding to the diver-
sifying needs of customers and markets, optimizing distribution opera-
tions and resource saving. However, as numerous items must be handled
in such integrated DC, adequate design of the components such as ware-
house, sorting-packing lines and their operation system is a critical issue
for maintaining competitive advantage. This paper, focusing on the last
issue, discusses the utility of TCM (Target Chasing Method) based dis-
patching policies for bottleneck management of entire process. Various
PCVs (Pseudo Control Variables) are examined to utilized, which repre-
sent the state of the system, easy to observe but indirectly controllable
(e.g. number of items and waiting time in queue, work load rate in each
sorting-packing line). Performance analysis is examined by simulation
experiments and obtained results suggest that smooth flow of items in
each stage of sorting-packing line is realized by the proposed dispatch-
ing policies. Especially, dispatching policy for Just-In-Time delivery to
customers that intends to adapt delivery time to customers, reveals the
weak portion of the entire system.
Keywords: Automatic warehouse system · Sorting-packing opera-

tions · Multi-stage optimization · Heijunka production · Target chasing
method · Just-in-time delivery
1 Introduction
In recent years, responding to the diversifying needs of customers and markets
with optimizing distribution operations and resource saving are the trends among
logistics industries. However, it requires to handle various types of products in
small lots and the efficiency of each operation are decreased in conventional
supply chain with plural local DCs (Distribution Centers). Thus, integration of
local DCs into a large-scale DC with AWS (Automatic Warehouse System) is
considered as an effective strategy for improvement of products flow in entire
Bottleneck Management of Multi-stage Sorting-Packing Operations 1055
supply chain. As numerous items must be handled in such integrated DC, ade-
quate design of the components such as warehouse, sorting-packing lines and
their operation system is a critical issue for maintaining competitive advantage.
For optimizing products flow in entire process of large-scale integrated DC,
this paper discusses the utility of TCM (Target Chasing Method) [2,4–7] based
dispatching policies. TOC (Theory of Constraints) [1] identifying and improving
the bottleneck processes in multi-stage operations is an effective way for better-
ment of productivity. In this study, various PCVs (Pseudo Control Variables)
[3,8] are examined to utilized, which represent the state of the system, easy to
observe but indirectly controllable. Performance analysis is examined by sim-
ulation experiments and obtained results suggest that smooth flow of items in
each stage of sorting-packing line is realized by the proposed dispatching poli-
cies. Especially, dispatching policy for Just-In-Time delivery to customers that
intends to adapt delivery time to customers, reveals the weak portion of the
entire system.
2 Experimental Method
2.1 Target Chasing Method (TCM)
TCM, which is frequently used in automotive industry, is a heuristical algorithm

of sequencing method. Basically, TCM controls the order of job entry for min-
imizing the gap between ideal and real resource usage of the facility at each
timing. Figure 1 shows the schematic diagram of TCM. Vertical and horizontal
axes are the order of job entry and resource usage in the facility, respectively.
From this figure, it is noticed TCM intends to pursue Heijunka (load levelling)
production by controlling the order of job entry. The objective function of TCM
is expressed in Eq. (1).
Objective Function:

R 2
k × Nj
min (Dk ) = min − Xjki . (1)
i∈I
j=1
Q
Symbols:
k : Order of job entry;
i : Lot number;
j : Resource type;
I : Set of the products;
R : Number of total resource;
Q : Total number of all products;
Nj : Usage of resource j for manufacturing all products;
Xjki : Usage of resource j for manufacturing from 1st to k th product i;
Dk : The gap between ideal and real resource usage after k th job entry.
1056 T. Sato and H. Katayama
Fig. 1. Schematic diagram of target chasing method
2.2 Pseudo Control Variables (PCVs)
This paper focuses on the utility of PCVs [3,8], which represents the state of
the system, easy to observe but indirectly controllable (e.g. number of items
and waiting time in queue, work load rate in each sorting-packing line), for
TCM-based dispatching policy. Figure 2 shows the schematic diagram of the
controlled behavior of TCM using PCVs. Vertical and horizontal axes of Fig. 2
are the PCV values of each resource. In this case, there are two facilities in the
same operational stage and the target of PCVs (diagonal line in Fig. 2) is defined
as the same values between PCV1 and PCV2. Therefore, TCM controls current
status of PCVs to the target line for Heijunka production. For example, the
value of PCV2 is higher than PCV1 on the point P of Fig. 2, thus TCM tries to
control PCV2 to decrease and PCV1 to increase for reducing the gap between
actual and target PCVs.
PCV2
Target of PCVs
Control
P
Time evolution
P'
0 PCV1
Fig. 2. Schematic diagram of target chasing method using pseudo control variables
2.3 Objective System
The objective system which is the multi-stage sorting-packing system in the

large-scale DC with AWS is shown in Fig. 3. Dotted arrows in this figure are
the routes of products flow from arrival to shipment. At first, handled products
are arrived from plural local suppliers and warehoused to AWS. Products in
the warehouse are sorted by AWS and dispatched to packing area. Products are
re-installed to AWS after packing operation. This process is repeated in terms
of multi-stage operation and packaged products are shipped to local assemblers.
Table 1 shows the facilities and details of the operations in each stage of the
objective DC. The 1st and 2nd stage of sorting and packing operations are the
target to optimize by TCM based dispatching policies in this study, which is
indicated by asterisks in Table 1. From Fig. 3, bottleneck management of the
products flow at each packing station is required for efficient Just-In-Time deliv-
ery to the customers.
Local DCs
Integrate
Large-scale DC
Arrival &
Suppliers Warehousing Automatic Warehouse 1
Stock
Yard 1st Sorting Stage ･･･
Automatic Warehouse 2
Packing Stations
Area 1
2nd Sorting Stage
･･･
Assemblers Shipment
Chute
Yard
1st Packing Stage
& Re-Warehousing
Packing
･･････
Area 2
Shipping Stations 2nd Packing Stage
Fig. 3. Schematic diagram of objective large-scale distribution center
2.4 Simulation Model
This paper analyzes the performance of proposed TCM-based dispatching poli-

cies by simulation experiments for a downscaled problem, as original large-scale
system is too complicated to tackle with. Figure 4 shows the schematic diagram
of the defined simulation model. The buffers before each conveyor and AWS are
inserted for dividing and identifying the bottlenecks of products flow in each
stage.
Table 1. Details and facilities of the operations in each stage of the objective distrib-
ution center
Operation Detail of operation Facility

Arrival Products arrive from suppliers Stock Yard
Warehousing Arrived products are stored to buckets and Stock Yard
transferred to Automatic Warehouse
1st sorting stage∗ Warehouse system controls the dispatch Automatic Warehouse 1
order and the dispatched station of each
product by the unit of a bucket
1st packing stage∗ Products are extracted from the buckets Packing Area 1
and labeled. After labeling operation, the
products are packed to inner boxes and
re-stored to the buckets
Re-warehousing Products are re-installed to Automatic Packing Area 1
Warehouse
2nd sorting stage∗ Warehouse system controls the dispatch Automatic Warehouse 2
order and the dispatched station of each
product by the unit of a bucket
2nd packing stage∗ Required quantity of inner boxes with Packing Area 2
products are packaged to shipping boxes (or
pallets)
Shipping Products are shipped to assemblers Shipment Yard
∗
Table 1 indicates the sub-processes of “Warehousing” and “Re-Warehousing”
Aisle 1
In order to input jobs to buffer.
Shuttle Racks Shuttle
Stock Yard Buffer Conveyor Buffer Buffer Conveyor
Aisle 2
Packing Area 1
Aisle 3
Buffer Station 1
Buffer Station 2 Buffer Conveyor Buffer
Buffer Station 3 Shuttle Racks Shuttle

Aisle 4
Packing Area 2
Buffer Station 4
Buffer Conveyor Shipment Yard
Buffer Station 5
Fig. 4. Schematic diagram of the downscaled simulation model
The aim of this paper to analyze the performance of dispatching policies in

AWS at each sorting stage; therefore, this paper unconsiders warehousing policy
of AWS. The parameters of the simulation experiments are defined by properly
deformed figures as follows:
• Number of product types: 3 (Product 1–3);

• Maximum load of the conveyors: 3 buckets;
• Supplying pitch of the conveyors: 1 bucket per unit time;
• Number of racks in each aisle: 100 racks;
Table 2. Facilities, processing time and lot size of each operation
Operation Facility Processing time Lot size

1st sorting stage Aisle 1 in Automatic Warehouse 1 (2 + 3 =)5∗ 1
Aisle 2 in Automatic Warehouse 1 (2 + 3 =)5∗ 1
1st packing stage Station 1 in Packing Area 1 3 1
Station 2 in Packing Area 1 3 1
2nd sorting stage Aisle 3 in Automatic Warehouse 2 (2 + 3 =)5∗ 1
Aisle 4 in Automatic Warehouse 2 (2 + 3 =)5∗ 1
2nd packing stage Station 4 in Packing Area 2 15 6
∗
Processing time of the aisles in each sorting stage are the sum of traveling times
on the shuttles and the conveyors, respectively.
Table 3. Part of the generated lists of job entry and shipping order
(a) The list of job entry (Input) (b) The list of shipping order (Output)
Time of Product Station of Station of Shipping Product Lot Station of Station of
job entry code packing packing order code size packing packing
Area 1 Area 2 Area 1 Area 2
1 1 3 4 1 3 4 2 5
2 2 3 4 2 1 4 1 5
3 1 3 4 3 1 6 2 4
4 2 3 4 4 2 4 2 5
5 1 3 4 5 2 4 2 5
6 3 3 5 6 3 6 2 4
7 2 3 4 7 2 6 1 4
8 3 3 5 8 2 4 1 5
9 3 3 5 9 2 6 3 4
10 3 3 4 10 3 6 2 4
11 3 2 4 11 3 6 1 4
12 2 2 5 12 3 6 3 4
13 3 2 5 13 2 6 3 4
14 2 2 5 14 1 4 1 5
15 1 2 4 15 3 6 2 4
··· ···
2000 1 2 5 410 2 6 2 4
• Buffer capacity of each packing station: 50 buckets;

• Number of job input to a queue: 2,000 products (expected as warehousing
products in a week);
• Facilities, processing time and lot size of each operation: shown in Table 2.
The lists of job entry and shipping order of packaged products have been gen-
erated randomly in advance, which are shown in Table 3. Note that all products
are assigned to operational stations in each packing stage beforehand of ware-

housing based on the job entry list which described the details of the packing
operations in each station.
Performances of the proposed dispatching policies are analyzed by the num-
ber of items and waiting time in queue and work load rate in each sorting-packing
line. Work load rate in sorting line (AWS) is defined as the rack occupancy. The
definitions of waiting time and work load rate in packing lines are expressed in
Eqs. (2) and (3), respectively.
Waiting time in queue:
xj x x
QTi = OTi j − ITi j . (2)
Work load rate in packing line:

N xj xj
xj i=1 STi
WR = . (3)
T
Symbols:
xj : The j th facility x;
xj
OTi : Dispatched time of product i from facility xj ;
QTixj : Waiting time of product i in facility xj ;
ITixj : Job entry time of product i into facility xj ;
W Rx j : Work load rate in facility xj ;
T : Total processing time of entire operations;
N xj : The number of products of which operation is finished in facility xj .
The proposed TCM-based dispatching policies in each stage of sorting oper-
ation using AWS are shown in Table 4. The multi-stage operations are improved
by a combination of the policies in the 1st and 2nd sorting stage. Regarding to
the policy code S − Xa or S − Xb, its components, i.e. S, X, a or b, represent
the stage of the sorting operations, the index of the policies and number of items
or waiting time in queue for PCVs, respectively. For example, the combination
of policy codes 1-1a and 2-0 represents adoption of indicated dispatching rule to
the 1st sorting stage and dispatching rule to the 2nd sorting stage described in
Table 4, respectively.
3 Experimental Results and Discussion

3.1 Analysis of the Present Dispatching Policy (Benchmark)
At first, this paper analyzes the performance of the present dispatching policy
that is used as a benchmark. Figure 5 shows the time series of the number of
items in queue at each packing station, where the policies 1-0 and 2-0 are applied.
As a result, the number of items in each queue are changed independently by
the corresponding policies in each packing stage. Especially, from Fig. 5(a), it is
observed that buffer capacities of each station that is defined as 50 buckets in
Packing Area 1 tend to be fully occupied.
Table 4. Proposed TCM based dispatching policies for each stage sorting operations
1st Sorting stage 2nd Sorting stage Notes

Policy Detail Policy Detail
1-0∗ : <Benchmark> To dispatch 2-0∗ : <Benchmark> To dispatch In case of the order of
the transactions by FCFS the transactions lot where all shipping schedule is
(First-Come, First-Served) the items are ready to pack not considered.
principle in the 2nd packing stage
1-1a: To dispatch the transactions 2-1a: To dispatch the transactions
to the station by TCM, where to the station by TCM, where
number of items in queue is number of items in queue is
less than other stations at less than another station at
the 1st packing stage the 2nd packing stage
1-1b: To dispatch the transactions 2-1b: To dispatch the transactions
to the station by TCM, to the station by TCM, where
where waiting time in queue waiting time in queue is
is shorter than other stations shorter than another station
at the 1st packing stage at the 2nd packing stage
1-2a: To dispatch the transactions 2-2a: To dispatch the transactions In case of the order of
to the station by TCM with to the station by TCM with shipping schedule is
consideration of shipping consideration of shipping considered.
schedule, where number of schedule, where number of
items in queue is less than items in queue is less than
other stations at the 1st another station at the 2nd
packing stage packing stage
1-2b: To dispatch the transactions 2-2b: To dispatch the transactions
to the station by TCM with to the station by TCM with
consideration of shipping consideration of shipping
schedule, where waiting time schedule, where waiting time
in queue is shorter than other in queue is shorter than
stations at the 1st packing another station at the 2nd
stage packing stage
∗
Current dispatching policies at each sorting stage.
Fig. 5. The time series of the number of items in queue at each packing station in case
of policies 1-0 and 2-0
Table 5 shows the summary of criteria values in each operation in case of

policies 1-0 and 2-0. The number of items and waiting time in queue of Packing
Area 1 are both worse values than Packing Area 2 as well as the work load rate.
Thus it is noticed that the 1st packing stage is the worst bottleneck operation
in entire system.
Table 5. Averages and standard deviations of the criteria values in case of policies 1-0
and 2-0
(a) Number of items in Queue [units] (b) Waiting time in Queue [unit time] (c) Work load rate
AWS1 Packing AWS2 Packing AWS1 Packing AWS2 Packing AWS1 Packing AWS2 Packing
Area 1 Area2 Area 1 Area2 Area 1 Area2
Ave 0.661 38.681 44.011 21.503 1.000 57.524 30.803 31.534 0.33% 50.01% 22.01% 83.44%
SD 0.813 52.552 35.066 18.402 0.000 55.398 30.315 18.612 0.46% 49.94% 7.80% 37.17%
3.2 Bottleneck Management of the 1st Packing Stage

Based on the results in the last section, this paper analyzes the performance of
proposed TCM-based dispatching policy applied to the worst bottleneck oper-
ation observed in the 1st packing stage. Figure 6 shows the time series of the
number of items in queue at each packing station where the policies 1-1a and
2-0 are applied. In this case, policy 1-1a uses the number of items in queue as
PCVs for TCM-based dispatching policy. From Fig. 6(a), the number of items
in queue try to mutually equalize the same PCV values of other stations at the
1st packing stage by TCM-based dispatching. Note that the products are re-
installed after the 1st packing operation and dispatched from AWS2 by policy
2-0, as in the last section. Thus, Fig. 6(b) are slightly changed from Fig. 5(b);
nevertheless, the tendency is similar to the benchmark result.
of policies 1-1a and 2-0

policies 1-1a and 2-0. The number of items and waiting time in queue at Packing
Area 1 are extremely decreased from the benchmark result described in Sect. 3.1
and the work load rate in AWS1 is increased. For this reason, it is expected that
the proposed TCM-based dispatching policies using PCVs facilitates Heijunka
production effectively in the objective system.
In addition, policy code 1-1b uses the waiting time in queue as PCVs for
TCM-based dispatching policy; however, the tendency of the criteria values is
similar. Thus, this paper discusses only the cases that use the number of items
in queue as PCVs in the following sections.
Table 6. Averages and standard deviations of the criteria values in case of policies
1-1a and 2-0
Ave 68.191 2.967 45.355 17.671 58.185 3.527 32.874 25.967 34.10% 49.58% 22.68% 82.73%
SD 72.045 3.912 36.089 15.591 75.736 3.209 34.045 17.020 21.77% 49.94% 7.73% 37.80%
3.3 Bottleneck Management of Entire System

From the benchmark result in Sect. 3.1, the second worst bottleneck operation
in the system is identified, that is Packing Area 2. Due to this reason, this paper
analyzes and discusses the proposed TCM-based policy for improving products
flow of both the 1st and 2nd packing stages in this section. Figure 7 shows the
time series of the number of items in queue at each packing station where the
policies 1-1a and 2-1a are applied. From this figure, the number of items, which
is defined as PCV in the policies, in each station at both the 1st and 2nd packing
stages try to mutually equalize the same PCV values by TCM-based dispatching.
of policies 1-1a and 2-1a

policies 1-1a and 2-1a. The number of items and waiting time in queue of the
2nd packing stage are decreased from the result of the last section, not only the
1st stage. Therefore, proposed TCM-based dispatching policies are utilized for
bottleneck management with Heijunka concept of entire operation. On the other
hand, the work load rate of Packing Area 2 in Tables 5, 6 and 7 are almost the
same level. This is caused by the lot sizes and the processing time of each station
at Packing Area 2 which are larger and longer than other processes, that is, after
the improvement of this section, the bottleneck operation of the entire system,
which is the 2nd packing stage, has not been transferred to other operations from
the result of the last section. Thus, more improvement of operations throughput
in Packing Area 2 is required for efficient improvement of the entire system, not
only the preceding stages.
Table 7. Averages and standard deviations of the criteria values in case of policies
1-1a and 2-1a
Ave 68.191 2.967 54.340 8.358 58.285 3.527 39.689 11.755 34.10% 49.58% 27.17% 82.73%
SD 72.045 3.912 44.170 7.497 75.736 3.209 47.651 7.435 21.77% 49.94% 10.19% 37.80%
3.4 Bottleneck Management with Consideration of the Shipping

Schedule
Finally, this paper analyzes bottleneck management with consideration of the
shipping schedule. The assignment of all products to the operational stations at
each packing stage in Table 3(a) decreases the degree of freedom in optimization
of dispatching policy; therefore, product jamming is occurred by optimization
with consideration of the shipping order. For this reason, the investigation of
this section focuses the utility of dynamic reassignment of the products to each
operational station. Figure 8 shows the time series of the number of items in
queue at each packing station considering scheduled shipping order as Table 3(b),
where the policies 1-2a and 2-2a are applied. As a result, the number of items
in each station at both the 1st and 2nd packing stages try to mutually equalize
the same PCV values, which is similar to the analyzed result of the last section.
8 Station1 20 Station4
Station2 Station5
Number of Items in Queue
Number of Items in Queue
Station3
6 15
4 10
5
2
0
0
0 500 1000 1500 2000 2500 3000 0 500 1000 1500 2000 2500 3000
Processing Time [unit time] Processing Time [unit time]
(a) Number of Items in Queue at the 1st Packing Stage (b) Number of Items in Queue at the 2nd Packing Stage
of policies 1-2a and 2-2a
Table 8 shows the criteria values in each stage in case of policies 1-2a and 2-
2a. In addition, the proposed dispatching policy dependent characteristics of the
number of items in queue, waiting time in queue and the work load rate at each
stage are summarized in Fig. 9. The bar and line charts in this figure are the aver-
ages and standard deviations of the criteria values, respectively. From Table 8,
the work load rate in AWS 1 is decreased from the result of the last section
shown in Table 7(c); nevertheless, the number of the items and the waiting time
in queue of AWS 1 are extremely decreased from Table 7(a) and (b). This is
Table 8. Averages and standard deviations of criteria values in case of policies 1-2a
and 2-2a
Ave 4.807 3.554 50.920 10.444 5.154 4.180 36.578 14.222 2.40% 51.92% 25.46% 86.62%
SD 11.048 4.554 46.495 8.753 20.815 3.276 74.151 7.768 7.18% 49.90% 12.27% 34.04%
Fig. 9. Dispatching policies dependent characteristics of criteria values at each stage
caused by efficient dispatching is realized by dynamic reassignment of products

to each operational station at the 1st packing stage. In contrast, the criteria value
of AWS 2 and Packing Area 2 in Table 8 are almost the same level as in Table 7.
This is caused by the lot sizes and the processing time of each station at Pack-
ing Area 2 which are larger and longer than other processes as similar to the
last section. Therefore, proposed TCM-based dispatching policies with dynamic
reassignment are utilized for efficient bottleneck management with Heijunka pro-
duction of the entire operation.
This paper examined the performance and the effectiveness of the proposed
TCM-based dispatching policy using PCVs for bottleneck management of multi-
stage sorting-packing operations in large-scale AWS-based DC. From the results,
Heijunka production in each packing stage is expected to realize by the proposed
dispatching policies. Moreover, this paper analyzed the case of given scheduled
shipping order. In this case, dynamic reassignment of the transactions to each
operational station at the 1st and 2nd packing stage facilitates optimization
of dispatching operation effectively. Therefore, the proposed TCM-based dis-
patching policies with dynamic reassignment are utilized for effective bottleneck
management for realizing Heijunka operation of the entire system.
The aim of this paper is the performance analysis of dispatching policies at
each sorting stage in AWS for its relevant installation. Heijunka of the work load
rate of each aisle at multi-stage AWS and operations at each stage are required
for further improvement of throughput; which is a possible subject for future
analysis.
Acknowledgements. This work was partially supported by JSPS KAKENHI

(Grants-in-Aid for Scientific Research of Japan Society for the Promotion of Science)
Grant Number JP16K17199.
References
1. Goldratt EM, Cox J (1984) The goal: a process of ongoing improvement. The North
River Press, Great Barrington
2. Hwang R, Katayama H (2009) Integrated procedure of balancing and sequencing
for mixed-model assembly lines: a multi-objective evolutionary approach. Int J Prod
Res 40(21):6417–6441
3. Ishikawa S, Murata K, et al. (2012) Bottleneck management on multi-process
sorting-packing system in distribution centre TCM implementation and its applica-
bility reinforcement. In: Proceedings of The 15th J.S.L.S. national conference. The
Japan society of logistics systems, 2nd–3rd June, pp 143–148 (in Japanese)
4. Katayama H (2006) Lean technology and its extensions-a case of target chasing
method. In: Proceedings of the 11th Cambridge symposium on international manu-
facturing (CSIM 2006), Institute for Manufacturing, University of Cambridge, UK,
27th–28th September, 15 pages in CD-ROM
5. Katayama H, Tanaka M (1996) Target chasing method for work load stabilisation
of production line-some advances of lean management technology. In: Proceedings
of the 20th international conference on computers & industrial engineering (20th
ICC & IE). Korean Institute of Industrial Engineers, Kyongju, 6th–9th October, pp
425–428
6. Kotani S (1983) Sequencing algorithms for the mixed-model assembly line. Toyota
Technol 33(1):31–38 (in Japanese)
7. Monden Y (1983) Toyota production system: an integrated approach to just-in-time,
2nd edn. Industrial Engineering & Management Press, Norcross
8. Sato T, Ishikawa S, Katayama H (2016) Bottleneck management on multi-process
sorting-packing system in large-scale distribution centre. In: Proceedings of the 2016
annual autumn meeting of japan industrial management association. Japan Indus-
trial Management Association, Tokyo, pp 188–189 (in Japanese)
A Carbon-Constrained Supply Chain Planning Model
Zhimiao Tao and Jing Xu(B)

taozhimiao@scu.edu.cn, 20210431@qq.com
Abstract. This paper focuses on a simplified supply chain planning network

consisting of manufacture and distribution in the context of carbon emission con-
straint. The problem is formulated as a bilevel programming model mathemati-
cally. An effective solution strategy is then presented to handle the bilevel model.
A numerical example is proposed to illustrate the proposed model and algorithms.
This study highlights the impact of carbon emissions constraint on the operational
decisions of supply chain.
Keywords: Bilevel programming model · Supply chain planning · Carbon emis-

sions · Carbon footprint
1 Introduction
More evidences show that human activity leads to global climate change via carbon
emissions. Curbing the carbon emissions and controlling climate change have become
a great challenge that all humanity has to face. As the main carbon emitters, firms must
take their carbon emissions seriously.
A conventional thinking to curb the carbon emissions is to improve energy effi-
ciency and install emissions control devices. However, such initiatives require capi-
tal investments. Meanwhile, efficient business practices and operational policies are
another way to change carbon emissions. Different manufacture or order frequency
may result in different emissions. Obviously, this way is more cost-effective, in con-
trast to the way of investing novel equipments. Some researchers have focused on the
strategy to reduce the carbon emission through supply chains. Benjaafar et al. [5] inte-
grated the carbon emission concerns into the operational decision-making with regard
to procurement, production, and inventory management. They provided a series of oper-
ational models under different regulatory policies. Lee [15] used empirical case study
of HMC (Hyundai Motor Company) and its key supplier’s front bumper product to
illustrate the issue of carbon footprint in supply chain management. Absi et al. [1] mod-
eled the integration of carbon emission constraints in lot-sizing problems. Helmrich
et al. [11] showed that lot-sizing with an emission capacity constraint is NP-hard and
proposed several solution methods. Hua et al. [13] investigated how firms manage car-
bon footprints in inventory management under the carbon emission trading mechanism.
Chen et al. [9] provided analytical support for the notion that it may be possible, via
1068 Z. Tao and J. Xu
operational adjustments alone, to significantly reduce emissions without significantly

increasing cost. He et al. [10] examined the production lot0sizing issues of a firm under
cap-and-trade and carbon tax regulations, respectively.
All the models in the literature mentioned above assume that there is a only deci-
sion maker and formulate as the problems as single level models. Single level decision-
making structures don’t reflect the real supply chain management. A supply chain natu-
rally involves multiple decision components, which are collected in a hierarchical way.
Individual activities are often governed by separate supply chain components which
have their own, often mutually conflicting, objectives [7, 17, 19]. In addition, different
participant components in the supply chain network may not operate with the same
level of information. Some may possess more information, while other may possess
less information, this may lead to information distortion [12, 22, 23].
Using a bilevel programming model, we study how carbon emission concerns can be
integrated into supply chain planning. We examine how the values of emission control
policies affect cost and emissions. An interactive approach is proposed to solve the
model.
The rest of this paper is organized as follows. In Sect. 2, we make a problem state-
ment as the motivation. In Sect. 3, we formulate a bi-level programming model for the
supply chain planning problem under with strict carbon caps. An interactive approach
is depicted in Sect. 4. In Sect. 5, we use the model to obtain managerial insights from
the numerical experiment and discuss the implications of these insights to management
practice and to public policy making. We summarize key findings and offer ideas for
future research in Sect. 6.
2 Problem Statement
A simple supply chain normally consists of several manufacturing plants and one ware-
house or distribution centre (DC), with several products (see Fig. 1).
Plant 1 DC 1 Customer 1
Plant 2 DC 2 Customer 2
. . .
. . .
. . .
Plant I DC J Customer K
Fig. 1. A supply chain network
Given there is common decision maker who manages the whole supply chain, the
supply chain planing problem can be formulated as an single level decision-making
A Carbon-Constrained Supply Chain Planning Model 1069
model. However, An alternative, more realistic, way to pose this problem is by recog-
nizing: (i) the natural hierarchy that exists between the production and the distribution
part of the supply chain and (ii) the fact that full information against the distribution
part may not be really available at the production part and vice versa.
Without carbon emission considerations, the plants and the DCs determine the pro-
duction amount and the inventory amount, respectively. Their objective In the presence
of carbon emission considerations, the plants and the DCs must account for the emis-
sions associated with various decisions regarding ordering, production, and inventory
holding. In order to develop the mathematical model, we make some condition assump-
tions as follows:
(1) The emissions are linearly increasing in the associated decision variables;
(2) All the plant belongs one owner and all the DCs belongs to the other owner. DCs’
owner has priority in decision;
(3) Customers’ demands are stochastic;
(4) There is only one period in a planning horizon.
3 Model Formulation
In this section, we develop a mathematical model for the problem described above.
3.1 Notations
Before formulating the mathematical model, the needed notations are listed as follows.
Indices:
i: Plant (1, · · · , I);
j: DC ( j, · · · , J);
k: Customer (1, · · · , K);
l: Product (1, · · · , L).
Parameters:
αil : Capacity coefficient of product l at plant i;
βil : Resource coefficient of product l at plant i;
γ jl : Resource coefficient of product l at plant j;
ail : Production cost coefficient for product l at plant i;
bi jl : Transportation cost coefficient for product l from plant i to DC j;
h jkl : Inventory holding cost coefficient for product l at DC j for customer k;
tr jkl : Transportation cost coefficient for product l from DC j to customer k;
ãil : Emission coefficient due to production for product l at plant i;
b̃i jl : Emission coefficient due to transportation for product l from plant i to DC j;
h̃ jkl : Emission coefficient due to inventory for product l at DC j for customer k;
jkl : Emission coefficient due to transportation for product l from DC j to customer k;
tr
Mkl : Demand of product l at customer k;
Pi : Production capacity of plant i;
Q: Resources available to all the plants;
CP : Emission cap for plants;

CD : Emission cap for DCs.
Decision variable:
Yi jl : Production amount of l at plant j for DC j;
X jkl : Inventory of product l at DC j for customer k.
3.2 Modelling
For the owner of the plants, his (or her) to minimize the costs, which consists of its
manufacturing cost and distribution cost between plants and DCs. Hence the objective
of the production part ZPC can be developed as
I J L I J L
min ZPC = ∑ ∑ ∑ ailYi jl + ∑ ∑ ∑ bi jlYi jl . (1)
Yi jl
i=1 j=1 l=1 i=1 j=1 l=1
We consider the strict emission cap regulation, the cap on carbon emissions of produc-
tion part cannot be exceeded, i.e.,
I J L I J L
∑ ∑ ∑ ãilYi jl + ∑ ∑ ∑ b̃i jlYi jl ≤ CP . (2)
i=1 j=1 l=1 i=1 j=1 l=1
Since production amounts from the plants should meet the levels required at the DC,
we have
I K
∑ Yi jl ≥ ∑ X jki , ∀ j, l. (3)
i=1 k=1
The plant capacities are formulated as
L J
∑ ∑ ailYi jl ≤ Pi , ∀i. (4)
l=1 j=1
Common used resources may be shared by all the plants, that is

L I J
∑ ∑ ∑ βilYi jl ≤ Q. (5)
l=1 i=1 j=1
For the distribution part, the decision maker’s objective is to minimize the costs, then
J K L J K L
min ZDC =
X jkl
∑ ∑ ∑ h jkl X jkl + ∑ ∑ ∑ tr jkl X jkl , (6)
j=1 k=1 l=1 j=1 k=1 l=1
in which the first term denotes inventory holding cost including material handling cost
at DCs and the second indicates transportation cost from warehouses to customers. The
carbon emissions from the distribution part also satisfies the emission cap, so
J K L J K L
∑ ∑ ∑ h̃ jkl X jkl + ∑ ∑ ∑ tr jkl X jkl ≤ CD . (7)
j=1 k=1 l=1 j=1 k=1 l=1
Sums of individual DCs’ holding should meet demands of customers, that is

J
∑ X jkl ≥ Mkl , ∀k, l. (8)
j=1
Demands’ constraints (8) are stochastic types. In other words, we cannot ensure that the
kl are known. Some techniques have
constraints are satisfied before the exact values of M
been developed to handle the stochastic case like this. Here, the chance-constrained
technique proposed by Charnes and Cooper [8] is applied. Let θkl denote prescribed
confidence levels. Applying chance-constrained technique to (8), the corresponding
constraints are
J
Pr ∑ X jkl ≥ M kl ≥ θkl , ∀k, l, (9)
j=1
where Pr means probability.
Inventory levels are limited by individual DC capacities, so
K L
∑ ∑ γkl X jkl ≤ R j , ∀ j. (10)
k=1 l=1
In many cases, the decision maker has priority decisions on the production part are
affected by parameters which are decided by the distribution part. For instance, produc-
tion levels are decided from given information regarding the inventory conditions. Thus
the supply chain planning model can be posed as the following bilevel decision-making
problem: ⎧ J K L J K L
⎪ min Z = ∑ ∑ ∑ h X + ∑ ∑ ∑ tr X
⎪
⎪
⎪ DC jkl jkl jkl jkl
⎪
⎪ X jkl
⎧ J Kj=1Lk=1 l=1 j=1 k=1 l=1
⎪
⎪
⎪
⎪ ⎪
⎪ ∑ ∑ ∑ h̃ jkl X jkl + ∑ ∑ ∑ tr
J K L
jkl X jkl ≤ CD
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪ j=1
⎪
⎪ ⎪
⎨
k=1 l=1 j=1 k=1 l=1
⎪
⎪ J
⎪
⎪ s.t. Pr ∑ X jkl ≥ M kl ≥ θkl , ∀k, l
⎪
⎪ ⎪
⎪
⎪ ⎪
⎪ j=1
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪
K L
⎪
⎪ ⎩ ∑ ∑ γkl X jkl ≤ R j
⎪
⎪
⎪
⎪
k=1 l=1
⎪
⎪ where
⎧ Yi jl solves
⎪
⎪
⎨⎪ ⎪
I J L I J L
⎪
⎪ min ZPC = ∑ ∑ ∑ ail Yi jl + ∑ ∑ ∑ bi jl Yi jl
⎪ ⎪ i jl
⎪ Y (11)
⎪
⎪ ⎪
⎪ ⎧ I Ji=1Lj=1 l=1 i=1 j=1 l=1
⎪
⎪ ⎪
⎪ ⎪
I J L
⎪
⎪ ⎪
⎪ ⎪
⎪ ∑ ∑ ∑ ãil Yi jl + ∑ ∑ ∑ b̃i jl Yi jl ≤ CP
⎪
⎪⎪ ⎪
⎪ ⎪
⎪
⎪
⎪ ⎪ ⎪
⎪
i=1 j=1 l=1 i=1 j=1 l=1
⎪
⎪ ⎪
⎪ ⎪
⎪ I K
⎪
⎪ ⎪
⎨ ⎪
⎪ ∑ Yi jl ≥ ∑ X jki , ∀ j, l
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎨ i=1 k=1
⎪
⎪ ⎪
⎪
L J
⎪
⎪ ⎪
⎪ s.t. ∑ ∑ ail Yi jl ≤ Pi , ∀i
⎪
⎪ ⎪
⎪ ⎪
⎪
⎪ ⎪ ⎪
⎪ l=1 j=1
⎪
⎪ ⎪
⎪ ⎪ L I J
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ∑ ∑ ∑ βil Yi jl ≤ Q
⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ l=1 i=1 j=1
⎪
⎪⎪ ⎪ ⎪ X ≥ 0, ∀ j, k, l
⎪
⎩⎪
⎪ ⎪
⎩
⎪
⎩ jkl
Yi jl ≥ 0, ∀i, j, l.
4 Algorithm
Algorithms are the bridges between the theoretical models and the practical applica-
tions. The NP-hardness of bilevel programming problems make it full of challenges
to design efficient algorithms. Researchers have proposed many approaches to solve
bilevel programming problems [2, 4, 6, 14, 16, 18]. Among these algorithms, the interac-
tive fuzzy programming methods are popular in recent years [3, 20, 24].
Model (11) can be reformulated in a general bilevel programming framework as
follows: ⎧
⎪ maxn F(x, y)
⎪
⎪ x∈R⎧ G(x, y) ≤ 0
⎪ 1
⎪
⎨ ⎪ ⎪
⎪
⎨ where y solves: (12)
⎪
⎪
⎪ s.t. max f (x, y)
⎪ ⎪ y∈Rn2
⎩ ⎪
⎪ ⎪
⎩
s.t. g(x, y) ≤ 0,
where, x ∈ Rn1 is the decision vector for the upper-level decision maker (leader) and
y ∈ Rn2 is decision variable for the lower-level decision maker (follower); F(x, y) is
the objective function of the upper-level model and f (x, y) is objective function of
the lower-level model; G(x, y) is the constraint of the upper-level programming and
g(x, y) is the constraint of the lower-level model.
The key idea is that the decision makers at both levels consider individual objectives
separately as the initialization. Let D = {x, y|G(x, y) ≤ 0, g(x, y) ≤ 0}. For each of
the objective functions, assume that the decision makers have fuzzy goals such as “the
objective should be substantially less than or equal to some value”. The individual best
values
F min = min F(x, y), f min = min f (x, y)
(x,y )∈D (x,y )∈D
and the individual worst values
F max = max F(x, y), f max = max f (x, y)

(x,y )∈D (x,y )∈D
are referred to when the decision makers elicit membership functions prescribing the
fuzzy goals for the objective functions. The decision makers determine the member-
ship functions μ1 (F(x, y)) and μ2 ( f (x, y)) which are strictly monotone decreasing for
objective functions, consulting the variation ratio of degree of satisfaction in the interval
between the individual best values and the individual values. The domains of the mem-
bership functions of leader and follower are the interval [F min , F min ] and [ f min , f min ],
respectively. For the sake of simplicity, in this paper, we adopt a linear membership
function, which characterizes the fuzzy goal of the decision maker at each level. The
corresponding linear membership function μ1 (F(x, y)) and μ2 ( f (x, y)) are defined as:
⎧
⎨ 0 max if F(X1 , X2 ) > F max
F −F(x,y )
μ1 (F(x, y)) = if F min ≤ F(x, y) ≤ F max (13)
⎩ F max −F min
1 if F(x, y) < F min ,
and ⎧
⎪
⎨ 0 max if f (X1 , X2 ) > f max
f − f (x,y )
μ2 ( f (x, y)) = f max − f min
if f min ≤ f (x, y) ≤ f max (14)
⎪
⎩1 if f (x, y) < f min .
After eliciting the membership functions, leader subjectively specifies a minimal
satisfactory level δ̂ ∈ [0, 1] for his/her membership function μ1 (F(x, y)). Then, fol-
lower maximize his/her membership function subject to the condition that leader’s
membership function μ1 (F(x, y)) is larger than or equal to δ̂ under the given con-
straints, that is, followers solves the following problem:
⎧
⎪
⎨ xmax μ2 ( f (x, y))
∈Rn1
μ1 (F(x, y)) ≥ δ̂ (15)
⎪
⎩ s.t.
(x, y) ∈ D.
If an optimal solution to problem (15) exists, it follows that DM1 obtains a satisfac-
tory solution having a satisfactory degree larger than or equal to the minimal satisfactory
level specified by leader’s own self. However, the larger the minimal satisfactory level
is assessed, the smaller follower’s satisfactory degree becomes. Consequently, a relative
difference between the satisfactory degrees of leader and follower becomes larger and
it is feared that overall satisfactory balance between both levels cannot maintain.
To take account of overall satisfactory balance between both levels, leader needs to
compromise with follower on leader’s minimal satisfactory level. To do so, a satisfac-
tory degree of both decision makers is defined as
λ = min(μ1 (F(x, y)), μ2 ( f (x, y))).
and the following problem is substituted for problem (15):

⎧
⎪
⎪ max⎧λ
⎪
⎪
⎨ ⎪ ⎪ μ (F(x, y)) ≥ δ̂ ≥ λ
⎨ 1
μ2 ( f (x, y)) ≥ λ (16)
⎪
⎪ s.t.
⎪
⎪ ⎪ λ ∈ [0, 1]
⎩ ⎪ ⎩
(x, y) ∈ D.
For problem (16), consider the auxiliary problem

⎧
⎪
⎪ max⎧λ
⎪
⎪
⎨ ⎪ ⎪ μ1 (F(x, y)) ≥ λ
⎨
μ2 ( f (x, y)) ≥ λ (17)
⎪ s.t.
⎪
⎪
⎪ ⎪ λ ∈ [0, 1]
⎩ ⎪ ⎩
(x, y) ∈ D.
By solving problem (17), we obtain a solution maximizing a smaller satisfactory

degree between those of both decision makers.
If an optimal solution (x∗ , y ∗ ) to problem (17) satisfies the condition that
μ1 (F(x∗ , y ∗ )) ≥ δ̂ , it follows that leader obtains a satisfactory solution. However, the
solution (x∗ , y ∗ ) does not always satisfy the condition. Then the ratio of satisfactory
degree between both levels
μ2 ( f (x∗ , y ∗ ))
Δ= (18)
μ1 (F(x∗ , y ∗ ))
is adopted.
If Δ > 1, i.e., μ2 ( f (x∗ , y ∗ ) > μ1 (F(x∗ , y ∗ )), then leader updates the minimal satis-
factory level δ̂ by increasing its value. Receiving the updated level δ̂ , follower solves
problem (15) with δ̂ and then leader obtains a larger satisfactory degree and fol-
lower accepts a smaller satisfactory degree. Conversely, if Δ < 1, i.e., μ2 ( f (x∗ , y ∗ ) <
μ1 (F(x∗ , y ∗ )), then leader updates the minimal satisfactory level δ̂ by decreasing its
value, and leader obtains a smaller satisfactory degree and follower accepts a larger
satisfactory degree. The interactive process terminates when two conditions hold [21].
5 Numerical Example
Consider supply chain consisting of three plants, two DCs, three customers, with two
products. The values of parameters are listed in Table 1. The demands follow the normal
distribution, defined as
Table 1. Parameter values
α11 = 4 α12 = 3 α21 = 5 α22 = 2 α31 = 3 α32 = 3

β11 = 7 β12 = 10 β21 = 12 β22 = 6 β31 = 8 β32 = 9
γ11 = 6 γ12 = 8 γ21 = 10 γ22 = 4
a11 = 1 a12 = 3 a21 = 2 a22 = 4 a31 = 3 a32 = 2
b111 = 2 b112 = 3 b121 = 3 b122 = 4 b211 = 2 b212 = 2
b221 = 1 b222 = 5 b311 = 4 b312 = 2 b321 = 3 b322 = 3
h111 = 5 h112 = 3 h121 = 6 h122 = 3 h131 = 5 h132 = 5
h211 = 5 h212 = 3 h221 = 6 h222 = 3 h231 = 5 h232 = 5
tr111 = 5 tr112 = 3 tr121 = 6 tr122 = 3 tr131 = 5 tr132 = 6
tr211 = 4 tr212 = 4 tr221 = 4 tr222 = 3 tr231 = 5 tr232 = 4
ã11 = 0.1 ã12 = 0.2 ã21 = 0.2 ã22 = 0.4 ã31 = 0.3 ã32 = 0.2
b̃111 = 0.2 b̃112 = 0.3 b̃121 = 0.3 b̃122 = 0.4 b̃211 = 0.2 b̃212 = 0.2
b̃221 = 0.1 b̃222 = 0.5 b̃311 = 0.4 b̃312 = 0.2 b̃321 = 0.3 b̃322 = 0.3
h̃111 = 0.5 h̃112 = 0.3 h̃121 = 0.6 h̃122 = 0.3 h̃131 = 0.5 h̃132 = 0.5
h̃211 = 0.5 h̃212 = 0.3 h̃221 = 0.6 h̃222 = 0.3 h̃231 = 0.5 h̃232 = 0.5
111 = 0.5 tr
tr 112 = 0.3 tr
121 = 0.6 tr
122 = 0.3 tr
131 = 0.5 tr
132 = 0.6
˜ 211 = 0.4 tr
tr ˜ 212 = 0.4 tr
˜ 221 = 0.4 tr
˜ 222 = 0.3 tr
˜ 231 = 0.5 tr
˜ 232 = 0.4
11 ∼ N (200, 10),

M 12 ∼ N (300, 20),
M 21 ∼ N (240, 10),
M
22 ∼ N (180, 10),
M 31 ∼ N (280, 20),
M 32 ∼ N (160, 10).
M
The resources available to all the plants are 10000. The production capacities of the
plants are 1800, 1500 and 1200, respectively. Assume that all θkl = 0.8.
Let CP = CD = 6000. Suppose that the initial minimal satisfactory level as δ = 1.0,
and the lower and the upper bounds of Δ as [0.6, 1.0]. Using the algorithm described in
Sect. 4, the decisions are obtained after three iterations:
∗ ∗ ∗ ∗ ∗ ∗
X111 = 52.5, X112 = 22.1, X121 = 12.8, X122 = 26.8, X131 = 72.3, X132 = 44.2,
∗ ∗ ∗ ∗ ∗ ∗
X211 = 12.6, X212 = 62.5, X221 = 45.2, X222 = 92.5, X231 = 12.5, X232 = 14.5,
∗ ∗ ∗ ∗ ∗ ∗
Y111 = 82.5, Y112 = 72.5, Y121 = 62.5, Y122 = 20.2, Y211 = 32.0, Y212 = 26.5,
∗ ∗ ∗ ∗ ∗ ∗
Y221 = 65.3, Y222 = 34.5, Y311 = 17.5, Y312 = 42.5, Y321 = 54.3, Y322 = 62.5.
The corresponding costs of leader and follower are ZDC ∗ = 60512.8 and Z ∗ = 32746.8,
PC
∗
respectively. The amounts of their emissions are E DC = 26128.2 and EPC ∗ = 11321.4,
respectively.
In order to test the effect the emission caps on costs, we use different emission caps
to calculate the costs and actual emissions. For simplicity, let CDC = CPC . The results
are shown by Fig. 2.
Fig. 2. The effect of emission cap on cost and carbon emission
Figure 2 shows the impact of varying the emission cap on the total cost and total
emissions for the examples considered. Both the leader and follower have the similar
trends. As expected, reducing the emission cap increases total cost and reduces total
emissions. However, what is perhaps surprising is the fact that the emission cap can
be significantly reduced without significantly affecting the total cost. In the example
shown, reducing the emission cap from 6000 to 3000 reduces the average total amount
of emissions by 50% but increases the average total cost by only 10%. Note that this
reduction is obtained through adjustments in operational decisions only, namely, modi-

fying order quantities in each period, could alone lead to significant reductions in carbon
emissions while not significantly compromising overall cost.
6 Conclusions and Future Research

In this paper, we presented a bilevel decision-making model to illustrate how carbon
footprint considerations could be incorporated into supply chain model. These insights
highlight the impact of operational decisions on carbon emissions and the extent to
which adjustments to operations can mitigate emissions. Our results show that oper-
ational adjustments alone can lead in some cases to significant emission reductions
without significant increases in cost. More significantly, the results highlight the oppor-
tunity that exists to leverage collaboration across the supply chain to mitigate the cost
of reducing emissions. Insights described in this paper revealed the importance of inter-
actions between firms in determining overall emission levels and corresponding costs.
The results in this paper will be helpful for the decision makers of supply chains to
address challenges from strict emission caps regulation.
However, there are still some issues to worth attention.
(1) More quantitative analysis. The results in this paper are based on a numerical
example. More quantitative analysis can reveal the essence behind the phenom-
enon, which is conducive to the application of the model.
(2) Scenario analysis under different regulatory emission control policies. In this
paper, only the strict emission cap policy is considered. However, policies of the
carbon tax, carbon offset and cap-and-trade are also used wildly. It is necessary to
compare them deeply.
Acknowledgements. This research was supported by the Programs of NSFC (Grant No.
71401114), the Fundamental Research Funds for the Central Universities (Grant No.
skqy201524).
References
1. Absi N, Dauzère-Pérès S et al (2013) Lot sizing with carbon emission constraints. Eur J Oper
Res 227(1):55–61
2. Angelo JS, Barbosa HJ (2015) A study on the use of heuristics to solve a bilevel program-
ming problem. Int Trans Oper Res 22(5):861–882
3. Baky IA (2014) Interactive topsis algorithms for solving multi-level non-linear multi-
objective decision-making problems. Appl Math Model 38(4):1417–1433
4. Beheshti B, Özaltın OY et al (2015) Exact solution approach for a class of nonlinear bilevel
knapsack problems. J Glob Optim 61(2):291–310
5. Benjaafar S, Li Y, Daskin M (2013) Carbon footprint and the management of supply chains:
insights from simple models. IEEE Trans Autom Sci Eng 10(1):99–116
6. Bialas WF, Karwan MH (1984) Two-level linear programming. Manage Sci 30(8):1004–
1020
7. Chang X, Xia H et al (2015) Production decisions in a hybrid manufacturing-

remanufacturing system with carbon cap and trade mechanism. Int J Prod Econ 162:160–173
8. Charnes A, Cooper WW (1959) Chance-constrained programming. Manage Sci 6(1):73–79
9. Chen X, Benjaafar S, Elomri A (2013) The carbon-constrained eoq. Oper Res Lett
41(2):172–179
10. He P, Zhang W et al (2015) Production lot-sizing and carbon emissions under cap-and-trade
and carbon tax regulations. J Cleaner Prod 103:241–248
11. Helmrich MJR, Jans R et al (2015) The economic lot-sizing problem with an emission capac-
ity constraint. Eur J Oper Res 241(1):50–62
12. Hsueh CF (2015) A bilevel programming model for corporate social responsibility collab-
oration in sustainable supply chain management. Transp Res Part E Logistics Transp Rev
73:84–95
13. Hua G, Cheng T, Wang S (2011) Managing carbon footprints in inventory management. Int
J Prod Econ 132(2):178–185
14. Kuo R, Lee Y et al (2015) Solving bi-level linear programming problem through hybrid of
immune genetic algorithm and particle swarm optimization algorithm. Appl Math Comput
266:1013–1026
15. Lee KH (2011) Integrating carbon footprint into supply chain management: the case of
hyundai motor company (hmc) in the automobile industry. J Cleaner Prod 19(11):1216–1223
16. Lv Y, Wan Z (2015) Solving linear bilevel multiobjective programming problem via exact
penalty function approach. J Inequalities Appl 1:1–12
17. Ma W, Wang M, Zhu X (2014) Improved particle swarm optimization based approach for
bilevel programming problem-an application on supply chain model. Int J Mach Learn
Cybern 5(2):281–292
18. Rahmani A, MirHassani S (2015) Lagrangean relaxation-based algorithm for bi-level prob-
lems. Optim Methods Softw 30(1):1–14
19. Sadigh AN, Mozafari M, Karimi B (2012) Manufacturer-retailer supply chain coordination:
a bi-level programming approach. Adv Eng Softw 45(1):144–152
20. Sakawa M, Matsui T (2014) Interactive fuzzy stochastic multi-level 0–1 programming using
tabu search and probability maximization. Expert Syst Appl 41(6):2957–2963
21. Sakawa M, Nishizaki I, Uemura Y (1998) Interactive fuzzy programming for multilevel lin-
ear programming problems. Comput Math Appl 36(2):71–86
22. Wang D, Du G et al (2016) A stackelberg game theoretic model for optimizing product
family architecting with supply chain consideration. Int J Prod Econ 172:1–18
23. Yeh K, Whittaker C et al (2015) Two stage stochastic bilevel programming model of a pre-
established timberlands supply chain with biorefinery investment interests. Comput Chem
Eng 73:141–153
24. Zheng Y, Liu J, Wan Z (2014) Interactive fuzzy decision making method for solving bilevel
programming problem. Appl Math Model 38(13):3136–3141
Supply Chain Coordination by Revenue Sharing
Contract Under Different Carbon Emission
Policies
Li Lu(B)
Tourism School, Sichuan University, Chengdu 610065, People’s Republic of China

luli rudy@126.com
Abstract. To solve the problem of global warming, governments have

carried out some corresponding policies to reduce carbon emissions. In
this paper, we considered a two-echelon supply chain consisting of a
supplier and a manufacturer, and researched the supply chain coordi-
nation by revenue sharing contract under two different carbon emission
policies, Mandatory carbon emissions capacity and Cap-and-trade. The
study indicates that: under both two carbon policies, the revenue shar-
ing contract can coordinate the supply chain. Manufacturer’s revenue
sharing proportion under cap-and-trade is always less than the situation
of no carbon emissions constraint. The expected profits of manufacturer
under cap-and-trade can be greater than the situation without carbon
emissions constraints in certain conditions.
Keywords: Supply chain coordination · Revenue sharing contract ·

Mandatory carbon emissions capacity · Cap-and-trade
1 Introduction
It is a common knowledge that large amount of carbon emissions directly leads
to global warming which has brought serious challenges to human survival and
development. The best way to solve this problem is changing the way of pro-
duction and life. So realize the sustainable development of low-carbon economy
is becoming a hot spot of global attention. Nowadays, different governments
including China have carried out the corresponding carbon emissions regulation
policy according to their national conditions. Typical carbon policies are Manda-
tory carbon emissions capacity (Cap) and Cap-and-trade. Both the two policies
are carbon constraints under the regulation of government, and the government
will assign different initial carbon emissions quota to different enterprises. As we
know, cap policy is a kind of mandatory constraint. The enterprises cannot sale
or buy the carbon emissions quota in carbon trading market when the initial is
overmuch or insufficient. However, cap-and-trade policy allows enterprises sale
or buy the carbon emissions quota in carbon trading market when the initial
is overmuch or insufficient, in order to meet the production requirements and
Supply Chain Coordination by Revenue Sharing Contract 1079
improve yields [13]. The earliest implementation of Cap and Cap-and-trade poli-
cies is European Emissions Trading System (EU ETS), and has now become the
world’s largest carbon emissions trading market [2].
The implementation of carbon policy in operation management has brought
new challenges to the enterprise. It makes enterprise management more complex
in performance target for decision-making (dual goals of increasing profits and
reducing carbon emissions), the decision variables (production, order, pricing and
other traditional decision variables and carbon emissions, carbon trading deci-
sion variables) and decision environment (production capacity, capital and other
traditional and limitation of carbon constraints). As we all know, carbon emis-
sions are throughout the entire supply chain, however, the existing researches
on carbon policies in practice are mainly based on the perspective of individual
enterprise. If only focus on individual enterprise, and cannot effectively solve the
supply chain coordination problem of upstream and downstream enterprises, we
cannot fundamentally promote carbon policy’s implementation, and achieve the
ultimate goal of reducing carbon emission [15].
Revenue sharing contract is a common supply chain coordination mechanism.
Downstream enterprise buys goods from upstream enterprise at a relatively low
wholesale price, but shares the sales revenue with upstream enterprise [3]. In
this paper, we analyze the supply chain coordination problem based on revenue
sharing contract, and research to solve the following problems:
(1) Whether revenue sharing contract could coordinate the supply chain under
different carbon policy?
(2) How does the different carbon policy affect the revenue sharing ratio?
(3) How does the different carbon policy affect the manufacturer’s expected
profit?
The rest of this paper is organized as follows: Sect. 2 reviews the relevant
literatures; Sect. 3 describes the problem and assumptions; Sect. 4 presents the
model and analysis; and discussion is presented in Sect. 5; finally Sect. 6 con-
cludes this paper.
2 Literature Review
In this paper, we review literature spanning across two streams. The first section
addresses literature on supply chain coordination by revenue sharing contract.
The second section addresses literature on supply chain coordination under dif-
ferent carbon policies.
2.1 Supply Chain Coordination by Revenue Sharing Contract
From the perspective of economics, Mortimer et al. [16] carried on the empirical
research on the application of revenue sharing contract in the DVD rental indus-
try. Giannoccaro et al. [10] systematically studied the problem of supply chain
1080 L. Lu
coordination by revenue sharing contract. Cachon and Lariviere [4] comprehen-

sively analyzed of the advantages and limitations of revenue sharing contract.
The results showed, although the revenue sharing contract could achieve supply
chain coordination, the upstream enterprise did not achieve the optimal as this
mechanism was instable. Kong et al. [14] analyzed the supply chain coordina-
tion by revenue sharing contract under the circumstance of information disclo-
sure. Govindan et al. [11] took PC industry as an example, and analyzed the
reverse supply chain coordination problem based on revenue sharing contract.
The results indicated that revenue sharing contract could coordinate the reverse
supply chain effectively. Arani et al. [1] introduced a novel mixed revenue-sharing
option contract to coordinate a retailerCmanufacturer supply chain.
2.2 Supply Chain Coordination Under Different Carbon Policies
Ghosh et al. [8] took garment industry as an example, and studied the influence
of green level on prices and profits, then put forward a kind of contract includ-
ing two parts of tax to coordinate the green supply chain. Swami et al. [17]
considered the consumers’ environmental consciousness, and thought that green
operation efforts of the retailers and manufacturers in the supply chain would
affect consumer demand, then designed a cost sharing contract to realize supply
chain coordination. Jaber et al. [12] considered a two-stage supply chain con-
sisting a manufacturer and a retailer, and studied the supply chain coordination
mechanism under Cap-and-trade policy when manufacturer undertook the cost
of carbon emission. Based on the newsboy model, Choi [6] analyzes the impact
of wholesale price and price subsidies contract under Cap-and-trade policy on
retailers procurement source selection. Du et al. [7] considered a two-stage supply
chain consisting of a carbon emission dependent enterprise and a carbon emission
supplier under Cap-and-trade policy, and designed a supply chain coordination
mechanism using the non-cooperative game theory. Ghosh and Shah [9] explored
supply chain coordination issues arising out of green supply chain initiatives and
explore the impact of cost sharing contract on the key decisions of supply chain
players undertaking green initiatives. Xu [18] analyzed the decision behavior
and coordination mechanisms for a two-echelon sustainable supply chain under
a cap-and-trade regulation. Cao et al. [5] investigated the government’s role in
allocating the appropriate emission quota to maximize social members’ utilities
and analyzes how the emission-dependent enterprise improves revenues of both
itself and the whole system through supply chain collaboration.
Through literature review, we found that the revenue sharing contract was a
widely used supply chain coordination mechanism. But most of the researches did
not consider the impact of different carbon policy on supply chain coordination
mechanism. In order to fill the gap, we will analyze the supply chain coordination
problem based on revenue sharing contract under different carbon policies.
3 Problem Description and Assumption

In this paper, taking the two echelon supply chain (one supplier and one man-
ufacturer) which is dominant by supplier as the research object, we analyze the
supply chain coordination problem based on revenue sharing contract under Cap
and Cap-and-trade policy.
Before the sales season, the manufacturer purchases raw materials Q at
wholesale price w from supplier. The manufacturer faces the random demand
D with probability density f (x) and distribution function F (x), and sells the
product at retail price p. At the end of the sales season, the salvage value per
unit product is v, the shortage cost per unit product is g. The manufacture keeps
ϕi ∈ [0, 1] (i = 1, 2, 3) part of own revenue, and gives 1 − ϕi part to the supplier,
where i = 1, 2, 3 respectively represents the situation of no carbon constraint,
Cap policy and Cap-and-trade policy. The supplier’s purchasing cost is cs , the
manufacture’s producing cost is cm , carbon emission for producing unit product
is e, trading price for unit carbon emission quota is k, and the carbon Cap set
by the government is E.
For the convenience of description, considering the decentralized decision,
let Q∗ϕ1 , Q∗ϕ2 , Q∗ϕ3 present manufacture’s optimal order quantity under no car-
bon constraint, Cap and Cap-and-trade policy respectively; considering the inte-
grated decision, let Q∗1 , Q∗2 and Q∗3 present supply chain’s optimal order quantity
under no carbon constraint, Cap and Cap-and-trade policy respectively. In addi-
tion, for the convenience of research, we make the following assumptions:
(1) The supplier and the manufacturer are risk neutral and perfectly rational,
and the information between supply chain enterprises is completely symmet-
rical.
(2) The product’s selling cycle is short, the market is stable, and during the
sales period, the product’s price remains the same.
(3) Carbon policy only applies to manufacturer.
(4) p > w + cm indicates that the sale price of product is always greater than
the manufacturer’s production costs, and the manufacturer is profitable.
(5) w > cs indicates that supplier’s wholesale price is always greater than cost,
supplier is also profitable.
4 Model and Analysis

4.1 The Base Model Without Carbon Emission Constraint
In this scenario, we get the manufacturer’s expected profit
E [πm1 (Qϕ1 )] = ϕ1 pS (Qϕ1 ) + ϕ1 vI (Qϕ1 ) − gL (Qϕ1 ) − (w + cm ) Qϕ1 , (1)

Q
where S (Qϕ1 ) = Q−∫0 ϕ1 f (x) dx presents the expected sales volume, I (Qϕ1 ) =
Q
∫0 ϕ1 (Qϕ1 − x) f (x) dx presents the expected inventory volume, and L (Qϕ1 ) =
∫Q
+∞
ϕ1
(x − Qϕ1 ) f (x) dx presents the expected shortage volume.
1082 L. Lu
While the supplier’s expected profit is
E [πs1 (Qϕ1 )] = [w − cs + (1 − ϕ1 ) v] Qϕ1 + (1 − ϕ1 ) (p − v) S (Qϕ1 ) . (2)

dE[πm1 (Qϕ1 )]
As dQϕ1 = [ϕ1 (p − v) + g] [1 − F (Qϕ1 )] − (w + cm − ϕ1 v),
2
d E[πm1 (Qϕ1 )]
d2 Qϕ1 = − [ϕ1 (p − v) + g] f (Qϕ1 ) < 0, there exists a unique optimal
order quantity Q∗ϕ1 which maximizes E [πm1 (Qϕ1 )].
= 0, we get optimal order quantity Q∗ϕ1 =
dE[πm1 (Qϕ1 )]
Let
dQϕ1
F −1 ϕ1ϕp+g−w−c m
. To coordinate the supply chain, we have to make Q∗1 =
1 (p−v)+g

p+g−cs −cm p+g−cs −cm
Q∗ϕ1 , that is F −1 p+g−v = F −1 ϕ1ϕp+g−w−c
1 (p−v)+g
m
, we get p−v+g =
ϕ1 p+g−w−cm
ϕ1 (p−v)+g , so ϕ1 = (p−v)(cm +w)+(w−cs )g+vg
(p−v)(cm +cs )+vg .
In order to guarantee both sides accept this contract, we must ensure that
their respective benefits under the revenue sharing contract are greater than
those before, namely realizing Pareto improvement. We should have
E [πs1 (Q∗1 )]
> 1, (3)
E πs1 Q∗ϕ1
E [πm1 (Q∗1 )]
> 1. (4)
E πm1 Q∗ϕ1
As E [πs1 (Q∗1 )] + E [πm1 (Q∗1 )] = E [πT 1 (Q∗1 )], E [πs1 (Q∗1 )] = (1 − ϕ1 )

E [πT 1 (Q∗1 )], E [πm1 (Q∗1 )] = ϕ1 E [πT 1 (Q∗1 )], put into the formula, we get
(1−ϕ1 ) E[πT 1 (Q∗1 )] ϕ1 E[πT 1 (Q∗1 )]
E [πm1 (Q∗ϕ1 )]
∗
E [πs1 (Qϕ1 )]
> 1, ∗
E [πm1 (Qϕ1 )]
> 1. To simplify, we get E [πT 1 (Q∗
< ϕ1 <
1 )]
E [πs1 (Q∗ )]
1 − E π Qϕ1∗ .
[ T 1 ( 1 )]
The result indicates although revenue sharing contract can coordinate sup-
ply chain without carbon emission constraint, sharing proportion must meet a
certain relationship to make the Pareto improvement for both sides; otherwise
the contract shall be invalid.
According to the assumption, ϕ1 ∈ [0, 1], that is 0 ≤ (p−v)(c m +w)+(w−cs )g+vg
(p−v)(cm +cs )+vg
≤ 1, then (p − v) (cm + w)+(w − cs ) g ≤ (p − v) (cm + cs ), we can proof w ≤ cs .
It shows that supplier must make the wholesale price less than or equal to the
purchase price to achieve supply chain coordination, then the supplier’s profits
drop. Therefore, the manufacturer must share enough revenue to make up for
the loss of supplier, and make him gain more profits.
4.2 The Cap Model


E [πm2 (Qϕ2 )] = ϕ2 pS (Qϕ2 ) + ϕ2 vI (Qϕ2 ) − gL (Qϕ2 ) − (w + cm ) Qϕ2
(5)
s.t. Qϕ2 ≤ Ee
E [πs2 (Qϕ2 )] = [w − cs + (1 − ϕ2 ) v] Qϕ2 + (1 − ϕ2 ) (p − v) S (Qϕ2 ) . (6)
Proposition 1. There exists a unique optimal order quantity which maximizes

manufacturer’s expected profit by revenue sharing contract
under Cap policy.
∗ ∗ −1 ϕ2 p+g−w−cm
When Qϕ2 ≤ e , the optimal order quantity is Qϕ2 = F
E
ϕ2 (p−v)+g , when
Q∗ϕ2 > E
e, it is Q∗ϕ2 = E
e.
dE[πm2 (Qϕ2 )]
Proof. As dQϕ2 = [ϕ2 (p − v) + g] [1 − F (Qϕ2 )] − (w + cm − ϕ2 v), then
2
d E[πm2 (Qϕ2 )]
d2 Qϕ2 = − [ϕ2 (p − v) + g] f (Qϕ2 ) < 0, there exists a unique optimal
order quantity Q∗ϕ2 which maximizes E [πm2 (Qϕ2 )].
= 0, we get optimal order quantity Q∗ϕ2 =
dE[πm2 (Qϕ2 )]
Let
dQϕ2
F −1 ϕ2ϕp+g−w−c
2 (p−v)+g
m
.
∗
When Qϕ2 ≤ e , the Cap policy doesn’t play a role to the supply chain,
E
it can be seen as the situation without carbon emission constraint, Q∗ϕ2 =

F −1 ϕ2ϕp+g−w−c
2 (p−v)+g
m
.
However, when Q∗ϕ2 > Ee , the Cap policy plays a role to the supply chain,
as the manufacturer’s profit function is concave function, and it is monotone
increasing at the left hand of Q∗ϕ2 . That makes the optimal order quantity be
Q∗ϕ2 = Ee . This completes the proof.

Proposition 2. When Q∗2 ≤ Ee , and Q∗2 ≤ Ee , the supply chain can coor-
dinate by revenue sharing contract in the scenario
of Cap policy where ϕ 2
(p−v)(cm +w)+(w−cs )g+vg E [πm2 (Q∗ϕ2 )] E [πs2 (Q∗ϕ2 )]
= (p−v)(cm +cs )+vg . And when ϕ2 ∈ E [πT 2 (Q∗
, 1 − E π Q∗ ,
2 )] [ T 2 ( 2 )]
it can realize Pareto improvement for both supplier and manufacturer; when
g [1−F ( E )]−w−cm
Q∗ϕ2 ≤ Ee < Q∗2 , the supply chain can coordinate where ϕ2 = F E e(p−v)−p ,
(e)
when Q∗2 > Q∗ϕ2 > Ee , the supply chain can coordinate where ϕ2 = 1.
Proof. (1) When Q∗2 ≤ Ee and Q∗ϕ2 ≤ Ee , the Cap policy doesn’t play a role.
To achieve supply chain coordination, we should make Q∗2 = Q∗ϕ2 , that is,

s −cm s −cm
F −1 p+g−c
p+g−v = F −1 ϕ2 p+g−w−cm
ϕ2 (p−v)+g , we get p+g−c
p−v+g = ϕ2ϕp+g−w−c
2 (p−v)+g
m
.
To simplify, we obtain ϕ2 = (p−v)(c m +w)+(w−cs )g+vg
(p−v)(cm +cs )+vg .
In order to make sure that both sides realize Pareto improvement,
E[πs2 (Q∗ E[π (Q∗ )]
we should have E [πs2 (Q∗
2 )]
> 1, E π m2 Q∗2 > 1. As E [πs2 (Q∗2 )] +
ϕ2 )] [ m2 ( ϕ2 )]
E [πm2 (Q∗2 )] = E [πT 2 (Q∗2 )], E [πs2 (Q∗2 )] = (1 − ϕ2 ) E [πT 2 (Q∗2 )],
∗ ∗ ϕ2 E[πT 2 (Q∗ )]
E [πm2 (Q2 )] = ϕ2 E [πT 2 (Q2 )], put into the formula, we get E π Q∗ 2 > 1.
[ m2 ( ϕ2 )]
E [πm2 (Q∗ )] E [πs2 (Q∗ )]
To simplify, we get E π Qϕ2 ∗ < ϕ2 < 1 − E π Qϕ2∗ .
[ T 2 ( 2 )] [ T 2 ( 2 )]
1084 L. Lu
(2) When Q∗ϕ2 ≤ Ee < Q∗2 , the Cap policy plays a role. Let Q∗2 = Q∗ϕ2 , that
E ϕ2 p+g−w−cm
is F −1 ϕ2ϕp+g−w−c
2 (p−v)+g
m
= E
e . To simplify, we obtain F e = ϕ2 (p−v)+g , then
E
g [1−F ( )]−w−cm
ϕ2 = F E e(p−v)−p .
(e)
(3) When Q∗2 > Q∗ϕ2 > Ee , as Q∗2 = Q∗ϕ2 = Ee , the supply chain doesn’t need
to coordinate, the manufacture keeps all the profit, now ϕ2 = 1.
This completes the proof.

This proposition suggests that, when carbon cap doesn’t play a role, supply
chain can coordinate by revenue sharing contract under Cap policy and realize
Pareto improvement for both sides; when carbon cap plays a role, although
we can find the condition which can achieve supply chain coordination, it is
unable to realize the Pareto improvement for both sides. Because the Cap policy
damages the whole supply chain’s profit, it is difficult to achieve supply chain
coordination in reality.
4.3 The Cap-and-Trade Model

⎧
⎪ E [πm3 (Qϕ3 )] = ϕ3 pS (Qϕ3 ) + ϕ3 vI (Qϕ3 ) − gL (Qϕ3 ) when Q∗ϕ3 ≥ E
⎪
⎨
− (w + cm ) Qϕ3 − k (eQϕ3 − E) e
(7)
⎪
⎪ E [πm3 (Qϕ3 )] = ϕ3 pS (Qϕ3 ) + ϕ3 vI (Qϕ3 ) − gL (Qϕ3 )
⎩ whenQ∗ϕ3 < Ee .
− (w + cm ) Qϕ3 + ϕ3 k (E − eQϕ3 )
When eQ∗ϕ3 ≥ E, k (eQϕ3 − E) presents the cost of buying carbon quota,

and when eQ∗ϕ3 < E, k (E − eQϕ3 ) presents the revenue of selling carbon quota.
⎧
⎪
⎪ E [πs3 (Qϕ3 )] = [w − cs + (1 − ϕ3 ) v] Qϕ3
⎪
⎪ when Q∗ϕ3 ≥ Ee ;
⎨ + (1 − ϕ3 ) (p − v) S (Qϕ3 )
E [πs3 (Qϕ3 )] = [w − cs + (1 − ϕ3 ) v] Qϕ3 (8)
⎪
⎪ ∗
⎪
⎪ + (1 − ϕ 3 ) [(p − v) S (Qϕ3 ) when Q < E
.
⎩ ϕ3 e
+k (E − eQϕ3 )]
Proposition 3. There exists a unique optimal order quantity which maxi-

mizes manufacturer’s expected profit by revenue sharing contract under Cap-
and-trade policy. When Q∗ϕ3 ≥ Ee the optimal order quantity is Q∗ϕ3 =

m −ke −1 ϕ3 p+g−w−cm −ϕ3 ke
F −1 ϕ3 p+g−w−c
ϕ3 (p−v)+g ; when Q∗
ϕ3 < E
e , it is Q∗
ϕ3 = F ϕ3 (p−v)+g .
Proof. (1) When Q∗ϕ3 ≥ E

e, E [πm3 (Qϕ3 )] = [ϕ3 (p − v) + g] S (Qϕ3 ) −
dE[πm3 (Qϕ3 )]
(w + cm + ke−ϕ3 v) Qϕ3 − gu + kE. As dQϕ3 = [ϕ3 (p − v) + g]
2
d E[πm3 (Qϕ3 )]
[1 − F (Qϕ3 )] − (w + cm +ke − ϕ3 v), d2 Qϕ3 = − [ϕ3 (p − v) + g]
f (Qϕ3 ) < 0, there exists a unique optimal order quantity Q∗ϕ3 which maximizes
E [πm3 (Qϕ3 )].
= 0, we get optimal order quantity Q∗ϕ3 = F −1

dE[πm3 (Qϕ3 )]
Let
dQϕ3
ϕ3 p+g−w−cm −ke
ϕ3 (p−v)+g .
(2) When Q∗ϕ3 < E
e, E [πm3 (Qϕ3 )] = [ϕ3 (p − v) + g] S (Qϕ3 )
dE[πm3 (Qϕ3 )]
− (w + cm + ϕ3 ke − ϕ3 v) Qϕ3 − gu + ϕ3 kE. As dQϕ3 =
d2 E[πm3 (Qϕ3 )]
[ϕ3 (p − v) + g] [1 − F (Qϕ3 )] − (w + cm +ϕ3 ke − ϕ3 v), = d2 Qϕ3
− [ϕ3 (p − v) + g] f (Qϕ3 ) < 0, there also exists a unique optimal order quan-
tity Q∗ϕ3 which maximizes E [πm3 (Qϕ3 )].
0, we get optimal order quantity Q∗ϕ3
dE[πm3 (Qϕ3 )]
Let = =
dQϕ3
ϕ3 p+g−w−cm −ϕ3 ke
F −1 ϕ3 (p−v)+g . This completes the proof.

Proposition 4. When Q∗ϕ3 ≥ Ee , the supply chain can coordinate by

revenue sharing contract in the scenario of Cap-and-trade policy where
ϕ3 = (p−v)(c m +w+ke)+(w−cs )g+vg
(p−v)(cm +cs +ke)+vg ; when Q∗ϕ3 < Ee , the supply chain can
m −w)−(p−ke)g+(w−cs )g
coordinate where ϕ3 = (p−v)(g−c (p−v)(g−cm −cs )−(p−ke)g . And when ϕ3 ∈
∗ ∗

E [πm3 (Qϕ3 )] E [πs3 (Q )]
E [πT 3 (Q∗
, 1 − E π Qϕ3 , it can realize Pareto improvement for both sup-
3 )] [ T 3 ( ∗3 )]
plier and manufacturer.
p+g−cs −cm −ke
Proof. (1) When Q∗ϕ3 ≥ E
e, let Q∗ϕ3 = Q∗3 , that is p−v+g =
ϕ3 p+g−w−cm −ke (p−v)(cm +w+ke)+(w−cs )g+vg
ϕ3 (p−v)+g , to simplify, we have ϕ3 = (p−v)(cm +cs +ke)+vg .
∗ E ∗ ∗ p+g−cs −cm −ke
(2) When Qϕ3 < e , let Qϕ3 = Q3 , that is p−v+g =
ϕ3 p+g−w−cm −ϕ3 ke (p−v)(g−cm −w)−(p−ke)g+(w−cs )g
ϕ3 (p−v)+g , to simplify, we obtain ϕ3 = (p−v)(g−cm −cs )−(p−ke)g .
In order to make sure that both sides realize Pareto improve-
E[π (Q∗ )] E[π (Q∗ )]
ment, we should have E π s3 Q∗3 > 1, E π m3 Q∗3 > 1. As
[ s3 ( ϕ3 )] [ m3 ( ϕ3 )]
E [πs3 (Q3 )] + E [πm3 (Q3 )] = E [πT 3 (Q3 )], E [πs3 (Q3 )] = (1 − ϕ3 ) E [πT 3 (Q∗3 )],
∗ ∗ ∗ ∗
(1−ϕ )E[π (Q∗ )]
E [πm3 (Q∗3 )] = ϕ3 E [πT 3 (Q∗3 )], put into the formula, we get E 3π QT∗3 3 >
[ s3 ( ϕ3 )]
∗
ϕ3 E[πT 3 (Q∗ )] E [π m3 ( Q )] E [πs3 (Q∗ )]
1, E π Q∗ 3 > 1. To simplify, we obtain E π Q∗ < ϕ3 < 1 − E π Qϕ3
ϕ3
.
[ m3 ( ϕ3 )] [ T 3 ( 3 )] [ T 3 ( ∗3 )]
This completes the proof.

This proposition shows that, whether carbon emission quota is enough or not,
supply chain can coordinate by revenue sharing contract under Cap-and-trade
policy. However, sharing proportion must meet a certain relationship to make
the Pareto improvement for both sides; otherwise the contract shall be invalid.
5 Discussions
Through the analysis above, we notice that revenue sharing contract can coordi-
nate the supply chain under no carbon emission constraint, Cap and Cap-and-
trade policy. Next, we will discuss the effect of carbon policy on revenue sharing
ratio and manufacturer’s expected profit.
1086 L. Lu
5.1 Effect of Carbon Policy on Revenue Sharing Ratio
Proposition 5. When eQ∗2 ≤ E ≤ EQ∗3 , ϕ2 > ϕ3 .
Proof. We have known when Q∗2 ≤ E

e , ϕ2 =
(p−v)(cm +w)+(w−cs )g+vg
(p−v)(cm +cs )+vg , when Q∗3 ≥
(p−v)(cm +w+ke)+(w−cs )g+vg
E
e, ϕ3 = (p−v)(cm +cs +ke)+vg , ϕ2 − ϕ3 = ke (p − v) (p − v + g) (cs − w),
according to the assumption p > v, we know p − v > 0, p − v + g > 0. So we
mainly discuss whether cs − w is positive or negative.
As ϕ2 ∈ [0, 1], that is 0 ≤ (p−v)(c m +w)+(w−cs )g+vg
(p−v)(cm +cs )+vg ≤ 1, then
(p − v) (cm + w) + (w − cs ) g ≤ (p − v) (cm + cs ), we can prove w ≤ cs . As
ϕ3 ∈ [0, 1], the same theory proves that w ≤ cs .
In conclusion, as cs − w > 0, that is ϕ2 − ϕ3 > 0, then ϕ2 > ϕ3 . This
completes the proof.

This proposition shows that, under cap policy, when the carbon cap doesn’t
play a role (that is no carbon emissions constraint), the manufacturer’s revenue
share proportion is greater than it under Cap-and-trade policy. Obviously, in the
situation of no carbon emission constraint, manufacturer’s profit is bigger, so he
can share a higher proportion of revenue.
5.2 Effect of Carbon Policy on Manufacturer’s Expected Profit
When the initial carbon quotas assigned by government are enough under cap
policy, it seems to be a problem without carbon emission constraint. In this
paper, taking the manufacturer’s expected profit without carbon emission con-
straint as a benchmark, we want to study the effect of carbon policy on man-
ufacturer’s expected profit. In addition, as the supply chain is coordinated by
revenue sharing contract, we adopt the whole supply chain profit to represent
the manufacturer’s expected profit.
Proposition 6. In the situation Q∗2 ≤ Ee , when E > E ∗ , E [πT 3 (Q∗3 )] >

E [πT 2 (Q∗2 )]; when E = E ∗ , E [πT 3 (Q∗3 )] = E [πT 2 (Q∗2 )]; when E < E ∗ ,
E [πT 3 (Q∗3 )] < E [πT 2 (Q∗2 )]; where E ∗ = EQ∗3 + k1 [E [πT 2 (Q∗2 )] − E [πT 2 (Q∗3 )]].
Proof. Compare E [πT (Q2 )] and E [πT (Q3 )].

If E ≤ EQ∗3 , E [πT 3 (Q∗3 )] = E [πT 2 (Q∗3 )] − k (eQ∗3 − E) ≤ E [πT 2 (Q∗3 )] <
E [πT 2 (Q∗2 )], we can get E [πT 3 (Q∗3 )] < E [πT 2 (Q∗2 )].
If E ≥ EQ∗3 , E [πT 3 (Q∗3 )] = E [πT 2 (Q∗2 )] + k (eQ∗2 − E) > E [πT 2 (Q∗2 )], we
can get E [πT 3 (Q∗3 )] > E [πT 2 (Q∗2 )].
As a result, there always exists E ∗ ∈ (eQ∗3 , eQ∗2 ), satisfy E [πT 3 (Q∗3 )] =
E [πT 2 (Q∗2 )], that is E [πT 2 (Q∗3 )] − k (eQ∗3 − E) = E [πT 2 (Q∗2 )].
So we get E ∗ = EQ∗3 + k1 [E [πT 2 (Q∗2 )] − E [πT 2 (Q∗3 )]]. This completes the
proof.

This proposition indicates that the manufacture’s expected profit under Cap-
and-trade policy is greater than it in the situation without carbon emissions
constraint under certain conditions. This is because the Cap-and-trade policy is
a kind of mixed strategy combining market operation and government regulation.
It not only control the total emission amount, but also offer an opportunity for
the enterprises to buy or sell the carbon emission quota, and make them to
guarantee production or gain more profits.
6 Conclusion and Future Research
In this paper, taking the two echelon supply chain (one supplier and one man-
ufacturer) which is dominant by supplier as the research object, we analyze the
supply chain coordination problem based on revenue sharing contract under Cap
and Cap-and-trade policy. It provides several interesting observations.
(1) Revenue sharing contract can coordinate the supply chain under both Cap
and Cap-and-trade policy, and can realize the Pareto improvement for both
sides under certain conditions.
(2) When the carbon policy plays a role, the manufacture’s revenue share pro-
portion under Cap-and-trade policy is always less than it without carbon
emission constraint. The manufacturer would buy carbon emission quota for
the limitation of it. It produced a certain cost and reduced profits, so he
would distribute less profit to the supplier.
(3) Under certain conditions, the manufacture’s expected profit under Cap-and-
trade policy is greater than it in the situation without carbon emissions
constraint. Because Cap-and-trade policy not only control the total emission
amount, but also offer an opportunity for the enterprises to buy or sell the
carbon emission quota, and make them to gain more profits. However, the
Cap policy do not has such function. As a result, theoretically speaking,
Cap-and-trade policy is better than Cap policy.
The conclusions of this paper is beneficial to the upstream and downstream

enterprises in such supply chain, at the same time they can also provide theoret-
ical guidance for carbon emissions policy for the government. In future study, we
can extend our research from one-to-one two-stage supply chain to one-to-many
or three-stage supply chain.
Acknowledgements. This research is supported by Fundamental Research Funds

for the Central Universities (funded by Sichuan University, No. 2015SCU11072 and
No. skzx2015-gb26), and Soft Science Project (No.2015-RK00-00244-ZF) funded by
Chengdu Science and Technology Bureau.
1088 L. Lu
References
1. Arani HV, Rabbani M, Rafiei H (2016) A revenue-sharing option contract toward
coordination of supply chains. Int J Prod Econ 178:42–56
2. Böhringer C (2014) Two decades of european climate policy: a critical appraisal.
Rev Environ Econ Policy 8(1):1–17
3. Cachon GP (2003) Supply chain coordination with contracts. Handbooks Oper Res
Manag Sci 11:227–339 Supply Chain Management
4. Cachon GP, Lariviere MA (2005) Supply chain coordination with revenue-sharing
contracts: strengths and limitations. Manag Sci 51(1):30–44
5. Cao J, Zhang X, Zhou G (2016) Supply chain coordination with revenue-sharing
contracts considering carbon emissions and governmental policy making. Environ
Prog Sustain Energy 2(35):479–488
6. Choi TM (2013) The international journal of advanced manufacturing technology.
Int J Prod Res 1(6851):835–847
7. Du SF, Zhu LL, Liang L (2013) Emission-dependent supply chain and environment-
policy-making in the cap-and-trade system. Energy Policy 57:61–67
8. Ghosh D, Shah J (2012) A comparative analysis of greening policies across supply
chain structures. Int J Prod Econ 2(135):568–583
9. Ghosh D, Shah J (2014) Supply chain analysis under green sensitive consumer
demand and cost sharing contract. Int J Prod Econ 164:319–329
10. Giannoccaro I, Pontrandolfo P (2004) Supply chain coordination by revenue shar-
ing contracts. Int J Prod Econ 89(2):131–139
11. Govindan K, Popiuc MN (2014) Reverse supply chain coordination by revenue
sharing contract: a case for the personal computers industry. Eur J Oper Res
2(223):326–336
12. Jaber MY, Glock CH, Saadany AME (2013) Supply chain coordination with emis-
sions reduction incentives. Int J Prod Res 1(51):956–968
13. Keohane NO (2009) Cap and trade, rehabilitated: using tradable permits to control
us greenhouse gases. Rev Environ Econ Policy 3(1):42–62
14. Kong G, Rajagopalan S, Zhang H (2013) Revenue sharing and information leakage
in a supply chain. Manag Sci 59(3):556–572
15. Matthews HD, Gillett N et al (2009) The proportionality of global warming to
cumulative carbon emissions. Nature 459:829–832
16. Mortimer JH (2002) The effects of revenue-sharing contracts on welfare in vertically
separated markets: evidence from the video rent industry. University of Californian
Los Angeles
17. Swami S, Shah J (2013) Channel coordination in green supply chain management.
J Oper Res Soc 3(64):336–351
18. Xu J, Chen Y, Bai Q (2016) A two-echelon sustainable supply chain coordination
under cap-and-trade regulation. J Cleaner Prod 135:42–56
Research on Integration of Livestock Products
Supply Chain Based on the Optimal Match
Between Supply and Demand
Liang Zhao, Yong Huang(B) , Zhusheng Liu, Mingcong Wu, and Lili Jiang

huangyong scu@aliyun.com
Abstract. China is a big country on livestock production and Chinese

livestock products output value is increased yearly. However, Chinese
livestock products are still mainly dominated by self-employed farmer’s
production mode which leads to imbalance between supply and demand,
periodic fluctuations in prices. In this paper an integration of livestock
products supply chain is proposed to solve the imbalance problem. A
sample of a duck farm is chosen to show the process of the integration.
First a discrete-event simulation model is built by the supply chain which
is composed of hatching, breeding, and packaging and transporting. Then
production plan, transport routes and the number of fixed assets are
adjusted to achieve the goal that the mean of supply-demand ratio is
close to 1 and the variance is close to 0.
Keywords: Livestock products · Supply chain · Duck farm · Integra-

tion · Discrete-event simulation
1 Introduction
China is a big country on livestock production and Chinese livestock products
output value is increased yearly. However, the construction of livestock products
supply chain (LPSC) is inefficient. At present, Chinese big enterprises in the
livestock products market have just started the integration of supply chain, and
the livestock products market is still mainly dominated by self-employed farmer’s
production mode. Farmers neither share their information nor have any plans for
production, so it leads to the imbalance between production and market demand.
Once a product is sold well, farmers flock to produce it blindly. Then the price
will fall, and farmers will switch to produce other products sold well, finally
becoming a vicious circle. In addition, the production of livestock products has
many shortcomings, such as low levels of automation, decentralized farmers. If we
want to integrate the supply chain, the issue that matches supply with demand
must be met. When the farmers make production plans, they’re faced with the
situation that demand is stochastic and production plans must be decided before
they know the real demand [1]. From the experience of developed countries
1090 L. Zhao et al.
for agricultural development, scale management on livestock products has an

incomparable cost advantage. Therefore, this paper considers building integrated
livestock products supply chain (ILPSC) to solve the mismatch between supply
and demand.
The process of production, cultivation, processing, transportation and sale
in LPSC will produce added value. Therefore, the LPSC is a value-added supply
chain from upstream to downstream. However, if a peasant-oriented platform is
just simply built, the information provided for each other may not be entirely
true due to the opportunism between enterprises in supply chain. It is difficult
to realize the information sharing in practice and this will cause supply and
demand imbalance, eventually to make market price fluctuations. Building the
ILPSC and formation of close cooperation are the emphases, which can integrate
the information flow in the supply chain. But in the implementation process,
there are still many issues to be resolved, such as the determination of the assets
number, production planning and transportation routes. This paper builds an
ILPSC model with simulation software, aimed at matching supply with demand,
and eventually using a sample to prove the effectiveness and feasibility of the
ILPSC model.
This paper is organized as follows. In Sect. 2, relevant literatures are reviewed.
In Sect. 3 the ILPSC model is shown and analyzed. In Sect. 4, an empirical model
of duck supply chain is presented and the supply-demand radio of the model is
optimized. Section 5 gives some conclusions.
2 Literature Review
LPSC is a kind of agricultural products supply chain. Chinese research on agri-

cultural supply chain started in 1999, and LPSC was later introduced. Chen and
Luo [2] put forward the idea of building a new Chinese LPSC by adding the
information agent to provide a scientific decision-making for strengthening the
supply chain management of Chinese LPSC. Zhou [15] proposed some conclu-
sions to promote the healthy and stable development of the pig industry. Yu and
Li [14] proposed the market operation mode and structure of livestock products
were no longer a single entity competition of production, processing and sales.
The survival and development of livestock products were more and more depen-
dent on the overall quality and comprehensive strength of the whole industry
chain. In the basis of this, the supply chain model of the combination of third-
party logistics and livestock products was used to improve the whole efficiency
of LPSC. Wang and Huang [12] put forward the SCOR model of agricultural
products based on information network, and constructed the agricultural supply
chain logistics management mode under the information network environment.
Then through the empirical analysis of Shanghai freshwater supply chain logistics
management mode, they proved the effectiveness and practicability of agricul-
tural logistics supply chain management mode. Liu, Sun and Wang [4] found
that the researches of Chinese agricultural supply chain were mainly concen-
trated on the traditional mode of operation of agricultural supply chain, fresh
Research on Integration of Livestock Products Supply Chain 1091
agricultural products at the end of the supply chain supermarket and informa-
tion technology under the traceability system and e-commerce environment. And
he believed that the establishment of a new integrated production and market-
ing supply chain and circulation system to achieve the circulation of agricultural
products and information network management, could make Chinese agricultural
products enter further into the global supply chain.
Compared with China, the United States, the Netherlands, the European
Union, Japan and other developed countries have been perfect on the develop-
ment of agricultural supply chain, so there are many researches about LPSC,
especially about beef and pork supply chain. Birthal and Pratap [10] proposed
that the revolutionary progress in livestock production was demand-driven. The
income elasticity of demand was higher compared to most other food commodi-
ties. And livestock would have a larger effect on poverty reduction compared
to crop sector. Facchioli [3] used a computational simulation system tool to
coordinate its flow of matter and information. Piewthongngam [8] developed a
system dynamics model as a tool for managers to visualise the movement of the
entire production chain. William, Norina and Cassavant [13] thought SCRA was
a “production adjustment & consumer-driven” systems. Van Roekel, Kopicki
and Broekmans [11] gave the basic four steps in the development of agricultural
supply chains. Simchi and Kaminsky [9] pointed out that information sharing
and task planning was the key to integrating supply chain. Min and Zhou [5]
summarized the methods and applications of integrated supply chain modeling.
It provided a reference to establishment of the ILPSC. Pan and Jean [7] analyzed
the link between pork supply chain in China and the United States, and put for-
ward that a mechanism of enterprises and peasants, a sound logistics operation
system and an information network platform should be established so that the
efficiency of Chinese pork processing supply chain could be improved. Robert and
Jon [6] used a visual modelling environment to overcome the problems involved
in implementing simulation models.
3 Model Description
3.1 Integrated Livestock Products Supply Chain Model
The ILPSC model refers to the integration of all members in the LPSC, based
on the common goals, to achieve true sharing of information and integration of
supply, production and sale. Figure 1 shows the whole ILPSC model, and there
are mainly four parts. The first part is the process of producers which con-
tains capital goods suppliers and flows. Capital goods suppliers weekly produce
pups as per production plan which is designed by demand. Then, a collection
of several suppliers makes up a flow. The combined production of a flow is kept
together when it moves through the model. The second part is the process of
manufacturers for a period of breeding activities. The third part is the process
of processors where livestock products are processed and packaged. The fourth
part is the process of sending products to market.
1092 L. Zhao et al.
Fig. 1. The integrated livestock products supply chain model
In the whole model, the information flow center makes information flow in
the whole supply chain smoothly and shares the information with every part
in the model. At the same time, reasonable decisions are made on all links of
the supply chain through the integrated information, such as production plan
and transportation routes. And early in the supply chain construction, it can
determine whether the number of assets to meet the market demand number by
the weekly market demand plan. Monitoring center is mainly for flows, manu-
facturers and processors. Among them, the flows and manufacturers are mainly
concerned about the health condition of livestock. What we want to monitor is
whether the livestock are sick, whether the disease is contagious and whether
the sick livestock are dead. In addition, the weekly sales of processors should be
monitored, because the processor needs to sell products to the market weekly.
And a batch of products is sent to market several weeks.
The ILPSC model is a two-way driving model. That is, on the one hand, the
model needs the production plan of the most upstream companies to push the
entire supply chain, on the other hand, it needs the final market demand to pull.
Therefore, it’s necessary to forecast market demand to pull the production to
guarantee the matching of supply and demand information, and then push the
production through the production plan. Events in the model take 1 week.
The goal of the model is to strike a balance between the supply of livestock
products to market and the demand of the market. The supply-demand ratio can
be used to response the situation about balance of supply and demand. There-
fore, whether the target is achieved or not can be determined by analyzing the
mean and variance of all processors’ corresponded supply-demand ratio. If the
mean of it is close to 1 and the variance is close to 0, it can be proved that the
target is achieved. And a sub-goal may exist, such as the cost. The rule to trans-
port products can be considered, when the transporter works. Distance between
different parts in the model is affected by the rule, and the cost is affected by
the distance. There are some uncertain factors in the model which is concerned
with the goal of the model. The first one is the uncertainties associated with
the health condition of livestock. It includes whether the livestock are diseased,
which type of health episodes is caught, whether it has died after the illness and
which week after the illness is the time livestock dead. Secondly, the week when
the livestock products are sent to the market is also an uncertainty factor. All
of them will affect the number transported to the market, through the uncertain
production losses for the entire system caused by themselves. In this paper, these
uncertainties are solved by simulation model.
The ILPSC model is suitable for livestock products, which mainly refer to
livestock and poultry meat, and livestock including pigs, cattle, sheep and so on,
poultry including chickens, ducks, geese and other poultry meat and wild birds.
The characteristics of livestock products include that the number of production
can be counted, production cycle is long, the sickness can result in uncertain pro-
duction losses, it’s easily restricted by the size of the farm, it will be transported
between different farms and processing and packaging are needed.
3.2 Model Entities, Variables
To build a discrete-event simulation model, entities and variables must be deter-

mined.
Model entities. Livestock, transporters, capital goods suppliers, flows, man-
ufacturers and processors are used in the ILPSC model as model entities.
Model variables. There are 12 variables in the ILPSC model, and the details
3.3 Analysis of Model Process
There are many process in the model, and the processes of health condition and
selection of flows to manufacturers are the most important and analyzed.
(1) Health Condition Process
Livestock may get sick during the breeding process, and the type of diseases
may be different. The infectivity of different types of diseases is also different.
If the disease is infectious, the diseased livestock need to be isolated from other
healthy livestock. The flow chart of the process is shown in Fig. 2.
(2) Selection of Flows to Manufacturers Process (based on the shortest distance
rule)
When livestock are transported from flows to manufacturers, a transportation
rule can be considered such as the shortest distance rule. It is preferable to
transport the products to the available manufacturer who is the nearest to the
flow. At the same time, the number of livestock transported will be constrained
by the capacity of transporters and manufacturer’s assets. The flow chart of the
process is shown in Fig. 3.
1094 L. Zhao et al.
Table 1. Variables of ILPSC model
Variable name Variable description Notes

Transfer week The number of weeks in the model Updated weekly when the
before to be sent to the market value is 0, it will be sent
to the market in the week
Health condition 0 is showed healthy, 1 is showed sick
Health episode The type of illnesses, using 1 n as
disease 1 disease n
Sick week The number of weeks before a sick Updated weekly when the
livestock dead value is 0, it will be dead
due to illness
Week The number of weeks the model runs
Quarter The number of quarters the model
runs
Min mile# The nearest distance between the flow # is the number of flows,
and the nearest available 1, 2, · · · , m
manufacturer
Supply# The number of livestock products # is the number of
supplied to markets weekly processors, 1, 2, · · · , t
Transporter# The number of livestock delivered by # is the number of flows,
a transporter 1, 2, · · · , m
Ill livestock# The number of sick livestock per week # is the number of flows,
for each flow 1, 2, · · · , m
Rate# Weekly supply-demand ratio # is the number of
processors, 1, 2, · · · , t
SD Rate# The difference between the mean of # is the number of
the Rate# in the simulation period processors, 1, 2, · · · , t
and 1 the value is closer to 0, the
matching rate is higher
Fig. 2. The flow chart of the health condition process

Fig. 3. The flow chart of the selection of flows to manufacturers process
3.4 Model Optimization Rules
To achieve the goal, the model needs to be optimized. There are many parts in
the supply chain will have an impact on the goal. In this paper, three factors
are mainly considered including the asset numbers of manufacturers, production
plans and transportation routes.
(1) The asset numbers of manufacturers
It refers to the number of assets in the supply chain where manufacturers breed
livestock. Manufacturers are used to breed pups provided by capital goods sup-
pliers, and it’s the key to connecting upstream and downstream of the supply
chain. Therefore, when the production plans of capital goods suppliers have been
decided with the change of the market demand, the asset number of manufactur-
ers should change accordingly. So, whether the asset number of manufacturers is
enough must be first determined in the establishment of the model. Only when
the number of manufacturers’ assets is sufficient can the model continue to be
optimized.
(2) The production plans
A production plan is a schedule of the number of pups produced by capital goods
suppliers. It directly determines the approximate size of the livestock products
that are ultimately sent to the market. Therefore, the sum of livestock products
transported to market will be closed to the sum of market demand by optimizing
the production plan.
1096 L. Zhao et al.
(3) The transportation routes

Transportation routes refer to the routes that transporters deliver livestock
between the links of the supply chain. This paper mainly considers the route
between manufacturers and processors. Since processors have their correspond-
ing markets, processed livestock products can be distributed evenly by adjusting
the routes, when the sum of supply is close to the demand.
4.1 Model Overview
The empirical data in this article refers to the competition of the 2016 Arena
Simulation Student Competition: Rockwell Duck Farm Supply Chain Optimiza-
tion. Two schedules will drive the simulation model of the farm system: a market
schedule, and a hatching farms schedule.
There are currently about 75 hatching farms in the system which produce a
total of approximately 100,000 ducks each week. Hatching farms breed ducks per
the hatching farms schedule, and a collection of several hatching farms makes
up a flow to move together. The order for transport is carried out in descending
order of flows and finds the nearest available growth/finish farms to receive the
ducks with the shortest distance rules. Production will be affected by the health
condition of ducks. There are four diseases to ducks showing in Table 2. The
probability of getting sick in four quarters of every hatching farm and prevalence
of the four diseases are based on the data provided in the competition.
Table 2. Four diseases to ducks
Disease types Requires Sick weeks Output loss (%)

separation? (week)
DED Exposed (Disease1 ) No 4 0.25
DED Naive (Disease2 ) No 4 1
DRRS Non-Vaccinated (Disease3 ) Yes 8 0.4
DRRS Vaccinated (Disease4 ) Yes 8 0.1
There are currently about 230 growth/finish farms in the system. The farms
receive the ducks from the hatching farms until it is full. Ducks are breeding
22 weeks in the growth/finish farms. Because disease 3 and disease 4 need to be
separated from other ducks, the growth/finish farm 27, growth/finish farm 199,
growth/finish farm 1564, growth/finish farm 65, and growth/finish farm 179 are
selected to receive the ducks suffering from the two diseases, which are relatively
closer to each station.
There are 11 packers/plants in the system. After packaged and processed,
ducks are sold for 6 weeks, the proportion shown in Table 3.
Table 3. Proportion of ducks sent to market weekly
Sales week Sales ratio

Week 1 0.04
Week 2 0.08
Week 3 0.2
Week 4 0.25
Week 5 0.25
Week 6 0.18
Sum 1
Ducks are transported by truck in the system and each truck is loaded with
2,600 ducks. A distance matrix defines the miles between a central point of a
flow’s location to a particular growth/finish farm. The simulation time is two
years or 104 weeks. Figure 4 shows the supply chain model of the duck farms
established with visual simulation software. The lower left part of the facility
interface is 75 hatching farms, the top part is 230 growth/finish farms and the
lower right part is 11 packers/plants.
Fig. 4. The ILPSC model of duck farms established with visual simulation software
4.2 Model Optimization

When the model is run, it’s found to be incorrect. The number of farms is insuf-
ficient to accommodate the duck. Therefore, there need more farms to meet the
extra expectations. 16 growth/finish farms near the 16 flows are selected to form
a group, three groups added. After running the model again, it’s found that the
supply of packer/plant5 is zero. Therefore, transportation routes are adjusted per
1098 L. Zhao et al.
the results of operations. It’s mainly solved through adjusting the corresponding
packers/plants of the new three groups to the packers/plants whose supply is
obviously insufficient. To find the optimal production plan, an experiment needs
setting up. The controlled variable for the experiment is the number of hatchings
per week for 75 hatching farms. The target is SD Rate# of 11 packers/plants.
SD Rate# in 40 scenarios is shown in Fig. 5.
Fig. 5. SD Rate# in 40 scenarios after running experiment
Since the ultimate goal is to achieve the match of supply and demand, it
can be achieved by finding a scenario where the supply-demand ratio is closer
to 1 and the SD Rate# is closer to 0. Therefore, by comparing the mean and
coefficient of variation (CV) of the absolute value of the SD Rate#, the scenario
can be found where the mean and CV are closer to 0. It proves that the number
of supply and demand is closer. An mean-CV line chart is shown in Fig. 6.
Fig. 6. Mean-CV line chart for 40 scenarios

Table 4. The adjustment scheme based on the result of experiment
The mean and CV of scenario 2,4,16,36 are all suitable. Finally, the pro-
duction plan of scenario 4 is selected. Then the model is ran again and the
transportation routes from the growth/finish farms to the packers/plants are
adjusted based on the results. The adjustment scheme is shown in table 4(G/F
farm represents Growth/Finish farm). Since the supplies of packer/plant5 and
packer/plant9 are sufficient, they are not shown in Table 4.
The line graphs between the supply and demand of the 11 packaging plants
in the original and the optimized model are shown in Fig. 7.
From Fig. 7, it can be seen the fluctuation range of the optimized model’s
supply is decreased compared with the original model, and supplies of every
packer/plant are basically fluctuating up and down in demand. The supply-
demand ratio of the original model and the optimized model is compared, as
shown in Table 5. And the line chart is shown in Fig. 8.
According to Fig. 8, it can be seen that the overall supply-demand ratio of
the 11 packers/plants of the optimized model is close to 1, and the fluctuation
is obviously reduced. There are three changes seen from Table 5 the mean of
supply-demand ratio of 11 packers/plants is reduced by 0.0836, the standard
deviation is reduced by 0.6979 and the variation coefficient is reduced by 0.6478.
Although the mean of the original model is also close to 1, the standard deviation
is 0.8142 which means it is very volatile. And the standard deviation and the
coefficient of variation after optimization is closer to 0, it can be seen that the
optimized model can make the match of supply and demand better.
1100 L. Zhao et al.
Fig. 7. The line graphs between the supply and demand of the 11 packaging plants in
the original and the optimized model during 28-104th weeks
Table 5. The supply-demand ratio of the original model and the optimized model
Packer number 1 2 3 4 5 6 7
Supply-demand ratio 2.6161 1.0385 0.6777 1.1512 0 2.1518 0.6294
Optimized supply-demand ratio 0.994 1.0853 0.832 0.9447 1.0613 0.8967 0.7984
Packer number 8 9 10 11 Mean SD CV
Supply-demand ratio 0.4204 1.8445 0.7463 0.4056 1.062 0.8142 0.7667
Optimized supply-demand ratio 1.0332 1.1317 1.1111 0.8735 0.9784 0.1163 0.1189
Fig. 8. The line chart for the supply-demand ratio of the original model and the opti-
mized model
5 Conclusion
This paper establishes an ILPSC to integrate the entire supply chain, which ulti-
mately matches the supply and demand of livestock products. The whole process
of the ILPSC from the upstream capital goods suppliers to the downstream mar-
ket is introduced, including the information flow center and the control center to
control the information flows and uncertain factors. Then, an integrated supply
chain model for a duck farm is constructed with simulation software. In the exam-
ple, the number of assets in the growth farms is added, and the production plan
of the hatching farms and the transportation routes between the growth farms
and the packing plants are optimized. Finally, the mean of supply-demand ratio
of the 11 packing plants is reduced by 0.0836, the standard deviation is reduced
by 0.6979 and the variation coefficient is reduced by 0.6478. The construction of
the ILPSC is a quite complex problem. Though the ILPSC can represent most
of the integrated process of livestock products, it also can be extended to other
aspects, for example, the question on reduction of transportation costs, and the
question on increasing utilization of growth farms.
References
1. Cachon G, Terwiesch C (2009) Matching supply with demand, 2nd edn. McGraw-
Hill, Singapore (in Chinese)
2. Chen C, Luo Y (2003) Construction on supply chain model of livestock products
in China. J Nanjing Agric Univ 26:89–92 (in Chinese)
3. Facchioli P, Severino G et al (2015) Use of a simulation system tool at the logistic
of a sugarcane company. Rev Metropolitana Sustentabilidade 5:112–127
4. Liu Z, Sun S, Wang J (2009) The trend of agricultural supply chain management
in China. Commercial Res 3:161–164 (in Chinese)
5. Min H, Zhou G (2002) Supply chain modeling: past, present and future. Comput
Ind Eng 43:231–249
1102 L. Zhao et al.
6. Muetzelfeldt R, Massheder J (2003) The simile visual modelling environment. Eur

J Agron Off J Eur Soc Agron 18:345–358
7. Pan C, Kinsey J (2002) The supply chain of pork: U.S. and China. Technical report,
University of Minnesota
8. Piewthongngam K, Vijitnopparat P et al (2014) System dynamics modelling of an
integrated pig production supply chain. Biosyst Eng 127:24–40
9. Simchi L, Kaminsky P, Simchi L (2000) Designing and managing the supply chain:
concepts, strategies and case studies
10. Staal S (2015) Livestock marketing and supply chain management of livestock
products. Indian J Agric Econ 70:89–92
11. Van R, Kopicki R et al (2002) Building agri supply chains: issues and guideline
[eb/ol]
12. Wang N, Huang L (2005) A study on agricultural product logistics supply
chain management mode based on information network. Res Agric Modernization
26:126–129 (in Chinese)
13. William C, Norina L, Cassavant K (2002) The use of supply chain management to
increase exports of agricultural products
14. Yu X, Li C (2012) Research on supply chain of livestock products mode based
on third-party logistics. In: Proceedings of 2011 International Conference on Engi-
neering and Information Management (ICEIM 2011)
15. Zhou C (2016) An improved quantitative model for evaluating pork price fluctua-
tion trend and the impaction in China: an empirical analysis between 2004–2014.
Int J Future Gener Commun Netw 9:199–210 (in Chinese)
A Novel Multi-Objective Programming Model
Based on Transportation Disruption in Supply
Chain with Insurance Contract
Kai Kang(B) , Jing Zhao, and Yanfang Ma
School of Economics and Management, Hebei University of Technology,

Tianjin, China
kkang2000@sina.com
Abstract. Supply chain risk management is becoming increasingly a

hot research issue. When the disruption occurs, insurance can be used
to reduce the risk. In this paper, we set a novel multi-objective program-
ming model based on transportation disruption in supply chain with
insurance contract in order to protect the whole supply chain members.
Then use a case to investigate the applicability of this model. From this
mathematical model, we obtain the optimal order quantity and the pure
premium, which can help the manufacture and the retailer make a better
plan. And the insurance can protect the manufacturer and the retailer’s
profits. In the future, the insurance will be used in supply chains widely.
Keywords: Supply chain risk management · Transportation disrup-

tion · The optimal order quantity
1 Introduction
Since the 21st century, with the global procurement, non-core business outsourc-
ing, and development of business models in supply chain management, such as
the lean management, the space distance on the supply chain becomes longer
and longer, while the time distance becomes shorter. The change of the tem-
poral and spatial variation of the supply chain improves the possibility of an
interrupt occurance. Every firm’s supply chain is susceptible to a diverse set of
risks, such as natural disasters, terrorism, war, financial crisis, supplier bank-
ruptcy and transportation delays. A lot of strategies used by firms to mitigate
disruption risks include emergence purchases, multi-sourcing, inventory reserve
and some reliability improvement of supply process [6]. The awareness regard-
ing the importance of supply chain risk management (SCRM) has grown in the
recent years. Supply chains are becoming increasingly competitive and complex
in order to effectively meet customer demands. Supply chain disruptions often
lead to declining sales, cost increases, and service failures for the company. There
are several kinds of mode of supply chain disruptions. Different scholars have dif-
ferent classifications for it. It can be clearly seen in Table 1.
1104 K. Kang et al.
Table 1. The classifications of supply chain disruption
Year Author Classification

2013 Hishamuddin [3] Supply disruption, transportation disruption
2015 Nooraie [7] Financial risk, transportation disruption
2015 Heckmann [2] Financial risk, demand disruption
2016 S. Tong [8] Node disruption, chain disruption
Transportation disruption, in particular, is slightly different from other forms

of SC disruptions, in that it only stops the flow of goods, whereas other disrup-
tions may stop the production of goods as well. The transportation disruption of
the supply chain is a non-negligible factor that it can reduce the manufacturer
and the retailer’s profits. For instance, in October 2000, two typhoons damaged
the oil gas pipeline belonging to Shanghai Petroleum Natural Gas Co., Ltd. in
China, leading to a 178-days service disruption of oil gas. Terrorist attacks on the
world trade center on September 11, 2001, causing a stagnation of goods which
came from the border of America. This led to a force to the Ford motor com-
pany to stop production activities temporarily. In March 2011, an earthquake
happened in Japan, which caused 22 automobile manufacturing companies’ shut-
down, such as Toyota, Honda, nissan. Affected by the earthquake the Japanese
exports of cars fell 12.5% in April compared with the year 2010, being the biggest
drop since the 18 months. As you can see from the above cases, transportation
disruption caused by emergency events not only leads manufacturers to stop
production, but also brings irreparable damages to the downstream enterprises
of the supply chain. Furthermore, a transportation disruption may affect the
condition of the valuable goods in transit.
Once transportation interrupts, it has a great influence for each member on
the supply chain, leading to transport delay, increasing the transportation time
and the transportation cost, even bringing serious damage to related companies.
Although realizing that there is a disruption risk in the supply chain, most of the
enterprises in the aspect of risk management of supply chain disruptions have
very little money and resources. To make sure for the firm’s profit, we tend to
make the insurance strategy according to the degree of risk. Purchasing an insur-
ance for the emergency is a kind of protection for enterprises. The model that
we have developed in this research addresses this vital aspect of transportation
disruption.
This paper proposes a newly multi-objective programming model for a supply
chain system subject to transportation disruption. In this model, we consider the
insurance contract to deal with the transportation disruption risk.
The contents of this paper are organized as follows. Section 2 presents the
related literature review. A description of the model and its formulation are given
in Sect. 3. Section 4 shows the analysis of numerical example. Section 5 provides
a conclusion of this paper and offers potential directions for future research.
A Novel Multi-Objective Programming Model Based on Transportation 1105
2 Literature Review
There is a great deal of literature on supply chain. In recent years, the research in
supply chain disruption is becoming a hot spot. Disruption has the characteristics
of low probability of occurrence and big influence on the part of the supply chain,
even the whole supply chain. Many scholars classified the types of supply chain
disruption, and provided several operational strategies for managing disruption.
But there are few papers which are mentioned the insurance to solve the risks.
Our study is related to studies focusing on the transportation disruption in
supply chain. Zhen et al. [9] investigated four strategies: basic strategy, BI insur-
ance strategy, backup transportation strategy, and mixed strategy, in order to
deal with distribution centers’ daily risk management. They used the mathematic
model to compare BI insurance strategy with backup transportation strategy,
and found that the choice of BI insurance strategy and the backup transportation
strategy depended on transportation market, insurance market and distribution
center’s operational environments. Hishamuddin et al. [3] built a recovery model
for a two-echelon serial supply chain when transportation disruption occurred.
This model determined the optimal ordering and production quantities with the
recovery window, and ensured the minimum total relevant costs. They developed
an efficient heuristic to solve the problem. In 2015, they developed a simulation
model of a three echelon supply chain system with multiple suppliers subject
to supply and transportation disruptions [4]. The objective of the paper is to
examine the effects of disruption on the system’s total recovery costs. Hernn
et al. [1] provided a novel simulation-based multi-objective model for supply
chains with transportation disruptions, aiming to minimize the stochastic trans-
portation time and the deterministic freight rate.
In this paper, we apply a novel multi-objective programming model for sup-
ply chain with transportation disruption. We reference about the method of
calculating profit of Lin’s paper [5] to study the problem of the manufacturer
and retailer’s profit when there is a transportation disruption. And we introduce
the insurance contract to the model.
3 System Description and Modeling
In the following subsections, we address the system’s description of the model.

Then present the mathematical representation of the model.
3.1 System Description
In this study, we consider a single supply chain model which has a manufacturer
and a retailer. We assume that the information between them is symmetrical.
The manufacturer has production, while the retailer has inventory. Before the
selling season, the manufacturer and the retailer agree on an insurance contract,
but the retailer needs to pay part fees to the manufacturer. And the premium
1106 K. Kang et al.
is decided by the insurance company. In our model, we assume that the trans-
portation disruption occurs in the delivery from the manufacturer to the retailer,
which interrupts the timely delivery of goods to the retailer. The transportation
disruption may be caused by an accident or a natural disaster, such as a earth-
quake, or flood. In addition, the goods in transit may or may not be damaged
during the disruption. Because of the transportation disruption, the insurance
contract comes into action.
In this paper, we use a multi-objective programming model to solve the
problem. The notations are explained as follows:
Denotes
m: represent the manufacturer;

r: represent the retailer;
I: represent the insurer;
D: the market demand;
πm : the expected profit of the manufacturer;
πr : the expected profit of the retailer.
Parameters
x: a random variable presenting the market demand D;

c: a unit cost of the product provided by the manufacturer;
w: a wholesale price that the manufacturer charges the retailer;
p: a retail price that the retailer serves the customers;
h: the shortage cost per unit;
ct : the transport cost per unit;
s: the salvage value of any unsold product per unit;
α: the retailer’s share of losses generated by the deviation of his order quantity
from the market demand, and α ∈ [0, 1], so the manufacturer’s share is 1 − α;
P : the pure premium that the insurer guarantees to pay the manufacturer with
the transportation interruption insurance, and it’s set by the insurer;
R: the rate of the pure premium to pay to the insurer;
β: the ratio that the retailer takes from the compensation, and β ∈ [0, 1], so the
manufacturer takes (1 − β)L;
f (x): probability density function with the market demand D;
F (x): cumulative density function with the market demand D.
Decision variable
q: an order made by the retailer and the manufacturer based on their forecast
of the market demand D.
When the transportation disruption occurs, the manufacturer and the retailer’s
expected profit will be decreased. But with the insurance contract, this situation
can be mitigated. So the objective of the insurance contract is to maximize their

profit. The objective functions are given as follows:
q
max πm =(w − c)q − qct − (1 − α) (p − s)(q − x)f (x)dx
0
∞ (1)
+ h(x − q)f (x)dx − P R + (1 − β)P.
q
Objective function (1) maximizes the manufacturer’s expected profit, where (w−
c)q is the earnings of the product for the manufacturer q with the order quantity
from the retailer, qct is the cost of transportation, 0 (p − s)(q − x)f (x)dx +
∞
q
h(x − q)f (x)dx is the expected losses generated by the deviation of the
retailer’s order quantity from the market demand, and PR is the cost of the
manufacturer paying to the insurer.
q ∞
max πr =(p − w)q − α (p − s)(q − x)f (x)dx + h(x − q)f (x)dx
0 q (2)
+ βP.
Objective function (2) maximizes the retailer’s expected profit, where (p−w)q
is the earnings of the product for the retailer while the assumption is that the
whole product can be sold out. When the transportation disruption happens,
the retailer can get a compensation βP. When the transportation disruption
happens, the insurance contract belongs to the third party liability insurance,
and its rate is usually a fixed value.
In order to guarantee the integrity of the model, there are constraints as
follows:
0 < s < c < w < p. (3)
If the product have not been sold off, the salvage value of the unsold product
s is less than its cost. This is equivalent to a kind of punishment for the exceed
quantity. And to make sure for the manufacturer and the retailer’s profit, the
relationship of unit cost c, the wholesale price w and the retail price p need to
be c < w < p.
0 < h < w. (4)
The shortage cost h is less than the wholesale price w.
P ≥ qc. (5)
The purpose to purchase the insurance contract is to minimise the manufac-
turer and the retailer’s profit, so the pure premium P is not less than the cost
of the product.
alpha ∈ [0, 1], (6)
beta ∈ [0, 1], (7)
where α is the proportion of the expected losses, and β is the proportion of the
pure premium.
1108 K. Kang et al.
3.3 Global Model

From the formulation above, a multi-objective model under transportation dis-
ruption for the supply chain has been deducted. The aims are to maximize the
profit of the manufacturer and the retailer. The global model is given:
q
max πm =(w − c)q − qct − (1 − α) (p − s)(q − x)f (x)dx
∞ 0
+ h(x − q)f (x)dx − P R + (1 − β)P
q
q
max πr =(p − w)q − α (p − s)(q − x)f (x)dx
0
∞ (8)
+ h(x − q)f (x)dx + βP
q
⎧
⎪
⎪ 0 <s<c<w<p
⎪
⎪
⎨ <h<w
0
s.t. P ≥ qc
⎪
⎪
⎪
⎪ α ∈ [0, 1]
⎩
β ∈ [0, 1].
Theorem 1. The optimal order quantity for the supply chain system will
achieve the maximum only if the following condition is satisfied:
p−w
α∗ = . (9)
p − c − ct
Proof. With the insurance contract, the system’s expected profit is
q
max π = max πm + max πr = (p − c − ct )q − (p − s)(q − x)f (x)dx
0
∞ (10)
+ h(x − q)f (x)dx − P R + P,
q
∂π
= p − c − ct − [(p − s + h)F (q) − h] = 0. (11)
∂q
Through the function of max πm and max πr , we can acquire the optimal
∗
order quantity qm and qr∗ solving the following equation:
∂πm
= w − c − ct − (1 − α)[(p − s + h)F (q) − h] = 0, (12)
∂q
∂πr
= p − w − α[(p − s + h)F (q) − h] = 0. (13)
∂q
Then the optimal order quantity is

p − c − ct + h
q ∗ = F −1 . (14)
p−s+h
The the optimal order quantity of the manufacturer and the retailer is
showed by:

∗ −1 (1 − α)h + w − c − ct
qm = F , (15)
(1 − α)(p − s + h)

αh + p − w
qr∗ = F −1 . (16)
α(p − s + h)
∗
Generate Eq. (9) into Eqs. (15) and (16), then we can get qm =qr∗ =q ∗ . So
∗ p−w
only when α = p−c−ct is established, the order quantity is to the most optimal
value.

4 Numerical Analysis
In this section, we examine the model by numerical analysis. The data was
obtained from the Lin’s paper [5], which assumed that the market demand fol-
lowed uniform distribution D ∼ [400, 500]. The other parameters were as follows:
p = 18, w = 15, c = 12, s = 8, h = 3, β = 0.6, R = 15%, and ct = 1.
Take the data into the formulations, and draw the trend chart Fig. 1, which
shows that the manufacturer’s optimal order quantity increases as α increases,
while the retailer’s optimal order quantity decreases as α increases.
∗ p−w
Let qm =qr∗ , then we get α∗ = p−c−c t
= 0.6. We can find α∗ = 0.6 from
Fig. 1 that proves the model. And from Fig. 1, we can see that the crossing is
the optimal order quantity q ∗ as Table 2, which is equal to 461. This is the best
quantity for them. So the pure premium is not less than 5532, and the purchase
to the insurer needs to be 830.
manufacturer
560
retailer
the optimal order quantity
520
480
440
0.0 0.2 0.4 0.6 0.8 1.0

a
Fig. 1. The effects of α on the optimal order quantity
Table 2. The quantity under different α
Decision 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
m 438.46 440.17 442.31 445.05 448.72 453.85 461.54 474.36 500 576.92 ∞
r ∞ 653.85 538.46 500 480.77 469.23 461.54 456.04 451.92 448.72 446.15
1110 K. Kang et al.
5 Conclusion
In this paper, we describe the transportation disruption in supply chains. We set
a mathematical model which we introduce the insurance contract to deal with
transportation disruption the way from the manufacturer to the retailer, and
use a case to investigate the applicability of this model. Although the insurance
contract is effective in coordinating the supply chain, it also has some limitations.
The most critical limitation is that the supplier incurs an administrative cost
in monitoring the retailer’s sales situation. The objective of the study is to
determine the optimal order quantity in the case of the uncertainty of the market
demand. This insurance contract transfers the risk from the manufacturer and
the retailer to the insurance company, which protects the manufacturer and the
retailer’s profits and improves the efficiency of the supply chain. In particular,
how much to purchase the insurance is discussed in this paper. The model is
useful for decision makers to determine the product quantity.
There are several directions for this study to continue. We can extend the
model to a complex supply chain with multiple manufacturers or retailers. In
addition, we can apply different strategies to deal with transportation disruption.
Acknowledgments. This research was supported by National Natural Science Foun-

dation of China (Grant No. 71640013, and Grant No. 71401020), and System Science
and Enterprise Development Research Center (Grant No. Xq16C10), National philos-
ophy and Social Science Foundation of Tianjin (TJGL15-007); Provincial Science and
Technology Foundation of Hebei Province (16457643D).
References
1. Chávez H, Castillo-Villar KK et al (2016) Simulation-based multi-objective model
for supply chains with disruptions in transportation. Robot Comput Integr Manu-
fact 43:39–49
2. Heckmann I, Comes T, Nickel S (2015) A critical review on supply chain risk—
definition, measure and modeling. Omega 52:119–132
3. Hishamuddin H, Sarker RA, Essam D (2013) A recovery model for a two-echelon
serial supply chain with consideration of transportation disruption. Comput Ind
Eng 64(2):552–561
4. Hishamuddin H, Sarker R, Essam D (2015) A simulation model of a three echelon
supply chain system with multiple suppliers subject to supply and transportation
disruptions. IFAC Papersonline 48(3):2036–2040
5. Lin Z, Cai C, Xu B (2010) Supply chain coordination with insurance contract. Eur
J Oper Res 205(2):339–345
6. Lynch GS (2012) Supply chain risk management 192:33–40. London: Springer
7. Nooraie SV, Parast MM (2015) Mitigating supply chain disruptions through the
assessment of trade-offs among risks, costs and investments in capabilities. Int J
Prod Econ 171:8–21
8. Shu T, Gao X et al (2016) Weighing efficiency-robustness in supply chain disruption
by multi-objective firefly algorithm. Sustainability 8(3):250
9. Zhen X, Li Y et al (2016) Transportation disruption risk management: business
interruption insurance and backup transportation. Transp Res Part E 90:51–68
An Interval-Parameter Based Two-Stage
Stochastic Programming for Regional Electric
Power Allocation
Jingqi Dai and Xiaoping Li(B)

lixiaoping@scu.edu.cn
Abstract. Electricity has become an indispensable part of human

being’s life. In China, regional electric power allocation, involving mul-
tiple groups who pursue the biggest social benefit, has been facing some
serious problems and raising people’s more and more concern. In this
paper, an interval-parameter based two-stage stochastic model is built to
solve the problem of planning of electric power allocation target in the
perspective of the State Grid Corporation. In the model, the interval-
parameter is used to deal with uncertainty of the problem and the two-
stage decision making helps take corrective actions after a random events
occurs. The application of the model was demonstrated efficiency and
practicality with the case study in real world. Combining models pro-
posed in this paper with some related policy implications, it is useful in
improving actual operation and helping better decision.
Keywords: Interval-parameter · Two-stage stochastic programming ·

Electric power allocation
1 Introduction
The second industrial revolution marked the people entered the age of electricity.
In recent years, the world electricity industry has experienced a great develop-
ment and it played a vital role in human beings’ life. According to Key World
Energy Statistics by International Energy Agency (IEA), world electricity gen-
eration increased form 6131 TWh in 1973 to 23816 TWh in 2014, where the
growth rate is as high as 288% [8]. However, electrical power systems are fac-
ing different problems such as economical and environmental issues. As for the
electric power distribution, it is also a complex giant system which has been
confusing electricity operators, and researchers have been committed to study it
for a long time.
In China, during the last decades, a lasting market-oriented reform has been
carried out in electric power industry. Improving the operation efficiency of elec-
tric power industry is the primal goal of the government of China, and 2016
1112 J. Dai and X. Li
marks the beginning of the second round of such reforms and the 13th Five-year
Plan of China [3,4]. Besides, “Further strengthening the institutional reform
of the electric power industry was promulgated by the State Council of China
whose No. 4 attachment file is about opening electricity plan orderly to promote
optimum allocation of electric power resources” [3]. Therefore, on the perspec-
tive of the national grid, how to guarantee the power supply and obtain more
benefit at the same time are important questions they should face.
In the past years, many scholars studied problem of electric power alloca-
tion plan from different perspectives. Many innovative optimization techniques
have been developed for allocating and managing electric power in more effi-
cient benign ways [5,7,9,10,13]. Among these methods, linear programming is
the basic and prior way to solve this problem. Recently, F. Chen et al [2] used
fuzzy chance-constrained programming model to study electric power generation
systems planning problem which can reflect uncertain interactions among ran-
dom variables directly. For power generation systems planning, this approach
can have a wider application scope than existing optimization models. However,
Huang and his group proposed a new method that is interval-parameter two-
stage stochastic program to solve resource management problem, such as water
resource management and solid waste management [11,12,14]. This inspired us
to apply this kind method to other resource management problem, for example,
electric power allocation.
Birge and Louveaux introduced various kinds of two-stage and multi-stage
stochastic programming problems [1]. Among them, two-stage stochastic pro-
gramming is effective for analysing medium-to-long-term planning problem
where the system data is characterized by uncertainty [12]. Decision makers
will make an initial decision based on the future events and this is called first
stage decision. Then, after the uncertainty things are resolved, the second stage
decision will be made to correct the results of the first stage decision. Besides,
introducing interval parameter into two-stage stochastic model can potentially
address uncertainties of the problem.
The goal of the paper is to apply interval-parameter two-stage stochastic
program to study problem of electric power allocation while simultaneously tak-
ing system uncertainties into account. It will be demonstrated that this kind
of method can help decision makers to design more efficient plans and obtain
more benefits. The paper is structured as follows: The next section is the state-
ment of key problem. Section 3 presents process of modelling through using the
method introduce in this paper. A simple case study was used to demonstrated
the effective of the method in Sect. 4. In the end, the paper summarizes the main
contents and contributions of this paper in Sect. 5.
2 Key Problem
This paper builds a interval-parameter two-stage stochastic programming model

to better solve the problem of electric power allocation. Therefore, it is necessary
to introduce some basic background of this problem.
TSSP Regional Electric Power Allocation 1113
In China, generally speaking, the State Grid Corporation and the govern-
ment have joint ownership of electric power allocation right. They will make
allocation decision overall considering the historical data, requirements of the
local government and production capacity of electric power. The authority has
an obligation to make a reasonable and scientific electric power allocation plan
to achieve the best economic benefits with better resources saving because it will
take a lot of resources in the process of electric power production.
Electric power allocation involves a region’s residents life, industrial produc-
tion and other unpredictable situations, so it is a complex system. All the groups
need to know how much electric power they can expect because if the guaran-
teed amount cannot be provided, they have to make a change of their initial
planning which may bring them a large loss. For example, the factories have to
halt production and the residents have to stop their work that needs use electric
power. At the same time, it is also hard for decision makers to make the alloca-
tion plan. In practice, there are many stochastic factors that will influence power
supply which directly generates to the complexity of the problem. In this situ-
ation, this paper uses two-stage stochastic method to solve the problem where
the authority will make two decisions. The first decision is to make allocation
plan to guarantee basic demands according to historical data and the second
decision is to make an compensate that is resulted by the first decision. At the
same time, when the State Grid Corporation makes decisions, there are some
uncertain factors such as weather, breakdown of machinery that would lead to
a gap between demand and supply. Thus, treating uncertain number as interval
parameter is a reasonable method to deal with problem’s uncertainty.
Therefore, from this discussion, the problem can be solved by building a
interval-parameter two-stage stochastic model to obtain more benefits. On one
hand, the first decision of insufficient power supply can be made up by the second
decision. On the other hand, by adjusting two-phase decision-making, the State
Grid Corporation can acquire maximum benefits, which ensures the harmonious
development of power industry to a certain extent. The mathematical form for
this problem is given in the next section.
3 Modelling
In this section, a interval-parameter two-stage stochastic model with uncer-

tain variables for electric power distribution is constructed. The mathematical
description of the problem and the transformation process is given as follows:
3.1 Assumptions
(1) There is no loss during the electric power transmission process;

(2) Interval parameter of the upper and lower values can be obtained through
the historical data.
3.2 Notations
To facilitate the problem description, the notations are introduced firstly.
Indices
i: index of provincial power user, where i = 1, 2, · · · , n.
Decision variables
Xi : the first-stage decision variable, which is allocation target for electric power
that is promised to user i;
YiS : the second-stage decision variable, which is shortage of electric power to
user i when the actual allocation amount is S.
Uncertain parameters
Bi : net benefit to user i per billion kilowatt hour;
Ci : punish coefficient, which is loss to user i per billion kilowatt hour;
Di min : minimal demand of user i to ensure the normal operation of the society;
Sih : the amount where the electric power allocation target is not meet for user
i when actual amount of electric power supply is Si with probabilities ph ,
S: random variable, actual total amount of electric power generation; letting S
take values Shi with probabilities Ph .
Certain parameters
Si max : maximum allowable allocation amount for user i;
ph : the probability of occurrence of shortage level h, where h = 1, 2, · · · , H,
H
h=1 ph = 1;
h = 1: actual power supply is very close to the demand and the shortage is
smallest;
h = 2: actual power supply is relatively close to the demand and the shortage is
medium;
h = H: actual power supply is far from to the demand and the shortage is
highest.
Symbols
+: represent the upper limit of parameters;
−: represent the upper limit of parameters.
3.3 The Modeling Process

(1) Objective function
In this paper, we consider regional power grid company has the authority to
allocating electric power to multiple users, with the objective of maximizing the
total net benefit by optimizing allocation scheme. As these users need to know
how much electric power they would be distributed so that they can make sound
plans for their production activities to obtain the best social benefit. At the same
time, a prescribed amount of electric power is promised to each user according

to related provisions because both government and grig company have right to
ensure basic function of the society. If the promised demand is met, there will
bring benefits to the local economy, otherwise, power shortages could lead to
many production activities cannot be carried out, resulting in economic loss.

n
For objective target Xi , the grid company can obtain benefit Bi Xi . However,
i=1
when uncertain of uncovered electric power appears, the second-stage decision
will be made to compensate for any adverse effect which was caused by the first

n
stage of the decision. Therefore, the loss can be expressed as E Ci YiS , where
i=1
E [·] represent expected value of a random variable. So, the total net benefits is
n
n
max f = Bi Xi − E[ Ci YiS ].
i=1 i=1
(2) Limitations
Owing to resource constraints and other reasons, there are 4 kinds of con-
straints and they are detailedly introduced as follows:
1 Minimum power demand constraints
Local government has relative policies to promise minimum power demand
of the area according to the historical data of electricity consumption so as to
ensure normal social life without causing unrest. Based on the above description,
we can get the constraint as follow:
Xi ≥ Di min .
2 Allowance electric power allocation constraints
For one region, the ability of electric power production is limited. Therefore,
allowance allocation amount is also limited so that grid company can’t distribute
more than actual electric power generation volume to all users. This kind of
constraints can be expressed as following:
YiS ≤ Xi ≤ Si max .
3 Available power constraints
In first-stage decision making process, Xi must be determined before actual
total electric power supply S are known, while the shortage Yi S are determined
during the second stage when electric power supplies are known but allocation
amounts have been fixed. This kind of constraints mean that first-stage decision
variable subtract second-stage decision variable should not be more than actual
total supply amount. We can have this relationship expressed as:
n

(Xi − YiS ) ≤ Shi .
i=1

4 Non-negative constraints
YiS ≥ 0.
(3) Overall model

Based on the previous discussion, a two-stage stochastic programming can be
built. In this problem, first stage decision must be made before unknown actual
total supply amount. When the uncertainty of supply is uncovered, a second-
stage decision can be made to compensate influence that is the result of the first-
stage decision. We can formulate overall model using two-stage programming:
n
n
M axf = Bi Xi − E Ci YiS
⎧ i=1 i=1
⎪
⎪ Xi ≥ Di min
⎪
⎪
⎪
⎪ YiS ≤ Xi ≤ Si max
⎨ n
s.t. (Xi − YiS ) ≤ Shi
⎪
⎪ i=1
⎪
⎪
⎪
⎪ YiS ≥ 0
⎩
∀i = 1, 2, · · · , n.
(4) Model transformation

There are many uncertain factors in the power system. In real world problem,
it is difficult to generate probability distributions for these parameters with small
sample size. In this complex giant system, decision makers cannot make accu-
rate prediction of available allocation of electric power in the system. Therefore,
the target allocation Xi , net benefit Bi and punish coefficient Ci are uncer-
tain. We introduce the concept of interval parameters to solve the problem of
uncertainty, where interval number is defined as a range known upper and lower
bounds [6]. Using “+” and “–” represents upper limit and lower limit of para-
meters respectively. Besides, to solve the problem through liner programming,
the allocation S must be approximated as discrete values. In this paper, we
Sni with probabilities ph (h = 1, 2, · · · , H). Therefore, we have
letS take value
n n n H
E Ci YiS = Ci · E [ i=1 YiS ] = i=1 Ci ( h=1 ph Sih ). Thus, the previ-
i=1 i=1
ous overall model can be converted to the following model:

n n H
M axf ± = Bi± X± i − ph Ci± Sih
±
⎧ ± i=1
±
i=1 h=1
⎪
⎪ X±i ≥ Di min
⎪
⎪ ±
⎪
⎪ Sih ≤ Xi ≤ Si max
⎨ n
s.t. (X± ±
i − Sih ) ≤ Sh
i±
⎪
⎪ i=1
⎪
⎪
⎪
⎪ S± ≥ 0
⎩ ih
∀i = 1, 2, · · · , n.
According to research of Wang and Huang [15], this model can be solved
through convert interval-parameter two-stage stochastic model into two sub-
models which correspond to the upper and lower bounds of the object-function
value. It is difficult for decision makers to determine whether the upper bound
X + or the lower bound X − of the uncertain variable X ± correspond to the upper
bound of the total net benefit. Thus, maximized total benefit can be obtained
by optimized target value. Letting Xi± = Xi− + ΔXi zi where ΔXi = Xi+ − Xi−
is a fixed value and zi is range from 0 to 1. Here, we introduce a new decision
variable zi to identify the optimized target value. Based on above statement, we
can get a new transformation model as follows:
n
n H
max f ± = Bi± (Xi− + ΔXi zi )− ph Ci± Sih
±
⎧ − i=1 i=1 h=1

⎪
⎪ Xi + ΔXi zi ≥ Di±min
⎪
⎪ ±
⎪
⎪ Sih ≤ Xi− + ΔXi zi ≤ Si max
⎪
⎪ n
⎨ (X− + ΔXi zi − S± i±
ih ) ≤ Sh
s.t. i=1 i
⎪
⎪ ±
⎪
⎪ Sih ≥0
⎪
⎪
⎪
⎪ 0 ≤ zi ≤ 1
⎩
∀i = 1, 2, · · · , n.
It can be solved by reformulating above model as two submodels, and the
lower bound of objective function value can be first introduced as follows:

n
n H
max f − = Bi− (Xi− + ΔXi zi )− ph Ci+ Sih
+
⎧ − i=1 i=1 h=1

⎪
⎪ Xi + ΔXi zi ≥ Di+min
⎪
⎪
⎪
⎪ Sih
+
≤ Xi− + ΔXi zi ≤ Si max
⎪
⎪ −
n
⎨ (X + ΔXi zi − S+ i−
ih ) ≤ Sh
s.t. i=1 i
⎪
⎪
⎪
⎪ Sih
+
≥0
⎪
⎪ 0 ≤ zi ≤ 1
⎪
⎪
⎩
∀i = 1, 2, · · · , n.
In addition, the upper bound of objective function value can be expressed as:
n n H
M axf + = Bi+ (Xi− + ΔXi zi )− ph Ci− Sih
−
⎧ − i=1 i=1 h=1

⎪
⎪ Xi + ΔXi zi ≥ Di−min
⎪ −
⎪
⎪
⎪ Sih ≤ Xi− + ΔXi zi ≤ Si max
⎪
⎪
⎨ n (X− + ΔX z − S− ) ≤ S i+
i i h
s.t. i=1 i ih
⎪ −
⎪
⎪
⎪ Sih ≥ 0
⎪
⎪
⎪
⎪ 0 ≤ zi ≤ 1
⎩
∀i = 1, 2, · · · , n.
±
In the above two models, we can get the optimal solution Sihopt and ziopt ,
±
respectively. Then optimal objective function value is fopt and it is easy to obtain
±
Xiopt = Xi− + ΔXi ziopt . At the same time, optimal electric power allocation is
± ± ±
Wiopt = Xiopt − Sihopt .
4 Case Study
A real case study is given in this section to demonstrate the efficiency of this
method.
Fig. 1. The location of the case
(1) Description of case

In China, the State Grid Corporation has right to allocate and schedule elec-
tric power, which plays an important role in China’s power industry. Owning to
China covers a large territory, the State Grid Corporate is divided into six sec-
tors in order to better and easier to management, which are North China branch,
East China branch, Central China branch, Northeast China branch, Northwest
China branch and Southwest China branch, respectively. Each branch has ability
to finish their work to provide stable and reasonable power supply and in this
process they can obtain benefit. Among them, Southwest China branch is mainly
responsible for electric power business of Sichuan province, Tibet and Chongqing
city, to manage such work as regional power grid scheduling, operation control
and electric power trade.
Southwest China branch is a new sector which was founded in December,
2014. This paper chose Southwest China branch as case to study because this
sectors is key to promote economic and social development of southwest of China,
and to ensure the safety of national energy efficient supply in strategic height.
Besides, this branch only has jurisdiction over the three region which can reduce
the computation burden. And the location can be shown in Fig. 1. Therefore,
through the case, we want to prove the method used in this paper is feasible.
(2) Data collection and dispose
The data was obtained from home page of State Grid Corporation, govern-
ment report and market surveys. Table 1 shows the related electric power and
economic data which contains demand of different user, punish coefficient and
net benefit and so on. Besides, probabilities of degree of electric power shortage
and the maximum allowable allocation amount are given in Table 2.
(3) Result analysis
Table 3 shows the results that indicates the most optimal electric power allo-
cation under the different situation of power shortage. It can be seen that these
interval solutions stem from uncertainty in input parameters. If the available
electric power allocation can’t meet the promised amount, there would bring
huge loss to the users, resulting in competition among them for the limited
Table 1. Allowable electric power allocation and related economic data
Parameters Sichuan provinve Chongqing Tibet

Net benefit [0.3612,0.6015] [0.510,0.743] [0.3018,0.6631]
(Bi± /yuan·kwh−1 )
Punish coefficient [0.52,0.71] [0.55,0.76] [0.39,0.69]
(Ci± /yuan·kwh−1 )
Minimal demand of user [2514.79,2697.41] [606.4,645.87] [42.23,43.19]
(Di±min /billion kwh)
Maximum allowable allocation 3129.44 679.81 44.77
amount
(Si max /billion kwh)
Electric power allocation target [2679.64,2991.56] [631.85,667.93] [41.25,44.08]
(X ±
i /billion kwh)
Table 2. The relevant probabilities and actual electric power generation
Probability (ph ) Low (h = 1) Medium (h = 2) High (h = 3)

0.85 0.10 0.05
Actual total amount of Sh1± [2699.45,2734.21] [2517.57,2683.82] [2397.10,2503.78]
Electric power generation Sh2± [627.56,661.08] [608.49,625.24] [588.67,605.23]
(Shi± )/MKh Sh3± [42.13,43.42] [40.16,42.09] [38.56,39.97]
electric power. Therefore, the decision making about electric power allocation
target plays a vital role in the whole process of the allocation problem. The opti-
±
mized targets for three uses can be obtained by letting Xiopt = Xi− + ΔXziopt ,
and it is 2679.42 billion kwh, 945.92 billion kwh and 43.2 billion kwh for Sichuan
province, Chongqing and Tibet, respectively. At the same time, each shortage
and the total benefit can be easy to know form the Table reftab:daijingqi3. From
the results, the electric power would be first allocated to Sichuan province, next
for Chongqing and the last for Tibet when the scarcity occurs because Sichuan
province will bring State Grid Corporation the highest benefit.
In real world, suitable policy is important for decision makers to make plan
of an area’s sustainable electricity use. From the analysis above, solutions can
be obtained by letting the different allocation targets under various policies.
Different policy orientation will make the aspect of the consideration vary when
decision makers think about the problem. Some policy implications will be given
to help making better decision as follows. Firstly, to establish a relatively stable
long-term trading mechanism where suppliers and demanders can trade on their
own in the competition market. Secondly, improving the mechanism of inter-
district electric power deals across the province which permits power generation
companies, power users and electricity sell bodies can make occasional trade
when electricity is in short supply in a day. Thirdly, environmental protection
also is the topic that policy makers need to be concerned with. Thus, forming
Table 3. Electric power allocation solutions under different probabilities
User Probability i = 1 (Sichuan provence) i = 2 (Chongqing) i = 3 (Tibet)

±
Sihopt h = 1 (0.85) 0 [6.85,18.36] [0.66,1.07]
h = 2 (0.10) [179.83,307.74] [37.43,42.69] [1.99,3.04]
h = 3 (0.05) [300.30,487,78] [57.25,62.70] [4.11,4.64]
±
Xiopt 2697.42 945.92 43.2
ziopt [0.057,1] [0.39,1] [0.69,1]
±
Wiopt h = 1 (0.85) [2697.42,2734.82] [627.56,661.08] [42.13,43.432]
h = 2 (0.10) [2517.57,2683.82] [608.49,625.24] [40.16,42.09]
h = 3 (0.05) [2397.10,2503.78] [588.67,605.23] [38.56,39.97]
±
fiopt [974.31,1799.42] [317.56,494.69] [12.40,29.14]
market mechanism to promote renewable energy use to promise priority use of

clean energy.
5 Conclusions
This paper studied interval-parameter two-stage stochastic model applying to
the problem of electric power allocation. By using the proposed model, we con-
sidered not only how to make a more stable consumer plan, but also let decision
makers can benefit more form it. In the process of model building, we overcame
uncertainties of the problem through introducing interval-parameter into the
model, which at the same time got over complexity of solving process. The ini-
tial model was transferred into two submodels so as to obtain solutions, avoiding
system failure risk of interval solution. Of course, the approach was demonstrated
effective by using data form Southwest branch of the State Grid Corporation of
China as a case study. Through analysing result of the case study, some pol-
icy implications were proposed in the end which may offer help to regulators of
electric power allocation to make decision.
Compared with previous studies, this kind of method is innovative to be
applied in the electric power allocation which has good practicality value. How-
ever, this method was built based on the assumption that the upper and lower
boundary is known. Besides, the probabilities of degree of electric power short-
age are also directly given. The future research may focus on these two points
to make the study more rigorous, which can better deal with the uncertainty.
Besides, some suitable policy simulation also can be put forward to help better
decision making.
Acknowledgements. We thank the support of the Social Science Planning Annual

Project of Sichuan Province (SC16C011), Soft science research plan project of Sichuan
province (2017ZR0098), Soft science research project of Chengdu city (2015-RK00-
of Sichuan University (skqy201623).
References
1. Birge JR, Louveaux F (1997) Introduction to Stochastic Programming. Springer,
New York
2. Chen F, Huang GH, Fan YR, Chen JP (2017) A copula-based fuzzy chance-
constrained programming model and its application to electric power generation
systems planning. Appl Energy 187:291–309
3. CCCCP (Central Committee of the Chinese Communist Party) SSC (2016) Further
strengthening the institutional reform of the electric power industry. In Chinese
4. Development NN, Reform Commission NNEA (2016) Six subsidiary documents for
the new electric reforms were released. In Chinese
5. Fukuta N, Ito T (2012) A preliminary experimental analysis on combinato-
rial auction-based electric power allocation for manufacturing industries. In:
IEEE/WIC/ACM International Conferences on Web Intelligence and Intelligent
Agent Technology, pp 394–398
6. Huang GH (1998) A hybrid inexact-stochastic water management model. Eur J
Oper Res 107(1):137–158
7. Iba K, Suzuki H et al (1988) Practical reactive power allocation/operation planning
using successive linear programming. IEEE Trans Power Syst 3(2):558–566
8. Iea IEA (2016) Key world energy statistics 2016. Int Energy Agency 1:37
9. Jordehi AR (2016) Allocation of distributed generation units in electric power
systems: a review. Renew Sustain Energy Rev 56:893–905
10. Kayal P, Chanda CK (2016) Strategic approach for reinforcement of intermittent
renewable energy sources and capacitor bank for sustainable electric power distri-
bution system. Int J Electr Power Energy Syst 83:335–351
11. Li YP (2014) Municipal solid waste management under uncertainty: an interval-
fuzzy two-stage stochastic programming approach. J Environ Inform 12(2):96–104
12. Maqsood I, Huang GH, Yeomans JS (2005) An interval-parameter fuzzy two-stage
stochastic program for water resources management under uncertainty. Eur J Oper
Res 167(1):208–225
13. Shang JC, Zhang LQ (2007) Research and application of technologies in energy-
saving emission-reducing and optimal resource allocation of electric power system.
Power Syst Technol 31(22):58–63 In Chinese
14. Wang S, Huang GH (2013) An interval-parameter two-stage stochastic fuzzy pro-
gram with type-2 membership functions: an application to water resources man-
agement. Stochast Environ Res Risk Assess 27(6):1493–1506
15. Wang S, Huang GH (2015) A multi-level taguchi-factorial two-stage stochastic
programming approach for characterization of parameter uncertainties and their
interactions: An application to water resources management. Eur J Oper Res
240(2):572–581
Pricing Strategies of Closed Loop Supply Chain
with Uncertain Demand Based
on Ecological Cognition
Dongjing Yu(B) and Chunxiang Guo

1452702457@qq.com
Abstract. With the consumer’s ecological cognition becoming more and

more higher, remanufacturing has became a trend. It is imperative for
contemporary enterprises to find out some ways to balance the profit of
supply chain and the government. In this paper, we consider the closed
loop supply chain that consists of the manufacturer, retailer and con-
sumers. The retailer is responsible for the acquisition of used products
from consumers, and sold them together to the manufacturer. Then the
manufacturer producing new products and remanufactured products,
they were sold together in the market. Based on the ecological cogni-
tion and value judgment of consumers, it is easy to derive the demand
quantity of remanufactured products and new products. First, we for-
mulate the cooperative and competitive game model to investigate the
manufacturer’s and retailer’s pricing strategy. Next, in the case of differ-
ent ecological cognition, we studied the behavior of government subsidy.
A mathematical model is presented to formulate the problem to get the
optimal decision for the government. Finally, we compare the character-
istics of the two models, and draw the corresponding conclusions.
Keywords: Ecological cognition · Cooperative and competitive game

model · Closed loop supply chain · Government subsidy
1 Introduction
In the competitive business environment, it is imperative for businesses to con-

sider the resource shortage and environmental pollution. In our country, they
are seriously restricting the development of society. Without doubt that recy-
cling of solid waste is an effective way to save resources and reduce pollution.
In the recent years, the government has advocated a resource-saving and envi-
ronmentally friendly mode of production, and resulting in large-scale academic
researches on remanufacturing. Therefore, Wei et al. [25] analyzed the obstacles
and the incentive factors that should be given in the manufacturing process of
China.
Pricing Strategies of Closed Loop Supply Chain 1123
In the process of reverse logistics, the factors affecting the recycling of used
products are numerous. Pankaj et al. [6] have done a research on the possibility of
three way recycling options. Llgin et al. [13] analyzed four major categories about
remanufactured products. Fleischmann et al. [8] studied quantitative models for
reverse logistics, and [5,11] researched on sustainable development of supply
chain systems. In addition, Thierry et al. [24] researched strategic issues on recy-
cling and management product. The spare parts research is also very important,
so Ruud H et al. [23] considered the uncertain quality of the used products.
Chari et al. [2] addressed a problem about the repairable spare parts to find
the cost-optimal production strategy. In order to further study the acquisition
price and the quantity of remanufactured products, Shaligram [21] studied the
impact of the consolidation center and collection centers. Based on the study of
Shaligram et al., Gönsch Jochen [10] supposed that the quantity of used products
and replacement parts are unrestricted. Sung et al. [14] devoted to developing
optimal production planningnings for different production strategies.
The production demand of remanufactured products is extremely uncertain,
so Mitsutaka [17] and others predict the demand of the product by time series
analysis. Shi et al. [22] developed an option pricing model to evaluate the acqui-
sition price of used products under uncertain demand and return. For the sake
of considering the actual situation of the remanufacturing process will be inter-
rupted, B.C. Giri and S. Sharma [9] studied the complex production planningn-
ing under the uncertain demand when the product supply is interrupted. Because
there is always a variety of products were produced at the same time, so Shi et al.
[22] considered the demand of many kinds of products are uncertain. For solving
the problem systematically, Cheng [26] studied the influence of the new product
design on the pricing strategy. In order to solve the pricing strategy, a model is
established by using the Brown model [16]. Chen et al. [4] used the dynamic pro-
gramming method to obtain dynamic pricing strategy. Also consider the study
of competition, such as Mitra [18] researched the competition strategy under
the monopoly industry. Adem et al. [20] researched the competition between
an original equipment manufacturer and an independent remanufacture. In the
closed-loop supply chain, Chen and Chang [3] considered a remanufacturing
models under cooperative and competitive.
Due to the emergence of remanufactured products, many consumers have a
different treatment of remanufactured products and new products. Ferrera et
al. [7] chosen differentiated prices for new products and remanufactured prod-
ucts. Benjamin et al. [12] considered the role of ambiguity tolerance in consumer
perception of remanufactured products. [1,19] consider the impact of govern-
ment subsidies on the remanufacturig activities. The scholar Li [15] divided the
demand of products into two categories by the ecological cognition of consumers,
and made a research on the policy of the government subsidy. In this paper, based
on the above, we discuss the problem of remanufacturing in closed-loop supply
chain, based on ecological cognition and government subsid, then formulate the
cooperative and competitive game model to investigate the manufacturer’s and
retailer’s pricing strategy.
1124 D. Yu and C. Guo
This paper is organized as follows: model description is reviewed in Sect. 2

and the proposed model is given in Sect. 3. The key parameters are analyzed
in Sect. 4 and the comparison of models will be given in Sect. 5. Finally, some
conclusions are summarized in Sect. 6.
2 Model Description
In this paper, we consider three stakeholders, they are manufacturer, retailer and
consumer. The manufacturer will produce remanufactured products and new
products, and then sold them together to the retailer. The retailer sold remanu-
factured products and new products to consumers, and collected used products
from the consumers, then provide them to manufacturers. The reverse logistics is
commissioned by the manufacturer, the manufacturer gives the acquisition price
to stimulate the retailers to collect the used products. Thus, the closed-loop
supply chain model is shown in Fig. 1.
Fig. 1. Model structure of closed loop supply chain
The model structure studied the case of a single manufacturer and a single
retailer. Besides, the manufacturer is the leader in the closed loop supply chain.
In this paper, we only consider the government takes the corresponding pol-
icy mechanism to stimulate the manufacturer. The remanufacturer derives the
market demand function of remanufactured products and new products based
on the cognitive degree of consumers, then makes the production planning of
the product and determines the wholesale price of the product. On the other
hand, in order to ensure the orderly conduct of the manufacturing process, the
acquisition price is determined by the manufacturers, and the corresponding
used product planning is established. Retailers are the followers, according to
the manufacturer’s wholesale price and acquisition price to determine the retail
price of finished product and acquisition price of used products, and optimize
their own profit. Finally, we explore the effects of cooperative and competitive
model on pricing strategies.
2.1 Notations
The following notations would be used throughout the model

Θ : Ecological cognition of remanufactured products;

pn : The unit selling price of new product;
pr : The unit selling price of remanufactured product;
C1 : The unit cost of sales for new product;
C2 : The unit cost of sales for remanufactured product;
p1 : The unit acquisition price of the retailer to user;
p2 : The unit acquisition price of the manufacturer to the retailer;
Cr : The unit production cost of remanufactured product;
Cn : The unit production cost of new products;
k : Government subsidies for the remanufacturer productions;
qr : The demand quantity of remanufactured products;
qn : The demand quantity of new products;
ωr : The unit wholesale price of remanufactured products;
ωn : The unit wholesale price of new product;
Πr : The profit function for retail;
Πn : The profit function for manufacturer;
Π : The profit function for the supply chain.
2.2 Assumptions
Assumption 1. The decision of the consumer to purchase the product is based
on the utility of the product. Then we supposed the consumers’ evaluation of
new products is v, and supposed the v is uniformly distribution, its probability
density function is:
1, x ∈ [0, 1]
f (x) =
2, x ∈
/ [0, 1] .
Supposed that the market size is 1, and each consumer only buy one product.
Assumption 2. We defined the θ is the degree that consumer based on environ-
mental awareness and willing to buy the remanufactured products. We supposed
0 ≤ θ ≤ 1, when pn > θv, customers will not buy new products, when pr > θv,
customers will not buy remanufactured products. And if pn ≤ v, pn > θv,
meanwhile, θv − pr ≤ θv − pn , consumers will choose the new products. When
pn ≤ θv, pr ≤ θv, and v−pr ≤ θv−pn , consumers will choose the remanufactured
products.
Assumption 3. Due to make the production continue, here, we assume that
p1 ≤ p2 , Cr ≤ Cn .
Assumption 4. Manufacturers and retailers are based on complete information
symmetry, it means that each of them fully learn of the cost of each other, pricing,
strategy, and other relevant information, so that the next Steinberg game model
can work well.
Assumption 5. Suppose that in the process of remanufacturing, the rate of
production is 100%. It means that no waste products in the process of manufac-
turing. And used production can meet the market demand.
Assumption 6. According to the degree of consumers’ cognition, the market

demand function of remanufactured product and new products is derived:
pn − pr pr
qn =1− , ≤ pn ≤ pr + 1 − θ,
1−θ θ
θpn − pr
qr = , pn + θ − 1 ≤ pr ≤ θpn .
θ(1 − θ)
3 Model Development
Based on the above, we obtain the profit function of the manufacturer and the
retailer, the profit function of manufacturer:
πn = (ωn − cn )qn + (ωr − cr − p2 )qr + kqr . (1)
The first element on the right side of the equation is the profit of the man-
ufacturer for selling new products; The second element on the right side of the
equation is the profit of the manufacturer for selling remanufactured products;
the third element on the right side of the equation is the profit of the manu-
facturer produces the remanufactured products and then gets the government
subsidies.
The profit function of retailer:
πr = (pn − ωn − c1 )qn + (pr − p1 − c2 − ωr + p2 )qr . (2)
The formula on the right side of the first represent the profits of retailer
for selling new products, of which second is stand for the profits for selling
remanufactured products. So there is a total profit of the closed-loop supply
chain:
π = πn + πr = (pn − cn − c1 )qn + (pr − p1 − cr − c2 + k)qr . (3)
In this paper, we consider the following two cases in the closed loop supply
chain.
3.1 Competitive Pricing Strategy
In the competitive pricing strategy, the manufacturer and the retailer take their
own benefit maximization as the decision-making goal respectively. In this deci-
sion models, the manufacturer is in the dominant position, and the retailer is
the follower. So the paper uses the reverse induction method to solve the model.
The derivation of the above formula, we can get:
2
∂ πr /∂ 2 pn = 2/(θ − 1) ∂ 2 πr /∂pn ∂pr = −2/(θ − 1)
.
∂ 2 πr /∂pr ∂pn = −2/(θ − 1) ∂ 2 πr /∂ 2 pr = 2/(θ(θ − 1))
In order to verify whether there is the optimal solution of the equation, we

need to calculate the characteristic value of Hessian matrix and judge that they
are negative. After that we get Matrix eigenvalue:

−(θ − (5θ2 − 2θ + 1)(1/2) + 1)/(−θ2 + θ) 0
2 (1/2) 2 .
0 −(θ + (5θ − 2θ + 1) + 1)/(−θ + θ)
Next, we just need to prove it: (θ − (5θ2 − 2θ + 1)(1/2) + 1)/(−θ2 + θ) > 0.

And then we can get the conclusion that the Hessian matrix is negative, so there
is a corresponding maximum value. Due to the θ ∈ [0, 1], the denominator is
greater than zero, it is necessary to prove that the molecule is greater than zero.
Easy to get 4θ − 4θ2 , so in this model, the retailer has the optimal retail price.
We take the partial derivative of Eq. (2):
∂πr /∂pn = 0∂πr /∂pr = 0.
And obtain that:
pn = (c1 − c2 − θ − p1 + p2 + 2pr + ωn − ωr + 1)/2, (4)

pr = (c2 + p1 − p2 + ωr − c1 θ + 2θpn − θωn )/2. (5)
Using formula (4) and (5), we can get:
pn = (c1 + ωn + 1)/2, (6)

pr = (c2 + θ + p1 + ωr − p2 )/2. (7)
Put pr and pn into the Πn , and then consider the equation:

∂πn /∂ωn = 0,
.
∂πn /∂ωr = 0
At last, we can obtain that:
ωn∗ = (1/2)(cn + 1 − c1 ), (8)

ωr∗ = (1/2)(cr − c2 + θ − k − p1 + p2 ). (9)
Combine Eqs. (6), (7), (8) and (9), we can get that:
p∗n = (1/4)(c1 + cn + 3), (10)

p∗r = (1/4)(c2 + 3θ − k + p1 − p2 + cr ). (11)
∗
In this paper, use to represent the optimization.
3.2 Cooperative Pricing Model

Under this model, the manufacturer and the retailer take the Π as the objective
function. This problem can be translated to the maximum value of the equation:
π = πn + πr = (pn − cn − c1 )qn + (pr − p1 − cr − c2 + k)qr . (12)

The Hessian matrix of the equation is obtained as follows:

2
∂ π/∂ 2 pn = 2/(θ − 1) ∂ 2 π/∂pn ∂pr = −2/(θ − 1)
.
∂ 2 π/∂pr ∂pn = −2/(θ − 1) ∂ 2 π/∂ 2 pr = 2/(θ(θ − 1))
In the same way, we can know that there is a optimal solution. And solved
it, we can get that:
p∗r = (1/2)(c2 + cr + θ − k + p1 + p2 ), (13)

p∗n = (1/2)(c1 + cn + 1). (14)
4 Numerical Example
In order to describe the effect of θ on pn , pr , qn , qr , we fixing the other parameters

as the base setting: C1 = 0, C2 = 0.05, Cn = 0.2, Cr = 0.15, k = 0.04, p1 = 0.05
and p2 = 0.08. Then we made the Tables 1, 2, and compared the different in the
two model. By the way, and in the next graphic analysis, if there is no special
note, the value of the parameters as the same as the table. From the following
table: In the cooperative pricing model, the government’s regulation should be
greatly strengthened. Because in the Table 2, the number of remanufactured
products appear negative growth, and their sales prices are gradually increasing.
It means that the remanufactured products is no profit at all, the only way is to
continuously improve the sales price of remanufactured products to get profit.
As can be seen from the table, the sales price of new product does not change
with the increase of θ. The reason is that we use the reverse induction method
to solve this model. The retailers is the first to determine the market price,
then manufacturer get the price of wholesale. As a result, the sales price of new
products is only related to the cost of sales and production cost of new products.
First, we make some graphic analysis in the model of competition. The effect
of θ on market demand and market profits is now studied and the results are
shown in Figs. 2 and 3 respectively.
For the analysis of the Fig. 2, we have the following conclusions: at the inter-
section, when the θ = 0.8781, the market demand for new and remanufactured
Table 1. The effect of θ on parameter in competition
Variable θ Value of the corresponding parameter variation

parameter
θ 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95 1.0
qr −0.0104 0.0000 0.0119 0.0267 0.0469 0.0784 0.1389 0.3158 −Inf
qn 0.2063 0.2000 0.1917 0.1800 0.1625 0.1333 0.0750 −0.1000 Inf
pr 0.5200 0.5575 0.5950 0.6325 0.6700 0.7075 0.7450 0.7825 0.8200
pn 0.8000 0.8000 0.8000 0.8000 0.8000 0.8000 0.8000 0.8000 0.8000
Table 2. The effect of θ on parameter in cooperation
Variable θ Value of the corresponding parameter variation

parameter
θ 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95 1.0
qr −0.3542 −0.3516 −0.3571 −0.3733 −0.4063 −0.4706 −0.6111 1.0526 −Inf
qn 0.6125 0.6286 0.6500 0.6800 0.7250 0.8000 0.9500 1.4000 −Inf
pr 0.4450 0.4700 0.4950 0.5200 0.5450 0.5700 0.5950 0.6200 0.6450
pn 0.6000 0.6000 0.6000 0.6000 0.6000 0.6000 0.6000 0.6000 0.6000
0.35
0.3
qr
0.25 qn
0.2
Volume of business
0.15
0.1
0.05
-0.05
-0.1
0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1
Cognitive level of remanufacturing products θ
Fig. 2. The effect of θ on market demand.

0.1
0.08
πr
0.06
πn
0.04
0.02
Profits
-0.02
-0.04
-0.06
-0.08
0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1
Cognitive level of remanufacturing products θ
Fig. 3. The effect of θ on market profits
products is equal. At this case, when θ > 0.8781, consumers tend to buy manu-
factured products. Manufacturer can only producing the remanufactured prod-
ucts when the remanufactured product is profitable enough. When θ = 0.9300,
manufacturers no longer produce the new products, because at this point, the
demand for new products is zero. Why it happened? Maybe there is no bias
towards the new products and remanufactured products, and the price of the
remanufactured product is lower than the new product. In Fig. 3 we can draw
that: when θ = 0.9735, the remanufacturer and the retailer’s profit is equal, and
in this mode, the manufacturer produces remanufacturer products are negative
profits. As in Fig. 2, with the increase number of remanufacturer products, the

manufacturer’s profits have been negative growth. And it is found that they
can produce remanufacturer products to adjust the profit distribution between
manufacturer and retailer. When θ is larger than 0.9901, the manufacturer is
no longer profitable, if retailer and government do not adjust it in this case, the
mode of production will no longer continue. It can be seen that the remanu-
facturing does not allow manufacturers to make a profit. As a result, the main
incentive for manufacturers to make the profit will be to improve remanufactur-
ing technology, save remanufacturing costs and get more government subsidies.
Here, we consider the situation: when the degree of ecological cognition is
very low, in order to advocate environmental protection activities, how should
the government make the decisions? We considered the following three cases, see
the chart below.
From the Fig. 4, it is observed that when θ = 0.6000, k1 = 0.05, the retailer’s
profit appears to the minimum value. With the increase of K, which can reduce
the number of new product and increase quantity of remanufactured products.
The wholesale price of products will be reduced with the increase in government
subsidies. As a result, the unit selling price of remanufactured product also begins
to reduce. Due to the decrease in the number of new product, retailers’ profits
0.09
πr
0.085
πn
0.08
0.075
Profits in the maket
0.07
0.065
0.06
0.055
0.05
0.045
0.04
0 0.05 0.1 0.15 0.2
θ =0.6 unit subsidy for remanufactured product k1
Fig. 4. The impact of government subsidies on market profit when θ = 0.6

0.09
0.085 πr
πn
0.08
0.075
0.07
0.065
0.06
0.055
0.05
0.045
0.04
0 0.05 0.1 0.15 0.2

0.1
0.09 r
0.08

0.07
0.06
0.05
0.04
0 0.05 0.1 0.15 0.2
continue to decline. Only when the government subsidy is greater than 0.05, the
retailer’s profit began to grow. We can find the selling profits mainly come from
new products. The pr decreases with the k1 increase, that makes the consumer
tends to buy remanufactured products. As a result, the demand of new product
begin to decline. When the k = 0.13, because the government subsidies make
up for the loss in remanufacturing, the manufacturers’ profit function begins to
grow, so the retailer’s selling price also increased significantly. From the Figs. 4,
5 and 6 we can draw that: in the case of other variables are constant, the smaller
the value of θ, the more serious decline in the profit function. It should be noted
that with the decrease θ the lowest point of red and blue line start move to
the right. We can draw that the higher cognition of remanufactured products,
the government’s subsidy will be smaller. The model also reveals that manufac-
turers and retailers will not be able to make a remanufacturing activity if the
government does not take the policies and give subsidies.
0.65
0.6 θ =0.4
θ =0.6
0.55 θ =0.8
0.5
0.45
r
p
0.4
0.35
0.3
0.25
0 0.05 0.1 0.15 0.2
unit subsidy for remanufactured product k
Fig. 7. The variation of pr with respect to K under different θ
Next, we make some graphic analysis in the model of cooperation. In order to

make more profits, the manufacturer and retailer set prices together. We assumed
that other factors are constant, and then discuss the k if the government wants
0.8
0.6
q n θ =0.8
0.4
Market demand
q r θ =0.8
0.2 q n θ =0.6
q r θ =0.6
0 q n θ =0.4
q r θ =0.4
-0.2
-0.4
-0.6
0 0.05 0.1 0.15 0.2
unit subsidy for remanufactured product k
Fig. 8. The variation of market demand with respect to K under different θ
to make remanufacturing continue. From the previous discussion, we learn that

the change of k does not affect the pn . So we compared the effect of k on pr
in the following cases. As follows: from the Fig. 7, we can clearly see a linear
relationship between the pr and k, the ratio is −0.5. The more greater the k, the
more smaller the unit retail price of remanufactured products. At the same time,
we can also find that in the process of pricing, government subsidies are partic-
ularly important. At the same time, we also compared the demand function of
new products and remanufactured products in the Fig. 8. It is obvious that when
the θ is bigger, the demand market of remanufactured products or new prod-
ucts are more sensitive to the k. Because with the increase of θ, the more steep
line. When the θ = 0.8, the government needs to provide at k = 0.17, because
the quantity of remanufactured products is zero at this point of k = 0.17. Only
when k0.17, the closed loop supply chain can continue to remanufacturing. For
the sake of reducing the subsidy, the government can do the following adjust-
ments: to enhance ecological cognition of remanufactured products and increase
the value of θ. Because in the Fig. 8, we can know that the θ is more bigger, the
possibility is more higher for manufacturer to remanufacturing. Secondly, our
government should establish a complete production system of closed loop sup-
ply chain, it can reduce the unnecessary cost waste. Besides, remanufacturing
costs account for a large proportion of the cost of production, and it seriously
hinders the development of remanufacturing. It is necessary for government to
provide technical support for the remanufacturing. At last, if the government
can destruct the joint policy of manufacturers and retailers, it will take profit
to him. Because in the joint pricing strategy, the new products can bring more
profits to them.
5 Comparison of Models
In this paper, we study cooperative and competitive game model in the closed
loop supply chain. In the model of competition, due to the manufacturers and
retailers make decisions respectively, there is competition between them. Pro-
ducing the remanufactured products can adjust the interests of the relationship
between the manufacturer and the retailer. It is benefit for the government to
promote remanufacturing. And the retailer more like to make remanufactured
products to compete interests with manufacturer. But From Tables 1 and 2 we
can obviously find remanufacturing hurts the interests of manufacturer, retailer
and consumer. So them will choose cooperative pricing model. In the Fig. 7, it
is not difficult to find that the total profit of cooperation is always higher than
the competition. Producing remanufactured products could not make profits, so
the profits always decrease with the increase of the degree of cognition. When
the θ from 0.9 to 0.95, there is a sharp decline in the total profit of the two
models. The reason is the rapidly decrease in the number of new products and
the number of remanufactured products rapidly increase. It proves again that
produces the remanufactured products will not bring profit to manufacturers.
The main reason of remanufacturing could not make profits is that the remanu-
facturing technology has not kept pace with the society. It also reveals that if the
government not take any actions, it will happen nothing on remanufacturing.
6 Conclusion
In this paper, we consider the closed loop supply chain that consists of the
manufacturer, retailer and consumers, discuss two models of cooperative and
competitive game model. At the same time, we using the game theory to analysis
the model and the conclusions are the following.
In the cooperative and competitive game model, if the government no sup-
port or no provide subsidy, remanufacturing will not be able to carry on. And
in the cooperative game model, the government’s regulation should be strength-
ened. Because the cooperation is not conducive to producing remanufactured
products. As the number of new products and profits are declining, manufactur-
ers have a reason to down the price of new products to get more quantity. But
the optimal wholesale price in this paper has nothing to do with θ, maybe in
this model we use the reverse induction method to solve the model, the retailers
is the first to determine the market price, then manufacturer make the price
of wholesale. What’s more, producing remanufactured products can adjust the
profit distribution between manufacturer and retailer. And in cooperation, it
can bring higher profit to the supply chain. But it does not accord with the
interests of the government. In order to save economic costs, the government is
more willing to see them make decisions separately. With the θ raised, it will not
good for manufacturers and retailers, but it is good for the government to take
actions. The discussion of these two models is very important for our research
in the future, and give the direction of the government to advocate green envi-
ronmental protection. In the future work, we could consider the complex closed
loop supply chain systems with multiple manufacturers and multiple retailers,
and the used products will be pre-sale to the manufacturers. What’s more, used
products will be divided into different grades by the consolidation center. The
consolidation center will be the new participator in the future model.
Acknowledgements. The authors wish to thank the Major International Joint

Research Program of the National Natural Science Foundation of China (Grant
No. 71020107027), the National Natural Science Foundation of China (Grant No.
71001075), Doctoral Fund of Ministry of Education of China (No. 20110181110034),
and Social sciences program of Chengdu (No. SKNHR13-07), Central University Fund
of Sichuan University (No. skqy201112) under which the present work was possible.
References
1. Aksen D, Aras N, Karaarslan AG (2009) Design and analysis of government
subsidized collection systems for incentive-dependent returns. Int J Prod Econ
119(2):308–327
2. Chari N, Diallo C et al (2016) Production planning in the presence of remanufac-
tured spare components: an application in the airline industry. Int J Adv Manuf
Technol 87(1):957–968
3. Chen JM, Chang CI (2012) The co-opetitive strategy of a closed-loop supply chain
with remanufacturing. Transp Res Part E Logist Transp Rev 48(2):387–400
4. Chen JM, Chang CI (2013) Dynamic pricing for new and remanufactured products
in a closed-loop supply chain. Int J Prod Econ 146(1):153–160
5. Choi TM, Li Y, Xu L (2013) Channel leadership, performance and coordination in
closed loop supply chains. Int J Prod Econ 146(1):371–380
6. Dutta P, Das D et al (2016) Design and planning of a closed-loop supply chain
with three way recovery and buy-back offer. J Clean Prod 135:604–619
7. Ferrer G, Swaminathan JM (2010) Managing new and differentiated remanufac-
tured products. Eur J Oper Res 203(2):370–379
8. Fleischmann M (2004) Quantitative models for reverse logistics. Springer, Heidel-
berg
9. Giri BC, Sharma S (2015) Optimal production policy for a closed-loop hybrid
system with uncertain demand and return under supply disruption. J Clean Prod
112:2015–2028
10. Gönsch J (2015) A note on a model to evaluate acquisition price and quantity of
used products for remanufacturing. Int J Prod Econ 169:277–284
11. Guide VDR, Wassenhove LNV (2009) The evolution of closed-loop supply chain
research. Oper Res 57(1):10–18
12. Hazen BT, Overstreet RE et al (2012) The role of ambiguity tolerance in consumer
perception of remanufactured products. Int J Prod Econ 135(2):781–790
13. Ilgin MA, Gupta SM (2010) Environmentally conscious manufacturing and product
recovery (ecmpro): a review of the state of the art. J Environ Manage 91(3):563–591
14. Jung KS, Dawande M et al (2016) Supply planning models for a remanufacturer
under just-in-time manufacturing environment with reverse logistics. Ann Oper
Res 240(2):1–49
15. Li XQ (2015) Research on the policy of remanufacturing subsidy based on the
ecological cognition of consumers. Prod Res 2015(9):12–19 in Chinese
16. Liang Y, Pokharel S, Lim GH (2009) Pricing used products for remanufacturing.
Eur J Oper Res 193(2):390–395
17. Matsumoto M, Komatsu S (2015) Demand forecasting for production planning in
remanufacturing. Int J Adv Manufact Technol 79(1):161–175
18. Mitra S (2015) Models to explore remanufacturing as a competitive strategy under
duopoly. Omega 20:215–227
19. Mitra S, Webster S (2008) Competition in remanufacturing and the effects of gov-
ernment subsidies. Int J Prod Econ 111(2):287–298
20. Örsdemir A, Parlaktürk AK (2014) Competitive quality choice and remanufactur-
ing. Soc Sci Electron Publ 23(1):48–64
21. Pokharel S, Liang Y (2012) A model to evaluate acquisition price and quantity of
used products for remanufacturing. Int J Prod Econ 138(1):170–176
22. Shi J, Zhang G, Sha J (2011) Optimal production planning for a multi-product
closed loop system with uncertain demand and return. Comput Oper Res
38(3):641–650
23. Teunter RH, Flapper SDP (2011) Optimal core acquisition and remanufacturing
policies under uncertain core quality fractions. Eur J Oper Res 210(2):241–248
24. Thierry M, Salomon M, Nunen JV, Wassenhove LV (1995) Strategie issues in
product recovery management. Calif Manag Rev 37(2):114–135
25. Wei S, Cheng D et al (2015) Motives and barriers of the remanufacturing industry
in China. J Clean Prod 94:340–351
26. Wu CH (2012) Product-design and pricing strategies with remanufacturing. Eur J
Oper Res 222(2):204–215
Beds Number Prediction Under Centralized
Management Mode of Day Surgery
Jianchao Yang1 , Luo Li1 , Hongsheng Ma2 , and Yong Luo1(B)

1
scuyangjc@163.com
2
West China Hospital, Sichuan University,
Abstract. In this paper, we firstly use simulation methods to reach

the conclusion that the beds allocation scheme under centralized man-
agement mode is more advantageous to either elective surgery patients
or day surgery patients, and the beds allocation scheme under decen-
tralized mode is more equitable. Then we do a large number of experi-
ments to explore the beds number allocate to day surgery patients under
centralized management mode, and seek the allocation scheme which is
advantageous to day surgery patients and the equity is closest to the
same situation under decentralized mode. At last, based on the data we
obtain from the experiments, regression and neutral network are used to
predict the beds number allocated to day surgery patients under central-
ized management mode.
Keywords: Day surgery · Beds number prediction · Linear regression ·

Neural network
1 Introduction
There are three types of surgery management modes adopted overseas: day
surgery center inside a hospital, free-standing day surgery center and opera-
tion room at the clinic [7]. The mainly adopted mode in our country is day
surgery center inside a hospital under three different management modes: cen-
tralized management mode, decentralized management mode and the manage-
ment mode combining centralized mode and decentralized mode. Centralized
management mode is an integrated management mode where day surgery cen-
ter, as the centralized management platform, assembles all patients together to
conduct centralized admission, scheduling and follow-up visit. While decentral-
ized management mode is a mode where day surgery is managed by departments.
In this paper, we study hospitals under the mode where centralized and
decentralized management coexist. Under centralized management mode, day
surgery patients are admitted in day surgery center where their admission, oper-
ation and discharge are arranged. Day surgery center is self-contained with wards
Beds Number Prediction under Centralized Management Mode 1137
and operation room. The beds and operation rooms of day surgery patient are
dedicated, not shared with elective patents. Under decentralized mode, wards
and operation rooms of day surgery and elective surgery are managed by depart-
ments. The beds and operation rooms has no boundary between day surgery
patients and elective surgery patients, i.e. to some extent, beds and operations
are shared by day surgery patients and elective surgery patients.
We conduct simulation study on centralized and decentralized management
modes in this paper. Through a large number of simulation experiments, we
reach the conclusion that centralized management mode is more advantageous
to either elective surgery patients or day surgery patients. And decentralized
management mode is more equitable. While the development of day surgery
is encouraged, centralized management mode is a better choice. But if beds
allocation is far from equitable under centralized management mode, it will be
exceedingly advantageous to one of the two types of patients. Thus, in this paper,
we seek the beds number allocate to day surgery patients, under Centralized
management mode, that is advantageous to day surgery patients and has Equity
Closest to the same situation under Decentralized mode (CECD). Based on the
large amount of simulation data, we screen out characteristic variables and use
regression and neural network to predict the beds number allocate to day surgery
patients under CECD. And then we verify the effectiveness of the regression
equation and neural network model.
Technological and theoretical advances in computer science and mathemat-
ics offer new options to complement traditional statistical analysis [7]. A major
focus of machine learning research is to automatically learn to recognize com-
plex patterns and make intelligent decisions based on data. In the past decade,
machine learning algorithms have revealed previously undetected trends in his-
torical data [9,13]. Especially, artificial neural network and traditional linear
regression are effective method. There are some studies use linear regression and
neural network to predict the objective function value. Such as, Menke et al.
[7] designed an artificial neural network to predict emergency department vol-
ume. Shi et al. [11] performed hierarchical linear regression and propensity score
matching to test hospital/surgeon volume for associations with breast cancer
surgery costs. Tsai [12] developed artificial neural network models to predict
length of stay for inpatients with one of the three primary diagnoses: coronary
atherosclerosis, heart failure, and acute myocardial infarction in a cardiovascular
unit in a Christian hospital in Taipei, Taiwan. Li et al. [15] proposed an artificial
neural network model to predict the severity of menopausal symptoms. Launay
et al. [5] used artificial neural networks to predicate of prolonged length hospi-
tal stay in older patients hospitalized in acute care wards after an emergency
department. Gholipour et al. [4] used a neural network for predicting survival
and length of stay of patients in the ward and the intensive care unit of trauma
patients and to obtain predictive power of the current method. Wise et al. [14]
used artificial neural network model to provide vascular surgeons a discriminant
adjunct to assess the likelihood of in-hospital mortality on a pending Ruptured
abdominal aortic aneurysm admission. Some studies use linear regression and
1138 J. Yang et al.
neural network to study the influencing factors of disease or identify influencing

factors. Such as, Shakerkhatibi et al. [10] were conducted to evaluate the rela-
tionship between air pollutants and hospital admissions for cardiovascular and
respiratory diseases using the artificial neural networks and conditional logis-
tic regression modeling. Abdullah [1] aimed to model the relationship between
Health Related Quality of Life variables using an integrated model of fuzzy
inference system and linear regression. Bonellie [3] used regression and logistic
regression models to investigate changes over time in size of babies particularly in
relation to social deprivation, age of the mother and smoking. Some studies use
linear regression and neural network do clustering and classification, Aguiar et
al. [2] developed one artificial neural network model for classification and another
risk group assignment for Pulmonary tuberculosis in hospitalized patients. And
some studies focus on the methods improvement. Russell et al. [8] through their
research discovered that bayesian linear regression provides a significantly more
accurate estimate of the rate of change in mean sensitivity than the standard
ordinary least squares linear regression approach.
Artificial neural network and traditional linear regression can be well applied
to the prediction of the objective function value, the identification of the influenc-
ing factors, clustering and classification. In our paper, we use linear regression
and neural network to predict the objective function value, the beds number
under CECD.
2 Evaluation of Two Management Mode

of Day Surgery Based on Simulation
2.1 Parameters
In this paper, taking Laparoscopic Cholecystectomy (LC) as an example, we

build two simulation models. Model I is the centralized management model,
where beds and operation rooms occupied by day surgery patients and elective
surgery patients are independent. And there are dedicated beds and operation
rooms for day surgery patients. That is day surgery patients wait in a queue for
admission and operation according to the First Come and First Served (FCFS)
policy, they can only occupy beds and operation rooms dedicated to day surgery
patients on admission. And elective surgery patients wait in other queue for
admission and operation according to the FCFS policy, they can only occupy
beds and operation rooms dedicated to elective surgery patients on admission.
Model II is the decentralized management mode, where beds and operation
rooms occupied by day surgery patients and elective surgery patients are shared,
i.e. day surgery patients has no dedicated bed or operation room. Day surgery
patients and elective surgery patients wait in one queue for admission and oper-
ation according to the F CF S policy.
Significant input parameters of the model are as follows.
B : total number of beds for day surgery patients and elective surgery
patients;
BD : number of beds for day surgery patients;
BE : number of beds for elective surgery patients, BD + BE = B;
λD : arrival rate of day surgery patients (unit: person per day);
λE : arrival rate of elective surgery patients (unit: person per day);
OTD : operation time of day surgery patients (unit: hour);
OTE : operation time of elective surgery patients (unit: hour);
RTD : length of hospital stay after operating day surgery (unit: hour);
RTE : length of hospital stay after operating elective surgery (unit: hour);
W T1 : average waiting time before admission (unit: hour), in this paper
we define
W T1 = time of admission − time of arrival;
AD : number of day surgery patients arrived (unit: person);
AE : number of elective surgery patients arrived (unit: person);
SD : number of day surgery patients served (unit: person);
SE : number of elective surgery patients served (unit: person);
S : total number of surgery patients served (unit: person),
S = SD + S E
2.2 Simulation Results and Analysis
In this paper, we define the regular opening hour of each operation room as 8 h,
at the end of each workday, operations still in progress should be finished with
extra work, newly-opened operations are not allowed, and operation rooms are
closed for weekends. When we conduct simulation in this paper, OTD , RTD ,
OTE and RTE are considered as random numbers generated according to the
mean values and standard deviations in Table 1.
Table 1. Descriptive statistics result [6]
Surgery type N Mean value Min Max SD SE mean

OT Elective surgery 1841 72.2819 20 1116 56.92553 1.32672
Day surgery 682 64.0176 20 674 33.30224 1.27521
RT Elective surgery 1841 2696.3781 920 43285 2391.37266 55.73402
Day surgery 682 1198.2361 600 8355 713.50039 27.32134
The simulation period of Model I and Model II are both 1 year, and the
conclusion we reach through a large number of simulation experiments are in
accordance. In this paper we only elaborate on the simulation result when D =
8, E = 30. The number of elective surgery patients arrived (AD ), number of
elective surgery patients arrived (AE ), number of day surgery patients served
(SD ), number of elective surgery patients served (SE ), total number of surgery
1140 J. Yang et al.
Table 2. Simulation result when λD = 8, λE = 30 (simulation period: 1 year, approx-

imately 260 days)
Model AD (p) AE (p) SD (p) SE (p) S(p) W T1 (h)

Model I (BD = 4) 2080 7800 1118 4324 5442 1397
Model I (BD = 5) 2080 7800 1404 4247 5651 1308
Model II 2080 7800 1148 4306 5454 1382
patients served (S), average waiting time before admission (W T1 ) are shown in
Table 2.
According to Table 2, under the condition where BD = 4, Model I has
30 fewer day surgery patients served than Model II, 18 more elective surgery
patients served than Model II. To sum up, Model I has 12 fewer patients served
than Model II, and 15 h longer average waiting time before admission than
Model II.
While, under the condition where BD = 5, Model I has 256 more day-surgery
patients served than Model II, 59 fewer elective-surgery patients served than
Model II. To sum up, Model I has 197 more patients served than Model II, and
74 h shorter average waiting time before admission than Model II.
The result shown in Table 2 indicates that, under centralized management
mode, bed and operation room resources of day surgery and elective surgery
are respectively independent, and it is advantageous either to elective surgery
patients (under the condition where BD ≤ 4 in Table 2) or day surgery patients
(under the condition where BD ≥ 5 in Table 2). While, under decentralized
mode, bed and operation room resources are shared between day surgery patients
and elective surgery patients, so it is more equitable. Under centralized manage-
ment mode, when allocation is more advantageous to elective surgery patients
(i.e. fewer beds are allocated to day surgery patients, and more elective surgery
patients are served at the cost of having fewer day surgery patients served than
decentralized mode), fewer patients are served within equal time period and the
average waiting time before admission is much longer when compared to decen-
tralized mode. When allocation is more advantageous to day surgery patients
(i.e. more beds are allocated to day surgery patients, and more day surgery
patients are served at the cost of having fewer elective surgery patients served
than decentralized mode), a lot more patients are served within equal time period
and the average waiting time before admission is much shorter when compared
to decentralized mode. While the development of day surgery is encouraged,
centralized management mode is a better choice. However, if bed allocation is
far from reasonable under centralized management mode, it will be exceedingly
advantageous to one of the two types of patients, such as the result in Table 2
when BD < 4 or BD > 5. For the convenience of applying the predictions in
hospital management, the following work in this paper use simple calculation
to explore the beds number allocate to day surgery patients under CECD, i.e.
situation in Table 2 where BD = 5, that is advantageous to day surgery and
has equity closest to decentralized mode. Taking no account of limitation to

operation room resource, we conduct a large number of simulation experiments
in this paper and obtain the data of beds number under CECD with distinct
arrival and service time. By screening out characteristic variables, we predict the
beds number under CECD through linear regression and neural network in the
following sections.
3 Linear Regression Prediction

In this section, we use linear regression to predict the beds number allocate to
day surgery patients under CECD. Table 3 shows the value of parameters used in
regression, where input parameters are shown in the first 11 columns, M VOTD
denotes the mean value of OTD , SDOTD denotes the SD of OTD and so on.
BCECD denotes the output of predictions. 100 sets of experiment data are used.
Table 3. Value of input and output parameters used in regression
λD λE M VOTD SDOTD M VOTE SDOTE M VRTD SDRTD M VRTE SDRTE B BCECD

10 13 45 21 130 55 892 430 3470 1626 26 4
25 9 95 13 125 43 1098 132 1427 1046 24 13
13 16 44 25 165 58 744 451 1645 2394 11 2
··· ··· ··· ··· ··· ··· ··· ··· ··· ··· ··· ···
26 11 53 17 78 22 1059 150 1151 949 18 9
25 11 55 26 67 54 1161 397 3167 1825 27 10
Let all the input and output in Table 3 stay unchanged, we enter them
into regression equation and find redundant variables in the input parameters.
Through stepwise regression, we identify M VOTD , SDOTD , M VOTE , SDOTE ,
SDRTD and SDRTE as redundant variables. In hospital management practice, the
operation time is usually shorten than the length of hospital stay after surgery, so
M VOTD , SDOTD , M VOTE , SDOTE are identified as redundant variables.
We let C0 to C5 represents constant, coefficient of λD , coefficient of λE ,
coefficient of M VRTD , coefficient of M VRTE and coefficient of B. The results of
linear regression are shown as Tables 4 and 5.
Table 4. Coefficient of linear regression of the input
C0 C1 C2 C3 C4 C5
0.5248 0.2386 −0.4378 0.0036 −0.0008 0.2726
Table 5. Test indicators of linear regression
R2 statistic F statistic p value An estimate of the error variance

0.9162 205.5406 0.0000 1.0990
1142 J. Yang et al.
As can be seen from Table 5, the regression equation is valid. So BCECD , the
beds number allocate to day surgery patients under CECD can be represented
by the formula (1).
BCECD = 0.5248 + 0.2386λD − 0.4378λE + 0.0036M VRTD

(1)
− 0.0008M VRTE + 0.2726B.
The actual beds number allocate to day surgery patients under CECD is inte-
ger, thus, we rewrite Eq. (1) by rounding it to integer and we obtain it denoted
as BCECD ∗ by the formula (2).
∗
BCECD = 0.5248 + 0.2386λD − 0.4378λE + 0.0036M VRTD
(2)
− 0.0008M VRTE + 0.2726B + 0.5.
The actual and the predicted beds number allocate to day surgery patients
under CECD calculated according to formula (2) are shown in Figs. 1 and 2:
The pentagrams in Fig. 2 denote entirely accurate predictions. From Figs. 1
and 2, we can see that not much predictions are entirely accurate, but the maxi-
mum error is 3. The results of the key parameters which reflect regression effect
are: R2 = 0.9162, F = 205.5406, p = 0.0000. Parameter R2 = 0.9162 indicates
Fig. 1. The predictions with fluctuation using linear regression
Fig. 2. The entirely accurate predictions using linear regression

that the fitting result of the model is satisfied, parameter p < α, which indicates
that selection of every variable in the regression equation is significant. The
results of Table 5 indicate that Eq. (1) has a satisfied fitting result. The regres-
sion equation can provide decision supports for beds allocation under centralized
management mode.
∗
Next section, we predict the number of BCECD with neural network method.
4 Neural Network Prediction

It has identified the input layer has five units. And there is one output unit
represents the beds number allocate to day surgery patients under CECD. Neural
network theory has been proved that if the number of hidden layer units can
be set free, then with three’s like I/O node characteristics can approximate any
continuous function with any precision. Therefore, we choose networks with three
architectures, that is only one hidden layer. And there are eight units at hidden
layer.
The error trend after we trained 500 times are seen as Fig. 3.
Based on the neural network above, the results of prediction we obtained are
as follows (Tables 6, 7 and 8).
Fig. 3. The error trend using Neural Network Prediction
Table 6. The value of w1
B1 B2 B7 B9 B11
−0.4018 0.3646 −0.2478 0.1699 −0.1319
−0.2532 0.2661 −0.0389 0.067 −0.3323
−0.2306 0.2747 −0.0048 0.0475 −0.3204
−0.1817 0.2614 0.0323 0.0675 −0.3431
−0.353 0.3008 −0.1178 0.0916 −0.2154
0.6947 −0.9056 0.4409 −0.2699 0.7933
−0.3098 0.2657 −0.0423 0.0847 −0.2985
0.0418 0.0656 0.1519 0.0017 −0.1027
1144 J. Yang et al.
Table 7. The value of w2
−1.2730 −0.9122 −0.7375 −0.6258 −1.1009 1.2214 −0.9976 −0.0026
Table 8. Threshold of hidden layer and output layer
B1 B2
−1.0442 −1.6046 −1.7828 −1.8679 −1.3586 −1.7432 −1.4882 −2.0515 0.5287
Fig. 4. The predictions with fluctuation using Neural Network Prediction
Fig. 5. The entirely accurate predictions using Neural Network Prediction
The pentagrams in Fig. 5 denote entirely accurate predictions. From Fig. 4

and Fig. 5, we can see that not much predictions are entirely accurate, but the
maximum error is 3. And the neural network above can provide decision supports
for beds allocation under centralized management mode.
5 Conclusion
In this paper, we firstly conduct a large number of simulation experiment using
Matlab and prove that centralized management mode is more advantageous
to either elective surgery patients or day surgery patients, it depends on the
allocation of beds; while decentralized management mode is the most equitable.

Then, based on the large amount of simulation data, we screen out characteristic
variables and obtain the regression equation of the CECD beds number, the
general linear regression equation is rounded to integer as
∗
BCECD = 0.5248 + 0.2386λD − 0.4378λE + 0.0036M VRTD
− 0.0008M VRTE + 0.2726B + 0.5.
The regression equation satisfied fitting results, the maximum error is 3, and
only a few recorded errors are 3, most are 1 and 2 besides the accurate pre-
dictions. And the beds prediction equation in this paper, which is simple and
convenient for calculation. Then we use artificial neural network to do the same
prediction. And the results are similar to linear regression. The maximum error
is 3, and only a few recorded errors are 3, most are 1 and 2 besides the accu-
rate predictions. Both linear regression equation and the neural network above
can provide decision supports for beds allocation under centralized management
mode.
The deficiency in this paper is that we only consider beds number without
considering the limitation to operation room resource. We will take account of
the limitation to operation room resource to enrich our study next. Moreover,
we will figure out other methods to improve accuracy of the prediction.
Acknowledgement. This paper is supported partly by the National Natural Science

Foundation of China under Grant Nos. 71532007, 71131006 and 71172197.
References
1. Abdullah L (2014) Modeling of health related quality of life using an integrated
fuzzy inference system and linear regression. Procedia Comput Sci 42:99–105 (in
Chinese)
2. Aguiar F, Torres R et al (2016) Development of two artificial neural network models
to support the diagnosis of pulmonary tuberculosis in hospitalized patients in Rio
de Janeiro, Brazil. Med Biol Eng Comput 54:1–9
3. Bonellie S (2012) Use of multiple linear regression and logistic regression models
to investigate changes in birth weight for term singleton infants in Scotland. J Clin
Nurs 21:92–114
4. Gholipour C, Rahim F et al (2015) Using an artificial neural networks (anns) model
for prediction of intensive care unit (icu) outcome and length of stay at hospital
in traumatic patients. J Clin Diagn Res 9:1096–1105
5. Launay C, Rivière H et al (2015) Predicting prolonged length of hospital stay in
older emergency department users: use of a novel analysis method, the artificial
neural network. Eur J Intern Med 26:478–482 (in Chinese)
6. Luo L, Luo Y et al (2014) Difference analysis of day surgery’s and elective surgery’s
duration associated with surgery. Stat Inf Forum 29:104–107 (in Chinese)
7. Menke N, Caputo N et al (2014) A retrospective analysis of the utility of an
artificial neural network to predict ed volume. Am J Emerg Med 32:614–617
1146 J. Yang et al.
8. Russell R, Malik R et al (2012) Improved estimates of visual field progression

using bayesian linear regression to integrate structural information in patients with
ocular hypertension. Invest Ophthalmol Vis Sci 53:2760–2769
9. Schmidt M, Lipson H (2009) Distilling free-form natural laws from experimental
data. Science 324:81–85
10. Shakerkhatibi M, Dianat I et al (2015) Air pollution and hospital admissions for
cardiorespiratory diseases in Iran: artificial neural network versus conditional logis-
tic regression. Int J Environ Sci Technol 12:504–523
11. Shi H, Chang H et al (2013) Breast cancer surgery volume-cost associations: hier-
archical linear regression and propensity score matching analysis in a nationwide
taiwan population. Surg Oncol 22:178–183 (in Chinese)
12. Tsai P, Chen P et al (2016) Length of hospital stay prediction at the admission
stage for cardiology patients using artificial neural network. J Healthc Eng 6:5–8
(in Chinese)
13. Waltz D, Buchanan B (2009) Computer science. Automating science. Science
324:43–44
14. Wise E, Hocking K, Brophy C (2015) Prediction of in-hospital mortality after
ruptured abdominal aortic aneurysm repair using an artificial neural network. J
Vasc Surg 62:8–15 (in Chinese)
15. Xian L, Feng C (2015) Predicting menopausal symptoms with artificial neural
network. Exp Syst Appl 42:8698–8706 (in Chinese)
A Standardization Methodology for Visual
Management in Lean Supply Chain
Environments
André Simas(B) and Virgilio Cruz-Machado
UNIDEMI (Department of Mechanical and Industrial Engineering),

Faculdade de Ciências e Tecnologia, Universidade Nova de Lisboa,
2829-516 Caparica, Portugal
af.simas@campus.fct.unl.pt
Abstract. A rather important tool of Lean used in a plant is Visual

Management, because people pay more attention to everything that is
visual than any other sense. This kind of eyes management allows the
abnormalities to stand out and simplify the work that needs to be done.
Visual Management within an organization, assumes that all the tools
used in the plant are standard, to be the same in the rest of the organi-
zation. In order to help organizations about the develop of visual stan-
dards, the author proposed to develop a Standardization Methodology
divided in two sections, the Current Situation Analysis Methodology
and a Future Design Methodology. The first one refers to all stages that
precede the analysis of instruments of some plants, while the second
one concerns the creation of improvements. The enhancement proposals
were focused on processes that could be an advantage to the existence
of a standard tool. Most of these tools are low cost of employment and
there are some advantages in their use: they allow more transparency on
processes and they make the shop floors look like cities.
Keywords: Lean · Visual management · Standardization ·

Methodology
1 Introduction
Nowadays, the consequences of the Great Recession are still quite noticeable, a
downfall that started in 2008 with the Lehman Brothers Holding Inc., an invest-
ment bank, declaring bankruptcy. Several other banking institutions followed
the same path after the beginning of crisis. After that period, the companies
focused on a clear objective, use innovation to be more competitive.
The automotive industry has also suffered some financial distress, leading
manufacturing organizations to start to think of new production methods and
a better application of the tools used at the shop floor. This allows companies
to pursue operational excellence with an improvement of their performance,
1148 A. Simas and V. Cruz-Machado
providing high quality products earlier in order to reduce costs [2]. Worldwide,
the companies start to encourage a Lean culture on their employees due to the
competitive nature of this industry, caused by emerging markets which produce
similar products at lower prices. To face those markets, the automotive industry
must be constantly updated, developing new products to customers.
Lean Thinking is the basis of implementation within companies, providing a
guide on how they can do more with less, less effort, equipment, time and space.
This can be divided in a set of five principles: value, value stream, flow, pull and
perfection [10].
Implementing Lean can be a difficult process, but when it is well applied
becomes an essential characteristic within an organization, playing a crucial role
leading up to the intended results. Some authors said this philosophy is deter-
minant for the survival and constant innovation of organizations. One tool from
Lean culture is Visual Management, an important tool used at the shop floor,
which is defined as signs and other forms of visual information used to simplify
the workplace and make it easy to recognize abnormalities on processes [4].
2 Visual Management
Visual Management is a Lean tool that has the objective of giving visual infor-
mation or displaying requirements to set directions [3]. This tool is frequently
used at manufacturing industries, but has currently expanded to other business
industries [1]. This concept was created to highlight abnormalities directly in the
workplace, thus helping operations and processes as soon as a problem occurs
[8]. Giving the right information to employees at the right moment is a vital
characteristic to improve the performance of a company [3]. That kind of infor-
mation can be provided by signs, labels or a colour code. The use of this type of
information eliminates the “guess”, searching and a cumulative of information
or material [6].
Many authors, such as Wilson [9] don’t use the term “visual management”
but rather transparency, because this tool allows the observation of processes on
time. An operator can see what is happening in a process and he can change or
adjust it if an abnormality is to occur.
Within Visual Management there are tools that support operators’ tasks and
highlight any abnormalities on processes. There are two different sets of tools [3]:
• Understanding processes: tools mainly focused on visualising the information

to allow a better understanding of the processes. Ex.: value stream mapping,
flow charts, A3 and area name boards.
• Performance of processes: tools that are used to communicate requirements
and manage the efficiency of the processes. Ex.: andon lights and boards,
Kanban and KPIs screens.
Many of these signals were created to control robust processes, allowing to

draw attention to many processes at once and act immediately [7].
A Standardization Methodology 1149
Some authors argue that Visual Management is the basis of sustainable

improvement, promoting the contribution of all operators in management activ-
ities and a quality enhancement of processes. This can only be obtained when
the information is conveyed in a succinct and clear manner [5].
The use of Visual Management has many benefits:
• less time spent understanding the information;
• visible abnormalities, installation of devices/signals;
• speed, the problems are eliminated earlier;
• involvement of all, promoting a continuous improvement;
• standardization, processes up to date.
On this list, only some of the most important benefits were referred, there
are many others that are related with this Lean technique. For Parry & Turner
[8] visual management must be simplified, only information that adds value to
the processes should be displayed.
3 An Approach to Establish a Standardization

of Visual Tools
Mostly organizations that have visual management on the mindset of their
employees and use it constantly, can make the process flow through these visual
aids. Employees must interact with the tools from visual management to continue
the process, if they don’t have that support the process doesn’t flow smoothly
until the end.
Every workplace has visual aids. This aids can be presented in many shapes,
for example: instructions of how to do a stage of process; directional signals;
photographs and boards. However, this information is often neglected because
many organizations don’t follow the requirements of visual management. These
requirements are all information to bring added value, easily understandable and
must be placed at visible places and updated.
An organization that defines very good standards for visual controls has
more benefits and employee efficiency. Then a methodology will be described,
that provides a simple guide for organizations who want to standardise their
visual tools.
3.1 Methodologies
A methodology characterizes all the stages to follow from a process. It is a
detailed explanation of all actions of the work until the expected result is reached.
For better support, a Standardization Methodology divided in two sections
was developed:
• Current Situation Analysis Methodology - this refers to all stages that precede
the analysis of instruments of some plants;
• Future Design Methodology - this refers to the creation of improvements.
All stages must be strictly followed for the success of a project, if the
implementer doesn’t respect the stages, the results can’t be truly trusted. This
Methodology was designed for organizations that have multiple plants worldwide
and the tools are often unequal, but can be used in all organizations with few
adjustments.
(1) Current Situation Analysis Methodology
This Methodology is the initial phase of a standardization, representing
the analysis of information that the implementer collects from an organization
(Fig. 1).
Research of organization
Visualize tools at the shop floor
Identification of possible tools to improve
Data Collection
Data analysis
Fig. 1. Stages of current analysis methodology
1 Research of organization
This phase must be the most detailed one considering the resources available,
documents or procedures that are used by the organization in the composition of
the workstation, which incorporates the visual controls. The person who wants
to implement this methodology should know if the organization has some Lean
policy that can match with this methodology.
2 Visualize tools at the shop floor
The objective of this stage is to get used to the tools that are being used
at the shop floor, through its visualization in operation. This knowledge is the
transition from theory, studied in the previous phase, to practice. The need
of seeing tools at the shop floor can clarify some misunderstandings of theory
information from documents and can be a way to identify abnormalities on tool
applications.
3 Identification of possible tools to improve
After observing the tools in operation, the implementer should use critical
sense to identify possible tools for improvement. It is important to keep in mind
the limitations of the organization, starting with a few number of improvements
is better than start with a huge change on tools. Employees tend to accept
changes if they are implemented slowly.
4 Data Collection
This is the critical phase of all the methodology, because at this stage is where
the implementer gathers the information from plants related to the tools for
improvement. This data collection combines interviews, by email or telephone,
with photos to support what they said, and visits to other plants from the
organization to check on how they apply the tools. The characteristics required
from plants are based on how the tool is used, its appearance, to whom it is
intended and where it is used inside the shop floor.
5 Data analysis
After the data collection is concluded, the information of all plants must be
placed on paper at the beginning (like a draft). When the draft is completed, the
implementer can introduce that in a computer and print it. It is recommended
to print in large formats for easier comprehension. The disposal of information
provides a good view of how tools are being used, if similar, or not, to the other
plants. When a plant has a good practice of some tool, it must be used in total
or partially for improvement of that tool.
(2) Future Design Methodology
This part of the Methodology is the development phase, which include the
conception and the implementation at the shop floor (Fig. 2).
Define assumptions/Set criteria
Develop improvements
Improvement validation process
Improvement implementation
Fig. 2. Stages of future design methodology
1 Define assumptions/Set criteria

The conception of improvements is a subject that has some rules and must
be studied before presenting the improvements. If there aren’t rules or a previous
study those tools can be worse than the actual. For an initial phase, it should
be guaranteed that some criteria are settled by the organization that uses this
methodology. The criteria are:
• Objectivity: must be clear, to not confuse employees;

• Similar: if a plant has a good tool, that tool should suffer a few adjustments
to be similar;
• Reliable: improvements proposals shall consider their economic reliability;
• Simple and intuitive: visual tools are easier to retain;
• Respect layouts and colours: when developing improvements, the rules estab-
lished by organization must be respected.
The need to establish criteria is the basis for a successful implementation of

a strategy and for a best probability of being validated by the Central Team.
2 Develop improvements
Through analysis of collected information from plants, a brainstorming must
be done for chosen tools, following the previously explained aspects. That brain-
storming should be done with a team of 2 or more people. A team with more
people can obtain multiple outputs, hence different tools. The improvements
derive, more or less from the previous presentation, depending on the good level
of effectiveness of the tool. The team presents all ideas for improvements. It is
recommended to have more than one solution, but it isn’t required.
3 Improvement validation process
This phase occurs when the team presents the improvements to the Central
Team (team responsible for the application of Lean at plants, with the objec-
tive of choosing the best ones to apply). The validation process is the approval
and standardization within an organization. A tool is accepted when the pro-
posal shows a better potential than the tool being currently used and will bring
more benefits. This comparison is made using a table with advantages and dis-
advantages of the old tools and the new ones. The Table 1 must fit the following
layout:
Table 1. Table template for comparison of tools
Characteristics New tool Tool plant 1 Tool plant 2

Characteristic 1
Characteristic 2
There are no predefined characteristics, the organization must adapt that

subject according to the nature of the tool. To fulfil the Table 1 above the orga-
nization must mark with “” if the tool has the characteristic or “♣” if not.
Through this analysis it’s possible to verify if the new tool has, in theory, more
potential than the implemented tool.
4 Improvement implementation
After being approved by the Central Team, it’s the right moment to start
standardizing on all plants. It’s a slow process because implementations take
time, so it’s recommended to act as quickly as possible. The output resulting
from the application of this methodology is the implementation of the tools
that were subject to improvements at the shop floor. This is the last step of
information collection and analysis regarding plants from different regions.
3.2 Application of Methodologies
This subchapter will give the reader a better comprehension of this methodology
applied in an organization from the automotive industry. Two topics will be
approached where the methodology can be applied, for example, Stopping Points,
a subject from routing, and the color code from plants.
Every plant has routes and this routes need stopping points, a specific loca-
tion near the workplace where the tuggers’ drivers do the collection or replace-
ment of materials during the route. After getting the information from plants,
through photos of the application of the tools at plants, it was verified that the
tool’s presentation is different in all plants within an organization. Figures 3, 4,
5, and 6 show the differences in stopping points between some plants.
Fig. 3. Stopping point plant 1

The existence of several ways of presenting the stop signs shows the need to
standardize this tool, so every plant can use the same tool.
Another concept that many plants have huge differences between them is
the color code, normally used in floor marking. This tool is important to have
for a well-organized workplace so it can prevent accidents and injuries. Many
organizations use these markings with the intention of enhancing the visual
management of the organization. To obtain the data from the plants interviews
by phone were held, with photos to support what has been told on interviews.
These photos in Figs. 7 and 8 show that there are enormous differences between
the color code from each plant.
Fig. 7. Orange color for dangerous material [plant 1] (at left) and rework [plant 2] (at
right)
Fig. 8. Differences between two plants at aisles markings

When all information was collected, it was observed that there wasn’t a
common color code for all plants, every plant has their own code.
3.3 Results
Now, the improvements for the tools mentioned above will be described, includ-
ing Stop Points and Color Code.
The appearance of Stop Points is a very useful visual tool in routing, they
let the driver of the tugger know which route belongs to what and the number
of the stop (usually are numbered with 1, the nearest stop from warehouse and
so on). The decision was to design an easily visible tool with all the necessary
information, so the driver doesn’t have issues performing the route. The new
tool is shown in the Fig. 9.
Fig. 9. New stop sign developed
In this case, this tool has the name of the route it belongs to, a frame with
the color of the route and the number of the stop, the other aspects are only
aesthetics. For the design of this tool, the criteria established by the methodology
were considered. To compare this tool with the ones already implemented, the
methodology requests a table with the comparison:
Table 2. Comparison of stop points
Characteristics New tool Tool plant 1 Tool plant 2 Tool plant 3 Tool plant 4
Fast interpretation ♥ ♥ ♣ ♣ ♣
Visible ♥ ♥ ♣ ♥ ♥
Low cost ♥ ♣ ♥ ♣ ♥
This tool can be quite financially viable. It is made with paper (A4 format),
with the colors from the respective routes allowing a fast interpretation for the
tugger’s driver (Table 2).
As I said before, the color code is essential to mark workplaces and pedes-
trian access areas. Those marks should be visible and noticeable by operators,
employees and visitors that walk at the shop floor. If a workplace has a tidy
Fig. 10. New color code developed
working environment that stimulates a good performance of an operator. So, a

color code was designed to be adopted by all plants within an organization:
From Fig. 10, the objective is the adoption by all plants of a single stan-
dardized color code in the organization. For the design of the code, the criteria
established by the methodology were considered and was decided not to do the
comparative table. Despite the differences, there is no comparison term for the
color code. The colors present in the code are the minimum required for a plant,
however, if any plant needs more combinations of colors they can define them.
4 Conclusions
Standardization of visual management within an organization, with a lot of
plants, is a complex and long journey process. You need to transform an entire
organization, from tools that are being standardized to the mindset of employees,
who will contact with the new tools. The tools to improve must be requested as
they come across a tool that can be improved.
The benefits of having all visual management standardized are huge: reduc-
tion of wasting time using confuse tools; an increased efficiency of operators, tools
are designed for them; simplifies the work environment and the seek for excel-
lence. Normally the output (improvement tools) manages to convey the basic
principles of visual management, through “eye management”, always prevailing
simplicity, transparency, and clarity.
References
1. Bateman N, Philp L, Warrender H (2016) Visual management and shop floor
teams - development, implementation and use. Int J Prod Res 7543:1–14
2. Belekoukias I, Garzareyes JA, Kumar V (2014) The impact of lean methods and
tools on the operational performance of manufacturing organisations. Int J Prod
Res 52(18):5346–5366
3. Eaidgah Y, Maki AA et al (2016) Visual management, performance management

and continuous improvement: a lean manufacturing approach. Int J Lean Six Sigma
7(2):187–210
4. Emiliani ML, Stec DJ (2005) Leaders lost in transformation. Leadersh Organ Dev
J 26(5):370–387
5. Hines PA, Found P et al (2008) Staying Lean: Thriving, Not Just Surviving. Society
for American Archaeology
6. Machado VC, Leitner U (2010) Lean tools and lean transformation process in
health care. Int J Manag Sc Eng Manag 5(5):383–392
7. Ortiz CA, Park M (2010) Visual Controls: Applying Visual Management to the
Factory. Productivity Press, New York
8. Turner GC, Parry CE (2006) Application of lean visual process management tools.
Prod Plann Control 17(1):77–86
9. Wilson L (2010) How to Implement Lean Manufacturing. McGraw-Hill, New York
10. Womack J, Jones D (2003) Lean Thinking, Banish Waste and Create Wealth in
Your Corporation. Free Press, New York
Resource Optimization Management
Online Fault Detection in Solar Plants Using a Wireless
Radiometer in Unmanned Aerial Vehicles
Carlos Quiterio Gómez Muñoz(B) , Alfredo Peinado Gonzalo, Isaac Segovia Ramirez,
and Fausto Pedro Garcı́a Márquez
Ingenium Research Group, Castilla-La Mancha University, Ciudad Real, Spain

CarlosQuiterio.Gomez@uclm.es
Abstract. A novel Non-Destructive Test (NDT) is presented in this paper. It

employs a radiometric sensor that measures the infrared emissivity of the solar
panel surface embedded in an unmanned aerial vehicle. The measurements pro-
vided by the sensor will determine if the panel is healthy, damaged or dirty. A
thermographic camera has been used to check the temperature variations and val-
idate the results by the sensor. The study shows that the amount of dirt influences
the temperature on the surface and the energy generated. Similarly, faults in pho-
tovoltaic cells influence the temperature of the panel. The NDT system is less
expensive than traditional thermographic sensors or cameras. Early detection of
these problems, together with an optimal maintenance strategy, allows to reduce
costs and increase the competitiveness of this renewable energy source.
Keywords: Infrared sensor · Radiometry · Solar plants · Photovoltaic panels ·

Fault detection and diagnosis
1 Introduction
1.1 Renewable Energies
The whole energy produced in the Earth has been obtained, directly or indirectly, from
the Sun. Even fossil fuels, which come from the decomposition of plants that have
previously needed the sun to perform photosynthesis. These fossil fuels are oil, coal
and natural gas, and they are in limited reserves.
From the industrial revolution, fossil fuels consumption has increased without con-
sidering the environmental effects, such as global warming or climate change. However,
as they are finite resources, alternative forms of fossil fuels must be sought to obtain
energy and to avoid the pollution they produce. These measures are being carried out
globally, led by the Intergovernmental Panel on Climate Change (IPCC).
The renewable energies are being developed and employed to use a clear production,
e.g. in Spain they produce 28.6% of the energy consumed [9, 16]. Renewable energies
are those based on the use of virtually endless resources, e.g. the sun (Fig. 1), wind
or ocean currents, or resources that are renewed periodically if they are used sustain-
ably, such as biomass or biofuels. In general, its production is more expensive than
1162 C.Q.G. Muñoz et al.
Fig. 1. China and Pakistan project for Solar plant Solar Quaid Azam installed in Pakistan (http://
www.qasolar.com/)
fossil fuels, but if does not generate polluting emissions during their production. These
include: Wind energy [14, 22], tidal, solar (Fig. 1), biomass and biofuels, geothermal
and hydraulic.
1.2 Non-Destructive Experimenting

Non-Destructive Experimenting (NDT) is a type of experiment that is performed on
parts and pieces to verify its state and operation without interfering in the integrity
of the material or its usefulness to the service [6, 10, 19, 24]. They are carried out in
recognition and preventive maintenance work in all types of industrial areas. These
experiments are used to detect internal and surface defects [8, 18, 20, 21]. Some of these
experiments are: Radiography with X-ray and Gamma, magnetic particles, Eddy cur-
rents, ultrasound experimenting [7, 11, 17] and infrared thermography [1]. Infrared ther-
mography consists in checking the state of a part or a structure from a thermal image.
They are used in many industries and they can be applied in any production phase.
2 Infrared Thermography
2.1 Radiation
There are only three methods of heat transmission in physical terms: conduction, con-
vection and radiation. In the case of radiation, a transmission medium is not necessary,
Online Fault Detection in Solar Plants 1163
the energy is transmitted due to the temperature of the emitter body. Depending on the
temperature, the radiation emitted changes. The law of Stefan-Boltzmann, Eq. (1), is
used to quantify it. It must be adjusted for real bodies, by adding the emitter parameter
to the equation.
EN (T ) = σ · T 4 , (1)
where EN is the emittance of the body, is the Stefan Boltzmann constant, value 5,6710 B
−8 W/m2 K4 )) and T the body temperature. This emittance is for of a black body, which
must be corrected for real bodies through the emissivity. The Planck law is given by
Eq. (2).
C1
ENλ (λ , T ) = C2 , (2)
λ 5 (e λ T − 1)
where EN is the spectral emitted, C1 and C2 are constants, λ the wavelength and T the
body temperature. Eq. (2) shows that most of the radiation emitted by bodies at low
temperatures is found in the infrared region [15].
Emittance is the proportion of energy that a body can emit relative to the total. In a
black body (maximum emitter) is the unit. In any other body, this magnitude is smaller
and depends on the temperature. In addition, the characteristic wavelength of the bodies
also depends on their temperatures, generating the electromagnetic radiation spectrum.
The thermal behavior of a real body is different from a black body. While the black
body absorbs all the energy that reaches its surface and emits throughout the radiation
region equally, a real body does not absorb all the energy and does not emit in the entire
electromagnetic spectrum.
A real body does not absorb all the radiation, but part of it passes through it or
reflects on its surface. In addition, this behavior also depends on the wavelength being
studied, that depends the concept of emittance, given by Eq. 3. Emittance is the ratio
of energy that a body emits with respect to the total that it can emit with a certain
temperature and wavelength [15].
E θ ,ϕ ,λ (θ , ϕ , λ , T ) I θ ,ϕ ,λ (θ , ϕ , λ , T )
ε θ ,ϕ ,λ (θ , ϕ , λ , T ) = θ ,ϕ ,λ
= θ ,ϕ ,λ
(3)
EN (θ , λ , T ) IE,N (λ , T )
being ε the emissivity, θ and ϕ emission angles, E emittance and I the radiation inten-
sity of the body. For this research work, infrared radiation is the most relevant, since it
is the dominant thermal radiation in this temperature range. Therefore, the camera and
the sensor are designed to collect this type of radiation.
2.2 Applications
The main function in this paper of infrared thermography is the visualization of ele-
ments whose surface temperature variations show their state [2, 3, 12]. Among its appli-
cations include the following:
(1) Electrical and mechanical systems: It is possible to observe the hot spots of the sys-
tems and areas subjected to higher thermal stresses by comparison on the surface
of the system [13].
(2) Structures: Any irregularity in the temperature of the facade can correspond to
different conditions in buildings, such as poor insulation, the emittance of an area
with a greater amount of bricks than concrete, or even an insect nest within the
facade.
(3) Welds: Temperature variations in a pipe are usually due to areas where welds have
been made. If these variations show a very irregular appearance, it is probably due
to a bad union
(4) Surveillance: Control of emissions, traffic, fire prevention, etc.
(5) Medicine: Monitoring of diseases and anomalies in the human body.
The infrared thermography is done by a non-radiometric camera to check the results
obtained with the sensors.
3 Experimental Platform
3.1 Radiometer
They are used to check the temperature and the radiance of the measured targets. Both
sensors are used to contrast the data obtained with each one in the different case studies.
(1) Campbell SI-111 Sensor
This sensor is connected to an Arduino board, which processes the voltage data
and process that information to obtain the temperature in Celsius degrees. The system
is capable of making accurate measurements to the bodies that are within its field of
vision. The SI-111 Precision Infrared Radiometer features a thermistor to measure its
internal temperature, and a thermopile for the target temperature using a germanium
lens.
Using Stefan-Boltzmann’s law of radiation, and the voltages in the thermopile and
the thermistor, the target temperature is calculated [23].
The temperature measured by the sensor is the average of the temperatures of its
view range, receiving between 95 and 98% of the infrared radiation of the view range
and between 2 and 5% from outside. Figure 2 shows the view range of the SI-111 sensor
with half-angle of 22.5o. The view range depends of the distance to the area and the
inclination with respect to this area (see Fig. 2).
This sensor has been chosen because of its high accuracy, reaction rate to changes
in temperature and the large range of temperatures that can operate. Its size and weight
allow a great manageability. Its installation on the unmanned aerial vehicle allows the
global system to performance more efficient and flexible inspections.
3.2 Arduino Platform and Wi-Fi Shield

For this research, Arduino Uno G3 motherboard is used and the Shield Wi-Fi extension,
which grants wireless features to the original board. This device has been chosen due
to its great extension in this type of applications and its low consumption. The Arduino
Fig. 2. SI-111 sensor field of view
Fig. 3. Arduino Uno G3 (a) and Wi-Fi Shield expansion card (b)
Uno R3 (Fig. 3a) is a motherboard that uses the ATmega328 microcontroller and is
powered by either USB cable connected to a computer or external power from a battery.
The Wi-Fi Shield component is an Arduino expansion card. It allows the transmis-
sion of data via wireless connection by Wi-Fi (Fig. 3b).
Data obtained by the sensor is processed and sent via Wi-Fi to the online platform
Thingspeak to monitor the results (Table 1).
3.3 Thermographic Camera
The camera is non radiometric used to check the data obtained with the sensors and to
see the temperature differences of the observed pieces. This camera consists in assign-
ing a determined colour to each level of radiation that it receives. These levels are cal-
culated for each resolution pixel that the camera has, creating a thermographic picture
(Figs. 4 and 5).
Table 1. Technical characteristics of the Campbell SI-111 sensor
Input voltage 2.5V excitation for the thermistor.

Absolute accuracy ± 0.2 ◦ C between [−10 and 60] ◦ C ± 0.5◦ C between [−40 and
70] ◦ C
Response time Less than 1 s for changes in target temperature.
Field of view 22◦ of half-angle
Operating environment Highly water resistant designed for continuous outdoor use.
Operating range −55 ◦ C to 80 ◦ C 0 to 100% relative humidity
Fig. 4. Thermographic camera Flir VUE
Fig. 5. Solar panels used in the experiments
3.4 Photovoltaic Solar Panels
These panels were used in the experiments to check their temperature and voltage at the
terminals. Their technical characteristics are shown in Table 2.
Table 2. Technical characteristics of photovoltaic solar panels
Manufacturer’s name TSM-170D

Max power 2.5 W
Maximum voltage 35.8 V
Maximum intensity 4.76 A
Open circuit voltage 43.6 V
Short circuit current 5.25 A
Maximum system voltage 700 V
Work temperature [−40 ◦ C,90 ◦ C]
Dimensions 1581 × 809 × 40 mm
Cell type Monocrystalline
Number and size of cell 72 pieces, 125 × 125 mm
Type of glass and width 3.2 mm tempered
Weight 15.6 kg
Fig. 6. Representation of the air vehicle taking a measurement with the wireless sensor mounted
on the Gimbal mount
3.5 Unmanned Aerial Vehicle DJI Spreading Wings S900
This unmanned aircraft operated by remote control is designed for reconnaissance,

monitoring and taking of measurements and images, thanks to its control system that
allows a great stability and handling.
The hexacopter can operate around 18 min autonomously, following pre-established
routes. It is made of carbon fiber to lighten weight, which allows a maximum load
capacity of 8.2 kg. The gimbal is a crucial component for this application for increasing
the stability and accuracy of the measurements. This support contains the measurement
equipment (sensor, camera, Arduino card and electronic components) in the drone. A
schematic representation of the drone collecting data in a photovoltaic plant is shown
in Fig. 6.
This system allows to inspect solar plants in a totally automatic and autonomous
way. It can inspect large areas, and a very large park can be inspected by sections. The
system sends online all the information about the state of the plates (dirt and defects),
generating alarms and reports of each module.
4 Case Study
The objective of the experiments was to detect dirt on the solar panels [4]. The experi-
ments were carried out in three different scenarios: In the first scenario, the solar panel
was clean; the second one, mud was added in half solar panel; third scenario, the solar
panel was totally covered by mud. In each experiment, the temperature values collected
by the radiometer and the volts generated by the solar panel were obtained.
The incidence of the sun is crucial in the emissivity values collected by the radiome-
ter. For this reason, they were carried out three cases of study at different hours to study
the incidence of the sun in the results. The first case study was performed at 10:00 am.
The second case study at 12:00 noon. The last case study was carried out with absence
of direct light because of clouds (Figs. 7, 8 and 9).
Fig. 7. Solar panel without dirt on its surface
Fig. 8. Solar panels covered in half of mud

Fig. 9. Solar panels totally covered of dirt
Table 3. Results obtained in the case study 1
Temperature (◦ C) Voltage (V)

Clean panel 46.2 36.2
Partially covered panel 44.7 35.7
Fully covered panel 44.1 35.3
4.1 Case Study 1: Incidence of Sunlight at 10:00 AM
Experiments have been performed at 10:00 am to check the temperature reached by the
panels when, due to the incidence of light, the panel does not reach its maximum power
generation. Table 3 shows the obtained temperature and generated voltage by the panel.
It is observed an evident difference of temperatures between the clean panel and
the panel with dirt on the half of its surface. There are also minor differences between
the half dirty panel and the panel totally covered by mud. Figure 10 shows the different
temperatures obtained in the series of experiments represented by columns. The ambient
temperature series has been added to contrast with those acquired by the panels in each
case shown in the form of a line.
These relationships are proportional to the voltages generated by the solar panel.
Figure 11 shows the influence of the mud in the voltage at 10:00 am [5].
4.2 Case Study 2: Incidence of Sunlight at 12:00 PM
These experiments have been performed to observe the differences in temperature and
energy production of the panels when the panel reach its maximum of production, i.e.
when the light strikes perpendicularly on the panels (Table 4).
The temperature variations are bigger than in the first case study, as well as the
voltage. This is because of the intensity of the light is much greater and it makes more
evident the cases where the panel is covered with dirt. Greater differences in tempera-
ture are observed in Fig. 12, as well as higher temperatures.
A greater voltage difference is shown in Fig. 13. It is due to a greater inequality of
sunlight incident between the panel without dirt and completely covered.
Fig. 10. Comparison between the different panel temperatures in case study 1
Fig. 11. Comparison between different panel temperatures and voltages in case study 1
Table 4. Results obtained in the second experiments

4.3 Case Study 3: Absence of Direct Light (18:00 P.M.)

These experiments have been performed at 18:00 pm with clouds to study the effec-
tiveness of the system for the detection of dirt under these conditions. The temperature
results obtained are shown in Table 5.
Fig. 12. Comparison between different panel temperatures in case 2
Table 5. Measurements obtained in the case study 3

The differences of temperatures are lower, with a large margin of error that does not
allow to know exactly the condition of the panel. Figure 14 shows the low temperature
variations between the different experiments.
The voltage in each case are similar for all case studies. In these conditions, it is
concluded that the system is not effective to detect the dirt in the solar panels.
Fig. 14. Comparison between the different panel temperatures in case study 3
5 Conclusions
Dirt and dust in solar panels is a common problem in photovoltaic plants, which are
usually located in desert areas with lots of sand and dust. A novel non-destructive test
system is proposed in this paper, based on: a radiometer; Arduino; a Wi-Fi Shield, and;
a UAV for inspecting the state of the photovoltaic panels.
The low weight and cost of the system can reduce the costs in operation and main-
tenance of the whole solar plant. The system can be automated, reducing the inspection
time and the costs (Fig. 15).
Three different conditions of the solar panel were analysed: Clean, half covered by
mud and totally covered. The infrared radiometry of the panels was read and the voltage
of the panels. The experiments were performed in three different solar conditions to
increase the actuary of the experiments: at 10 am, 12:00 pm and with clouds and without
direct sunlight. The absence of direct light does not allow identification among the three
cases.
Acknowledgements. The work reported herewith has been financially by the Spanish Ministerio
de Economı́ay Competitividad, under Research Grant Ref.: RTC-2016-5694-3.
References
1. Acciani G, Simione G, Vergura S (2010) Thermographic analysis of photovoltaic panels. In:
International conference on renewable energies and power quality (ICREPQ10). Granada,
Spain, March, pp 23–25
2. Ancuta F, Cepisca C (2011) Fault analysis possibilities for pv panels. In: International youth
conference on energetics, pp 1–5
3. Bazilian MD, Kamalanathan H, Prasad DK (2002) Thermographic analysis of a building
integrated photovoltaic system. Renew Energ 26(3):449–461
4. Dorobantu L, Popescu M et al (2011) The effect of surface impurities on photovoltaic panels.
In: International conference on renewable energy and power quality
5. Fares Z, Becherif M et al (2013) Infrared thermography study of the temperature effect on
the performance of photovoltaic cells and panels. In: Sustainability in energy and buildings,
Springer, pp 875–886
6. Garcı́a Márquez FP, Muñoz G et al (2014) Structural health monitoring for concentrated
solar plants. In: 11th International conference on condition monitoring and machinery failure
prevention technologie. Manchester, UK
7. Gómez CQ, Villegas MA et al (2015) Big data and web intelligence for condition monitoring:
a case study on wind turbines. In: Handbook of research on trends and future directions in
big data and web intelligence; information science reference. Hershey
8. Gómez Muñoz CQ, Garcı́a Márquez FP (2016) A new fault location approach for acoustic
emission techniques in wind turbines. Energies 9(1):40
9. Heras-Saizarbitoria I, Cilleruelo E, Zamanillo I (2011) Public acceptance of renewables
and the media: an analysis of the spanish pv solar experience. Renew Sustain Energ Rev
15(9):4685–4696
10. Ruiz de la Hermosa González-Carrato R, Garcı́a Márquez FP et al (2015) Acoustic emission
and signal processing for fault detection and location in composite materials. In: Global
cleaner production & sustainable consumption conference. Elsevier
11. Jiménez AA, Muñoz CQG et al (2017) Artificial intelligence for concentrated solar plant
maintenance management. Springer, Singapore
12. Maldague XP (2002) Thermographic inspection of cracked solar cells. In: Aerosense, p 185
13. Maldague XPV (2002) Introduction to ndt by active infrared thermography. Mater Eval 60(9)
14. Márquez FPG, Tobias AM et al (2012) Condition monitoring of wind turbines: techniques
and methods. Renew Energ 46(5):169–178
15. McAdams WH (1958) Heat transmission. McGraw-Hill
16. Montoya FG, Aguilera MJ, Manzano-Agugliaro F (2014) Renewable energy production in
spain: a review. Renew Sustain Energ Rev 33(2):509–531
17. Munoz CQG, Arenas JRT, Marquez FPG (2014) A novel approach to fault detection and
diagnosis on wind turbines. Glob Int J 16(6):1029–1037
18. Muñoz CQG, Marquez FPG et al (2015) A new condition monitoring approach for mainte-
nance management in concentrate solar plants. Springer, Heidelberg
19. Muñoz CQG, Márquez FPG, Tomás JMS (2016) Ice detection using thermal infrared radiom-
etry on wind turbine blades. Measurement 93:157–163
20. Muñoz G, Quiterio C et al (2015) Energy environment maintenance management. In: Gfeecc
2015 Gfeecc International conference forenergy, environment and commercial civilization.
Chengdu, China
21. Papaelias M, Cheng L et al (2016) Inspection and structural health monitoring techniques
for concentrated solar power plants. Renew Ener 85:1178–1191
22. Pliego Marugán A, Garcı́a Márquez FP, Pinar Pérez JM (2016) Optimal maintenance man-
agement of offshore wind farms. Energies 9(1):46
23. Quinn TJ, Martin JE (1985) A radiometric determination of the stefan-boltzmann constant
and thermodynamic temperatures between -40 degrees c and +100 degrees c. Philos Trans R
Soc B Biol Sci 316(1536):85–189
24. Ramirez IS, Muñoz CQG, Marquez FPG (2017) A condition monitoring system for blades
of wind turbine maintenance management. Springer
Demand Response Mechanism of a Hybrid
Energy Trading Market for Residential
Consumers with Distributed Generators
Nastaran Naseri, Babak Rezaee(B) , and Shahin Kazemzadeh
Industrial Engineering Department, Faculty of Engineering,

Ferdowsi University of Mashhad, PO Box: 91775-1111, Mashhad, Iran
brezaee@um.ac.ir
Abstract. Smart grid (SG) as a complex advanced electricity system is

capable of facing with the growing energy demand, where energy con-
sumers are connected not only to the conventional grid but also to some
local energy markets for bidirectional energy trading. In this context, this
paper presents an optimization problem the demand response problem
where a hybrid energy trading system consisting of a conventional grid, a
local trading system, a number of energy consumers, having distributed
generators and battery, and some generator companies. In this case, it
has been considered that the local trading center (LTC) is a non-profit
oriented LTC which aims at benefiting the energy consumers and energy
sellers. The optimization model is formulated as a mixed integer Linear
Program (MILP) with a bounded number of variables and constraints.
Furthermore, the solution can be obtained in polynomial time and pro-
vides the optimal scheduling for the problem such that the costs were
minimized while the total demand is satisfied.
Keywords: Hybrid energy trading market · Demand response mecha-

nism · Local trading center · Distributed generators · Battery bank
1 Introduction
Smart grid (SG) as a complex advanced electricity system is capable of fac-
ing with the growing energy demand in a reliable sustainable and economical
manner [3]. Advanced two-way communication infrastructure of smart grids and
efficient demand response mechanism (DR), according to which the energy con-
sumers and energy sellers can schedule their energy consumption and energy
supply, respectively, call attention to SG for achieving better performance than
the conventional grid. Exclusively, bidirectional energy trading has been pro-
vided by expanded utilization of advanced smart metering systems in future SG
and deployment of distributed energy sources. Accordingly, a precise design of
control mechanism for both economical optimization and energy scheduling for
energy consumers and sellers has been needed.
1176 N. Naseri et al.
Nowadays, photovoltaic (PV) generation has been considered as the most

promising source of renewable energy. Moreover, the wide exploitation of PV
storage facilities integrated into the utility power grid has been provided by the
development of energy storage. On the top of that, the batteries can be used in
order to shift peak PV generation to be in phase with peak energy demand [8].
There have been several related researches examining the energy trading
mechanism and the related DR mechanism. These researches have been catego-
rized into two groups. The first category includes those works, mainly, investi-
gating optimal energy scheduling for energy consumers in response to the pricing
of the retail market. Particularly, Mohsenian-Rad and Leon-Garcia [5], proposed
an optimal and automatic residential energy consumption scheduling framework
to achieve the desired trade-off between minimizing the electricity payment and
minimizing the waiting time for the operation of each appliance in a household
with considering a real-time pricing tariff combined with inclining block rates.
They focused on a scenario which real time pricing is combined with inclining
block rates to get a more balanced residential load with a low peak to average
ratio. Chen et al. [2], formulated a problem to minimize the cost of energy drawn
from the external grid while usage of appliances was subjected to individual delay
constraints and a long-term average delay constraint. They supposed that the
end user has an energy storage battery as well as an energy harvesting device lest
harvested renewable energy can be stored and later used when the price is high.
Kim et al. [4], formulated the energy scheduling problem as a non-cooperative
game among self-interested customers, where each customer determines its own
load scheduling and energy trading to maximize its own profit. They proposed a
tiered billing scheme that can control the electricity rates for customers accord-
ing to their different energy consumption levels in order to resolve the unfairness
between heavy and light customers in the non-collaborative approach. Vaziri
et al. [9], developed bi-objective integer linear programming model in order to
minimize the energy costs with considering the revenue, renewable energy subsi-
dies, and overtime costs, and minimize the displeasure of surgeons and patients
in a hospital as a non-isolated micro grid. Ren et al. [7], proposed an optimiza-
tion model for smart homes. The model finds out the optimal running schedule
with considering annual running cost or annual CO2 emissions as the objective
function to be minimized. Besides the energy flows among the equipment within
the hybrid energy system, the economic information including electricity tariff
and natural gas price, as well as some policy issues (e.g., buy-back price) are
also accounted.
The second category subtends researches which investigated the pricing
strategies for energy retailers as well as optimizing certain objectives, based
on price-dependent energy scheduling. Qian et al. [6], proposed a real-time pric-
ing scheme in order to reduce the peak-to-average load ratio through demand
response management. The proposed scheme solved a two-stage optimization
problem. Then, they proposed an algorithm based on simulated annealing to
solve the non-convex price optimization problem. Chai et al. [1] studied DR
mechanism with considering multiple utility companies and multiple residen-
Demand Response Mechanism of a Hybrid Energy Trading Market 1177
tial users and a two-level game was proposed to model the interaction between
these two levels. The competition among the utility companies was formulated
as a non-cooperative game, while the interaction among the residential users
was formulated as an evolutionary game. Then, they proved that the proposed
strategies were able to make both games converge to their own equilibrium. Wu
et al. [10] focused on a hybrid energy trading market consisting of an external
utility company and a local trading market managed by a local trading center.
First, they quantified the respective benefits of the energy consumers and the
sellers from the local trading and then investigated how they can optimize their
benefits by controlling their energy scheduling in response to the LTC’s pricing.
The main contributions of this paper are summarized as follows. This paper
first models a hybrid trading market that is comprised of multiple generator
companies, multiple energy consumers, external power grid, and a local trad-
ing center. In the consumers’ side, some distributed generators are considered.
The consumers’ appliances are categorized into three groups according to their
features. The battery as a storage facility is considered for consumers so as to
reduce the average peak load ratio.
The paper is organized as follows. Sect. 2 proposes the overview of the system
and presents the mathematical model. Sect. 3 shows the numerical example and
the results. Finally, the paper is concluded in Sect. 4.
In this paper, a real-time scheme has been considered for a hybrid trading market
in order to reduce the peak to average load ratio and maximize each user’s
objectives. A hybrid trading market consists of the number of Energy Consumers,
Energy Sellers or Generator Companies (GenCo), a Local Trading Center (LTC)
and Conventional Power Grid. Figure 1 shows a simplified illustration of the
hybrid energy system.
Every energy consumer and energy seller is connected to both local trading
center and the conventional power grid. Energy consumers are residential con-
sumers which have the different type of electrical appliances, these appliances
Fig. 1. A simplified illustration of the hybrid energy system

Fig. 2. An example of appliances type B
Fig. 3. An example of appliances type C
can be categorized into three main groups. The first category, A, named as non-
flexible appliances, includes background appliances which are needed to be used
at specific times and cannot be shifted; refrigerator, lightening appliances are
some examples of this group.
The second category, B, named as semi-flexible appliances, consists of appli-
ances which the time of using of these appliances can be shifted in a specific
period of time. Figure 2 shows an example of these type of appliances. α and β
are the start and end of the preferred time interval, respectively, for an appliance
type B.
The third category of appliances, C, which is called as flexible appliances
includes those appliances which their time of use can be shifted during a specific
period of time as well as be disarticulated in the definite number of sub-activities.
Figure 3 shows an example of appliance type C.
It is assumed that a number of energy consumers have Distributed Generators
(DG). These DGs consist of a Photovoltaic system with a definite capacity and
load availability, and a battery with a specific capacity. The battery is used to
store energy surpluses during off- peak time, therefore, it can be charged via
PV, LTC, and the grid; and it would be discharged only to supply power to load
during peak demand hours.
In the GenCos’ side, it is assumed that in every time slot h, there is a certain
and constant amount of load available.
The objective of this problem is to minimize the cost of energy for consumers
and maximize the income for GenCos. To implement the model, a day is divided
into 24 time slots. To model the problem, the following notations are defined:
Indices:
i : Number of energy consumers i = 1, · · · , I;

j : Number of generator companies j = 1, · · · , J;
k : Number of sub-activities k = 1, · · · , KaC ;
h : Number of time slots h = 1, · · · , H;
aA : Appliances of type A aA = 1, · · · , m;
aB : Appliances of type B aB = 1, · · · , n;
aC : Appliances of type C aC = 1, · · · , r.
Parameters:
[αaB , βaB ] : Start and finish time of appliance aB ;
[λaC ,k , γaC ,k ] : Start and finish time of subactivity k for appliance aC ;
laB : Activity time of appliance aB ;
la C ,k : Activity time of sub-activity k for appliance aC ;
T LaC ,k : Upper limit of the gap between two sub-activity of
appliance aC ;
T LaC ,k : Lower limit of the gap between two sub-activity of
appliance aC ;
di,h : Total energy demand of consumer i at time h;
P Vi,h : Available load from PV system of customer i at time h;
EaA : Total energy needed for appliances type A;
ε : Self-discharge rate of the battery;
η discharge : Discharging efficiency of the battery;
η charge : Charging efficiency of the battery;
B
Ecapacity : Rated capacity of the battery (kWH);
μ : Minimum residue coefficient of the battery
BaB ,h : Energy consumption of appliance aB at time h;
M : A large number;
dj,h : Available load from GenCo j at time h;
CaC ,k,h : Activity k for appliance aC at time h;
ph : Energy sell-out price for the grid at time h ($ /kW);
Wi,h (.) , Wj,h (.) : A linear function for energy transmission loss ;
qh : Buy-back price for the grid at time h ($ /kW);
ph : Energy sell-out price for LTC at time h. ($ /kW);
qh : Buy-back price for LTC at time h ($ /kW)
uh : Buy-back price for the grid from consumers at time
h ($ /kW);
uh : Buy-back price for LTC from consumers at time h ($ /kW)
Decision variables:

1, if appliance aB starts at time h,
STaB ,h =
0, otherwise;
1, if appliance aB finishes at time h
ETaB ,h =
0, otherwise;
1, if sub-activity k of appliance aC starts at time h,
STaC ,k,h =
0, otherwise;
1, if sub-activity k of appliance aC finishes at time h,
ETaC ,k,h =
0, otherwise;
1, if the status of appliance aB at time h is on
SaB ,h =
0, otherwise;

1, if the status of sub-activity k of appliance aC at time h is on
SaC ,k,h =
0, otherwise;
grid
Ei,h : Total power purchased from the grid for customer i at time h
(kW);
grid
Ei,h,sto : Power purchased from the grid for battery storage of customer i
at time h (kW);
grid
Ei,h,self : Power purchased from the grid for self-use of customer i at time
h (kW);
LTC
Ei,h : Total power bought from LTC for customer i at time h (kW);
LTC
Ei,h,self : Power purchased from LTC for self-use of customer i at time h
(kW);
LTC
Ei,h,sto : Power purchased from LTC for battery storage of customer i at
time h (kW);
PV
Ei,h : Total power of PV unit of customer i at time h (kW);
PV
Ei,h,self : Power out of PV unit of customer i at time h for self-use (kW);
PV
Ei,h,sto : Power out of PV unit of customer i at time h for battery
charging (kW);
PV
Ei,h,salegrid : Surplus electricity from PV unit of customer i at time h sell to
the grid (kW);
PV
Ei,h,saleLTC : Surplus electricity from PV of customer i at time h sell to LTC
(kW);
B
Ei,h : Available power stored in the battery of customer i at time h
(kW);
B
Ei,h,self : Power discharged from the battery of customer i at time h for
self-use (kW);
B
Ei,h,outLTC : Power discharged from the battery of customer i at time h sell to
LTC (kW);
B
Ei,h,outgrid : Power discharged from the battery of customer i at time h to the
grid (kW);
fi,h grid : Binary variable indicating the energy purchase from the grid for
customer i
fi,h B : Binary variable indicating the charging state of the battery of
customer i
fi,h LTC : Binary variable indicating the energy purchase from the LTC for
customer i
fi,h grid,sto : storage for battery from the grid for customer i
fi,h LTC,sto : Binary variable indicating the energy storage for battery from
the LTC for customer i
fi,h P V,sto : Binary variable indicating the energy storage for battery from
the PV for customer i
Xj,h : Total energy sold to the grid from GenCo j at time h (kW);
Yj,h : Total energy sold to LTC from GenCo j at time h (kW)
The mathematical model for the aforementioned problem is proposed in the
coming 24 h as follows. The objective function is minimizing the cost of buying
Energy for customers and maximizing the income for energy consumers and
sellers, respectively. These two objectives have been formulated in the following
expression.
grid
min Z = Ei,h × ph + LT C
(Ei,h − Wi,h (Ei,h
LTC
)) × p h
h i h i

− PV
Ei,h,salegrid × uh + PV
Ei,h,saleLTC × u h (1)
h i h i

− Xj,h × qh + (Yj,h + Wj,h (Yj,h )) × q h
h j h j
βaB −laB

s.t. STaB ,h = 1; ∀aB , ∀h ∈ H (2)
h≥αaB
βaB

ETaB ,h = 1; ∀aB , ∀h ∈ H (3)
h≥αaB +laB
γaC ,k −l aC ,k

ST aC ,k,h = 1; ∀aC , k, ∀h ∈ H (4)
h≥λaC ,k
γaC ,k

ET aC ,k,h = 1; ∀aC , k, ∀h ∈ H (5)
h≥λaC ,k +l aC ,k
STaB ,h = ETaB ,h+laB ; ∀aB , ∀h ∈ H; αaB ≤ h ≤ βaB − laB (6)

STa C ,k,h = ETa C ,k,h+l a ,k ; ∀aC , k, ∀h ∈ H; λaC ,k ≤ h ≤ γaC ,k − la C ,k (7)
C
SaB ,h = SaB ,h−1 + STaB ,h − ETaB ,h ; ∀aB , ∀h ∈ H (8)

Sa C ,k,h = Sa C ,k,h−1 + STa C ,k,h − ETa C ,k,h ; ∀aC , k, ∀h ∈ H (9)

h
T LaC ,k ≤ h × STa C ,k,h − h × ET aC ,k−1,h ≤ T LaC ,k ;
h =1
(10)
∀aC , k, k ≥ 2, ∀h ∈ H
PV PV PV PV
Ei,h,self + Ei,h,salegrid + Ei,h,saleLTC + Ei,h,sto = P Vi,h ;
(11)
∀i ∈ ΩC , ∀h ∈ H
grid
(1 − ε) × Ei,h
B
+ (Ei,h,sto LTC
+ Ei,h,sto PV
+ Ei,h,sto ) × η charge
B
(Ei,h,outgrid B
+ Ei,h,outLTC B
+ Ei,h,self ) (12)
− discharge
B
= Ei,h+1 , ∀i ∈ ΩC , ∀h ∈ H
η
μ × Ecapacity
B
≤ Ei,h
B
≤ Ecapacity
B
; ∀i ∈ ΩC , ∀h ∈ H (13)
grid
Ei,h,sto LT C
+ Ei,h,sto PV
+ Ei,h,sto ≤ M × fiB ; ∀i ∈ ΩC , ∀h ∈ H (14)
B
Ei,h,outgrid B
+ Ei,h,outLTC PV
+ Ei,h,sto ≤ M × (1 − fi B ); ∀i ∈ ΩC , ∀h ∈ H
(15)
grid
Ei,h,sto ≤M× figrid,sto ; ∀i ∈ ΩC , ∀h ∈ H (16)
LTC
Ei,h,sto ≤ M × fiLTC,sto ; ∀i ∈ ΩC , ∀h ∈ H (17)
PV
Ei,h,sto ≤ M × fiP V,sto ; ∀i ∈ ΩC , ∀h ∈ H (18)
figrid,sto + fiLTC,sto + fiP V,sto = 1; ∀i ∈ ΩC (19)
grid grid
Ei,h,self + Ei,h,sto ≤ M × fi grid ; ∀i ∈ ΩC , ∀h ∈ H (20)
PV
Ei,h,salegrid B
+ Ei,h,outgrid ≤ M × (1 − fi grid ); ∀i ∈ ΩC , ∀h ∈ H (21)
LTC
Ei,h,self LTC
+ Ei,h,sto ≤M× fi LTC
; ∀i ∈ ΩC , ∀h ∈ H (22)
PV B
Ei,h,saleLTC + Ei,h,outLTC ≤ M × (1 − fi ); ∀i ∈ ΩC , ∀h ∈ H
LTC
(23)

CaC ,k,h × S aC ,k,h + grid
BaB ,h × SaB ,h +EaA + Ei,h,sto LTC
+ Ei,h,sto
aC k aB (24)
PV
+ Ei,h,sto = di,h , ∀i ∈ ΩC , ∀h ∈ H
grid
Ei,h,self PV
+ Ei,h,self B
+ Ei,h,self LTC
+ (Ei,h,self − Wi,h (Ei,h,self
LTC
)) = di,h ;
(25)
∀i ∈ ΩC , ∀h ∈ H
Xj,h + (Yj,h + Wj,h (Yj,h )) ≤ dj,h ; ∀j ∈ ΩG , ∀h ∈ H (26)
B grid LTC P V,sto
STaB ,h , ETaB ,h , ST aC ,k,h , ET aC ,k,h , SaB ,h , S aC ,k,h , fi,h , fi,h , fi,h , fi,h ,
fi,h grid,sto , fi,h LTC,sto ∈ {0, 1}
(27)
grid grid grid LTC LTC LTC PV PV PV PV
Ei,h , Ei,h,sto , Ei,h,self , Ei,h , Ei,h,sto , Ei,h,self , Ei,h , Ei,h,sto , Ei,h,self , Ei,h,salegrid ,
PV
Ei,h,saleLTC B
, Ei,h B
, Ei,h,self B
, Ei,h,outgrid B
, Ei,h,outLTC , Xj,h , Yj,h ≥ 0.
(28)
Constraints (2)–(5) represent the start and finish time of each appliance. Con-
tinuity of appliances’ activities have been avouched by constraints (6) and (7).
The situation of appliances’ activities (on or off) have been showed in Eqs. (8)
and (9). Constraint (10) represents the limitation for the gap between each sub-
activity for appliances type C; it means that the finish time of sub-activity k − 1
and start time of sub-activity k must be in a definite limit. Constraint (11)
shows that the total load available from PV system is consumed for self-use,
storing in the battery, selling to the grid and LTC. Constraint (12) represents
available load stored in the battery at time h + 1 during charge and discharge
process. It remarks that the stored load at time h + 1 is equal to initial load at
time h with considering the self-discharge percentage, plus incoming load from
PV system, the grid, and LTC deducting the amount of load needed for the
customers’ demand. As the full discharge of the battery reduces its lifetime con-
straint (13) confines the state of charge and discharge of the battery. Constraints
(14) and (15) prevent the battery from charging and discharging simultaneously.
Constraints (16)–(19) prevent the process of charging from three sources (grid,
LTC, and PV) simultaneously for the battery. Constraints (20)–(23) illustrate
the limit of simultaneous buying and selling electricity load process for both grid
and LTC. Equation (24) confines that the total energy demand for an energy
consumer is coming from the total energy needed for appliances type A, B, and
C, and the amount of energy needed for charging the battery. Equation (25)
expresses that the energy demand can be met through the energy purchased
from the grid or LTC, or the energy supplied through PV or battery; taking
into consideration the amount of energy transmission loss from the LTC. The
last expression (26) represents the constraint for the generators’ side. Total load
available from every GenCo at time can be sold to the grid or LTC. Transmission
loss has been considered in the process of selling electricity to the LTC.
3 Numerical Example
In order to show the performance of the hybrid energy system consisting of the
conventional grid, LTC, energy consumers with different types of appliances and
GenCos, a numerical example is presented. Based on the structure of the hybrid
energy system illustrated in Fig. 1, the numerical example is utilized to verify
the optimization model. Figure 4 shows the energy demand for appliances of
type B and the cumulative energy demand. It has been considered there are four
appliances of type B.
Fig. 4. Energy consumption of appliances type B and the total demand
Figure 5 shows the Energy sell-out and buy-back prices for both grid and
LTC. In this example, it is assumed that the energy sell-out price for LTC is
less than the energy sell-out price for the grid. Besides, the energy buy-back
price for LTC is greater than energy buy-back price for the grid. The numerical
example is carried out by using IBM ILOG CPLEX Optimizer v12.3 on the PC
with Intel(R) Core(TM) i7-4770 CPU@3.4 GHz and 8 GB RAM. Tables 1 and
2 represent the results. The consumer uses the PV energy and provides more
energy with buying energy from LTC in order to meet energy demand in peak
time.
Fig. 5. Energy sell-out and buy-back prices for grid and LTC
Table 1. Energy buy-back schedule for the consumer
Time slots
1 2 3 4 5 6 7 8 9 10 11 12
LTC
Ei,h,self 0.58 0.58 0.58 0.55 0.00 0.00 0.00 0.25 2.11 3.34 4.34 3.34
PV
Ei,h,self 0.00 0.00 0.00 0.99 2.72 3.66 4.46 4.41 2.16 0.00 0.00 0.00
sum 0.58 0.58 0.58 1.54 2.72 3.66 4.46 4.66 4.26 3.34 4.34 3.34
Table 2. Energy buy-back schedule for the GenCo
Time slots
1 2 3 4 5 6 7 8 9 10 11 12
Yj,h 0.13 0.13 0.13 0.51 0.82 0.99 0.79 0.95 0.88 0.54 0.99 0/89
4 Conclusion
Based on MILP theory, this paper formulated an optimization model to inves-
tigate the optimal operation strategy of the PV, battery, conventional grid, and
LTC based residential hybrid energy trading system. Besides satisfying the resi-
dential electricity demands, total costs were minimized; moreover, the optimiza-
tion scheme maximized the profit of GenCos. The LTC provides new opportuni-
ties for the energy consumers and GenCos to perform the local energy trading in
a cooperative manner, as a result, they all can benefit. In this case, it has been
considered that the LTC is a non-profit oriented LTC which aims at benefiting
the energy consumers and energy sellers.
Numerical results are presented in order to validate the benefits of the con-
sidered hybrid energy market and the related DR mechanism.
References
1. Chai B, Chen J et al (2014) Demand response management with multiple utility
companies: a two-level game approach. IEEE Trans Smart Grid 5(2):722–731
2. Chen S, Shroff NB, Sinha P (2013) Heterogeneous delay tolerant task scheduling
and energy management in the smart grid with renewable energy. IEEE J Sel Areas
Commun 31(7):1258–1267
3. Keshav S, Rosenberg C (2010) How internet concepts and technologies can help
green and smarten the electrical grid. In: ACM SIGCOMM workshop on green
NETWORKING 2010. New Delhi, India, pp 35–40
4. Kim BG, Ren S et al (2013) Bidirectional energy trading and residential load
scheduling with electric vehicles in the smart grid. Sel Areas Commun IEEE J
31(7):1219–1234
5. Mohsenian-Rad AH, Leon-Garcia A (2010) Optimal residential load control with
price prediction in real-time electricity pricing environments. IEEE Trans Smart
Grid 1(2):120–133
6. Qian LP, Zhang YJA et al (2013) Demand response management via real-time
electricity price control in smart grids. IEEE J Sel Areas Commun 31(7):1268–
1280
7. Ren H, Wu Q et al (2016) Optimal operation of a grid-connected hybrid pv/fuel
cell/battery energy system for residential applications. Energy 113:702–712
8. Torres D, Crichigno J et al (2014) Scheduling coupled photovoltaic, battery and
conventional energy sources to maximize profit using linear programming. Renew
Energ 72(4):284–290
9. Vaziri SM, Rezaee B, Monirian MA (2017) Bi-Objective integer programming of
hospitals under dynamic electricity price. Springer, Singapore
10. Wu Y, Tan X et al (2015) Optimal pricing and energy scheduling for hybrid energy
trading market in future smart grid. IEEE Trans Ind Inf 11(6):1585–1596
A Fuzzy Multi-Criteria Evaluation Method
of Water Resource Security Based
on Pressure-Status-Response Structure
Talat Qadeer and Zongmin Li(B)

lizongmin@scu.edu.cn
Abstract. In many circumstances, water resources advancement and

usage has come to or surpassed the global warning level. With expand-
ing water request and shortage, utilities require compelling techniques
for ideal utilization of accessible water resources. This paper proposes
a fuzzy multi-criteria evaluation method for supporting the evaluation
of water resources security. Firstly, this paper establishes a novel water
resource security evaluation index system based on the Pressure-Status-
Response (PSR) structure. Then, a revised TOPSIS method was utilized
to handle fuzzy ratings and integrate the ranking outcomes. After that,
data for 9 provinces from Yellow River basin are collected and the pro-
posed methodology is applied to test the feasibility.
Keywords: Multi-criteria decision making · Water resource security ·

Fuzzy evaluation · Pressure-status-response · TOPSIS
1 Introduction
Water is a fundamental resource and indispensable material basis for energy,

production, human society, and economic development [8,13,19]. The pivotal
role of water is beyond doubt. However, when enjoying the benefits of water,
the water conflicts between human and ecosystems are becoming increasingly
acute. Despite having suffered the same dilemma many times throughout history
and the rapidly developed technology, sharing water resources is not something
societies tend to do well [13]. The long existing water conflicts, rapid population
increases, climate change, water shortage and pollution have made water resource
security an indispensable component of sustainable water resource management
[20]. People in developing countries are particularly at risk because of severe
water pollution, environmental damages, poor water supply conditions and even
social conflicts caused by water problems.
Water resource security is a concept that was proposed in the late 20th
century [6,7,21]. Nowadays, water security issue is becoming more and more

A Fuzzy Multi-Criteria Evaluation Method of Water Resource Security 1187
prominent and has attracted a worldwide attention and emphasis [3]. The eval-
uation and insurance of water security are the core issues of sustainable water
resources management. There are increasing studies about water resource secu-
rity evaluation. For example Jiang [6] studied on water resource safety strat-
egy for China in the 21st century. Bitterman et al. [2] proposed a conceptual
framework and candidate indicators for water security and rainwater harvest-
ing. Chen [4] studied on water resources security concept and its discussion. Xia
and Zhang [21] worked on water security in north China and countermeasure to
climate change and human activity. Hall and Borgomeo [9] worked on risk-based
principles for defining and managing water security. Norman et al. [14] worked
on water security assessment: Integrating governance and freshwater indicators.
Qian and Xia [5] worked on risk assessment of water security in Haihe River
Basin during drought periods based on D-S evidence theory. However, a com-
prehensive evaluation of water resources security is such a complex, vague and
multi-level evaluation process. Multi-criteria evaluation method of water secu-
rity is still worthy depth research. In addition, due to the increasingly severe
climate change, water scarcity and pollution, the problem of water resources
security has becoming more and more important. Therefore, this paper presents
a fuzzy multi-criteria evaluation method, which include indicator system under
Pressure-Status-Response framework, uncertainty rating analysis, and aggrega-
tion method of TOPSIS with fuzzy judgments.
2 Evaluation Index System Establishment

Water resource security evaluation is multi-criteria in nature. This paper estab-
lishes the evaluation index system for water resources security evaluation based
on the Pressure-Status-Response (PSR) approach, which is a causal one that
covers causes and effects influencing a measurable state. In this sense, three
categories of indicators are distinguished [15]:
• Indicators of environmental Pressure describe pressures on the environment
originating from human activities, including quality and quantity of natural
resources (e.g., emissions, mining of raw materials, fertilizer input).
• Indicators of environmental State are designed to describe the status quo of
the environment and the quality and quantity of resources and their changes
over time (e.g., forest area, protected areas).
• Indicators of societal Response show to which degree society is responding
to environmental changes and concerns. This could be the number and kind
of measures taken, the efforts of implementing or the effectiveness of those
measures. Responses may range from public (e.g., legislation, taxation, pro-
motion) to private sector activities (e.g., reduced consumption, recycling)
[12].
The originally PSR approach is in this manner changed by presenting the
monetary, biological and social perspectives of water resources security into
the PSR model created on the established of key variables. The model still
1188 T. Qadeer and Z. Li
Pressure State Response
Water function and benefits
Current water
Human-induced resource Action to reduce
influences management influences
unit key figures
impacts information
resources measures
Actions and measures reducing influences
Fig. 1. Adapted pressure, state and response framework for evaluating security on the
water resource management unit level
expands on the outstanding PSR structure (Fig. 1) including three boxes for
Pressure, State and Response objects with the cases practically connected.
Human-initiated Pressures incite impacts on State limitations by making utili-
tarian and monetary advantages additionally influencing the (societal) Response
factors. The other way around Pressures utilize and deplete resources both sub-
jectively and quantitatively (State) and undertaking information, choices and
activities from Responses to control them. State and Response are connected
by figures and material interchange on the one side and directly state-changing
procedures on the other. Inside the three boxes, the structure is changed. Asso-
ciations and impacts between variables are involved into evaluation by making
system subsystems inside every case (Fig. 1) each pointer is appointed to each
of the cases however conveying a different idea for the evaluation. Really, the
examples of connotation will change among the groups since the pointers cover
dissimilar PSR viewpoints for every indicator. This adjustment has been picked
because a strict arrangement of indicator into boxes would mean loss of fig-
ures about the interconnections amongst indicator and their pertinence as per
pressure, states and responses [18]. The interconnections between pointers are
indicated by arrows in (Fig. 1).
Based on the PSR structure, we consider five criteria for water resource
security, which are: water resources pressure (C1 ), social-economic pressure
(C2 ), water resources state (C3 ), social-economic state (C4 ) and social-economic
response (C5 ).
3 Fuzzy Statement
The subjective evaluation part of water resource security evaluation, inevitably

contains the evaluators’ subjective judgments and preferences. This situation has
further amplified the uncertainty of assessments inherent in the water resources
security evaluation process. Evaluators cannot apply precise numbers to describe
Table 1. Linguistic variables and triangular IFNs for rating under the subjective eval-
uation criteria
Linguistic variables Triangular IFNs

Very good (VG) (0.9, 1.0, 1.0)
Good (G) (0.7, 0.9, 1.0)
Medium good (MG) (0.5, 0.7, 0.9)
Fair (F) (0.3, 0.5, 0.7)
Medium poor (MP) (0.1, 0.3, 0.5)
Poor (P) (0.0, 0.1, 0.3)
Very poor (VP) (0.0, 0.0, 0.1)
Table 2. Linguistic variables and triangular IFNs for rating the importance
Linguistic variables Triangular IFNs

Very high (VH) (0.7, 0.9, 1.0)
High (H) (0.6, 0.7, 0.8)
Medium high (MH) (0.4, 0.5, 0.6)
Medium (M) (0.1, 0.3, 0.5)
Medium low (ML) (0.1, 0.2, 0.3)
Low (L) (0.0, 0.1, 0.2)
Very low (VL) (0.0, 0.0, 0.2)
their assessments, however, they can utilize linguistic variables according to their
professional knowledge and experience. Hence, the concept of fuzzy numbers can
be integrated into the multi-criteria evaluation of water resources security.
Evaluators first make their own judgments of water resource security based
on subjective evaluation criteria C1 –C5 . Ratings under subjective evaluation
criteria are considered as linguistic variables. A linguistic variable is a variable
whose value is a natural language phrase. It is very useful in dealing with situa-
tions which are ill-defined to be described properly in conventional quantitative
expressions. Water resource security performance under each subjective evalu-
ation criteria can be expressed on a 7-point rating scale: “very good”,“good”,
“medium good”, “fair”, “medium poor”, “poor”, and “very poor”. Such linguis-
tic variables are converted into triangular intuitionistic fuzzy numbers (IFNs) [1]
as shown in Table 1. The linguistic variables and triangular IFNs for rating the
importance are shown in Table 2. Figures for membership functions are shown
in Figs. 2 and 3. IFNs are commonly used for solving decision-making problems,
where the available information is imprecise. There are different shapes or forms
of IFNs, among those, trapezoidal IFNs and triangular IFNs are the most com-
monly used. For example, Shaw and Roy [17] used trapezoidal IFNs for analysing
fuzzy system reliability, while Vahdani et al. [22] applied triangular IFNs to fuzzy
Fig. 2. Figure for membership functions of linguistic variables of rating under the
subjective evaluation criteria
Fig. 3. Figure for membership functions of linguistic variables for rating under the
importance
group decision-making problems with an application to the contractor selection.

This paper chooses to use the triangular IFNs because of their conceptual and
computational simplicities [11].
4 Aggregation Method
TOPSIS (Technique for Order Preference by Similarity to an Ideal Solution)

method proposed by Hwang and Yoon [10] is adopted to aggregate all ratings.
The basic principle of TOPSIS is that the chosen alternative must have the
shortest distance from the ideal solution and the farthest distance from the
negative-ideal solution [16]. The original TOPSIS is only applicable for crisp
ratings. In this paper, all ratings are represented by triangle fuzzy numbers.
Therefore, this paper extents the original TOPSIS to fuzzy environment. The
detailed processes are listed as follows.
Suppose the fuzzy ratings under all criteria are denoted as A = [aij ], (i =
1, · · · , m, j = 1, · · · , n).
Step 1. Normalize the evaluation index as: Compute the normalized fuzzy deci-
sion matrix (V = [vij ]). The normalized value nij is calculated as:
aij
vij = . (1)
a2ij
Step 2. Calculate entropy measure of every index the following equation:

aij
pij = m , (2)
aij i=1
1
Ej = − K [Pij × ln Pij ], where K= . (3)
i
ln(m)
Step 3. Define the divergence through:
d j = 1 − Ej . (4)
Step 4. Obtained the normalized weights of indexes as:

di
Wj = . (5)
dj
Step 5. Let ã = (a1 , a2 , a3 ) and b̃ = (b1 , b2 , b3 ) be two triangular fuzzy numbers,

then the vertex method is defined to calculate the distance between them:

1
d(ã, b̃) = [(a1 − b1 )2 + (a2 − b2 )2 + (a3 − b3 )2 ]. (6)
3
Step 6. Determine the ideal and negative-ideal solution:
vj+ = vi+ , · · · , vn+ = [(max vij |i ∈ I), (min vij |i ∈ I)], (7)
vj− = vi− , · · · , vn− = [(min vij |i ∈ I), (max vij |i ∈ I)]. (8)
Step 7. Calculate the separation measures, using the n-dimensional Euclidean

distance. The separation of each alternative from the ideal solution is given
as:

n
d+
j = (vij − vj )2 , i = (1, 2, · · · , m), (9)
j=i

n
d−
j = (vij − vj )2 , i = (1, 2, · · · , m). (10)
j=i
Step 8. Calculate the relative closeness to the ideal solution. The relative close-
ness of the alternative aj with respect to A∗ is defined as:
d−
j
C.Li = , I = 1, 2, · · · , m. (11)
d− +
i + di
Step 9. Rank the preference order. By ordering C. Li in descending order,

provinces can be ranked from the best to the worst.
5 Case Study
This study presents a case study of 9 provinces from Yellow River basin. These
9 provinces of China are Qinghai, Sichuan, Gansu, Ningxia, Inner Mongolia,
Shaanxi, Shanxi, Henan and Shandong. The original data are from Liu et al.
[13]. Based on the original data, we have experts to make their fuzzy judgements
for all the criteria (C1 –C5 ). The decision table is Table 3.
We also have experts give their ratings about the importance of all criteria
Based on Eqs. (1) and (2), we obtained results shown in Tables 5 and 7.
By using Eq. (3) results are E1 , E2 , E3 , E4 , E5 = (0.9, 0.9, 0.6), (0.8, 0.9, 0.9),
(0.8, 0.9, 0.9), (0.6, 0.8, 0.9), (0.8, 0.8, 0.9) respectively (Table 6).
Table 3. Fuzzy judgements for all criteria
Criteria Water Socio- Water Socio- Socio-

alternatives resources economic resources economic economic
pressure (C1 ) pressure (C2 ) state (C3 ) state (C4 ) response (C5 )
Qinghai MH ML G MP MG
Sichuan MH L G P Vp
Gansu H H P P F
Ningxia VH ML MG G F
Inner H MH F G VP
Mongolia
Shaanxi H ML MG P MP
Shanxi H L MP VP F
Henan MH MH MP MP MP
Shandong MH F VP P P
Table 4. Decision matrix for importance
Criteria Water Socio- Water Socio- Socio-

alternatives resources economic resources economic economic
pressure (C1 ) pressure (C2 ) state (C3 ) state (C4 ) response (C5 )
Qinghai MH ML G MP MG
Sichuan MH L G P Vp
Gansu H H P P F
Ningxia VH ML MG G F
Inner H MH F G VP
Mongolia
Shaanxi H ML MG P MP
Shanxi H L MP VP F
Henan MH MH MP MP MP
Shandong MH F VP P P
Table 5. The quantitative decision matrix
Criteria C1− C2− C3+ C4+ C5−

alternatives
Qinghai 0.4, 0.5, 0.6 0.1, 0.2, 0.3 0.7, 0.9, 1.0 0.5, 0.7, 0.9 0.5, 0.7, 0.9
Sichuan 0.4, 0.5, 0.6 0.0, 0.1, 0.2 0.7, 0.9, 1.0 0.0, 0.1, 0.3 0.0, 0.0, 0.1
Gansu 0.6, 0.7, 0.8 0.6, 0.7, 0.8 0.0, 0.1, 0.3 0.0, 0.1, 0.3 0.3, 0.5, 0.7
Ningxia 0.7, 0.8, 0.9 0.1, 0.2, 0.3 0.5, 0.7, 0.9 0.7, 0.9, 1.0 0.3, 0.5, 0.7
Inner 0.6, 0.7, 0.8 0.4, 0.5, 0.6 0.3, 0.5, 0.7 0.7, 0.9, 1.0 0.0, 0.0, 0.1
Mongolia
Shaanxi 0.6, 0.7, 0.8 0.1, 0.2, 0.3 0.5, 0.7, 0.9 0.0, 0.1, 0.3 0.1, 0.3, 0.5
Shanxi 0.6, 0.7, 0.8 0.0, 0.1, 0.2 0.1, 0.3, 0.5 0.0, 0.0, 0.1 0.3, 0.5, 0.7
Henan 0.4, 0.5, 0.6 0.4, 0.5, 0.6 0.1, 0.3, 0.5 0.1, 0.3, 0.5 0.1, 0.3, 0.5
Shandong 0.4, 0.5, 0.6 0.1, 0.3, 0.5 0.0, 0.0, 0.1 0.0, 0.1, 0.3 0.0, 0.1, 0.3

a 4.7, 5.6, 6.5 1.8, 2.8, 3.8 2.9, 4.4, 5.9 2, 3.2, 4.7 1.6, 2.9, 4.5
ij
m 2
i=1 aij 1.6, 1.9, 2.2 0.8, 1.1, 1.4 1.3, 1.7, 2.2 1.1, 1.5, 1.9 0.7, 1.2, 1.7
Note: − and + signs denote the negative and positive criteria respectively.
Table 6. Normalize the evaluation index
Criteria C1− C2− C3+ C4+ C5−

alternatives
Qinghai 0.3, 0.3, 0.3 0.1, 0.2, 0.2 0.6, 0.5, 0.5 0.5, 0.5, 0.5 0.7, 0.6, 0.0
Sichuan 0.3, 0.3, 0.3 0.0, 0.1, 0.1 0.6, 0.5, 0.5 0.0, 0.1, 0.2 0.0, 0.0, 0.1
Gansu 0.4, 0.4, 0.4 0.7, 0.6, 0.6 0.0, 0.1, 0.1 0.0, 0.1, 0.2 0.4, 0.4, 0.4
Ningxia 0.4, 0.4, 0.4 0.1, 0.2, 0.2 0.4, 0.4, 0.4 0.6, 0.5, 0.5 0.4, 0.4, 0.4
Inner 0.4, 0.4, 0.4 0.5, 0.5, 0.4 0.2, 0.3, 0.3 0.6, 0.5, 0.5 0.0, 0.0, 0.1
Mongolia
Shaanxi 0.4, 0.4, 0.4 0.1, 0.2, 0.2 0.4, 0.4, 0.4 0.0, 0.1, 0.2 0.1, 0.3, 0.3
Shanxi 0.4, 0.4, 0.4 0.0, 0.1, 0.1 0.1, 0.2, 0.2 0.0, 0.0, 0.1 0.4, 0.4, 0.4
Henan 0.3, 0.3, 0.3 0.5, 0.5, 0.4 0.1, 0.2, 0.2 0.9, 0.2, 0.3 0.1, 0.3, 0.3
Shandong 0.3, 0.3, 0.3 0.1, 0.3, 0.4 0.0, 0.0, 0.1 0.0, 0.1, 0.2 0.0, 0.1, 0.2
By using Eq. (4) divergence are d1 , d2 , d3 , d4 , d5 = (0.0, 0.0, 0.4), (0.3, 0.1, 0.1),
(0.2, 0.1, 0.1), (0.4, 0.2, 0.1), (0.3, 0.2, 0.1) respectively.
According TOPSIS, we obtained weights: W = [(0.005, 0.005, 0.154), (0.099,
0.038, 0.020), (0.084, 0.050, 0.024), (0.184, 0.095, 0.038), (0.104, 0.068, 0.033)].
By using Eq. (5) the normalized weights are W1 , W2 , W3 , W4 , W5 =
(0.0, 0.0, 0.2),
(0.1, 0.0, 0.0), (0.1, 0.1, 0.0), (0.2, 0.1, 0.0), (0.1, 0.1, 0.0) respectively.
Table 7. Calculate entropy measure
Criteria alternatives C1− C2− C3+ C4+ C5−

Qinghai 0.1, 0.1, 0.1 0.1, 0.1, 0.1 0.2, 0.2, 0.2 0.3, 0.2, 0.2 0.3, 0.2, 0.2
Sichuan 0.1, 0.1, 0.1 0.0, 0.0, 0.1 0.2, 0.2, 0.2 0.0, 0.0, 0.1 0.0, 0.0, 0.0
Gansu 0.1, 0.1, 0.1 0.3, 0.3, 0.2 0.0, 0.0, 0.01 0.0, 0.0, 0.1 0.2, 0.2, 0.2
Ningxia 0.2, 0.1, 0.1 0.1, 0.1, 0.1 0.2, 0.2, 0.2 0.4, 0.3, 0.2 0.2, 0.2, 0.2
Inner Mongolia 0.1, 0.1, 0.1 0.2, 0.2, 0.2 0.1, 0.1, 0.1 0.4, 0.3, 0.2 0.0, 0.0, 0.0
Shaanxi 0.1, 0.1, 0.1 0.1, 0.1, 0.1 0.2, 0.2, 0.2 0.0, 0.0, 0.1 0.1, 0.1, 0.1
Shanxi 0.1, 0.1, 0.1 0.0, 0.0, 0.1 0.0, 0.1, 0.1 0.0, 0.0, 0.0 0.2, 0.2, 0.2
Henan 0.1, 0.1, 0.1 0.2, 0.2, 0.2 0.0, 0.1, 0.1 0.0, 0.1, 0.1 0.1, 0.1, 0.1
Shandong 0.1, 0.1, 0.1 0.1, 0.1, 0.1 0.0, 0.0, 0.0 0.0, 0.0, 0.1 0.0, 0.0, 0.1
After the weighted normalized decision matrix we obtain a fuzzy weighted

decision table. Taking Qinghai as example, by applying the Eq. (6) results are
given below:

1
+
D1 = + [(0 − 0.001)2 + (0 − 0.001)2 + (0 − 0.042)2 ]
3

1
+ [(0 − 0.012)2 + (0 − 0.007)2 + (0 − 0.004)2 ]
3

1
+ [(1 − 0.047)2 + (1 − 0.026)2 + (1 − 0.011)2 ]
3

1
+ [(1 − 0.083)2 + (1 − 0.045)2 + (1 − 0.019)2 ]
3

1
+ [(1 − 0.071)2 + (1 − 0.040)2 + (1 − 0.002)2 ]
3
=2.756,

− 1
D1 = + [(1 − 0.001)2 + (1 − 0.001)2 + (1 − 0.042)2 ]
3

1
+ [(1 − 0.012)2 + (1 − 0.007)2 + (1 − 0.004)2 ]
3

1
+ [(0 − 0.047)2 + (0 − 0.026)2 + (0 − 0.011)2 ]
3

1
+ [(0 − 0.083)2 + (0 − 0.045)2 + (0 − 0.019)2 ]
3

1
+ [(0 − 0.071)2 + (0 − 0.040)2 + (0 − 0.002)2 ]
3
=2.112.
Calculate the relative closeness to the ideal solution using Eq. (4), Qinghai’s
results are given below:
2.112
CLQinghai = = 0.434.
2.112 + 2.756
By the TOPSIS method, we get the rank of 9 provinces as shown in Table 8.
Table 8. Table of ranks
Alternatives TOPSIS index Rank

Qinghai 0.434 1
Sichuan 0.403 5
Gansu 0.394 9
Ningxia 0.417 2
Inner Mongolia 0.405 4
Shaanxi 0.399 7
Shanxi 0.402 6
Henan 0.409 3
Shandong 0.396 8
Based on the ranking result of TOPSIS method, the water resources security
degrees of the 9 provinces.
6 Conclusions
This paper presented a fuzzy multi-criteria evaluation TOPSIS method for the
water resource security evaluation. 9 provinces within the Yellow River basin was
ranked for the water resources security degree based on the proposed method.
The structure of “pressure-state-response” was embedded in developing the
methodology. A fuzzy multi-criteria was proposed not just to deal with the eval-
uation created on the established indicators, additionally to tackle the inherent
suspicions. Based on the case study results, Qinghai, Ningxia, Henan, Inner
Mongolia and Sichuan are ranked top 5 of the water security degree, with TOP-
SIS index greater than 0.4. Future research will be focused on the uncertainty
analysis, innovative evaluation index system and aggregation methods of water
resource security evaluation.
Acknowledgements. We are thankful for financial support from the Research Center
for Systems Science & Enterprise Development (Grant No. Xq15C01), the National
Natural Science Foundations (Grant No. 71671118, Grant No. 71601134) and Project
funded by China Postdoctoral Science Foundation.
References
1. Atanassov KT (1986) Intuitionistic fuzzy sets. Fuzzy Sets Syst 20:87–96
2. Bitterman P, Tate E et al (2016) Water security and rainwater harvesting: a con-
ceptual framework and candidate indicators. Appl Geogr 76:75–84
3. Cook C, Bakker K (2012) Water security: debating an emerging paradigm. Glob
Environ Change 22:94–102
4. Chen SJ (2004) Water resources security concept and its discussion. China Water
Res 17:13–15
5. Dong QJ, Liu X (2014) Risk assessment of water security in Haihe River Basin
during drought periods based on D-S evidence theory. Water Sci 7:119–132
6. Jiang WL (2001) Study on water resource safety strategy for China in the 21st
century. Adv Water Sci 1:66
7. Jia SF, Zhang JY, Zhang SF (2002) Regional water resources stress and water
resources security appraisement indicators. Progr Geogr 21:538–545
8. Koehler A (2008) Water use in LCA: managing the planet’s freshwater resources.
Int J Life Cycle Assess 13:451–455
9. Hall J, Borgomeo E (2013) Risk-based principles for defining and managing water
security. Philos Trans R Soc London A Math Phys Eng Sci 371:20120407
10. Hwang CL, Yoon K (1981) Multiple attributes decision making methods and appli-
cations. Springer, Heiddberg
11. Li Z, Liechty M et al (2014) A fuzzy multi-criteria group decision making method
for individual research output evaluation with maximum consensus. Knowl Based
Syst 56:253–263
12. Linser S (2001) Critical analysis of the basics for the assessment of sustainable
development by indicators. Schriftenreihe Freiburger Forstliche Forschung, Bd. 17,
Freiburg
13. Liu KK, Li CH et al (2014) Comprehensive evaluation of water resources security in
the Yellow River basin based on a fuzzy multi-attribute decision analysis approach.
Hydrol Earth Syst Sci 18:1605–1623
14. Norma ES, Dunn G et al (2013) Water security assessment: integrating governance
and freshwater indicators. Water Res Manag 27:535–551
15. OECD (1993) Core set of indicators for environmental performance reviews: a syn-
thesis report by the group on the state of the environment. Environment mono-
graphs, vol. 83. Organization for Economic Co-operation and Development, Paris
16. Shirouyehzad H, Dabestani R (2011) Evaluating projects based on safety criteria:
using TOPSIS. In: 2011 2nd International conference on construction and project
management IPEDR, vol. 15. Singapore
17. Van Raan AFJ (1996) Advanced bibliometric methods as quantitative core of peer
review based evaluation and foresight exercises. Scientometrics 36:397–420
18. Vacik H, Wolfslehner B et al (2007) Integrating the DPSIR-approach and the
analytic network process for the assessment of forest management strategies. In:
Reynolds K, Rennolls K et al (eds) Sustainable forestry: from monitoring and
modelling to knowledge management and policy science. CABI Publishing
19. Voeroesmarty CJ, McIntyre PB et al (2010) Global threats to human water security
and river biodiversity. Nature 467:555–561
20. Wolf A (2007) Shared waters: conflict and cooperation. Annu Rev Environ Res
32:269–279
21. Xia J, Zhang YY (2007) Water security in north China and countermeasure to
climate change and human activity. Phys Chem Earth Parts A/B/C 33:359–363
22. Xu JP, Li ZM et al (2013) Multi-attribute comprehensive evaluation of individual
research output based on published research papers. Knowl Based Syst 43:135–142
Management of Technological Modes of System
Distributed Generation Electric Energy
on the Basis of Daily Schedules
of Electric Loadings
Abdulla Arifjanov1(B) and Romen Zakhidov2

1
Center for the Development of Software Products and Hardware-Software
Complexes, Tashkent University of Information Technologies, Tashkent, Uzbekistan
arifjanov@yandex.ru
2
Institute of Power Engineering and Automation,
Academy of Sciences of Uzbekistan, Tashkent, Uzbekistan
Abstract. Currently the concept of distributed power with application

of renewable energy sources is gaining popularity and is becoming not
only a supplement but also a real alternative of traditional energy supply
from centralized networks. The distributed generators of electricity inte-
grated into Unified Energy System (UES), represent difficult system in
which communication between observable parameters and a system con-
dition has difficult and ambiguous character. Therefore, the approach to
management in the automatic control systems, based on management on
output parameters, is inadequate as a matter of fact. In this case for-
mation of operating influence should be carried out on the basis of data
about a current condition of object of management. The daily schedules
of electric load allow properly assess the mode of the work of electrical
equipment and the enterprise as a whole, identify the bottlenecks and
reserves, to set the optimal mode of the work. In this paper are con-
sidered an issue of the development of methods for the estimation of
electrical loads for the various types of electricity customers in the res-
idential and public buildings and addresses issues of the integration of
distributed generators to the UES and ensuring the optimal management
of technological modes of their work.
Keywords: Electric power system · Hybrid system of power supply ·

Renewable energy sources · Distributed generation electric energy · Elec-
trical loads · Daily schedule load · Management
1 Introduction
The future power supply system should combine large power stations without
which are problematic electro supply of large consumers and maintenance of
growth of power consumption. Large power stations allow raising voltage with
Technological Modes of System Distributed Generation Electric Energy 1199
generating to voltage of the main network on which there is a transport of the

electric power to the large centers of consumption. Installations of the distrib-
uted generation, including on renewable energy sources (RES), work through
distributive networks. The following level of a power supply system of the future
will make mini-and micro-installations (mini-and micro-hydroelectric power sta-
tion, wind generators, solar energy installations, fuel elements, etc.) which are
connected on a low voltage and are established at small consumers, for example,
in separate houses or on small enterprises, including agricultural [1,7].
In this regard, is currently the concept of distributed power with application
of renewable energy sources (RES) is gaining popularity and is becoming not
only a supplement but also a real alternative of traditional energy supply from
centralized networks.
Distributed generation systems represent a hybrid power system consisting
of various energy sources and are built in close proximity to the customer and
to the maximum extent take into account the individual characteristics of the
consumer’s by power and profile (Fig. 1). They have such advantages as control-
lability, reliability, economy (including by eliminating the losses of transporting),
scalability and flexibility [7].
Fig. 1. A structural diagram a centralized and distributed power generation
In these systems electric energy is generated with use of solar photo-electric

panels, wind turbines or other energy installations. Generating of thermal energy
for systems of heating, hot water supply and technological processes is carried out
with use of solar collectors (flat and vacuum tubular), geothermal systems, and
also other converters of thermal energy. The combination of various renewable
energy sources is not only presence of such elements, as solar collectors, photo-
electric panels, wind turbines, thermal pumps, but also use of a uniform control
system for maintenance of effective work of these elements that makes a basis of
stable hybrid system of power supply.
1200 A. Arifjanov and R. Zakhidov
As of today the distributed generation systems in most cases presented in the

form of autonomous energy centers, as the connection of distributed generators
to centralized networks so far is limited by the absence or imperfection of the
regulatory framework and effective methods of controls of technological modes.
For an effective utilization of hybrid system of power supply with applica-
tion of renewable energy sources (RES) it is necessary to co-ordinate intensity
of receipt of energy with its expense consumers. Such coordination can be car-
ried out on the basis of monitoring of the current information on receipt and
power consumption and the management which is carried out by uniform system,
constructed on the basis of modern information-communication technologies [2].
Control of the distributed generators is carried out as control of a micro net-
work (micro grid) of the “virtual” power stations integrated into a global network
(Unified Energy System-UES) and, thereby, in the electric power and capacity
market that will promote increase of a role of the consumer in management of a
power supply system [2].
On a way of creation of “virtual” power stations it is necessary to solve a
number of the organizational-technological problems, one of which is the problem
of connection of the distributed generators to a uniform network and manage-
ment of their work. The distributed generators of electricity integrated into UES,
represent difficult system in which communication between observable parame-
ters and a system condition has difficult and ambiguous character. Therefore,
the approach to management in the automatic control systems, based on man-
agement on output parameters, is inadequate as a matter of fact. In this case
formation of operating influence should be carried out on the basis of data about
a current condition of object of management [2].
The specific issues on usage of renewable energy sources as a distributed
generator, as well the approach to the synthesis of the operational level of the
active-adaptive system of automatic operation control of the combined plant as
part of the UES were reviewed in the report [2].
The effectiveness of control by system of distributed generation of electrical
energy, comprising renewable energy sources (RES), energy storages and intelli-
gent control system is achieved through the ability to complex influence on the
technological parameters of the generator modes.
For example, voltage regulation apart from the use of traditional means,
may be implemented by change the generation of active or reactive power, by
exposure on the energy storage means or by management of load.
This set of measures together with hardware and software represents the
control system of system of distributed generation of electric energy.
The main purpose of this control system is the integration of distributed
generators to the UES and to ensure the optimal management of technological
modes of their work to meet the needs of consumers in the electrical energy.
Needs for electric power supply are characterized by the schedule of electrical
loads. Dependence of load on time of day is called the daily schedule of electric
load. It can be compiled for individual and group of electricity consumers, and for
the Energy System, which includes a power plant or for large of the incorporated
power system [6].
The daily schedules of electric load allow properly assess the mode of the work
of electrical equipment and the enterprise as a whole, identify the bottlenecks
and reserves, to set the optimal mode of the work.
Most fundamentally change in the electrical load associated with communal
general needs. Fig. 2 represents the daily schedule, from which we see that the
electrical load in the winter more than the summer and is sharply reduced at
night hours [6]. The smallest of its value called of the minimum load. In the
afternoon and evening hours is observed increasing of load and a more consid-
erable change - in winter. There are two maximum loads-morning and evening.
Schedule of electrical loads should be provided (“covered”) on a mandatory basis.
Therefore, aspire carry out all necessary renovations during the summer in order
to virtually all equipment of power plant could be used to provide winter peak.
This maximum is called the peak load.
Fig. 2. Daily schedule of electric load
Calculation of the urban network load includes determining the load of indi-
vidual consumers (residential buildings, public buildings, municipal services,
etc.) and electrical elements of the system (distribution lines, transformer sub-
stations, distribution centers, power centers, etc.) [3].
In connection with the appearance of part of the population the possibility of
use in the home of a wide range of modern household appliances and equipment,
as well as in connection with the construction of urban and rural buildings on

individual projects with luxury apartments, new specific standards of electrical
loads should be determined based on the actual measurements electrical loads,
taking into account their probabilistic nature and features of modern luxury
apartments in new buildings of the city, analysis of electrical appliances market
development and technology, the degree of saturation of apartments, both now
and in the future.
In the cities, the problem of reliable determination of residential and public
buildings loads, and provide energy savings is especially difficult, because of
population growth and an increase the number of used electrical appliances.
Along with the increasing number of electrical appliances used by the population
grows and their installed capacity.
Modern residential buildings full of a lot of different power consumers. These
include lighting and household appliances and power equipment. There is a con-
stant process of improving housing comfort, and this in turn increases the amount
of domestic power consumers and increased domestic electricity consumption.
Increasing the number of storeys homes tightens the requirements for reliability
and continuity of supply of electric networks.
The same features are characteristic and for public buildings, which include:
a variety of institutions and the organization of management, finance, credit,
state monopolies, education; pre-schools, libraries, archives, enterprises of trade,
public catering, consumer services; hotels, hospitals, museums, entertainment
and sports facilities.
In this regard, the reliable definition of electric loadings of residential and
public buildings in the cities became very actual problem on which solution the
end results of actions for energy saving in many respects depend. The normative
document The electric facilities of residential and public buildings. Construction
norms and rules using to calculate of loadings of residential buildings doesn’t
solve the designated problem, because the specific norms of electric loadings
given in this document for various types of consumers of energy have been defined
in the 80s last century and don’t correspond to today’s condition of power con-
sumption in residential and public buildings. In many cases the calculated values
of electric loadings of residential and public buildings defined according to this
normative document with use of various correction coefficients aren’t true [7].
2 The Problem Statement
Analysis of the existing regimes of power consumption in the urban electric

network shows the need to develop new accounting standards specific to electrical
loads for different types of residential and public buildings consumers, taking into
account the current realities of power consumption in the residential and public
sector [3–5].
Estimated electrical loads and dynamics of their growth in the future is the
foundation that determines the nature of construction and the development of
urban electrical networks. The electrical loads of houses and public buildings are
random and depend on some of factors: the way of life of the various families,
the number of power consumers and power and others. Therefore, the basis for
determining the load is used the probabilistic and statistical approach to the
load as a non-stationary random variable. Hence, the estimated electrical load
or network element assumes to be probable maximum load value for an interval
of 30 min.
To obtain reliable data while a design of standard charts and determine
their numerical characteristics it is necessary correctly handle the experimental
results, based on the position of mathematical statistics and probability theory.
It should be borne in mind that, according to the law of large numbers and the
theory of probability, the results determine the average aggregate will be valid
when the number of tests or, in other words, the number of surveyed members
together will be quite large. On the other hand, with an increase in the number
of aggregate members, their examination is a huge work, and we have a problem
of determining a sufficient number of members in the aggregate, that allow to
obtain mean values with sufficient accuracy.
Therefore, for an assessment of this loading very often used some general-
ized indicators, coefficients, specific loadings and specific expenses of the electric
power.
The existing algorithms of load calculation of the industrial power systems,
determination of the maximum loads and the choice of electric equipment don’t
consider dynamics of growth and character of household loading, laws of its
functioning, do not allow to take into account in calculating of the actual loading
the relationship of load changes on a day of the week, time of day. Despite
the large number of works on the subject, the daily schedule load model of
residential and public buildings and their practical implementation are not well
developed. There is not a program that gives to specify the load calculation data
of residential and public buildings on the current state of the loads.
The method widely used for determining the maximum electrical load of elec-
trical networks are based on the measurement of the average load of electrical
consumers over a given time period t (t = 8 h or 0.5 h) with variable initial mea-
suring point. The total number of electrical consumers, on which measurements
are made should not be less than 20% of the total number of electrical con-
sumers connected to the electric network (but not less than 15). Measurements
should be carried out repeatedly and for a long time [2,7]. All these methods are
characterized by a long time and a low accuracy of measurements; by a signif-
icant difference from the actual loads, especially the total; by significant errors
in calculation (do not include the probabilistic nature of the electrical loads in
the urban network). These methods also do not take into account the time fac-
tor of maximum electric load of each customer. In addition, these techniques
are designed for installations with a regular operating mode and require large
amounts of additional measurements and not possible to determine the desired
value for a predetermined time period including at least three units.
The way of definition of the maximum load of electric consumers [6] according
to which measure once individual loadings in various technological operating
modes, total time of each technological mode for basic time and calculate the
load group with a probability of exceeding no more than the required conditions
of the problem.
However, this method does not provide sufficient accuracy while determining
the estimated load for a group of different types of electric consumers, in partic-
ular, if even in the same technological modes individual electrical loads are not
constant.
3 A New Method of Solving the Problem

A method for determining the maximum electrical load of electrical networks
and individual electrical consumers patented by one of author of this paper [5].
This method significantly reduces the numbers and time of measurements
and improve the accuracy and reliability of determining the maximum electrical
load of consumers.
The essence of a way is electric loads of the certain consumers are measured
in various technological operating modes for basic time. Then from set of the
measured loads the maximum size of load of the certain consumers is determined.
The maximum value of the electric load is defined as the highest value for all
of the measured periods of typical daily schedules, with a maximum electrical
load in a time interval 15-minute and 30-minute. Further calculation of loads of
a network is conducted in a general view on a formula [4].
Let’s us consider the essence and use of this new method. The essence of the
new method for calculating the load of 0.4 kV and 6–10 kV electrical networks
is to determine estimated load for each time period of the day according to
the on typical daily schedule of customer load. It is found by summing of the
mathematical expectations and the mean square deviations of all consumers. For
the calculated maximum value of the total load is accepted its greatest value from
all the calculating for hour or half-hour periods of the daily schedule.
The mathematical expectation for the construction of a typical load graph
of homogeneous consumers will be equal [4]:
m l n
β=1 × γ=1 × j=1 Yijγβ
Pcpi = ,
m ln
where
Pcpi : mathematical expectation of the random load;
Yij : matrix of initial values of the loads daily schedule;
n : the number of points measured value of the daily schedule for a
half-hour or hour interval;
l : the number of load measurements (in days) in each point at
electrical network;
m : the number of homogeneous consumers, accepted as experimental.
So, the root-mean-square deviation equals:
m 2
m l n l n
√ 2 2
β=1 × γ=1 × j=1 Yijγβ β=1 × γ=1 × j=1 Yijγβ
σpi = Di = − .
m ln m ln
The value of measured daily schedule load can be given as a matrix-current

or power, where:
i : rows (at half-hour intervals i = 48 per day, at hourly intervals i = 24).
j : columns of the matrix (at 5-minute increments of the load curve in a
half-hour interval j = 7, j = 13 in an hour interval).
Investigation of the electrical load of public buildings has its own peculiar-
ity. Average maximum values of different groups of public buildings cannot be
subjected to similar statistical processing as the load of residential buildings,
so there are large variations in the average value of apartments load. Because
absolute values of electrical loads for different by construction and character of
public building have a very large variation of average values and do not provide
sufficient accuracy of calculations. Therefore, it is necessary to conduct process-
ing of public buildings statistical data for homogeneous consumers.
4 The Algorithm of a New Method Implementation

The algorithm of the new method is as follows [4]:
• The experimental investigations of urban power consumers and electrical
loads conduct.
• Using the mathematical apparatus of the theory of probability and math-
ematical statistics are being built characteristic (model) daily schedules of
electrical loads for all types of urban power consumers and the electrical
load.
• On the basis of these daily schedule characteristic, using Lagrange, Newton,
Stirling, Bessel, Chebyshev and others methods are determined analytical
functions (equations, formulas) of curves (daily schedules) of mode changes
each consumer.
• Next, by functions (equations, formulas), which are describes curve of the
daily schedule load, analytically are defined of maximum coefficients Km.
In general, the equation describing the daily schedule of electric load has the
form:
S (x) = Ui =1,··· ,24 Si (x) ,
or
fi−1
S3 (x) = x3i − 3x2i x + 3xi x2 − x3 + x3 − 3x2 xi−1 + 3xx2i−1 − x3i−1 + xi
hi
Mi−1 · h2i fi−1 Mi−1 · h2i fi f Mi · h2i Mi · h2i
− xi − x+ x+ x − i xi−1 − x+ xi−1
6hi hi 6hi hi hi 6hi 6hi
x3 3
= (Mi − Mi−1 ) · + · (Mi xi − Mi−1 xi−1 )x2 + (Mi · x2 i−1 − Mi−1 · x2 i − 2fi−1 ,
6hi 6hi
+ Mi−1 · h2 i − Mi · h2 i + 2fi )x + (Mi−1 · x3 i − Mi · x3 i−1 + 6fi−1 · xi − 6fi · xi−1
1
− Mi−1 · h2i · xi + Mi · h2i · xi−1 )
6hi
= ai0 · x3 + ai1 · x2 + ai2 · x + ai3 ,
Fig. 3. The characteristic (typical) daily schedule of electric load of the shop manufac-
tured goods: 1- constructed from experimental data, 2- built by Lagrange polynomial
where
Mi − Mi−1
a0 = ,
6hi
1
a1 = (Mi−1 xi − Mi · xi−1 ),
2hi
1
a2 = (Mi · x2 i−1 − Mi−1 · x2 i + Mi−1 · h2 i − Mi · h2 i − 2fi−1 + 2fi ),
2hi
1
a3 = (Mi−1 · x3 i − Mi · x3 i−1 + 6fi−1 · xi − 6fi · xi−1
6hi
+ Mi · h2i · xi−1 − Mi−1 · h2i · xi ).
On the base of this algorithm, substituting the values Mi , fi , hi , xi from the

table, we obtain a system from 24 equations in 3rd order.
As a result, for the curve of the daily schedule of electric load we obtain a
system of equations, which has the following general form:
S3 (x) = ai0 x3 + ai1 x2 + ai2 x + ai3 ,
where ai0 , ai1 , ai2 , ai3 : coefficients of each equation.

Substituting the data from the daily schedule in these equations, we can
obtain the curve segments of the daily schedule, and combining all of the 24
segments, as a result, get the curve of daily schedule, that is almost completely
coinciding with the characteristic of daily schedule of consumer loads.
For each electrical consumer or appliances construct its equation of the curve
and own formula for the maximum electrical load. As a result, the calculation
is obtained precise and more authentic and closer to the actual load, taking
into account the peculiarities of load conditions change, the time factor for each
appliance and consumer included in the city’s power system.
Example 1. The characteristic (typical) daily schedule of electric load of shop

manufactured goods, built using the Lagrange polynomial for x0 = 1, x1 = 10,
x2 = 16, x3 = 22; y0 = 10, y1 = 54, 98 = v2 , v3 = 10 is as follows (Fig. 3) [4]:
P (t) = 21.95057 − 15.91852t + 2.72962t2 − 0.26084t3 .
5 Conclusions
The distributed generators of electricity integrated into Unified Energy System

(UES), represent difficult system in which communication between observable
parameters and a system condition has difficult and ambiguous character. There-
fore, the approach to management in the automatic control systems, based on
management on output parameters, is inadequate as a matter of fact. In this
case formation of operating influence should be carried out on the basis of data
about a current condition of object of management.
The daily schedules of electric load allow properly assess the mode of the work
of electrical equipment and the enterprise as a whole, identify the bottlenecks
and reserves, to set the optimal mode of the work.
This paper develops an improved method for determining the estimated elec-
trical loads in residential and public buildings with using characteristic of daily
load curves of the same type of consumers, constructed from experimental mea-
surements of the characteristics of the typical power consumption of residential
and public buildings consumers.
The use of the developed methods, models and algorithms allow more reliably
determine the electrical loads in modern residential and public buildings taking
into account peculiarities the load condition change each electrical consumer and
load, included in the city’s power supply system.
References
1. Arifjanov AS, Ayupov AS (2015) Forecasting and rationalization of energy con-
sumption for megacities taking into account the factors caused by scientific and
technical progress, with application of informational - analytical technologies. In:
A report on the global forum for energy, environment and commercial civilization
(GFEECC)
2. Arifjanov AS, Zakhidov RA (2017) An approach to the creation of the adaptive
control system for integration of nonsteady power sources into a common electric
power grid. In: Proceedings of the tenth international conference on management
science and engineering management. Springer, pp 563–574
3. Arifjanov AS, Zakhidov RA, Saidkhodjaev AG (2014) Determination of electrical

loads in modern residential and public buildings. In: First all-Russian scientific-
practical conference “Energy efficiency and energy saving: theory and practice”
4. Arifjanov AS, Yusupov ZE, Saidkhodjaev AG (2016) The methods, models and
algorithms of electrical loads estimation for different type’s electricity consumers
in the residential and public buildings. In: Proceedings of the 3rd international
symposium on environment and morality, SEM 2016
5. Saidkhodjaev AG, Teshabayev BM (2010) A method for determining the maximum
load of electric network and the maximum electrical load of individual consumers
6. Volobrinsky SD, Kayal GM et al (1971) The electrical load of the industrial enter-
prises. Leningrad branch
7. Zakhidov RA (2015) Development of renewable energy and implementation of pro-
grams of energy saving in Uzbekistan as a basis of ensuring power safety. In: A report
on the global forum for energy, environment and commercial civilization (GFEECC)
Haze-Related Air Pollution and Impact
on the Stock Returns of the Listed Steel
Companies in China
Kai Liu1 , Ying Li1(B) , and John Thomas Delaney2

1
liyinggs@scu.edu.cn
2
Kogod School of Business, American University,
4400 Massachusetts Ave, Washington, D.C. 20016-8044, USA
Abstract. The purpose of this paper is to study the effects of air pollu-
tion, especially the haze, on the stock returns of steel mills. This study
collects data from air quality index, variables represent characteristic of
the eleven steel mills in China and the stock returns ratio from the stock
market etc. SPSS19.0 is utilized to conduct a descriptive analysis of the
correlation between the principal air pollution(including PM2.5 , PM10 ),
the variables represent charactering(monetary funds, net assets, liabili-
ties, operating margin, financial leverage, and total asset growth) of the
elven steel mills in China and the stock returns ratio from the stock mar-
ket etc. Hence, through the research on the air pollution index, analysis
of the listed company’s earnings and stock price, and by using the linear
regression analysis method, research shows that the serious air pollution
have a negative impact on the profitability of iron and steel enterprises
through the emotion and expectations of investors. It is imperative for
people to tackle air pollution urgently.
Keywords: Haze related air pollution · Stock return rate · Listed steel
companies
1 Introduction
With the rapid economic development and fast expansion of productive enter-
prises, China is increasingly facing the large-scale severe air pollution. Air pol-
lution with PM2.5 and PM10 as the main pollutants troubles the people’s life,
threatening their health. According to a report released by the WHO in May,
2016, seven of the world’s 10 cities with the most severe air pollution are in
China (China Environmental Report 2016). PM2.5 has an essential impact on
air quality and visibility, and the haze may last longer based on the different
geographical and meteorological conditions, impairing the people’ health consid-
erably. The report further shows that less than 1% of China’s 500 largest cities
1210 K. Liu et al.
meet the air quality standards recommended by the WHO (China Environment
Report 2016).
The severe air pollution is caused principally by the rapid growth of pro-
ductive enterprises which use coal as the main raw material and highlight high
energy consumption. To maximize the profits, these steel companies have ignored
environmental protection, and even sacrificed the environment for a long time,
causing serious environmental problems, threatening future health of the Chinese
people and sustainable development of the ecosystem. As a resource-intensive
industry, the iron and steel industry, using coal as the main source of energy,
explores a variety of powdery and massive ferrous metals and non-metallic min-
erals through large-scale production and complex processes. The emissions of
the iron and steel industry can be divided into three categories: The first is the
exhaust gas caused in the production process, such as the smoke, sulfur diox-
ide and other harmful gases generated through the sintering, smelting, and steel
rolling process. The second type is the dust and sulfur dioxide produced by burn-
ing fuels, for instance, coal and coal gas, in the furnace. The third category is
the dust generated during the processes of transporting, loading and discharging
and processing of raw materials and fuels.
The smog not only severely impairs people’s health but limits their outdoor
activities as well. It affects individual mental activities and emotional states,
resulting in more negative emotions. This paper seeks to study whether air pol-
lution, especially the smog, affects the steel plants’ stock returns. Through exam-
ining the air pollution index, the paper analyzes the earnings and stock prices
of listed steel companies, and uses the linear regression analysis to investigate
whether the air pollution has relevance to the stock returns of these enterprises.
The structure of this paper is as follows: Introduction comes the first, which is
followed by a literature review–the second section. The third part is the research
model, data and the method. Then, the paper presents the empirical research
results and the corresponding analyses and discussions, which constitute the
fourth section. Last, it draws conclusions and provides suggestions.
2 Literature Review
Studies have shown that air pollution directly or indirectly affects people’s psy-
chological conditions and emotions. Lepori [8], using different trading techniques
in Italy and samples from major international stock exchanges, demonstrated the
relationship between air pollution, people’s emotions, as well as stock returns.
Evans [5] find that exposure to polluted air increases the level of depression,
anxiety, helplessness, and anger. Emotion affects people’s assessment of both
future prospects and risks [13]. Some studies reveal that air pollution is posi-
tively correlated with people’s negative emotions. Levy [9] examined the rela-
tionship between air pollution and stock returns. Using air quality indices and
stock return data from four US stock exchanges, they argue that air pollution
is negatively correlated with stock returns. Mehra [11] observed, in their study,
Haze-Related Air Pollution and Impact on the Stock Returns 1211
that small emotional volatility has significant impacts on the fluctuation of cap-
ital prices. Lucy [10] held that emotions, which are not caused by the investors’
future decisions, may still affect the decision makers.
To address the deteriorating air quality and ease the increasing public con-
cern, relevant government departments may reinforce stringent emission stan-
dards and formulate strict environmental policies to control the high pollution
and high emissions behavior. The regulatory policies strengthening environmen-
tal protection will affect the stock returns of iron and steel companies with the
such characteristics as high input, high pollution, and high energy consumption
[12]. Taking China’s steel mills as examples, considering the serious environmen-
tal protection attitudes and stringent environmental standards, steel mills will
face greater pressure from environmental protection and environmental costs,
and the stock returns of the listed steel companies will be affected accordingly.
Meanwhile, air pollution may easily trigger industrial policy adjustments, which,
in turn, lead to the re-allocation of resources. This may impact the listed com-
panies’ operating performances and growth prospects, resulting in stock price
fluctuation [4]. In addition, other government policies, for instance, security reg-
ulations, energy prices, credit rationing, and international cooperation, which
are closely related to air quality, also have a significant impact on the stock
prices of the listed companies [6].
It is worth noting that air quality is regional and it largely affects the sen-
timent of local investors, especially individual investors. Likewise, air pollution
may impact the stock market through influencing the stock traders’ emotion.
Currently, the two major stock exchanges and three futures exchanges in China
adopt the command-driven system rather than the quotation-driven system, and
all trading quotations should be input to the matching system by the investors
with the help of the agent brokers (exchange members). As for the command
trading system operators, when their emotions are affected by local air pollution,
their rational judgment and selection ability may be reduced, resulting in irra-
tional trading behavior and causing stock price fluctuation. In the information
age, individual investors are still the largest fund providers and the main traders
in the current stock market [3], among whom there is a “herd effect” [1,2,7]. In
this context, when outside investors make trade decisions (of the listed compa-
nies) based on the air quality of the major cities (e.g. Beijing, Shanghai), the
“herd effect” may further amplify the impact of air quality on the stock market
of the steel industry.
3 Model and Assumption
3.1 Data
This paper studies 11 Chinese steel plants, which are Shanghai Baosteel
(600019), Shougang Group (000959), Ansteel Group Corporation (000898),
Wuhan Iron and Steel (Group) Company (600005), Jiangsu Shagang Group Co.,
Ltd. (0020), Taiyuan Iron & Steel (Group) Co., Ltd. (000825), Jinan Iron and
1212 K. Liu et al.
Steel Co., Ltd. (600022), Panzhihua Iron and Steel (Group) Co. (000629), Hes-
teel Group Tangsteel Company (000709), Benxi Iron and Steel (000761), and
Nanjing Iron and Steel Group Co., Ltd. (600282). The paper gathers the stock
returns and feature variables (monetary funds, net assets, liabilities, operating
margin, financial leverage, and total asset growth) from the CSMAR database.
Moreover, some data of financial leverage and total asset growth are collected
from the WIND database and China Securities Depository and Clearing Corpo-
ration Limited (CSDC).
This paper uses six variables (monetary funds, net assets, liabilities, operating
margin, financial leverage, and total asset growth), which are borrowed from
Berger and Tuttle, to examine the stock information of the steel companies from
different dimensions. The annual financial statements are released with a time
lag. That is, the announcements are usually released in March of the next year. In
order to ensure the relative effectiveness of the financial information, the paper
uses the financial information (July, t, ∼ June, t + 1) as the stock transaction
data of the t − 1 year.
Table 1. The definitions of explained variables
Explained variables Comments

Monetary funds Indicates the size of the company
Net profit assets Indicates the company’s ability to develop
Debt ratio Indicates the company’s ability to repay the debt
Operating margin Indicates the company’s growth characteristics
Financial leverage Indicates the company’s profitability
Total assets growth rate To determine the stability of the stock return rate
The urban air quality data used in this study are gathered at the sites of
the steel plants. Partly of the data are from the PM2.5 hourly data monitored
by the five US embassies and Consulates in China. The 2013–2015 air quality
index data (such as SO2 , NO2 , CO, PM2.5 , PM10 , and AQI) are obtained from
the Ministry of Environmental Protection of the People’s Republic of China.
3.2 The Model
In order to investigate the impact of air pollution on the earnings of the steel mills,
the paper uses an ordinary least squares (OLS) regression model, which is as follows:
For Returns = αX1 + β X2 + e.
For Returns represents the dependent variable and refers to the monthly rate
of return of the steel plant.
Fig. 1. A scatter plot of returns and PM2.5
In the model, α denotes a constant term; e is the random error term; β

signifies the weight associated with the explanatory variable; X1 is PM2.5 and
X2 represents PM10 .
The purpose of this paper is to explain whether and how air quality affects the
profitability of the iron and steel plants. The explanatory variables involved are
mainly the air quality indicators. The explained variables include data such as
monetary funds, net profit assets, debt ratio, operating margin, financial leverage
and total assets growth rate (Table 2).
The linear model: Y = −0.179 × X1 + 0.233 × X2 + e, where Y represents
the returns, X1 is PM2.5 , X2 denotes PM10 , and e signifies the error.
In Table 3, the correlation coefficient R2 is 0.015, which is very small and
close to 0. Therefore, the returns are not significantly correlated to PM2.5 and
PM10 , and the e of the model is larger.
According to the analysis of the above theoretical background, in this paper,
the assumption is as follows serious air pollution have a negative impact on the
profitability of iron and steel enterprises.
Table 2. Model summary
Model R R Adjusted Std. Error Change statistics

Square R Square of the
estimate
R Square Change F df1 df2 Sig. F
change change
1 0.122a 0.015 0.007 0.160499 0.015 1.972 2 260 0.141
1214 K. Liu et al.
Fig. 2. a scatter plot of returns and PM10
Table 3. Coefficient analysis
Model Unstandardized Standardized t Sig. Correlations Collinearity

coefficients coefficients statistics
B Std. error Beta Zero-order Partial Part Tolerance VIF
1 (Constant) 0.003 0.028 0.123 0.903
PM2.5 −0.001 0.001 −0.179 −1.505 0.133 0.02 −0.093 −0.093 0.267 3.749
PM10 0.001 0 0.233 1.958 0.051 0.08 0.121 0.121 0.267 3.749
4 Empirical Results and Analysis

4.1 Descriptive Analysis
General descriptive statistical analysis & descriptive statistical analysis by com-
pany names (Table 4).
Table 4. The overall descriptive statistics are summarized
N Min Max Mean Std. Variance Skewness Std. error Kurtosis Std. Error
deviation of skewness of kurtosis
Returns 263 −0.3729 1.0142 0.0392 0.1611 0.026 1.006 0.15 5.056 0.299
PM2.5 275 17 182 67.95 30.103 906.212 1.072 0.147 1.311 0.293
PM10 275 40 259 113.84 42.381 1796.152 0.729 0.147 0.241 0.293
Valid N (list wise) 263
The population standard deviation of net profit is very large, indicating sig-
nificant fluctuations of the net profit. The variance of PM2.5 and PM10 is rela-
tively small, but also they are over 30, reflecting obvious fluctuations of the air
quality. The skewness of PM2.5 and PM10 is greater than 0 and has a maximum,
which should be removed, suggesting the right deviation. The skewness of the
net profit is less than 0, showing left deviation. The skewness has a minimum,
which should be removed (Table 5).
Table 5. The conclusion and analysis of the overall descriptive statistics
No N Min Max Mean Std. Variance Skewness Std. Error of Kurtosis Std. Error
deviation skewness of kurtosis
1 Returns 25 −0.2231 0.3908 0.0193 0.1249 0.016 0.958 0.464 2.783 0.902
PM2.5 25 33 124 55.56 19.812 392.507 1.865 0.464 4.881 0.902
PM10 25 49 ‘ 75.64 22.381 500.907 1.143 0.464 1.176 0.902
Valid N 25
(listwise) 25
2 Returns 21 −0.3728 0.2349 0.0420 0.1466 0.022 −0.832 0.501 1.706 0.972
PM2.5 25 44 154 80.84 28.868 833.39 1.19 0.464 1.043 0.902
PM10 25 58 173 108.28 30.893 954.377 0.275 0.464 −0.774 0.902
Valid N 21
(listwise) 21
3 Returns 25 −0.2137 0.4106 0.0230 0.1227 0.015 1.075 0.464 3.175 0.902
PM2.5 25 40 144 64.64 24.605 605.407 1.609 0.464 3.206 0.902
PM10 25 66 188 109.72 29.389 863.71 0.736 0.464 0.576 0.902
Valid N 25
(listwise) 25
4 Returns 25 −0.2859 0.3299 0.0257 0.1435 0.021 0.323 0.464 0.131 0.902
PM2.5 25 35 182 78.44 38.448 1478.257 1.523 0.464 2.355 0.902
PM10 25 67 218 113.08 37.627 1415.827 1.214 0.464 1.502 0.902
Valid N 25
(listwise) 25
5 Returns 19 −0.1911 1.0142 0.1490 0.2862 0.082 1.622 0.524 3.538 1.014
PM2.5 25 33 111 63.12 22.769 518.443 0.728 0.464 −0.191 0.902
PM10 25 66 179 113.04 27.902 778.54 0.658 0.464 0.022 0.902
Valid N 19
(listwise) 19
6 Returns 24 −0.3657 0.3861 0.0370 0.1712 0.029 −0.192 0.472 1.102 0.918
PM2.5 25 38 105 64.72 21.084 444.543 0.562 0.464 −1.091 0.902
PM10 25 82 176 122.12 27.039 731.11 0.123 0.464 −0.863 0.902
Valid N 24
(listwise) 24
7 Returns 25 −0.3012 0.3698 0.0337 0.1503 0.023 0.447 0.464 0.851 0.902
PM2.5 25 57 159 91.68 26.009 676.477 1.326 0.464 1.194 0.902
PM10 25 98 259 173.4 38.249 1463 0.414 0.464 0.157 0.902
Valid N 25
(listwise) 25
8 Returns 25 −0.2836 0.3081 0.0278 0.1369 0.019 0.121 0.464 0.439 0.902
PM2.5 25 17 73 35.72 13.296 176.793 1.233 0.464 1.615 0.902
PM10 25 40 144 74.56 26.013 676.673 1.182 0.464 1.261 0.902
Valid N 25
(listwise) 25
9 Returns 24 −0.3657 0.3861 0.0370 0.1712 0.029 −0.192 0.472 1.102 0.918
PM2.5 25 39 147 92.04 29.464 868.123 0.211 0.464 −0.82 0.902
PM10 25 80 204 153.68 37.201 1383.893 −0.393 0.464 −0.962 0.902
Valid N 24
(listwise) 24
10 Returns 25 −0.2762 0.3809 0.0324 0.1365 0.019 0.357 0.464 1.303 0.902
PM2.5 25 26 111 53.32 22.527 507.477 0.791 0.464 −0.108 0.902
PM10 25 55 146 94.08 27.412 751.41 0.326 0.464 −0.935 0.902
Valid N 25
(listwise) 25
11 Returns 25 −0.3672 0.4124 0.0312 0.1607 0.026 0.067 0.464 0.942 0.902
PM2.5 25 30 155 67.32 29.22 853.81 1.387 0.464 2.432 0.902
PM10 25 59 243 114.6 44.507 1980.917 1.188 0.464 1.735 0.902
1216 K. Liu et al.
4.2 Descriptive Statistical Analysis by Companies

In terms of companies, the average net income of Panzhihua Iron and Steel
(Group) Co. is negative, suffering loss, while the average PM2.5 and PM10 of
this enterprise are 35.72 and 74.56, respectively, considerably lower than the
overall PM2.5 and PM10 average, namely 67.95 and 113.84. It shows that the
air quality of Panzhihua Iron and Steel (Group) Co. is better than the overall
situation. The Shanghai Baosteel Group has the highest net income, the average
PM2.5 and PM10 of which are 55.56 and 75.64, lower than the overall PM2.5
and PM10 average. As for other data, the net income fluctuates substantially,
and the PM2.5 and PM10 are close to the overall mean. Therefore, the relation-
ship between PM2.5 and PM10 and steel plants’ net income cannot be analyzed
directly (Table 6).
Table 6. Descriptive statistical test: Chi-Square tests
Test of PM2.5 Value df Asymp. sig.(2-sided)

a
Pearson Chi-Square 23098.601 23100 0.501
Likelihood ratio 2243.815 23100 1
Linear-by-linear association 0.109 1 0.741
N of valid cases 263
Test of PM10 Value df Asymp. Sig. (2-sided)
Pearson Chi-Square 30356.775a 30492 0.707
Likelihood ratio 2395.163 30492 1
Linear-by-linear association 1.666 1 0.197
a
30856 cells (100.0%) have expected count less than 5. The minimum
expected count is .00.
b
23432 cells (100.0%) have expected count less than 5. The minimum
expected count is .00.
The Table 1 is a chi-square test of PM2.5 , while the second is a chi-square

test of PM10 . It can be seen from the tables that sig < 0.05, meaning that there
are significant differences between PM2.5 , PM10 and the net income (Table 7).
The Table 1 is about PM2.5 and the second is about PM10 . The values of
the symmetric measures are both around 0.99, which is much larger than 0.1,
indicating correlation between PM2.5 , PM10 , and the net income.
4.3 Correlation Analysis

As shown in the PM10 and net profit correlation Table 8, for most areas, the
absolute value of the Pearson correlation coefficient between the net profit and
PM10 is relatively low, suggesting an insignificant impact of PM10 on the net profit.
Only a few areas are excepted: For instance, Jiangsu shows strong correlation.
Table 7. Symmetric measures
Test of PM2.5 Value Asymp. Std. errora Approx. Tb Approx. sig

Nominal by nominal Phi 9.372 0.501
Cramer’s V 0.937 0.501
Contingency coefficient 0.994 0.501
Ordinal by ordinal Kendall’s tau-b 0.022 0.042 0.534 0.593
Kendall’s tau-c 0.022 0.042 0.534 0.593
Gamma 0.022 0.042 0.534 0.593
Spearman correlation 0.036 0.062 0.575 .566c
Interval by interval Pearson’s R 0.02 0.051 0.33 .742c
Test of PM10 Value Asymp. Std. errora Approx. Tb Approx. sig
Nominal by nominal Phi 10.744 0.707
Cramer’s V 0.935 0.707
Contingency coefficient 0.996 0.707
Ordinal by ordinal Kendall’s tau-b 0.055 0.041 1.328 0.184
Kendall’s tau-c 0.055 0.041 1.328 0.184
Gamma 0.055 0.041 1.328 0.184
Spearman correlation 0.083 0.062 1.341 .181c
Interval by interval Pearson’s R 0.08 0.053 1.292 .197c
a
Not assuming the null hypothesis.
b
Using the asymptotic standard error assuming the null hypothesis.
c
Based on normal approximation.
This correlation varies from region to region. For example, Shanghai and
Jiangsu show a positive correlation. That is, the net profit increases with the
rise of PM10 . In contrast, Nanjing shows a strong negative correlation. In other
words, the net profit reduces with an increase in PM. Overall, the majority of
the cities show a negative correlation. That is, when the PM10 increases, the net
profit decreases.
According to the correlation analysis of PM10 and the total assets growth, dif-
ferent from the net growth rate, most regions show strong positive correlations.
That is, the total asset ratio will rise with the increase in PM10 . For exam-
ple, Jinan, Panzhihua, Tangshan, and Benxi show a strong positive correlation.
Jiangsu, however, shows a strong negative correlation.
The correlation coefficient between PM2.5 and the return is less than 0.2,
reflecting a weak correlation. Although some correlation coefficient of PM10 is
over 0.2, but they are less than 0.5, indicating its weak correlation with the
return. PM2.5 and PM10 , hence, impact the return of the steel companies.
1218 K. Liu et al.
Table 8. Correlations
Returns PM2.5 PM10

Pearson correlation 1 0.02 0.08
PM2.5 0.02 1 0.856
PM10 0.08 0.856 1
Sig. (1-tailed) Returns . 0.371 0.099
PM2.5 0.371 . 0
PM10 0.099 0 .
N Returns 263 263 263
PM2.5 263 263 263
PM10 263 263 263
Returns PM2.5 PM10
Returns Pearson correlation 1 0.02 0.08
Sig. (2-tailed) 0.742 0.197
N 263 263 263
PM2.5 Pearson correlation 0.02 1 .857a
Sig. (2-tailed) 0.742 0
N 263 275 275
PM10 Pearson correlation 0.08 .857a 1
Sig. (2-tailed) 0.197 0
N 263 275 275
a
Correlation is significant at the 0.01 level (2-tailed).

This paper, based on a theoretical analysis, argues that air quality may affect
stock market participants through emotions, policies, and expectations, ulti-
mately impacting the steel companies’ rate of return.
The results of this study show that severe air pollution has a negative effect
on the profitability of the steel companies. The air pollution is not only an
environmental problem, but also an economic and financial issue. Therefore, it
is imperative for people to tackle air pollution urgently. The steel companies
should increase the public transparency of information, enhance the develop-
ment of corporate social responsibility activities, improve the corporate image.
Air pollution event occurs, investors should keep rational judgment and choice
ability, make relevant rational trading behavior. Relevant government depart-
ments should reinforce stringent emission standards and formulate strict envi-
ronmental policies to control the high pollution and high emissions behavior to
address the deteriorating air quality and ease the increasing public concern.
The research enriches the literature regarding the influence of investor senti-
ment on the capital market, and provides the empirical estimation of the impact
of air pollution on the Chinese capital market. It has certain reference value for
the relevant government departments in terms of policy evaluation and decision-

making.
Acknowledgements. This research was supported by System Science and Enterprise

Development Research Center Project (Grant No. Xq16C12), Sichuan University Cen-
tral University Research Fund project (Grant No. skqy201653), Soft Science Project
of Sichuan Science and Technology Department (No. 2017ZR0033) and Sichuan Social
Sciences Research 13th Five-Year annual planning project, (NO. SC16A006).
References
1. Banerjee AV (1992) A simple model of herd behavior. Q J Econ 107(3):797–817
2. Cao M, Wei J (2005) Stock market returns: a note on temperature anomaly. J
Bank Finance 29(6):1559–1573
3. Chen YA, Tan S (2005) Estimation of the steady inflation rate of economic growth.
Econ Res J 4:002
4. Esposito P, Patriarca F et al (2013) Economic convergence with divergence in envi-
ronmental quality? Desertification risk and the economic structure of a mediter-
ranean country (1960–2010). Mpra Paper 102(4):715–721
5. Evans GW, Jacobs SV et al (1987) The interaction of stressful life events and
chronic strains on community mental health. Am J Community Psychology
15(1):23–34
6. Fong WM, Toh B (2014) Investor sentiment and the max effect. J Bank Finance
46(3):190–201
7. Kamstra MJ, Kramer LA, Levi MD (2003) Winter blues: a sad stock market cycle.
Am Econ Rev 93(1):324–343
8. Lepori GM (2015) Air pollution and stock returns: Evidence from a natural exper-
iment. J Empir Finance 35:25–42
9. Levy T, Yagil J (2011) Air pollution and stock returns in the US. J Econ Psychology
32(3):374–383
10. Lucey BM, Dowling M (2005) The role of feelings in investor decision-making. J
EconSurv 19(2):211–237
11. Mehra R, Sah R (2002) Mood fluctuations, projection bias, and volatility of equity
prices. J Econ Dynamics Control 26(5):869–887
12. Oberndorfer U (2006) Environmentally oriented energy policy and stock returns:
an empirical analysis. ZEW-Centre for European Economic Research Discussion
(06–079)
13. Slovic P, Finucane ML et al (2007) The affect heuristic. Eur J Oper Res
177(3):1333–1352
Low Carbon-Oriented Coupling Mechanism
and Coordination Model for Energy Industry
Investment and Financing of China
Yanfei Deng1 and Lei Xu2(B) and Ning Hou3

1
School of Management, Southwest University for Nationalities,
2
School of Economics, Xihua University,
leihsu@163.com
3
Faculty of Engineering, Imperial College London, London, UK
Abstract. Based on the perspective of low carbon and greenhouse gas

emission reduction, this paper utilizes the synergy theory to analyze the
coupling and coordination characteristics of China’s energy investment
and financing, and argues that energy investment and energy financing
constitute a composite system and find that it has a stress effect on the
climate environment, and the climate environment has a constraint effect
on energy investment and financing. According to the analysis of the
coupling relationship between the two, the energy industry investment
and financing coupling coordination model and evaluation index system
are established. By revealing the development law and evolution trend
between the two, we will actively promote the coordinated and coordi-
nated development of economic growth, energy supply and climate and
environment, and provide new research ideas and analysis methods.
Keywords: Renewable energy · Investment and financing · Coupling

mechanism · Coordination model
1 Introduction
In recent years, affected by the combined effects of resource endowment con-
straints and economic growth, China’s energy production and energy consump-
tion is dominated by coal, greenhouse gas emissions are rising rapidly has become
the first emissions of global carbon dioxide and sulfur dioxide [5]. In recent years,
China’s atmospheric pollutant emissions continue to increase, the composite air
pollution has become increasingly prominent, sulfur dioxide, nitrogen oxides and
volatile organic compounds led to the second pollution was intensified, which
include Fine particulate matter, ozone, acid rain and so on. The formation cause
of air pollution is complex energy production and processing conversion is one of
the important source. So it is imperative to realize energy saving and emission
Low Carbon-Oriented Coupling Mechanism and Coordination Model 1221
reduction and low-carbon development in the energy industry. Energy indus-

try investment and financing is to regulate and control the energy industry to
improve the level of its clean production, “the total valve”, both the coupling
coordinated development will help to accelerate the existing technology and
equipment to upgrade and clean production projects and technology operation,
to fundamentally reduce the energy consumption, reduce industrial emissions, to
promote the coordinated development of energy, economy and environment, the
purpose of climate. Investment in energy industry and energy industry financing
interaction coupling relationship as follows: Energy industry investment through
the production of investment and pollution control investment in two aspects of
stimulating financing development, energy industry financing through the sup-
port of the above two aspects to promote investment expansion. When the energy
industry investment expansion too fast and beyond the ecological environment
carrying capacity limit, the government based on the reality of environmental
degradation will significantly enhance the policy constraints of environmental
protection and greenhouse gas emission reduction. Under the comprehensive con-
trol of industrial policy, credit policy and environmental protection policy, the
energy industry financing will continue to shrink and restrict the energy industry
investment. Instead, consistent with the national industrial policy, climate and
environment protection policy of renewable energy projects in the production,
energy production source pollution control and cleaner production projects will
receive financing inclination, thereby stimulating investment in areas related to
energy industry development. When the pollution control of energy production
of excessive growth of investment, the economic principle of optimal use of funds
is not consistent, the development of energy industry investment will be blocked
due to the lack of funds, climate and ecological environment protection also lost
the support of the construction. With the rapid growth of economy, the increas-
ingly intensified climate environment has become an important bottleneck in
the development of energy economy. This paper is based on the macro policy
background of greenhouse gas emission reduction from the theory of synergetic
and coupled coordination evaluation method the energy industry investment and
energy industry financing can be considered as two subsystems, in the analysis of
the two coupling based on the coordination of the relationship, the establishment
of energy industry investment and financing coupled coordination model and the
order parameter evaluation index system. By revealing the law of development
and evolution trend of the two, to provide a new research ideas and analysis
methods for the effective treatment of economic growth energy supply and cli-
mate environment deterioration of the prominent contradictions and actively
promote the coordinated development of coupling.
Coupling coordination theory is mainly used to study economic manage-
ment in technology innovation and technology management, industry cluster
and industry chain and ecological environment and regional economic problems
[1,3,6,7]. American scholar Weick [8] first used coupling theory to study eco-
nomic and social problems, he explained the relationship between the school
members contact each other but keep each other independent through loose
1222 Y. Deng and L.N. Xu Hou
coupling theory. Norgaad [4] put forward the theory of coordinated develop-
ment, think through the feedback loop and can achieve the coordination and
common development between social system and ecological system. Guneralp [2]
constructed the important infrastructure and the national security, the ecologi-
cal environment, the economical coupling effect model and so on. The domestic
scholars based on the measure method of coupling coordination, researches on
the problems of ecological environment and economic development coordination.
Binding energy industry investment and financing characteristics, we believe
that they may exist coupling interaction and coordinated operation mechanism
and the coupling coordination theory and method shall also apply in the energy
industry investment financing in the field of research on the coordination mecha-
nism. In the plane of the sustainable development, if the interaction of investment
in the energy industry and the energy industry financing has a higher level of
coordination of the coupling of, there will be helping to improve the efficiency
of the compound system, and ultimately the economy energy low carbonization
and intensive development.
2 Key Problem Statement

Coupling as a physical concept, the extent to which two (or more than two)
systems or forms of motion are affected by each other. For example, between
two connected to a spring pendulum, they shake each other ups and downs,
mutual influence, this interaction is called coupling pendulum. Similarly, you can
put two sub investments in energy industry and energy industry financing sys-
tem through their respective coupling elements produce interaction and mutual
influence of the phenomenon is defined for the coupling of the investment in the
energy industry and the energy industry financing, both of which constitute the
energy industry investment and financing system. Energy industry investment
subsystem as the guide and engine of energy economic development, determines
the quality, scale and benefit of the energy industry financing subsystem, and at
the same time, it is influenced by the movement of the order parameter of the
financing of the energy industry. The mutual coupling situation of the invest-
ment and financing of the energy industry is decided by the investment and
financing gap, which is equal to the time accumulation of investment minus the
financing. Any investment factors that affect the investment and financing gap
will contribute to the change of the financing. Under the background of green-
house gas emissions reduction, investment in the energy industry in addition
to the category of traditional energy industry investment should also cover for
prevention and control of pollution source, newly started construction projects
“three simultaneous” project, renewable energy development, carbon capture
and sequestration (CCS) and integrated gasification combined cycle technology
(IGCC) and other new production projects and processing technology of the
investment. The productive capacity of these investments will contribute to a
significant reduction in industrial emissions and greenhouse gases, and improve
the quality of regional climate and environment. Information related to invest-
ment will continue to feedback to departments of environment and trade market,
through policy intervention and price signals to attract financing support, and
achieving the energy industry investment and financing coupling coordinated
development goals.
In the energy industry investment subsystem, the long-term implementation
of the energy industry investment government approval, corporate finance, bank
loans investment and financing model, the lack of flexible, effective and clean
investment and financing mechanism. This type of investment is the subject
of the behavior of state-owned energy enterprises, national investment of this
kind of enterprises lack of supervision method and risk control means, lack of
supervision and evaluation mechanism of enterprise investment decision, enter-
prise project approval lack of benefit and risk evaluation mechanism. Especially
in the oil and gas, electric power and other industries, to a large extent there
are administrative intervention, unclear definition of the functions of the govern-
ment, macro-control offside, resulting in energy companies are often blind invest-
ment behavior. In the energy industry financing subsystem, in recent years, the
energy industry financing channels tend to be diversified, but still in the enter-
prise internal self financing and bank credit as the main. Commercial banks in
the energy companies monopoly profits and national climate and environmental
protection policies to make a difficult choice between. Leading to the ups and
downs of the size of the finance. As shown in Fig. 1, the energy industry invest-
ment and financing gap has increased in recent years. At the same time, by
stimulating the impact of energy production and consumption, energy industry
investment growth rate increased year by year. Overall, the interaction between
the two is weakened, and the coordinated development is facing challenge.
Fig. 1. Energy industry investment and financing, energy production and the evolution
of energy consumption
The investments of oil and gas extraction, petrochemical, coking and nuclear
fuel processing industry are unstable, lack of financing support. Overall, the
energy industry internal financing coupling degree is not high.
3 The Coordination Model and Index System

According to the theory of synergetic, system always has a spontaneous irregu-
lar independence movement, while at the same time by the interaction of other
subsystems, the connection between the seed system formed a cooperative move-
ment. In the cooperative movement, there are many control parameters, which
are divided into ”fast variable” and “slow variable”, that is, the order parameter
is in the dominant position. With the constant change of the order parameter,
the correlation between the subsystems is gradually increased before the critical
point of the system; when the order parameter reaches the “threshold”, plays
a leading role in the association between subsystems system from disorder to
order, from the chaos produced one kind of order structure. It can be said that
the order parameter is the most prominent sign of the qualitative leap before
and after the phase transition, which is about the characteristics and rules of the
whole system transformation, find the order parameter of system is equivalent
to find the observation of interaction between composite system movement and
each subsystem of the key. Under the background of the greenhouse gas emission
reduction and environmental protection policy, the energy industry investment
subsystem has a multi-dimensional meaning, mainly including the existing pro-
duction project pollution sources of investment, the new construction project
“three simultaneous” investment, resource reserves and climate environmental
protection policy, and so on four order parameters. Energy industry financing
system is subject to monetary policy, energy industry financing and energy enter-
prise financial situation and so on three sequence parameters of the integrated
drive. As a result, the coupling relationship of the energy industry investment
and financing system is the sum of the various nonlinear relationships of the
interaction between the order parameters. In order to eliminate the correla-
tion between the subsystems within the control parameters, in this paper, the
extraction and control parameter based, determine the order parameter esti-
mated weight and measurement subsystem of order parameter, and then the
composite system coupling coordination degree was evaluated.
3.1 The Theoretical Model of Investment and Financing in the

Energy Industry
The coupling degree is used to describe the interaction degree of the energy
industry investment and financing system, and to judge the degree of coordina-
tion and the stage of the coupling effect. First, establish the efficacy function.
xij are the first I index of the control parameter of J, and xij is the function
value of the function after the standardization. αij and βij are the upper and
lower limit of the critical point order parameter of the system stability. Thus,
“investment in the energy industry, energy industry financing coupled system
orderly efficacy coefficient xij can be expressed as follows:

(xij − βij )/(αij − βij )
xij = (1)
(αij − xij )/(αij − βij ),
where xij is the variable xij contribution to the system’s function, which reflects
the degree of satisfaction of the control parameters to achieve the goal, its range
of values for [0, 1], 0 for the most dissatisfied, 1 for the most satisfactory.
If M (j = 1, 2, m) order parameter is extracted from the H control parame-
ters in a subsystem U (k = 1, 2), the order parameter matrix (Fij )n×m with
n(i = 1, 2, · · · , n) is formed. Because the investment in energy industry and
energy industry financing are two different but interacting subsystems within
the system, to achieve each order parameter of the order degree by the inte-
grated method, the model is as follows:
m
m

uk = θj × Fij , θj = 1, (2)
j=1 j=1
where θj are order parameter weights. Using the concept of capacity coupling
and coefficient model for reference, the energy source is obtained, industrial
investment and financing coupling degree function is:
1
C = {(U1 × U2 )/ [(U1 + U2 ) × (U1 + U2 )]} 2 , (3)
where the letter C is the system interaction coupling degree, C ∈ (0, 1); U1
and U2 represent energy industrial investment sub-system and the energy indus-
try finance system on the system total contribution, namely investment in the
energy industry comprehensive sequence parameters and energy industry financ-
ing order parameters.
For the energy investment and financing system, the significance of the cou-
pling degree model is: quantitative description the interaction of between the two
subsystems; reflect the order parameter of each subsystem in a certain period
of time, the relationship between the number of regional interaction and adjust-
ment process, in order to provide a basis for evaluation of the evolution trend of
the interactive coupling of the composite system.
3.2 The Coordination Model of Investment and Financing in the

Energy Industry
Coordination of each subsystem in the evolution process and factors of each
part of the composition of different germplasm (order parameter), in which a
unified whole showed mutual cooperation and harmony of property. If the order
parameter of each subsystem is coordinated, it can promote the coordinated
development of the whole complex system. In this paper, the difference of each
subsystem in the composition of the composite system, the harmonious and
consistent attribute is defined as the energy industry investment and financing
Association scheduling. With respect to the coupling degree model, the coordi-
nation degree model considers the factors that influence the development speed
of the overall order parameter of each subsystem, which makes up the coupling
degree model only focusing on the horizontal coupling, and ignores the prob-
lem of the speed of development. It can be more comprehensive and accurate
to judge the degree of coupling coordination between regional energy industry

investment and financing system:
√
D =√C × T
(4)
T = ρ1 U1 × ρ2 U2 ,
where D is the coordination degree, D ∈ (0, 1); C is the coupling degree; T is a

comprehensive coordination index of energy industry investment and financing
system, reflect the overall energy industry investment and financing synergy; ρ1
and ρ2 are undetermined coefficient; U1 and U2 are respectively the comprehen-
sive order parameter of the energy industry investment and the order parameter
of the energy industry financing. Coordination degree can be divided into 3 sec-
tions and 8 stages. As shown in Table 1 shows, through the comparison of the
investment and financing system comprehensive order parameter, the sample
can be roughly divided into three types: energy industrial investment excessive
type, energy industry financing advanced type and investment and financing
balanced development type, the following research will accordingly be identified
and evaluated.
Table 1. The coupling coordination system and criteria of energy industry investment
and financing
According to the connotation and characteristics of the investment and

financing system of the energy industry, the comprehensive measurement index
system is determined according to the principle of integrity, hierarchy and oper-
ation, see Table 2.
Table 2. Energy industry investment and financing coupling coordination system index
system
Subsystem Control parameter Order parameter Comprehensive order

parameter
Energy industry Total investment in Existing production Comprehensive order
investment subsystem environmental pollution project pollution source parameter of investment
control (100 million yuan) control investment subsystem
Actual implementation of
the “three simultaneous”
project number
The actual
implementation of the
“three simultaneous”
project investment (100
million yuan)
The actual
implementation of the
“three simultaneous”
project environmental
protection investment
(100 million yuan)
Investment in energy New construction
industry (100 million project “three
yuan) simultaneous”
investment
Source of industrial
pollution control projects
completed investment in
the year (100 million
yuan)
Basic reserves of energy Resource reserve
(100 million tons of
standard coal)
Carbon intensity of Regulation of climate
energy industry environmental
protection policy
Financing subsystem of Total energy industrial Financial situation of Comprehensive order
energy industry assets (100 million yuan) energy enterprises parameter of financing
subsystem
Main business income of
energy industry (100
million yuan)
Total profit of energy
industry (100 million
yuan)
Energy industry should
pay value-added tax (100
million yuan)
RMB loan balance (100 Monetary policy
million yuan)
RMB loans added (100
million yuan)
Balance of energy Financing of energy
industry loans (100 industry
million yuan)
4 Summary
Based on the system theory and synergetic theory, this paper establishes a
coupled coordination model based on low carbon sustainable development and
greenhouse gas emission reduction to analyze their interactions. The main con-
clusions are obtained as follows: (1) The investment and financing of the energy
industry has a stress effect on the climate environment, and the climate environ-
ment has a binding effect on the investment and financing of the energy indus-
try. The energy industry investment and financing stress effects on climate and
environment are mainly: the energy industry investment and financing of energy
production capacity increase the emissions of air pollutants; the constraint effect
of the climate environment on the investment and financing of the energy indus-
try is mainly caused by the severe climate and environmental protection policy.
(2) Under the conditions of the policy of reducing greenhouse gas emissions,
the interaction between the energy investment and financing has weakened year
by year. In considering the greenhouse gas reduction factors, energy industrial
greenhouse gas emissions in areas of high energy investment and financing cou-
pling is relatively low, relatively poor coordination; Energy development and
construction may not be an important cause of regional complex atmospheric
pollution; The climate and environment protection policy constraints on energy
investment is not strong, but has an important impact on energy financing.
Acknowledgement. This research is supported by the Program of National Natural

Science Foundation of China under Grant [No. 71503216] and [No. 71401136], the
Social Science Planning Project of Sichuan Province under Grant [No. SC15C051], the
Central Universities Project under Grant [No. 2016NZYQN13], the Technology Finance
and Entrepreneurship Financial Research Center Project under Grant [No. JR201507],
and the System Science and Enterprise Development Research Center Project under
Grant [No. xq15c02].
References
1. Du J, Bo Y, Yao X (2012) Selection of leading industries for coal resource cities
based on coupling coordination of industry’s technological innovation. Int J Min Sci
Technol 22(3):317–321
2. Gneralp B, Seto KC (2008) Environmental impacts of urban growth from an inte-
grated dynamic perspective: A case study of shenzhen, south china. Global Environ
Change 18(4):720–735
3. Li Y, Li Y et al (2012) Investigation of a coupling model of coordination between
urbanization and the environment. J Environ Manage 98(1):127
4. Norgaard RB (1990) Economic indicators of resource scarcity: a critical essay. J
Environ Econ Manage 19(1):19–25
5. Smale R, Hartley M et al (2006) The impact of CO2 emissions trading on firm
profits and market prices. Climate Policy 6(1):31–48
6. Tang Z (2015) An integrated approach to evaluating the coupling coordination
between tourism and the environment. Tourism Manage 46:11–19
7. Wang SJ, Ma H, Zhao YB (2014) Exploring the relationship between urbanization

and the eco-environmentla case study of beijing-tianjin-hebei region. Ecological Indi-
cators 45(5):171–183
8. Weick KE (1976) Educational organizations as loosely coupled systems. Adminis-
trative Sci Q 21(1):1–19
Space-Time Analysis for Water Utilization
Efficiency of Urban Agglomeration and Its
Economic Influence Factors llin Chengdu Urban
Agglomeration
Yeyun Huang1 and Yunqiang Liu2(B)

1
College of Economics, Sichuan Agricultural University, Chengdu 611130,
2
College of Management, Sichuan Agricultural University, Chengdu 611130,
liuyunqiang@sicau.edu.cn
Abstract. This paper uses the DEA method to establish the evalu-
ation model of water resources efficiency to make an empirical study
on the urban agglomeration of Chengdu city in 2008–2014. The results
show that the efficiency is increasing year by year, and the economic
efficiency and environmental efficiency of the water resources utilization
are improved. Meanwhile, under the control of the technology, educa-
tion and transportation level to analyze of efficiency from the economic
factors, this paper gets the conclusion that: (1) the city water use effi-
ciency and the level of economic growth showed a significant linear cor-
relation. (2) Technology shows no obvious linear relationship with the
efficiency of water use. Traffic enhances the effect of openness on water
efficiency. Education is negatively correlated to the efficiency. It indicates
that the urban agglomeration should enhance the traffic infrastructure
on its regional economic advantage, optimizing the industrial.
Keywords: DEA data envelopment · Water use efficiency · Urban

agglomeration · Economic development · Econometric analysis
1 Introduction
With the economic development and population growth, water shortage is a
severe problem. According to the National Bureau of statistics report, Sichuan
water consumption for each industrial added value dropped from 135 m3 per
million yuan in 2008 to 46 m3 per million yuan in 2014, a decline of 66%, achiev-
ing the requirements of the State Council in advance. However, compared with
Beijing, Suzhou and other “national water-saving city” in recent years, which
maintain the value of 16–19 m3 per million yuan, the efficiency of water use needs
improvement.
Space-Time Analysis for Water Utilization Efficiency 1231
The shortage of water resources is a global problem, and how to make efficient
use of water to ease the supply and demand of resources is urgent. The coordi-
nated development of economy and environment is the foundation of building a
resource-saving society. On the one hand, the main branches of the economy have
an impact on the efficiency of water use [2], on the other hand, the improvement
of water use efficiency is conducive to the local food exports [7].
Therefore, the quantitative analysis of economic development with water
use for sustainable development is particularly important. Previous studies had
focused on the evaluation of the efficiency of water use on agriculture, industry
and mixed model, or on the profound the model. The present studied the use of
water in the economic output perspective, ignoring the ecological impact, this
paper aims to evaluate the Chengdu city group water utilization through two
aspects of economic and environmental efficiency. Few existing research attached
quantitative analysis of factors affecting the efficiency of water use, this paper
based on the empirical analysis study the efficiency of water use and its economic
factors, prolonging sustainable development for the construction of water-saving.
2 Research Objectives
There are many ways to analyze the utilization of water resources, in view of the
water potential of the unified understanding degree is not high. The range of fluc-
tuation is difficult to quantitative analysis of the decoupling theory. This article
based on the perspective of the utilization efficiency of water resources to study
the water use in the city of Sichuan province. This article unfolds as follows.
Primarily, In order to explore the relationship between water use efficiency and
economic development, this article based on the Cobb-Douglas production func-
tion, to set the input-output model to measure the efficiency of water resources
utilization. Secondly, according to the difference of water use in time and space,
this paper divides the comprehensive utilization efficiency of water resources into
economic efficiency and environmental efficiency. Thirdly, this article analyzes
the impact of economic development on the efficiency of water use, to discover
the relationship between economic development and water consumption in urban
agglomeration. Lastly, In order to reflect the economic indicators in a better way,
this paper under the control of the city of scientific technology, education and
water transportation level, to do further analysis between economic level and
efficiency of water use. It’s of great importance to profoundly analyze economic
development and resource allocation. Figure 1 shows the construction.
This paper chooses BBC model to analyze the relative efficiency of variable
returns to scale. The BBC model increases the convexity assumptions on the
basis of the CCR model, which will be helpful to analyze the environmental and
economic output [4,8,9]. We add the curve measure to the BBC model. The
curve measure is the environmental efficiency evaluation method of nonlinear
analysis, analyze the output with radial infer, use reciprocal curve to measure the
pollutants, in order to achieve efficiency evaluation in the process of increasing
the output and reduce the pollution.
1232 Y. Huang and Y. Liu
Opening
degree
Economic
Consumption
structure
Urban
Transportation Education Technology
Water
Capital Population
resource
Economic Comprehensive Environmental

efficiency efficiency efficiency
GDP Sewage
Fig. 1. The construction of the paper
In order to fully reflect the input and output of water resources in cities, the
paper based on the Cobb-Douglas production function, constructed an input
output analysis framework including water use and water pollutant emissions.
The exact input-output index system followed in Table 1. The fixed assets invest-
ment of the whole society reflects the scale of capital investment of the whole
society in a period. The local population, compared to the employment popula-
tion, better expresses the contribution to the increase in water consumption and
economic output. We use the annual water supply as the water consumption
of each grade city. All the data comes from the Sichuan statistical yearbook.
This paper adds city sewage emissions as undesirable outputs. In case of pollu-
tion and other undesirable products are produced, based on the previous [3,4]
for pollutants treatment, this paper inserts the reciprocal evaluation which has
comparative advantage to DEA model as the environmental impact.
Table 1. Input-output index system
Input Water resource Cities’ water consumption

Capital size Social investment in fixed assets
Population Current population
Output Economic output Current GDP
Environmental effects Current sewage discharge
Table 2. The unit root test for panel data
Structure Export Consumption

Statistic −5.0904 −5.7898 −1.30E+02
Z −4.5241 −3.7001 −4.9723
Prob. 0 0 0
3 Space-Time Analysis for Efficiency
The paper uses the DEAP 2.0 to set the model, lists the results in Table 3. It is
obvious from Table 3 The water use efficiency of Chengdu urban agglomeration
increased gradually in 2008 to 2014. The water utilization efficiency of Ziyang
in 2008 to 2012 showed the optimal level 5 times, indicating that the capital
investment, labor and water resources allocation in Ziyang city is in a good
condition and take the best place among Chengdu urban agglomeration. We use
ArcGIS to demonstrate the variation of the comprehensive efficiency.
Table 3. 2008–2014 Chengdu urban agglomeration water use comprehensive efficiency
2008 2009 2010 2011 2012 2013 2014

Chengdu 0.76 0.69 0.81 0.87 0.9 0.92 1
Deyang 1 0.63 0.84 1 0.98 0.99 1
Mianyang 0.72 0.53 0.62 0.75 0.79 0.83 0.87
Leshan 0.81 0.73 0.76 0.9 0.89 0.86 0.9
Meishan 0.71 0.66 0.72 0.85 0.81 0.81 0.85
Yaan 0.65 0.73 0.64 0.8 1 1 1
Ziyang 1 0.86 0.67 1 1 1 1
On Fig. 2, these directly show the trend of water utilization efficiency. From
the perspective of time series, the financial crisis in 2008 affected the economic
development and water use of Chengdu urban agglomeration which led to the
tendency of water utilization efficiency declined in 2009. In 2011, with economic
recovery and city industrial construction, water utilization efficiency and the
indirect influence gradually recovered to the level of 2008.
From the perspective of the spatial distribution of the industry, Chengdu
urban agglomeration pillar industries (electronic information, water and elec-
tricity, pharmaceutical and chemical industry, machinery, metallurgy, food and
beverage, etc.) of large enterprises were mainly distributed in Chengdu, Deyang,
Mianyang, Ziyang, However, the development of industry in Meishan, Leshan
lagged behind, which lacked the support for large enterprises and projects.
Fig. 2. Chengdu city group water use comprehensive efficiency from 2008–2014
This paper analyzes the comprehensive utilization efficiency of water

resources from two aspects of economic efficiency and environmental efficiency.
Environmental efficiency reflects the scale is insufficient or redundant. Based
on the conventional expectations output, we adds the new sewage emissions as
undesirable output. It not only reflects the moderation between water supply
quantity and capital investment, but also reflects the reduction of potential pol-
lutant emission under the same factor input and output conditions [6,10]. It is
of great help for the various regions to establish future planning objectives in
the existing economic and environmental conditions. Economic efficiency refers
to the existing economic input, the best use of technology or the ability to deal
with pollutants. It embodies the ability to achieve the maximum output under
the conditions of a variety of input factors. In this paper the DEA-BBC input-
output model, the economic efficiency reflects the investment of fixed assets and
water resources social investment which leads to the ability to achieve the max-
imum economic output in the certain number consumption of a city. Analysis
of the economic efficiency and environmental efficiency of water resources in
Chengdu urban agglomeration by DEAP 2 is shown in Fig. 3. Based on the time
series, it is obvious that 2009 was a turning point due to the financial crisis in
2008. In terms of the city’s water efficiency, Yaan’s economic efficiency had been
maintained at a high level. The reason is that the economic development of Yaan
lagged behind by the state in the Chengdu urban agglomeration. Its economic
output and degree of industrialization were low, thus embodying water utiliza-
tion efficiency good situation in the input and output. Ziyang and Deyang kept a
stable efficiency at the high level in 2011. Deyang’s pillar industry was mechani-
cal, which accounted for nearly half share. Although the development of machin-
ery industry was still in the extensive stage, under the atmosphere of 2011–2014,
tap estates had a greater contribution to pull the local economic output. Ziyang
on the location was the unique regional center city which connected Chengdu
and Chongqing the “dual core”. It’s known as its automobile industry the same
as Deyang, which drove the rapid economic development through increasing
industrial output then indirectly improved the economic efficiency. Compared to
Mianyang, the city manifested in industrialization as well, it existed an obvious
gap. Mianyang played a relatively low ecoefficient role among Chengdu urban
agglomeration with its electronic as pillar industry. The falling behind velocity
of the increasing development restricted its economic efficiency to a higher level.
Although Mianyang regarded as the second biggest city in Sichuan, the lowest
water environmental effect dragged it at the bottom and pull down its water
utilization comprehensive efficiency. Similarly, Chengdu, which was specialized
in electronics, medicine and so on, because it was the urban agglomeration cen-
ter city and the capital city of Sichuan Province, its economic development was
much higher than that of other prefecture level cities which raised its water use
efficiency. In terms of the environmental efficiency of water use, Chengdu urban
agglomeration showed the effective in environmental efficiency but relatively low
valid in eco-efficiency. The reason is that it didn’t make the best use of factors.
For illustration, Yaan and Meishan did not serve its turn. Meanwhile, the plenty
of cities had already taken a big scale, such as Chengdu, Deyang, Leshan, Ziyang
and Mianyang. In general, the input-output ratio was reasonable, and the com-
bination of factors achieved a certain scale economy. Therefore, it is close to the
optimal efficiency on the scale of the city but insufficient in eco-efficiency.
1 1
0.9 0.9 Chengdu
0.8 0.8 Deyang
0.7 0.7 Mianyang
0.6
0.6 Leshan
0.5
2008
2009
2010
2011
2012
2013
2014
2008
2009
2010
2011
2012
2013
2014
Eco. Meishan
Env.
eff Yaan
eff.
Year Year Ziyang
Fig. 3. 2008–2014 Chengdu urban agglomeration water use Economic and Environ-
mental efficiency
4 Economic Effects
4.1 Hypothesis
Based on the above empirical, it is visible that economic output and
its level of city development has close relationship with water utilization.
This paper assumes that the explanatory variables are: economic structure,
opening degree, consumption level. At the same time, this paper selects the
technology, transportation and education level as the control variables.
Hypothesis 1 : economic structure has a negative impact on water use efficiency.
The economic structure is the composition and structure of the national econ-
omy. The influence of economic structure on water resources utilization efficiency
can be analyzed from two aspects. On the one hand, the industrial production
have positive effects on the economy of a region, the economic structure in the
industry tend to have higher economic output which shows high utilization effi-
ciency of water. On the other hand, the increase in industrial investment will
increase the impact of the ecological environment, resulting in a larger proportion
of environmental output, thereby reducing the city’s water use efficiency. Now
the negative effect of the development of Chengdu city group is the ecological
environment influence over its economic effects is unknown.
Hypothesis 2 : the degree of opening to the outside world has a positive impact
on the efficiency of water resources utilization.
The degree of opening effects the import of resources and the introduction of new
technologies. Chengdu is located in China’s western region, the western energy
technology itself is relatively backward. The change in the degree of dependence
on foreign trade has great influence on the west. Therefore, the study on the
degree of opening to the outside world is of great significance to the utilization
efficiency of water resources in Chengdu urban agglomeration.
Hypothesis 3 : consumption level has a positive impact on water use efficiency.
To some extent, the consumption level reflects the scale of production and indi-
rectly affects the efficiency of resource utilization. On the one hand, consump-
tion is an important booster of economic growth. On the other hand, all kinds
of garbage generated in the process of consumption also makes the decline in
environmental quality [1]. Whether the ecological impact of rising consumption
level is beyond its economic effects? This paper assumes that the consumption
level is a positive impact on water use efficiency.
4.2 Analysis of Indicators
This paper will explain the following three aspects on the selection of economic
factors and the setting of variables: As explained variable, utilization represents
the regional comprehensive efficiency of water use. The data is calculated by
the DEA-BBC model. Other data comes from the 2009–2015 Sichuan statistical
yearbook. For Explanatory variables, structure refers to the economic structures,
this paper uses the sum of the primary and secondary industries with the ratio
of the service industries to express the economic structure. Set the ratio due
to the water consumption of primary and the secondary industriesshare more,
the service industrial (including commodity trade, catering and accommodation,
transportation organizations and other services industry) accounted for only 3%
(Sichuan water resources bulletin 2014). In terms of the gross national product
structure, in the Chengdu urban agglomeration, the primary industry occupies
less compared to the secondary, the service industries. Therefore, use the sum
of primary and secondary industrial added value divided by the added value of
the third industry to represent the economic structure and investment allocation
direction. Consumption indicates the level of consumption, this paper selects the
total retail sales of social consumer goods to express the transaction situation
of various organs, enterprises and institutions, in order to reflect the domestic
consumption level and demand of the city. Export is used to indicate the degree
of opening up. In this paper, the ratio of total import and export to GDP is
used to reflect the degree of opening up. The city’s total import and export of
goods better embodies the foreign trade level. Because of its joint venture, said
the unit is million, taking into account other indicators to billions of units, the
direct use of the total import and export data analysis model will have a certain
deviation, and the change of exchange rate is floating, it is difficult to calculate
by import and export the amount of US dollars into RMB. This paper selects
the import and export volume and the current ratio of GDP to reflect the effect
of the level of international trade.
Table 4. Definition and description of variables
Variables Definition Description Effect

Utilization Water utilization efficiency Local comprehensive −
efficiency of water use
Structure Economic structure (Primary + +/−
Secondary)/(Service)
Consumption Level of consumption Total retail sales of social +/−
consumer goods
Export Degree of opening up (Import + +
Export)/(Current GDP)
Technology Level of technology The ratio of technology +
expenditure
Education Level of education The ratio of education +
expenditure
Transportation Level of transportation The ratio of +
transportation
expenditure
For the control variables, this paper use the proportion of public finance
expenditure to reflect the levels. The city’s technology investment has positive
effect on improving the efficiency of water resources utilization. The progress of
technology can create more water-saving facilities, increase the circulating water,
thereby increasing the utilization efficiency of water resources. The educational
level of the city reflects a city’s progress and the civilized degree. The level of
urban transportation will affect the degree of opening up. At the same time,
the improvement of the level of transport will also play a role in alleviating the
imbalance of water resources distribution [5], so as to improve the utilization

efficiency of water resources (Table 4).
4.3 Model Description
The econometric model of panel data includes fixed effect model, random effect
model and least square regression method. The equitation is as followed.
Utilization = b1 + b2 × Structure + b3 × Export + b4 × Consumption + ε, (1)
where b1 stands for the intercept, b2 , b3 , b4 represent the coefficient of the eco-
nomic structure, the degree of opening up and the social consumption. ε is the
stochastic error. Because of the regional difference of the urban agglomeration
is relatively small, the difference of the water resources utilization efficiency
between different sections of the cross section can be ignored. The time span
is small, and the variation of parameters can be ignored as well. In this paper,
WLS regression is used to fit and analyze the influencing factors and weights,
and the control variables are introduced gradually.
Utilization = b1 + b2 × Structure + b3 × Export + b4 × Consumption

(2)
+ b5 × Technology + ε,
(3)
+ b5 × Technology + b6 × Transportation + ε,
+ b5 × Technology + b6 × Transportation (4)
+ b7 × Education + ε,
where b5 , b6 , b7 represent the effective coefficients of the technology, transporta-

tion and education towards the water use efficiency.
(1) Test of stationary

In order to avoid spurious regression, and ensure the validity of the estimation
results, we must check the stability of each panel sequence. We use Stata 11.0
to perform unit root test for short panel data. The sample capacity is small, the
effect of the unit root test for single variable is weak, thus we use the LLC test
to analyze the data. The original hypothesis H0 has the same unit root. The test
results are shown in Table 2, the data has a high degree of significance, so the
result is to reject the original hypothesis. The perturbation terms of different
individuals are independent of each other, and the data are stable.
(2) Results and analysis
According to the significance of the data, the mixed OLS model is better than the
fixed effect and random effect by Stata 11, and the Table 5 is the regression fitting
result of the three models. It was the primary and the secondary industry high
proportion improve the utilization efficiency of water use. In addition, the further
openness would introduce new energy, advanced technology and talents, and it
could largely reduce economic output and has high impact on the ecological
environment. The regression coefficient of the level of consumption is significantly
positive. We use stepwise least squares regression, in the case of abandoning the
constant, to analyze the standardized coefficient.
Table 5. Comparison of regression coefficients
Before the insert of control variables After the insert of control variables
Pooled Fixed Random WLS + Technology + Transportation + Education
OLS effect effects
b2 0.217a 0.230a 0.2259a 1.041a 0.404 0.379 0.700a
b3 4.162b 4.421 4.3012b 0.532b 0.671a 0.939a 1.253a
a b a b b a
b4 0.001 0.001 0.001 0.637 0.981 0.607 0.542a
b5 − − − − −0.1 − −
b6 − − − − − 0.418a 0.471a
b7 − − − − − − −0.338a
2
R 0.412 0.411 0.411 0.407 0.354 0.503 0.577
F 55.31 6.61 − 10.759 6.995 12.895 13.649
a b
Correlation is significant at the 0.01 level, Correlation is significant at the 0.05 level.
From Table 5, it is obvious that the economic structure takes the greatest
place among the economic indicators to demonstrate the trend of the water
use efficiency. The reason is that the Chengdu urban agglomeration has a higher
degree of industrialization. The domestic consumption promotes economic diver-
sification of functions of the city through the sharing of resources among cities,
which indirectly improve the utilization of water. Meanwhile, international trade
will bring the inflow of technology and talent, and promote the development of
economy.
In this paper, the coefficient of technology is negative as control, and it didn’t
pass the significance test. This doesn’t mean that there is no linear relationship
between the technology and the efficiency of water use. It is based on the tech-
nology proportion of public expenditure. In fact, many of the city’s scientific
development depend not only on the public finance expenditure, it depends on
technological innovation in enterprises and large companies. The traffic level
shows that increasing its proportion of public finance expenditure has a positive
effect on improving the efficiency of water use. Under the influence of the traffic,
the effect of the economic structure weakened, and the effects of opening up rise
up. The coefficient of educational level is negative.
(3) Heteroscedastic test

In order to check the deviation between the fitting value and the real value
of the model and to ensure the regression parameter estimation has good sta-
tistical properties, the paper makes use of the residual graph to illustrate the
heteroscedasticity test. From the residual plot of heteroscedasticity test that, in
general, the fitting degree of the model is good. The fit of economic structure is
good, the scattered distribution is random and uniform (Fig. 4).
Fig. 4. The residue of structure, consumption, openness and fitted values
5 Conclusion
The water use efficiency of Chengdu urban agglomeration has difference in time
and space. Attach the time series as the axis, since the financial crisis into the
economic buffer period, the efficiency of water use was gradually increasing. From
the perspective of spatial difference, because of its geographical advantages, in
the electric, automotive and other industries as a pillar industry in the city,
its water use efficiency was better. For some cities, although the marginal scale
investment was higher, due to its serious environmental pollution and relatively
weak economic growth, it shared the low utilization efficiency of water use. The
data shows that the water use of Chengdu urban agglomerations still had some
room to improve, the government departments should make efforts to increase
sewage treatment.
The water utilization efficiency from the perspective of economic level analy-
sis, found the water resource utilization and economic development is linear, the
economic structure takes the most profound effect. Increasing the utilization rate
of resources depends largely on the green industry output; the opening degree
and city’s consumption affected the input of technology and talent to indirectly
affect the efficiency of water use. Meanwhile it directly affect the water use
through the flow of resources among the agglomeration.
This paper makes a further analysis on the urban development, and finds that
the level of urban transportation has a positive effect on the efficiency of water
resources utilization and will increase the effective proportion of the degree of
opening up. Raising the level of education in the short term is not conducive to
the water use. Due to the long period of payback in the educational investment,
it requires for a higher levels of openness to compensate the effect for extra
education spending of the utilization efficiency of water resources.
Acknowledgements. This study is supported by Ministry of Education Humanities

and Social Sciences Youth Project (15YJC630081); Sichuan Provincial Social Science
Research “12th Five-Year” planning project (SC14C027); Sichuan Provincial Institute
of systems science and Enterprise Development Research Center (Xq14B08).
References
1. Kox HLM (2013) Export decisions of services firms between agglomeration effects
and market-entry costs. Service industries and regions. Springer, Heidelberg, pp
177–201
2. Kulmatov R (2014) Problems of sustainable use and management of water and
land resources in Uzbekistan. J Water Resour Prot 6(1):35–42
3. Liao H (2011) Utilization efficiency of water resources in 12 western provinces of
China based on the DEA and malmquist TFP index. Resour Sci 33(2):273–279
4. Mai Y, Sun F et al (2014) Evaluation of China’s industrial water efficiency based
on dea model. J Arid Land Resour Environ 11:008
5. Malaeb L (2004) Constituent transport in water distribution networks
6. Melo PC, Graham DJ, Noland RB (2009) A meta-analysis of estimates of urban
agglomeration economies. Reg Sci Urban Econ 39(3):332–342
7. Taheripour F, Hertel TW et al (2016) Economic and land use impacts of improving
water use efficiency in irrigation in South Asia. J Environ Prot 7(11):1571–1591
8. Thanassoulis E (2000) Dea and its use in the regulation of water companies. Eur
J Oper Res 127(1):1–13
9. Veettil PC, Ashok A et al (2011) Sub-vector efficiency analysis in chance con-
strained stochastic dea: an application to irrigation water use in the Krishna River
Basin, India. In: The 122nd Eaae Seminar “Evidence-Based Agricultural And Rural
Policy Making: Methodological And Empirical Challenges of Policy Evaluation”
10. Wu F, Liu Z (2008) Research on the mechanism of how city group drive eco-
nomic growth-empirical evidences from 16 cities of Yangtze River delta. Econ Res
J 11:126–136
Exploring Linkages Between Lean and Green
Supply Chain and the Industry 4.0
Susana Duarte(B) and Virgilio Cruz-Machado
UNIDEMI, Departamento de Engenharia Mecânica e Industrial,

Faculdade de Ciências e Tecnologia,
Universidade Nova de Lisboa, 2829-516 Caparica, Portugal
scd@fct.unl.pt
Abstract. A new paradigm with a considerable influence on the indus-

trial value creation is begun to evolve. Industry 4.0 represents the devel-
opment to the fourth industrial revolution. A new shift in business and
technology trends is confirmed. With this new paradigm of high-tech
industry, supply chains will develop into highly adaptive networks. Lean
and Green supply chain management should have better opportunities
to be performed, becoming more efficient and competitive. The aim of
this study is to link the lean and green supply chain characteristics to
the Industry 4.0. This paper has twofold: first, presents an overview of
the phenomenon Industry 4.0 and the lean and green supply chain man-
agement. Next, a conceptual model is developed which incorporates the
Industry 4.0 topics to the well-known lean and green supply chain. This
paper provides an understanding of the role of lean and green paradigms
in the new era of industrialization.
Keywords: Industry 4.0 · Lean paradigm · Green paradigm · Supply

chain
1 Introduction
Today we face a new paradigm introduced by Germany in 2011, the Indus-

try 4.0. This represents the beginning of the fourth industrial revolution and is
driven by modern information and communication technology (ICT) [21,26,27].
By Industry 4.0 intends the optimization of value chains by implementing an
autonomously controlled and dynamic production [21], through a full automa-
tion and digitalization processes. This paradigm is based on the idea that com-
munication via internet allows a continuous exchange of information applying
cyber-physical systems (CPS) [27,30]. The CPS provides the source for the cre-
ation of internet of things and services and their combination makes possible
the Industry 4.0 [15,27]. CPS integrates networking, computation and physi-
cal processes and take them together to create a global value chain networked
[15,27,30].

Exploring Linkages Between Lean and Green Supply Chain 1243
According to [34] to establish the global value chain networks, the Industry
4.0 describes a production oriented CPS that integrates production facilities,
warehousing and logistics systems and even social requirements. In addition,
Germany Trade & Invest (GTAI) [15] mention that the industrial value chain,
product life cycles and business information technology combination must inte-
grate the processes from the product design to production, supply chain man-
agement, aftermarket service and training [15]. An intelligent factory is in devel-
opment and is coined as smart factory. With this concept, others appear and
are important for the establishment of implementation of Industry 4.0, as for
example smart products, smart manufacturing and smart data.
The authors [27] mentioned that this implementation is still in progress.
That’s why it is important to understand the role of lean and green supply chain
management in Industry 4.0. For example, the lean waste would be recognized
with the smart factory implementation [28]. The resource efficiency which is a
lean and green concept, are in the focus of the design of smart factories [27].
This study intends to understand if the Industry 4.0 allows the lean and green
supply chains concepts become more important. That is, if it enables more easily
the deployment of lean and green characteristics. A number of characteristics
were presented on model, namely: (i) manufacturing, (ii) logistics and supply,
(iii) product and process design, (iv) product, (v) customer, (vi) supplier, (vii)
employee, (viii) information sharing and (ix) energy.
The remainder of this paper is organized as follows: in Sect. 2, a theoretical
background on Industry 4.0 and lean and green supply chain are presented; in
Sect. 3, a combination between lean and green supply chain and Industry 4.0 is
developed; Finally, some concluding remarks are drawn.
2 Industry 4.0 Paradigm
The Industry 4.0 is considered the paradigm of the fourth stage of industri-
alization and describes a vision of future production [21,30]. The core idea of
Industry 4.0 is the integration and application of information and communica-
tion technologies to implement Internet of Things and Services so that business
process and engineering process are deeply integrated making an environment
intelligent [28,34]. The concept of industry 4.0 which represents the integration
of the virtual and physical worlds in a way that together create a truly networked
environment and where intelligent objects communicate and interact with each
other, is a Cyber-physical systems [15]. According to [19] the Industry 4.0 “will
involve the technical integration of CPS into manufacturing and logistics and
the use of the Internet of Things and Services in industrial processes. This will
have implications for value creation, business models, downstream services and
work organization.”
The Industry 4.0 is represented by three features [19,30,34]: (i) horizontal
integration across the entire value networks; (ii) vertical integration and net-
worked manufacturing system; and (iii) end-to-end digital integration of engi-
neering across the entire value chain or product life cycle.
1244 S. Duarte and V. Cruz-Machado
The horizontal integration across the entire value network refers to the inte-
gration of the various systems used in the different stages of the manufacturing
and business planning processes that involve an exchange of materials, energy
and information both within a company as logistics, production, and marketing,
and between several different companies; The idea is that information, mate-
rial and money can flow easily among different companies creating new value
networks as well as business models. This can result in an efficient ecosystem
[19,30,34].
The vertical integration refers to the integration of the various information
and physical systems at the different hierarchical levels, as for example the pro-
duction management, manufacturing and execution, and corporate planning.
This integration is inside a factory to create flexible and reconfigurable manu-
facturing system [19,30,34].
The goal of the horizontal and vertical integration is to deliver an end-to-end
solution. The end-to-end solution refers to the digital integration of engineering
across the entire value chain to support product customization: from the raw
material acquisition to manufacturing product, and product in use and in the
end of life [19,30,34].
Through these features, the Industry 4.0 expects to implement an environ-
ment more flexible, efficient, and sustainable. The idea is to individualize the
customer requirements, as a customized product through a mass customization,
improving productivity and achieving higher levels of quality with a manufac-
tured profitably result [6,19]. Indeed, by applying advanced information and
communication technologies and systems in the manufacturing and supply chain
operations, the industry 4.0 addresses the smart factory [28]. Smart factory is
designed according to sustainable and business practices, insisting upon, flex-
ibility, adaptability and self-adaptability, learning characteristics, fault toler-
ance, and risk management [15]. Therefore, standards are essential to ensure the
exchange of data between machines, systems and software and guarantee that
product moves within a network value chain [6].
That is, high levels of automation come as standard [15]. Automation sys-
tems, manufacturing and product management are integrated and are the base
of the smart factory [6]. Manufacturers can now add sensors and microchips
to tools, materials, machines, vehicles and buildings to communicate with each
other in real-time to make smart products [15].
According to [36], “products know their histories and their routes, and
thereby not only greatly simplify the logistic chain but also form the basis
for product life cycle data memories”. Also, the products can be manufactured
because smart factory is being supplied with energy from smart grids [30].
Not only smart factory and smart product were defined in this new industri-
alized era. Others concepts connected to them are considered in the literature.
For their work development, Kolberg and Zuhlke [21] considered four different
smart concepts to define the smart factory, namely, smart planned, smart prod-
uct and smart machine, and smart operators. The authors [30] make mention
of the smart grid, smart logistics and smart data. Sanders et al. [28] mention
Table 1. Concepts of Industry 4.0
Concepts Description
Smart factory Smart factory represents the key characteristic of Industry 4.0 [15]. The
smart factory will be more flexible, dynamic and intelligent [27], where
people, systems and objects communicate with each other [15]. The
internet of things and services are the main enabler technology for a
smart factory [15, 29]
Smart manufacturing Manufacturing will be equipped with sensors and autonomous systems
which allow that operations can be optimized with a minimum
employee’s intervention [27, 29]. It produces small-lot products of
different types, more efficiently [34]
Smart product A smart product is a product with sensors and microchips that allow
communication via the internet of things with each other and with
employees [27]. It holds the information about its requirements for the
manufacturing processes and manufacturing machines [21, 30]
Smart logistics It is one of sustainable mobility strategies [15]. Smart logistics will use
CPS for carrying the material flow within the factory and in the supply
chain (between factories, customers and other stakeholders) [30]. The
transport equipment is a part of smart logistics that is able to react to
unexpected and autonomously should be able to drive between the
starting point and the destination [30]. Distribution and procurement
will increasingly be individualized [27]
Smart engineering Includes product design and development, production planning and
engineering, production and after sales service [29]
Smart data Smart data is structured information of data that can be used for
decision-making [30]
Smart machine Machines and equipment will have the ability to improve processes
through an intelligent decision-making, instead of being directly
instructed [27, 34]. The smart machines should have additional
autonomy and sociality capabilities to adapt and reconfigure to
different types of products [34]
Smart planner Smart Planner optimizes processes in real-time [21]. Decentralized
self-organization [27]
Smart operator smart operator is an employee who supported by ICT, control and
supervise ongoing activities [21]. Employees can be quickly directed to
the right tool [6]
Smart customer Customers’ needs and behaviors are analyzed in real-time in way to
provide them with new and more sustainable products and services [29]
Smart supplier Based on factory needs it is possible to select the best supplier (which
allows higher flexibility) and strengthen a sustainable relations with
suppliers (by increase information sharing in real-time) [29]
Smart grid Responsible to supply energy to a factory [30]. Energy management [15]
Smart energy Monitor and provide feed-back on energy production and use [23]
others concepts as the smart systems, smart environment, smart machine and
smart devices, and smart task. Table 1 compiles several concepts of Industry 4.0.
Through the integration of the industry concepts and technologies it should
be possible provide a customized or individualized product or service and at the
same time be highly adaptive to demand changes [15]. These changes must be
made on all stages of product life cycle: design phase, raw material acquisition
phase, manufacturing phase, logistics and supply phase, and the use and end of
life phases [15,30]. Therefore, the requirements for design and operations of our
factories become crucial for the success [36].
3 Lean and Green Supply Chain Paradigms
Nowadays, lean and green supply chain is an integrated approach; they have
different objectives and principles but they complement each other [7,9,10,12–
14,31]. Lean supply chain is about to increase value for customers by adding
product or service features, with the elimination of waste or non-value steps
along the value chain [11]. Green supply chain regards to reducing environmen-
tal impacts and risks while improve ecological efficiency of the organizations
and their partners, and try to achieve corporate profit and market share objec-
tives [35].
These two paradigms are often seen as compatible because of their joint
focus on waste reduction [5]. Lean paradigm is concerning to the elimination of
waste in every area of design, manufacturing, and supplier network and factory
management [13]. The basic forms on the reduction and elimination of waste
are [17]: production, waiting, transportation, unnecessary inventory, inappropri-
ate processing, defects and unnecessary motions. One more waste is pointed by
[31] as the unused employee creativity. Green considers ways to eliminate waste
from the environment’s perspective [11]. The waste generation have the form of
[16]: Greenhouse gases, eutrophication, excessive resource usage, excessive water
usage, excessive power usage, pollution, rubbish and poor health and safety.
In their research [12] mention that the two paradigms have the same type of
wastes: (i) inventory; (ii) transportation, and (iii) the production of by-product
or non-product output. According to [5] the removal of non-value adding activ-
ities suggested by lean paradigm can provide substantial energy savings which
integrates the principles of green paradigm.
The combination of lean and green supply chain practices have better results
than the total from the implementation of each, but separately [12]. The two
paradigms have similar characteristics. According to [4] both paradigm prac-
tice contribute for: (i) the increase of information frequency, (ii) the increase
of the level of integration in supply chain, (iii) the decrease of production and
transportation lead time, (iv) the reduction in the supply chain capacity buffers;
(v) and the decrease of inventory levels. Another practice that contributes for
the better employ and use of all tools is the involvement of the employees [12].
Both paradigms look into how to integrate product and process redesign in order
to prolong product use, to allow easily the recycling or re-use of products, and
to make processes with less wasteful [12]. In the supply chain both paradigms
ask for a closed collaboration with partners [22]. In addition, waste reduction,
lead time reduction, and use of techniques and approaches to manage people,
organizations, and supply chain relations are synergies mention by [13].
Commitments must be made within factory, supplier network and customer,
for the better deployment of lean and green practices in way to achieve the best
supply chain efficiency. In the authors previous study [11] it was presented a table
Table 2. Lean and green paradigms characteristics in supply chain
Characteristics Lean Green

Philosophy focus Long-term thinking [2] Long-term thinking since
environmental impacts affects the
natural environment for many years
[18]
Market focus Serve only the current market Demands from at least some
segments with predictable demand customer segments for more
[33] environmentally friendly practices
[35]
Manufacturing focus Lean manufacturing use Lean Green manufacturing focuses on the
Automation [28]. Lean manufacturing environment and identifies waste as
focus on maximizing productivity by excessive use of resources or
increasing output per unit of input, substances released to the air, water
conserving resources, reducing waste, or land that could harm human
and minimizing costs [7] health or the environment [24]
Product design focus Product design sets the conditions Green design or Eco-design; Selection
for manufacturing: process and of low impact materials, reduction of
product development [18]; Eliminate materials usage, optimization of
waste, increase productivity and production techniques, improvement
reduce cost [7] of distribution system, reduction of
impact during usage, improvement of
initial life time, and improvement of
end of life system [18]
Inventory strategy Minimizes inventory throughout the Reductions in inventory levels [4]
chain [4]
Supply focus Inter-organizational involvement [22]. Inter-organizational collaboration
Close cooperation with all supply involves activities to deploy green
chain entities [18] issues in supply chain entities [22]
Information sharing Demand high levels of information Information sharing is important for
sharing [22] the successful implementation [22]
Employees Employee involvement and Involvement of employees [18]. The
empowerment [12]. The education education of the employees
and training of the employees to concerning green issues [12]
execute multiple tasks [2]
Customer Satisfying the customers by reducing Satisfying the customers by the
costs and lead times [12] implementation of green issues [12]
Processes Focus on product [2]. Continuous Focus on processes and products [18].
improvement culture [24] Process optimization through the
implementation of green issues [31]
comparing the different characteristics between lean and green. Others important
studies [3,4,12,18,22] were inspire for the development of a comparison between
lean and green paradigms. Several lean and green supply chain characteristics
are considered in Table 2.
4 Combining Industry 4.0 and Lean and Green Supply

Chain
The future must become leaner in organization accurately in planning and tech-
nology [36]. The authors [21] mentioned that the paradigm Industry 4.0 solutions
not only can be integrated in lean manufacturing but can be beyond that improve
lean manufacturing. Also Sanders et al. [28] considered that industry 4.0 and lean
manufacturing can be integrated to achieve a successful production management.
However, they are not mutually exclusive [28].
Definitely, several researches mention the benefits of the integration of lean
and green in different stages of the company or the supply chain [1,9,12,13,20].
Kainuma and Tawara [20] studied the lean and green supply chain incorporating
there cycling or re-use during the life cycle of products and services. It represents
the different phases of a product life cycle, consisting in [20]: (i) the acquisition of
the raw material, (ii) the manufacturing, (iii) the distribution, (iv) the retailer,
(v) the use, (vi) the collection, (vii) the transportation, (viii) the dismantling
and (ix) the decomposition.
Stock and Seliger [30] presented the opportunities for the realization of a
sustainable manufacturing in the Industry 4.0. For them the life cycle (in the
end-to-end solution) consists in different phases [30]: (i) the raw material acqui-
sition, (ii) the manufacturing, (iii) the transport (between all phases), (iv) the
use and the service phase, and (v) the end-of-life phase (containing the reuse,
remanufacturing, recycling, recovery and disposal). In addition, the environ-
mental/green dimension of the sustainability is better considered because the
allocation of resources as products, materials, energy and water can be realized
in a more efficient way [30]. The adoption of smart energy systems facilitates the
energy use [23].
In fact energy models would assist the analysis of green factory designs,
especially for evaluating alternatives during early design stages [25]. The design
of lean and green supply chain, special in the early design stages for the products
and the processes is a very important issue for the elimination of waste. In a lean
and green environment [7] mention that “eliminating the use of toxics through
product or process re-design could mean reduced worker health and safety risks,
reduced risks to consumers and lower risk of product safety recalls and reducing
process wastes in manufacturing often find more opportunities to reduce waste
throughout the life cycle of the product, thereby having a possible domino effect
on the entire supply chain”. Industry 4.0 is in line with these ideas. According to
[27], Industry 4.0 processes will change the entire supply chains, from suppliers
to logistics and to the life cycle management of a product. It helps to streamline
the process, with more transparency and flexibility.
Lean and green supply chain requires manufacturing technologies to make
processes and products more environmentally responsible [22]. In addition they
ask for a flexible information system [17]. The technology is a driver of the
Industry 4.0 [15,34]. With smart technologies which include the use of electronics
and information technologies [27] will help the implementation of a more efficient
lean and green supply chain.
Also the collaboration with suppliers which is a lean and green characteristic
is considered by Industry 4.0. Through a better communication mechanisms,
with a high compatibility issues of hardware and software which should required
standardized interfaces, and synchronisation of data, allow that lean and green
suppliers get better synchronisation with manufacturers [28,32].
The author [14] concludes that lean and green “is an effective tool to improve
processes and reduce costs, by not only reducing non-value-added activities but
also physical waste created by systems”. Industry 4.0 is in line with this state-
ment due this paradigm make all but in a better way, more sustainable, faster
and efficient. According to [21] lean allows the organizations to be more stan-
dardized, transparent and having only the essential work which result in an
organization less complex and support the installation of industry 4.0 processes
and solutions. The green also support the implementation of the Industry 4.0
due it allows to reduce the negative environmental impacts.
The customer type is a concern in the lean and green supply chain. Of course
that lean and green aim is to satisfy the customer needs, but this satisfaction
is relative to: in the lean paradigm is based on cost and lead time reduction
[9] and in the green paradigm is based on helping customers to being more
environmentally friendly [13]. The Industry 4.0 will go to improve in this subject.
It allows a better understanding of the customer needs and allows the immediate
sharing of the demand data throughout complex supply chains [15]. According to
[27] with the full automation and digitalization systems, it allows an individual
customer-oriented adaptation of products that will increase the value added for
organizations and customers. Customers instead of choose from a fixed product
spectrum set by the manufacturer, they will be able to individually combine
single functions and components and define their own product [15].
Another characteristic of lean and green supply chain is the employee involve-
ment and empowerment [12]. According to [8] employee commitment and moti-
vation, and employee empowerment and participation are elements of lean
and green organization. Also [24] mention that connections between lean and
green practices are shown through: (i) employee involvement, (ii) learning by
doing, (iii) continuous improvement, and (iv) problem-solving tools. [36] men-
tion that lean means reducing complexity, avoiding waste and strictly supporting
the employees in their daily work. Also the reduction of environment impacts
improves the health and safety of employees [31]. These aspects are in line of
what is an employee in the four industrial revolution. Indeed, employees may
find greater autonomy and more interesting or less arduous work [6]. Industry
4.0 needs employees not only with creativity and decision-making skills, encoun-
tered as a lean and green supply chain characteristic, as well as technical and
ICT expertise [6].
There are in literature some studies that try to make the bridge between lean
paradigm and green paradigm with Industry 4.0. [28] used 10 lean concepts in
their research in way to validated for attainability through Industry 4.0 para-
digm. Kolberg and Zühlke [21] described the lean automation and Industry 4.0
and give an overview of the links between them. [30] present an overview of sus-
tainable manufacturing with the future requirements of Industry 4.0. Figure 1
illustrates an attempt to link lean and green supply chain characteristics to
Industry 4.0 concepts.
Lean and Green Supply Chain Industry 4.0 Concepts
Manufacturing Smart Manufacturing
Logistics and Supply Smart Logistics
Product and Process Design Smart Engineering
Product Smart Product
Customer Smart Customer
Supplier Smart Supplier
Employee Smart Operator
Information Sharing Smart Data
Energy Smart Energy
Fig. 1. Linking the lean and green supply chain characteristics to the Industry 4.0
concepts
5 Conclusion
Today, the term Industry 4.0 describes a vision of future of the supply chains.
There is a strong conviction that the definition of lean and green supply chain
will not disappear, it will be evolve and adapt to the new trends that the new
industrial era will require. Lean and green supply chain is focussed on organiza-
tion and in the flow of information, material and money between partners. That
is, more directed to physical processes and less for virtual and technology. Even
so there are in literature some examples that try to make the bridge between
lean paradigm or green paradigm with Industry 4.0. This paper bridges the gap
between the well-known lean and green supply chain management and the new
era of industrial revolution.
A conceptual model was developed linking the lean and green supply chain
characteristics to the Industry 4.0 concepts. Several characteristics were pre-
sented on model, namely: (i) manufacturing, (ii) logistics and supply, (iii) prod-
uct and process design, (iv) product, (v) customer, (vi) supplier, (vii) employee,
(viii) information sharing and (ix) energy. Those who understand the relation-
ships between these two topics will have a greater chance of influencing their
supply chains into a source of competitive advantage and help in a better way
on the deployment of the Industry 4.0 paradigm.
Future research is needed. Understand which lean and green characteristics
are more important for the development of Industry 4.0 is required. It would be
also beneficial to understand the priority between characteristics on the imple-
mentation of this new paradigm and in different entities in the supply chain.
Industry 4.0 will be a step forward for the effectiveness and competitiveness of
the lean and green supply chains.
Acknowledgements. Authors would like to acknowledge to UNIDEMI–Research and

Development Unit for Mechanical and Industrial Engineering (UID/EMS/00667/2013).
References
1. Azevedo SG, Carvalho H et al (2012) Influence of green and lean upstream supply
chain management practices on business sustainability. IEEE Trans Eng Manage
59(4):753–765
2. Bortolotti T, Boscari S, Danese P (2015) Successful lean implementation: organi-
zational culture and soft lean practices. Int J Prod Econ 160(12):182–201
3. Carvalho H, Azevedo SG, Cruzmachado V (2010) Supply chain performance man-
agement: lean and green paradigms. Int J Bus Perform Supply Chain Model
2(3):304–333
4. Carvalho H, Duarte S, Machado VC (2011) Lean, agile, resilient and green: diver-
gencies and synergies. Int J Lean Six Sigma 2(2):151–179
5. Carvalho H, Azevedo S, Cruz-Machado V (2014) Trade-offs among lean, agile,
resilient and green paradigms in supply chain management: a case study approach.
Lect Notes Electr Eng 242:953–968
6. Davis R (2015) Industry 4.0, digitalisation for productivity and growth. European
Parliamentary Research Service (EPRS), Members’ Research Service, European
Union
7. Dhingra R, Kress R, Upreti G (2014) Does lean mean green? J Cleaner Prod 85:1–7
8. Duarte S, Cruz-Machado V (2013) Lean and green: a business model framework.
Lect Notes Electr Eng 185:751–759
9. Duarte S, Cruz-Machado V (2013) Modelling lean and green: a review from busi-
ness models. Int J Lean Six Sigma 4(3):228–250
10. Duarte S, Cruz-Machado V (2015) Investigating lean and green supply chain link-
ages through a balanced scorecard framework. Int J Manage Sci Eng Manage
10(1):20–29
11. Duarte S, Machado VC (2011) Manufacturing paradigms in supply chain manage-
ment. Int J Manage Sci Eng Manage 6(5):328–342
12. Dües CM, Tan KH, Ming L (2013) Green as the new lean: how to use lean practices
as a catalyst to greening your supply chain. J Cleaner Prod 40(2):93–100
13. Garzareyes J (2014) Lean and green-synergies, differences, limitations, and the
need for six sigma. IFIP Adv Inf Commun Technol 439:71–81
14. Garzareyes JA (2015) Green lean and the need for six sigma. Int J Lean Six Sigma
6(3):226–248
15. Germany Trade & Invest (GTAI) (2014) Industry 4.0-Smart Manufacturing for the
future. Germany Trade and Invest
16. Hines P (2009) Lean and Green, 3rd edn. Sapartner, New York
17. Jasti NVK, Kodali R (2015) Lean production: literature review and trends. Int J
Prod Res 53(3):867–885
18. Johansson G, Winroth M (2009) Lean vs. green manufacturing: Similarities and
differences. In: International euroma conference: implementation - realizing oper-
ations management knowledge
19. Kagermann H, Helbig J et al (2013) Recommendations for implementing the strate-
gic initiative german industrie 4.0. final report of the industrie 4.0 working group.
Technical report, Forschungsunion
20. Kainuma Y, Tawara N (2006) A multiple attribute utility theory approach to lean
and green supply chain management. Int J Prod Econ 101(1):99–108
21. Kolberg D, Zühlke D (2015) Lean automation enabled by industry 4.0 technologies.
IFAC Papers Online 48(3):1870–1875
22. Mollenkopf D, Stolze H et al (2010) Green, lean, and global supply chains. Int J
Phys Distrib Log Manage 40(1/2):14–41
23. Noppers EH, Keizer K et al (2016) The importance of instrumental, symbolic, and
environmental attributes for the adoption of smart energy systems. Energy Policy
98:12–18
24. Pampanelli AB, Found P, Bernardes AM (2014) A lean & green model for a pro-
duction cell. J Cleaner Prod 85:19–30
25. Prabhu VV, Jeon HW, Taisch M (2012) Modeling green factory physics - an ana-
lytical approach. In: IEEE international conference on automation science and
engineering, pp 46–51
26. Qin J, Liu Y, Grosvenor R (2016) A categorical framework of manufacturing for
industry 4.0 and beyond. Procedia Cirp 52:173–178
27. Roblek V, Meško M, Krapež A (2016) A complex view of industry 4.0. Sage Open 6
28. Sanders A, Elangeswaran C, Wulfsberg J (2016) Industry 4.0 implies lean manu-
facturing: research activities in industry 4.0 function as enablers for lean manufac-
turing. J Ind Eng Manage 9(3):811
29. Shrouf F, Ordieres J, Miragliotta G (2014) Smart factories in industry 4.0: A
review of the concept and of energy management approached in production based
on the internet of things paradigm. In: IEEE international conference on industrial
engineering and engineering management, pp 697–701
30. Stock T, Seliger G (2016) Opportunities of sustainable manufacturing in industry
4.0. Procedia Cirp 40:536–541
31. Verrier B, Rose B, Caillaud E (2015) Lean and green strategy: the lean and green
house and maturity deployment model. J Cleaner Prod 116:150–156
32. Veza I, Mladineo M, Gjeldum N (2016) Selection of the basic lean tools for
development of croatian model of innovative smart enterprise. Tehnicki Vjesnik
23(5):1317–1324
33. Vonderembse MA, Uppal M et al (2006) Designing supply chains: towards theory
development. Int J Prod Econ 100(100):223–238
34. Wang S, Wan J et al (2016) (2016), Implementing smart factory of industrie 4.0:
an outlook. Int J Distrib Sensor Netw 4:1–10
35. Zhu Q, Sarkis J, Lai KH (2008) Confirmation of a measurement model for green
supply chain management practices implementation. Int J Prod Econ 111(2):
261–273
36. Zuehlke D (2010) Smartfactory-towards a factory-of-things. Annu Rev Control
34(1):129–138
A Model of Maker Education
in Chinese Universities:
The Perspective of Innovation Ecosystem
Qinglong Zhan1(B) and Mengjia Yang2

1
School of Information, Tianjin University of Technology and Education,
Tianjin 300222, People’s Republic of China
qlzhan@126.com
2
Tianjin Nankai High School, Sino-Singapore Tianjin Eco-city,
Tianjin 300222, People’s Republic of China
Abstract. Maker education is an important mission of Chinese univer-

sities for the entrepreneurship and innovation. However, current maker
education failed to meet the makers’ demands for real-time awareness
of space, adaptive automatic resources, diverse innovation activities and
tools with intelligent support in the makerspace. Therefore, the inno-
vation ecosystem has emerged to be the best choice to solve those
problems. This paper first reviews literatures on maker education, and
then analyses fundamental principles and elements of the innovation
ecosystem. After that, essentials and properties are discussed for innova-
tion ecosystem-oriented maker (IEOM) education. At last, the model of
IEOM education is developed, including innovation subject, innovation
activity, and smart innovation environment.
Keywords: Maker education · Innovation ecosystem · Chinese

universities
1 Introduction
Maker education is an important mission of Chinese universities for the entre-

preneurship and innovation. However, current maker education in Chinese uni-
versities has the limitations of closed innovation, efficiency-driven, focal product,
specialized organization, and so on. They use a hierarchical management team
as the subject of innovation, and use virtual laboratories, heavyweight wearable
devices and other technologies to build an intelligent and well-defined physi-
cal space. This failed to meet the makers’ demands for real-time awareness of
space, adaptive automatic resources, diverse innovation activities and tools with
intelligent support in the makerspace. It also does not comply with their hopes
of cooperative innovation, wisdom communications and results-sharing. There-
fore, the innovation ecosystem has emerged to be the best choice to resolve

1254 Q. Zhan and M. Yang
those problems. At the same time, it needs to construct a model of innova-

tion ecosystem-oriented maker (IEOM) education as an important accelerator
of the sustainable development of university makerspace. In IEOM education,
any maker can be innovative as long as there are a true interest and consistency
in establishing a favorable environment to support it [4].
In this paper, we first review literatures on maker education, and then analy-
ses fundamental principles and elements of the innovation ecosystem. After that,
essentials and properties are discussed for innovation ecosystem-oriented maker
(IEOM) education. At last, the model of IEOM education is developed, including
innovation subject, innovation activity, and smart innovation environment.
2 Literature Review
Maker education of universities arises from the maker movement with the learn-
ing by doing culture. The maker movement has provided a perspective on learn-
ing that differs from the traditional learning practices taking place in schools and
universities [14]. Barrett, Pizzico, Levy and Nagel [1] discussed the benefits of
university makerspaces that are primarily focused around building physical mod-
els and the inherent of informal learning environments and community. Physical
models can increase the effectiveness and quality of the final design for the devel-
opment of undergraduates by linking the material covered in the classroom to the
real world. Informal learning environments are more open environment allows
them to be used more freely and interwoven into the class structure for multiple
classes without typical classroom scheduling constraints.
The best practices of university makerspace are investigated by several
researchers. Wilczynski [23] explored academic maker spaces in universities,
such as Arizona State University, Georgia Institute of Technology, Massachusetts
Institute of Technology, Northwestern University, Rice University, Stanford Uni-
versity, and Yale University. He recommended the best practices as follows: The
mission of the academic makerspace must be clearly defined; Ensure that the
facility is properly staffed; Open environments promote collaboration; Aligning
access times with the student work schedules; Providing user training is essential;
Attention must be devoted to establish a maker community on campus. Katona,
Tello, et al. [10] put forwards four best practices: flexible models for interdisci-
plinary faculty hiring and engagement, development of student entrepreneurs,
integrating cross-campus curricula, and the development of cross-campus col-
laborations. Myers [12] encouraged interdisciplinary research and idea, which
elements were student led engagement, access to the latest technology, and key
partnerships. Bieraugel [2] found that the on-campus makerspace located outside
the university library encouraged the most innovative behaviors and exploration
of new ideas, and within the library, collaboration rooms were the best spaces
for encouraging creativity.
As to Chinese researchers, their focuses are: relationships between maker edu-
cation and innovation education [8,25]; Essentials and functions of maker educa-
tion [27–29], double helix model [24] and design-based learning [9,30]; augmented
A Model of Maker Education in Chinese Universities 1255
virtual technology and Internet plus in maker education [19,20,27], development

paths of maker education [18,21] and system framework [26].
Obviously, innovation ecosystem that can be aligned with university maker
education currently seems to be not involved in the literatures. In fact, it is
a good idea to solve problems that development of maker education faces in
Chinese universities.
3 Theoretical Roots
The innovation ecosystem is a newer and mainstream concept discussed widely
by scholars and practitioners. However, innovation ecosystem is often described
in different ways, and thus understanding it remains a challenge. Luomaaho and
Halonen [11] defined innovation ecosystem as a permanent or temporary sys-
tem of interaction and exchange among an ecology of various actors that enable
the cross-pollination of ideas and facilitates innovation. Nordfors [13] referred
that innovation ecosystems embody technology and information flow between
those needed to turn ideas into processes, products or services. Bulc [4] pointed
out that an innovation ecosystem is a system made for innovation creation in an
open, natural manner, which enables a holistic understanding of needs, solutions,
and consequences related to innovation processes and innovation itself, and is an
interaction between people, enterprises, and institution. Ritala [15] had a view
that innovation ecosystem refers to clusters (physical or virtual) of innovation
activities around specific themes (e.g., biotechnology, electronics and software).
The complex innovation ecosystem where networks of innovations and commu-
nities of people and organizations interact to produce and use the innovations
[22]. Taken all together, Innovation ecosystem is an interactive work environment
which strengthens actors of different disciplines to co-create new ideas, and then
to construct prototypes and commercialize them in a way of ongoing innovation.
An innovation ecosystem has several important elements. The basic compo-
nents of the ecosystem must be a high quality (universities, funding possibili-
ties, specialized services, talent pool and regional dynamics) [7]. Bulc [4] thought
that key elements can be put into 4 major groups: participants, tools, content,
and principles. Haines [6] identified the following key ingredients of an innova-
tion ecosystem: culture, champion(s), network, stakeholder engagement, process,
physical space and events. Obviously, the innovation ecosystem includes funds,
material resources (equipment or facilities), and the innovative actors such as
students, teachers, researchers, industry specialists, venture capitalists. They
participate in the ecosystem to enable technology development and all kinds of
innovations.
Creating an innovation ecosystem needs to think about some basic princi-
ples: getting rid of the box, core value, competences, dynamic structure, and
constant change [4]. To make innovation, a suitable innovation ecosystem must
meet different conditions such as natural, structural, organizational and cul-
tural factors. Those factors can be grouped based on the following dimensions:
resources, governance, strategy and leadership, organizational culture, human
resources management, people, partners, technology and clustering [5]. A flow

of technology and information is the key factor among the actors: the people,
enterprises, and institutions [16].
4 Essentials and Properties of IEOM Education

4.1 Essentials of IEOM Education
IEOM education creates a new paradigm for a large-scale cooperative design, cre-
ation, invention, and innovation. It develops a novel way of intellectual associa-
tion, transparent operation, and collaborative participation in the innovation. In
this novel approach, creative ideas, services, and products can be implemented,
and the greatest value can be achieved.
IEOM education takes innovative actors as a community and utilizes new
technologies to create a real world of participatory innovation. These technolo-
gies are emotional computing, open source hardware, holographic environment,
3D printing, social media and adaptive technologies, etc. IEOM education is not
only the emergence of new technologies but also the emergence of transformative
innovation activities participated by makers. Firstly, due to the implementation
of the open source hardware and open resources, IEOM education provides mul-
tiple paths involved in innovation activities for makers full of creative thinking
and creativity. Secondly, makers can not only connect with people, machines
and objects to the physical space, but also seamlessly connect all the infor-
mation sources outside the space. Makers’ creative partners are also just from
members of a team to global actors in the project.
4.2 Properties of IEOM Education

(1) Wisdom circle
The wisdom circle is proposed to be a new concept in the development of the
biosphere, emphasizing the powerful forces of human wisdom in modern evolu-
tion. Human beings give full play to collective wisdom to explore new methods,
to create innovative technologies, and to develop new tools. They use created
products to change the makerspace, further enhance the intelligence of the group,
expand the scope of the role. Ultimately, the makerspace evolves into the human
wisdom circle containing capacity for development and creativity. The wisdom
circle relies on the rapid flow and superior connectivity of data, information
and wisdom. The flow and connectivity produce great transformation power,
and establish a fresh creative field of relating to wisdom individuals, enhancing
group wisdom and activating participatory innovation experiences.
(2) Micro-innovation
While micro-innovation is characterized by a small team, it shares the com-
mon trait with macro innovation: creative thinking based on recurring actions,
re-engineering an opportunity to make every thought new, and building cre-
ative solutions that bring true progress. Micro-innovation is a local action and a
global innovation. It is a decentralized, network-based innovation. Combinations

of indigenous micro-innovation will have a direct impact on the overall innova-
tion. Micro-innovation is not holistically planned and guided. In this sense, the
innovation is a daily habit. Micro-innovation encourages new technologies and
processes to be extended in IEOM education, and offers sustainable innovative
results for any size of the makerspace.
(3) Cross-border
As a creative scene, the makerspace in IEOM education is cross-organizational,
inter-spatial and interdisciplinary. It breaks the closed boundaries in order to
truly achieve the integration of innovation. Firstly, the makerspace is not only
beyond the visible border of physical space, but also is across the invisible bound-
ary of the digital space. Secondly, the wisdom flow and the information flow con-
tinually mobilize among the enterprises, research groups, project groups, inno-
vative communities and other organizations in IEOM education. Thirdly, the
acquisition of innovation is not confined to a specific discipline. Therefore, the
IEOM education emphasizes cross-border, connection and integration among the
different disciplines and spaces.
(4) High value-added
There are ongoing associations, high circulation and self-enhancement in IEOM
education. It is the circulation and re-creation of data and information, and envi-
ronmental innovation that produces the most valuable knowledge. For those data
and information of weak circulation and low re-creation, it is not able to create
leapfrog development and high valued innovation. Therefore, data is needed to
be reprocessed deeply, and the knowledge flow is needed to speed up. Further
to this, individual wisdom and the power of group innovation are converged to
achieve a high value-added and an upward spiral of innovation capabilities show-
ing the development of creativity. For example, scientific knowledge and big data
generated by activities, and deeper structure for mining methods, are combined
with language understanding and image recognition to provide value-added ser-
vices for makers.
(5) Strong lean
IEOM education of strong lean emphasizes scientific analysis of big data in the
creative process, and optimizes decision-making. Strong lean can improve the
performance and reduces the bias, with the aim of discovering potential prob-
lems, assessing the progress of creation, and predicting future performances.
Among them, lean thinking is scientific and accurate, thereby being efficient and
optimal. In strong lean, analysis technologies of data and information visualiza-
tion can be used to present the results of complexity analysis such as charts or
stereoscopic images. In doing so, relationships and patterns of things are identi-
fied quickly and easily.
(6) Multi-interactions
Multi-interactions include internal interactions between human-to-thing (H2T),
human-to-machine (H2M), thing-to-machine (T2M), human-to-human (H2H),
thing-to-thing (T2T). They provide more maker-oriented services for IEOM

education and meet the maker’s strong demands of highly interactive experi-
ences. The friendly multi-interactions are inseparable for designing human-thing
communication devices, machine operation objects, and human-computer inter-
action. It provides natural, convenient and efficient interactions among makers,
machine and things.
5 Design Principles of IEOM Education

IEOM education takes Web 4.0 as the basis and creation as being the engine
to establish the innovative ecosystem of the overall situation. Its mission is to
create an innovative ecosystem that integrates the various innovation activities
into IEOM education. The core is to create synergies among the innovative col-
laborators, such as industries, educational institutions and research institutions.
The paradigm is an innovative network which is composed of innovative actors.
IEOM education focuses on new technologies and innovation strategies. It
also concentrates on innovators, new technology adoption, cross-cutting issues,
learning organizations, forms of activity, and innovative business environments.
IEOM education balances process innovation and creativity by establishing inno-
vative space, and builds a network of innovative culture and a collaborative
innovation model. Thus by strengthening the creative culture and interdiscipli-
nary innovative solutions, IEOM education brings the integration of innovation
strategy and digital technology to an unprecedented level.
IEOM education produces a combination of innovative technologies and pro-
vides better user experiences of contextualization and value-added, supported
by digital technologies such as social media, mobile computing, big data, cloud
computing, sensor networking, Internet and machine intelligence. For example,
social media networks enhance coexistence; big data provides a unique analysis,
targeting the recommendations and preferences in the situations; mobile technol-
ogy creates location context services and awareness; cloud computing enlarges
access to resources and services; sensors create real-time feedback and response;
machine intelligence improves the quality of insight and decision-making.
The digital space is a new method to drive the innovative wisdom of indi-
viduals, organizations, and communities to establish a network of cross-industry
and actors’ innovation. It expands the ecosystem leading to innovation and cre-
ation. Digital space in IEOM education connects various things and the actors
to enhance the relevance and importance of digital technology. Through newly
connecting and integrating digital technologies aforementioned, the digital space
can enhance collaboration, and access the results of high value-added.
With the integration of digital technologies, IEOM education takes Internet
into the real-world, to form the human networking, things networking, and ser-
vice networking. This integration can configure technology, media, market and
actions of the actor altogether, and cross-links smart applications of real-world
products, services, and location by Web 4.0.
6 Building a Model of IEOM Education

IEOM education consists of three elements of innovation subject, innovation
activity, and innovation environment. Innovative subjects include experiencers,
innovators, researchers and experts, namely, the inner creative knowledge space.
Innovation activities are developed by the innovation subjects including inter-
est, research, practice and smart creation, namely, external creative knowledge
space. The innovation environment is a creative scenario of a whole new perspec-
tive, reflecting innovative subjects’ rich creation in participating in innovation
activities. The model of IEOM education is shown in Fig. 1.
Fig. 1. Development model of IEOM education
6.1 Innovation Subjects

(1) Experiencers
The experiencers are the subjects of intellectual creation, are the explorers of
the unknown world, and are the perceivers of facing the future. On the basis
of light capital and low investment, numerous experimenters learn, experience
and absorb in IEOM education, and then redesign and transform knowledge and
activities. In IEOM education, experiencers affect the development of the real
industry remodeling the physical world, to generate digital, open and popular
desktop manufacturing industry. Experiencers take the dexterous participation
and agile action to input new perceptions and ideas for the intellectual cre-
ation activities. At the same time, intellectual activities also provide feedback
on innovative activities for the experiencers.
(2) Innovators
In IEOM education, innovators with the similar interests gather together to
try unique combinations of multi-domain, multi-entity, multi-method and cross-
linking innovations. They blur boundaries, continuously redefine, and constantly
remodel works. Innovators connect IEOM education to nature and life in order
to design a unique perspective of the new works, such as technology + nature
+ art, digitization + virtualization + audiovisual performances, and other new
explorations.
(3) Researchers
Facing the unknown future, researchers use systematic thinking and scientific
methods to deepen the essence of things and the main contradictions. They cap-
ture new knowledge, generate new ideas, utilize new technologies, and explore
new approaches. Rather than the reality of unchanging model, researchers study
the possibility of the future to help the development with the IEOM education.
It is important for researchers to migrate and remix technologies, which are an
evolution path of new technologies. Similarly, the study of innovative activi-
ties can also stimulate the creative thinking and technological innovation of the
researchers.
(4) Experts
Experts emphasize the professionalism of their subject knowledge and the sensi-
tivity to complex things. They are not to solve problems according to the steps,
but to focus on asking questions, mining data, clarifying the relationship and
finding a solution. In IEOM education, the experts transcend over-fitting (self-
perceived vertexes) through cross-border connectivity and cyclic learning. At the
same time, on the basis of strong information filtering, experts show innovative
wisdom of research on product-updating, platform-upgrading, and entity-high-
value-reusing.
6.2 Innovation Activities
(1) Interest
Interest is the engine of makerspace development. Innovators gather in the mak-
erspace, generate creative ideas and innovative products in a more personalized
and independent way, based on aspirations, visions and interests. In IEOM edu-
cation, interest-based innovation activities, are not a clear bounded knowledge
or experience passed to the creators, but a creative enthusiasm for innovation,
social responsibility and other elements to affect makers. Innovation activities
emphasize cross-border, connectivity, collaboration and sharing, to enhance the
cohesion of makers’ innovations. They highlight the motive of exploration and
innovation, to encourage self-learning, decision-making and creativity. Innova-
tive activities show the powerful forces of innovative self-development and co-
creation through the de-centralization, peer organization. They offer high-yield
innovations with substantial value-added.
(2) Research
Research is the foundation of the development of IEOM education. Through the
unknown, undetermined, complex, chaotic and other advanced technology explo-
ration and the transfer of key technologies, research achieves changes and tech-
nological breakthroughs. As for the entities, products, and services, a researcher
can flexibly design, massively customizes and agilely develops. The aspects of
design, construction, and testing may be updated or changed iteratively.
(3) Practice
Makerspaces are central resource spaces offering various making practices in dif-
ferent contexts. Practice is an acceleration power of innovation. Practice is the
only way to produce knowledge creation and innovation. Creative activity in
practice refers to the fact that the experts absorb, interact, reflect, criticize and
recreate knowledge and products through the context-based collection, corre-
lation and integration. It stresses the unity of theory and practice, and shows
the great wisdom of innovation and creation. In IEOM education, cross-domain
experts carry out innovative ideas and actions to the practice of product develop-
ment. This practice appears in a circle of knowledge production and application
transformation.
(4) Smart creation
In industry, smart creation means to quickly and adaptively fabricate prod-
ucts using automatic, networked and intelligent information technologies such
as Internet of things. In IEOM education, smart creation can be all a mat-
ter of structural change, production optimization, efficiency enhancements and
value-added. It is decided on core skills, intellectual capital and other top fac-
tors of the practitioners. Smart creation is the dexterous behavior, which the
individual and the collective in the creative network culture focus on the tiny
creativity, the agile product development, the custom individuality service, and
the creative behavior reshaping. Smart creation demonstrates the interaction
and collaboration between makers by enhancing their dexterity behavior and
the creativity practice. With IEOM education, it increases the possibility of the
creative exploration, scientific attempt and innovative practice for makers.
6.3 Innovation Environments
(1) Cloud services of IEOM education

Hyper-scale data gathered in a creative environment is inseparable from the bear-
ing and calculation capacity provided by the cloud. Cloud supports the integra-
tion and sharing of large-scale resources and has a low requirement on equipment,
multi-channel access to resources, low-cost storage of large-scale data, and effi-
cient use of data resource to deliver long-tail stable service. Cloud has features
of on-demand access, ubiquitous access, resource pooling, elasticity, measurable
use, recoverability, high security and high scalability. These features of cloud
computing are essential to the reality of demand for IEOM education.
(2) Smart objects of IEOM education

As the cognized smart objects, they empower objects or things corresponding
cognitive abilities and reflect decentralized control features of the non-linearness,
openness and flatness. As an entity embedded in the autonomously distributed
intelligent system, smart objects can intelligently act on the human-machine-
thing objects based on the sensor’s perception in IEOM education. Multi-object
cooperation is able to globally optimize, to percept real laws and to solve the
major complexity of the problem by a distributed cooperative work.
(3) Big data of IEOM education
Big data in IEOM education aims at all the context data generated by the inno-
vation subject in the innovation activity. Through batch data processing and
unique perspective analysis of all the relational and non-relational data, it pro-
vides accurate forecast and personalized real-time suggestions for the innovation
subject. It can be a keen insight into the quality of data, data-mining in the
hidden information, and the true reflection of the state of activity. At the same
time, weak related data floods with sparse value density can still be appropri-
ately associated and deeply mined, in order to create a new situation and a new
development stage of IEOM education.
(4) Sensors of IEOM education
Sensors provide large amounts of real metadata for big data by detecting objects,
observing events, tracking targets, describing or quantifying activities and cap-
turing information. In IEOM education, the main function of sensors is per-
ception and cognition. The perception is the fact that the different sensing
elements perceive the context to obtain the parameters of the quantized and
non-quantized. Cognition is tantamount to acquire the state information or
knowledge, such as subjective emotion, psychological state, and another individ-
ual implicit information. Cognition can abstract the relationship of information
recorded in the communication, properties and location of devices and tools, and
others. Sensors own the functions of self-capturing, self-tuning, self-identification,
and logical judgments. They provide on-line or off-line production and establish
real-time information feedback, moving towards the depth of miniaturization
and high-level intelligence.
(5) Communities of IEOM education
Members of a community are technologically sophisticated do-it-yourself (DIY)
enthusiasts interested in activities related to engineering, electronics, robotics,
and desktop fabrication [3]. In the inter-connected times, the community recon-
structs the circle of interaction, re-establishes the role of identity, reconstructs the
cooperative relationship, and creates the group relationship of real and virtual
coexistence. The community is connected. The connection is namely develop-
ment. In IEOM education, the innovative subjects connected with interest will
establish the creative community and is involved in the innovation activities.
This community establishes a co-existence and co-construction, and shares in
loyalty or game behaviors.
(6) Network of IEOM education

IEOM education networks provide links be-tween specific makerspace workshops
and projects that permit forms of identities to flow and mobilize [17]. Innovation
networks in IEOM education give innovation subjects insights into temporal and
spatial contexts, in which to capture preferences, goals, paths, activities, and
locations. They can automatically connect appropriate information, resources
and nodes with humanized supports. Location-based networked services are not
just based on the physical space of the subject, but also the mind and emotional
states displayed by the subject of activities.
(7) Machine intelligence and human-machine collaboration in IEOM education
IEOM education can be inextricably linked to cyber-physical systems (CPS),
which is the in-depth integration of computation, communication, and control
technologies. CPS couples network and computing to the physical environment,
and to achieve full autonomy, efficient collaboration, mobile control, dynamic
adaptation, and human services. On the one hand, human-machine collabora-
tion transforms the interaction between human and machine and object, and
generates new methods. On the other hand, it associates individual objects or
subsystems into a unified entity, setting up a fusion of the natural world, the vir-
tual world, and the thought world. CPS emphasizes the communication between
human and machine, and scientifically controls the movement of objects through
precise calculation of data parameter. These communications and controls are
woven into innovation activities and continuously facilitate innovative subjects
sensing objects and knowing laws.
7 Conclusion
IEOM education not only drives the development of the world but also the
progress of humanity. It transforms the external world, while at the same time
shaping the inner world of mankind which conforms to the characteristics of
the new era. IEOM education takes individual intelligence to be gathered into
a powerful force, and to create a new situation of humanization, open-sources,
and co-creation. It breaks the limitations of traditional innovation and closed
research, to change people’s ideas and thoughts in the traditional innovation
environment, and promotes the development of the makerspace.
Acknowledgements. This work is supported by a grant from the Research Program

of Science and Technology Development Strategy of Tianjin “Research on Mechanism
and Model of Innovation Ecosystem-oriented Makerspace in Tianjin universities” (No.
16ZLZXZF00410).
References
1. Barrett T, Pizzico M et al (2015) A review of university maker spaces. Georgia
Institute of Technology
2. Bieraugel M, Neill S (2017) Ascending bloom’s pyramid: fostering student creativ-

ity and innovation in academic library spaces. Coll Res Libr 78:3–52
3. Brown A (2015) 3D printing in instructional settings: identifying a curricular hier-
archy of activities. TechTrends 59(5):16–24
4. Bulc V (2011) Innovation ecosystem and tourism. Academica Turistica-Tourism
Innov J 4(1):27–34
5. Durst S, Poutanen P (2013) Success factors of innovation ecosystems. http://www.
academia.edu
6. Haines T (2016) Developing a startup and innovation ecosystem in regional Aus-
tralia. Technol Innov Manage Rev 6(6):24–32
7. Hautamäki A (2006) Innovation ecosystem in city policy: the case of Helsinki.
http://www.edu.helsinki.fi
8. He K (2016) On maker education and innovation education. Educ Res 4:12–24 (in
Chinese)
9. Huang L, Bao X, Wang Y (2016) Learning by design: a new study on maker
education practice model in schools. China Educ Technol 11:18–22 (in Chinese)
10. Katona T, Tello S et al (2015) Pathways partners: entrepreneurial change across
campus. In: VentureWell. Proceedings of open, the annual conference, national
collegiate inventors & innovators alliance, p 1
11. Luoma-aho V, Halonen S (2010) Intangibles and innovation: the role of communi-
cation in the innovation ecosystem. Innov Journalism 7(2):3–20
12. Myers J (2015) Creating collaborative spaces at the university of Arizona: ways to
encourage interdisciplinary research and ideas
13. Nordfors D (2009) Innovation journalism, attention work and the innovation econ-
omy. A review of the innovation journalism initiative 2003–2009. Innov Journalism
6(1):5–46
14. Pont E (2015) Analysis and comparison of representative locations in the general
makerspace panorama. http://upcommons.upc.edu/handle/2117/87120
15. Ritala P, Agouridas V et al (2013) Value creation and capture mechanisms in
innovation ecosystems: a comparative case study. Jpn J Appl Phys 63(3–4):244–
267
16. Schwartz D, Barel R (2015) The role of a local industry association as a catalyst
for building an innovation ecosystem: an experiment in the state of ceara in Brazil.
Innovation 17(3):383–399
17. Smith A (2016) Grassroots digital fabrication and makerspace. https://ideas.repec.
org
18. Tian Y (2016) Maker education: its origin, connotation and possible paths. Int
Comp Educ 38(1):22–28 (in Chinese)
19. Wan L, Kang C (2016) Internet plus maker education: construction of the new
ecology for innovation and entrepreneurship education in universities. Res Educ
Dev 7:59–65 (in Chinese)
20. Wang D, Song S, Chen Z (2016) Application of virtual reality and augmented
reality in maker-driven education in universities. China Educ Technol 10:112–115
(in Chinese)
21. Wang M, Zhang L, Zheng Z (2016) The implementation approach of maker edu-
cation in our country. Mod Educ Technol 26(9):11–17 (in Chinese)
22. Wang P (2009) An integrative framework for understanding the innovation ecosys-
tem. http://doku.iab.de
23. Wilczynski V (2015) Academic maker spaces and engineering design. Am Soc Eng
Educ 26:1
24. Yang G (2016) The construction of the double helix model of maker education.
Mod Distance Educ Res 139:62–68 (in Chinese)
25. Yang G (2016) Maker education: the new path to the development of creative
education in China. China Educ Technol 3:8–13 (in Chinese)
26. Yang L, Zhang L, Wang G (2016) Study on the system framework of maker edu-
cation. Mod Distance Educ 3:28–33 (in Chinese)
27. Zhan Q, Yang M (2015) Research on maker education 2.0 and smart learning
activities from the perspective of “internet plus”. J Distance Educ 6:24–31 (in
Chinese)
28. Zhang M, Liu X, Zhang C (2016) Connotation, function and reflection of maker
education. Mod Educ Technol 26(2):14–19 (in Chinese)
29. Zhong B (2016) Discussion on background, essence, form, and support system of
maker education. Mod Educ Technol 26(6):13–19 (in Chinese)
30. Zhu L, Hu X (2016) Research on maker education oriented design based learning:
model and case. China Educ Technol 358:23–29 (in Chinese)
Pricing and Greening Policy for Manufacturer
with Low Carbon Strategic Customers
Wen Jiang(B)
College of Architecture and Urban-Rural Planning,

Sichuan Agricultural University, Chengdu 611830, People’s Republic of China
xuezhongsha wen@163.com
Abstract. Global warming is mainly caused by excessive emissions of green-

house gases especially the carbon dioxide, which makes the carbon reduction
become hot issues to businesses and consumers. This article considers a manu-
facturer produce one product and assumes the customers are of low carbon and
consisting of myopic customers and strategic customers. The customer decision
is affected by the price and unit carbon emissions of the product. The paper exam-
ines the optimal pricing and greening policies for the manufacturer with and with-
out green technology investment. It shows that the optimal prices in two scenarios
are decreasing in the ratio and patience of strategic customers. The presence of
strategic customers is detrimental to the manufacturer and the negative impact
increases as the ratio and patience of strategic customers increases. It also shows
that with the addition of carbon sensitive degree, the manufacturer will improve
the regular sale price and invest more in green technology when the customer
number exceed a certain threshold. Finally, the article analyzes the effect of green
technology investment on the optimal price and find that the optimal price with
green technology investment is lower than that without green technology invest-
ment.
Keywords: Strategic customer · Carbon sensitive demand · Pricing decision ·

Green technology investment
1 Introduction
Global climate change is a serious threat to human survival and development. Global
warming is mainly caused by excessive emissions of greenhouse gases especially the
carbon dioxide [9], which makes the carbon reduction become hot issues to businesses
and consumers [5, 10]. In response to the Government’s carbon reduction requirements
and low carbon preference of customers, manufactures are considering whether invest
in green technologies to reduce the carbon emissions of their products. With the imple-
mentation of the carbon footprint, it is becoming easier for customers to obtain the
information on carbon emissions and to take this information into account in purchas-
ing decisions. Therefore, it is very meaningful to study the optimal decisions of the
manufacturer with carbon emissions sensitivity demand. Another phenomenon worthy
Pricing and Greening Policy for Manufacturer 1267
of study is the presence of strategic customer behavior [7]. With the development of
e-commerce and the application of dynamic pricing strategies, customers are increas-
ingly able to anticipate future price cuts and chose to delay the purchasing. Such “smart”
customers are called strategic customer and customers who will not anticipate discount-
ing are called myopic customer. The strategic and myopic customers are co-exist in
real-life. Firms who ignore the presence of strategic customer will result in 20% losses
of their total profit [1].
Two research streams are closely related to this article and will be reviewed to high-
light our contributions. In the literature of models with carbon emissions demand or
green technology investment, Bansal and Gangopadhyay [2] found that consumers were
willing to pay a higher price for low-carbon products. Klingelhö [11] examined the effect
of carbon emissions trading on green technology investment. The paper found that tak-
ing carbon emissions trading into consideration had impact on the investment, but this
impact was not always positive incentives. Zhao et al. [17] investigated how did the man-
ufacture in perfectly competitive market chose a technology when green technology was
available. Yalabik and Fairchild [16] examined the manufacturer’s optimal green tech-
nology investment of environment-friendly products with consumer choice was taken
into consideration. Xia et al. [15] analyzed the interrelations among the carbon emission
reduction, green technology investment choices and enterprise performance. These study
provided a reference for the study of this paper, but they all assumed the customers were
myopic and ignored the fact that myopic and strategic customers co-exists in real-life. In
addition, Su [14] examined the pricing model of a monopolist who sold finite inventory
to heterogeneous strategic customers in finite period. Jerath et al. [6] investigated the
impact of two sales approach (through an opaque intermediary versus last-minute sales
directly to consumers) on the service provider’s profit when strategic customer behavior
was considering. Lai et al. [12] investigated the effect of a posterior price matching pol-
icy on the customers’ purchasing behavior, and seller’s pricing decisions. They showed
that strategic consumers’ waiting incentive could be eliminated by this policy. Du et al.
[4] examined the single-period joint inventory and pricing strategies with strategic cus-
tomers and risk preference. Prasad [13] considering mix of myopic and strategic con-
sumers and studied the choices of mixed bundling pricing and reserved product pricing
for a firm who sold two products. Jiang and Chen [8] studied the optimal decision for
a low carbon supply chain consists of one supplier constrained by cap-and-trade policy
and one retailer facing homogeneous strategic customers. The literature above does not
consider carbon emissions, or assumes that customers are homogeneous.
In order to fill the gap present above, this paper examine the manufacturer’s pric-
ing and green technology investment policy when the customer are carbon emissions
sensitive and myopic and strategic hybrid. This paper aims to address following three
questions:
• What is the optimal pricing policy for the manufacture without green technology
investment? How does the presence of strategic customers affect the optimal pricing?
• What is the optimal pricing and greening policy for the manufacture with green
technology investment? How does the presence of strategic customers affect the
optimal pricing and greening policies?
• What effect does the green technology investment on the optimal policies?
1268 W. Jiang
2 Model Descriptions and Assumption

The paper considers one manufacturer who produce a product and direct sell it to the
end users. The whole sales period is divided into two stage. At the first stage, the manu-
facturer sells the products at regular price p which is a decision variable. In the second
stage, the manufacturer sells the products at lower price s which is an exogenous vari-
able and decided by the market. The price in second stage is still greater than the cost of
the product, i.e. s > c. The initial unit carbon emissions of the product is e0 . The manu-
facture can invest in green technology to reduce unit carbon emissions. The unit carbon
emissions after green technology investment is e(e < e0 ). Referring to the expression
of [16], the paper assumes the green technology investment function, denoted by I(e),
is:
I (e) = t(e0 − e)2 , (1)
where t represents the investment coefficient. So the manufacturer’s decision variables

is regular sales price p and unit carbon emissions e.
This paper characterizes customers as heterogeneous low carbon customers. The
heterogeneity behavior in two aspects. On the one hand, the customers include myopic
and strategic types. On the other hand, the customers have different valuation to the
product. The customers can also observe the unit carbon emissions during the sales
period and take it into consideration to make a buying decision. The paper assumes that
the customer size are fixed and denoted by N. The ratio of strategic customers is θ and
the ratio of myopic customers is 1− θ . The paper assumes that each customer purchases
a product at most. The valuation of the two type customers (denoted by v) is follows
uniform distribution in [0, v f ]. Strategic customers discount future consumption at δ ,
(0 < δ < 1). By delaying a purchase until the second stage, Strategic customers lose
out on some consumptive value, and hence their future utility is reduced to reflect this
loss [3]. δ can be see as patience of the strategic customer and higher δ implies more
patient customers. The unit cost is c. The shortage cost and salvage value is not take
into consideration.
In addition, v f > p+ η e0 > s+ η e0 , which indicates that the customers have positive
δ +θ δ
margin when they buy the products. In order to express convenient, set A = 1−1− δ .
3 The Base Model

3.1 Analysis of Decision Making Process
According to traditional study, the decision condition for myopic customers is v− p > 0.
This paper extends it as v − p − η e0 > 0 to depict the carbon sensitive characteris-
tics of customers, where η represents the carbon sensitive degree. The extension of
strategic customer decision rules is similar, that is v − p − η e0 > 0 and v − p − η e0 >
δ (v − s − η e0 ).
(1) The first stage
At the first stage, the manufacturer set p. Then the two type customers will response to
these decisions according to themselves rules.
As to myopic customers. According to the decisions rules, myopic customers will

buy products as long as v − p − η e0 > 0, Otherwise, they will not buy products and
leave the market. So the expected number of myopic customers who take the purchase
vf v f −p−η e0
action in the first stage is (1 − θ ) N ∫ p+ η e0 f (v) dv = (1 − θ ) N vf
.
As to strategic customers. According to the decisions rules, strategic customers will
buy products only if:
v − p − η e0 > 0, v − p − η e0 > δ (v − s − η e0 ) .
Then
p − δ s + (1 − δ ) η e0
0< < v < vf .
1−δ
So the expected number of strategic customers who take the purchase action in the
first stage is:
vf v f − p−δs
1−δ − η e0
θ N p−δ s+(1−δ )η e f (v) dv = θ N .
1−δ
0 vf
To sum up, the expected number of customers who take the purchase action in the first
vf v f −p−η e0
stage is (1 − θ ) N ∫ p+ η e0 f (v) dv = (1 − θ ) N vf
.
(2) The second stage
At the second stage, the sales price reduces to s. Then the rest of the two type customers
will response to this price according to themselves rules.
As to myopic customers. The expected number who take the purchase action in the
p+η e
second stage is (1 − θ ) N ∫s+η e00 f (v) dv = (1 − θ ) N p−s
vf
.
As to strategic customers. The product will exit the market after the second stage.
Waiting does not make sense for strategic customers now. As long as the valuation is
greater than s + η e, the strategic customers will buy products. Therefore, the expected
number of strategic customers who take the purchase action in the second stage is:
p−δ s+(1−δ )η e0 p−s
+η e0
θ N ∫s+η e01−δ f (v) dv = θ N 1−δv f .
Based on the above analysis, the number of customers who purchase products can
be obtained and listed in Table 1.
Table 1. The number of customers who purchased the product
First stage Second stage

f
−p−η e0
Myopic customers (1 − θ ) N v vf
(1 − θ ) N p−s
vf
δ
v f − 1−δ −η e0
p− s p−s
+η e0
Strategic customers θN θ N 1−δ v f
vf
Total N
v f − (1 − θ ) p − θ (p−δ s) − η e0 N (1−δ +θ δ )(p−s)
+ η e0
vf 1−δ vf 1−δ
1270 W. Jiang
3.2 Pricing Model Without Green Technology Investment
The manufacturers expected profit function without green technology investment,

denoted by πn (p), is:

N θ (p − δ s)
πn (p) = f v f − (1 − θ ) p − − η e0 (p − c)
v 1−δ
(2)
N (1 − δ + θ δ ) (p − s)
+ f + η e0 (s − c) .
v 1−δ
Lemma 1. πn (p) is a concave function of p.

Proof. According to formula (2), I have d πdnp(p) = vNf v f − 2(1−1− δ +θ δ )
δ p − η e0 +
1−δ +2θ δ
d 2 πn (p) N 2(1−δ +θ δ )
1−δ s , d p2 = − v f 1−δ < 0. So I can obtain that πn (p) is a concave func-
tion of p. This completes the proof.

Proposition 1. There exists an unique optimal pricing policy pn , which maximizes

vf
πn (p), where pn = 2A − η2Ae0 + 1 − 2A
1
s.
d πn (p) 2(1−δ +θ δ ) δ +2θ δ
Proof. Set dp = 0, I have v f − 1−δ p − η e0 + 1−1− δ s = 0. Assume A =
1−δ +θ δ f η e
1−δ , I can derive the optimal pricing policy pn = 2A v
− 2A0 + 1 − 2A 1
s by solving
the equation. This completes the proof.

Proposition 1 shows that the manufacturer’s optimal pricing policy without green
technology investment is unique exist.
Proposition 2. pn is decreasing in θ .
d pn d pn dA v f −η e0 −s δ
d θ = dA d θ = − 1−δ . I have v − η e0 − s > 0, 1 − δ > 0. Then I can
Proof. f
2A2
d pn
obtain d θ < 0, i.e. pn is decreasing in θ . This completes the proof.

This conclusion is intuitive. θ represents the ratio of strategic customer. When the
number of strategic customers is increasing, in order to enable more strategic customers
buying at regular price, the manufacture will reduce the regular price in first stage.
Proposition 3. pn is decreasing in δ .
f −η e
0 −s θ
Proof. d pn
dδ = d pn dA
dA d δ = −v 2A2
. I have v f − η e0 − s > 0. Then I can obtain
(1−δ )2
d pn
dδ < 0, i.e. pn is decreasing in δ . This completes the proof.

Proposition 3 shows that the optimal pricing of the manufacturer decreases in con-
sumer patience (i.e. δ ). More patience means loss out less consumptive value and more
willing to buy the product in second stage for the strategic customers. In order to stim-
ulate strategic customers buy early, the manufacturer have to reduce the price.
Proposition 4. pn is decreasing in η .
d pn e0 d pn
Proof. dη = − 2A 2 < 0. So dη < 0, i.e. pn is decreasing in η . This completes the proof.

Proposition 4 indicates that the optimal pricing of the manufacturer decreasing in the
carbon sensitive degree (i.e. η ). Without green technology investment, the unit carbon
emissions is fixed and higher η means the customers less willing to buy the products.
In order to stimulate more customers to buy products, the manufacturer will reduce the
optimal price in first stage.
4 Pricing and Greening Model with Green Technology Investment

4.1 The Optimal Policy
The analysis of decision making process is similar to the base model and the differ-
ence is that the unit carbon emissions is no longer constant but a decision variable of
the manufacturer. According to the analysis in Sect. 2, I can obtain the manufacturer’s
expected profit function, denoted by πg (p, e), with green technology investment.

N θ (p − δ s)
πg (p, e) = f v f − (1 − θ ) p − − η e (p − c)
v 1−δ
(3)
N (1 − δ + θ δ ) (p − s)
+ f + η e (s − c) + t(e0 − e)2 .
v 1−δ
Lemma 2. If 4tA − η 2 > 0, πg (p, e) is a joint concave function of p and e.
Proof. I can obtain the partial derivatives of profit function πg (p, e) with respect to p
and e:

∂ πg (p, e) N f 2 (1 − δ + θ δ ) 1 − δ + 2θ δ
= f v − p − ηe + s , (4)
∂p v 1−δ 1−δ
∂ 2 πg (p, e) 2 (1 − δ + θ δ )
=− < 0,
∂p 2 1−δ
∂ πg (p, e) N
= − f (p − s) + 2t (e0 − e) , (5)
∂e v
∂ 2 πg (p, e)
= −2t < 0,
∂ e2
∂ 2 πg (p, e) ∂ 2 πg (p, e)
= = −η .
∂ p∂ e ∂ e∂ p
I have 4tA − η 2 > 0, then I can show that
∂ 2 πg (p,e) ∂ 2 πg (p,e)
∂ p2 ∂ p∂ e
∂ 2 πg (p,e) ∂ 2 πg (p,e)
= 4tA − η 2 > 0.
∂ e∂ p ∂ e2
So πg (p, e) is a concave function of p and e. This complete the proof.

1272 W. Jiang
Proposition 5. The optimal pricing and greening policy (denoted by pg and eg ) of the
manufacturer are:

2tv f v f − e0 η + (2A − 1) s − N η s
pg = ,
4Atv f − N η
4Ae0tv f − N v f − s
eg = .
4Atv f − N η
∂ πg (p,e) ∂ π (p,e)
Proof. Let ∂p = 0 and g∂ e = 0, from equation (4) and (5), I get pg =
2tv f [(2A−1)s−e0 η +1]−N η s 4Ae tv f −N (v f −s)
4Atv f −N η
and eg = 04Atv f −N η . This complete the proof.

Proposition 5 show that the optimal pricing and greening policy for the manufacturer
with green technology investment exist and are unique.
Proposition 6. pg is decreasing in θ ; eg is increasing in θ .
Proof. From Proposition 5, I get the value of pg and eg , then I can show
2
d pg d pg dA 8t 2 v f v f − η e0 − s δ
= =− ,
dθ dA d θ (4Atv f − N η )
2 1−δ
deg d pg dA 4Ntv f v f − η e0 − s δ
= = .
2 1−δ
dp de
I have v f − η e0 − s > 0. Then I can obtain d θg < 0 and d θg > 0. Therefore, pg is
decreasing in θ and eg is increasing in θ . This complete the proof.

Proposition 6 shows that the higher the ratio of strategic customers, the lower the
manufacturer’s optimal price. This conclusion is in line with the scenario without green
technology investment.
Proposition 7. pg is decreasing in δ ; eg is increasing in δ .
Proof. I can obtain
2
d pg d pg dA 8t 2 v f v f − η e0 − s θ
= =− ,
dδ dA d δ (4Atv − N η )
f 2
(1 − δ )2
deg d pg dA 4Ntv f v f − η e0 − s δ
= = .
2 1−δ
dp de
I have v f − η e0 − s > 0. Then I can obtain d δg < 0 and d δg > 0. Therefore I get pg
is decreasing in δ and eg is increasing in δ . This complete the proof.

Proposition 7 indicates that more patience the strategic customer have, lower the
optimal pricing with green technology investment is and higher the optimal unit carbon
emissions is with green technology investment. As to the impact of δ on the optimal
unit carbon emissions, I find that more “strategic” customers will lead to less investment
that the manufacturer do in carbon reduction.
4Atv f
Proposition 8. If N > η , pg and eg are decreasing in η .
4Atv f
Proof. N > η , then 4Atv f < N η . Recall v f − η e0 − s > 0, so
2
d pg 2Ntv f v f − s − 8Ae0t 2 v f 2Ntv f v f − η e0 − s
= > > 0,
dη (4Atv f − N η )
2
(4Atv f − N η )
2
deg N 2 v f − s − 4NAe0tv f N 2 v f − η e0 − s
=− <− < 0.
dη (4Atv f − N η )
2
(4Atv f − N η )
2

Proposition 8 shows that with the addition of carbon sensitive degree (i.e. η ), the
manufacturer will improve the regular sale price and invest more in green technology
investment to reduce carbon reduction when the customer number exceed a certain
threshold.
4.2 Effect of Green Technology Investment

In order to analyze the effect of green technology investment on the manufacture’s
optimal pricing, I compare the manufacturer’s optimal prices with and without green
technology investment.
4Atv f
Proposition 9. If N > η , pg < pn .
2tv f [v f −e0 η +(2A−1)s]−N η s N η (v f −η e0 −s)

− η2Ae0 + 1 − 2A
f
Proof. pg − pn = 4Atv f −N η
− 2A
v 1
s= 2A(4Atv f −N η )
. If
4Atv f
N> η , i.e. 4Atv f − N η < 0, then pg − pn < 0, i.e. pg < pn . This complete the proof.

Proposition 9 indicates that the optimal price of the manufacturer with green tech-
nology investment is lower than that of without green technology investment when the
number of customers exceed a certain threshold. Considering green technology invest-
ment, the manufacturer reduces the unit carbon emissions and then get the room for
price reduction. In order to maximize the expected profit, the manufacturer need to
reduce the optimal price to enable more customers to buy the products.
5 Conclusion and Future Research

This paper investigates the pricing and greening policy of a manufacturer and analyze
the impact of green technology investment on the optimal price. The manufacture pro-
duction a product and sells them to the mix myopic and strategic customers. Customers
will be affected by unit carbon emissions of the product when making a purchase deci-
sion. The manufacture can invest in green technology to reduce unit carbon emissions.
To the best of our knowledge, this is the first study on the model incorporate mix myopic
and strategic customers and green technology investment together. This article provides
several interesting observations.
1274 W. Jiang
(1) There is an unique optimal pricing policy for the manufacturer without green tech-
nology investment and there is also an unique optimal pricing and greening policy
for the manufacture with green technology investment. Therefore, by deriving the
optimal policies in different scenarios, the manufacturer can behave appropriately
based on the findings to maximize its expected profit.
(2) I find that the optimal price without green technology investment is decreasing in
the ratio of strategic customers, patience of strategic customers and carbon sen-
sitive degree. This conclusion means that the presence of strategic customers is
detrimental to the manufacturer and the negative impact increases as the ratio and
patience of strategic customers increases.
(3) I find that the optimal price with green technology investment is also decreasing
in the ratio and patience of strategic customers and carbon sensitive degree; the
optimal unit carbon emissions is increasing in the ratio and patience of strategic
customers but decreasing in carbon sensitive degree.
(4) The optimal price with green technology investment is lower than that without
green technology investment when the number of customers exceed a certain
threshold. This conclusion is interesting. The cost of the product is increasing when
the manufacturer invest in green technology. According to the experience, the opti-
mal price should be improved. However, the optimal pricing of the manufacturer
not increased but reduced instead. This conclusion can guide firms make decisions
scientific.
The paper formulates the model by considering only one manufacturer. But supply
chain is more common in real-life. Extending this assumption will have a knock-on
effect on supply chain decisions. So one key research direction is to consider a sup-
ply chain scenario. The manufacturer decide the wholesale price and green technology
investment and the retailer decide selling price and order quantity. Second, this paper
assume that the manufacturer only produces one product. While this simplified setting
provides some interesting insights, we have to acknowledge that usually two or more
products is produced by a manufacturer. These products will substitute or complement
each other. So to assume that the manufacturer produces two or more products, which
have different unit carbon emissions in production, is one important extension of this
work.
Acknowledgements. This research is partially supported by National Natural Science Founda-

tion of China (No. 71602134) and Research Interest Training Program of Sichuan Agricultural
University (No. 2016133).
References
1. Aviv Y, Pazgal A (2008) Optimal pricing of seasonal products in the presence of forward
looking consumers. Manufacturing Serv Oper Manag 3(10):339–359
2. Bansal S, Gangopadhyay S (2003) Tax/subsidy policies in the presence of environmentally
aware consumers. J Environ Econ Manag 45(2):333–355
3. Cachon GP, Swinney R (2011) The value of fast fashion: quick response, enhanced design,
and strategic consumer behavior. Manag Sci 4(57):778–795
4. Du J, Zhang J, Hua G (2015) Pricing and inventory management in the presence of strategic
customers with risk preference and decreasing value. Int J Prod Econ 164:160–166
5. IPCC (2007) Cliamte change 2007 synthesis report. IPCC pp 1–13
6. Jerath K, Netessine S, Veeraraghavan SK (2010) Revenue management with strategic cus-
tomers: Last-minute selling and opaque selling. Manage Sci 3(56):430–448
7. Jiang W, Chen X (2012) Manufacturers production and pricing policy with carbon constraint
and strategic customer behavior. Adv Inf Sci Serv Sci 23(4):231–238
8. Jiang W, Chen X (2015) Supply chain decisions and coordination with strategic customer
behavior under cap-and-trade policy. Control Decis 31(3):477–485
9. Jiang W, Chen X (2016) Optimal strategies for low carbon supply chain with strategic cus-
tomer behavior and green technology investment. Discrete Dynamics Nat Soc, 1–13
10. Jiang W, Chen X (2016) Optimal strategies for manufacturer with strategic customer behav-
ior under carbon emissions-sensitive random demand. Ind Manag Data Syst, 4(116):759–776
11. Klingelhö FHE (2009) Investments in eop-technologies and emissions trading-results from
a linear programming approach and sensitivity analysis. Eur J Oper Res 1(196):370–383
12. Lai G, Debo LG, Sycara K (2010) Buy now and match later: Impact of posterior price match-
ing on profit with strategic consumers. Manufacturing Serv Oper Manag 1(1):33–55
13. Prasad A, Venkatesh R, Mahajan V (2015) Product bundling or reserved product pricing?
price discrimination with myopic and strategic consumers. Int J Res Mark 32(1):1–8
14. Su X (2007) Intertemporal pricing with strategic customer behavior. Manage Sci 5(53):726–
741
15. Xia D, Chen B, Zheng Z (2015) Relationships among circumstance pressure, green technol-
ogy selection and firm performance. J Clean Prod 106:487–496
16. Yalabik B, Fairchild RJ (2011) Customer, regulatory, and competitive pressure as drivers of
environmental innovation. Int J Prod Econ 2(131):519–527
17. Zhao J, Hobbs BF, Pang JS (2010) Long-run equilibrium modeling of emissions allowance
allocation systems in electric power markets. Oper Res 3(58):529–548
Evolutionary Game Analysis of the Reservoir
Immigrants’ Joint Venture
Xiaofeng Liu(B) , Tingting Song, and Yanhua Liu
The Management Department in Northwest University for Nationalities,

Gansu, People’s Republic of China
liuxiaofengmdqk@126.com
Abstract. This paper combined the previous results with the actual
situation of the immigrant area, proposed the main factors which would
influence the joint venture, such as technology level, the familiar degree
of both sides of cooperation, the costs and benefits of cooperation and
coordination. To establish the evolutionary game model of immigrant
joint venture, Finding the influence factors and the “free-rider” punish
coefficient, the investment allocation coefficient, the income distribution
coefficient how to affect the direction of the joint venture for immigrants
according to payoff matrix and the Jacobi matrix.
Keywords: The reservoir immigrants · The joint venture · Evolutionary

game
1 Introduction
With the time increase of reservoir immigrants living in the settlement, the pos-
sibility of its entrepreneurship. Many immigrants chose to entrepreneurship on
some basis of technical in Yongjing county in Linxia Hui Autonomous Prefecture,
Gansu province, such as tourism, fisheries and greenhouse cultivation. It still
exists social employment problem now, the government vigorously encouraging
entrepreneurship. There are many immigrants started the business in Yongjing
county, but to reduce risk, choosing partners to joint venture will reduce the risk
of entrepreneurship. Therefore, it is very necessary to study on the joint venture.
Yang [6] used the binary Logistic regression method to study the influence
factors of immigrant entrepreneurs. Yang [7] re-examined the “bounded ratio-
nality” of venture capitalists. Su [3] argued that the policy of encouraging coop-
eration that government put forward to has not work, suggesting that the gov-
ernment should intensify the policy support and increasing investment. Chou [2]
suggested when cooperating, both firms need to contribute sufficient and com-
plementary efforts when they choose to cooperate partner. When entrepreneurs
choosing a partner, relation network is the key variables to affect entrepreneurial
decision [1]. Szolnoki [4] thouhgt that a carefully chosen threshold to establish
a joint venture could efficiently improve the cooperation level. Wu and Wang
[5] thought that the harder the joint venture is, the higher level of cooperation
will be.
Evolutionary Game Analysis of the Reservoir Immigrants’ Joint Venture 1277
2 Evolutionary Game Model

Evolutionary game theory is the result of the evolution, which is a kind of
dynamic equilibrium.
2.1 The Assumption of Model Variables

In this paper, the mode of entrepreneurship is individual, cooperation means
people to do a project together. Assuming that the probability of individual A
choose to cooperation is x, uncooperative probability is 1−x, and the probability
of individual B choose to cooperation is y, uncooperative probability is 1−y. The
rest of the variables are as follows: The technical level Ti , Resource utilization
ability αi , The cost of coordination C, the prior information Qi , The credibility
of the prior information βi , the profit R, the free-rider penalty coefficient d. To
make the following Hypothesis about the problem of the joint venture:
(1) This study involved two entrepreneurs, assuming that there are two individ-
uals A and B, and the two individuals are in the same industry, they have
their own worker sources and experience. In their cooperative game has two
choices: cooperation and noncooperation.
(2) When the cooperative game beginning, both sides will take the degree of
familiar with each other into account, a person’s credit affects whether other
people can choose to believe him. Both sides have some prior information
about each other, which is the understanding of each other, but there is a
credibility in prior information, and some of the information may be wrong.
So it is important to put the prior information reliability coefficient as βi .
That is to say, A has the prior information with B, which is QB , The credi-
bility of the information is βA .
(3) Both the game sides have a certain technology. What’s more, technical degree
is one of the factors that both sides of game will consider in choosing cooper-
ation. Assuming that technical level is Ti , there also is exists a certain degree
of resource utilization while working together, assuming that resource uti-
lization coefficient is αi .
(4) Assuming that the investment apportion coefficient are a and 1 − a, which
can get the income distribution coefficient b and 1 − b.
(5) When there is the phenomenon of “free-rider”, it will have some punishment
d, the punish coefficient is established on the basis of the technical level.
2.2 Evolutionary Game Model
According to the hypothesis and model variables, we can get the payoff matrix
of A and B, as shown in Table 1.
In order to facilitate discussion, assuming that the value of x and y is invari-
able. To calculate the expectations of A and B in the case of cooperation and non-
cooperation respectively, and then get replicated dynamic equation of A and B:
1278 X. Liu et al.
Table 1. The payoff matrix
B
Cooperation y Noncooperation 1 − y
A Cooperation x bR + βA QB + αA TB − aC, bR − aC − TA
(1 − b)R + βB QA + αB TA − (1 − b)R + αB TA − (1 − a)
(1 − a)C C − dTA
Noncooperation bR + αA TB − aC − dTB , 0,0
1−x (1 − b)R − (1 − a)C − TB
dx
A: F (x) = = x(UA1 − UA ) = x(1 − x)[y(βA QB + dTB − bR + aC + TA )
dt
+ (bR − aC − TA )],
dy
B: F (y) = = y(1 − y){x[(βB QA + (1 − a)C + dTA + TB − (1 − b)R]
dt
+ (1 − b)R − (1 − a)C − TB }
To obtain the evolutionary stable strategy of the game, First of all, we need
to get the stable point of replicated dynamic equation, namely it need to meet
such conditions:
No. 1 F (x) = 0; No. 2 Evolutionary stable strategy needs to have some
F (x) = 0
anti-interference. For , we can get five stable points, O(0, 0), A(0, 1),
F (y) = 0
B(1, 0), C(1, 1), D(XD ,YD ).
−(1 − b)R + (1 − a)C + TB

XD = x∗ = ,
βB QA + dTA − (1 − b)R + (1 − a)C + TB
−bR + aC + TA
YD = y∗ = .
βA QB + dTB − bR + aC + TA
Assuming that M = βA QB + dTB and W = bR − aC − TA in F (x); E =

βB QA + dTA and F = (1 − b)R − (1 − a)C − TB in F (y).
According to the methods put forward by Friedman, the dynamic equilibrium
of participants’ activities can be got according to the part of the system stability
of
the Jacobi
matrix. We can get the Jacobi matrix based on the above. J =
X11 X12
, Among them,
X21 X22
X11 = (1 − 2x)[y(M − W ) + W ],
X12 = x(1 − x)(M − W ),
X21 = y(1 − y)(E − F ),
X22 = (1 − 2y)[x(E − F ) + F ].
We use trJ and det J to express the trace of matrices and determinant. When
the stable points meet the conditions: trJ > 0 and det J > 0, The stable point
is the evolutionary stable strategy of the system (ESS), the value of the Jacobi
matrix and determinants are as follow:
trJ=X11 + X22 ,
det J = X11 X22 − X12 X21 .
Based on the contents mentioned above we can get the local stability of equilib-
rium under different parameters:
(1) When W < 0, F < 0, M < 0, E < 0, the evolutionary stable strategy of the
system is O(0, 0).
(2) When M < 0, E < 0, W > 0, F > 0, the evolutionary stable strategy of the
system is A(0, 1), B(1, 0).
(3) When W > 0, E > 0, M > 0, F > 0, the evolutionary stable strategy of the
system is C(1, 1).
Fig. 1. System dynamic phase diagram of three different conditions
2.3 The Influence Factors Analysis of Evolution Path

Because of the coefficient set in the evolutionary game model that may influence
the final result, the probability of system dynamic evolution reached to O(0, 0),
A(0, 1), B(1, 0), C(1, 1). Probability of cooperation can be determined according
to the area of the quadrilateral and quadrilateral. Therefore, for discussing the
effect of various parameters on the dynamic evolution of the payoff matrix can be
converted into the effects of various parameters on the area of the quadrilateral.
The area of can be expressed as:
(1 − y ∗ ) (1 − x∗ ) E M
SADBC = + = + ,
2 2 2(E − F ) 2(M − W )
1280 X. Liu et al.
(1) Request the first partial derivative and second derivative of SADBC about
∂2S )C E 2
−W )C M 2
a, we can get ADBC
∂a2 = (E−F
(E−F )4
+ (M(M −W )4
> 0.
• If E > M , that is to say when investment apportion coefficient get greater,
the both sides of game tend to be non-cooperative.
• If E < M , that is to say when investment apportion coefficient get greater,
the both sides of game tend to be cooperative.
(2) Request the first partial derivative and second derivative of SADBC about
∂2S )R E 2
−W )M R 2
b, we can get ADBC
∂b2 = (E−F
(E−F )4
+ (M(M −W )4
.
• If E > M , that is to say when income distribution coefficient get greater,
the both sides of game tend to be cooperative.
• If E < M , that is to say when get income distribution coefficient greater,
the both sides of game tend to be non-cooperative.
∂SADBC
(3) Request
the first partial
derivative SADBC about C, then can get ∂C =
1
2
(a−1)E
(E−F )2 + (M−aM
−W )2 > 0, that is to say when the coordination cost get
greater, the evolutionary games tend to be non-cooperative.
(4) Request the first partial derivative SADBC about d, when the penalty coef-
ficient get greater, the evolutionary game tends to be the cooperative.
(5) Request the first partial derivative SADBC about βA and βB , we can conclude
that when the familiar degree between A and B is more higher, the more
tendency for them to cooperate.
(6) Request the first partial derivative SADBC about αi , the result is 0, so the
ability to absorb and utilize resources in the cooperation is not the main
factors.
2.4 The Analysis of Model Results

Based on the analysis above, we can conclude that:
(1) When the value of a and b equal to 0.5, it will reach the steady state, namely
when both the investment allocation coefficient and income distribution coef-
ficient are 0.5, the probability of cooperation is the largest.
(2) The less the coordination costs are, the more the probability of cooperation is.
(3) The bigger the “free-rider” punish coefficient is, the bigger the probability
of cooperation is.
(4) The higher the credibility of a prior information is, the bigger the probability
of cooperation is.
(5) The ability to absorb and utilize resources in the cooperation is not the main
factors.
3 The Analysis of an Empirical Example

Through investigation and interview in Yongjing county in Linxia Hui
Autonomous Prefecture, Gansu province, Some of immigrants are first to
migrate, some are the second. Expert scoring method is used to evaluate several
Table 2. The value of the various factors
The symbol The value

Ti TA = 5, TB = 6
αi αA = 0.6, αB = 0.7
C 500
Qi QA = 2, QB = 3
βi βA = 0.2, βB = 0.1
R 5000
d 0.3
related aspects on some individuals according to the assumption. We found out

two individual households at same level, The value of the various factors are as
follows (Table 2):
According to the conclusion of evolutionary game model, assuming that a =
b = 12 , and setting the value of the different factors according to the hypothesis,
and then we can get the payoff matrix with different value:
Table 3. The No. 1 payoff matrix
B
A Cooperation x 2254.2, 2253.7 2245, 2252
Noncooperation 1 − x 2251.8, 2244 0, 0
It is the best choice for A and B to cooperate with each other. When TA =
1, TB = 9, other value isn’t changed, the payoff matrix is as follow:
Table 4. The No. 2 payoff matrix
B
A Cooperation x 2256, 2250.9 2249, 2250.4
Noncooperation 1 − x 2252.7, 2241 0, 0
Comparing Table 3 with Table 4, it is not difficult to find that, in Table 4, the
benefits is similar for B to choose cooperation or not, so the smaller technology
differences between both sides of the game is, the more tendency to cooperation.
It’s easy to prove the other results from model analysis with this methods.
1282 X. Liu et al.
4 Conclusion
According to the example of Yongjing county in Linxia Hui Autonomous Pre-
fecture, Gansu province, we can get the main factors which would influence the
joint venture in this area, such as technology level, the familiar degree of both
sides of cooperation, the costs and benefits of cooperation and coordination. we
can also get the direction of every factors and the “free-rider” punish coefficient,
the investment allocation coefficient, the income distribution coefficient influence
on joint venture.
(1) When the both sides of the game have a big difference on technical level,
the high level of technical would not choose joint venture.
(2) The most residents in settlement came from different areas, and some of
them experienced two or more times of migration. When they familiar with
each other, they will have a sense of security, and the higher the credibility
of the information is, the more possibility the choice of the joint venture will
be.
(3) The higher costs and benefits of cooperation is, the greater the risk will be,
and the probability of joint venture will decline.
(4) The greater the benefits of joint venture is, the greater the probability of
cooperation will be.
(5) The bigger the “free-rider” punish coefficient is, the bigger the probability
of cooperation is.
(6) When both the investment allocation coefficient and income distribution
coefficient are 0.5, the probability of cooperation is the largest.
Acknowledgement. This research was financially supported by Ministry of Educa-

tion of the People’s Republic of China Project (12YIA630080).
References
1. Ardichvili A, Cardozo R, Ray S (2003) A theory of entrepreneurial opportunity
identification and development. J Bus Ventur 18(1):105–123
2. Chou PB, Bandera C, Thomas E (2017) A behavioural game theory perspective on
the collaboration between innovative and entrepreneurial firms. Int J Work Innova-
tion 2(1):6–31
3. Su L, Yang Z et al (2013) The influence factors of green technology cooperation
between enterprises, based on the perspective of supply chain. The Population of
China Resources and environment, pp 149–154 (in Chinese)
4. Szolnoki A, Chen X (2016) Cooperation driven by success-driven group formation.
Phys Rev E 94(4):042–311
5. Wu T, Fu F, Zhang Y et al (2013) The increased risk of joint venture promotes
social cooperation. PLoS One 8(6):1–10
6. Yang X, Wang C, Xiong Y (2015) The influence factors study of the three gorges
reservoir immigrant entrepreneurship decision. The Rural Economy pp 120–124 (in
Chinese)
7. Yang Y (2016) Based on evolutionary game of joint venture formation mechanism
research. J Technol Econ Manag Res 12:20–24 (in Chinese)
Solid Waste Management
in the Republic of Moldova
Gheorghe Duca(B) and Aliona Mereuţa
Academy of Sciences of Moldova, Chisinau, Republic of Moldova

malvina.condratiuc@mail.ru
Abstract. Waste management refers to all operations implemented for

the proper management of solid waste: methods of collecting, recycling
and recovery, processing, treatment, storage, transportation, analysis
and monitoring. The main topics in this paper are the following: insti-
tutional framework, national legislative and regulatory frameworks of
waste management; Integrated management of municipal waste; Haz-
ardous waste management; Secondary wine product management; Man-
agement of plastic waste; Infectious medical waste management. Thus,
ASM in collaboration with a private company implemented two tech-
nologies for complex treatment of solid waste. The following goals were
currently reached in RM: a sector of complex full processing of plastic
waste was created at industrial scale, a pilot line was put into operation,
for sorting, processing and obtaining of new products from secondary
polymer raw materials and an authorized point was created for manag-
ing infectious waste resulting from medical activities in medical insti-
tutions and hospitals, by collecting, sorting, packing, transporting and
neutralizing these wastes to prevent environmental contamination. The
neutralization of infectious medical waste is performed by vacuum steril-
ization method using the STERISHRED 250, which is the most modern
method of solving local problems of infectious medical waste processing
without incineration.
Keywords: Waste management · Secondary wine product manage-

ment · Integrated management of municipal waste · Management of plas-
tic waste · Infectious medical waste management · Republic of Moldova
1 Introduction
Transforming waste management and recycling practices in Republic of Moldova
is one of the most important challenges we face of the next decade. Inad-
equate waste management affects our communities, threatens our environ-
ment, and contributes to the global emissions of greenhouse gases. Thus, solid
waste management and recycling is a local, national and international priority
[1,3,5,6,9,10,12]. Government will establish the legal and institutional frame-
work necessary to support the gradual alignment of our waste management prac-
1284 G. Duca and A. Mereuţa
tices to those of the European Union: Government Action Programme (2016–

2018); National Strategy on Waste Management (2013–2023): Government Deci-
sion no. 428 of 10 April 2013; Strategy on Environment Protection: Government
Decision no. 301 of 24 April 2014; National Plan on Implementation of the Asso-
ciation Agreement (2014–2016); Association Agreement between the European
Union and the Republic of Moldova, ratified by Law no. 112 of 02.07.2014.
1.1 Waste Management Planning in Moldova
Municipal waste generation is influenced by many factors, the most significant

being the income of the population, consumer behavior, appearance on the mar-
ket of new packed products and the demographic evolution. A high level of the
population’s income and urbanization has resulted in the generation of large
quantities of waste per capita, which is in the rural areas typically range from
0.3 to 0.4 kg/inhabitant/day, 0.9 kg/place/day or more in urban areas, accord-
ing to the World Bank studies. The increasing number of supermarkets, along
with increased GDP per capita has increased purchases of packaged products
and hence the waste generated. Demographics also influence a waste genera-
tion, people in the urban areas producing much more waste than those in the
rural areas. The most common method of treating household waste in Moldova
is landfill, which often represents an important source of pollution of soil and
groundwater. Annually through the sanitation services from urban localities are
transported about 1144–2131 thousands m3 of solid waste [2]. Another nega-
tive aspect of inadequate waste management is that many useful and recyclable
materials are stored together with the non-recyclable thus is losing a large part
of their useful potential (paper, glass, metals, plastics); being mixed, chemically
and biologically contaminated, their recovery is very difficult. Effective waste
management planning certainly is focused on the morphological composition of
solid household waste [11]. The waste morphological analysis is imperative in
recycling operations establishing and in generated waste eliminating. This data
has been collected from newsletters, reports, publications, while in the national
statistics such information is missing (Fig. 1).
Possible scenarios of development of the domain of reference in the Republic
of Moldova were determined based on statistical data analyzes and the experts
assumptions, taking into account the macro-economic indices, indicate the exist-
ing strategies for development of some sectors (industry, agriculture, etc.), and
namely the Gross Domestic Product Value, demographic trends, trends of indus-
try development, tendencies of the agricultural sector development and waste
management system development. According to statistical data analysis, we
can ascertain that the solid household waste tendencies increasing volumes for
Chisinau city are 5–15% and in rural areas, there is a solid household waste
volume increase by approximately 10%. These increases are explained by the
fact that in localities, as well as in regional centers was improved the sanitation
services activity. Currently, an increased attention is given to the creation of the
solid household waste storage capacity and respective the annual growing trends
Solid Waste Management in the Republic of Moldova 1285
120,00
100,00
80,00
60,00
40,00
20,00
0,00
1986 1993 1996 1999 2001 2003 2005 2012 2016
Food waste Paper & carton Plasc Texles Leather Yard wastes Wood Other waste
Fig. 1. Municipal waste composition
of 10% will be maintained in the 2011–2015, subsequently is expected a contin-

ued growth of 5% for both categories of localities. The solid waste generation
trends are presented based on these assumptions (Fig. 2).
8000
6000
4000
2000
0
2010 2011 2012 2013 2014 2015 2016 2017 2017 2018 2019 2020 2021 2022 2023 2024 2025
Fig. 2. The forecasts on the solid household waste generation, thousand m3
The regional approach for waste management is essential for attracting nec-
essary investment, as well as for the high-cost maintenance through realizing
the implementation on the economical level. It is unacceptable and at the same
time economically unjustified the enterprise’s construction of recovery or waste
disposal in every city, without taking into account the rural areas of the district,
including the specific waste generated. The experience of neighboring countries
shows that the object’s financial sustainability of recovery and waste disposal is
economically feasible when it has a territorial coverage of at least 200–300 thou-
sand people. In this context, it is proposed the waste management regionalization
by the territorial division of the country into 8 regions for waste management.
The basic criteria for regional planning are the geographic position, economic
development, the existence of access roads, soil, and hydrogeological conditions,
the number of population, etc. (Fig. 3).
The local public authorities are expected to establish the waste management
associations at a regional level, as recommended by the Ministry of Environment
regarding the regional waste management planning. The role of associations
is to establish and approve terms of reference for selecting the company that
will manage waste in the region, as well as tariffs for the waste collection and
disposal, etc.
Fig. 3. Regional waste management planning in the Republic of Moldova
Considering the country’s economic development, morphological amount and

composition of the generated waste, as well as pedological and geological condi-
tions, were developed three alternatives for establishing integrated waste man-
agement system:
• Alternative 1 consists in the construction of eight solid household waste

deposits for each region. The cost of implementation is approximately 293
million e.
• Alternative 2 envisions the building of 2 mechanical-biological waste treat-
ment stations in Region 4 Chisinau and Region 7 Balti and 6 landfills for
solid domestic waste for remaining regions. The cost of implementation is
approximately 370 million e.
• Alternative 3 proposes the construction of 2 incinerators in Region 4 Chisinau
and Region 7 Balti and 6 deposits for solid domestic waste for remaining
regions. The cost of implementation is approximately 476 million e.
These options will be subject to public debate in order to select the optimal
alternative from the perspective of environmental protection and cost–benefit
analysis, to encourage the creation of partnerships at international, national and
local level, and to attract the necessary investment that will allow the sustainable
development of sector in accordance with the priority needs and at a rhythm
accessible for society.
2 Advanced Treatment of Secondary Wine Products in

the Republic of Moldova
For the Republic of Moldova, oenology is one of the most substantial key in the
national economic sectors, thus this sector requires continuous implementation of
new performances achieved in science and technology. The wine industry in the
country has a potential of grape processing up to 1 million tons, located in over
100 companies. Recent years, Moldovan Wine industry attention is increasingly
higher. As a result of the implementation of the State Program, the areas planted
with vines have increased significantly, in the quantity of grapes processed rose
as well. But it is known that about 20–25% of the processed grapes turns in are
secondary wine products. If you calculate the amount of waste that result after
processing 300 thousand tons of grapes, we get 60–75000 tones of products:
the cake, yeast, vinasse, wastewater, etc. All these products are a source of
environmental pollution and create major environmental problems. Therefore,
it is important to use complex raw material (grapes), implementing technological
processes with minimum waste, as well as processing products of wine in order
to obtain a wide range of valuable products.
Products of vinification are an important source for obtaining special natural
products with specific properties that cannot be obtained synthetically. Such
products as tartaric acid, oil seed polyphones, colorings, tannins, etc. have a
wide field of use in various industries.
Through the years 1980–1990, Moldova was the country that provides the raw
tartar material (calcium tartrate, tirighia, calcareous sediments) for countries
that produced directly tartaric acid (Ukraine, Odessa, Yerevan, Tbilisi, Italy,
etc.). Prime cost of the product is quite high (100 g of tartaric acid (99.9%) used
in the pharmaceutical industry costs 80–100 $, or 1 kg of natural tartaric acid
used in food costs 30–40 $.
Tartaric acids are reductions that can be used in food, wine, bakery, pharma-
ceutical, photochemical, chemical, textile, construction, electro-technical indus-
try, yet their production in the country is missing, although the country prevails
large stocks of raw materials and annual it could produce 100 tons up to 350
tons of tartaric acid.
An effective stabilizer used in the wine industry is the met tartaric acid,
which inhibits the process of submitting wine stone in wine and juice, do not
change the taste of wine or juice qualities and is not toxic because it represents a
modifying tartaric acid. Compared with the cold treating method, stabilization
of the wines and juices with mesotartaric acid it has a much better economic
effect. The local market is not assured, with this product, although it is requested
on the domestic market than tartaric acid.
Grapes seed represents a precious waste for our national economy. The con-
tent of the seeds in one tone of grapes is 7% by weight. The process of obtaining
seeds and their conditioning is a complicated technological process. In the years
1982–1986 in Moldova obtained to 10,000 tons seeds annually, which were trans-
ferred to processing in the Bender city (oil mill), Odessa, Armavir (Krasnodar
region), Tbilisi (Georgia) and Cocand (Uzbekistan). Partly some seeds were
exported to France with a price of $400 per ton.
Currently, grape dried seeds, oil seeds as well have high demand nationally
and internationally and we are also having factories producing food seed oil.
Getting antimicrobial chemotherapeutic remedies, particularly in natural raw
materials, is current and prospective permanent. Also of major importance is get-
ting antifungal properties and antioxidative substances from plant raw materials.
Research results on the activity of dissoluble etanon in water, allow production
of preparations with pronounced biological activity, useful in medicine and vet-
erinary use in combating bacterial diseases, fungal infections etc.
Currently, after processing the grapes are produced about 59 thousand tons
of pomace. Marc contents of seeds make up 40–65%. So annually in the country,
the grape processing plants can be harvested 20–25 thousand tons of seeds. Given
the fact that enotatin content in grape seed makes up on average 8% annually
could get about 2–2.5 thousand tons of enotanin. Prior research showed that
leaching enotaninului yield is 65–70%. Thus, it may get about 1.2–1.3 thousand
tons of active ingredient to changing them to produce medicinal preparations,
veterinary, agricultural, etc.
Among all waste formed after the vinification process is the marc, which
is the result of distilling alcohol from wine products. The perspective method
of treating this waste concentrates worldwide is considered anaerobic digestion,
which allows not only solve environmental challenges but also allows the conver-
sion of organic pollutants into biogas, which can still be used as an alternative
source of heat and electric.
The data presented above demonstrates that Moldova must build a com-
plex factory for processing secondary wine products. An argument convincing
enough to build such an enterprise serve the scientific data presented in the
monograph [4].
2.1 Plastic Waste Management in the Republic of Moldova
The plastic materials represent about 10–12% of the amount of household gener-
ated waste in the country. Not all the categories of plastic materials are subject
to recycling. The plastic materials in the majority of PVC degrade over sev-
eral centuries, between 100 and 1000 years. Therefore we ascertain how great
the pollution that they produce being persistent in time. The basic problems of
recycling of plastics are in the variety of their frequencies, which are recycled
separately and mixing them make it impossible to recover. The most often is
considered optimal recovery separated from other plastic materials of the recep-
tacles from PET by producers and consumers of soft drinks, which prevents
mixing with other resins.
Beneficiaries dealing with plastics processing decide the categories of waste
accepted for collection through specialized points.
For these reasons, it is necessary to organize the collection of differentiated
plastics and training potential generators of waste plastic materials, including
the general public. Concomitantly with the increasing amount of organic plastic
materials in use, the problem of recycling is becoming more topical. Proceed-
ing from the particularities of the composition of the waste polymers, plastics
processing at low temperatures, which allowing us to obtain the polymer raw
material (for example, construction materials, ramps, pillars, etc.) is the most
promising.
Plastic deformation of solid bodies usually leads not only to changing its
shape but also to the defects that change the physicochemical properties, includ-
ing the reactivity. The accumulation of defects is used in chemistry to accel-
erate reactions involving the solid substances, for reducing the temperature
processes and for other manipulations for increasing chemistry interaction in
the solid-phase. The mechanical energy absorption initiates destruction of poly-
mers, polymorphic transformations, and other reactions. In order to communi-
cate the mechanical energy of plastic materials, the last are processed in extru-
sion reactors by using the crushers and conveyors. During of such a treatment,
in addition to shredding polymers, are taking place the structural changes of
substance, there are formed a lot of flaws and the material becomes reactive. In
the simultaneous processing of several types of polymers, between them occur
interactions–chemical reactions. However, analogous in the case of the thermal
activation in the solid-phase reactions, for the initiation of mechanical-chemical
reactions it is necessary to transmit the dust a sufficient quantity of mechanical
energy. The energy can be communicated through the crusher and polymer screw
reactor. The screws acting through the impulses on the polymer shredding. The
strength and the result (the degree of transformation after the chemical reac-
tion) of such actions depend on the acceleration of the movement of the screw
within the reactor. Previously, the use of mechanical activation of substances in
the production of new materials has been limited by the lack of reactors oper-
ating safely. Currently, such reactors exist and can be used in such a process
harmless for the environment as a secondary plastic processing [7]. By a project
of technology transfer, in the Republic of Moldova was implemented a complex
process of waste plastics [8].
Waste of plastics separated from polyolefin (PE, PP), unwashed are intro-
duced into the hammer mill where it is comminuted up to 50 mm parallel the
polyolefin are crushing separately, and then they are crushing separately until
the fraction 0.5 mm up to 10 mm (Fig. 4 and 5).
Comminuted and mixed waste in the proportion of 1/3 of the polyethylene
and polypropylene: 3/4 of plastics waste with unidentified composition is loaded
into the bunker with a speed of 0.2 m/s through the screw charger.
Fig. 4. Waste of plastics
Fig. 5. Waste of plastics comminuted by PP comminuted with containing unidentified
Thus, in the extruder (stainless reactor) the mixture is heated to a tem-

perature of 180–2100 C. The speed of charging of comminuted waste of plastic
materials are adjusted, for the temperature of self-heating of the mixture in the
extruder (steel reactor) to vary in the range of 180–2100 C and no more.
At this temperature (polyethylene) (polypropylene) is liquidized and other
compounds remain as viscous (polystyrene PS, polyvinyl chloride, PVC, poly-
ethylene terephthalate, PET serve as compounds of liaison and polycarbonates,
polyamides, etc. constitute the filler mixture. The obtained viscous mass is
homogenized and transferred to the calibrator.
Calibrated mass of material is transported in the matrix pressing machine,
where under the action of pressure from 3 bars is obtained the finite product:
forms for plastic pallets (Fig. 6).
Respecting the same regularity of the technological process, but varying the
report of homogenization of plastic waste comminuted in a proportion of 1/3
parts of polyethylene and polypropylene 3/4 of plastics waste with unidenti-
fied composition may be obtained new forms of the sidewalk (Fig. 6). To the
Fig. 6. New products obtained from waste of plastic mass
processing of 1,000 kg of the waste of plastics, may be obtained 950–970 kg of

new products from plastics.
In the Republic of Moldova, in cooperation with the Academy of Sciences of
Moldova for the first time, it was implemented a complex technology of process-
ing from waste plastics with obtaining of new products. This technology creates
no other waste and at the same time don’t generate any wastewater (Fig. 7).
Fig. 7. The sector of complex processing of plastic waste
2.2 Infectious Medical Waste Management in Republic of Moldova
Medical wastes management activity is a part of nosocomial infection control in

health facilities (IMS). Nosocomial infections are listed as special health issues
(Regulation on the national epidemiological surveillance and control of commu-
nicable diseases and health events, Gov.Dir nr.951 of 25.11.2013).
According to the Statistical Yearbook of the health system for 2014 in
Moldova (Right Bank) are 348 state medical institutions (legal entities) and
1029 are private.
There are 3 categories of infectious medical waste generators:
Large manufacturers: Republican, municipal, district and departmental hos-
pitals; University of dispensary clinics, Medical Research Institute, Institute of
Forensic Medicine, Forensic Regional Services, Pre-Clinical Departments of the

University of Medicine, Faculty of Medicine and Pharmacology, etc.
Medium manufacturers: Medical institutions of primary care clinics of hos-
pitalization, blood transfusion centers, microbiological laboratories and clinics,
nursing, medical and dental offices, private hospitals and clinics, etc.
Small producers: dental laboratories, rehabilitation centers, funeral services,
vision care centers, acupuncture clinics, preventive medicine centers, private clin-
ics, etc.
Among all categories of waste products of medical activity in medical insti-
tutions as a result of daily care activities, are the following categories of waste
(Regulation of medical waste management no. 06.8.3.45 of 10.12.2001, approved
by the Doctor Chief State Sanitary no. 05-00 of 14 December 2001 (Official
Monitory of the Republic of Moldova, 2002, No. 13–15 art. No.:29).
“Hazardous waste” is the waste resulting from medical activities which
present a real risk to human health and the environment, which are gener-
ated in the course of diagnosis, treatment, surveillance, disease prevention and
rehabilitation, including medical research and production, testing, storage and
distribution of medicines and biological products;
“Medical waste assimilated with domestic waste” are non-hazardous waste
whose composition is similar to waste and no major risk to human health and
the environment;
“Pathological waste and parts (parts) anatomical” these include tissues and
organs, anatomical results of surgical instruments, the autopsy and other medical
procedures; this category also includes animals used in laboratory diagnostic
work, research and experimentation;
“Pathological waste and anatomical parts” are wastes include tissues and
organs, anatomical results of surgical instruments, the autopsy and other medical
procedures; this category also includes animals used in laboratory diagnostic
work, research and experimentation;
“Infectious waste” are liquid and solid wastes containing or contaminated
with blood or other body fluids and materials that contain or have come in
contact with viruses, bacteria, parasites and/or toxins microorganisms;
“Chemicals and pharmaceutical waste” posed by chemicals are solid, liquid or
gas that can be toxic, corrosive or flammable; expired drugs and chemotherapy
residues of substances that can be cytotoxic, genotoxic, mutagenic, teratogenic
or carcinogenic.
“stinging-cutting waste” is waste that can cause mechanical damage by punc-
turing or cutting;“Radioactive waste” is solid, liquid and gaseous waste resulting
from nuclear activities medical diagnosis and treatment, containing radioactive
material;
According to the World Health Organization, 80% of all waste resulted from
medical activities are considered inoffensive and are managed as waste and other
hazardous materials, 20% are infectious, toxic or radioactive.
The survey data conducted in 2004 with the support of the World Health
Organization show that in Republic of Moldova the estimated amount of haz-
ardous waste from medical activity is 0.05 kg/patient/day treated at home for
hospitalized cases–average 0.44 kg/bed/day. Thus the total production of haz-
ardous waste was estimated between 10 and 11 tons of waste per day 4000 tons
per year.
Currently, the total amount of waste reported by medical institutions has
increased, accounting for about 8.7 million tons per year, including infectious
waste (syringes, infusion catheters, etc.).
In Chisinau daily it is accumulate about 3 tons per day of waste medical activ-
ity annually–about 1095 tons of wastes. At the same time, 120 million syringes
annually and 6 million infusion systems are used throughout the country.
According to the World Health Organization, the estimated operational costs
for treating 1 kg of infectious waste varies between 0.13–2.2 US $, depending on
the applied method. The lowest costs for treating infectious waste is the method
by autoclaving (0.13–0.36 US $/kg).
Academy of Sciences of Moldova and the Ministry of Health and Ministry
of Environment in 2015 implemented a pilot project for treatment of infectious
medical waste. It was developed an authorized treatment of infectious medical
waste (Figs. 8 and 9).
Fig. 8. Plan scheme of waste treatment point
In Fig. 8, 1 is working room of the plant; 2, 3, 4, 6, 7, 10, 11 are public

spaces in the clean and dirty; 8 is room for receiving and temporary storage of
waste (dirty area); 9 is room for washing and disinfecting bins; 12 is room for
temporary storage of bins after washing and disinfection process; 13 is room for
temporary storage of waste processed-disinfected; 5, 15 are rooms for personal
and workplace.
Fig. 9. Authorized infectious medical waste processing point
STERISHRED technology and facility for treating hazardous waste from

medical activities implemented at this enterprise, allows the obtaining of the
final product environmentally friendly and health. Processing infectious waste
takes place at low temperatures (136◦C) with grinding, sterilization and destruc-
tion of confidential information storage units, in one compact space. The device
converts waste “yellow bags” (biologically hazardous), including stinging and
sharp objects in non-hazardous municipal waste-sterilized. Waste results, up to
70% can be processed at other companies for the obtaining of new products.
3 Conclusion
• Moldova has resources and scientific potential for capitalizing secondary wine
products.
• Create a pilot plant for treating waste plastics complex served as an impetus
for development of the field of waste recovery in Moldova.
• Academy of Sciences collaboration with the private sector has led to an autho-
rized point of treating infectious medical waste, which resolved partially the
pressing problem of neutralizing medical waste in Moldova.
References
1. Braunegg G, Bona R et al (2004) Solid waste management and plastic recycling in
Austria and Europe. Polymer-plastics technology and engineering 43(6):1755–1767
2. Chisinau Statistica (2010) Anuarul statistic al republicii moldova
3. Damani N, Koolivand A et al (2013) Hospital waste generation and management
in some provinces of Iran. Toxicological Environ Chem 95(6):962–969
4. Duca G (2011) Produse vinicole secundare
5. Ikeda Y (2016) Current home medical care waste collection status by nurse in
Japan. J Air Waste Manag Association
6. Kumar S, Mukherjee S, Chakrabarti T, Devotta S (2007) Hazardous waste manage-
ment system in india: an overview. Critical Rev Environ Sci Technol 38(1):43–71
7. Macaev F, Bujor S, Mereuţa A (2011) Reciclarea deşeurilor din mase plastice prin
procedee mecanochimice
8. Macaev F, Mereuţa A, et al (2016) Procedeu de reciclare a deşeurilor de mase
plastice/brevet de invenţie md 949 z
9. Sivapullaiah PV, Naveen BP, Sitharam TG (2016) Municipal solid waste landfills
construction and management—a few concerns. Int J Waste Resources 6(214):
10. So WMW, Cheng NYI, et al (2015) Learning about the types of plastic wastes:
effectiveness of inquiry learning strategies 44
11. Tugui T, Duca G, et al. (2013) Municipal solid waste composition study
12. Vyas P (2011) Municipal solid waste in India. J Ind Pollut Control 27(1):79–81
How to Popularize Green Residential Buildings
in China: A Survey Study from Sichuan
Jun Gang1(B) , Dirong Xu2 , and Ying Wei1

1
Sichuan Institute of Building Research,
467488122@qq.com
2
College of Architecture and Urban-Rural Planning, Sichuan Agricultural
University, Chengdu 610003, People’s Republic of China
Abstract. Green buildings have been ushered in a period of rapid devel-

opment in China. For there are few researches on economic benefit analy-
sis, a result that most of Chinese people don’t want to purchase residen-
tial buildings because of misunderstanding that green buildings have a
higher cost. This paper focuses on how to popularize green residential
building in China. In order to achieve the goal, the field survey is car-
ried out, qualitative and quantitative descriptive methods are used to
analyze the statement of green residential building in Sichuan Province.
There are two important conclusions. Firstly, the incremental cost con-
trol of green building in our province is relatively reasonable. However,
the incremental selling price is much higher than the incremental cost.
Secondly, ordinary people can accept is much lower than the selling price.
Finally, in order to improve the popularization of green residential build-
ing, some suggestions are provided.
Keywords: Green residential buildings · Survey study · Statistical

analysis · Popularization
1 Introduction
China has been the world’s largest carbon emitter and the world’s largest energy
consumer country since 2011. Building is one of the main contributor of energy
consumption and pollution emission. China’s building energy consumption has
more than industrial and transportation, as the first source of energy consump-
tion. To reduce the energy consumption, China had proposed the concept of
green buildings in 2004 [7], which aim to save resources (including: energy, land,
water and materials) and reduce the environmental impact. In the past five years,
China’s green building ushered in a period of rapid development. Green Building
has first stated explicitly in the Five-Year (from 2011 to 2015) Plan of China
[3]. In 2013, China’s Green Building Action Plan has been launched [5]. And in
2015, the new version of “assessment standard for green building” was formally
How to Popularize Green Residential Buildings in China 1297
implemented [2]. All these have brought positive results to promotion of green
building. And also, more and more scholars have paid their attention to the
green building in China, not only involving technology [4], but also containing
cost [1] and management [8].
Although more and more peoples are showing solicitude for green building,
but most of them are architects, engineers, scholars, and others related to the
construction industry. Green buildings are not yet popularized since most com-
mon peoples don’t still understand its concept and value. In China, all evaluation
standards on green buildings don’t contain economic indexes. The new version
of “green building evaluation criteria” explicitly proposed that green building
evaluation should take the economic analysis into consideration, but it did not
put the economic analysis into the comprehensive evaluation system. Researches
on economic benefit analysis are also seldom found [6]. As a result, most of Chi-
nese people believe that green buildings (especially residential buildings) always
have higher cost, so they don’t want to purchase residential buildings. That an
important reason why residential buildings cannot be popularized in China.
In order to understand the common peoples’ attitude to green residential
buildings fully, we did carry out an investigation to 34 green residential buildings
and 935 peoples in Sichan in January, 2016. The following several sections will
show the investigation data and some statistical analysis in detailed.
2 Survey Statement
This investigation mainly includes three aspects: (1) the public foundation of
green residential buildings; and (2) the incremental cost and incremental price of
green residential buildings; and (3) the people’s acceptability to the incremental
price of green residential buildings. Among them, the first part mainly surveys
local residents’ cognitive level and purchase intention about Sichuan’s Green
Buildings; the second part mainly investigates the incremental cost and the
incremental price of green residential buildings in Sichuan; the third part mainly
researches the public acceptability to the price of green residential buildings,
and analyzes the people’s needs and purchase intentions to green buildings in
different hierarchy.
The questionnaire method was used to the ordinary peoples who are chosen
randomly. This sampling survey was issued 1000 questionnaires, which 935 were
recovered, and 901 are valid. The effective recovery rate of the questionnaires
was 90.1%. The basic information of the respondents is as follows Table 1.
In addition, we also generally investigated all the 34 green residential build-
ings in Sichuan, which information is showed as Table 2.
3 Popularization Degree
3.1 Knowledge Level About Green Residential Buildings
In order to find out the popularized degree of green residential buildings in
Sichuan, three questions are designed: (1) Do you understand the conception of
1298 J. Gang et al.
Table 1. Basic information of the respondents
Statistical variables Sampling Percentage (%) Statistical variables Sampling Percentage (%)
Sex Education
Male 599 66.48 High school or below 104 11.55
Female 302 33.52 College graduate or bachelor 632 70.14
Profession Bachelor or above 165 18.31
Self-employed 101 11.21 Monthly income RMB
White-collar 119 13.43 7000 417 46.28
Ordinary worker 338 37.51 7000 14000 303 33.63
Student 273 30.3 14000 21000 99 10.99
Other 68 7.55 21000 82 9.1
Table 2. The basic information of the 34 green residential buildings
One-star Two-star Three-star

Design identification 23 6 3
Operation identification 2 1 0
Total 25 6 3
green residential buildings? (2) Do you understand the functions of green resi-
dential buildings? (3) Do you know the policy of green building? The statistical
results of the two questions are shown in Figs. 1, 2 and 3:
Understand
completely
Understand Don't 1%
14% understand,
58, 28%
Understand a
little, 117,
57%
Fig. 1. People’s understanding to the concept of green buildings
Understand
completely Don't
Understand 0.4% understand,
24% 32, 16%
Understand a
little, 122,
60%
Fig. 2. People’s understanding to the functions of green buildings

Know, 11, Very Know,

5% 2, 1%
Unknown,
97, 48%
Know a little,
94, 46%
Fig. 3. People’s understanding to the policies of green buildings
The survey results show that the people who know green buildings very
clearly were accounted for only about 1%, some clearly only 14%; and under-
standing for the concept of green buildings, most people yet stayed in the “a
little” and “a little more” dimensional level. Similar results are found about
the functions and policies. 76% of the people don’t understand or only under-
stand a litter about the function of green residential buildings, while 94% of the
people don’t clearly understand the related policies about green buildings. How-
ever, in the process of investigation, we found that some people believe that the
“green buildings” is the energy-saving buildings, and some people think “green
buildings” is the buildings with many greenbelts. As a whole, the popularization
of green residential buildings is serious shortage, although there is good policy
support.
3.2 Purchase Intention to Green Residential Buildings

The purchaser of green residential buildings are the ordinary peoples, the govern-
ment only plays a promotion role. Only when the public really understand what
is the green building and how important to their live, it is possible to popularize
green buildings. After analyzing the people’s knowledge levels about green build-
ings, a question involving people’s purchase intentions to the green buildings is
designed, and survey results are showed in Figs. 3 and 4. It can be seen that 29.8%
of the people think there is no necessary to develop green buildings in the first
round of the investigation. However, after listening to the investigation team’s
explanation about the concept and functions of green buildings, 93.7% of people
express their intentions to support the development of green buildings, which
shows a good and extensive public foundation to popularize green residential
buildings. Of course, it is necessary to enhance the publicity before populariza-
tion. Similarly, when investigated whether are willing to buy green residential
buildings, after listening to the investigating group’s carefully explanation, in the
situation without considering incremental price, about 93.14% of respondents are
willing to buy green residential buildings. In addition, nearly 7% of respondents
are unwilling to buy green buildings, because of the worry about the incremental
price. Indeed, the prices of green residential buildings will increase with the rise
1300 J. Gang et al.
Very
unnecessary
Unnecessary
Necessary
Very
necessary
0 100 200 300 400 500 600

Befor explanation After explanation
Fig. 4. Necessity of popularizing green residential buildings
100% 93.14%
80%
58.33%
60%
41.67%
40%
20% 6.86%
0%
Willing Unwilling
After explaration Before explaration
Fig. 5. People’s willingness to buy green residential buildings
of design and construction cost. However, cost is not the unique factor which
affects the selection to green buildings. It is found from existing literature that
the main factors about the popularization of green residential buildings mainly
include: government funding, comfort levels, crowding effects, cost factors and
so on. Specifically, the survey found that most people give priority to benefit and
comfort (36.82%). Followed by is the relevant government funding (29.1%). And
22.89% of the respondents tend to the benefit because of the diminishing cost
in the operation phase. In addition, 10.2% of the respondents would buy green
buildings if their relatives and friends have bought. In summary, the public has
a high recognition for green residential buildings. They think it is necessary to
vigorously promote and build the buildings, and also almost all are willing to buy
this kind of building. This shows that the green buildings have a stable audience
groups, there is a huge potential market to be tapped. But how to activate the
vitality of the market, to stimulate people’s actual purchase desire. The most
important thing is that balance the relationship between the construction cost
and the final benefit, that is, not only let people realize the bringing benefits of
green residential buildings, and not simple to raise the price too much.
Government‘s
funding
29.10%
Other reasons 36.82% More comfortable

1.00%
10.20%
22.89%
Cost reduction in Oriented by
use phase peoples around
Fig. 6. Main factors affecting the purchase intention
4 Cost and Price

4.1 Incremental Cost and Incremental Price Statistics
From the statistics in Table 3, it can be seen that the incremental cost of green
residential buildings in Sichuan is similar to that in China, which is controlled
within 70 (CNY/m2 ), the average incremental cost about 35 (CNY/m2 ). It is
less than 1% of total housing construction costs. It shows that the incremental
cost of green buildings in Sichuan is reasonable.
Table 3. Incremental cost statistics
One-star (CNY/m2 ) Two-star (CNY/m2 ) Three-star (CNY/m2 )

Average value Scope Average value Scope Average value Scope
China 33 10 70 73 23 138 222 36 492
Sichuan 35 0–67 68 45–132 – –
Notes: 1 The nationwide data derives from the research results of “Post-assessment of
green building in China”, which is supported by the Ministry of Housing and Urban-
Rural Development of China and led by of Chinese Society for Urban Studies; 2 The
data of Sichuan from the statistical results to 34 green residential projects from 2011
to 2015
At the same time, we also carried out statistics to the price of one-star
green residential buildings in Sichuan, the results as shown in Table 4. It shows
Sichuan Province’s the average price of one-star green residential buildings is
7235 (CNY/m2 ), and the average incremental price is 284 (CNY/m2 ). Among
them, the average price and average incremental price of the city of Chengdu are
the highest. They are reached 9345 (CNY/m2 ) and 385 (CNY/m2 ), respectively.
Compared the data in Tables 3 and 4, it can be found that the incremental price
of green residential buildings in Sichuan is much higher than the incremental
cost. The main reasons are as follows: (1) there is no a unified statistics standard
to the incremental cost of green buildings in Sichuan, which leads the reported
1302 J. Gang et al.
Table 4. Incremental price statistics
Actual selling price (CNY/m2 ) Incremental price (CNY/m2 )

Scope Average value Scope Average value
Sichuan 4200–11500 7235 94–684 284
Chengdu 7800–11500 9345 136–684 385
Mianyang 4200–5100 4780 94–346 106
Nanchong 4900–5900 5360 115–526 218
incremental cost of some project seriously distorted; and (2) the current statistics
to the incremental costs of green buildings only contain one-time investment
cost, without considering the cost in the whole life cycle, such as cost in project
planning, consulting, and operation stage. For example, although the consulting
fees of green building must be produced for all the green buildings labeling
programs, they are not counted in the reported incremental cost; and (3) for
the same real estate, green buildings than other buildings often are occupied
better natural resources (such as location, direction, greening, ventilation and
etc.), and thus prices will be increased; and (4) housing prices than the cost of
housing will usually retain a certain profit space, thus widening the direct gap
between the price and the cost.
0, 500)
[220 [500, + )
7%
7 %
2%
0)
[140, 220 [0, 70)
18% 29%
Unit:
CNY/m m2
[70, 140)
44%
Fig. 7. People are willing to pay the proportion of the incremental price range
4.2 The Acceptability to the Incremental Price
In order to clarify the cost and efficiency of green residential buildings, a survey
about public’s acceptability to the incremental price is carried out. A question
including five incremental price levels is set. The survey results are showed in
Figs. 6–8. The results show that: (1) 73% of people are only willing to buy green
residential buildings with increase price less than 140 (CNY/m2 ); (2) only 8.8%
of people can accept the increase price more than 220 (CNY/m2 ). Thus, except
Mianyang, the incremental prices of green residential buildings are slightly high
in Sichuan. Especially in Chengdu, the incremental prices of green buildings
are much higher than the acceptable price of common people. Therefore, with
enhancing the propaganda of green buildings, the price of green residential should
also be reduced appropriately.
450
4
400
4
350
3 %
28%
300
3 Agges:
%
5%
3
30-50
250
2
1
10-20
7%
37
200
2 2
20-30
5
50-70
150
1 4%
14 64%
% 32%
%
100
1 %
11%
47
7% 25%
%
50 54%
% 25%
8%
67%
2%
% 2%
% 3% 75%
0
0-70
0 70-140
0 0
140-220 220-500 500-1000 (Unit: CNY
Y/m2)
Fig. 8. Acceptability to incremental prices for different ages
5 Policy Proposals
In order to popularize green residential buildings, based on the survey results and
statistic analysis above, the following policy recommendations are put forward.
5.1 Strengthen Publicity of Green Residential Buildings
From the above statistical analysis, it can be found that one of main reasons for
hindering the development of green residential buildings in Sichuan is that the
common people know too little about green buildings. The survey found that
the vast majority of people are unclear about “what is a green building”, “what
can it bring what”, “whether it increases the cost or not” and other core issues,
which lead to a very low public acceptance for green buildings. At present, in
China, only a small circle of people are working for green buildings. There are
no specialized agencies or platform to popularize and propagate green buildings.
1304 J. Gang et al.
1
100%
10.5% 14.3%
90%
80% 13.6%
70%
60%
50%
40%
30%
20%
10%
0%
<7000 7000-14000 14000-2100
00 >2100
00
<70 70-140 140-2
220
Fig. 9. Acceptability to incremental price for different income levels
Also, there is no corresponding funding support. Therefore, it is recommended

that the government should increase budget to organize or guide some capable
and responsible organizations or enterprises to give publicity to green buildings,
including the conception, technology, production and benefit, to improve people’s
cognitive level for green buildings.
5.2 Build the Basic Database of Green Buildings
By the end of 2015, 3036 projects have obtained the identification of green
buildings in China, and 56 in Sichuan. However, this information is only posted
on few government websites, including the official website of ministry of housing
and urban rural development of China, and some local official website. Moreover,
these data are often published as public announcements, which are often too
simple to understand. The project information in detail is always put in the
attachment, and there is an open database to store them. As results, there are
no channels to know and identify the green buildings’ information, which hinder
the further development of green building badly. So it is suggested that the
government should establish official database on green buildings and open them
to common people.
5.3 Drawing up Incremental Cost Statistical Standard
As describe above, the statistics on incremental cost of green buildings is variety

in China. Unlike the German DGNB and Canada GBC, there is no strict require-
ments and uniform standard on the statistics of incremental cost in China. Some
projects are counted rigorously, so reported relatively high incremental costs;
however, some other projects are only counted minimal incremental cost items,
some even no statistic, which lead to the unreasonable incremental cost. This
brings serious difficult to correctly estimate the green buildings’ costs and ben-
efits, also to the management and service. In this situation, it is recommended
that the unified standard about the incremental cost statistics of green build-
ings should be drawn up. In addition, the local government should require all
green building projects must announce the data on incremental cost. Although
this will increase cost slightly, it will pay an important role to popularize green
buildings.
5.4 Enhance Cost and Price Control

From the above statistics analysis, it can be seen that the incremental price of
green residential buildings in Sichuan was significantly higher than the afford-
able price of common people. Therefore, in order to promote the comprehensive
development of green buildings, the price must be controlled. According to sta-
tistical data, the incremental price of green residential buildings should be con-
trolled within 140 (CNY/m2 ) as far as possible, especially for the one-star green
residential buildings. In addition, the developers should also consider different
demander according to the different price levels. For example, the projects which
incremental prices are more than 140 (CNY/m2 ) should focus on high-income
groups. If the incremental price cannot be controlled to less than 500 (CNY/m2 ),
then the average monthly income of potential buyer should be more than 21000
CNY.
6 Conclusion
Although green residential buildings as the trend of the future development,
many people still don’t know exactly what’s the green buildings. A green build-
ing not only means the vertical greening or roof garden, but also refers in the
construction of the whole life cycle, to maximize the energy saving, land saving,
water saving, material saving, environmental protection and pollution reduc-
tion. It is also called the sustainable development construction, ecological con-
struction, energy saving and environmental protection construction etc. Looking
around the property market in Sichuan, the green buildings are still very lit-
tle. Three main reasons have been found according to the survey study in this
paper. First, the common people know too little about green residential build-
ings. Second, the incremental prices of green residential buildings have exceeded
the buyers’ psychological expectations. Third, the environmental benefit and
economic benefit for buyers are not widely accepted. Therefore, to popularize
the green residential buildings, at the same time of improving the comfort, it is
also necessary to strengthen the science propaganda and control the incremen-
tal price.
Acknowledgement. This research was supported by the special funds for building
energy efficiency of Sichuan, and system science & enterprise development research
center of Sichuan (No. Xq15C14).
1306 J. Gang et al.
References
1. Dwaikat LN, Ali KN (2015) Green buildings cost premium: a review of empirical
evidence. Energy Buildings, 110:396–403
2. GB, T (2014) Assessment standard for green building. China Building Industry
Press, Beijing
3. (2011) The twelfth five-year plan for national economic and social development of
the People’s Republic of China. Chin Nurs Res 25(3):207–214
4. Olubunmi OA, Xia PB, Skitmore M (2016) Green building incentives: a review.
Renew Sustainable Energy Rev 59:1611–1621
5. The Ministry of Housing and Urban-Rural Development of the People’s Republic of
China (2013) Green building development plan. http://www.gov.cn/zwgk/2013-01/
06/content 305793.htm
6. Yang Y (2015) Study on economic cost benefit evaluation of green building technol-
ogy. Master’s thesis, Southwest Jiao Tong University, Chengdu (in Chinese)
7. Ye L, Cheng Z et al (2013) Overview on green building label in China. Renew Energy
53(9):220–229
8. Ye L, Cheng Z et al (2015) Developments of green building standards in China.
Renew Energy 73:115–122 (in Chinese)
Prognostics of Lithium-Ion Batteries Under
Uncertainty Using Multiple Capacity
Degradation Information
Fan Li1(B) and Yusheng Wang2

1
School of Economics, Sichuan University,
lifanscu@sina.com
2
Uncertain Decision-Making Laboratory, Sichuan University,
Abstract. Prognostics of lithium-ion batteries play an important role in

the intelligent battery management systems. The State of health (SOH)
estimation for batteries often needs to be implemented under uncertain
situations for the complicated operating conditions. In this work, a novel
integrated approach based on a mixture of Gaussian process (MGP)
model is presented for lithium-ion battery SOH estimation under uncer-
tain conditions, where the degradation process parameters distribution
can be learnt from the multiple available capacity measurements. To
capture the time-varying degradation behavior, the distribution infor-
mation of the degradation model parameter is extracted by fusing the
training data from different battery conditions with the MGP model.
Moreover, by exploiting the degradation model parameter distribution
information, the PF algorithm is employed to predict the battery SOH.
The correlation case study and comparison analysis are provided to show
the efficiency and effectiveness of the proposed prediction method.
Keywords: Prognostics · Lithium-ion batteries · State estimation ·

Uncertainty
1 Introduction
With the advantages of high energy density, high galvanic, low-temperature per-
formance, and long lifetime, lithium-ion batteries play a more important role
in electronics system energy supply, and therefore have been widely used in
communication, aerospace avionics, portable devices and other industrial areas
[5,8]. However, battery deterioration and battery failure commonly occur, which
can lead to a reduction in systems performance, result in increased costs and
catastrophic failure [14]. Therefore, for the electronics device, prognostics and
health management (PHM) has received increased attention in battery manage-
ment system to determine the advent of systems failure and to mitigate system
1308 F. Li and Y. Wang
risk through the evaluation of system reliability in terms of the current life-cycle
conditions [13,16].
In order to evaluate the current health situation, some indicators such as
the state-of-charge (SOC), the state-of-health (SOH) and the state of life (SOL)
are commonly used to quantify the degradation of lithium-ion batteries [17]. To
provide useful prognostics information in the reliability monitoring for battery
systems health management, current researches have focused on SOC and SOH
estimation for batteries. The main measurements required to effectively measure
battery SOH are impedance and capacity [7]. In this study, battery capacity
percentages relative to the initial capacity are adopted to measure the SOH.
In recent years, there are many valuable research works on the battery degra-
dation modeling and prognostics for battery SOH estimation [9,15]. For example,
neural networks and artificial intelligent methods have been rapidly used to esti-
mate the battery SOH and predict the remaining useful life [11]. Stochastic fil-
tering approaches such as Kalman filtering [1], extended Kalman filtering [4,10],
unscented filtering and Bayesian filtering [2] are widely adopted in the SOH esti-
mation of battery, in which the empirical degradation models are often used to
build the dynamical system equation. Usually, the stochastic filtering methods
based on the parameter process or degradation model can show good perfor-
mance only when the degradation model used represents the actual system’s
behavior effectively. However, in practical battery conditions such as operating
conditions, environmental conditions, and other complicated inherent system, it
is difficult to obtain accurate state process model or parameter description when
the uncertainties are considered. To address this issue, many approaches have
been proposed such as a prognostic algorithm based on a Bayesian Monte Carlo
method and the Dempster-Shafer theory for battery SOH estimation and RUL
prediction [3]. Liu et al. [7] used the Gaussian process regression (GPR) to per-
form SOH prediction for lithium-ion battery, where the degradation trends are
learned from battery data sets with the combination of Gaussian process func-
tions. It should be noticed that uncertainty representation of the degradation
model has not yet been fully investigated.
As mentioned above, it is urgent to develop the battery prognostics under
the uncertainties. In this work, a novel approach to lithium-ion battery SOH
estimation is presented through the integration of the MGP model learning and
particle filtering. The proposed method consists of two phases, and the first is the
MGP is used to learn the statistical properties of the degradation model para-
meter combining training data sets from uncertain battery conditions. Secondly,
based on the parameter distribution information for the degradation process,
particle filtering is exploited to obtain the battery SOH estimation. In the train-
ing phases, the GPR is exploited to initialize the distribution parameters for
each component. Then, the MGP learning and the PF updating are recursively
implemented. Finally, a case example is provided based on the NASA battery
data sets to show the performance of the new prognostics method. The con-
tributions of this study can be summarized in two points: the first is a fusion
prognostics framework for the lithium-ion battery SOH estimation is developed
Prognostics of Lithium-Ion Batteries Under Uncertainty 1309
by combining the degradation conditions from different batteries, the second

is the proposed algorithm implements distribution learning for the multimode
process under uncertainty.
The rest sections of this paper are organized in the following order: the GPR
and the MGP model are introduced in Sect. 2. Section 3 details the proposed SOH
prediction method which incorporates MGP learning and PF. Experiments and
analysis are given in Sect. 4 to show the performance of proposed prognostics
algorithm. Finally, we make the conclusions in Sect. 5.
2 Related Work
2.1 Gaussian Process Regression
The degradation model parameter vectors need to be treated as a dynamic
process which captures the time-varying situations in the battery degradation
cycles. To represent the system degradation behavior, The Gaussian process is
considered. A stochastic process {g(x) : x ∈ X }, indexed by elements from some
set X , is a Gaussian process with mean function m(x) and covariance function
k(x, x ). It means for any finite set of elements x1 , · · · , xm ∈ X , the associated
finite set of random variables g(x1 ), · · · , g(xm ) have a multivariate Gaussian
distribution, i.e.,
⎡ ⎤ ⎛⎡ ⎤ ⎡ ⎤⎞
g(x1 ) m(x1 ) k(x1 , x1 ) · · · k(x1 , xm )
⎢ .. ⎥ ⎜⎢ . ⎥ ⎢ .. .. .. ⎥⎟
⎣ . ⎦ ∼ N ⎝⎣ .. ⎦ , ⎣ . . . ⎦⎠ .
g(xm ) m(xm ) k(xm , x1 ) · · · k(xm , xm )
It can be denote by g(x) ∼ GP (m(x), k(x, x )). Actually, the mean function
and covariance function are defined as
m(x) = E[g(x)],
k(x, x ) = E[(g(x) − m(x))(g(x ) − m(x ))],
for any x, x ∈ X .
Gaussian process represent distributions over functions, it provide a method
for modeling probability distribution under multiple corruptions in complicated
or uncertainty situations. When the accurately describe for the dynamical para-
meter process is hard to be obtained in advance, the Gaussian process regression
(GPR) can be exploited to supply the approximation distribution of the para-
meter process through learning from the training data available [6].
Consider a set of training data S = {xi , yi }Ni=1 , the relationship between
input xi and output yi can be modeled by yi = g(xi ) + i , where i is zero mean,
Gaussian white noise with variance σn2 . From the GPR, if the prior distribu-
tion over g(xi ) be assumed as Gaussian process, the posterior distribution over
outputs conditioned on sample set S and the test input x∗ is also a Gaussian
process, and the mean and variance can be given by
g¯∗ = E[g∗ |x∗ , S] = k∗T K −1 y, (1)
Cov(g∗ ) = k(x∗ , x∗ ) − k∗T K −1 k∗ , (2)
where, K is the covariance matrix whose entries is determined by the kernel

function i.e. Kij = k(xi , xj ). In order to capture the uncertainty of state process
and the influence of noise, the kernel function can be expressed as k(xi , xj ) =
kg (xi , xj ) + kn (xi , xj ), and the squared exponential kernel and constant kernel
function are chosen as the covariance function of Gaussian process:

(xi − xj )2
kg (xi , xj ) = σg2 exp − , (3)
2
kn (xi , xj ) = σn2 δij . (4)
The parameters Θ = [σh , σn ] are so-called hyper-parameters of the Gaussian

process, which are actually the parameters of the covariance function. The hyper-
parameters can be determined by exploiting numerical optimization techniques
which is to maximize the log-likelihood function given by
1 1 N
logp(y|X, Θ) = − y T (K + σn2 I)−1 y − log(|K + σn2 I|) − log(2π),
2 2 2
where, I is the unit matrix of N dimensions, and N is the number of training
data. Therefore, the learning process with GPR need to select the mean function
and covariance firstly, then the hyper-parameters can be determined with the
training data. After that, the prediction of the distribution parameter can be
yielded with the new inputs.
2.2 Mixture of Gaussian Process Models
Consider the uncertainties in the degradation of different batteries, the available

training data usually comes from several sources which may have different initial
capacities or were collected from different measurement conditions. Therefore,
to make full use of the various monitoring data, the MGP model is utilized to
model the prediction distribution of the degradation model parameters by learn-
ing the mean and covariance from the multiple input space. With the obtained
SOH measurements, the training set S is supposed to be composed of known
K-components, S1 , · · · , SK , where Sk = {xi , yi }N k
i=1 . For each training subset Sk ,
considering that there is a latent degradation model with uncertainty that can
be used to describe the degradation trend for different battery components, the
initial Gaussian processes are first used to learn the distribution parameters of
the degradation process. Then, the MGP model which is obtained as a weighted
combination of finite Gaussian process models can be exploited to fuse the dis-
tribution information from different components. Therefore, denote p(z(x) = k)
as the probability that sample x corresponds to the k component, then, with
the mixture weights πi , i = 1, · · · , K, the posterior probability of the samples
associated with the k th component is given by:
πk N (x|μk , Σk )
p(z(x) = k|x) = K . (5)
i=1 πi N (x|μi , Σi )
Let p(y∗ |Sk , x∗ ) be the predictive pdf for the output variable y∗ on the con-
dition that the new sample x∗ is obtained from the k th component. Thus, the
predictive pdf from the MGP models can be given as:
K

p(y∗ |S, x∗ ) = p(z(x∗ ) = k|x∗ )p(y∗ |Sk , x∗ ) (6)
k=1
where, p(z(x∗ ) = k|x∗ ) is the posterior probability that x∗ belongs to k th com-

ponent, and it can also be denoted as p(k|x∗ ).
Based on the structure of the MGP models, the state process for the battery
degradation parameters can be given through the distribution learning which
combines multiple possible degradation processes. It can be seen from Eq. (6),
the predictive distributions from the MGP models can be determined from the
posterior probability of the component the test sample belongs to and the asso-
ciated distribution parameters update from the Gaussian process components
when the new test sample is obtained. Commonly, the Expectation-Maximization
algorithm, which is based on the maximization of the log-likelihood, is used for
MGP model parameter optimization. To represent the time-varying process of
the degradation model parameters in the training phase, recursive learning for
the Gaussian mixture model can be exploited. In this way, denote the distrib-
ution learnt from k th GPR component at time t as {πk , μk , Σk }, when the
(t) (t) (t)
test sample x (t+1)

arrives at time t + 1, the mean and covariance of the k th
component, and its mixture weights can be updated as follows [18].
Step 1. Compute the posterior probability of the component:
(t) (t) (t)
π N (x(t+1) |μk , Σk )
p(k|x(t+1) ) = K k (t) . (7)
(t+1) |μ(t) , Σ (t) )
i=1 πi N (x i i
Step 2. Update the mixture weights:
= πk + γ(p(k|xt+1 ) − πk ).
(t+1) (t) (t)
πk (8)
Step 3. Update the distribution parameters:
p(k|xt+1 )
(xt+1 − μk ),
(t+1) (t) (t)
μk = μk + γ (t)
(9)
πk
t+1
p(k|x )
((xt+1 − μk )(xt+1 − μk )T − Σk ),
(t+1) (t) (t) (t) (t)
Σk = Σk + γ (t)
(10)
πk
where, γ can be viewed as the control parameter which can be set as 1/N where
N is the number of acquired samples.
3 The Proposed Method Based on MGP

The proposed method of SOH prediction has two phases. For the uncertainties in
the battery capacity degradation, the MGP is able to combine the correspond-
ing degradation process information under different conditions, thus, the MGP
model is used to learn the appropriate distribution to represent the transition

of the degradation model parameters in the training phase. In each mixture
component, the nonlinear and time-varying parameter processes are treated as
Gaussian process with associated means and covariance, which means the GPR is
exploited for initialization. In the prediction phase, with the training data under
the current condition, the trained MGP model is used to supply the impor-
tance density needed to produce the weighted samples in the Bayesian filtering
implementation.
In fact, there are two parts in the training phase. Firstly, in order to capture
the statistical properties for the degradation model parameters under uncertain
state transition, the GPR is utilized to learn the distribution information from
the different training sets. In each component, the training data can be given
by Dk = {(xi , yi )}N i=1 , where yi = Δxi = xi+1 − xi , thus, the parameter at
k
the next time instant can be expressed by adding the corresponding output
to the current state, as the output with each input xi is set as the change
of the parameters. Then, for each component, the mean ḡNk and covariance
(0)
cov(gNk ) are determined through the GPR Eqs. (1)–(2), and denoted μk = ḡNk
(0)
and Σk = cov(gNk ), which means that the initial mean and covariance are
given in the Gaussian process components. Secondly, when the initial distribution
parameters of MPG are given, the degradation model parameters pdf can be
represented by using the recursive algorithm for MPG with the current training
data. For the battery capacity data available from the current battery condition,
denote the current training data DT = {Δx(t) }Tt=1 as the parameter transition,
(t) (t) (t) (t)
where Δx(t) = x(t+1) − x(t) , and Θk = {p(k|Δx(t) ), πk , μk , Σk } are the
current distribution parameters for each component, then, the pdf of MGP can
be computed from the MGP parameters updating using Eqs. (7)–(10). After
that, the importance sampling and the resampling procedure are implemented
from the standard particle filter algorithm. With the particles and associated
[i]
weights, the battery capacity prediction SOHl+p at p step after current cycle l
[i]
can be computed by exploiting the degradation parameter samples x̃l for the
ith trajectory. Meanwhile, the SOH prediction at cycle l can be estimated by
combining the Ns samples of capacity measurements with their corresponding
weights. The steps of our proposed method are summarized as follows.
Step 1. Initialization:
Given the trained data set SL = {(xi , zi )}L
i=1 and the trained parameter
set in different components Dk = {(xi , yi )}N i=1 , (k = 1, · · · , K).
k
Step 2. Gaussian process regression-based learning in each component: For

(0) (0)
each component k = 1, · · · , K, {μk , Σk }=GPR(Dk ).
Step 3. Update the MGP parameters with current training data DT :
(t+1) (t)
Step 3.1 For each componentk = 1, · · · , K Θk = MGP(Θk ,
(t+1)
Δx )
Step 3.2 Compute the density of MGP using Eq. (6)
[i]
Step 4. Denote MGP(Δxl−1 ; Θ(T ) ) as the MGP density with the parameters
[i] [i]
Θ(T ) , and N (zl ; GPμ (xl , SL ), GPΣ (xl , SL )) represents the Gaussian
[i] [i]
pdf with the mean GPμ (xl , SL ) and the covariance GPΣ (xl , SL ), then
the importance sampling and resampling can be implemented:
Step 4.1 FOR i = 1, · · · , Ns
[i] [i] [i]
sample xl ∼ xl−1 + M GP (Δxl−1 ; Θ(T ) ),
[i] [i] [i] [i]
ωl ∝ ωl−1 N (zl ; GPμ (xl , SL ), GPΣ (xl , SL )).
END FOR
[i]
Step 4.2 Calculate the normalization weights ω̂l and effective sample
size Nef f
Step 4.3 IF Nef f < Nth
[{x̃l , ω̃l }N N
[i] [i] [i] [i]
i=1 ] = resampling[{xl , ω̂l }i=1 ].
Step 4.4 IF l < L, let l = l + 1, turn to Step 4.1;

[i] [i] [i] [i] [i]
Step 5. Estimate the capacity with weighted samples x̃l = (al , bl , cl , dl ):
[i] [i] [i] [i] [i]
SOHl+p = al × exp[bl · (l + p)] + cl × exp[dl × (l + p)],
Ns
[i] [i]
SOHl+p = ω̃l SOHl+p .
i=1
4 Case Example
4.1 Battery Data Set
In this section, a case study is conducted to validate the proposed approach.
The battery capacity data used in this paper are collected by the NASA Ames’
Prognostics Center of Excellence (PCoE) [12]. The lithium-ion batteries were run
through different operational profiles, such as charge, discharge and impedance
at room temperature. The battery capacity at cycles was adopted to measure the
SOH, and the end-of-life criterion in the repeated charge and discharge cycles
for the accelerated aging of the batteries was a 30% fade in rated capacity. Four
batteries No. ∼5, No. ∼6, No. ∼7, No. ∼18 are acquired through the discharge
with 2 A constant current level at an ambient temperature of 24 ◦ C until the
battery voltage fell to 2.7V, 2.5V, 2.2V and 2.5V, respectively.
To represent the empirical capacity degradation model, the conditional three-
parameter capacity degradation model as follow was considered, where the con-
stant variable κ was set at 2.
SOH = τ · exp(ι · l) + ρ · lκ . (11)
Using the Matlab curve fitting toolbox, the real degradation and the curve
fitting for battery No. ∼6 with the degradation model (11) are shown in Fig. 1.
The empirical degradation model parameters are (1.0263, 0.0037, 0), which were
estimated by curve fitting with the capacity samples given from cycle 1 to cycle
100. It can be seen from Fig. 1 that the real SOH at prediction cycles are higher
than the capacity degradation estimation. The root mean squared error is 2.8919
and the adjusted R2 is 0.947, which are the goodness-of-fit statistics of the
empirical model.
Fig. 1. The SOH of battery No. 6
4.2 The SOH Prediction
For the different operating profiles, the data from battery No. ∼6 was assumed
as the capacity samples under the current condition, and the data from other
batteries, which represented the different operating conditions, were treated as
the historical training data. In the training phase, there are two aspects. Firstly,
with historical training data sets for different batteries, the GPR was exploited
to learn the initial distribution parameters, where the mean function was cho-
sen from the linear function m(x) = ax + b and the covariance function was
chosen as the exponential function expressed by Eqs. (3)–(4). Therefore, the
hyper-parameters Θ = [a, b, σn , σg ] were optimized with the maximization of the
log-likelihood function. Secondly, the mixture density of MGP model is learnt by
exploiting the current battery training set with the obtained distribution para-
meters from the different components. After that, the obtained distribution from
data can be treated as the importance density, and then the particles and associ-
ated weights were updated using importance sampling and resampling procedure
implemented as the standard particle filter algorithm.
In this study, the number of particles was set at 200. Consider the current
battery No. ∼6, the prediction began at cycle 100, which means the training
data obtained from the cycle 1 to cycle 50 was used as the training set for the
MGP learning, and the SOH measurements from cycle 51 to cycle 100 were used
for the particles updating, and then the prediction results are shown in Fig. 2.
As shown in Fig. 2, the prediction combining three different training com-
ponents can capture the degradation trends of SOH in most cycles. Therefore,
from these results, the proposed method shows the effective prediction by fusing
the multiple training data which represent the uncertain capacity degradation
conditions.
Fig. 2. Prediction results for battery No. 6 at cycle 100 with different training sets
from batteries No. 5, No. 7 and No. 18
In addition, to further evaluate the proposed SOH prediction approach under

uncertain condition, the method based on MGP learning and particle filtering
(MGP–PF) was compared with the method which only considers the data from
current battery. The GPR was used for parameter distribution learning, then the
SOH estimation could also be implemented with the PF, thus the latter method
denoted GPR–PF. Consider the current battery No. ∼6, the prediction cycle
was set at cycle 100, then the prediction results with two methods are shown
in Fig. 3.
Fig. 3. Prediction results based on GPR-PF and MGP-PF for battery No. 6
For the battery No. 6, it can be seen from the Fig. 3 that the SOH prediction
based on MGP and PF performed batter than the methods which only took
the current capacity data into account. This indicates that the method which
combines the priori information of capacity degradation process can improve the
estimation performance in the current experiments. The reasons for this may lie
in that distribution approximate which only considers the current measurements
of battery ignores the useful fading information under different degradation con-
ditions. Therefore, the method which fuses knowledge from multiple capacity
degradation conditions is more suitable for lithium-ion battery SOH estimation
with the uncertain battery degradation situations.
5 Conclusion
In this paper, a new model for the battery SOH estimation is presented based on
the MGP model which was exploited to learn the distribution parameters from
multiple training sets on the different degradation conditions. To represent the
density of the degradation model parameters under uncertain conditions, the ini-
tialized distribution parameters can be updated recursively through MGP learn-
ing from the current capacity measurements of battery. Based on the degradation
model parameter distribution information, importance sampling and resampling
in the particle filtering framework was implemented. Our method is based on
distribution learning from training data and does not assume any certain state
model of degradation parameter, which is usually hard to be obtained in advance.
Therefore, the proposed approach can also be applied to other estimation prob-
lems with unknown state space models under uncertain conditions.
Lithium-ion battery degradation under complicated conditions or working
environments occurs in practice and different battery conditions can cause dif-
ferences in the degradation. The prognostics for battery SOH estimation under
uncertain conditions have many challenges, as different capacity degradation
models with various degradation features cannot be obtained in advance. The
uncertainties for the PHM of batteries is far more complex than the investigation
in this paper, so further research on the implementation of fusion models under
other uncertain conditions will be the focus of our future works.
References
1. Burgess WL (2009) Valve regulated lead acid battery float service life estimation
using a kalman filter. J Power Sources 191(1):16–21
2. Cui H, Miao Q et al (2012) Application of unscented particle filter in remaining
useful life prediction of lithium-ion batteries 2012:1–6
3. He W, Williard N et al (2011) Prognostics of lithium-ion batteries based on
dempster-shafer theory and the bayesian monte carlo method. J Power Sources
196(23):10314–10321
4. Hu C, Youn BD, Chung J (2012) A multiscale framework with extended kalman
filter for lithium-ion battery soc and capacity estimation. Appl Energy 92(4):694–
704
5. Kim IS (2010) A technique for estimating the state of health of lithium batteries
through a dual-sliding-mode observer. IEEE Trans Power Electron 25(4):1013–1022
6. Ko J, Fox D (2009) Gp-bayesfilters: Bayesian filtering using gaussian process pre-
diction and observation models. Autonomous Robots 27(1):75–90
7. Liu D, Pang J et al (2013) Prognostics for state of health estimation of lithium-ion
batteries based on combination gaussian process functional regression. Microelec-
tron Reliab 53(6):832–839
8. Nishi Y (2001) Lithium ion secondary batteries; past 10 years and the future. J
Power Sources 100(1–2):101–106
9. Piller S, Perrin M, Jossen A (2001) Methods for state-of-charge determination and
their applications. J Power Sources 96(1):113–120
10. Plett GL (2004) Extended kalman filtering for battery management systems of
lipb-based hev battery packs: Part 3. state and parameter estimation. J Power
Sources 134(2):277–292
11. Qian K, Zhou C et al (2010) Temperature effect on electric vehicle battery cycle
life in vehicle-to-grid applications. In: CICED 2010 Proceedings, vol 2010, pp 1–6
12. Saha B, Goebel K (2007) Battery data set. NASA AMES prognostics data repos-
itory
13. Vichare NM, Pecht MG (2006) Prognostics and health management of electronics.
IEEE Trans Components Packaging Technol 29(1):222–229
14. Wakihara M (2001) Recent developments in lithium ion batteries. Mater Sci Eng
R Reports 33(4):109–134
15. Williard N, He W, Pecht M (2012) Model based battery management system for
condition based maintenance. In: Proceedings of the MFPT, vol 2012
16. Xu L, Xu J (2014) Integrated system health management-based progressive diag-
nosis for space avionics. IEEE Trans Aerosp Electron Syst 50(2):1390–1402
17. Zhang J, Lee J (2011) A review on prognostics and health monitoring of li-ion
battery. J Power Sources 196(15):6007–6014
18. Zivkovic Z, Ferdinand VDH (2004) Recursive unsupervised learning of finite mix-
ture models. IEEE Trans Pattern Anal Mach Intell 26(5):651–656
Analysis of the Effect of Low-Carbon Traffic
Behavior on Public Bicycles
Jing Ma1 , Xin Liu2 , Zhineng Hu1(B) , and Yong Cheng2

1
Cengdu 610065, People’s Republic of China
huzn@scu.edu.cn
2
Abstract. With increasing concern about PM2.5 , global warming,

greenhouse gas emissions and the problem of the “last mile”, the public
bicycle system is gaining importance as viable mode in urban transporta-
tion worldwide. In order to more successfully promote the public bicycle
to the residents, it is obviously necessary to understand in-depth what
can affect the residents’ usage of public bicycle. The study elicit the influ-
ence of demographic, attitude towards the behavior, subjective norm and
perceived behavioral control on the usage of public bicycle. Based on the
theory of planned behavior (TPB) and the binary logistic model, the
paper conclude that the variable of gender, age, occupation and income
are closely contacted with the usage of public bicycle. And the education
qualification has a little effect on it. And there are significant relationship
between usage public bicycle and environmental friendly, flexible, find-
time, cleanness, the bicycle lanes, signals for cyclist, suitable weather.
Keywords: Public bicycle · Factor analysis · Low-carbon · Traffic

behavior
1 Introduction
Though there are more and more convenient bus routes today in daily life, some
back streets and new residential areas do not share public transport, and the
residents often take more than 10 min to reach the bus station. To solve the
public “last mile” problem, the community buses driving around the community
entrance came into being in Chengdu in 2012. But because of the limitation of
community buses, such as fixed lines and fewer bus stops, the residents were
not satisfied with community buses, so the traffic pattern could not solve the
problem perfectly. As the replacing for community buses, the public bicycles are
more flexible and more popular.
Generally, the problem of traffic congestion and heavy emissions of CO2
become more and more serious and people are increasingly worried about the
PM2.5 emissions and the smog problem in China. Therefore, low-carbon traffic
Analysis of the Effect of Low-Carbon Traffic Behavior on Public Bicycles 1319
has been becoming increasingly popular in many cities. The local authorities
have been paying more and more attention to public transport, and cycling in
daily mobility is considered according to efficient means to reduce air pollution,
traffic jams, and carbon emissions. Public bicycle systems (PBS) have turned out
to be efficacious in promoting riding in many cities, particularly when contacted
with public transportation.
Chengdu seems to be one of the most suitable cities for the development of
shared bicycle in China due to flat terrain, mild weather, little sunshine and less
wind as well as wide non-motor vehicle lanes. In addition, it is densely populated
and there are about 238834 persons per square kilometers. Therefore, the public
bicycle solve the Chengdu “last kilometers” of the problem effectively.
Despite public bicycle sharing system existing in many cities in China, cycling
still does not become the habit of most residents, and researches for the influ-
encing factors of PBS is not enough much and perfect. A study summed up the
impact of spatio-temporal interactions on bicycle sharing system demand [10].
And a find pointed out bicycle sharing stations and greenness were motivators
for bicycle commuting and public transport stations and elevation were deter-
rents for bicycle commuting by multinomial logistic regression models [6]. A
research about public bicycle system in Beijing found that the bike-share choice
is most sensitive comfort, user demographics, however, do not affect strongly on
the bike-share choice, indicating the mode will draw users from across the social
spectrum [4]. Firstly, this paper investigated the demographic factors related
the usage of public bicycles. Then based on the theory of planned behavior the
binary logistic model, this paper explored the factors that may affect the pub-
lic bicycle selection of residents from the perspective of attitude towards the
behavior, subjective norm and perceived behavioral control.
2 Literature Review
Attitudes and perception of an individual towards a particular mode would influ-
ence the behavior of that individual [8]. Commonly, the cycling requires a phys-
ical effort among the cyclist. The public bicycles are more and more popular
today. The norm is also important for the usage of public bicycles. So it is neces-
sary to understand which factors make a positive impact on the usage of public
bicycles.
As the history of public bicycle system is not long, the researches on theory
and practice of PBS are still immature. And most of the researches on public
bicycle selection of residents concentrated on characters of public bicycle travel,
the method or theory of travel choice and factors affecting public bicycle selec-
tion.
For the characters of public bicycle travel, a study found that the average
travel distance of bicycle users is about 3.7 km, and the male users were 3.8 km,
the female users are 3.6 Km through a survey towards the American travelers [1].
New findings unequivocally demonstrate that the mean journey length made by
private bicycle is 700–800 m (0.44–0.5 miles) greater than those made by public
1320 J. Ma et al.
bicycle [5]. And a survey found that the average distance traveled by bicycles for
fitness was 14 km [9]. What’s more, through the development of public bicycle
planning in Hangzhou, Shi [16] pointed out that the public bicycle system facil-
ities construction, the usage and operational mechanisms should be carried out
to research from the perspective of public bicycle facilities. In addition, a paper
that examines the tensions between the glamour of a global competitive city
explained what roles might public bicycle paly and how might a public bicycle
system business model operates if it were to serve the transport disadvantage
[13].
For the method and theory of travel choice, based on the information theory,
a selection procedure is proposed to select important variables in the logistic
regression model, and the strong consistency of these programs is proved [2].
And Yang [17] built a structural equation model of resident’s travel choice in
the process of urbanization based on out chain. Sakari [12] put forward that
public bicycle system should be viewed as part of public transport rather than
as a separate cycling scheme. Furthermore, a study developed a methodology for
categorizing bicycling environments defined by the bicyclist’s perceived level of
safety and comfort [14]. In addition, there is a research that develop an estimation
method using sinusoidal model to fit the typical pattern of seasonal bicycle
demand [11].
For the factors affecting public bicycle selection, it turned out that lower edu-
cation and absence of docking stations within walking distance were associated
with lower likelihood of awareness of the public bicycle system programs [3].
And a research explored the impact of specific weather conditions and calendar
events on the spatio-temporal dynamics of the case study public bicycle sharing
programs employing novel spatial analytical techniques [7]. Besides, there is the
study agreed that residents’ travel choices are mainly driven by environmental
conditions and personal travel habits [4]. A survey outcomes in India revealed
perceived benefits, physical barriers, safety hazards, social barriers, and road
condition may be the major factor classes influencing bicycle mode choice used
in exploratory factor analysis [15].
The current literature review reveals a dearth in studies that comprehensively
elicit the influence of demographic, attitude towards the behavior, subjective
norm and perceived behavioral control on the usage of public bicycle. Aptly, the
current paper aims to identify the effect in those aspects. Moreover, the study
also points out the determinants and correlates of public bicycle commuting.
3 Methodology
The factors were identified in the literature as influential in public bicycle system
are numerous and diverse [15]. The study focuses on commuting trips in order to
analysis the factors that influence people use the public bicycle based on theory
of planned behavior (TPB). In this paper, the factors have been divided accord-
ing to this rule that whether it is belong to “attitude towards the behavior”,
“subjective norm” and “perceived behavioral control”, they are the three main
variables that determine behavioral intentions. Where there is more positive atti-
tude, the more support of important others, and the stronger the control of per-
ceived behavior, there is more the behavior intentions about using public bicycle.
And vice versa. The factors of personal, social and cultural (such as personality,
intelligence, experience, age, gender, cultural background, etc.) indirectly affect
attitude towards the behavior, subjective norms and perceived behavior control
by influencing the behavioral beliefs, and ultimately affect behavioral intentions
and behavior. Attitude towards the behavior, subjective norms, and perceived
behavioral controls are conceptually distinct, but sometimes they may have a
common basis of beliefs, so they are both independent and related.
We first investigated the factors of demographic by a simple distribution
analysis specified for bicycle-shared and non-bicycle-shared commuters. Rele-
vance analysis (e.g. Person’s chi-square test) were applied to find the relation-
ships between the factors of demographic and commuting by public bicycles or
not.
After that, we investigated the determinants of public bicycle commuting
with the model of binary logistic regression from the perspectives of attitude
towards the behavior, subjective norms, and perceived behavioral controls.
4 Data Collection
4.1 Study Location

Chengdu is a large-size city of Sichuan province in the southwest of China with
the resident population of 14.428 million. And it is about 604.1 square kilometers.
Economic development is very rapid and it is one of the fastest growing cities
in China. What’s more, the environmental pollution is very serious, so it is
important to develop low-carbon traffic in the city. The terrain of Chengdu is
relatively flat so that the citizen develop riding habit is possible. And public
bicycles became more and more popular in the city since 2016.
4.2 Questionnaire Survey
The questionnaire includes four sections. The first section focuses on the demo-
graphic variables, and the respondents were asked to report their gender, age,
occupation, educational qualification and income. In the second, third and fourth
section, the questions about attitude towards the behavior, subjective norms,
and perceived behavioral controls are descripted in the questionnaire. Table 1
shows the characteristics about the questions.
The data for the study is collected usage face-to-face questionnaire survey
by a sample collected using a random sampling method. And it was at subway
station, university, attractions and business district we did this survey.
1322 J. Ma et al.
Table 1. Summary of questions
Attitude towards the Subjective norm Perceived behavioral

behavior controls
(1) I think riding public (1) I do not use the public (1) I would not use the
bicycle is a cool choose to bicycle because I do not public bicycle because it is
commute how it works laborious
(2) I want to use the (2) I would not use the (2) I would not use the
public bicycle because it is public bicycle because the bicycle because there is
so flexible registration fee is too not bicycle lane
high(higher 300 yuan)
public bicycle because it is public bicycle because the bicycle because of heavy
time-saving unit price is too traffic
high(higher 2 yuan per
hour)
public bicycle because it is public bicycle if the bicycle because there is no
environmentally friendly destination is not near to signal for cyclist
my house(without 10 km)
public bicycle because it is public bicycle if it takes bicycle if it is night
entertaining too much time to find the
bike.(beyond 5 min)
public bicycle because it is public bicycle because the bicycle if weather is not
good for health bicycle is not clean pleasant
(7) I would not use the (7) I would not use the
public bicycle because it is public bicycle because the
not consistent with bicycle is not beautiful
identity
(8) I would not use the
public bicycle because I
have private bicycle and
e-bicycle
4.3 Sample Characteristics
A total of 633 people are in the survey and 79% respondents have been used
the public bicycle. Most of the respondents from various income classes use the
public bicycle as transportation to school and to work. Before using the public
bicycles, 73% of respondents used subway or bus as the main mode, 60% od
them walked, 26% of them take a bike, 34% of them use car or taxi. In addition,
it is at school and community sa well as subway and bus station where most
respondents used the public bicycle.
5 Statistical Analysis
All data analysis were performed using SPSS. For the demographic data, we
test whether it is independent between usage public bicycle and demographic
character by Person’s chi-square test (The significance level α = 0.05).
Attitude towards the behavior that may affect propensity of using public
bicycle were evaluated by using a binary logistic regression model (95% con-
fidence). Participant characteristic were dichotomized before testing as poten-
tial confounders, including whether participants think it is cool, flexible, time-
saving, environmentally friendly, entertaining, good for health, not consistent
with identity. In addition, the potential confounders for subjective norm include
the understanding about PBS, the registration fee, the unit price, the distance,
the find-time, outward appearance, cleanliness, and whether the participants
have private bicycle and e-bicycle. And the potential confounders for perceived
behavioral controls include whether it is laborious, whether there is bicycle lanes,
signals for cyclist, heavy traffic, visibility and weather. Final model fitness was
checked with a (post-estimation) generalized Hosmer–Lemeshow goodness-of-fit
test.
6 Results and Discussion

6.1 The Relationship Between Public Bicycle Selection and
Demographic Factors by Chi-Square Test
When the demographic data was analyzed, Table 2 shows the sample distribution
according to demographic factors. It reveals that the public bicycles are very
popular among people. There are 503 respondents usage the public bicycle. The
majority of cyclists who use public bicycle are women (65%). And for those
who do not use public bicycle, it is equivalent between men and women. The
Asymp.sig. (2-sided) is equal to 0.002, so it explain that the use of public bicycles
is related to gender at the 99.5% confidence level.
For groups of aged 10–48, there is no significantly difference in column pro-
portions between use and non-use. However, the proportion of the use aged 49–65
age (4%) is significantly lower than the non-use. What’s more, among the use,
the groups of aged 10–35 are significantly higher than the other age groups. The
group mainly comprises students and the young worker who are more likely to
take bikes to school or to work. In addition, the Asymp.sig. (2-sided) is equal to
0.006, so it explain that the use of public bicycles is related to age at the 99.5%
confidence level.
For the group of students, it is obviously that the proportion of the use
(78%) is higher than the non-use (49%). After all, the students are easier to
accept new things, like public bicycle. And it is usually not too far from their
home (dormitory) to school (classroom), so the bike is the best transportation
for them. Besides, among the use, the groups of retirees (1%) are significantly
lower than the other groups. The group mainly comprises elder who may feel
more difficult to ride a bike. And the Asymp.sig. (2-sided) is equal to 0.003, so
1324 J. Ma et al.
it explain that the use of public bicycles is related to occupation at the 99.5%
confidence level.
For those whose income is less 3000, the proportion of the use is slightly
higher than the non-use. But for those whose income is more than 3000, the
proportion of the use is higher than the non-use. And for those who use the
public bicycle, the groups whose income is less 1500 and between 1500 and 3000
are significantly higher than the other income groups. The group mainly prefer
riding than driving in order to reduce the cost. What’s more, the Asymp.sig.
(2-sided) is equal to 0.005, so it explain that the use of public bicycles is related
to income at the 99.5% confidence level.
For those whose educational qualification is at the same level, the proportion
of the use and the non-use is equal on the whole. And the Asymp.sig. (2-sided)
is equal to 0.345, so there is slight connection between educational qualification
and the use of public bicycles at the 99.5% confidence level.
Table 2. Distribution of the sample according to demographic factors.
Variables Total Use Non-use Asymp.sig. (2-sided)

Gender 633(100%) 503(100%) 130(100%) 0.002
Male 241(61.9%) 176(35%) 65(50%)
Female 392(38.1%) 327(65%) 65(50%)
Age 633 (100%) 0.006
10–22 270 (42.7%) 227(45%) 43(33%)
23–35 257(40.6%) 213(42%) 44 (34%)
36–48 67(10.6%) 49(9%) 16 (5%)
49–65 39(6.1%) 14(4%) 25(18%)
Occupation 633(100%) 0.003
Student 444(70.1%) 391(78%) 63(49%)
Office worker 138(21.8%) 93(18%) 45(36%)
Self ?C employed 24(3.8%) 13 (3%) 11(8%)
Retirees 13(2.1%) 4(1%) 9(7%)
Income 633(100%) 0.005
≤1500 256(40.5%) 201(40%) 54(41%)
1501–3000 184(29.1%) 155(31%) 29(22%)
3001–6000 155(18.2%) 91(18%) 24(19%)
≥6001 78(12.2%) 56(11%) 22(18%)
Educational qualification 633(100%) 0.345
High school student 8(40.5%) 5(1%) 3(2%)
College students 30(29.1%) 22(4%) 8(6%)
Undergraduate 411(18.2%) 324(51%) 87(67%)
Graduate or doctor 184(12.2%) 152(30%) 32(25%)
Note: the Chi-square statistic is significant at 0.05 level
6.2 The Analysis of the Influencing Factors of Public Bicycle

Selection by Binary Logistic Regression
The development of binary logistic model:
exp(β0 + β1 x1 + · · · + βk xk ) 1
p= = .
1 + exp(β0 + β1 x1 + · · · + βk xk ) 1 + exp(−β0 − β1 x1 − · · · − βk xk )
In the model, p is the probability of selecting public bicycles, x1 , x2 , · · · , xk

are explanatory variables, β0 , β1 , · · · , βk are the coefficients of the corresponding
explanatory variables.
For the factors of attitude towards the behavior, a regression analysis were
gone on according to whether the use of public bicycle as the dependent vari-
able, whether participants think it is cool, flexible, time-saving, environmental
friendly, entertaining, good for health, not consistent with identity as the inde-
pendent variable. For the factors of subjective norms, according to whether the
use of public bicycle as the dependent variable, the knowledge of PB, the regis-
tration fee, the unit price, distance, find-time, cleanness, appearance, and private
bicycle and e-bicycle as the independent variable, a regression model get. For
the factors of perceived behavioral controls, regressed according to whether the
use of public bicycle as the dependent variable, laborious, the bicycle lane, the
heavy traffic, signal for cyclist, visibility and weather as the independent vari-
able. Then we develop binary selection model and select independent variable
by “input” method to research what factors affect the choice of public bicycle.
Table 3 shows that the final chi-square statistics is 10.483 after iteration from
Hosmer and Lemeshow Check and the critical value: CHINV(0.05, 8) = 15.507,
so the chi-square statistics is lower than the critical value. From the perspective
of significance, 0.233 is higher than 0.05. It indicates that the model can be a
good fit and there is no significant difference.
Table 3. Hosmer and Lemeshow check (attitude towards the behavior)
Step Chi-square df Sig

1 10.483 8 0.233
Table 4 elicits the factors of significant contribution for chose of public bicy-
cles about attitude towards the behavior. Among the attitudinal characteristics,
the characteristics of “environmental friendly” and “flexible” turned out to be
the prominent variables with the Exp (B) value of 2.626 and 2.190 when com-
pared with the others. From the perspective of significance, the significance of
“environmental friendly” and “flexible” are lower than 0.05 and the significance
of constant is 0.008 lower 0.005, so it explains that there are significant relation-
ship between using public bicycle and environmental friendly as well as flexible.
Therefore, a binary selection model can be shown below as following.
1326 J. Ma et al.
Table 4. Binary selection model regression results summary (attitude towards the
behavior)
Independent variable B S.E Wals df Sig Exp (B)

Environmental friendly 0.816 0.237 11.881 1 0.001 2.262
Good for health 0.123 0.231 0.283 1 0.595 1.130
Entertaining 0.229 0.212 1.163 1 0.281 1.257
Cool 0.238 0.240 0.985 1 0.321 1.269
Flexible 0.737 0.215 11.733 1 0.001 2.090
Time-saving 0.002 0.217 0.000 1 0.993 1.002
Not matching identity 37.232 68459278.282 0.000 1 1.000 1.478E16
Constant 0.619 0.235 6.930 1 0.008 1.858
1
p= .
1 + exp(−0.619 − 0.816x1 − 0.737x2 )
In the model, p is the probability of selecting public bicycles, x1 is 1 when

the participants use the public bicycle because it is environmentally friendly and
is 0 when they do not. And x2 is 1 when the participants use the public bicycle
because it is flexible and is 0 when they do not.
so the chi-square statistics is lower than the critical value. From the significant,
0.064 is higher than 0.05. It indicate that the model can be a good fit and there
is no significant difference.
Table 5. Hosmer and Lemeshow Check (subjective norms)

1 19.108 8 0.064
Table 6 shows the binary selection model regression results of subjective

norms factors which can contribute to an increased public bicycle used. Obvi-
ously, the characteristics of “find-time” and “cleanness” are more important
than other variables because the Exp (B) value (2.262 and 1.775) of them is
higher than others. clearly, the significance of “find-time” and “cleanness” are
lower than 0.05 and the significance of constant is 0.000 lower 0.005, so it explains
that there are significant relationship between using public bicycle and find-time
as well as cleanness. Therefore, a binary selection model can be shown below as
following.
Table 6. Binary selection model regression results summary (subjective norms)

The knowledge of PB −1.146 0.218 27.554 1 0.172 0.318
The registration fee 0.252 0.230 1.203 1 0.273 1.287
The unit price 0.139 0.273 0.257 1 0.612 1.149
Distance −0.27 0.213 0.016 1 0.899 0.973
Find-time 0.816 0.237 11.881 1 0.001 2.262
Cleanness 0.574 0.212 7.347 1 0.007 1.775
Appearance 0.307 0.246 1.564 1 0.211 1.359
private bicycle and e-bicycle 0.416 0.333 1.564 1 0.211 1.516
Constant 1.019 0.200 26.044 1 0.000 2.770
1
p= .
1 + exp(−1.019 − 0.816x1 − 0.574x2 )
In the model, p is the probability of selecting public bicycles, x1 is 0 when the

participants do not use the public bicycle if he (she) can’t find public bicycles in
5 min and is 1 others. And x2 is 0 when the participants do not use the public
bicycle because it is not enough clean and is 1 when others.
so the chi-square statistics is lower than the critical value. From the perspective
of significance, 0.195 is higher than 0.05. It indicates that the model can be a
good fit and there is no significant difference.
Table 7. Hosmer and Lemeshow check (subjective norms)

1 11.125 8 0.195
In summary, Table 8 highlights the determinant variable of perceived behav-

ioral control that is closed linked with the use of public bicycles. As shown,
whether there are the bicycle lanes, signals for cyclist or suitable weather con-
tribute much for usage of the public bicycles according to the value (2.088, 1.861,
1.760) of Exp (B). And the significance of them (0.001, 0.006, 0.007) and constant
(0.001) is lower than 0.005, so a binary selection model can get as following.
1
p= .
1 + exp(−0.609 − 0.736x1 − 0.621x2 − 0.566x3 )
1328 J. Ma et al.
Table 8. Binary selection model regression results summary (subjective norms)

The bicycle lane −0.736 0.225 10.700 1 0.001 2.088
The heavy traffic −0.179 0.299 0.359 1 0.549 0.836
Laborious 0.533 0.336 2.522 1 0.112 1.704
Signal for cyclist 0.621 0.228 7.449 1 0.006 1.861
Weather 0.556 0.210 7.241 1 0.007 1.760
Visibility −0.090 0.209 0.185 1 0.667 0.914
Constant 0.609 0.187 10.635 1 0.001 1.838
In the model, p is the probability of selecting public bicycles, x1 is 0 when

the participants use the public bicycle if there are not bicycle lanes and is 1
when others. x2 is 0 when the participants use the public bicycle if there are not
signals for cyclists and is 1 when others. x3 is 0 when the participants use the
public bicycle if the weather is not suitable to ride and is 1 when others.
7 Conclusion
This study investigated the demographic factors related the usage of public bicy-
cles, and also explored the factors that affect the choice of using public bicycles
from the perspective of attitude towards the behavior, subjective norm and per-
ceived behavioral control according to TPB. The following are the conclusion
obtained from the study.
For the demographic factors, the variable of gender, age, occupation and
income are closely contacted with the usage of public bicycle. And the education
qualification has a little effect to it.
As for attitude towards the behavior, people are more concerned about the
convenience and low-carbon. Therefore, there are significant relationship between
usage public bicycle and environmental friendly or flexible. From the perspective
of subjective norm, people maybe care about the time-saving and comfort more
than appearance of the public bicycles so that they are more particular about
find-time or cleanness. What’s more, for perceived behavioral control, people
are more concerned about safety, and it is important for people to be competent
to cycle in some particular environment, so people would have a more positive
perception about cycling if there are the bicycle lanes, signals for cyclist or
suitable weather.
Acknowledgements. This research was supported by the projects XQ14B04 and

SKZX2015-SB87.
References
1. Aultman-Hall L, Hall F, Baetz BB (1997) Analysis of bicycle commuter routes
using geographic information systems: Implications for bicycle planning. Trans-
portation Res Record J Transportation Res Board 1578(1578):102–110
2. Bai Z, Krishnaiah P (1991) Variable seleetion in logistic regression. J China Univ
Sci Technol 21(2):23–31
3. Bernatchez AC, Gauvin L et al. (2015) Knowing about a public bicycle share
program in montreal, Canada: are diffusion of innovation and proximity enough
for equitable awareness? J Transp Health 2(3):360–368
4. Campbell AA, Cherry CR et al (2016) Factors influencing the choice of shared
bicycles and shared electric bikes in beijing. Transp Res Part C Emerg Technol
67:399–414
5. Castillo-Manzano JI, López-Valpuesta L, Sánchez-Braza A (2016) Going a long
way? On your bike! Comparing the distances for which public bicycle sharing
system and private bicycles are used. Appl Geogr 71:95–105
6. Cole-Hunter T, Donaire-Gonzalez D et al (2015) Objective correlates and determi-
nants of bicycle commuting propensity in an urban environment. Transp Res Part
D Transp Environ 40:132–143
7. Corcoran J, Li T et al (2014) Spatio-temporal patterns of a public bicycle sharing
program: the effect of weather and calendar events. J Transp Geogr 41:292–305
8. Davies D, Gray S et al. (2001) A quantitative study of the attitudes of individuals
to cycling. Trl Report
9. Dill JJ, Gliebe (2008) Understanding and measuring bicycling behavior: a focus
on travel time and route choice. Bicycle Lanes 29
10. Faghih-Imani A, Eluru N (2016) Determining the role of bicycle sharing system
infrastructure installation decision on usage: case study of montreal bixi system.
Transp Res Part A Policy Practice 94:685–698
11. Fournier N, Christofa E Jr, MAK (2017) A sinusoidal model for seasonal bicycle
demand estimation. Transp Res Part D Transp Environ 50:154–169
12. Jäppinen S, Toivonen T, Salonen M (2013) Modelling the potential effect of shared
bicycles on public transport travel times in greater helsinki: an open data approach.
Appl Geogr 43(43):13–24
13. Jennings G (2015) Finding our balance: considering the opportunities for public
bicycle systems in cape town, South Africa. Res Transp Bus Manag 15:6–14
14. Joo S, Oh C et al (2015) Categorizing bicycling environments using gps-based
public bicycle speed data. Transp Res Part C Emerg Technol 56:239–250
15. Majumdar BB, Mitra S (2015) Identification of factors influencing bicycling in
small sized cities: a case study of Kharagpur, India. Case Stud Transp Policy
3(3):331–346
16. Shi X, Cui D, Wei W (2011) The investigation of the relationship between using
and planning of public bicycle in Hangzhou. Urban Studies
17. Yang LY, Zhu XN (2012) Travel mode choice in the process of rapid urbanization.
China Soft Science
An Empirical Study on the Impact of Niche
Overlap of Tourism Enterprise on Tourist
Satisfaction
Shaojiang Lin1 , Jiaying Chen2(B) , and Jing Tang2

1
College of Tourism, Sichuan Agricultural University,
2
Business School, Sichuan Agricultural University,
cjysc@qq.com
Abstract. Vicious competition of tourism enterprises results from prod-

uct homogeneity deeply manifested as homogeneity of tourism enterprise
resource utilization, thus becoming the key impeding current tourism
development. This paper uses niche theory as reference, illuminates the
conception and connotation of niche overlap of tourism enterprise, and
proposes its three-dimensional measurement model by interviews and
questionnaire surveys of tourism enterprises, and this thesis also tests
the data using methods such as EFA analysis and CFA analysis. On this
basis, the influence mechanism of niche overlap of tourism enterprise on
tourist satisfaction degree is analyzed by using structural equation. Ver-
ification result shows overlapping degree of space-time resource niche of
tourism enterprises, niche overlap of market resource, niche overlap of
ecology resource are all significantly negatively related to tourism satis-
faction negatively.
Keywords: Tourism enterprises · Niche · Niche overlap · Measurement

model
1 Introduction
Tourism industry is still in the initial phase of development at present, and the
overall level of tourism enterprises’ utilization efficiency of resources is relatively
low, which leads to serious phenomena of product homogeneity and vicious com-
petition affecting tourist experience. For tourism industry is a complex business
ecology system which integrates many industries with blurred boundaries, prob-
lem of tourism enterprise competition cannot be explained if the competition
is confined to traditional competition theory within the industry, thus many
scholars have begun to use related theory of natural ecology system to research
tourism phenomena, such as niche theory [8].
Niche theory, as basic theory of ecology, was widely applied in the research of
related fields such as diversity of species, interspecies relation, etc. Niche refers
An Empirical Study on the Impact of Niche Overlap of Tourism Enterprise 1331
to a set of resources used in a community by a population or an individual, and

is a “way of living” of itself [9]. Since the conception of niche is blurred, the
theoretical cycle offers its measurement standard, namely niche measurements.
And niche overlap is one of the most important indexes. Niche degree is used
to measure the niche degree similarity of two (or more than two) populations,
representing common use degree of them on a kind of resource to evaluate their
level of competition [4].
In the application of the niche theory on tourism, foreign scholars prefer to
use niche theory to study the tourism market segmentation. Lew and Duval [10]
explained the Long Tail concept and proposed that the concept of the post-
tourist, for example, is a Long Tail phenomenon. Pitts [12]’s study was to deter-
mine if there is a lesbian and gay sports tourism industry. Vanessa [15] used the
Rocinha case, a favela in Rio de Janeiro, as a model for other communities to
show the favela as a new niche market for tourism activity in Brazil.
Different from the foreign scholars, domestic scholars mainly discuss the def-
inition and connotation of tourism niche, the construction of niche evaluation
system and put forward niche optimization strategy from the perspective of
tourism industry, tourism destination, tourism city and so on. Wang [18] firstly
summarized the correlated theory research on the niche, and proposed that the
industry cluster niche is its occupied resource space in the regional ecosystem
through its own conditions, and further more brings forward the concept and
features of the TIC niche based on its characteristics. And thus he held that the
TIC niche consists of tourism resources, the cluster management and the cluster
environmental niche. Wang et al. [16], by introducing the theory of ecological
niche and based on the analysis of application, constructed the index system
of the evaluation of theme parks in Pearl River Delta. Xiang [19] used Delphi
method to choose and establish the tourism niche evaluation index, then evaluate
the tourism niche of Zhangjiajie and Tianmenshan national forest park by means
of on-the-spot investigation and questionnaire. Wang and Yu [17] proposed that
the measurement system of the niche of regional tourism cities consists of 4
dimensions of tourism resources, market, socioeconomy and environment as well
as 32 variables. Chen [3] explored how Taiwan could build a niche in Asia’s cruise
tourism industry.
The above researches mainly concentrated in the tourism industry, tourism
market, tourism and regional tourism by using niche theory or thought. How-
ever, at present, the academic circle has little research on tourism enterprises,
the core population in tourism business ecosystem. In fact, in the circumstance
of explosive growth of tourism, vicious competition of tourism enterprises and
product homogeneity have become the key impeding tourism’s development. It is
urgently needed to research the resource competition situation of tourism enter-
prises and resource utilization effect from the perspective of ecological system so
as to optimize competition relationship.
1332 S. Lin et al.
The innovation of this paper is to applicate the niche theory to tourism enter-
prises, to construct a measuring index system of niche overlap degree of tourism
enterprises, and to analyze the influence of niche overlap degree on tourist
satisfaction by using SEM. This article mainly divides into the following sev-
eral parts: the first part mainly introduces the research background and research
progress, the second part mainly puts forward the theoretical model and research
hypothesis, the third part constructs tourism enterprises measuring index system
of niche overlap degree and tourist satisfaction, the fourth part is data collection
and analysis which is based on the questionnaires and interview survey, the last
part is the research conclusions and prospects.
2 Theory Model and Research Hypothesis
Scholars have identified niche factors from different aspects. People such as Han-
nan [6], Baum [1] elaborated the identification of niche from the aspects of enter-
prises’ market resource demand and ability of providing products. They think
enterprise niche, as the intersection of enterprise resource demand and enterprise
producing ability, relies on the position where the enterprise is located and what
it does. Yan A and Qing-Li DA proposed that enterprises’ niche is formed by
the interaction between other enterprises and themselves actively in special cir-
cumstances [21]. Shan M indicated that enterprise niche can be divided in three
dimensions: human resources, tangible resources and intangible resources [13].
The above scholars’ identification of niche can be summarized in two dimen-
sions: market resource and human resource. However, tourism enterprises are
service organizations satisfying tourists’ tourism experience and demand by pro-
viding products in certain time and space, and the elements of tourism time and
space of tourism enterprises are virtually tourism enterprises’ resource. On this
basis, this thesis proposes the research hypothesis as follows:
Hypothesis 1 : Niche overlap of tourism enterprises can consist of three dimen-
sions: space-time resource, market resource and productive resource.
Furthermore, in business ecosystem, competition within populations can not
only affect their internal development, but also impact other populations’ health.
Tourists, as one of core populations in scenic business ecosystem, their health
conditions are also certainly affected by the relationship of tourism enterprises.
Presentation of tourist population health condition in the whole scenic business
ecosystem is actually tourists’ experience condition, which can be measured with
the index of tourist satisfaction degree by us. Accordingly, this thesis proposes
research hypothesis as follows:
Hypothesis 2 : There is significantly negative influence of tourism enterprise
niche degree on tourist satisfaction degree.
According to the above analysis, this thesis proposes theory study model and
study hypothesis, as shown in Fig. 1.
Fig. 1. Theory study model of “enterprise niche overlap-tourist satisfaction degree”
3 Variation Measurement
3.1 Measurement of Enterprise Niche Overlap
(1) Space-time Resource Overlapping Degree
Baum argued that enterprise niche, as the intersection of enterprise resource
demand and enterprise producing ability, relies on the position where the enter-
prise is located and what it does [1]. Yan and Da [21] pointed out development
opportunity and ecology strategy should be sought in time factor and social
factor. Thus this thesis designs related measurement items to niche overlap, as
shown in Table 1.
Table 1. Measurement items of space-time resource overlapping degree
Number Measurement items Sources

NO11 There are many enterprises of similar operation Yan and Da [21];
time-limit to your enterprise Baum and Mezias [1];
NO12 There are many enterprises of similar life cycle
phase to your enterprise
NO13 There are many enterprises of same everyday
business hours as your enterprise
NO14 The enterprise is located in non-gold geographic
area in the scenic spot
NO15 There are many enterprises around your enterprise
Data source: Arranged according to this research.
(2) Market resource overlapping degree

From the perspective of niche theory development, the influence of market
resource dimension on enterprise niche has been emphasized more and more by
scholars, and plenty of related research results have occurred in recent years.
For example, scholars such as Dobrev [5], examined the effects of crowding in a
market center on rates of change in organizational niche width and on organiza-
tional mortality. Hou et al. [7] considered that exorbitant overlapping between
1334 S. Lin et al.
resources, abilities, markets, which populations need to exist and existing enter-
prises should be avoided to the greatest extent, for the niche overlapping will
lead to competition of populations. Therefore, this thesis also adopts market
resource dimension to measure the enterprise niche overlap (Table 2).
Table 2. Measurement items of market resource overlapping degree
Number Measurement items Resources

NO21 The enterprise is in bad situation in market Hou et al. [7];
competition Dobrev and Hannan [5];
NO22 There are many enterprises of similar target
customer groups to your enterprise
NO23 Visitors think the product price of your
enterprise is higher on equal basis of equality
NO24 The sales method of the enterprise is similar to
most enterprises
NO25 The marketing channel of the enterprise is
similar to most enterprises
Data source: Prepared arranged according to this research.
(3) Overlapping degree of productive resources

Baum [1] considered that niche of enterprise relies on its location and func-
tions, which is the intersection of enterprise resources and its productive capacity.
Stuart and Podolny [14] proposed that niche of enterprises shall be identified by
the number that the enterprise obtains in equivalent technical patents, while
niche overlap in the enterprise shall be identified by the number of patents’ rel-
ative references. According to the above analysis, the Thesis has extracted the
dimensionality of productive resources to measure niche overlap in the enterprise
Table 3. Measuring items of productive resources overlapping degree
No Measuring items Source

NO31 Staff’s professional skills are better than other Stuart and Podolny [14];
enterprises Hannan et al. [6];
NO32 Possessing similar techniques to most enterprises Baum and Podolny [1]
NO33 Develop and launch new products and new
services more often than other enterprises
NO34 New products or services are quicker to be
purchased by customers and seize the market
NO35 The scale of manufactures is larger
Data source: Sorted according to the research.
3.2 Tourist Satisfaction Degree
The research adopts a more mature SERVQUAL service quality scale to replace
a simple tourist satisfaction evaluation. Proposed in 1990, SERVQUAL model
included five dimensions, tangibles, reliability, responsiveness, assurance and
empathy, and measured the distance of service quality by a questionnaire form
[2]. At present, the model is considered as a most classic method to measure
all kinds of service qualities appropriately and used in the relevant studies of
all service industries [20]. In view of the particularity of different industries,
the dimensionality and index of SERVQUAL model must be amended before
it is applied to tourism industry [11]. Based on the previous studies, the paper
has fine tuned the model and formed a tourist satisfaction evaluation system as
shown in Table 4.
Table 4. Tourist satisfaction degree evaluation index system
Elements Component items

Tangibles 1. Complete servicing facilities
2. Modernized servicing facilities
3. Staff in tidy clothes and appearances
4. Comfortable and elegant decoration and environment
Reliability 5. Complete the things promised to the tourists without delay
6. Show concerns and provide helps when the tourists encounter
difficulties
7. Provide products with reliable qualities
8. Products safety can be trusted.
Responsiveness 9. Provide initiative and efficient products or service
10. Provide prompt service
11. Be always willing to help the tourists solve problems
12. Provide service promptly and meet tourists’ demands.
Assurance 13. Each staff is reliable
14. Tourists will be at ease in service process
15. The service attitude is passionate and polite
16. Staff’s communication skills
Empathy 17. Provide individual services for different customers
18. Give an individual care for customers
19. Pay attention to customer’s demands
20. Give priority to customer interests
Data source: Sorted according to the research.
1336 S. Lin et al.
4 Data Collection and Analysis

4.1 Data Collection
The research based on theoretical model and research hypothesis adopts one-
to-one matching questionnaire method, chooses and designs each variable mea-
surement scale and prepares and issues preliminary survey questionnaire. Due to
indistinct tourism boundaries, it is difficult to define tourism enterprises. Thus,
the selection of the research objects mainly focuses on the enterprises around
scenic spots regarding tourism service as a main purpose. The research issues 450
questionnaires. Removing 38 questionnaires with incomplete data and obvious
erring data, the research has recycled 402 valid questionnaires which indicate
the recovery rate is 89.33%.
4.2 Testing of Validity and Reliability About Niche Overlap in the

Enterprise
(1) Reliability test

The research adopts SPSS19.0 and AMOS17.0 software to test reliability and
validity of scale. According to KMO value and Bartlett spherical test, the results
show that KMO value is 0.917 and the significance probability of Bartlett spher-
ical test’s statistic is 0.000. The research adopts principal component analysis
to leach factors, then extracts 3 factors in total with approximate 70.577% of
gross number of variations according to the principle that characteristic root is
greater than 1. The characteristic value of fourth factor is obviously less than
1, which indicates that the enterprise’s niche overlap exists only 3 dimension-
alities. Orthogonal rotation shall be used on factor reference axis and Varimax
shall be used on the rotation. Loading coefficients of measuring items involved
in all factors are basically greater than 0.7 after the rotation, which indicates
that the scale has preferable discrimination validity and all measuring items are
in line with the theorized expectations of factor structure. The above is shown
in Table 5.
The Thesis adopts Cronbach’s α coefficient to conduct a reliability test after
the completion of factor analysis. Detailed analysis result is shown in Table 6.
It can be seen that internal consistency reliability of overall scale in official
scale of the research is 0.917 and internal consistency reliabilities of all potential
variable scales are all greater than 0.8 from reliability analysis. It indicates that
consistency among all items of scales is comparatively high and official scale has
a better internal consistency.
(2) Validity test
The research adopts SEM method to conduct construct validity test of all
scales and the analytical model is shown in Fig. 2.
In 3 factors confirmatory factor analysis models of overlapping degree scale of
niche, CMIN/DF value is less than 3, NFI, IFI and CFI values are all greater than
0.9, RMSEA value is less than 0.1 and fitting degree of all indexes is in better
adaptation. The results have verified that niche overlap is formed by overlapping
Table 5. Rotating component matrix of niche overlap of tourism enterprise
Dimensionality Measuring items Factor loading

1 2 3
Space-time resources overlapping degree NO11 0.198 0.186 0.798
NO12 0.286 0.272 0.758
NO13 0.284 0.217 0.795
NO14 0.258 0.254 0.76
Productive resources overlapping degree NO21 0.091 0.735 0.14
NO22 0.132 0.765 0.103
NO23 0.056 0.725 0.184
NO24 0.165 0.729 0.27
NO25 0.187 0.79 0.222
Market resources overlapping degree NO31 0.84 0.133 0.23
NO32 0.817 0.166 0.209
NO33 0.851 0.151 0.2
NO34 0.832 0.096 0.271
NO35 0.867 0.141 0.194
Extraction method: Principal Component Analysis
Rotation method: Orthogonal rotation method with Kaiser standardization
The rotation is converged after 5 iterations.
Table 6. Official scale reliability
Scale (number of items) Mean value Variance of items Internal consistency

reliability
NO1 (4) 2.763 0.003 0.874
NO2 (5) 2.746 0.004 0.841
NO3 (5) 2.603 0.001 0.928
ST (4) 3.267 0.002 0.898
SRT (4) 3.18 0.002 0.867
SRS (4) 3.069 0.002 0.841
SA (4) 3.176 0 0.871
SE (4) 3.133 0.003 0.879
Total (34) 3.144 0.095 0.917
degree of space-time resource, overlapping degree of productive resource and

overlapping degree of market resource.
Pearson’s correlation coefficient is used in this study to calculate the relation-
ship between niche overlap and tourist satisfaction. The empirical research indi-
cates, there is no multi-collinearity between the variables, as shown in Table 7.
1338 S. Lin et al.
.76
NO35 e1
.87
.72
.85 NO34 e2
.86 .73
productive overlap NO33 e3
.82
.67
.85 NO32 e4
.73
NO31 e5
.42 .74
.86
NO25 e10
.65
.81 NO24 e11
.62 .39
.62 Market overlap NO23 e12
.65
.42
.61 NO22 e13
.38
NO21 e14
.61
.63
.79 NO14 e16
.83 .69
Space-time overlap NO13 e17
.81
.66
.75 NO12 e18
.56
NO11 e19
Fig. 2. Confirmatory factor analysis of niche overlap of tourism enterprise
Table 7. Pearson’s correlation coefficient between niche overlap of tourism enterprise

and tourist satisfaction degree
Mean value Standard deviation NO1 NO2 NO3 TS

NO1 2.763 0.85794 1
NO2 2.7462 0.70047 .523∗∗ 1
NO3 2.603 0.9468 .558∗∗ .363∗∗ 1
∗∗ ∗∗
TS 3.1649 0.61748 −.432 −.314 −.577∗∗ 1
Note: ** p < 0:01, NO1 - space-time resource overlapping degree, NO2 - pro-
ductive resource overlapping degree, NO3 - market resource overlapping degree,
TS - tourist satisfaction degree.
Based on results above, structural equation model is presented in this thesis

to analyze the influence of overlapping degree of corporate niche on tourist sat-
isfaction degree, and as overlapping degree of space-time resource, overlapping
degree of market resource, overlapping degree of productive resource and tourist
satisfaction degree are incorporated into the structural equation model, model
fitting is well completed, as shown in Fig. 3.
As shown in analysis results above, overlapping degree of space-time resource,
overlapping degree of market resource, overlapping degree of productive resource
all have obvious negative influence on tourist satisfaction degree, with their stan-
dardized regression coefficient being β = −0.11∗∗∗ , β = −0.07∗∗∗ , β = −0.3753∗∗
(p < 0.001), respectively. It is thus clear that overlapping degree of corporate
niche has negative influence on tourist satisfaction degree, among which the
.31
e5 NO31 e12
.39 .95
e4 NO32 .95
.30 .95
e3 NO33 Productive overlap
.94 e11
.32 .29
e2 NO34 ST e6
1.00
.28 1.00 -.37 .24
e1 .85 SRT e7
NO35
.80 .22
.30 Tourist satisfaction SRS e8
.82
.49 e13 .23
e15 NO21 .86 SA e9
.46 .72 .26
-.07 SE e10
e16 NO22 .77
.52 .75
e17 NO23 Market overlap
.95
.28
e18 NO24 1.00
.20 -.11
e19 NO25
.45
.44
.35 e14
e20 NO11
.35 1.00
e21 NO12
.30 1.10
e22 NO13 1.07
Space-time overlap
.38 1.07
e23 NO14
(CMIN=220.912; DF=146; CMIN/DF=1.513;

NFI=.953; IFI=.983; CFI=.983; RMSEA=.036)
Fig. 3. The influence of niche overlap of tourism enterprise on tourist satisfaction

degree
influence of overlapping degree of market resource is greater. Therefore, research

Hypothesis H2, H21, H22 and H23 are all verified.
5 Research Conclusion and Prospect
The empirical research results indicate that overlapping degrees of corporate

niche include overlapping degree of space-time resource, overlapping degree of
productive resource and overlapping degree of market resource, and these three
dimensionalities all have significant negative influence on tourist satisfaction
degree. This is because: firstly, when business hours and location of corporations
is the same as that of majority of corporations in business ecosystem of scenic
spot, space-time resource overlap, as a result, competition in space-time resource
of scenic spot between corporations is more intense, and in order to seek for more
suitable time and space resource, the corporations will gradually consume their
own resources in other aspects to survive in the competition.
Secondly, there are tourism enterprises of similar market niche inevitably in
scenic business ecosystem, with different types of behavior occurring to fight
for the same target customer group, such as price war, soliciting, etc., which
will lead to decline of tourism satisfaction degree. Finally, tourists’ experience
of all products in the scenic area is actually experience of enterprise supply
chains, so enterprises owning non-homogeneity technology are more likely to
ally and integrate resources; and enterprises with homogenous technology tend
1340 S. Lin et al.
to be subject to mutual repulsion, failing to provide better tourism experience

for tourists.
In the future, research should be conducted further in combination with niche
width of tourism enterprises, for niche overlap reflects the competition condition
of resources used by tourism enterprises, while ecology width reflects the diversity
degree of tourism enterprise resource utilization, and both of them interact with
each other to influence tourists’ tourism experience.
Acknowledgements. This work was funded by Youth Program of Social Science

Research of Sichuan Province for the 13th Five-year Plan “Research on relationship
embedded model between agricultural e-commerce enterprises and customer under
the trend of industrial integration (SC16C022)”, Project of Education Department of
Sichuan Province “Research on the development choice between diversification or spe-
cialization of agricultural electricity supplier under the trend of industrial integration”
and the National Natural Science Foundation of China (71501019).
References
1. Baum JA, Mezias SJ (1992) Localized competition and organizational failure in
the manhattan hotel industry, 1898–1990. Adm Sci Q 37(4):580–604
2. Carman JM (1990) Consumer perceptions of service quality: an assessment of the
servqual dimensions. J Retail 66:33–55
3. Chen CA (2016) How can taiwan create a niche in Asia’s cruise tourism industry?
Tour Manag 55:173–183
4. Colwell RK, Futuyma DJ (1971) Application of genetic algorithm to aircraft
sequencing in terminal area. Ecology 52:567–576
5. Dobrev SD, Hannan MT (2001) Dynamics of niche width and resource partitioning.
Am J Sociol 106:1299–1337
6. Hannan MT, Freeman J (1993) Organizational ecology. Harvard University Press
7. Hou J, Lu Q, Shi YJ (2011) Study on the variation and survival factors in the
business evolution process based on organizational ecology. Ind Eng Eng Manag
(IEEM) 5:1922–1926
8. Huang F (2001) Eco-principles’ application in the optimum of tourist system. Ecol
Econ 11:19–20 (in Chinese)
9. Hutchinson GE (1957) Concluding remarks: population studies animal ecology and
demography. Cold Spring Harb Symp Quant Biol 22:415–427
10. Lew AA, Duval DT (2008) Long tail tourism: new geographies for marketing niche
tourism products. J Travel Tour Mark 25:409–419
11. Lin S, Ren P et al (2014) Research on evaluation of travel service quality for
jiuzhaigou valley on the basis of servqual. Wit Trans Inf Commun Technol 46:2157–
2162
12. Pitts BG (1999) Sports tourism and niche markets: identification and analysis of
the growing lesbian and gay sports tourism industry. J Vacat Mark 5:31–50
13. Shan M, Li G, Chen D (2006) Research on enterprises’ competitive strategies based
on ecology niche theory. Sci Sci Manag S & T 3:159–163 (in Chinese)
14. Stuart TE, Podolny JM (2007) Local search and the evolution of technological
capabilities. Strateg Manag J 17:21–38
15. Vanessa OM (2014) Favela tourism: a new niche to be developed by Brazil. In:
Eurochrie conference “hospitality and Tourism Futures”, Dubai 6–9 October
16. Wang G, Dong GZ, Zhao JL (2008) Research on competitive pattern of theme
parks based on ecological niche: take pearl river delta as an example. Tour Trib
12:45–51
17. Wang QR, Yu G (2008) A study on the measurement of the niche of regional
tourism cities and cooperation-competition model-a case study on the Pearl River
Delta region. Tour Trib 3:50–56 (in Chinese)
18. Wang ZF (2009) Research on the niche strategy of tourism industry cluster. Hum
Geogr:12–15
19. Xiang YP (2009) A study on the evaluation of the tourism niche-case of Zhangjiajie
and Tianmenshan national forest park. Issues For Econ 2:185–188 (in Chinese)
20. Xu M, Yu J (2001) Application of servqual scale to measure service quality. Ind
Eng Manag 6:6–9 (in Chinese)
21. Yan A, Da Q (2005) Study of enterprise’ niche and its motile selection. J Southeast
Univ Philos Soc Sci Ed 7:62–67 (in Chinese)
Long Term Scheduling for Optimal Sizing
of Renewable Energy Sources for Hospitals
Shabnam Mahmoudzadeh Vaziri1 , Babak Rezaee1(B) ,

and Masoud Amel Monirian2
1
Faculty of Engineering, Industrial Engineering Department,
Ferdowsi University of Mashhad, Mashhad, Iran
babak rezaee@aut.ac.ir
2
Masoud Amel Monirian, Nian Electronic Company, Mashhad, Iran
Abstract. Energy use is on the rise due to population increment and

technology development. Generating electricity energy with fossil fuels
causes environmental problems; moreover, their cost is high because of
their depletion. For solving energy crisis and environmental problems,
alternative methods for energy generation should be considered such as
Renewable energy sources (RESs) which are considered as a promising
alternative for fossil fuels. On the other hand, hospitals are one of the
major consumers of energy. Energy costs consist a large amount of hospi-
tals’ costs; therefore, hospitals aim to minimize these costs owing to com-
petition among health centers. The objective of this paper is to propose a
long term scheduling model for the optimal sizing of the RESs among can-
didate technologies to meet the hospital’s demand while minimizing the
costs and maintaining reliability subject to economic and technical con-
straints. The model is run with real data from a hospital in Iran.
Keywords: Long term scheduling · Optimal sizing · Renewable energy

sources · Hospitals
1 Introduction
Energy use has increased in last decades due to population and per capita usage
increment. Due to fossil fuel sources depletion, importance of environmental
issues and decrease in carbon emission, the energy crisis is on the rise. According
to the energy crisis, most of the countries intend to generate energy with alterna-
tive sources and change their energy portfolio. Renewable energy sources (RESs)
can reduce green gas emissions and environment pollution; therefore, they are
proper alternatives for current generation method, and some researches focused
on optimal sizing of RESs with different constraints. Moradi et al. [5] considered
renewable energy sources as a reliable source of electricity. In this paper, they
used Net Present Value (NPV) as an economic factor to determine type and
capacity of sources. Sharafi and Tarek [8] presented a multi-objective model for
determining the optimal size of renewable energy sources. They used ε-constraint
Long Term Scheduling for Optimal Sizing of Renewable Energy Sources 1343
method to minimize total cost, fuel emission, and unmet load. Yamada and Raji
[2] presented a model for determining the size of wind turbines, batteries, and
solar panels. The goal of their paper was to minimize the costs in three years time
horizon. Energy balance constraints, battery constraints, inverter model and load
model constraints were considered in the paper. Abd-el-Motaleb and Bekdach
[1] proposed stochastic model for optimizing size of wind turbines and energy
storage units. They considered energy balance constraints, storage constraints,
and also reliability constraints. Theo et al. [9] presented a linear optimization
model for a hybrid power system in order to minimize net present value of costs.
They chose wind turbines, solar panels and battery in an eco-industrial park as
a case study.
Health centers and hospitals are one of the major consumers of electricity,
where is necessary to generate energy in a continuous way and without black-
outs. Health care forms a significant cost for governments; however, thanks to
energy development, health care costs can be reduced [12]. Health care systems
are evolving rapidly, and the number of patients is increasing due to emergence
of new diseases and aging in society. Consequently, using modern medical equip-
ment and patient increment cause to the rise in energy use. The number of health
centers is increasing, and they evolved as a business industry because their com-
petition increases. In competition market, patients have to pay parts of the costs,
so health centers intend to decrease their costs as far as possible. Since hospitals
consist 31% of health care costs, hospitals are considered in this paper. One way
for reducing costs in hospitals is installation of RESs for energy generation [11].
Many countries allocate incentives to motivate investors for using RESs. There-
fore, hospitals can use government’s subsidy for supporting renewable energy
sources. In this paper, hospitals are considered as energy consumers. Although
supplying efficient energy is essential in hospitals, there are few researches on
energy use in hospitals. Some scholars research about heating and cooling sys-
tems in hospitals [3,6]. Mavrotas et al. [4] developed a bi-objective model for
minimizing costs and maximizing demand satisfaction. The technologies consid-
ered in their model were combined heat and power unit for providing power and
heat, an absorption unit and/or a compression unit for providing a cooling load.
Other papers focused on proper equipment selection in hospitals are [7,13].
The goal of this article is to determine optimal sizing of RESs in order
to minimize the costs and conserve reliability. The paper is organized as fol-
lows. Section 2 proposes problem definition and mathematical model. Section 3
presents the numerical example, analyses results, and discusses sensitivity analy-
sis. Finally, a conclusion is presented in Sect. 4.
Since a significant amount of costs in a hospital is related to energy costs, careful

energy planning is of a special importance in hospitals. In order to minimize the
costs and reduce greenhouse gas emissions, hospitals can use renewable energy
1344 S.M. Vaziri et al.
sources. Due to the importance of environmental issues, and consumer moti-

vates for using RESs, governments consider some incentives for consumers pro-
ducing renewable energies. Renewable energy sources are dependent on the nat-
ural situation for generation, and they are unpredictable; consequently, one way
to increase their generation reliability is to use two or more renewable sources
simultaneously [10]. For increasing generation reliability, wind turbines and solar
panels are used in this paper, which means that hospital, according to its budget,
can invest in solar panels and wind turbines for energy generation. Some part of
required energy can be produced by RESs in this paper, and other parts are sup-
plied by the grid. Moreover, the hospital can sell its extra produced energy to the
grid. Economic factors such as inflation rate and interest rate affect decisions, so
aforementioned factors are considered in this model by using NPV method for
calculating objective function. The purpose of the model is to minimize invest-
ment costs in a specific period while maintaining RESs system’s reliability. By
solving the model, questions like the optimal size of wind turbines and solar
panels and the optimum time for investment are answered.
Some assumptions are considered for the mathematical model as follows:
• The planning horizon is ten years which is divided into 120 months.
• According to geographic location, the hospital can invest in wind and solar
energy.
• The hospital is connected to the main grid. Purchased energy price from the
grid is varied in different time slots.
• Extra generated energy by RESs is sold to the grid with fixed price.
• Simultaneous buying energy from the grid and selling energy to the grid are
not allowed.
• The investment cost of RESs is varied in time slots which is increased or
decreased by the inflation rate.
• Interest rate and inflation rate are known and stable.
• The specific amount of money is considered as a budget for each time slot.
The extra budget does not transfer to next time slot.
• Since the goal of the hospital is to focus on medical problems, the allocated
area for installing RESs is limited.
• It is assumed that renewable energy sources use different rate of their capacity
in each time slot (βs, βw).
• In each time slot, minimum amount of energy has to be purchased from the
grid to be used in emergency situations.
• For motivating hospitals to invest in renewable energy sources, government
pays a% of sources investment cost.
To model the problem, the following notations are defined:

Indices:
t : Planning horizon (month) t = 1, 2, · · · , T .
Parameters:
cap : Capacity of 1 kW of solar panels and wind turbines in a month;
βst : Utilization percentage of 1 kW of solar panels at time t;
βwt : Utilization percentage of 1 kW of wind turbines at time t;
f s : Required area for installing 1 kW solar panel (Square meters);
f w : Required area for installing 1 kW wind turbine (Square meters);
Fs : Total area for solar panels (Square meters);
Fw : Total area for wind turbines (Square meters);
dt : Energy demand of hospital at time t;
rt : Reliability at time t;
smax : Maximum amount of allowable energy which can be bought from the grid;
zmax : Maximum amount of allowable energy which can be sold to the grid;
ps : Investment cost of 1 kW solar panel;
pms : Maintenance cost of 1 kW solar panel;
pw : Investment cost of 1 kW wind turbine;
pmw : Maintenance cost of 1 kW wind turbine;
ct : Cost of buying energy from the grid at time (cent);
w : Price of selling energy to the grid (cent);
It : Budget at time t;
ss : Energy bought for emergency situations;
μ : Interest rate in a month;
α : Inflation rate in a month;
a : Government subsidy.
Decision Variables:
xt : Installed capacity of solar panels at time t (kW);
yt : Installed capacity of wind turbines at time t (kW);
zt : Amount of energy bought from the grid at time t (kW);
st : Amount of energy sold to the grid at time t (kW);
ut : Binary variable (1 if energy has sold to the grid).
Scheduling model intends to determine optimal sizing of RESs during 10 years
time horizon as follows,
T

(ps + pms )(1 − a)(1 + α) xt
t t
(pw + pmw )(1 − a)(1 + α) yt
min t + t (1)
t=1 (1 + μ) (1 + μ)

ct (zt + ss) wst
+ t − t
(1 + μ) (1 + μ)
Subject to
t
t

βst × cap × xj + βwt × cap × yj + zt − st = dt , ∀t (2)
j=1 j=1
t t
(ps + pms )(1 + α) xt + (pw + pmw )(1 + α) yt + ct (zt + ss) − w × st ≤ It , ∀t
(3)
T

f s × xt ≤ Fs (4)
t=1
T

f w × y t ≤ Fw (5)
t=1
st ≤ smax × ut , ∀t (6)
zt ≤ zmax (1 − ut ) , ∀t (7)
t t
dt − βst × cap × j=1 xj − βwt × cap × j=1 yj + st
≤ (1 − rt ), ∀t (8)
dt
xt , yt ≥ 0, integer, ∀t (9)
zt , st ≥ 0, ∀t (10)
ut ∈ {0, 1} , ∀t. (11)
The objective function is minimizing the cost of supplying energy from renew-
able sources and grid. The two first expressions calculate the investment and
maintenance cost of the wind and solar energy sources by considering interest
rate and inflation rate. The third expression shows the cost of buying energy
from the grid, and the last one expresses hospital’s income from selling energy
to the grid.
Since supplying reliable energy is essential in hospitals, all the demand should
be met which is presented in constraint (2). Constraint (3) represents the limi-
tation of hospital’s budget in each time slot. Area limitation for installing solar
panels and wind turbines have been avouched by constraints (4) and (5), respec-
tively. In each time slot, it is not possible to buy energy from the grid and sell
energy to the grid simultaneously, which have been determined in constraints
(6) and (7). Constraint (8) represents renewable energy sources’ reliability. In
this equation, the left expression shows the amount of energy which has been
purchased from the grid which should be less than a specific amount. Constraints
(9)–(11) indicates variable types.
3 Numerical Example
To show the performance of the model, it is solved for a hospital in Iran. The
numerical example is carried out by using IBM ILOG CPLEX Optimizer v12.3.
As mentioned, the goal of this model is to find the optimal size of renewable energy
sources (RESs) installed in the hospital during the ten years time horizon. For
reducing forecasting errors, years are divided into months time slots (120 time
slots). Hospital’s monthly demand and budget are estimated during meetings with
its experts. Some of the data used in the model are shown in Table 1.
The results are shown in Fig. 1. With scrutiny, it is obvious that investment
is higher in first time slots. It is reasonable due to the positive inflation rate.
Moreover, investments in solar panels are more than wind turbines owing to the
fact that they have more productivity in the aforementioned hospital, and also
they need fewer investment costs.
Table 1. Data related to time slots
Time d(kW) I (dollars) βs βw

1 12000 50000 0.5 0.23
2 12000 50000 0.6 0.3
.. .. .. .. ..
. . . . .
120 15000 8750 0.25 0.3
Fig. 1. Installed kW of wind turbines and solar panels in each year
Total solar panels which will be installed during ten years is 56 kW and
installed wind turbines are 9 kW.
Sensitivity analysis is done for effective parameters on model such as inflation
rate, interest rate, demand, reliability and etc. Some of them which are more
important are elaborated in this paper.
In the aforementioned hospital, a high amount of budget is allocated to the
first time slot because the decision for investment is given in this time slot. It is
assumed that the budget is allocated to the hospital in the sixth time slot or at
the beginning of the second year (13th time slot) with considering its interest.
According to results, which is showed in Table 2, if the interest rate is between
15 and 20%, it is better to allocate budget in the first time slot. If the interest
rate is between 20 and 24%, it is better to allocate budget in the 6th time slot.
Table 2. Inflation rate sensitivity analysis
Capital allocating method Capital Capital Capital

allocation in the allocation by allocation by
first period considering its considering its
profit in the profit in the
sixth month second year
Objective Inflation rate = 15% 80226 82562 83379
function Inflation rate = 20% 67324 64824 67849
Table 3. Purchased price sensitivity analysis
40% 20% Current 20% 40% 250%

decrease decrease price increase increase increase
in price in price in price in price in price
Solar 56 56 56 56 56 56
panels kW
Wind 9 9 9 9 9 9
turbines
kW
Amount of 96460 90828 89744 88544 87416 53020
purchased
energy from
the grid
Fig. 2. Installed kW of wind turbines and solar panels in each year
When purchased price is increased, it is expected that investment in RESs is

increased too. However, other scenarios can be happened due to other parame-
ters. For example, it is possible that the number of RESs is decreased or even
stay stable, but the time of investment is changing in these scenarios. Accord-
ing to results, shown in Table 3, with increment or decrease in purchased price,
number of installed RESs is stable, and the time of investment is just changed.
For understanding such changes, variations of investment time for current price
and 250% price increment are shown in Fig. 2 for ten time slots. With current
price, investment in wind turbines is started from 66th time slot; however, if pur-
chased energy increases 250%, investment in wind turbines starts from 9th time
slot, which causes to take advantage of wind generation in 57 more time slots.
Energy price is low in Iran; however, in developed countries such price is a
significant amount. As a result, we compared current price in Iran with current
price in United States, which is about 250% higher than energy price in Iran.
Table 4. Wind turbine utilization percentage sensitivity analysis
30% 20% 10% Current 10% 20% 30%

increase wind wind utiliza- wind wind wind
in wind turbines turbines tion turbines turbines turbines
turbines utiliza- utiliza- percent- utiliza- utiliza- utiliza-
utilization tion age tion tion tion
tion percent- percent- percent- percent- percent-
percentage age age age age
age
Solar kW 48 48 51 56 57 60 61
panels
Wind 12 13 12 9 9 7 7
turbines
kW
Table 5. Wind turbine utilization percentage sensitivity analysis
20% increase 10% increase Current 10% decrease 20% decrease

in solar in solar utilization in solar in solar
panels panels percent- panels panels
utilization utilization age utilization utilization
percentage percentage percentage percentage
Solar panels 52 53 56 57 64
kW
Wind 4 7 9 13 13
turbines kW
Table 4 shows sensitivity analysis for wind turbine utilization percentage.

When wind turbine utilization percentage increases, investment in solar panels
decreases, and investment in wind turbines increases. Whatever the wind turbine
utilization percentage increases, there is less need for wind turbines. With wind
turbine utilization percentage reduction, the number of solar panels increases.
Table 5 presents sensitivity analysis for solar panels utilization percentage. By
increment in solar panels utilization percentage, amount of installed kW of wind
turbines decrease. Moreover, there is less need to install solar panels owing to the
fact that more demand will be met by less solar panels with higher utilization
percentage. On the other hand, by decrease in solar panels utilization percentage,
the number of wind turbines increases as far as area constraint permits.
4 Conclusion
In this paper, long term scheduling model was presented for determining optimal
sizing of renewable energy sources in hospitals. The model aimed at minimizing
costs while maintaining reliability according to specific constraints in hospitals
such as area limitation and budget constraint. By solving the model with real
data of the hospital in Iran, the model can reduce 10% of costs comparing with
the time when energy is only purchased from the grid. In presented model,
planning horizon was divided into monthly time slots which caused an increment
in forecasting errors. For further researches, authors can deal with daily or weekly
data. Also in this model, the price of required field for installing wind turbines
and solar panels was not considered. Authors can investigate the effect of field
price on model outputs.
Acknowledgements. The authors would like to acknowledge Professor Mitsuo Gen

for his scientific support.
References
1. Abd-El-Motaleb AM, Bekdach SK (2016) Optimal sizing of distributed generation
considering uncertainties in a hybrid power system. Int J Electr Power Energy Sys
82:179–188
2. Atia R, Yamada N (2016) Sizing and analysis of renewable energy and battery
systems in residential microgrids. IEEE Trans Smart Grid 7(3):1–10
3. Bizzarri G, Morini GL (2006) New technologies for an effective energy retrofit of
hospitals. Appl Therm Eng 26(2–3):161–169
4. Mavrotas G, Diakoulaki D et al (2008) A mathematical programming framework
for energy planning in services’ sector buildings under uncertainty in load demand:
the case of a hospital in athens. Energy Policy 36(7):2415–2429
5. Moradi MH, Eskandari M (2014) A hybrid method for simultaneous optimization
of dg capacity and operational strategy in microgrids considering uncertainty in
electricity price forecasting. Int J Electr Power Energy Syst 56(7):241–258
6. Pagliarini G, Corradi C, Rainieri S (2012) Hospital chcp system optimization
assisted by trnsys building energy simulation tool. Appl Therm Eng 44(6):150–
158
7. Renedo C, Ortiz A et al (2006) Study of different cogeneration alternatives for a
spanish hospital center. Energy Build 38(5):484–490
8. Sharafi M, Elmekkawy TY (2014) Multi-objective optimal design of hybrid
renewable energy systems using pso-simulation based approach. Renew Energy
68(68):67–79
9. Theo WL, Lim JS et al (2016) An milp model for cost-optimal planning of an
on-grid hybrid power system for an eco-industrial park. Energy 116:1423–1441
10. Thomaidis NS, Santos-Alamillos FJ et al (2016) Optimal management of wind and
solar energy resources. Comput Oper Res 66:284–291
11. Vaziri SM, Rezaee B, Monirian MA (2017) Bi-objective integer programming of
hospitals under dynamic electricity price. In: Proceedings of the tenth international
conference on management science and engineering management. Springer, pp 409–
417
12. Wang YC, Mcpherson K et al (2011) Health and economic burden of the projected
obesity trends in the USA and the UK. Lancet 378(9793):815–825
13. Yoshida S, Ito K, Yokoyama R (2007) Sensitivity analysis in structure optimization
of energy supply systems for a hospital. Energy Convers Manag 48(11):2836–2843
Life Cycle Assessment of Waste Mobile Phone
Recycling–A Case Study in China
Tingting Liu, Moudi Mahdi, and Liming Yao(B)
Uncertainty Decision-Making Laboratory, Sichuan University, Chengdu 610064,

lmyao@scu.edu.cn
Abstract. With the rapid pace of technological development, electronic

waste (e-waste) has become one of the fastest growing waste streams in
the world. The exponential growth of e-waste, which is disposed of in
landfills, incinerated, or dismantled, now significantly contributes to the
massive increase in overall waste. This paper assesses the waste produced
from recycling mobile phones in China and examines the influence of the
waste that is released into the environment. The mobile phone recycling
process at formal recycling facilities in China is investigated as a case
study. The results highlight the toxicity of the waste and the influence
such waste has had on the environment. Life cycle assessment (LCA)
was employed to analyze the mobile phone recycling process and Gabi
was used to quantitatively indicate the flows and environmental impacts.
The results showed that resources recovery, human health, and ecosystem
quality are the main effects of the mobile phone recycling waste produced.
Based on this quantitative analysis, to reduce the environmental impact
of mobile phone recycling, the demand for recycled materials, environ-
mental awareness, law enforcement, and the e-waste recycling system
were found to be significant drivers.
Keywords: Waste mobile phone · Life cycle assessment · Recycling
1 Introduction
Electronic waste, also known as e-waste, has been one of the major contributors
to the waste stream since the rapid growth of advanced technological products
[6]. E-waste includes waste electrical and electronic products such as televisions,
desktop computers, microwaves and mobile phones. In recent years, economic
and technological development in China has seen as massive growth in mobile
phone penetration and a consequent rise in mobile phone waste.
Currently over 40 million tonnes of e-waste is generated annually in the world
and is growing every year. Of all the kinds of e-waste, the growth in mobile phone
waste has been three times faster than total annual waste [19]. Current mobile
phone using data indicates that there are 7.2 billion active mobile phones in
2016, with over 1.305 billion of these being in China, a penetration rate of 93.1%
1352 T. Liu et al.
per 100 inhabitants subscripƟon（in millions)
120.0 8,000
Mobile-cellular telephone
per 100 inhabitants
subscripƟon(in millions)
7,000
100.0
6,000
80.0
5,000
60.0 4,000
3,000
40.0
2,000
20.0
1,000
- -
Fig. 1. Global mobile-cellular subscriptions, total and per 100 inhabitants, 2005–2016.
Note: * Estimate
100.00 Mobile-cellular telephone 1,400.00

subscripƟons
90.00
per 100 inhabitants 1,200.00
80.00
70.00 1,000.00
60.00
800.00
50.00
600.00
40.00
30.00 400.00
20.00
200.00
10.00
0.00 0.00
2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015
Fig. 2. China mobile-cellular subscriptions, total and per 100 inhabitants, 2005–2015
(see Figs. 1 and 2) [25]. About 77 million mobile phones are discarded every year
in China, a figure the United Nations environmental program has predicted to
be 7 times higher by 2020 [11].
The growing quantity of e-waste has begun to have a severe impact on the
environment; especially in developing countries. As the leading manufacturing
country of the world, China has become an “electronic waste treatment plant”,
with informal mobile phone recycling having become a major source of toxic
pollution [13]. With rising climate change and water and land pollution con-
cerns, mobile phone recycling waste has attracted increasing attention, forcing
the mobile phone production industry to find solutions to decrease the environ-
mental impact. In addition, various methods have been developed for mobile
phone recycling to reduce the environmental damage. There has been a growing
body of research into the recycling of waste mobile phones in China and the
inherent problems. Wang et al. evaluated the potential yield of indium recycled
from waste mobile phones in China [21], and Yi-Bo et al. and several others have
examined the recovery and recycling of old mobile phones in China [23], and the
other scholars did [4].
Life Cycle Assessment of Waste Mobile Phone Recycling–A Case Study 1353
The main methods that have been applied to waste management analy-
sis are data envelopment analysis(DEA) [2], the index decomposition analysis
method(IDA) [12], and the structure decomposition analysis method (SDA) [1].
Life Cycle Assessment (LCA) was first used in 1969 when the Midwest Research
Institute tracked and quantitatively analyzed the complete process of beverage
containers from raw materials extraction to final disposal [3]. With development,
the LCA has been included in the ISO14000 series of environmental management
standards and has become an important support tool for international envi-
ronmental management and product design [8]. LCA has been widely used in
environmental assessments in such areas as food waste management [14], waste
water [18], and plastic production as well as in waste electrical and electronic
equipment (WEEE) management [10].
As an environmental management tool, LCA has expanded the boundaries of
research systems, as it encompasses the entire life cycle of resource use, energy
consumption and waste discharge and provides an evaluation of the potential
environmental impact at each stage. This paper assesses the recycling of waste
mobile phones in China using LCA. To date, analyses of the environmental
impact of recycling waste mobile phones has lacked a systematic and scientific
focus and has tended to ignore the management of discarded mobile phones [15].
In this paper, the life cycle assessment of the discarded mobile phone recycling
process is quantitatively evaluated to identify the main factors influencing the
environment and provide decision support for the waste management of other
small electronic devices.
The remainder of this paper is structured as follows. In Sect. 2, the current
recycling process in China is outlined, an abandoned mobile phone path analysis
is conducted, and the latest recycling process explained. Section 3 applies the
LCA to the mobile phone recycling process and gives the results. Section 4 gives
suggestions and recommendations based on the analyses.
2 Mobile Phone Recycling in China
2.1 Recycling Status
China has had a rapid growth in e-waste streams over the past ten years [11],
with mobile phone waste being the most rapidly growing. Because there has
not been a regimented recycling system established in China, most waste mobile
phones are kept, donated to others or discarded with other house-hold waste; the
formal recovery rate has been only about 1% [25]. The recycling of batteries and
other waste mobile phones parts has mainly been conducted through nonpro-
fessional channels such as street vendors or small electronics shops or through
recycling platforms on the internet. After recovery, refurbished mobile phones
are generally sold in a second-hand market or sold at a low price in remote towns
and villages. Mobile phones that cannot be refurbished often end up in illegal
dismantling workshops, where workers extract the copper, gold, silver and other
valuable metals using such processes such as acid leaching and open burning,
1354 T. Liu et al.
processes which not only damage the health of workers but also severely pollute
the surrounding environment.
From a product to waste, a mobile phone can change hands hundreds of times;
they can be retained by consumers, returned to a retailer, sold to a second-hand
shop or sent to recycling. The specific circulation paths are shown in Fig. 3.
Recycling Centre
Waste CollecƟon
Recycling Factory
Retailer
Holder
Second-hand
Dealer/Trader
Shop
Junk Shop
Repair Centre Idle at Home
Fig. 3. The flow of waste mobile phones in China
2.2 Recycling Mode
This paper evaluates the pollution produced in the recycling phase from the
perspective of material flows using the standard life cycle assessment method,
examines the waste mobile phone recycling process in formal recycling facilities
in China as a practical case study (see Fig. 4) and identifies the factors that
adversely impact the environment in the recycling phase.
The brief description of the recycling process is as follows: after being trans-
ported to the waste treatment plant, the mobile phones are manually disas-
sembled and the phone casings, printed circuit boards, lithium batteries, LCD
screens and other main components extracted. Then, each of these components
is processed separately. The ABS/PC plastic casings are crushed into plastic
pellets then reused as raw materials, the stainless steel, copper and other metals
are melted, with a further smelting processing step needed for the printed circuit
boards and lithium ion batteries (recovery of electrolyte solvent) to recover the
copper, gold, silver, palladium, cobalt, lithium, tin and other metals. LCD dis-
plays are generally incinerated to reduce environmental waste and the remaining
residues are put into landfill.
Mentioned in the process: (1) the life cycle of the mobile phone at each
stage from production, use, and disposal including the transportation of the
various materials; however, the transportation and landfill residue stage after
disassembly are not taken into account. (2) only lithium ion scrap mobile phone
batteries and mobile phone LCD displays are considered.
Fig. 4. Recycling process
3 LCA on Recycling Waste Mobile Phone
The LCA has four interrelated steps, which we applied to the mobile phone
recycling process, the rough contents for which are shown in Fig. 5.
•goal:analyze the influence •input(energy & resources) •influence on environment

on environmental impact •output(emissions to •impact on human health
•scope:the recycling air/water&human health) •ecotoxcity(air/freshwater)
process for waste mobile •data collecƟon
phones •material flows
Goal &Scope Inventory Impact

definiƟon analysis assessment
•analysis
•suggesƟons
InterpretaƟon
Fig. 5. Detailed content of the four steps
3.1 Goal and Scope Definition
The study object in this paper is the abandoned mobile phone. The functional
unit is set to a tonne of phones of various brands and models, with the data for
each component being set at an average value. The study scope for the system
1356 T. Liu et al.
Fig. 6. Flows at a formal recycling facility
is set for disassembly, waste collection, sorting, processing and final disposal
(Fig. 6) and an inventory analysis and impact assessment for the energy and
resource consumption and pollutant emissions at each stage is established.
3.2 Inventory Analysis
From relevant research data regarding domestic and foreign waste mobile phone
processing and combined with the Chinese current waste recycling, the main
process inventory data related to the recycling process was determined (see
Table 1). The data in this paper was taken from the ecoinvent database [7];
Table 1. Inventory analysis
Process flow Inventory data

Disassembly GLO: manual treatment plant, WEEE scrap [Recycling]
= 2500 tone/yr. GLO: mechanical treatment plant,
WEEE scrap [Recycling]/GIO/IU = 50,000 tone/yr
Plastic recycling GLO: manual treatment plant, WEEE scrap [Recycling]
= 2500 tone/yr
Iron metal RER: steel, electric, un- and low-alloyed, at plant
[Beneficiation]
Metal SE: copper, secondary, from electronic and electric scrap
recycling, at refinery [Beneficiation]
Printed Circuit Boards (PCBs) GLO: disposal, treatment of printed wiring boards
[Recycling] SE: gold, secondary, at precious metal refinery
[Beneficiation] SE: silver, secondary, at precious metal
refinery [Beneficiation] SE: palladium, secondary, at
precious metal refinery [Beneficiation]
Tin RER: tin, at regional storage [Beneficiation]
Li-ions batteries GLO: disposal, Li-ions batteries, hydrometallurgical
[Recycling] GLO: cobalt, at plant [Beneficiation] GLO:
lithium hydroxide, at plant [inorganics]
LCD CH: disposal, LCD module, to municipal waste
incineration [municipal incineration]
Landfill Disposal, municipal solid waste, 22.9% water, to sanitary
landfill/CH U
as an overview and methodological framework for the LCA, the ecoinvent data-
base is the most consistent and transparent life cycle inventory database in the
world [22].
Inventory analysis involves the tracking of the input and output flows of a
product system including the materials, water and energy used and the waste
released to the air, land and water [17]. In the model developed in this paper,
energy and raw materials consumption is taken into account and GaBi, a pow-
erful life-cycle assessment tool, is used to generate the model from the inventory
data to assess the sustainability solutions [24].
Fig. 7. Gabi built model
As shown in Fig. 7, the input for this model is 1000 kg of waste mobile phones.
3.3 Life Cycle Impact Assessment
LCIA is a qualitative and quantitative analysis of environmental impacts based

on the elements listed in the inventory analysis (LCI) [20]. It is used to evaluate
and interpret the environment impacts of a product system by assigning quan-
tifiable measurements and connecting the input and output material inventories
to obtain the damage points of a specified boundary system.
Generally speaking, the LCIA includes the following steps: classification of
the elements listed in the inventory analysis process; a qualitative and quantita-
tive analysis of the elements; identification of the major environmental factors in
all aspects of the system; and analysis and judgment of the environmental fac-
tors. This paper identifies and analyzes the major environmental factors in the
mobile phone recycling process. The environmental factors include ecotoxcity,
cancer effects and human ecotoxicity [16].
3.4 Interpretation
In this section, the evaluation and improvements for the mobile phone recycling
process are discussed, after which, with the aim of reducing the burden on the
environment, suggestions are given to improve the current recycling mode of
1358 T. Liu et al.
China. In line with the goals and scope of the analysis, the recycling process
results were evaluated, from which we were able to come to conclusions and
develop recommendations.
The inventory results from the model focused on three main aspects. The
first was the recovery of resources and materials. Valuable materials, such as
copper, gold, and silver, which can be recycled and reused, were retrieved and
the other materials treated as waste. The second aspect was human health con-
cerns. Many related environmental issues such as carcinogenic materials, climate
change, ozone depletion, radiation, and organic and inorganic respiratory irri-
tants can cause various diseases, premature death and other non-normal deaths,
reducing life expectancy. This aspect was measured in units of disability adjusted
life years (DALY), which examined two kinds of health losses; disability and
death; to determine the social impact of possible medical conditions on pop-
ulation health. This indicator has also been adopted by the World Bank and
the World Health Organization. The third aspect focuses on ecosystem quality
which refers to the ecological environmental biodiversity loss within a given time
and space caused by environmental issues such as acidification/eutrophication
of land and water bodies, toxicity, land-use changes [5], and air emission [9].
The measurement units for this aspect were the potentially disappeared fraction
(PDF) m2 yr and the potentially affected fraction (PAF) m2 yr.
4 Discussion
The Chinese recycling waste mobile phone process was examined in this paper
using LCA as the evaluation tool. A model was proposed and the possible three
impact factors; resources, human health, and ecosystem quality; were analyzed to
assess the environmental impact of the recycling process. In the model building
procedure, resources consumption and material emissions were found to have a
certain degree of impact on the environment, and further research can extract
specific data to analyze the effects of specific substances such as heavy metals,
CO2 and NO2 . Because of the speed of mobile phone technological development,
the rate of mobile phone disposal is expected to increase, further impacting the
environment. If waste mobile phones can be effectively recycled, there will be less
environmental and health effects, improving both the economy and the residents
quality of life. This paper gave an assessment on the recycling process of waste
mobile phone in China, identified the main factors affecting the environment.
On the basis of this analysis, proposals were given to improve the eco-efficiency
in recycling process. First, relevant policies should be presented to standardize
the recycling behavior of citizens and market. Second, the recycling technology
should be improved. Besides, a more eco-efficient recycling model is needed. The
three main factors in this paper; resources, human health, and ecosystem quality;
will be taken into account in the model.
Acknowledgements. The work is supported by the National Natural Science Foun-

dation for Young Scholars of China (Grant No. 71301109), and Soft Science Program
of Sichuan Province (Grant No. 2017ZR0154).
References
1. Ang BW, Xu XY, Su B (2015) Multi-country comparisons of energy performance:
the index decomposition analysis approach. Energy Econ 47:68–76
2. Barba-Gutiérrez Y, Adenso-Dı́az B, Lozano S (2009) Eco-efficiency of electric and
electronic appliances: a data envelopment analysis (DEA). Environ Model Assess
14(4):439–447
3. Beccali M, Cellura M, Finocchiaro P (2014) Life cycle assessment performance
comparison of small solar thermal cooling systems with conventional plants assisted
with photovoltaics. Energy Procedia 104(1):893–903
4. Bian J, Bai H, Li W (2016) Comparative environmental life cycle assessment of
waste mobile phone recycling in China. J Clean Prod 131:209–218
5. Bieda B (2008) The use of life cycle assessment (LCA) conception for arcelormit-
tal steel poland S.A. energy generating-KRAKW plant case study. Wydawnictwo
Politechniki Krakowskiej Im Tadeusza Kociuszki
6. Buekens A, Yang J (2014) Recycling of weee plastics: a review. J Mater Cycles
Waste Manag 16(3):415–434
7. Domain TE (2016) Life cycle inventory database. http://www.ecoinvent.org/
8. Guinee JB (2002) Handbook on life cycle assessment operational guide to the iso
standards. Int J Life Cycle Assess 7(5):311–313
9. Gurjar BR, Butler TM, Lawrence M (2008) Evaluation of emissions and air quality
in megacities. Atmos Environ 42(7):1593–1606
10. Harris R (2015) Life cycle assessment of post-consumer plastics production from
waste electrical and electronic equipment (weee) treatment residues in a central
european plastics recycling plant. Sci Total Environ 529:158–167
11. Herat S, Agamuthu P (2012) E-waste: a problem or an opportunity review of issues,
challenges and solutions in Asian countries. Waste Manag Res J Int Solid Wastes
Public Clean Assoc Iswa 30(11):1113–1129
12. Korica P, Cirman A, Gotvajn A (2016) Decomposition analysis of the waste gen-
eration and management in 30 European countries. Waste Manag Res
13. Liu X, Tanaka M, Matsui Y (2006) Electrical and electronic waste management in
China: progress and the barriers to overcome. Waste Manag Res J Int Solid Wastes
Public Clean Assoc Iswa 24(1):92–101
14. Lundie S, Peters GM (2005) Life cycle assessment of food waste management
options. J Clean Prod 13(3):275–286
15. Middendorf A (2008) Assessment of toxicity potential of metallic elements in dis-
carded electronics. J Environ Sci 20(11):1403–1408
16. Navarro E, Baun A, Behra R (2008) Environmental behavior and ecotoxicity of
engineered nanoparticles to algae, plants, and fungi. Ecotoxicology 17(5):372–86
17. Organization IS, Iso ISO (2006) Environmental management-life cycle assessment-
principles and framework. International Standard ISO
18. Pradel M, Aissani L, Villot J (2015) From waste to product: a paradigm shift in
LCA applied to wastewater sewage sludge. In: Setac Europe meeting
19. Ramesh BB, Parande AK, Ahmed BC (2007) Electrical and electronic waste: a
global environmental problem. Waste Manag Res J Int Solid Wastes Public Clean
Assoc Iswa 25(4):307–318
20. Russo G, Vivaldi GA, Gennaro BD (2015) Environmental sustainability of different
soil management techniques in a high-density olive orchard. J Clean Prod 107:498–
508
1360 T. Liu et al.
21. Wang HG, Wu YF et al (2016) An evaluation of potential yield of indium recycled

from waste mobile phone in China. In: Materials science forum
22. Wernet G, Bauer C, Steubing Bea (2016) The ecoinvent database version 3 (part
i): overview and methodology. Int J Life Cycle Assess:1–13
23. He YB, Li LQ (2007) A study on recovery and recycle of old mobile phones in
China. Nonferrous Met Eng Res
24. Zhang H (2015) Product sustainability software. http://www.gabi-software.com/
25. Zhao H (2016) ITU World Telecommunication /ICT Indicators database. http://
www.itu.int/
A Malmquist Index-Based Dynamic Industrial
Green Efficiency Evaluation in Sichuan Province
Xuhong Liu(B) and Xiaowen Jie

liuxh 99@163.com
Abstract. In this paper, the DEA-Malmquist index is used to evalu-

ate the industrial green efficiency of 18 cities in Sichuan province during
2008–2014. The study found that in the overall level, the industrial green
efficiency of Sichuan prefecture level cities’ is low, but it is growing year
by year in recent years. The technical efficiency index grows rapidly, and
the technological progress index weakened the role of technical efficiency.
Industrial green efficiency of Chengdu Plain economic zone is the high-
est in the four economic zones, followed by the Panxi economic zone,
South Sichuan economic zone, lowest in Northeast Sichuan economic
zone. Industries need to increase the intensity of the green transforma-
tion in Sichuan province.
Keywords: Malmquist index · Dynamic evaluation · Industrial green

efficiency · Sichuan province
1 Introduction
The industrial developments of China have received great achievement recent
years, but the past extensive industrial mode was at the cost of environment
pollution and excessive resources consumption, which results in the problems
of resource constraint and environment pressure more visibly. And the resource
utilization has also become one important factor to measure the competitiveness
of a country’s manufacturing industry, and the industrial green development is
the necessary way to promote international competitiveness. In order to change
the industrial mode of high involvement, high consumption and high pollution, to
promote the region industrial economic development, the 13th Five Year Plan
outline of national economic and social development had put forward higher
requirement for industrial green development.
Sichuan province, as the most important economic area of China western, its
industrial green development is in what level, whether the industrial green effi-
ciency has improved. The deep research of these issues would be great theoretical
and practical significance. So based on the urgent need of green development, the
presented paper takes the prefecture-level city as research objects, by using DEA
research method to dynamic evaluate the green efficiency of industrial inputs and
its output. And propose countermeasures and suggestions to improve industrial
green efficiency of Sichuan province.
1362 X. Liu and X. Jie
2 Literature Review
Industrial green development is the inner requirement and inevitable choice
of new industrialization. United Nations Industrial Development Organization
(UNID) definition that green development is a sustainable industrialization mode
of the production and consumption by effective utilization energy resources, low
carbon emission, low waste reduction, during the industrialization development
process to expand in order to eliminate poverty and create employment oppor-
tunities. The proposed research consider that the industrial green development
is under the theory guidance of sustainable development and circular economy,
through green design, green production, green management and other aspects of
the control, to realize the high utilization rate of energy resource, low environ-
mental disruption, high comprehensive benefits development mode.
At present, the foreign scholars are focused to research the connotation of
industrial green development and evaluation of its efficiency. Eiadat et al. [2]
and Nagesha [11] selected patent number of green process and the carbon diox-
ide emissions to assess the industrial green efficiency, Honma and Hu [6] used
the random front model to evaluate the industrial green efficiency. There are
scholars had researched the influence factors of green development on the spe-
cific industry, Vasauskaite et al. [15] put forward R&D investment, energy saving
technology or energy management system can promote the sustainable develop-
ment of Lithuania furniture industry. Fleiter et al. [4] evaluated the 17 process
technologies of the pulp and paper industry in Germany, and considered that the
innovation of paper mills and paper drying technologies in heat recovery systems
is the most influential technology for energy conservation and green development.
Mukherjee et al. [9] proposed computer-aided process design can optimize the
energy and material flow to minimize the impacts from other harmful emissions.
Kanagaraj et al. [10] reviewed different cleaner technological methods of reducing
the generation of solid and liquid wastes in leather industry.
Domestic scholars Chen et al. [1] researched the influencing factors
of industrial green development. Wang and Huang [16] according to the
five indexes of industrial and social development, recycling, resource reduc-
tion, pollution reduction, resource and environmental security, used AHP
method to evaluate the development of industrial circular economy in
Jiangsu province from 1985 to 2003. And they introduced the concept of
“obstacle degree”, to quantitatively diagnosis obstacle in different stages
of industrial circular economy development in Jiangsu province quantita-
tive. Tu and Xiao [14] established frontier function model of environmen-
tal by directional distance function, and calculated the green total factor
productivity of industrial enterprises above designated size, with energy con-
sumption and environmental pollution in 30 provinces of China in 1998–2005.
As well the pollutants considered the emissions of industrial sulfur dioxide. It
has found that the industrial development In the eastern coastal areas was more
harmonious with the environment, and the environmental technology efficiency
was generally low in the central and western regions. Mei et al. [8] established
the evaluation index system of green development efficiency of coal industry
A Malmquist Index-Based Dynamic Industrial Green Efficiency Evaluation 1363
by improved DEA. Wu [17] analyzed China’s industrial total factor produc-

tivity (TFP) growth in 1998–2007 under environmental constraints through
Malmquist-Luenberger index. The results showed that the TFP of western indus-
try was growing fastest, eastern followed, while the central growth was the slow-
est. Zhang and Zhu [22] took the non-expected industrial exhaust emissions as
input variables, and measured the efficiency of industrial innovation based on
green growth. It was considered that there was a phenomenon of increasing
returns to scale in the Western region. Yang [19] used DEA method to measure
the efficiency of low carbon economy in Sichuan Province, based on the analysis
of energy consumption and carbon emissions, and concluded that the efficiency
of developing low-carbon economy was still not high in Sichuan Province, which
was mainly restricted by technology shortage and scale bottleneck.
Therefore, the researches of industrial productivity considering of resource
environmental factors are basically from the regional perspective. There are few
studies on the efficiency of industrial development from the perspective of Eco-
nomic Zone. Or mostly concentrated in the eastern region of China, such as the
Pearl River Delta urban agglomeration, Shandong peninsula urban agglomera-
tions and other places, the city of the western region was less discussed. So the
presented paper based on the original data of prefecture level cities in Sichuan
Province, evaluates the status of its industrial green development and lays the
foundation for further research.
3 General Situation of Industrial Green Development in

Sichuan
At present, industrial development in Sichuan Province is in the middle stage,
but according to the environmental Kuznets curve, the negative impact on energy
and the environment is still relatively large of the current industrial development
in Sichuan [7]. Moreover, Sichuan is a traditional resource and energy province,
in which has proven reserves of 132 kinds of mineral resources, accounting for
70% of the total resources in China. This will inevitably lead to its industrial
development mainly resource oriented. From 2005 to 2015, the average annual
growth of GDP in Sichuan province is 8.1%, but the average annual growth of
total energy consumption is 3.5%, which put enormous pressure on Sichuan’s
resources.
It can be seen from Fig. 1 that, energy consumption per unit of industrial
added value and energy consumption per unit of GDP in Sichuan continued
to decline (2005–2013), decreased by 48.64% and 32.67% respectively. This
shows that the efficiency of industrial energy utilization in Sichuan province
has increased rapidly, and the industrial structure of its low-end has improved.
But on the other hand, by the end of 2014, raw material resources traditional
industry in Sichuan province accounted for up to 43%, and low-end manufactur-
ing industry accounted for nearly 30%. It indicates that Sichuan province still
needs to further increasing the green structure transformation, so as to improve
the utilization efficiency of resources and energy. Figure 2 showed that the dis-
charge of industrial waste water, waste gas and solid waste (hereinafter referred
to as three wastes) per unit of GDP in Sichuan Province decreased year by year
during the period of 2005–2013, with the decreasing rate of 38.96%, 32.07% and
85.19%. These were higher than the decreasing rate of national industrial three
wastes emissions (30.63%, 21.85% and 72.89%). It indicates that in the process
of industrial development, Sichuan Province compared to the national, the level
of mitigation to environmental pollution is higher. In addition, although the
industrial waste water, waste gas per unit of GDP in Sichuan province is lower
than the national level after 2010, the industrial solid waste generated per unit
of GDP is still higher than the national recently. It showed that the situation of
environmental pollution in Sichuan province is still grim, and it is an important
way of strengthening the industrial green transformation to ease the pressure of
resources and the environment.
2005 2.94
1.01
2006 2.79
0.97
2007 2.60
0.93
2008 2.39
0.89
2009 2.25
0.84
2010 2.00
0.81
2011 1.84
0.77
2012 1.62
0.72
2013 1.51
0.68
0.00 1.00 2.00 3.00

Tons of standard coal / million
Energy consumpƟon per unit of industrial added value
energy consumpƟon per unit of GDP
Fig. 1. Resource consumption situation of Sichuan province in 2005–2013 (Source:

China Statistical Yearbook)
4 The Measurement of Industrial Green Efficiency
4.1 Malmquist Index Model
Malmquist index was proposed by the famous economist Malmquist in 1953, then
Cavers et al. and Fare applied this index to estimate productivity. It can describe
multiple input variables and multiple output variables of production technology,
and is a method of measuring total factor productivity using panel data by
16.00
14.00
12.00
10.00
8.00
6.00
4.00
2.00
0.00
2005 2006 2007 2008 2009 2010 2011 2012 2013
Industrial waste water discharged per per unit of GDP in Sichuan(Million tons / Billion yuan)
Industrial waste water discharged per unit of GDP in China(Million tons / Billion yuan)
Industrial waste gas discharged per unit of GDP in Sichuan(Billion standard cubic meters/Billion yuan)
Industrial waste gas discharged per unit of GDP in China(Billion standard cubic meters/Billion yuan)
Industrial solid wastes generated per unit of GDP in Sichuan(Million tons / Billion yuan)
Industrial solid wastes generated per unit of GDP in China(Million tons / Billion yuan)
Fig. 2. Comparison of “three wastes” emissions in Sichuan province and China in 2005–
2013 (Source: calculated according to the Chinese Statistical Yearbook 2006–2014)
nonparametric. In empirical analysis, researchers generally use Malmquist index

based DEA by Fare et al. [3]. Its true value is approximately:

D0 t+1 xt+1 , y t+1 D0 t (xt+1 , y t+1 ) D0 t+1 (xt+1 , y t+1 )
t+1 t+1 t+1 t t
M0 x ,y ,x ,y = × × .
D0 t (xt , y t ) D0 t (xt , y t ) D0 t+1 (xt , y t )
(1)
In formula (1), M0t+1 is the index for decision making unit DMU0 in t+1
period, (xt , y t ) and (xt+1 , y t+1 ) represent input-output vectors for period t to
period t + 1 respectively, D0t and D0t+1 represent the distance function taking
technology as a reference for period’s t and t + 1 techniques. Further, under the
assumption of variable returns to scale, formula (1) can be rewrited as follows:

t+1
t+1 t+1 t t D0 t+1 xt+1 , y t+1 |V
M0 x ,y ,x ,y =
D0 t (xt , y t |V )

D0 t+1 xt+1 , y t+1 |C D0 t (xt , y t |V )
× (2)
D0 t (xt , y t |C) D0 t+1 (xt+1 , y t+1 |V )

D0 t (xt+1 , y t+1 ) D0 t (xt , y t )
× t+1 × .
D0 (xt+1 , y t+1 ) D0 t+1 (xt , y t )
In formula (2), V represents a variable size, C represents the same size. The
first item represents the pure technical efficiency index (Pech), the second rep-
resents the scale efficiency index (Sech), and the third represents the technical
progress index (Techch), that is Tfpch = Effch × Techch = Pech × Sech ×
Techch. Among them, the technical efficiency index measures the change degree
to the production frontier of each decision unit from period t to period t + 1,

the technical progress index measures the moving degree of production frontier
from period t to period t + 1, the pure technical efficiency index reflects the tech-
nical efficiency when the scale returns unchanged, and the scale efficiency index
reflects the optimal production scale of the decision unit. This paper uses direc-
tional output variables to measure the industrial green efficiency of prefectural-
level cities in Sichuan. When the index is equal to 0, indicating it has no effect
on the industrial green efficiency. If the index is greater than 1, indicating a
promotion effect, or vice versa, showing there is a blocking effect.
4.2 Evaluation Indexes and Data Sources

The selection indexes in industrial green efficiency evaluation system are not
only to meet the needs of industrial economic development, but also to meet the
demands of environmental and social development. The presented paper selects
evaluation index of industrial green efficiency combining with previous studies
[13,18,21], and on the basis of following principles of scientific, systematic and
operational. As shown in Table 1, input indicators select key elements of indus-
trial development process. X1 and X2 represents human input and capital input
of industrial development respectively. X3 reflects investment in energy appli-
cation. Output indicators select total industrial output value (Y1) and total
Table 1. Evaluation index of industrial green efficiency
Indicator types Index code and index Unit Data sources

name
Input index X1: employment Ten thousand people Statistical Yearbook
numbera of Sichuan Province
X2: total investment Billion yuan China City
in fixed assetsb Statistical Yearbook
X3: industrial added Tons of standard
value of energy coal/million
consumption
Output index Y1: total industrial Billion yuan
output value
Y2: total discharge of Ton
industrial waste
pollutionc
Notes Based on data availability. a Employment number selects the number of employed
persons at the year-end in the mining, manufacturing and electricity, gas and water
production and supply industries. b Total investment in fixed assets selects gross fixed
assets of industrial enterprises above-scale on behalf of existing production and oper-
ation scale and technical level of industrial enterprises. c Because the comprehensive
utilization rate of industrial solid waste in cities are all high and have no remarkable
difference, so the output indicators mainly selects industrial emissions of waste water,
waste gas (SO2 ) and smoke (powder) dust.
discharge of industrial waste pollution (Y2) respectively reflects the industrial

output and pollution emissions of each cities. In order to eliminate the influ-
ence of inflation, the total industrial output value of current year was converted
to constant price in 2005, according to the ex-factory price index of industrial
products. And the index unit of total discharge of industrial waste pollution was
processed.
Each evaluation index data are from statistical yearbook of Sichuan and
City Statistical Yearbook during 2005 to 2014. Some missing data has been
processed accordingly to ensure the accuracy of indicators. Due to a large number
of missing data in Ganzi, Aba, Liangshan area, this paper takes other 18 cities
in Sichuan province as the research object.
Among them, total discharge of industrial waste pollution uses entropy weight
method to calculate the index weight, and take reciprocal at last as a bad output.
The process of calculating as follows:
(1) Standardizing the industrial emissions of waste water, waste gas (SO2 ) and
smoke (powder) dust. Xij Indicates J index values in I area:
(Xj max − Xij )

rij = .
(Xj max − Xj min )
(2) Calculating the entropy of each index, m is the number of evaluation units,
n indicates the number of indicators:
m
1 r rij rij rij
Hi = − mij × ln m while m = 0, ln m = 0.
lnm i=1 i=1 rij i=1 rij i=1 rij i=1 rij
(3) Obtaining the entropy weight of every index. To keep the longitudinal com-
parability of the dynamic evaluation results, according to the method of
Zhang and Zhou [23], the presented paper takes the average entropy weight
as the final entropy weight to avoid the differences in each year:
1 − Hj
Wj = n .
n − j=1 Hj
(4) Calculating the total discharge of industrial waste pollution (abbreviation

to TIW):
n
T IW = rij wj .
j=1
5 Result
5.1 Overall Analysis of Industrial Green Efficiency in Sichuan
The paper using DEAP 2.1 software to calculate the Malmquist index of indus-
trial green efficiency of Sichuan province and its decomposition during 2005–2014
as shown in Table 2. It can be seen that the level of Sichuan’s industrial green
efficiency is low, but the development trend is increasing year by year. This
is consistent with Jia [7] results, which the Sichuan’s green industrial develop-
ment rate was still lagging behind, energy efficiency has gradually increased.
But the energy pressure is still large, environmental pollution has been curbed
but the situation is still grim. From the index decomposition terms, the impe-
tus for growth of industrial green efficiency comes from technical efficiency, and
its average annual growth rate is 2.1%. The improvement of technical efficiency
is mainly due to pure technical efficiency and scale efficiency, which increased
by 1.6% and 0.6% respectively. But the technical progress efficiency is low, far
below the effective value of 1. It indicates that it should increase the R&D,
introduction, digestion and absorption of green technology, to drive the green
development of Sichuan industry. One of the highest levels of Sichuan indus-
trial green development is in 2013–2014 with an increase of 16.6%, and mainly
depends on the improvement of technological progress index. This may be due
to the governments and enterprises had paid more attention to the green devel-
opment of industry in recent years, as well as the lag effect of green technology
has begun to produce.
Table 2. Average growth rate per annum and decomposition of TFP in Sichuan
province (2005–2014)
Year Effch Techch Pech Sech Tfpch

2008–2009 1.071 0.779 1.046 1.024 0.835
2009–2010 0.953 0.804 1.001 0.952 0.767
2010–2011 1.083 0.759 1.001 1.082 0.822
2011–2012 0.963 0.973 1.001 0.962 0.937
2012–2013 1.06 0.845 1.013 1.047 0.896
2013–2014 0.994 1.173 1.016 0.978 1.166
Mean 1.021 0.887 1.016 1.006 0.906
5.2 Analysis of Industrial Green Efficiency Among the Cities of

Sichuan Province
According to the year data from 2005 to 2014, we further study the Malmquist
index of industry green efficient and its decomposition terms item of Sichuan
province; the results are shown in Table 3. It is easy inferred that there is big
difference between different regions from the table date. The industry green effi-
cient of Chengdu, Deyang, Mianyang, Yibin and Dazhou has increased, which
accounts for 27.8%, Chengdu is 6% with maximal growth rate. The industry
green efficient growth rate of rest of 13 regions is negative, Guangyuan is −20.5%
with the lowest growth rate. This indicates that the industry prosperous of the
latter is at the cost of taking up and consuming a large number of ecologi-

cal resources. And the former is the beneficiary of development experience and
mature technology learning.
Table 3. Average growth rate per annum and decomposition of TFP in cities (2005–
2014)
Region Effch Techch Pech Sech Tfpch

Chengdu Plain Chengdu 1.054 1.006 1 1.054 1.06
Deyang 1.055 0.974 1.081 0.976 1.028
Mianyang 1.02 0.98 1.02 1 1
Suining 1.025 0.864 1.056 0.971 0.886
Leshan 1.055 0.829 1.011 1.044 0.875
Meishan 1 0.822 1 1 0.822
Yaan 1.042 0.843 1.054 0.989 0.879
Ziyang 0.954 0.869 0.943 1.012 0.829
South Sichuan Guangyuan 0.986 0.806 0.988 0.998 0.795
Nanchong 1.064 0.868 1.056 1.008 0.924
Guangan 1.009 0.842 1 1.009 0.85
Dazhou 0.998 1.002 1 0.998 1
Bazhong 0.987 0.807 0.987 1 0.797
Northeast Sichuan Zigong 1.028 0.874 1.03 0.998 0.898
Luzhou 1.005 0.867 1.003 1.002 0.871
Neijiang 0.965 0.846 0.988 0.977 0.817
Yibin 1.037 0.968 1.037 1 1.004
Panxi Panzhihua 1.1 0.894 1.03 1.068 0.983
Sichuan Mean 1.021 0.887 1.016 1.006 0.906
Note: Because of the missing statistical data of Liangshan state, Ganzi state
and Aba state, this paper does not studied the West Sichuan Economic
Zone. And considered the large industrial scale of Panzhihua, it was chosen
to represent the overall level of Panxi Economic Zone.
Combined the Table 3 and Fig. 3, we analysis deeply the dynamic change
trend and the distribute state of the Malmquist index resolution item in Sichuan
province. Figure 3 reflects the growth pattern and characteristic of the industrial
green efficiency. The cross-point of the horizontal-vertical coordinate is the equal-
ization point of the technology efficient index and technology progress index, by
which the 18 prefecture-level cities of Sichuan province are divided into 4 quar-
tiles. In the first quartile, the Effch is higher and Techch is lower than the average.
It indicates that these regions with lower technology contribution and simplex
elements driven mode. The Effch and Techch are higher than equalization in the
second quartile, these regions with reasonable element input scale, high tech-
nology import and absorption contribution rate. It promotes the growth of the
Techch
1.03 III Da zhou Chengdu II
Mia nya ng
0.98 Yibin Deya ng
0.93
Pa nzhihua
Ziya ng LuzhouZigong Na nchong
0.88 Suining
Neijia ng Gua ng'a n Ya ‘a n
Meisha n Lesha n
0.83 Gua ngyua n
Ba zhong
IV I
0.78
Effch
0.90 0.95 1.00 1.05 1.10 1.15
Fig. 3. The Techch and Effch distribution of the industrial green efficiency of 18
prefecture-level cities of Sichuan province
industrial green efficiency. The Techch is higher than Effch in the third quartile,
it manifests that the technology level of these regions is above of production fron-
tier. But it needs to improve the scale efficiency and pure technology efficiency.
The Effch and Techch are lower than province equalization in the fourth quartile,
these regions with low industrial green level and weak sustainable development
due to the technical constraints and scale bottlenecks.
In the meantime, from the four major economic regions data of Chengdu
Plain of Sichuan, Northeast Sichuan, South Sichuan, and Panxi (the presented
paper have not studied the economic region of northwest of Sichuan due to the
large missing data), it indicates that the industrial green efficiency is a gradual
reduce step distribute from north to south, and from west to east, as show in
Table 4.
The industrial green efficiency of Panxi head the list, secondly are Chengdu
Plain of Sichuan and South Sichuan, the worst is Northeast Sichuan, its year
equalization reduce is 12.7%. The technology progress indexes are inefficiency
state of the four economic regions, from Malmquist index and its resolution
items. Besides the south of Sichuan, the technology efficiency and its resolution
items of the other three economic regions are progressing, which reveals that
its resource allocation ability, resource utilization rate, and scale efficiency are
valid.
The driving force of technical efficiency in Chengdu Plain Economic Zone
is mainly pushed forward by pure technical efficiency. This also shows that the
radiation driven role of Chengdu, as the capital of Sichuan Province, has signif-
icant advantages in terms of production factors, management level and system
construction. The improvement of technical efficiency of Panxi Economic Zone
mainly originates in its advantage of scale efficiency (an increase of 6.8%). It also
shows that the resource input and its utilization efficiency have achieved scale
benefit. Conversely, the scale efficiency of South Sichuan Economic Zone has
decreased by 0.6%, which may be related to the unreasonable input structure of
production factors.
Table 4. Average growth rate per annum and decomposition of TFP in economic zone
(2005–2014)
Region Effch Techch Pech Sech Tfpch

Chengdu Plain Economic Zone 1.022 0.883 1.023 1 0.903
Northeast Sichuan Economic Zone 1.009 0.865 1.006 1.003 0.873
South Sichuan Economic Zone 1.009 0.889 1.015 0.994 0.898
Panxi Economic Zone 1.1 0.894 1.03 1.068 0.983
6 Conclusions and Suggestion
Combined with the existing problems and deficiencies in the process of industrial
green development in Sichuan province, there are several measures could be
taken.
(1) Adjusting the industrial structure within industry.
To eliminate backward production capacity with high energy consumption
and low added value, and reduce the industries proportion of high energy con-
sumption and high pollution [8], such as steel, cement and others. Developing
renewable energy sources actively, like wind, solar, bio-energy and others to
improve coal-based industrial structure. Besides, it can expand the scale of envi-
ronmental protection industry, for instance, solid waste treatment industry and
sewage treatment industry.
(2) Promoting technological innovation.
It is beneficial to reform and upgrade traditional industries, to strengthen
technological transformation, thus promoting the development of green recy-
cling economy. Positively implementing environmentally friendly technologies
and industrial process, such as industrial energy and water saving, clean produc-
tion, recycling and others, to enhance green technology capability [12]. Encour-
aging university–industry collaboration to promote the industrialization of green
technology. Protecting intellectual property is to enhance the enthusiasm of
enterprises to develop green technology, and solve the market barriers. The gov-
ernment can set up innovation platform with the opportunities of the eastern
industry transfer to the west, to enhance the ability of independent innovation.
(3) Adjusting scale structure to improve industrial green efficiency.
Using “incentive” environmental regulation tools based on market tools,
and establish market based emission regulations and industry dynamics to pro-
mote industrial scale efficiency [5]. Increasing the environmental surveillance of
medium-sized enterprises and implementing environmental measures with pow-
erful binding. In the aspects of the industrial development policy, it’s should
play an exemplary role of leading enterprise, encourage the joint development of
small and medium enterprises, enhance the linkage with large enterprises in the
upstream and downstream industry chain. In addition, establishing green net-
work supply chain to reduce the damage to ecological environment and achieve
the goal of improving industrial green efficiency.
(4) Accelerating the construction of green manufacturing system.

Firstly is to improve “green performance” in product design, manufacturing,
transportation, application and other product life cycle. Secondly is to create
and promote green factories and green areas, with high technology content, high
resource utilization and low environmental pollution. Thirdly is to establish green
manufacturing standards and clean energy policies [20], which can evaluate the
resource utilization and the cost of social environmental, thus to constraint and
guide industrial green development.
The presented paper studies the industrial green efficiency in Sichuan
province by DEA-Malmquist index method. The study results indicate that the
overall level of industrial green efficiency in Sichuan province is not high, but
with a gradually increase trend in recent years. The reason is mainly caused by
technology efficiency, but the technology progress index retardants the promote
effect of technology efficiency. In the four major economic regions, the industrial
green efficiency of Panxi heads the list, next are Chengdu Plain of Sichuan and
South Sichuan, the worst is Northeast Sichuan.
Acknowledgements. This work was supported by the Key Program (Grant No.
16AGL003) of the National Social Science Foundation of China. And we thank the sup-
port of the Social Science Planning Annual Project of Sichuan Province (SC16C011),
Soft science research plan project of Sichuan province (2017ZR0098), Soft science
research project of Chengdu city (2015-RK00-00156-ZF), The Central University Basic
Scientific Research Business Expenses Project of Sichuan University (skqy201623).
References
1. Chen C, Han J, Fan P (2016) Measuring the level of industrial green development
and exploring its influencing factors: empirical evidence from China’s 30 provinces.
Sustainability 8(2):153
2. Eiadat Y, Kelly A et al (2008) Green and competitive? an empirical test of the
mediating role of environmental innovation strategy. J World Bus 43(2):131–145
3. Fare R, Grosskopf S, Norris M (1994) Productivity growth, technical progress, and
efficiency change in industrialized countries: reply. Am Econ Rev 84(5):1040–1044
4. Fleiter T, Fehrenbach D et al (2012) Energy efficiency in the german pulp and
paper industry-a model-based assessment of saving potentials. Energy 40(1):84–99
5. Fowlie M, Reguant M, Ryan S (2016) Market-based emissions regulation and indus-
try dynamics. J Polit Econ 124:249–302
6. Honma S, Hu JL (2014) A panel data parametric frontier technique for measuring
total-factor energy efficiency: an application to japanese regions. Energy 78:732–
739
7. Jia RM (2011) The ecological character of evolution in sichuan province and reg-
ulatory strategy. Ind Technol Econ 8:98–104
8. Mei S, Wang LJ, Li WB (2009) Efficiency assessment of coal industry sustainable
development by improved data envelopment analysis. In: International conference
on management science and engineering, pp 1830–1835
9. Mukherjee R, Sengupta D, Sikdar SK (2015a) Sustainability in the context of
process engineering. Clean Technol Environ Policy 17(4):833–840
10. Mukherjee R, Sengupta D, Sikdar SK (2015b) Sustainability in the context of

process engineering. Clean Technol Environ Policy 17(4):833–840
11. Nagesha N (2008) Role of energy efficiency in sustainable development of small-
scale industry clusters: an empirical study. Energy Sustain Dev 12(3):34–39
12. Ng R, Low JSC, Song B (2015) Integrating and implementing lean and green
practices based on proposition of carbon-value efficiency metric. J Clean Prod
95:242–255
13. Su L, Zheng H, Wang Y (2013) Evaluation on green development of China’s provin-
cial industry. Resour Environ (8):116–122
14. Tu Z, Xiao G (2009) Research on China’s industrial growth model under environ-
mental constraints. J World Econ 11:41–54
15. Vasauskaite J, Streimikiene D (2016) Energy efficiency development in lithuanian
furniture industry. Adv Manag Sci 4:9–15
16. Wang C, Huang XJ (2006) On the evaluation and diagnosis of obstacle for circular
industrial economy in Jiangsu Province. Ind Eng Manag 112(2):301–315
17. Wu J (2009) Tfp growth and convergence across China’s industrial economy con-
sidering environmental protection. J Quant Tech Econ 11:17–27
18. Xuxiao WU (2016) Research on the improving path and dynamic evaluation of
regional industrial green development efficiency: Taking the heavy chemical indus-
try zone of Qinghai, Henan and Fujian as examples. Ecol Econ 2:63–68
19. Yang Y (2012) Research on evaluation of sichuan low-carbon economy efficiency.
China Popul Resour Environ 6:52–56
20. Yi H, Liu Y (2015) Green economy in China: regional variations and policy drivers.
Glob Environ Change 31:11–19
21. Zhang C, Liu H (2014) The impact of unbalanced regional development on indus-
trial green total factor productivity: a new perspective based on three stages dea
adjustment efficiency. Contemp Econ Res 9:39–45
22. Zhang J, Zhu L (2012) Research on technological innovation efficiency of industrial
enterprises based on green growth of regions in China. J Quant Tech Econ 2:113–
125
23. Zhang M, Zhou ZF (2010) Research on dynamic evaluation on the independent
innovation capacity of high-tech industries in China. Soft Sci 95:242–255
Risk Control
Machine Learning and Neural Network
for Maintenance Management
Alfredo Arcos Jiménez(B) , Carlos Quiterio Gómez Muñoz,

and Fausto Pedro Garcı́a Márquez
Ingenium Research Group, Castilla-La Mancha University, Ciudad Real, Spain

Alfredo.Arcos@alu.uclm.es
Abstract. The paper presents a novel approach that allows to optimize

the ultrasonic wave sensors for a condition monitoring system employing.
It can detect and diagnosis different faults with a signal, such as delam-
ination, mud or ice on blades of wind turbines. This methodology allows
to avoid the redundancy of sensors, since a specific number of ultrasonic
transducers can determine the structural condition using guided waves.
The signal is pre-processed with the aim of removing the noise, then
extracted and selected features to be later classified by Machine Learn-
ing and Neural Networks. Finally, for each damage or anomaly, the best
classifier will be evaluated. The best classifier of each damage will act on
a parallel network that will process the signal sent by the sensor.
Keywords: Non-destructive testing · Fault detection and diagnosis ·

Condition monitoring system · Wavelet transforms · Machine learning ·
Neuronal network
1 Introduction
The objective in the next decades is to obtain a competitive and sustainable energy
to all countries, allowing to reduce dependence on fuels to households, industries
and transportation. Wind power and Concentrated Solar Power (CSP) are being
two of the main renewable energy sources. Its importance in the energy market is
being essential, and the forecasts show it will continue in the near future.
Renewable energy industry requires significant improvements in reliability,
lifetime or availability that it is done by an efficient maintenance based on Con-
dition Monitoring Systems (CMS). CMS is the process of determining the condi-
tion of system [37,38,49], where one of the main proposes is to identify changes
between two main states of the structure, damaged and undamaged [36,39,41].
They will provide different patterns that will be used in order to analyze the
condition of the system [19].
A complex CMS contains numerous sensors that generates different type of
signals with information and sampling frequencies [41]. It needs to be processed,
preferable on-line, with a minimum computational cost and with high accuracy
to reduce false alarms [39].
1378 A.A. Jiménez et al.
This paper present a novel approach to reduce the number of sensors, max-
imizing the signal processing analysis to detect different damages. The guided
wave ultrasonic signals are employed for fault detection and classification.
Structural Health Monitoring (SHM) is a technology that combines advanced
CMS, together with signal processing, to determine the condition of the struc-
tures on line or not [19,36,41]. SHM leads to increase the Reliability, Availability,
Maintainability and Safety (RAMS) of the system [16,41]. SHM allows also to
know the different levels of the defect severity. It will be useful for an optimal
maintenance management to reduce costs and increase the profitability.
Farrar and Sohn [16] considered pattern recognition in SHM. This method-
ology consists in four stages shown in Fig. 1.
Fig. 1. Methodology SHM
The problems caused by the implementation of damage identification system

is assessed in operational evaluation phase. Data acquisition and normalization
involves the choice of sensors and their location, the normalization procedure
of the data collected and the selection of the data to be used in the feature
selection process. Dimensionality reduction techniques have been developed to
remove irrelevant data (noise) and redundant features, and categorized mainly
into Feature Extraction (FE) and Feature Selection (FS). FE contains infor-
mation of the signal that distinguishes between a damaged or non-damaged
structures [57]. FS selects the data that are better for the detection of damage
and executes a condensation of the data [16,40]. Finally, classifiers, based on
statistical models, calculate and cluster the dataset depending on the damage.
These classifiers are grouped as: Supervised Learning (SL) and Unsupervised
Learning [35].
This paper proposed patterns recognition of each damage by Machine Learn-
ing and Neural Network. Then approach minimize the number of sensors, where
through an ultrasonic signal emitted by a sensor is able to classify different states
of different damages or anomalies.
This process is complex because each damage must be analyzed individually.
On the other hand, there is no generic classifier that provides the best results.
Even the same classifier for the same signal depends on the methodology of FE
and its subsequent FS.
2 Approach
Figure 2 shows the schematic approach for determining the level of damage or
anomaly based on SHM employing ultrasonic guided waves.
Machine Learning and Neural Network for Maintenance Management 1379
Fig. 2. Methodology schedule for determining the level of damage or anomaly
2.1 Signal Pre-Processing
Sensors generate noise from a variety of sources and therefore a signal pre-
processing is necessary to eliminate/reduce the dataset that does not provide
useful informatio [45]. Standard statistical techniques have been employed in
references [10,53]. Yu et al. [64] were able to reduce global noise using averag-
ing techniques and Daubechies Wavelet (DW) to eliminate local high-frequency
perturbations. Denoising and compression signal in Guided Waves (GW) based
on the Discrete Wavelet Transform (DWT) was employed by Rizzo and Scalea
[55]. There are a large number of research and reviews articles on filtering in the
treatment of ultrasonic guided waves [46]. Hamming [21] performs a review about
low-pass filters available for data smoothing. This paper will consider Wavelet
transforms for filtering the signals.
The denoising of the signal is performed employing a multilevel 1-D wavelet
analysis using Daubechies family. The wavelet decomposition structure of the
signal is extracted. The threshold for the de-noising is obtained by a wavelet
coefficients selection rule using a penalization method provided by Birgé-Massart
[47]. An overly aggressive filtering could eliminate data that should show, for
example, small echoes that come from defects. Figure 3 shows the original signal
and the de-noised signal when it is applied the wavelet de-noised filter. The
Wavelet de-noising filter does not produce an unwanted signal delay in contrast
to other digital filters.
It is observed that the filter removes noise significantly, and does not eliminate
information that is related to different structural features.
2.2 FE Methods
FE affects the learning and classification process. FE and FS are capable of

improving learning performance, presenting less computational complexity and
historical dataset [58].
Classical dimensional reduction techniques such as Principal Component
Analysis (PCA) [26] and Multidimensional Scale (MDS) [29] are extensively
employed. However, the Linear and Nonlinear Regression systems are used when
the signal type is complex and the first order interactions are not enough to derive
good results [20].
Two autoregressive models have been applied in this paper for FE: Linear
(AR) and nonlinear (NARX). AR is a lineal mathematical model commonly
employed for FE because of its high computational performance. AR method
-3
Noisy signal
x 10
2
Voltage
0
-2
200 400 600 800 1000 1200 1400 1600 1800 2000
Denoised signal
-3
x 10
Voltage
1
0
-1
200 400 600 800 1000 1200 1400 1600 1800
Residual noise
-5
x 10
5
Voltage
-5
200 400 600 800 1000 1200 1400 1600 1800 2000
Time (samples)
Fig. 3. Decomposition detail five (D5), De-noised D5 and extracted residual noise
is popular because estimation of AR parameters (features) is achieved easily

by solving linear Yule-Walker equations [13]. NARX presents a better behavior
in the most of the nonlinear systems. NARX was introduced by Leontaritis and
Billings [30] and it has been widely used in the identification of nonlinear dynamic
models and feature selection [6]. NARX presents a good generalization capability,
as effective training algorithm, a fast convergence rate, and can represent a wide
class on nonlinear systems [3,18,25].
Within the NARX model, the lineal-in-the-parameters structure is a method
widely used in nonlinear dynamics models. It is a linear combination of model
terms, or basis functions, that are some nonlinear-usually polynomial- functions
of the system variables [32]. In this case, the monomials of the linear combi-
nation are the features. The linear-in-the-parameter models are well structured
to adaptive learning, with learning conditions and demonstrable convergence
and capacity parallel processing, in addition to numbers applications in signal
processing and pattern recognition [24,36].
2.3 FS Methods
FS techniques can be categorized into three approaches: embedded, wrapper or

filter [33]. The main methods for FS are Information Gain, Relief, Fisher Score
and Lasso [27]. In this paper, the main objective to reduce dimensionality and
it is not intended to select the best features.
SHM literature propose several methods to select the AR model order: AIC
[2]; Akaike’s Final Prediction Error [1] (FPE); Partial Autocorrelation Function
(PAF) [60], and; Root Mean Square (RMS) [31]. AIC statistical technique has
been used to reduce the dimensionality of the feature extraction. The AIC is
one of the most efficient techniques for order optimization [35]. The AIC is a
measure of the goodness-of-fit of an estimated statistical model, based on the
trade-off between fitting accuracy and the number of estimated parameters.
2.4 Classifiers
(1) Machine Learning Approach
1 Decision Trees
Decision Tree (DT)is a classifier used in many fields to study if the data contains
different classes of objects that can be interpreted significantly in the context of
a substantive theory [34,48,51,52].
DT generates a split of space from a labeled training set. The objective
is to separate the elements of each class into different labeled regions (leaves)
minimizing the local error. Each internal node in the tree is a question (decision)
that determines which branch of the tree must be taken to reach a leaf. DT is
determined for; (1) how to split the space (Splitting Rules); (2) stop condition
of splitting; (3) labeling function of a region, and; (4) measurement of error.
The purpose of the Splitting Rule is to minimize the impurity of the node. The
recursive splitting algorithm stops when it finds any of the following conditions:
the node is pre-set maximum deep; all elements of the node are same class; there
is no empty sub node or; SR does not reach a pre-set value.
To label a leaf or region once it is already considered as a terminal is consid-
ered to develop a DT. Equation (1) establishes the labeling function.
k

l = arg min l NI × cl,l , (1)
l=1
where NI is the number of elements of class l in the region, l is the class to label
and c( l, l ) is the labeling cost.
c( l, l ) with all classes is calculated and l is selected, which minimizes the
error. The label that minimizes the error of a region is the most populated class.
In case of a tie, a random one is chosen. Equation (2) gives the classification
average error
N
1
ε= err(l, li ), (2)
N
l=1
where err(l, li ) is the error of labeling a class l as l . This error is solved by
splitting the space and assigned to each split a label.
2 Discriminant Analysis (DA)
The ultrasonic signals considered are homogeneous for the same frequency and,
therefore, the classifier that provides better results is LDA [15]. This classifier is
valid for ultrasonic signals obtained at different frequencies and used for speech
recognition. This type of classifier is based on a geometrical approach. A linear
function representing a decision limit divides a feature space into regions that
have common properties. The aim of this method is to find the expression of lin-
ear discriminant functions that enable the objects classification in the considered
classes.
3 Support Vector Machine (SVM)
SVM [9] is a supervised multivariate classification method. “Supervised” refers to
a training step where the algorithm learns the differences between pre-specified
groups to be classified [59]. SVM treats each feature as a point in a high-
dimensional space, being the number of dimensions the same to the number
of rating levels. Each feature is assigned to a group and the linear classification
function (3) learns the characteristics to discriminate among five groups. A limit
of decision, or hyperplane (a generalization of a plane of n − 1 dimensions which
divides an n-dimensional space), must be defined to separate the data based on
class membership and to classify linearly the dataset. However, for a linearly
separable problem, there are an infinite number hyperplanes correctly classified
data. SVM algorithm finds the optimal one characterized by the largest margin
between classes. The margin is defined as the distance of the closest training
data points of the hyperplane. These points are the most difficult to classify and
they are called support vectors. The hyperplane is defined by a weight vector,
which is a linear combination of the support vectors, and specifies both a direc-
tion and a displacement which together define the maximum margin classifier.
The decision function D(x) is given by Eq. (3),
D(x) = w∅(x) + b, (3)
where w and b are the SVM parameters, and ∅(x) is a kernel function
The hyperplane is defined by Eq. (3), and the distance between the hyper-
plane and pattern x can be written by Eq. (4).
D(x)
. (4)
w
A training classifier is designed to find the value of w that maximizes the

margin between the class boundary and the training patterns. The objective
function of the training algorithm is given by Eq. (5).
2
w
J= . (5)
2
In this case, the linear function Kernel and one vs one method have been
employed [43].
4 k-Nearest Neighbors
k-Nearest Neighbors(k-NN) [11,12] is a high-performance classifier widely used in
Machine Learning [42]. k-NN rule classification is an extension of the Nearest-
Neighbor (NN) rule. Given a set x of the k samples, the rule assign to each
sample is to label most frequently represented among the k nearest samples [15].
k-NN search technique and k-NN based algorithms are widely used as benchmark
learning rules. The relative simplicity of the k-NN search technique makes it easy
to compare the results from other classification techniques to k-NN results.
The accuracy of k-NN classification significantly depends on the metric used
to compute distances between different samples [61]. In most cases, the best
performing classifier is Fine k-NN [62], using metric Euclidean distance in Eq. (6).
d2st = (xs − yt ) · (xs − yt ) (6)
5 Ensemble Subspace Discriminant (ESD)

Ensemble methods are learning algorithms that construct a set of classifiers
whose individual decisions are combined in some way (generally by weighted or
unweighted voting) to classify the set of features of each pattern [14]. Bagging,
boosting, and random subspaces are general techniques that can be used with
any type of base classifier. Ensemble Bagged Tree (EBT) and Random Subspace
Method (RSM) have been the most successful methods. EBT uses the Breiman’s
‘random forest’ algorithm [7]. RSM is a parallel learning algorithm proposed by
Ho [23].
(2) Artificial Neuronal Network (ANN)
ANN unidirectional supervised through a MLP with training by backpropagation
algorithm [56] has been applied. Backpropagation with algorithm scaled conju-
gate gradient and performance Cross Entropy [28,44] with ‘Early Stopping’ to
avoid overfitting [50] has been training mode. ANN is given by Eq. (7),

zk = wkj yj − θi = wkj f wji xi − θj − θj , (7)
j j i
where xi is ANN input, yi is hidden layer output, zi is final layer output, tk is

targets output, wj i is hidden layer weight, wkj is final layer weight, θj is hidden
layer bias, θk is final layer bias and f (.) is the activation function of sigmoid
type, employed as the activation function of the ANN.
The MLP process tests initially with an ANN architecture and is trained with
70% of the total of the experiments and 30% of them to test the network. Then
the ANN architecture is chosen according to the accuracy and performance.
Finally, ANN is tested with different cases (30%) to know if the learning is right
and to the accuracy.
Backpropagation (BP) is one of the simplest and most general methods for
supervised training of multi-layer ANNs. Scaled, standardization, normalization
that perform pre-processing inputs techniques have been employed to accelerate
the BP training. Scaled conjugate gradient and performance Cross Entropy are
employed to increase the computational learning cost of ANN.
2.5 Evaluation Classifier

Analysis Receiver Operating Characteristic (ROC), as Confusion Matrix (CM),
is employed to evaluate of the classification. CM determines the accuracy of a
classifier and measures its performance. The main parameters in CM are:
True Positive (TP) : corresponds to real successes of the classifier

False Positives(FP) : is the sum of the values of a class in the
corresponding CM column, excluding the TP
False Negative(FN) : is the sum of the values of a class in the
corresponding CM row, excluding the TP
True Negatives(TN) : is the sum of all columns and rows, excluding that
class’s column and row
Recall function, R, given by Eq. (8), provides the probability to be correctly
classified, called true positive rate or hit rate:
TP
R= . (8)
TP + FN
Specificity, S, also called the true negative rate, measures the proportion of
negatives that are correctly identified as negatives.
TN
S= . (9)
FP + FN
Precision, P :
TP
P = . (10)
FP + TP
F-score, F [54]
2×P ×R
F = . (11)
P +R
There are two conventional methods to establish the average performance in
all categories: macro-averaging and micro-averaging [63]:
Macro-average: P M , RM and F M is obtained are the mean of all PiM , RiM
and FiM where M denotes Macro-average, and i is the
scenario, then for each category they are calculated, i.e.
they are valued are evaluated locally PiM , RiM and FiM ,
and then globally P M , RM and F M .
Micro-average: P M , RM and F M value is obtained as: i) T Pi , F Pi , F Ni
values is calculated for each of the scenario; ii) the value of
T P , F P and F N are calculated as the sum of T Pi , F Pi
and F Ni , and; iii) to applying the equation of the measure
that correspond.
There are several indices extracted from the ROC curve to evaluate the
efficiency of a classifier. Area Under Curve (AUC) represents the area between
the ROC curve and the negative diagonal [5,22], being between 0.5 and 1. Values
≤ 0.5 indicates that the classifier is invalid and a value of 1 indicates a perfect
rating because there is a region in which, for any point cut, the values of R and
Pare equal to unit. The statistical property of AUC is equivalent to the Wilcoxon
test of ranks [22]. The AUC is also related to the Gini coefficient [8], which is
twice the area between the diagonal and the ROC curve.
The recommendations by Demsar [4], and the extensions by Garcia and Her-
rera [17], have been employed to perform the comparative analysis of classifiers.
Friedman Test will be used to test the null hypothesis that all classifiers achieve
the same average. Bonferroni-Dunn Test is applied to determine significant dif-
ferences between the top-ranked classifier and the following. Holm Test is used
to contrast the results.
3 Approach Scheme
The methodological process is represented schematically in Fig. 4. Firstly, for
each damage or anomaly, the best classifier is selected.
Classifier
Feature (Machine
Data
Acquisition Extraction and Learning and Classifiers Decision
and Filtering Feature Neural Evaluation Making
Selecction Network)
Fig. 4. Methodology flow for determining the classifier
The methodological process will follow the scheme given in Fig. 5 when the
best classifier for each damage is set.
Classifier
Damage-1
Classifier
Damage-2
Data Acquisition Feature
and Filtering Extraction Feature Selection
⋮
Classifier
Damage-n
Fig. 5. Methodology flow for determining the damage
4 Conclusions
The paper presents a novel approach to optimize the sensors in a condition
monitoring system employing ultrasonic waves. The approach can detect dif-
ferent potential faults with a single signal emitted by a sensor, such as delam-
ination, mud or ice on blades of wind turbines. This methodology allows to
avoid the redundancy of sensors, since a specific number of ultrasonic trans-

ducers can determine the structural condition using guided waves. The signal is
pre-processed with the aim of removing the noise, then extracted and selected
features to be then classified by Machine Learning and Neural Networks. Finally,
for each damage or anomaly, the best classifier will be evaluated. The best clas-
sifier of each damage will work in a parallel network that will process the signal
sent by the sensor.
Acknowledgements. The work reported herewith has been financially supported by

the Spanish Ministerio de Economı́a y Competitividad, under Research Grant DPI2015-
67264-P and RTC-2016-5694-3.
References
1. Akaike H (1969) Fitting autoregressive models for prediction. Ann Inst Stat Math
21(1):243–247
2. Akaike H (1974) A new look at the statistical model identification. IEEE Trans
Autom Control 19(6):716–723
3. Alazrai R, Lee CSG (2012) An narx-based approach for human emotion identifi-
cation. In: Ieee/rsj international conference on intelligent robots and systems, pp
4571–4576
4. Ar J (2006) Statistical comparisons of classifiers over multiple data sets. J Mach
Learn Res 7(1):1–30
5. Bradley AP (1997) The use of the area under the roc curve in the evaluation of
machine learning algorithms. Pattern Recognit 30(7):1145–1159
6. Brankovic A, Falsone A et al (2016) Randomised algorithm for feature selection
and classification. Xiv preprint
7. Breiman L (2001) Random forests. Mach Learn 45(1):5–32
8. Breiman L, Friedman J, et al (1984) Classification and regression trees. CRC Press
9. Burges CJC (1998) A tutorial on support vector machines for pattern recognition.
Data Min Knowl Disc 2(2):121–167
10. Chopra I (2002) Review of state of art of smart structures and integrated systems.
AIAA J 40(11):2145–2187
11. Cover T, Hart P (1967) Nearest neighbor pattern classification. IEEE Trans Inf
Theory 13(1):21–27
12. Dasarathy BV (1990) Nearest neighbor (nn) norms: nn pattern classification tech-
niques. Los Alamitos IEEE Comput Soc Press 13(100):21–27
13. De Lautour OR, Omenzetter P (2010) Damage classification and estimation in
experimental structures using time series analysis and pattern recognition. Mech
Syst Signal Process 24(5):1556–1569
14. Dietterich TG (2000) Ensemble methods in machine learning. In: International
workshop on multiple classifier systems, pp 1–15
15. Duda RO, Hart PE, Stork DG (2012) Pattern classification. Wiley
16. Farrar CR, Doebling SW, Nix DA (2001) Vibration based structural damage iden-
tification. Philos Trans R Soc B Biol Sci 359(359):131–149
17. Garc S, Herrera F (2008) An extension on “statistical comparisons of classifiers over
multiple data sets” for all pairwise comparisons. J Mach Learn Res 9(12):2677–2694
18. Ghosh S, Maka S (2009) A narx modeling-based approach for evaluation of insulin
sensitivity. Biomed Signal Process Control 4(1):49–56
19. González-Carrato RRDLH, Márquez FPG, Dimlaye V (2015) Maintenance man-

agement of wind turbines structures via mfcs and wavelet transforms. Renew Sus-
tain Energy Rev 48:472–482
20. Guyon I, Elisseeff A (2006) An introduction to feature extraction. Springer,
Heidelberg
21. Hamming RW (1989) Digital filters, 3rd edn. Prentice Hall
22. Hanley JA, Mcneil BJ (1982) The meaning and use of the area under a receiver
operating characteristic (roc) curve. Radiol 143(1):29–36
23. Ho TK (1998) The random subspace method for constructing decision forests.
IEEE Trans Pattern Anal Mach Intell 20(8):832–844
24. Hong X, Mitchell R et al (2008) Model selection approaches for non-linear system
identification: a review. Int J Syst Sci 39:925–946
25. Jiang C, Song F (2010) Forecasting chaotic time series of exchange rate based on
nonlinear autoregressive model. In: Advanced Computer Control (ICACC), 2010
2nd international conference, pp 238–241
26. Jolliffe I (2002) Principal component analysis. Wiley Online Library
27. Jovic A, Brkic K, Bogunovic N (2015) A review of feature selection methods with
applications. In: International convention on information and communication tech-
nology, electronics and microelectronics, pp 1200–1205
28. Kroese DP, Rubinstein RY et al (2013) Cross-entropy method. In: Encyclopedia
of operations research and management science. Springer, pp 326–333
29. Kruskal JB, Wish M (1978) Multidimensional scaling. Book on Demand Pod
30. Leontaritis IJ, Billings SA (1985) Input-output parametric models for non-linear
systems part ii: stochastic non-linear systems. Int J Control 41(2):303–328
31. Levinson N (1949) The wiener (root mean square) error criterion in filter design
and prediction. Stud Appl Math 25(1–4):261–278
32. Li K, Peng JX, Bai EW (2006) A two-stage algorithm for identification of nonlinear
dynamic systems. Automatica 42(7):1189–1197
33. Liu H, Motoda H (2008) Less is more. Computational methods of feature selection.
In: H Liu, e Motoda, H (eds) Chapman and Hall/CRC, pp 3–17
34. Loh WY, Shih YS (1997) Split selection methods for classification trees. Stat Sinica
7(4):815–840
35. Lordo RA (2001) Learning from data: Concepts, theory, and methods. Technomet-
rics 43(1):105–106
36. Márquez FPG, Munoz JMC (2012) A pattern recognition and data analysis method
for maintenance management. Int J Syst Sci 43(6):1014–1028
37. Márquez FPG, Pedregal DJ, Roberts C (2013) New methods for the condition
monitoring of level crossings. Int J Syst Sci:878–884
38. Márquez FPG, Pardo IPG, Nieto MRM (2015) Competitiveness based on logistic
management: a real case study. Ann Oper Res 233(1):157–169
39. Márquez FPG, Pérez JMP et al (2016) Identification of critical components of
wind turbines using fta over the time. Renew Energy 87:869–883
40. Martinez-Luengo M, Kolios A, Wang L (2016) Structural health monitoring of off-
shore wind turbines: a review through the statistical pattern recognition paradigm.
41. Marugán AP, Márquez FPG, Pérez JMP (2016) Optimal maintenance management
of offshore wind farms. Energies 9(1):46
42. Michalski RS, Carbonell JG, Mitchell TM (2013) Machine learning: an artificial
intelligence approach. Springer Science and Business Media
43. Milgram J, Cheriet M, Sabourin R (2006) “one against one” or “one against all”:
which one is better for handwriting recognition with svms? In: Proceedings of
international workshop on frontiers in handwriting recognition
44. Møller MF (1993) A scaled conjugate gradient algorithm for fast supervised learn-
ing. Neural Netw 6(4):525–533
45. Muñoz JMC, Márquez FPG, Papaelias M (2013) Railroad inspection based on acfm
employing a non-uniform b-spline approach. Mech Syst Signal Process 40(2):605–
617
46. Munoz CG, Márquez FG (2016) A new fault location approach for acoustic emis-
47. Munoz CQG, Márquez FPG, Tomás JMS (2016) Ice detection using thermal
infrared radiometry on wind turbine blades. Measurement 93:157–163
48. Pal M, Mather PM (2003) An assessment of the effectiveness of decision tree meth-
ods for land cover classification. Remote Sens Environ 86(4):554–565
49. Papaelias M, Cheng L et al (2016) Inspection and structural health monitoring
techniques for concentrated solar power plants. Renew Energy 85:1178–1191
50. Prechelt L (1998) Automatic early stopping using cross validation: quantifying the
criteria. Neural Netw Off J Int Neural Netw Soc 11(4):761
51. Quinlan JR (1986) Induction of decision trees. Mach Learn 1(1):81–106
52. Quinlan JR (2014) C4. 5: programs for machine learning. Elsevier
53. Raghavan AC, Cesnik CES (2007) Review of guided-wave structural health moni-
toring. Shock Vib Dig 39(2):91–114
54. Rijsbergen CJV (1979) Information retrieval. Butterworth-Heinemann
55. Rizzo P, Scalea FLD (2004) Discrete wavelet transform to improve guided-wave-
based health monitoring of tendons and cables. Proc SPIE Int Soc Optical Eng
5391:523–532
56. Rumelhart DE, McClelland JL et al (1988) Parallel distributed processing, vol 1.
IEEE
57. Staszewski WJ, Robertson AN (2007) Time-frequency and time-scale analyses
for structural health monitoring. Philos Trans R Soc A Math Phys Eng Sci
365(1851):449
58. Tang J, Alelyani S, Liu H (2014) Feature selection for classification: a review. In:
Documentación Administrativa, pp 313–334
59. Vapnik VN (1999) An overview of statistical learning theory. IEEE Trans Neural
Netw 10(10):988–999
60. Wei WWS (1994) Time series analysis. Addison-Wesley publ Reading
61. Weinberger KQ, Saul LK (2009) Distance metric learning for large margin nearest
neighbor classification. J Mach Learn Res 10(1):207–244
62. Xu Y, Zhu Q et al (2013) Coarse to fine k nearest neighbor classifier. Pattern
Recognit Lett 34(9):980–986
63. Yang Y (1999) An evaluation of statistical approaches to text categorization. Inf
Retr J 1(1):69–90
64. Yu L, Bao J, Giurgiutiu V (2004) Signal processing techniques for damage detection
with piezoelectric wafer active sensors and embedded ultrasonic structural radar.
Proc SPIE Int Soc Optical Eng 5391:492–503
Volatility Spillover Between Foreign Exchange
Market and Stock Market in Bangladesh
Shibli Rubayat1(B) and Mohammad Tareq2

1
Faculty of Business Studies, University of Dhaka, Dhaka 1000, Bangladesh
shibli@du.ac.bd
2
Department of Accounting and Information Systems, University of Dhaka,
Dhaka 1000, Bangladesh
Abstract. This paper investigates the link between foreign exchange

market and stock market in Bangladesh through volatility spillover
between the markets. It also examines the volatility persistence and
asymmetric effect of information on the volatility of these two finan-
cial markets. Using data from January 1, 2009 to December 12, 2016
Taka/USD exchange rate’s volatility and the stock return’s volatility of
CSE General Index this paper utilizes ARCH and GARCH models to
investigate the spillover, persistence and asymmetry effect of volatility.
The results reveal high level of presence of volatility of CSE General
Index but not of the Taka/USD exchange rate. The result also finds
asymmetric effect of information i.e. bad news is followed by higher
volatility compared to that of after good news in the stock market. How-
ever, the result shows that volatility in the CSE General Index spillover
that of the Taka/USD exchange rate but not vice versa. The meaning
and significance of the study is also presented in the paper.
Keywords: Foreign exchange market · Stock market · Volatility ·

Spillover · Taka/USD (The local and United States Currency)
1 Background of the Study

Foreign exchange market and capital market are linked in many countries and
show inter-dependence [1,7,11]. Among others, volatility of these markets links
each other and spillover from one market to another. The volatility is a key
factor that causes risk in these markets and thus influences economic activities
in a country. Volatility information of these markets is used for risk management,
portfolio allocation and stock trading strategies [12].
Bangladesh, a south Asian country struggling for poverty reduction and eco-
nomic development, has a smaller and less equipped capital market than its
neighboring countries [10]. A developed capital market can aid a country’s eco-
nomic development by turning savings into investment. Researchers have con-
sistently found a positive relationship between a developed stock market with
economic growth and capital accumulation [8,9]. As mentioned above, knowledge
1390 S. Rubayat and M. Tareq
about the volatility help investors in their risk management, portfolio allocation
and stock trading strategies and is considered one of the ways of developing
a capital market. However, there is no notable study in Bangladesh that has
investigated the volatility of these financial markets, especially how these two
markets are linked through volatility.
The contribution of the research lies in the fact that no notable study has
been done so far that investigates the link between the volatility of the capital
market and foreign exchange markets in Bangladesh. It is hoped that this study
will inform policy-making decisions in relating to these two market and help
investors in making investment decision.
1.1 Methodology
The study uses CSE General Index of Chittagong Stock Exchange (the 2nd
largest in the port city) from January 1, 2009 to December 1, 2016. The official
exchange rate of Taka/USD of the same period was also used. These data were col-
lected from Chittagong Stock Exchange and from the central bank of Bangladesh,
‘Bangladesh Bank’ respectively. The Taka/USD exchange rates and the CSE Gen-
eral Index for the sample period are shown in Figs. 1 and 2 respectively.
The return of the stock market was calculated as follows:
Rt = ln (Pt /Pt−1 ) × 100.

85
85
80
80
USDSell
USDBuy
75
75
70
70
0 500 1000 1500 2000 0 500 1000 1500 2000

date day
Fig. 1. Taka/USD exchange rate (From January 1 2009 to December 12, 2016)
15000
CSE General
10000
5000
0 500 1000 1500 2000

days
Fig. 2. CSE General (From January 1, 2009 to December 12, 2016)

Volatility Spillover Between Foreign Exchange Market and Stock Market 1391
Daily spread between buy and sell rate of Taka/USD was used to measure
return on Taka/USD exchange rate. The Taka/USD exchange rates return and
the CSE General Index daily return is for the sample period are shown in Figs. 3
and 4 respectively.
1
.5
Return2
0
-.5
-1
0 500 1000 1500 2000

day
Fig. 3. USD/Taka exchange rate return (Dec 12, 16 to Jan 1 2009)

15
10
5
return x
0
-5
-10
0 500 1000 1500 2000

date
Fig. 4. CSE General Index stock return (Dec 12, 16 to Jan 1 2009)
To analyze volatility, asymmetric volatility and spillover effect, different

variation of the Generalized Auto Regressive Conditional Heteroskedasticity
(GARCH) were used in this study. Maximum Likelihood estimation method
was used to estimate the parameters of the GARCH models of this study. Aug-
mented Dickey–Fuller (ADF) unit root tests for the CSE General Index return
and Taka/USD exchange rate return series were done and the results reject the
hypothesis of unit root. GARCH is a model for time-series where volatility is clus-
tered together. The autoregressive term in a GARCH model takes into account
past values on the time-series. Heteroskedasticity means that data contains non-
constant variance, which by running a standard OLS would result in erroneous
standard errors and confidence intervals. GARCH models corrects these treat
heteroskedasticity as a variance inside the model.
The ARCH model was the forerunner to the GARCH specifications. Before
this, volatility used to be measured using rolling standard deviations over past
time periods. However, the problem with this approach is that each time period
receives equal weight in the estimation of the current period’s variance. The
ARCH process, proposed by Engle [5], solves this by letting these weights be
parameters to be estimated. This allows the data to determine the best weight
to be used on the present volatility [6]. An ARCH model is presented below:
r = π + εt ,

q
σt2 = c + αi ε2t−i ,
i
where εt = σt zt ; zt is an i.i.d random variable with mean zero and variance one
and σ 2 is the squared conditional variance in the model.
So the conditional variance is here described as a distributed lag of past
squared innovation [3]. That means that the variance of the current error term
is a function of the size of the previous period’s squared error term. In order to
avoid a very large number of coefficient in high order polynomial, Bollerslev [2]
developed the GARCH-model as a generalization of Engle’s [5] ARCH-model.
This makes a declining weight that never reaches zero [6]. GARCH (1, 1) is the
most common model to use in empirical research. A GARCH (1, 1) model is
presented below:
r = π + εt ,
σt2 = c + αε2t−1 + βσt−1
2
,
where, εt = σt zt ; zt is an i.i.d random variable with mean zero.
2 Persistence and Asymmetry of Volatility

The study utilize a GARCH (1, 1) model for the CSE General Index return and
Taka/USD exchange rate return. The result shows that the “α + β” estimated
from the CSE General Index is about .99. This means that volatility is highly
persistent for the CSE General Index. It means that if there is volatility in the
CSE stock market it decays slowly over a long period of time. However the study
did not find the persistence of volatility in the Taka/USD rate (See Table 1).
In the world of financial market the reality is that negative news often has
much bigger impact on volatility than good news has on relaxing these fluctua-
tions i.e. negative shocks at time t − 1 have a stronger impact on the variance
at time t than positive shocks. The GARCH model is symmetrical, negative and
positive shocks are given equal weight. The study also investigated the asym-
metry effect of positive news and negative news on the market volatility using
EGARCH-model. Result of the asymmetric GARCH shows that bad news causes
Volatility Spillover Between Foreign Exchange Market and Stock Market 1393
Table 1. Persistence and asymmetry of CSE General Index and Taka/USD exchange
rate
CSE General Index CSE General Index Taka/USD exchange rate

Con .005∗ (.027) .005∗ (.027) 0.00∗∗ (00)
α .16∗∗∗ (.014) .16∗∗∗ (.014) 1.23(1.15)
β .83∗∗∗ (.011) .83∗∗∗ (.011) .34∗∗∗ (.03)
γ (asymmetric term) - −.02∗ (.011) −.6(1.16)
n 1911 1911 1940
∗ p < .05; ∗∗ p < .01; ∗∗∗ p < .001
higher volatility (γ = −0.02∗ ).compared to good news in the stock markets in

Bangladesh (See Table 1).
Past studies have shown that volatility in one market can affect other mar-
kets [4,11]. The study also investigates whether the volatility crated in stock
market spills over the other and vice versa. To study this, MGARCH is used.
The result shows that volatility in the CSE General Index spillover that of the
USD/Taka exchange rate but not vice versa. Table 2 shows the spillover effect
of CSE General Index on the Taka/USD exchange rate market.
Table 2. Spillover effect of stock market on foreign exchange market
CSE General Index Taka/USD exchange rate

Volatility Taka/USD −.02(.40) -
Volatility CSE General Index - .0007∗∗∗ (.0002)
N 1911 1940
∗ ∗∗ ∗∗∗
p < .05; p < .01; p < .001
3 Conclusion
This study investigates and analyses the link between the stock market and for-
eign currency (ForEx) market of Bangladesh. The study was done to get insight
about on the volatility and its spillover effect between the stock and the foreign
exchange markets, and consequently the degree of their integration to expand
the information set available to international portfolio managers, multinational
corporations and policymakers for decision-making and policy formulation.
It was found that the persistence of volatility in stock market in Bangladesh
is very high which indicates that volatility clusters in stock market and decays
slowly. This study is similar to studies to Wu [11] and Jebran and Iqbal [7].
However, the study found low persistence in Taka/USD currency market. The
study also found that there is asymmetry in the volatility of in the stock market
i.e. negative news causes higher volatility compared to good news in the market.
The result shows that volatility in the CSE General Index spillover that of the
Taka/USD exchange rate but not the opposite.
References
1. Bala DA, Asemota JO (2013) Exchange-rates volatility in Nigeria: application of
garch models with exogenous break. J Appl Stat 4(1):89–116
2. Bollerslev T (1986) Generalised autoregressive conditional heteroskedasticity.
Econometrica 31(302):2–27
3. Campbell JY, Lo AW et al (1997) The econometrics of financial market. Princeton
University Press, Princeton
4. Cheung YW (2001) Equity price dynamics before and after the introduction of the
euro: a note. Multinational Financ J 5(2):113–128
5. Engle RF (1982) Autoregressive conditional heteroscedasticity with estimates of
the variance of United Kingdom inflation. Econometrica 50(4):987–1007
6. Engle RF, Patton AJ (2001) What good is a volatility model? Quant Financ
1(2):237–245
7. Jebran K, Iqbal A (2016) Dynamics of volatility spillover between stock market and
foreign exchange market: evidence from Asian countries. Financ Innov 2(1):1–20
8. Levine R (1999) Financial development and economic growth: views and agenda.
J Econ Lit 35(2):688–726
9. Levine R, Zervos S (1996) Stock markets, banks, and economic growth. Am Econ
Rev 88(3):537–558
10. Siddiqui J (2010) Development of corporate governance regulations: the case of an
emerging economy. J Bus Ethics 91(2):253–274
11. Wu RS (2005) International transmission effect of volatility between the financial
markets during the Asian Financial Crisis. Transition Stud Rev 12(1):19–35
12. Zivot E (2009) Practical issues in the analysis of univariate GARCH models.
Springer, Heidelberg
Cost/Efficiency Assessment of Alternative
Maintenance Management Policies
Diego Ruiz-Hernández(B) and Jesús Marı́a Pinar-Pérez
Department of Quantitative Methods, CUNEF-Ingenium, Madrid, Spain

d.ruiz@cunef.edu
Abstract. In this work we compare a number of maintenance schedul-

ing policies for a collection of deteriorating machines. We consider a
framework where a manager must allocate the effort of a small number
of technicians for providing preventive maintenance to a larger num-
ber of machines. The aim is mitigating wear and preventing failures or
breakdowns, guaranteeing the system’s availability, reliability and safety.
We evaluate different maintenance regimes that emerge in practice, dis-
tinguishing two main groups: time based and condition based policies.
The comparisons are conducted by means of an intensive computational
assessment of the performance of the different policies discussed.
Keywords: Machine maintenance management · Condition based

maintenance · Time based maintenance · Index policies · Simulation
1 Introduction
Maintenance interventions in deteriorating equipment are aimed at guaranteeing
its availability, reliability and safe performance whilst maximizing output and
minimizing waste. Such interventions may be predictive, preventive, or correc-
tive. Regarding the planning horizon, whereas corrective maintenance is reac-
tive (interventions come as a response to a failure), predictive and preventative
interventions are typically scheduled using either time-based or condition-based
schemes. Time-based interventions are usually scheduled according to an age-
based regime or following a pre-determined calendar (usually as indicated by
the producer or by certain legal requirements). In this work we focus our atten-
tion on preventative maintenance interventions.
There is a vast amount of literature addressing the problem of scheduling
preventative maintenance interventions in one single machine. Some recent ref-
erences are [3–6,12,13]. Good reviews of earlier contributions please see [1,8,20].
However, it is hard to find references in literature addressing the problem of
scheduling maintenance interventions to more than one machine, among them
[10,15,21]. Moreover, in most of the work addressing single machine mainte-
nance, the attention is focused on finding optimal intervention times given the
information gathered about the current condition of the machine without taking
1396 D. Ruiz-Hernández and J.M. Pinar-Pérez
into consideration the costs associated with the machine operation, intervention
and potential breakdowns. These values become particularly relevant when it
comes to deal with the design of multiple machines maintenance schedules, as
the trade-off intervention/operation-breakdown becomes more relevant.
Notwithstanding the shortage of academic work, technical literature shows
that maintenance tasks are typically allocated following either calendar based
programmes or, alternatively, regimes based on information about the current
state of the equipment [2,7,11,14,18,22]. Based in this observation, in this work
we compare the performance of a number of maintenance scheduling policies
which broadly replicate the approaches taking in real life problems.
2 Problem Formulation
In this work we address the problem of scheduling preventive maintenance inter-

ventions in a collection of independent machines by a limited number of techni-
cians. The aim is operating the system at the least operation/maintenance cost
whilst minimizing the cost of failures or breakdowns. At each period the man-
ager selects the number of machines that will be maintained during that period.
If necessary, technicians are allowed to remain idle during certain period.
As discussed in [19], the manager’s problem can be modeled a discrete-time
infinite-horizon Markov Decision Problem, where each machine is modeled as a
two action (operation/intervention) Markov Decision Chain that evolves over a
discrete state space1 . Roughly speaking, during the operation phase the machine
deteriorates, operation costs increase and the probability of a failure or break-
down becomes larger. Maintenance interventions, which are conducted at a cost,
aim at returning the machine to some pristine or as-good-as-new state. The
objective of the manager is to minimize the discounted operation/intervention
cost of equipment over an infinite horizon.
In order to characterize the decision maker’s problem we introduce the
following notation: let β, M and R represent the discount factor, and the
number of machines and repairmen, respectively. The system’s state space is
represented by the set S , with the state at any decision epoch t given by
S(t) = {sm (t), m = 1, · · · , M } ∈ S . The actions available are represented by
the set A , whose elements a = {am : am ∈ 0, 1, m = 1, · · · , M } represent the
admissible actions at any given state S(t); where am = 1 indicates that machine
m must be intervened during the current decision period, and am = 0 prescribes
operation. Finally, C is the collection of all possible vectors Ca (S), representing
the costs incurred from taking action a ∈ A when the system is in state S ∈ S .
With this elements we can now write down the objective of the decision maker
as the one of choosing the stationary, deterministic policy σ that minimizes
1
Please notice that, even if the machines’ condition or level of wear may not necessarily
be a discrete variable, if the deterioration of the machine occurs smoothly over
a linear space, then the state space can be discretized or uniformised in order to
obtain a discrete state space [17].
Cost/Efficiency of Maintenance Management 1397
the expected discounted operation/maintenance

∞ cost incurred over an infinite
horizon, namely V ∗ = minσ {Eσ [ t=0 β t Ca(t) S(t)]}.
Unfortunately, this problem suffers a dimensionality problem making stan-
dard dynamic programming techniques unavailable for finding an optimal solu-
tion. Instead, in real life, different approaches are taken (see [2,14,22]) in order
to obtain the most efficient maintenance scheduling regime possible. Most of
these approaches fall within the five broad categories described in the following
section.
3 Maintenance Policies
In this section we briefly describe the five maintenance regimes that have
been used in our work. All of them aim at minimizing the overall opera-
tion/maintenance cost of equipment. Three of them are condition based and
presume the existence of a condition monitoring system that provides timely
information about the system’s state. Most of the condition based policies that
are used in real life fit (with minimal adaptations) within the more general frame-
works described below. However, because of the large investment costs involved,
in some cases condition monitoring systems may not economically feasible (see,
for example, [16]). In such cases time based regimes are particularly useful. For
this reason, the fourth policy discussed in this work is a purely calendar based
one. Finally, the last policy discussed below combines features of both, condition
based and time based regimes.
(1) Threshold Policy
At each decision epoch, the manager observes the state of the system and
assigns maintenance tasks to those machines whose state is equal to or above
than certain threshold. If the number of candidate machines is larger than the
number of technicians, then the effort is allocated to those machines with larger
state of wear. Ties are broken randomly. More precisely, the action set of the
threshold policy for a threshold T at time t is given by a(t) = {am (t) : sm (t) ≤
M
T, m=1 am (t) ≤ R}.
(2) Pure Condition Based Policy

At each decision epoch, this policy prescribes to intervene those machines
with larger observed state of wear. As in the previous case, ties are broken
randomly. This kind of policies do not take into account other characteristics of
the machines as age, deterioration rate, operation and intervention costs, etc.
If we define an ordering sm,(i) , i = 1, · · · , M , where sm,1 is the machine with
largest state and so on, then we can build the following two sets ā = {sm,(i) , i =
1 · · · , R} and a = {sm,(i) , i = R, · · · , M }, therefore the condition based policy
is represented by: a(t) = {am |am = 1 ∀ m : sm,(i) ∈ ā; am = 0 ∀ m : sm,(i) ∈ a}.
(3) Index Policies
Index based families of maintenance scheduling policies have been proposed in
order to address the deficiencies of purely condition based policies, [9,10]. These
policies exploit all the information available for each machine at any time in order
to compute a so-called activity or intervention index. The indices are calibrated
for each machine/state pair and the policy prescribes to intervene those machines
with larger indices. Typically, the indices will have negative values for early states
of wear, when the machine is supposed to be in an as-good-as-new condition.
Consequently, only machines with positive indices will be intervened. This result
has been extended to all the other policies evaluated in this article, imposing
that no machine must be intervened whenever it lies in a state whose index has
negative value2 .
If we let Im, sm represent the index of machine m at state sm , and I is
the Rth largest index, then the actions available at time t are given by a(t) =
{am |am = 1 ∀ m : Im, sm > I ; am = 0 ∀ m : Im, sm ≤ I }.
(4) Calendar Based Policy
Periodical or calendar based policies follow a pre-defined intervention plan.
In our setting, a periodical intervention time is fixed for all the machines, with
the first intervention scheduled after certain sojourn time (when the machine is
operated from new). Even though in practice this calendar is determined by the
acquisition date and the periodical interventions prescribed by the maker, in our
formulation the interventions calendar is determined as follows: Once the sojourn
period is finished, at each period the decision-maker intervenes as many machines
as technicians are available according to their level of wear (pristine machines
are not intervened). Once a machine has been intervened, a new intervention
is scheduled at certain time in future (the interval between interventions is the
same for all machines and exogenously determined). Once all the machines have
been scheduled and intervened, a new cycle begins. For the sake of results’s
comparability with the index and other pure condition-based regimes, a variant
of this policy consisting of preventing the machine to be intervened in early
states of wear, has been included. These states coincide with the minimal state
for which all the indices computed for the index policy are non-negative.
(5) Dynamic Calendar Policy
When the collection of machines includes equipment from different versions,
ages or makers, their deterioration rates may differ and a rigid calendar-based
policies may not be completely appropriate. In order to address these cases, we
propose an extension of the pure calendar-based policy which is an hybrid of
calendar and condition based policies: at each decision epoch, the less deterio-
rated machines in the plan are substituted for other -more deteriorated ones-
which are not in the current period’s plan. The machines whose intervention is
postponed join a queue or buffer of delayed machines that will be included in
future plans.
This policy, depicted in Fig. 1, works as follows. First, all the machines are
assigned an intervention period as in the Calendar Based Policy. Once the inter-
vention stage has started, at each period the manager checks on the condition of
2
Please notice that this condition has only been imposed for the sake of results’
comparability and does not apply for more general cases.
all the machines. If any of the machines scheduled for intervention at the current
period is still at the pristine state, the machine is dropped-off from the list and
its identifier is included in the buffer. Simultaneously, if there are non-pristine
machines in the buffer, they are immediately included, by order of wear, in one of
the empty slots of the current plan. The decision maker then checks on the con-
dition of all the other, non-scheduled, machines and identifies those units whose
state of wear is worst than the best machine in the plan. If no machines are
identified, the intervention stage starts and the maintenance tasks are deployed.
If, otherwise, some machines are picked during the check, they are either allo-
cated to any empty slot in the current plan (if any), or substitute any scheduled
machines which are in a better condition. The machine (or machines if there
are ties) with the largest state of wear in the current intervention plan is never
substituted. The displaced machines are included in the buffer and the interven-
tion actions are deployed. As it was done for the purely calendar based policy,
a variant of this policy consisting of preventing the machine to be intervened in
early states of wear, has also been considered.
Fig. 1. Dynamic calendar policy
A number of numerical experiments was conducted in order to compare the
performance of each of the policies discussed in Sect. 3 under a collection of
different scenarios characterized by the following variables: maintenance and
operation costs; breakdown costs; transition rate; breakdown probability; and
number of machines technicians.
The value of the objective function, V ∗ is computed by means of the simu-

lation of 1040 decision periods. If we consider that interventions are conducted
weekly, this accounts for a planning horizon of 20 years, which corresponds to
the lifespan of a typical industrial equipment under normal conditions.
A total of 30 instances were simulated for each of 19440 possible scenarios.
The scenarios were obtaining as combinations of the following parameter specifi-
cations. Instances consisted of 30, 60, 100 and 150 machines with the number of
technicians given by ceiling of the 2%, 4% and 6% of the number of machines. The
machines evolve over a 52 states deterioration space, visiting a new state every 4
to 6.67 weeks in one case, and every 2 to 4 weeks in another. Therefore, the deteri-
oration probabilities are random extractions from U (0.15, 0.25) and U (0.25, 0.50)
distributions, respectively. Additionally, the deterioration probabilities increase
0.5% in every transition. For the breakdown probabilities we considered five cases
where each machine is first assigned a random value Bm obtained from one of the
following distributions U (0.05, 0.10, ), U (0.10, 0.15), U (0.15, 0.20), U (0.20, 0.25)
and U (0.25, 0.30)3 ; once this value is fixed, the breakdown probability of machine
m at state sm is computed as Bm e0.025sm .
Regarding the costs, three possible values were considered for the break-
down costs: 500, 1500 and 3000. A random linear operation costs func-
tion was defined with three alternative intersects obtained from distrib-
utions U (45, 55), U (135, 165) and U (270, 330), and slope extracted from a
U (11.25, 18.75) distribution. Finally, twelve different configurations of the inter-
vention costs where defined: three of them, constant random functions with
support on U (90, 110), U (225, 275) and U (450, 550); and other nine quadratic
configurations with the intersect following the same distribution as the constant
cases and the quadratic term obtained from one of the following distributions:
U (0.0075, 0.0125), U (0.0225, 0.0375) and U (0.0375, 0.0625).
In order to determine the states where intervention is allowed for all the
policies, a parameter Forbid was introduced. This parameter takes the value of
the minimal state for which all the indices in the index policy take non-negative
values for each particular instance. In order to do so, the indices corresponding
to the index policy are computed in the startup routine of each instance of the
numerical experiments. These values were obtained based on [19] by means of
the expression I(m, sm ) = G(s)(1 − E[β τ (0,s) ]) − K(0, τ (0, s)) − E[β τ (0,s) C(s̃)].
∗
In these expressions, with G(s) = K(s,1)+E[βC(s 1−β
)]−C(s)
. In these expressions,
τ (0,s)
E[β ] represents the accumulated discount during a transition from state 0
to state s, which takes an expected time given by τ (0, s); K(0, τ (0, s)) represents
the expected discounted intervention cost incurred in a transition between states
0 and s; E[β τ (0,s) C(s̃)] is the expected discounted cost of intervening a machine
after τ (0, s) periods of operation (s̃ represents the expected arrival state after the
operation sojourn); K(s, 1) is the expected discounted cost of operating during
one period when the initial state of the machine is s; E[βC(s∗ )] represents the
expected discounted cost of intervening the machine at the expected arrival state
3
Notice that in each case all the random values are extracted from the same distrib-
ution.
Table 1. Results of the illustrative example
Criteria Index Condit. Calendar Calendar- Dyn.Cal. Dyn. Threshold

policy Bsd policy F policy Cal.-F policy
policy policy policy
Average St. 2.25 2.26 2.93 3.33 3.30 3.37 2.26
Var. Av. St. 7.91e−02 1.24e−02 2.52e−01 2.05e−01 3.95e−02 4.35e−2 1.24e−2
Maximum St. 8 8 18 20 10 16 8
Breakdowns 2889 2878 2934 2956 2621 2671 2818
Av. St. Interv. 5.90 5.91 5.77 7.85 8.61 8.86 5.91
Var. Av. Interv. 3.69e−01 1.10e−02 8.24e−01 1.06e+00 1.61e−01 1.96e−01 1.08e−02
Total Cost 14220580 14293900 15610959 16366256 16140478 16310495 14293900
Numb. Interv. 3118 3118 3120 3120 3120 3120 3118
†The benchmark example consists of a collection of 60 machines and three technicians. The
operation costs are linear with constant parameter 150; the intervention costs are quadratic,
with intersect 500 and quadratic coefficient 0.03. The breakdown costs are 1500, with break-
down parameter 0.10. Parameter Forbid takes value 3.
after one period operation (notice that s∗ ∈ (0, s, s + 1)); finally, C(s) is the cost
of intervention when the machine is in s.4
Once the indices are obtained, the threshold policy is computed fixing the
threshold equal to the value Forbid ; likewise, the pure condition based policy is
only allowed to intervene machines with current state larger than Forbid. For
the calendar based policies two alternatives approaches were taken: in the first
one, the procedure was allowed to intervene any machine with state different
than zero; in the second one, intervention was only admitted for states larger
than Forbid.
A summary of the results obtained in a benchmark example is shown in
Table 1. The columns represent the seven alternative policies, with the suf-
fix (−F ) indicating that parameter Forbid has been used. The columns show,
respectively, the average visited state and its variance, the maximum visited
state, the number of breakdowns over 1040 periods, the average intervened state
and it variance, the total operation/maintenance cost and the total number of
interventions. The cost structure and the indices are illustrated in Fig. 2.
Figure 3 illustrates the evolution of an insulated machine under two of the
most relevant policies: Index and Calendar. It can be seen that, under the index
policy, the interventions tend to occur, in average, at state number 5, whereas the
evolution under the calendar policy tends to be more erratic, as this policy does
not take into consideration the state of the machine for scheduling interventions.
Given that the numerical values of the results of the experiments are not
as important as the patterns detected during the numerical computations, we
limit ourselves to presenting a graphical summary of our main findings. Firstly,
we notice that the average cost of the operation/maintenance schedule over the
4
Closed form expressions for this equations, which have been omitted for the sake of
briefness, are available from the authors upon request.
Maintenance and Operation Costs × 10 4 Value of the Intervention Indices
2500 5
4
2000
3
1500
Costs
Index
Operation Costs 2
1000
500 Maintenance Costs
0
5 10 15 20 25 30 35 40 45 50 5 10 15 20 25 30 35 40 45 50
States States
Fig. 2. Costs and Indices for 56 states in the illustrative example

Evolution of a Machine under the Index Policy Evolution of a Machine under the Calendar Policy
6 12
5 10
4 8
Machine State
Machine State
3 6
2 4
1 2
0 0
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Period Period
Fig. 3. Evolution of a random machine under index and calendar policies.
20 years planning horizon follows a very similar pattern for all policies, with
the time dependent policies showing larger costs (in average) than the state
dependent ones. This can be seen in panel (a) of Fig. 4, where the index (blue),
Calendar-F (red) and the threshold policies (green) are compared5 . This panel
shows the total cost incurred by the seven policies over the 19440 scenarios and
(the values are averages over 30 repetitions for each case)6 .
The sketch suggest the existence of four well defined blocks, corresponding to
each of the four different specifications for the number of machines. Each of these
main blocks is composed by three more or less differentiated segments, which
5
For the sake of simplicity, the dynamic calendar policies and the simple version of the
calendar policy have not been illustrated in this figure. However, their behavior does
not present significant differences with respect to the results shown in the displays.
6
Each point in the graph corresponds to the average cost, computed over 30 ran-
domly generated problems, of a particular combination of operation, maintenance
and breakdown costs, deterioration rates, breakdown probabilities and number of
machines and repairmen.
× 10 7 Total Costs for 19440 Different Cases

12 Average State Visited (60 Machines)
5.5
Index
Calendar Index
Threshold Calendar
5
Threshold
10
4.5
Operation/Maintenance Costs
8 4
3.5
Average State
6
3
2.5
4
2
1.5
2
0 0.5
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Cases × 10 4 Cases
(a) (b)
Number of Breakdowns (60 Machines) Number of Interventions (60 Machines)
7000 4500
Index Index
Calendar Calendar
Threshold Threshold
6000 4000
5000 3500
Interventions
Breakdowns
4000 3000
3000 2500
2000 2000
1000 1500
0 1000
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Cases Cases
(c) (d)
Fig. 4. Costs performance of the index, calendar-f and threshold policies over a 20 years
planning horizon
correspond to the three possible number of technicians used in our experiments.

The figure clearly shows that the costs are increasing on the number of machines
and decreasing on the number of technicians. In order to make our discussion
simpler, and given the fact that the results observed in one group follow the same
structure as the others (differing mainly in their absolute scale), in panels (b),
(c) and (d) we limit our discussion to the results observed for the cases with 60
machines. Panel (b) shows the average visited state during the planning horizon.
It can be appreciated that the time based regimes tend to visit larger states (i.e.
machines deteriorate more) than state based policies. It can also be observed
that the behavior of the threshold and the index policy is quite similar. This
suggests that, even in the absence of a well grounded index heuristic, a decision
maker will-in general-do better with a threshold policy than with a time based
one. In panel (c) we show the number of breakdowns suffered by the system
over the planning horizon. Again, the calendar policy shows a larger number
of breakdowns, and the index policy shows a smaller number of failures than
the other two. Finally, panel (d) depicts the number of interventions conducted
by each policy over the 20 years horizon. It can be appreciated that condition
based policies intervene much less than the time based ones. Moreover, the index
policy clearly outperforms the threshold one. It is also interesting to see that the
calendar policy intervenes a constant number of machines irrespectively of the
specific cost structure of the particular scenario, with interventions depending
only on the number of machines and technicians available.
× 10 7 Costs for 100 Machines and 4 Technicians × 10 7 Costs for 100 Machines and 4 Technicians
6 6
Bc=500 Bc=500
Bc=1500 Bc=1500
Bc=3000 Bc=3000
5 5
4 4
3 3
2 2
1 1
Det. Rate Case 1 Det. Rate Case 2 Det. Rate Case 1 Det. Rate Case 2
0 0
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
Cases Cases
(a) (b)
Fig. 5. Inner structure of the index policy results
We finally take a look inside of our plots in order to illustrate the inner
structure of the results. In order to do so, we pick the case where there are 100
machines, four technicians and the costs are quadratic. Figure 5 illustrates the
corresponding 810 results obtained for the index policy. Vertically, each panel
is divided in two sections, corresponding to each of the two alternative deteri-
oration rates proposed. Each section consists of five columns representing each
of the possible configurations of the breakdown probabilities. Horizontally, it is
possible to distinguish three blocks consisting of three groups of points (red, blue
and magenta). Each of these groups is associated with a different value of the
breakdown cost as illustrated by the legend. Additionally, each of the three main
levels (consisting of three colors each) correspond to a different operation cost.
For example, the southernmost red group corresponds to the smallest operation
and breakdown costs; the next one (blue) corresponds to the lowest operation and
the second breakdown cost, and so on. It can be seen that the colored points are
grouped in small clusters. Each of these clusters correspond to the nine different
combinations of the maintenance cost parameters. The second panel depicts the
cases of a fixed maintenance cost and two alternative (high and low) operation
costs, together with the three alternatives for the breakdown cost. This panel
allows a more clear visualization of the behavior of our formulation. It can be
seen, for example, that a larger deterioration rate implies larger costs for all the
configurations (comparing the right and the left hand side sections). In general,
the costs are larger as the likelihood of a breakdown increases, however, it is
worth noticing that for low breakdown costs, this relationship may become neg-
ative. This may be reflecting an incentive to intervene less when the replacement
cost of the equipment is small compared with the maintenance costs. This struc-
ture can be observed all over the set of experiments with only small variations
between policies.
5 Conclusion
In this article presented a collection of general purpose machine maintenance
scheduling policies that mimic the intervention strategies that are used in prac-
tice. These policies can be grouped in two broad categories: time and condition
based policies. We additionally suggest a novel alternative consisting of a mixture
of both. Our aim has been to obtain an assessment of their relative performance
before different combinations of operation, intervention and failure costs, dete-
rioration rates and breakdown probabilities.
With this aim, we conducted a large number of experiments. In general the
results are according to what intuition will suggest: condition based policies in
general perform better than time based ones. We also found that threshold poli-
cies in general outperform the pure condition based ones and are almost as effi-
cient as the more sophisticated index policies. Our results show that when inter-
vention and breakdown costs are low, time based policies may perform better
than the index one, as this will try to postpone the intervention tasks, increasing
the probabilities of breakdowns.
Our results are consistent with the trend that has been observed in the last
few years in heavy equipment based industries: the use of time based policies
has been gradually relegated in favor of more efficient condition based strategies.
However, time based policies are still relevant because of both guarantee, safety
and maintainability purposes, and the fact that in many cases condition moni-
toring systems require an important investment that cannot always be afforded
by small firms.
References
1. Ahmad R, Kamaruddin S (2012) An overview of time-based and condition-based
maintenance in industrial application. Comput Ind Eng 63(1):135–149
2. Ben-Daya M, Ait-Kadi D et al. (2009) Handbook of maintenance management and
engineering, vol 7. Springer
3. Dao CD, Zuo MJ, Pandey M (2014) Selective maintenance for multi-state series-
parallel systems under economic dependence. Reliab Eng Syst Saf 121:240–249
4. Do P, Voisin A et al (2015) A proactive condition-based maintenance strategy with
both perfect and imperfect maintenance actions. Reliab Eng Syst Saf 133:22–32
5. Doostparast M, Kolahan F, Doostparast M (2014) A reliability-based approach to
optimize preventive maintenance scheduling for coherent systems. Reliab Eng Syst
Saf 126:98–106
6. Fitouhi MC, Nourelfath M (2014) Integrating noncyclical preventive maintenance
scheduling and production planning for multi-state systems. Reliab Eng Syst Saf
121:175–186
7. Márquez FPG, Tobias AM et al (2012) Condition monitoring of wind turbines:

techniques and methods. Renew Energy 46:169–178
8. Garg A, Deshmukh S (2006) Maintenance management: literature review and direc-
tions. J Qual Maint Eng 12(3):205–238
9. Glazebrook K, Mitchell H, Ansell P (2005) Index policies for the maintenance of
a collection of machines by a set of repairmen. Eur J Oper Res 165:267–284
10. Glazebrook K, Ruiz-Hernandez D, Kirkbride C (2006) Some indexable families of
restless bandit problems. Adv Appl Probab 38:643–672
11. Hodge V, O’Keefe S et al (2015) Wireless sensor networks for condition monitoring
in the railway industry: a survey. IEEE Trans Intell Transp Syst 16(3):1088–1106
12. Khatab A, Aghezzaf EH (2016) Selective maintenance optimization when quality
of imperfect maintenance actions are stochastic. Reliab Eng Syst Saf 150:182–189
13. Liu B, Xu Z et al (2014) A value-based preventive maintenance policy for multi-
component system with continuously degrading components. Reliab Eng Syst Saf
132:83–89
14. Mobley R, Higgins L, Wikoff D (2008) Maintenance engineering handbook.
McGraw-Hill, New York
15. Moghaddam KS (2013) Multi-objective preventive maintenance and replacement
scheduling in a manufacturing system using goal programming. Int J Prod Econ
146(2):704–716
16. Pinar-Pérez J, Garcı́a-Márquez F, Ruiz-Hernández D (2016) Economic viability
analysis for icing blades detection in wind turbines. J Clean Prod 135:1150–1160
17. Puterman ML (1994) Markov decision processes: discrete and stochastic dynamic
programming. Wiley, New York
18. Riera-Guasp M, Antonino-Daviu J, Capolino G (2015) Advances in electrical
machine, power electronic, and drive condition monitoring and fault detection:
state of the art. IEEE Trans Industr Electron 62(3):1746–1759
19. Ruiz-Hernández D (2007) Indexable restless bandits. Index policies for some
stochastic scheduling and dynamic allocation problems. Verlag Dr. Müller,
Saarbrücken, Germany
20. Sharma A, Yadava G, Deshmukh S (2011) A literature review and future perspec-
tives on maintenance optimization. J Qual Maint Eng 17(1):5–25
21. Tao X, Xia T, Xi L (2014) Opportunistic preventive maintenance scheduling based
on theory of constraints. In: Proceedings of the IIE annual conference. Institute of
Industrial Engineers-Publisher, p 230
22. Wang H (2002) A survey of maintenance policies of deteriorating systems. Eur J
Oper Res 139(3):469–489
Heijunka Operation Management
of Agri-Products Manufacturing by Yield
Improvement and Cropping Policy
Ritsuko Aoki1(B) and Hiroshi Katayama2

1
Department of Industrial and Management Systems Engineering,
Graduate School of Creative Science and Engineering,
Waseda University, Tokyo, Japan
may-mar.1103@ruri.waseda.jp
2
Abstract. The environment of agricultural businesses in Japan is

changing drastically due to decreasing self-sufficiency rate and the num-
ber of farmers, increasing average age of concerned workers, and compe-
tition with imported products. To cope with these trends, national gov-
ernment revised agricultural land law which came into effect in 2009 and
some companies in this industry are implementing industrial manage-
ment methods such as lean management for better performance. Besides
the general common problems of manufacturing industry, agri-products
have distinctive features such as production in natural environment, per-
ishability of products, demand volatility although products are living
necessaries. Based on the above cognition, this study discusses implemen-
tation of Heijunka (the method of level production and supply) operation,
which is one important concept of lean management.
Keywords: Lean management · Supply chain · Agri-businesses · Hei-

junka production · Seeding schedule · Segmented farmland
1 Introduction
Because of the decreasing self-sufficiency rate and the number of farmers, increas-
ing of the average age of farmers and limited agricultural land, the environment
of agricultural business is in tough phase [6]. In addition, extreme narrowness of
agricultural land area hinders production cost reduction, which leads to decline in
price competitiveness of domestic agricultural products against imported goods.
For this reason, in recent years the government has considered the ways of agri-
cultural reform and promotion of more efficient production activities. As the
result, the revised agricultural land law was implemented in 2009. So many
companies try to adopt the management approach developed in manufactur-
ing industries such as TPS (Toyota Production System) and aim for stable and
efficient supply. However this scientific approach against hidden problem was
1408 R. Aoki and H. Katayama
not satisfactory performed in the conventional agriculture, because of managers’

leadership and employees’ understanding, intuitions and skills [4].
Based on the above cognition, this study investigates and analyzes cases of
farmers’ efforts which incorporate industrial management methods and guide to
effective management of agricultural production.
2 Related Study
Vassian [7] logically derived the periodic production ordering method that
achieves the minimization of inventory fluctuation by using the demand fore-
cast values, past production ordering and the term end stock amount. Based
on this, Chiyoma et al. [1] proposed the production planning model for agricul-
tural supply chain by the following mathematical Eqs. (1), (2), and (3), where
consideration is given to varied harvest quantity that peculiarly happens on
agricultural products.
L
L−1

Pt = D̂t,t+i − Pt−i − It + SS, (1)
i=1 i=1
It = It−1 + Qt − Dt , (2)
Qt = Pt−L + εt . (3)
Symbols
Pt : Production quantity planned at period t, completed at period t + L;
D̂t,t+i : Demand quantity of period t + i predicted at the end of period t;
It : Inventory quantity at the end of period t;
SS : Safety stock level;
L : Production lead time + 1;
Dt : Demand quantity at period t;
Qt : The yield quantity at period t, where it does not always coincide with
the planned quantity;
εt : The difference between Qt and Pt−L .
3 Background and the Production Process

of the Objective Products
In this study, research procedure includes case investigation of a company that
aggressively adopts lean management, model building of manufacturing process
focusing on loss structure, development of Heijunka contrivance for the objec-
tive system, of which idea is cropping scheduling based on segmented farmland.
Actually, this study focuses on the baby leaf growing and supply process at the
collaborating company H [2]. Among the people, who carefully consider their
health, baby leaf are now booming foods and this business has been growing
rapidly in Japan and many SMEs including this company were established. As
a result, the competitiveness of this field becomes serious. On the other hand,
Heijunka Operation Management of Agri-Products Manufacturing 1409
there is a burden on production and delivery activities because of the lack of

reliable workers and useful tools/weapons regarding agricultural management.
Companies trying to mitigate such obstacles have begun to learn from com-
panies in other industrial field and introduce their management methods and
techniques. For example, TPS is a typical lean management scheme born in
Japan. Company H started the TPS learning project in 2006 with Toyota Motor
headquarters under supervision of former Chairperson Mr. Cho and Senior Tech-
nical Executive Mr. Hayashi, who proposed how to introduce TPS during 2006–
2008 [2]. This study focuses on reaching the next stage of lean, namely realizing
level production and supply (that is called Heijunka) operations based on the
historical approaches such as 5S, visualization, yield improvement etc.
The objective process consists of 4 major operations, i.e. seeding, germina-
tion, harvesting and shipping, and it is illustrated in Fig. 1. Note that there are
some unknown mechanism creating losses in each sub-process, these losses are
called germination yield rate, harvest rate and sorting rate where weight gain
is increasing quantity of the baby leaf during the growing process. Models for
parameter estimation of these losses and weight gain are given in the follow-
ing 4 steps with Eqs. (4)–(10) and the result is given in Table 1 [5], where it is
known that germination, harvest and sorting rates are stochastically fluctuated
and weight gain rate stays constant which depends on climate (warmer or cooler
terms).
Fig. 1. Outline of the objective production process [5]
Step 1. Calculation formulas of δ̂i , θ̂k
Estimates δ̂i and θ̂k can be evaluated by least square method through mini-
mizing the total difference between actual data of harvest and its estimated
quantity as shown in Eqs. (4), (5), (6) and (7).
2
min (Ĉi − Ci ) , (4)
i∈N
Ĉi = Ai × δ̂i × θ̂k (k = 1, 2), (5)

s.t. 0.64 ≤ δ̂i ≤ 1, (6)

50 ≤ θ̂k ≤ 300. (7)
Step 2. Calculation formulas of standard deviation
Estimating standard deviation of δ̂i is also important as this value represents

the degree of internal turbulence that harms stable production.
σδ̂ = S(δ̂i ), (8)
where, sample data of δ̂i is obtained by the formula δ̂i = ACθ̂i , which is
i k
deduced from formula (5) and S is the operator for standard deviation.
Step 3. Calculation formula of γ̂
Point estimate of sorting rate γ̂ can be obtained by the quotient of actual

data of shipment and harvest quantities as shown in Eq. (9), where, E is the
operator for mean value.

Di
E(γ̂) = E . (9)
Ci
Step 4. Calculation formula of standard deviation
As the same meaning described in Step 3, standard deviation of γ̂ is also

important to estimate.

Di
σγ̂ = S . (10)
Ci
Symbols
Ai : Seeding quantity at planning term i [kg/a];
Ci : Harvest quantity at planning term i [kg/a];
Ĉi : Estimated harvest at planning term i [kg/a];
Di : Shipping quantity at planning term i [kg/a];
D̂i : Estimated shipping quantity at planning term i [kg/a];
N : Planning horizon [day];
αi : Germination yield rate at planning term i;
βi : Harvest yield rate at planning term i;
γ̂ : Estimated sorting yield rate;
θ̂k : Estimated weight gain rate in k, (k = 1 : warmer season, k = 2 : cooler
season)
δ̂i : αi × βi which is called estimated value of aggregated yield.
4 Concept of Heijunka and Its Realization

In this section, general approach and agri-specific production model for Heijunka
are discussed.
Table 1. Estimated statistics of parameters [5]
Germination Weight gain Weight gain Sorting rate

rate × harvest (warmer term) (cooler term)
rate
Average 0.63 78.29 128.58 0.57
Standard deviation 0.14 N/A N/A 0.21
4.1 General Approach for Heijunka
It is necessary that standard deviation and average of variable must be controlled

in this order to realize continuous and stable production, which is called Heijunka
production. General approach for Heijunka is illustrated in Fig. 2.
• Phase 1: Recognition of “Muri”; Firstly, current Muri must be recognized.

• Phase 2: Reduction of “Mura”; Secondly, variance reduction must be focused.
• Phase 3: Reduction of “Muda”; Finally, reduction of “Muda” must be con-
ducted after reduction of “Mura”, where level production and supply opera-
tions become much easier than before because internal turbulence was con-
trolled in Phase 2.
Fig. 2. General approach for Heijunka
4.2 Agri-Specific Production Model for Heijunka
For Heijunka of the objective operation, 3M (“Muri”, “Mura”, “Muda”) are

visualized and zero-tized, and seeding schedule based on segmented farmland
are discussed respectively.
(1) Visualizing and zero-tizing the phenomena of 3M (Conventional lean app-
roach)
Heijunka procedure regarding to Muri, Mura and Muda for the objective
process is illustrated in Fig. 3 where target variable and its factors are shown.
Stochastic simulation is a possible way to analyze the characteristics because
variables and parameters in this system are fluctuated. In this study, firstly,
current Muri is understood and assumed. In the second stage, variance control
must be focused, which consists of 2 experiments. Factors affecting variation of
harvest interval are considered as seeding interval and growth term, and that of
Fig. 3. Procedure of experiment
harvest quantity is considered as aggregated yield, and these standard deviation

must be reduced. In the final stage, average of seeding interval is controlled under
the idea of frequent small-lot production of industrial management method.
(2) Seeding schedule based on segmented farmland (Lean approach with case
depending procedure)
It is essential to introduce a production system that can meet demand with
limited field area because productivity improvement is indispensable in agricul-
tural business in Japan. Therefore, seeding schedule on segmented farmland that
can flexibly respond to demand quantity is considered. Design of seeding quan-
tity, harvest quantity and harvest date is defined by Eqs. (11), (12) and (13).
Ai = T /N, (11)
Ĉi = Ai δ̂i θ, (12)
dî = Ri + si . (13)
Symbols
N : Number of segments of farmland;
T : Seeding capacity of entire farmland [kg];
Ai : Seeding quantity per segment at planning term i [kg/seg].
Ĉi : Estimated harvest quantity at planning term i [kg/seg];
δ̂i : Estimated value of aggregated yield of seeding quantity at planning term i;
θ : Supposed weight gain rate (percentage of weight increase);
dî : Estimated harvest date of seeded plants at term i;
Ri : Growth duration of seeds at planning term i;
Si : Seeding date at planning term i.
In addition, an example of seeding schedule is shown in Fig. 4, and an example
of input and output dates are shown in Fig. 5, where the number of segments is 9.
5 Result
In this Sect. 2 results derived from the model/procedure described in the previous
chapter are summarized.
Fig. 4. Example of seeding schedule
Fig. 5. Example of input and output time series
5.1 Performance of Conventional Lean Approach

(1) Phase 1: Quantification of “Muri” and its effects
The lack of human resources and skills throughout the whole process from
seeding to shipment is recognized. Performance of current system is given in
Table 2.
Table 2. Condition and performance of current system
Variable Aggregated Result

yield
Seeding Seeding Growth Harvest Harvest
quantity interval duration quantity interval
[kg/a] [days] [days] [kg/a] [days]
Average 0.85 7.12 63.26 0.63 71.27 8.28
Standard 0.08 5.24 18.10 0.14 18.72 7.71
deviation
(2) Phase 2: Reduction of “Mura”

The purpose of the experiment is to evaluate the stability of harvest perfor-
mance by improving fluctuation of variables. Firstly, standard deviation of the
seeding interval and growth duration were reduced. Result is given in Table 3.
Secondly, in addition to the condition of the first experiment, standard devi-
ation of the aggregated yield is reduced. Result is given in Table 4.
Table 3. Result of “Mura” reduction: first experiment (∗ is focused variable and per-
formance)

yield
Average 0.89 7.05 63.26 0.63 73.35 7.06
∗ ∗
Standard 0 1.26 6.08 0.14 17.20 4.89∗
deviation
Table 4. Result of “Mura” reduction: second experiment (∗ is focused variable and

performance)

yield
Average 0.89 7.05 63.26 0.63 72.27 7.06
Standard 0 1.26 6.08 0.05∗ 5.66∗ 4.89
deviation
Table 5. Result of “Muda”reduction (∗ is focused variable and performance)

yield
Average 0.73∗ 5.78∗ 63.26 0.63 59.29∗ 5.82∗
Standard 0 1.26 6.07 0.05 4.89 4.32
deviation
(3) Phase 3: Reduction of “Muda”

The purpose of this experiment is to evaluate the influence of frequent small-
lot production (Heijunka). For a pragmatic improvement point of view, a case
that average seeding interval and quantity are 6/7 respectively is considered for
analysis. Result is given in Table 5. Where, it is assumed that necessary farmland
can be procured without any restriction but based on the contracts.
Changes in each value of standard deviation and average are described below.
• Before and after ratio of standard deviation of harvest interval is 0.56 approx-
imately (from 7.71 to 4.32) and that of harvest quantity is 0.26 approximately
(from 18.72 to 4.89).
• Before and after ratio of average of harvest interval is 0.7 approximately (from
8.28 to 5.82) and that of harvest quantity is 0.8 approximately (from 71.27
to 59.29).
Relationship of before and after improvement is given in Fig. 6. By the above
procedure, harvest date and quantity can be predicted more accurately by reduc-
ing Mura which is general turbulent factor, and stock volume and space are
expected to be improved by reducing both seeding interval and quantity as men-
tioned in Fig. 3. This procedure leads us to the Heijunka world. Worker’s skill
and the number of necessary workers can be encouraged to improve and then
Muri of current system must be disappeared.
Fig. 6. Relationship on before and after improvement
5.2 Performance of Lean Approach with Case Depending Procedure

In Sect. 5.1, condition of farmland is not considered. In this section, performance
of the proposed model is as described in Sect. 4.2, where farmland segmentation
and its scheduling procedure [3] are added to the conventional lean approach, is
evaluated by experiments. Relationship between number of segments and har-
vest interval and that of harvest quantity are given in Fig. 7. The star marks
(‫ ۻ‬and ) in these figures show the result of the model described in Sect. 5.1,
where ‫ ۻ‬and represent the performance described in Tables 3 and 4 respec-
tively. Obtained performance of the proposed model indicates that average har-
vest interval and quantity are both decreasing along with increase of number
of segments. Standard deviation of these variables has the same characteristics.
And there are significant gaps in the characteristics between standard deviations
of harvest interval obtained by Sects. 5.1 and 5.2. Namely, the proposed model
discussed in this section seems to provide worse performance than conventional
lean approach. Main reason will be consideration of the pragmatic restriction on
segment rotation in the proposed model, which simply describes seeding is the
precedence operation of harvesting in the same segment.
Fig. 7. Trial experiment results (five years)
In this study, agricultural production management method is suggested for
achieving Heijunka operation and applied to a production management division
of collaborating agri-company. The turbulent factors in the agricultural produc-
tion process were clarified and its performance is analyzed by simulation exper-
iments by controlling levels of these factors. The obtained results revealed that
elimination of turbulent elements such as seeding interval, growth duration and
aggregated yield are critical for performance improvement. Furthermore seeding
schedule on segmented farmland can respond flexibly to demand changes under
restricted usability of farmland. Obtained result can be a fine example of lean
management transfer to agriculture industry.
References
1. Chiyoma H, Murata K, Katayama H (2014) On performance analysis of sustainable
closed logistics system in agricultural industry. In: Proceedings of the 17th annual
conference of Japan society of logistics systems, pp 101–106
2. Hatake Company Co., Ltd. HP (2016) (in Japanese). http://www.t-k-f.jp/company.
html
3. Ito K (2006) Mathematical programming problems in agricultural management.
Oper Res Manage Sci 5:264–267 (in Japanese)
4. Katayama H (2016) Transfer activities of lean management to other industries-
transplanting Heijunka concept for leanised operations. In: Proceedings of the 20th
Cambridge international manufacturing symposium, pp 1–17
5. Katayama H, Aoki R (2016) On Heijanka manufacturing system for agricultural
products. In: Proceedings of the 46th computers and industrial engineering, p 8
6. Ministry of Agriculture, Forestry and Fisheries HP (2015). http://www.maff.go.jp/
7. Vassian HJ (1955) Application of discrete variable servo theory to inventory control.
J Oper Res Soc Am 3:272–282
Optimizing Reserve Combination with
Uncertain Parameters (Case Study: Football)
Masoud Amel Monirian(B) , Hamideh Razavi,

and Shabnam Mahmoudzadeh Vaziri
Industrial Engineering Department, Ferdowsi University of Mashhad,

Mashhad, Iran
masoudmonirian@yahoo.com
Abstract. Preventive maintenance is critical when non-stop operation

with maximum efficiency and reliability is required. In such conditions,
proper identification of reserve and spare parts in terms of types and
models, order and stock sizes as well as replacement timings become
highly important. Arranging reserve sets, many constraints such as costs,
space, obsolescence and deterioration must be met. Therefore, an inte-
grated model for appropriate selection of reserves and timing for substi-
tution is needed. This model can be similarly used in multiplayer sports
such as football games. In this paper we propose an uncertain mathe-
matical model to solve reserve combination strategy.
Keywords: Reserve combination · Optimization · Uncertainty theory ·

Uncertain programming
1 Introduction
Selecting an optimum reserve set and implementing the best replacement strat-
egy is a challenge for most industries. Likewise, in multiplayer sports, choosing
the right combination of the reserve team and taking a proper replacement strat-
egy needs an advance planning and optimization. Inaccurate arrangements can
cause significant financial loss as well as reputations.
When a team enters a tough league to play against different opponents with
different methods, these issues become even more critical. Success or failure
of each team depends on member’s ability and skills. Therefore, it is essential
to arrange the best combination both in main and reserve team in order to
obtain the optimum results [2,9]. A few structured models have addressed this
problem. Most of the studies emphasis on combination of the main team [1].
Since decisions are made prior to the games, we face probability and uncertainty
in the problem, for which we use uncertain programming. It is a powerful tool
in modeling industrial, human, natural and peddling and probability systems.
Certainly in intensive games, it is impossible to use 11 players. Players due
to fatigue, injury, disciplinary cases and other reasons aren’t available in all
1418 M.A. Monirian et al.
games, and coach should use all main and reserve players to reach victory and
profitability. Importance of reserve players when they replaced in proper time
with suitable person, is not hidden to anyone, because with their adequate energy
and by using skills and power can be influential. For instance, in final game of
world cup 2014 in extra times, substitution players led the German team won
the game.
1.1 Football Overview
Players are divided to 5 posts: goal keeper, defender, midfielder, sides, forward.
According to game’s technique, number of players in each post (expect goal
keeper) can be different. This classification is done according to each player’s
abilities. Players due to their skills can be suitable for several posts.
1.2 Problem Definition
Broadly speaking, in selecting reserve team we assume that we have a main team
and according to its constraints, situation and game circumstances, a reserve
team should be selected. Replacement strategy conforming the game process
and player’s situation, declares that which player in the pitch should replace
with which player in reserve team, in best time in order to preserve quality
of the team. A football team plays in diverse leagues. Sometimes the schedule
becomes so compressed, even they may have 3 or 4 games in a week. This event
causes player’s exhaustion. Consequently, generally main combination are not
considered fix, so reserve combination selection problem and also replacement
planning are very important issues. The first and main step for selecting reserve
team and replacement strategy is recognition and evaluation of main team and
desired options for reserve so that able us to compare different players in each
post with stable criteria. First, players evaluated with belief degrees and by using
mathematical model, choose reserve team to maintain team quality at maximum
state. For rating players we need to consider several number of quantitative and
qualitative characteristics in player’s selection, like individual skills, statistics
data in previous matches, physical readiness, psychological factors, injury and
also opponent’s condition [8]. In average, each football team, in normal situation,
has more than 24 players that 10 of them plays in main combination and 6
of them are in reserve team. In this problem, we do not consider goal keeper
selection. Because they run less during the game so their fatigue is less, but for
compulsive situations, injury or expulsion it is necessary to have a goalkeeper
in reserve team [3,8]. For scoring and evaluating players, 18 indexes and criteria
such as heading, jumping, dribbling, tackling and so on are defined, that these
criteria are approved by football simulation’s companies. We value these criteria
according to coach’s opinion. In Table 1 these criteria are presented.
Uncertainty theory was founded by Liu [5] in 2007 and studied by many
researchers. Uncertainty theory is a new method for modeling undetermined
phenomena based on normality, duality, subadditivity and product axioms.
Optimizing Reserve Combination with Uncertain Parameters 1419
Table 1. List of criteria
Criteria Number Criteria Number

Heading, jumping 1 Create goal scoring position 10
Shoot 2 Tackling 11
Short passing 3 Both feet 12
Crossing 4 Great stamina 13
Ball control 5 Height 14
Dribbling 6 Providing through (long) pass 15
Finishing (composure) 7 Technical ability 16
Speed 8 Create attacking opportunities 17
Creativity 9 Read the game 18
Uncertain programming presented by Liu [3], is a type of mathematical pro-

gramming which involves uncertain variables. This paper aims to model uncer-
tainty in football decisions based on uncertainty theory.
The rest of this paper is organized as follows. In Sect. 2, we review some basic
concepts about uncertainty theory. Then we introduce uncertain programming
in Sect. 3. After that, we introduce determining the reserve combination problem
and uncertain determining the reserve combination problem in Sects. 4 and 5,
respectively. In Sect. 6, we propose experimental results and solve the model. At
last, some remarks are made in Sect. 7.
2 Preliminary
Uncertainty theory is a branch of axiomatic mathematics based on normality,
duality, subadditivity and product axioms. In this section, we introduce some
concepts in uncertainty theory, which are used throughout this paper.
Definition 1 (Liu [7]). Let τ be a nonempty set, and £ be a σ− algebra on

Γ . A set function M is called an uncertain measure if it satisfies the following
axioms,
Axiom 1: (Normality Axiom) M {Γ } = 1;
Axiom 2: (Duality Axiom) M {Λ} + M {Λc } = 1 for any event Λ;
Axiom 3: (Subadditivity Axiom) For every sequence of, we have
∞ ∞

M Λi ≤ M {Λi }. (1)
i=1 i=1
In this case, the triple (Γ, £, M ) is called an uncertainty space.

In addition, a product axiom on the product of uncertainty spaces was defined
by Liu [7], which is the fourth axiom of uncertainty theory.
Axiom 4: (Product Axiom) Let (Γk , £k , Mk ) be uncertainty spaces for Then

the product uncertain measure M is an uncertain measure satisfying
∞

M Λk = Λ∞ k=1 Mk {Λk } , (2)
k=1
where Λk are arbitrarily chosen events from £k for k = 1, 2, · · · , respectively.
Definition 2 (Liu [7]). An uncertain variable ξ is a measurable function from

an uncertainty space (Γ, £, M ) to the set of real numbers, i.e., for any Borel set
ξ of real numbers, the set
{ξ ∈ B} = {γ ∈ Γ |ξ(γ) ∈ B} (3)
is an event.
Definition 3 (Liu [4]). The uncertain variables ξ1 , ξ2 , · · · , ξn are said to be

independent if
n

M (ξi ∈ Bi ) = Λni=1 M {ξi ∈ Bi } (4)
i=1
for any Borel sets B1 , B2 , · · · , Bn of real numbers.
Definition 4 (Liu [7]). The uncertainty distribution φ of an uncertain variable

ξ is defined by
φ(x) = M {ξ ≤ x} (5)
for any real number x.

If the uncertainty distribution φ(x) has an inverse function φ−1 (α) for each
α ∈ (0, 1), then φ−1 (α) is called the inverse uncertainty distribution of the uncer-
tain variable ξ. In this case, the uncertainty distribution φ is said to be regular.
An uncertain variable ξ is called linear if it has an uncertainty distribution
⎧
⎨ 0, if x < a
(x−a)
φ(x) = (b−a) , if a ≤ x ≤ b (6)
⎩
1, if x > b
denoted by £(a, b).
Theorem 1 (Liu [6]). Let ξ1 , ξ2 , · · · ., ξn be independent uncertain variables

with regular uncertainty distributions φ1 , φ2 , · · · , φn , respectively. If the func-
tion f (x1 , x2 , · · · , xn ) is strictly increasing with respect to x1 , x2 , · · · , xm and
strictly decreasing with respect to xm+1 , xm+2 , · · · , xn , then ξ = f (ξ1 , ξ2 , · · · , ξn )
is an uncertain variable with an inverse uncertainty distribution
φ−1 (α) = f (φ−1 −1 −1 −1

1 (α), · · · , φm (α), φm+1 (1 − α), · · · , φn (1 − α)). (7)
Definition 5 (Liu [7]). The expected value of an uncertain variable ξ is defined

by
+∞ 0
E [ξ] = M {ξ ≥ x} dx − M {ξ ≤ x} dx, (8)
0 −∞
provided that at least one of the two integrals is finite.

If an uncertain variable ξ has a regular uncertainty distribution φ, then
1
E [ξ] = φ−1 (α)dα (9)
0
provided that the expected value E [ξ] exists. Lease acknowledge collaborators
or anyone who has helped with the paper at the end of the text.
3 Uncertain Programming
Uncertain programming, which was first proposed by Liu [5], is a type of mathe-
matical programming involving uncertain variables. Assume that x is a deci-
sion vector, and ξ is an uncertain vector. Since the uncertain programming
model contains the uncertain objective function f (x, ξ) and uncertain constraints
gj (x, ξ) ≤ 0, j = 1, 2, · · · , p, Liu [5] proposed the following uncertain program-
ming model,
minx E [f (x, ξ)]

(10)
s. t. M {gj (x, ξ) ≤ 0} ≥ αj , j = 1, 2, · · · , p.
Definition 6 (Liu [6]). A vector is called a feasible solution to the uncertain

programming model if
M {gj (x, ξ) ≤ 0} ≥ αj (11)
for j = 1, 2, · · · , p.
Definition 7 (Liu [6]). A feasible solution x∗ is called an optimal solution to

the uncertain programming model if
E [f (x∗ , ξ)] ≤ E [f (x, ξ)] (12)
for any feasible solution x.
Theorem 2 (Liu [6]). Assume f (x, ξ1 , ξ2 , · · · ., ξn ) is strictly increasing with

respect to ξ1 , ξ2 , · · · , ξm and strictly decreasing with respect to ξm+1 , ξm+2 , · · · , ξn ,
and gj (x, ξ1 , ξ2 , · · · , ξn ) are strictly increasing with respect to ξ1 , ξ2 , · · · , ξk
and strictly decreasing with respect to ξk+1 , ξk+2 , · · · , ξn for j = 1, 2, · · · , p.
If ξ1 , ξ2 , · · · , ξn are independent uncertain variables with uncertainty distributions

φ1 , φ2 , · · · , φn , respectively, then the uncertain programming
minx E [f (x, ξ1 , ξ2 , · · · , ξn )]
(13)
s. t. M {gj (x, ξ1 , ξ2 , · · · , ξn ) ≤ 0} ≥ αj , j = 1, 2, · · · , p.
is equivalent to the crisp mathematical programming

⎧
⎨ minx 0 f (x, φ−1 −1
1 −1 −1
1 (α), · · · , φm (α), φm+1 (1 − α), · · · , φn (1 − α))dα
−1 −1 −1
s. t. gj (x, φ1 (αj ), · · · , φk (αj ), φk+1 (1 − αj ), · · · , φ−1
n (1 − αj )) ≤ 0,
⎩
j = 1, 2, · · · , p.
(14)
4 Mathematical Model for Determining the Reserve

Combination
In this model, due to the evaluations from players on the pitch and reserve players
and considering constraints, number of players are chosen to be in reserve team,
aims to maximize the power of reserve team and minimize the reduction of team’s
weakness. The model is linear with binary variables.
Indices:
n: number of players (in football it is usually 22 players);

k: number of features k = 1, 2, · · · , 18 (Table 2);
i: the number of fixed players i = 1, 2, · · · , 10;
j: the number of reserve players j = 1, · · · , n − i;
l: number of posts l = 1, · · · , 4 (for values of l we have:
l = 1 defender, l = 2 midfielder, l = 3 sides, l = 4 forward).
Parameters:
cjk : the value of criteria k of player;

M inR: minimum number of reserve players;
wkl : the value of criteria k in post l;
M axR: maximum number of reserve players;
Li : the post of player i;
Pi : probability of need to replace player i;
Rj : player accessibility characteristic (if player j is available Rj = 1, otherwise
Rj = 0).
Decision variable:
1, if there is need to replace player i with player j
xij =
0, otherwise.
The model is defined as follows,

max wkLi cjk pi xij (15)
k i j

s.t xij ≤ max R (16)
i j

xij ≤ Rj ∀j (17)
i

xij ≥ 1 ∀i, Li = 1, ∀i, Li = 2, ∀i, Li = 3, ∀i, Li = 4 (18)
j
xij ∈ {0, 1}. (19)
The proposed model is linear. The objective function (15) targets to maximize
the reserve team’s quality, which obtains from multiplying each player’s value in
desired post in probability of the need to replacement in that post. Constraint
(16) determines maximum number of allowable replacements. Constraint (17)
indicates the availability of players because it is possible that the player is not
in the team owing to Injury or deprivation of the game. Constraint (18) states
that each post should not be vacant. Constraint (19) indicates decision variable
type.
5 Uncertain Mathematical Model for Determining

the Reserve Combination
The player’s criteria are not constant always and they can change due to game
situation, fatigue and so on. To use the criteria in mathematical model, we use
belief degree and uncertain programming.
The parameters cjk , wkLi and pi are uncertain parameters with uncer-
tain
distributions
ϕjk , φkLi and ωi , respectively. We define f (x, C, W, P ) =
Cjk W kL Pi xij , whose inverse uncertainty distribution is γ −1 (x, α) =
k i j −1 −1 i −1
k i j ϕjk φkLi ωi xij .
To use uncertain programming the model changes as follows, Since γ −1 (x, α)
is the inverse uncertainty distribution of f (x, C, W, P ) , the mathematical model
is simplified as follows,
maxE [f (x, C, W, P )] (20)

s.t. xij ≤ M axR (21)
i j

xij ≤ Rj ∀j (22)
i

xij ≥ 1 ∀i, Li = 1, ∀i, Li = 2, ∀i, Li = 3, ∀i, Li = 4 (23)
j
xij ∈ {0, 1}. (24)

6 Experimental Results
The proposed model solved with IBM ILOG CPLEX-Optimizer v12. 3. We
choose a team with 24 players and solve the problem with these players. Due
to coach’s opinion the belief degrees are obtained. The results are presented in
appendix. In football games the maximum allowed replacements as 3 times. The
results of solving models using these data are examined in this section. The
main players’ number, which is determined according to coach’s opinion, is like
follows. This will be an input for model. The reserve team will be chosen due to
the input of model.
22 20 18 15 14 9 5 2 7 1
The reserve team, due to the model, is chosen as follows,

12 23 19 11 8 3
According to model suggested combination, average quality of the reserve will

be 83%. When only the coach’s opinion was important, and without using model,
the average quality of the reserve was 55% that brings a significant increase for
team. The reserve team, according to coach’s opinion, shows as follows,
10 24 19 11 4 3
Table 2 shows the comparison of objective functions in two cases which indi-
cates the model’s efficiency.
Table 2. Comparison of the first model in simulation environment
Number of f ∗ without the model f ∗ by the model Improvement of

simulation runs simulation runs
80 63% 91% 28%
7 Conclusion
The purpose of this article is to propose a model for choosing reserve team, so the
quality reduction will be minimum. We proposed a linear mathematical model.
To obtain uncertain parameters we use belief degree and uncertain programming.
For future works, authors can investigate main combination selection due to the
future games and simultaneous select reserve team.
Acknowledgement. This work was granted from Professor Baoding Liu, Tsinghua
University. The authors appreciate his financial and scientific supports. We also appre-
ciate Professor Mitsuo Gen for his guidance.
Appendix
See Tables 3, 4, 5, 6, and 7.
Table 3. The belief degrees of coach on all criteria
Number Defender Midfielder Sides Forwards Number Defender Midfielder Sides Forwards
of of
criteria criteria
C1 (0.9,1) (0.4,0.5) (0.6,0.75) (0.85,1) C10 (0.45,0.55) (0.6,0.75) (0.6,0.75) (0.85,1)
C2 (0.45,0.55) (0.6,0.75) (0.6,0.75) (0.85,1) C11 (0.85,1) (0.6,0.75) (0.45,0.55) (0.85,1)
C3 (0.85,1) (0.85,1) (0.85,1) (0.6,0.75) C12 (0.45,0.55) (0.45,0.85) (0.45,0.55) (0.6,0.75)
C4 (0.45,0.55) (0.6,0.75) (0.85,1) (0.6,0.75) C13 (0.4,0.5) (0.85,1) (0.4,0.5) (0.6,0.75)
C5 (0.6,0.75) (0.85,1) (0.85,1) (0.85,1) C14 (0.6,0.75) (0.45,0.55) (0.85,1) (0.85,1)
C6 (0.45,0.55) (0.9,1) (0.6,0.75) (0.85,1) C15 (0.3,0.45) (0.85,1) (0.6,0.75) (0.6,0.75)
C7 (0.45,0.55) (0.4,0.5) (0.6,0.75) (0.85,1) C16 (0.3,0.45) (0.6,0.75) (0.85,1) (0.85,1)
C8 (0.45,0.55) (0.6,0.75) (0.85,1) (0.85,1) C17 (0.6,0.75) (0.85,1) (0.6,0.75) (0.85,1)
C9 (0.45,0.55) (0.85,1) (0.9,1) (0.6,0.75) C18 (0.6,0.75) (0.85,1) (0.85,1) (0.6,0.75)
Table 4. The belief degrees of coach on all player
Number 1 2 3 4 5 6 7 8
of
players
P (0.6,0.75) (0.3,0.5) (0.1,0.3) (0.1,0.2) (0.3,0.5) (0.3,0.5) (0.1,0.3) (0.3,0.4)
Number 9 10 11 12 13 14 15 16
of
players
P (0.4,0.5) (0.5,0.6) (0.3,0.4) (0.55,0.8) (0.2,0.3) (0.3,0.5) (0.6,0.8) (0.3,0.45)
Number 17 18 19 20 21 22 23 24
of
players
P (0.3,0.4) (0.5,0.7) (0.1,0.3) (0.1,0.3) (0.3,0.5) (0.3,0.45) (0.3,0.5) (0.3,0.4)
Table 5. The belief degrees of coach on all criteria of different players
Number Number of players

of criteria
Defender
1 2 3 4 5 6 7 8
C1 (0.7,0.8) (0.7,0.8) (0.8,0.95) (0.6,0.7) (0.8,0.95) (0.45,0.6) (0.95,1) (0.7,0.8)
C2 (0.45,0.6) (0.95,1) (0.8,0.95) (0.25,0.4) (0.35,0.5) (0.7,0.8) (0.7,0.8) (0.45,0.6)
C3 (0.7,0.8) (0.7,0.8) (0.8,0.95) (0.7,0.8) (0.8,0.95) (0.35,0.5) (0.5,0.6) (0.7,0.8)
C4 (0.7,0.8) (0.7,0.8) (0.45,0.6) (0.25,0.4) (0.7,0.8) (0.95,1) (0.45,0.6) (0.7,0.8)
C5 (0.7,0.8) (0.45,0.6) (0.45,0.6) (0.7,0.8) (0.7,0.8) (0.95,1) (0.7,0.8) (0.7,0.8)
C6 (0.45,0.6) (0.7,0.8) (0.7,0.85) (0.7,0.8) (0.6,0.7) (0.6,0.7) (0.25,0.4) (0.7,0.8)
C7 (0.45,0.6) (0.6,0.7) (0.6,0.7) (0.25,0.4) (0.6,0.7) (0.7,0.8) (0.7,0.8) (0.7,0.8)
C8 (0.7,0.8) (0.45,0.6) (0.35,0.5) (0.8,0.95) (0.5,0.6) (0.6,0.7) (0.7,0.8) (0.95,1)
C9 (0.6,0.7) (0.45,0.6) (0.5,0.6) (0.6,0.7) (0.25,0.4) (0.7,0.8) (0.7,0.8) (0.95,1)
C10 (0.45,0.6) (0.7,0.8) (0.25,0.4) (0.45,0.6) (0.25,0.4) (0.45,0.6) (0.45,0.6) (0.7,0.8)
C11 (0.7,0.8) (0.7,0.8) (0.95,1) (0.95,1) (0.7,0.8) (0.95,1) (0.7,0.8) (0.7,0.8)
C12 (0.25,0.4) (0.7,0.8) (0.25,0.4) (0.45,0.6) (0.45,0.6) (0.7,0.8) (0.25,0.4) (0.6,0.7)
C13 (0.7,0.8) (0.7,0.8) (0.45,0.6) (0.7,0.8) (0.95,1) (0.95,1) (0.7,0.8) (0.7,0.8)
C14 (0.7,0.8) (0.7,0.8) (0.45,0.6) (0.95,1) (0.7,0.8) (0.95,1) (0.7,0.8) (0.7,0.8)
C15 (0.7,0.8) (0.7,0.8) (0.7,0.8) (0.45,0.6) (0.7,0.8) (0.45,0.6) (0.45,0.6) (0.45,0.6)
C16 (0.7,0.8) (0.25,0.4) (0.45,0.6) (0.6,0.7) (0.45,0.6) (0.6,0.7) (0.5,0.6) (0.6,0.7)
C17 (0.45,0.6) (0.7,0.8) (0.45,0.6) (0.25,0.4) (0.7,0.8) (0.95,1) (0.7,0.8) (0.95,1)
C18 (0.7,0.8) (0.7,0.8) (0.45,0.6) (0.7,0.8) (0.45,0.6) (0.7,0.8) (0.7,0.8) (0.6,0.7)

of
criteria
Midfielder
9 10 11 12 13 14 15
C1 (0.7,0.8) (0.35,0.5) (0.7,0.8) (0.45,0.6) (0.25,0.4) (0.6,0.7) (0.95,1)
C2 (0.8,0.95) (0.7,0.8) (0.95,1) (0.35,0.5) (0.7,0.8) (0.95,1) (0.7,0.8)
C3 (0.8,0.95) (0.7,0.85) (0.7,0.8) (0.95,1) (0.7,0.8) (0.45,0.6) (0.7,0.8)
C4 (0.7,0.8) (0.7,0.8) (0.45,0.6) (0.7,0.8) (0.95,1) (0.95,1) (0.7,0.8)
C5 (0.95,1) (0.6,0.7) (0.8,0.95) (0.7,0.8) (0.45,0.6) (0.5,0.6) (0.45,0.6)
C6 (0.7,0.8) (0.45,0.6) (0.95,1) (0.7,0.8) (0.7,0.8) (0.95,1) (0.8,0.95)
C7 (0.45,0.6) (0.7,0.8) (0.95,1) (0.7,0.8) (0.7,0.8) (0.45,0.6) (0.95,1)
C8 (0.7,0.8) (0.7,0.8) (0.45,0.6) (0.45,0.6) (0.95,1) (0.95,1) (0.7,0.8)
C9 (0.7,0.8) (0.7,0.8) (0.95,1) (0.7,0.8) (0.95,1) (0.45,0.6) (0.7,0.8)
C10 (0.7,0.8) (0.7,0.8) (0.45,0.6) (0.7,0.8) (0.95,1) (0.95,1) (0.7,0.8)
C11 (0.7,0.8) (0.7,0.8) (0.7,0.8) (0.95,1) (0.45,0.6) (0.45,0.6) (0.7,0.8)
C12 (0.6,0.7) (0.5,0.6) (0.5,0.6) (0.6,0.7) (0.6,0.7) (0.45,0.6) (0.7,0.8)
C13 (0.7,0.85) (0.6,0.7) (0.6,0.7) (0.6,0.7) (0.7,0.8) (0.7,0.8) (0.45,0.6)
C14 (0.45,0.6) (0.45,0.6) (0.7,0.8) (0.45,0.6) (0.8,0.95) (0.6,0.7) (0.6,0.7)
C15 (0.6,0.7) (0.8,0.95) (0.6,0.7) (0.6,0.7) (0.8,0.95) (0.7,0.85) (0.6,0.7)
C16 (0.7,0.8) (0.6,0.7) (0.6,0.7) (0.7,0.8) (0.7,0.85) (0.95,1) (0.7,0.8)
C17 (0.7,0.8) (0.45,0.6) (0.7,0.8) (0.7,0.8) (0.95,1) (0.95,1) (0.7,0.8)
C18 (0.8,0.95) (0.95,1) (0.45,0.6) (0.7,0.85) (0.7,0.8) (0.45,0.6) (0.7,0.8)

of criteria
Sides Forward
16 17 18 19 20 21 22 23 24
C1 (0.7,0.8) (0.7,0.8) (0.8,0.95) (0.7,0.8) (0.95,1) (0.8,0.95) (0.95,1) (0.95,1) (0.7,0.8)
C2 (0.95,1) (0.95,1) (0.8,0.95) (0.7,0.8) (0.95,1) (0.8,0.95) (0.6,0.7) (0.7,0.8) (0.7,0.8)
C3 (0.7,0.8) (0.95,1) (0.7,0.8) (0.95,1) (0.7,0.8) (0.7,0.8) (0.95,1) (0.7,0.8) (0.95,1)
C4 (0.45,0.6) (0.7,0.8) (0.7,0.8) (0.7,0.8) (0.95,1) (0.7,0.8) (0.7,0.8) (0.95,1) (0.7,0.8)
C5 (0.95,1) (0.7,0.8) (0.95,1) (0.7,0.8) (0.7,0.8) (0.8,0.95) (0.6,0.7) (0.7,0.8) (0.7,0.8)
C6 (0.7,0.85) (0.7,0.8) (0.95,1) (0.6,0.7) (0.8,0.95) (0.6,0.7) (0.7,0.8) (0.7,0.8) (0.95,1)
C7 (0.7,0.8) (0.7,0.8) (0.95,1) (0.95,1) (0.7,0.8) (0.7,0.8) (0.95,1) (0.7,0.8) (0.95,1)
C8 (0.45,0.6) (0.95,1) (0.7,0.8) (0.8,0.95) (0.95,1) (0.7,0.8) (0.8,0.95) (0.8,0.95) (0.7,0.8)
C9 (0.95,1) (0.95,1) (0.7,0.8) (0.95,1) (0.7,0.8) (0.7,0.8) (0.95,1) (0.7,0.8) (0.95,1)
C10 (0.95,1) (0.7,0.8) (0.7,0.8) (0.7,0.8) (0.95,1) (0.95,1) (0.45,0.6) (0.95,1) (0.45,0.6)
C11 (0.95,1) (0.7,0.8) (0.45,0.6) (0.7,0.8) (0.45,0.6) (0.7,0.8) (0.7,0.8) (0.7,0.8) (0.7,0.8)
C12 (0.25,0.4) (0.7,0.8) (0.7,0.8) (0.8,0.95) (0.7,0.85) (0.7,0.85) (0.5,0.6) (0.7,0.85) (0.6,0.7)
C13 (0.7,0.8) (0.45,0.6) (0.6,0.7) (0.8,0.95) (0.8,0.95) (0.7,0.85) (0.7,0.8) (0.95,1) (0.7,0.8)
C14 (0.8,0.95) (0.95,1) (0.7,0.8) (0.95,1) (0.95,1) (0.7,0.8) (0.7,0.8) (0.95,1) (0.7,0.8)
C15 (0.7,0.8) (0.95,1) (0.7,0.8) (0.7,0.8) (0.95,1) (0.95,1) (0.7,0.8) (0.7,0.8) (0.95,1)
C16 (0.7,0.85) (0.7,0.8) (0.8,0.95) (0.7,0.8) (0.95,1) (0.7,0.8) (0.7,0.8) (0.7,0.8) (0.95,1)
C17 (0.95,1) (0.7,0.8) (0.7,0.8) (0.7,0.8) (0.45,0.6) (0.95,1) (0.7,0.8) (0.7,0.8) (0.95,1)
C18 (0.95,1) (0.45,0.6) (0.7,0.8) (0.45,0.6) (0.7,0.8) (0.45,0.6) (0.95,1) (0.85,0.95) (0.7,0.8)
References
1. Abdulwhab A, Billinton R et al (2004) Maintenance scheduling optimization using a
genetic algorithm (GA) with a probabilistic fitness function. Electr Power Compon
Syst 32(12):1239–1254
2. Ahmed F, Deb K, Jindal A (2013) Multi-objective optimization and decision making
approaches to cricket team selection. Appl Soft Comput 13(1):402–414
3. Liu B (2002) Theory and Practice of Uncertain Programming. Physica-Verlag, Hei-
delberg
4. Liu B (2009) Some research problems in uncertainty theory. J Uncertain Syst 3(1):3–
10
5. Liu B (2009) Uncertain entailment and modus ponens in the framework of uncertain
logic. J Uncertain Syst 3(4):243–251
6. Liu B (2010) Uncertainty Theory: A Branch of Mathematics for Modeling Human
Uncertainty. DBLP
7. Liu B (2015) Uncertainty theory. Stud Comput Intell 154(3):1–79
8. Tavana M, Azizi F et al (2013) A fuzzy inference system with application to player
selection and team formation in multi-player sports. Sport Manag Rev 16(1):97–110
9. Trninić S, Papić V et al (2008) Player selection procedures in team sports games.
Acta Kinesiologica 2(1):24–28
A Bayesian-Based Co-Cooperative Particle
Swarm Optimization for Flexible Manufacturing
System Under Stochastic Environment
Lu Sun1 , Lin Lin1,2,3(B) , and Haojie Li1

1
lin@dlut.edu.cn
2
3
Dalian, People’s Republic of China
Abstract. In recent years, flexible scheduling attracted considerable

attention, motivated by both important practical issues and interest-
ing research problems, especially under stochastic environment. In this
paper we discuss a stochastic scheduling problem whose scale arranges
from small to large. The objective in the schedule is to minimize the
processing time over all of the jobs. However, stochastic environment
will bring more difficulty especially for large scale because the sudden
change of processing time for each operation may break the current opti-
mal solution and lose efficiency. So, we propose a Bayesian network based
particle swarm optimization (BNPSO) for solving this stochastic schedul-
ing problem. Firstly, we use new framework named co-cooperative (CC)
evolutionary framework which decompose all decision variables into sev-
eral small group containing part of decision variables to overcome large
scale problems. And then, BNPSO adjust the group scene according
to their interactive relationships based on Bayesian network structure.
Meanwhile, importing self-adaptive mechanism for parameters in order
to satisfy the stochastic environment. Some practical test instances will
demonstrate the effectiveness and efficiency of the proposed algorithm.
Keywords: Co-cooperative framework · Bayesian network structure ·

Stochastic environment · Flexible manufacturing system
1 Introduction
Over the past sixty years, a great number of researches have been conducted
on the job shop scheduling problem (JSP), which is a branch of the scheduling
problem and highly popular in the manufacturing industry. JSP is a classical
combinatorial optimization problem, and it is recognized as NP-hard under the
precedence and resource constraints [2,5]. The flexible job shop scheduling prob-
lem (FJSP) is a generalization of the classical JSP for flexible manufacturing
systems (FMSs).
A Bayesian-Based Co-Cooperative Particle Swarm Optimization 1429
In real-world problem, most of the scheduling problems are the stochastic

scheduling problems. As one of the newest issues, more and more attentions
focused on the problem with random processing time. As a result, in the last sev-
eral decades, a significant amount of results has been achieved on the stochastic
scheduling problems [3]. Moghaddam et al. [13] proposed a hybrid method using
a neural network approach and a simulated annealing algorithm in two stages,
in order to produce the optimal/near-optimal solution. Liu, Wang and Jin [9]
presented an approach named PSOSAHT that is hybrid with simulated anneal-
ing (SA) and hypothesis test (HT), for stochastic flow shop scheduling with
uncertain processing time. Zhou, Nee and Lee [17] developed an ant colony opti-
mization algorithm (ACO) with different levels of machine utilizations, process-
ing time distributions, and performance measures. Lei [6] developed an efficient
decomposition-integration genetic algorithm (DIGA) to minimize the maximum
fuzzy completion time. For solving realistic scheduling problems, intelligent man-
ufacturing planning and scheduling based on meta-heuristics, such as GAs and
ACOs have become common techniques for finding satisfactory solutions within
reasonable computational times. However, as the scale of problem increases,
the search space will increase exponentially, consequently and the probability of
obtaining optimal solution will be reduced. To overcome it, coevolution frame-
work was proposed firstly in 1994, which represents and solves more complex
problems by explicitly modeling the coevolution of cooperating species [11]. Up
to now, there exist various typical grouping mechanisms. Bergh et al. [14] pro-
posed a cooperative coevolutionary approach which used the fixed group size
to solve function problems. Li, et al. [10] thought more frequent random group-
ing mechanism could have better performance and proposed a cooperative co-
evolutionary framework and employs a technique called random grouping in
order to group interacting variables into one subcomponent for large scale con-
tinuous global function optimization. Yao et al. [8] balanced the random group-
ing with fixed grouping, and proposed a set-based grouping mechanism which
group set was given in advanced. When the global optimal solution does not
be optimized, to select a new group size randomly immediately. However, all of
the grouping strategies described so far use a predefined and fixed group size.
For example, random grouping decomposes an n-dimensional problem into k s-
dimensional problems. A major drawback of these techniques is that the user
needs to specify a value for either k or s or a set of potential s values and the
given values will effect the performance directly. Another drawback of this mul-
tilevel scheme is that once an s value is chosen, the decision variables are divided
into a set of equally sized subcomponents. It is unlikely that in most real-world
problems the sizes of interacting groups will be not equal. Hence, it is desir-
able that a decomposition strategy can automatically determine the number of
subcomponents and their sizes.
In recent years, probabilistic graphical models (PGMs), which are used to
represent the relationship among the discrete variables, have attracted a growing
attention. PGMs can be mainly categorized into two categories: Bayesian net-
works (directed) and Markov networks (undirected) [15]. In Bayesian network,
1430 L. Sun et al.
nodes represent variables and arcs shows the relationship between two linked
nodes. So, if we can use Bayesian network to shown the relationships among the
variables, we will no longer need to determine the number of subcomponents
and their sizes. In the application of Bayesian network, Bayesian optimization
algorithm (BOA) is widely used in scheduling, such as, resource assignment, task
scheduling and so on. In 2011, Yang et al. [15] present a novel scheduling algo-
rithm based on BOA for heterogeneous computing environments. Li et al. [7]
proposed a Bayesian optimization algorithm to solve task assignment problems
in heterogeneous computing systems. Hao et al. [4] proposed a novel coopera-
tive Bayesian optimization algorithm (CoBOA) to overcome the challenges men-
tioned in the field of multiple resources scheduling problem (MRSP). Inspired
by BOA, in this paper, we proposed a learning-based grouping mechanism in
which we apply BOA to get the optimal Bayesian network structure for find-
ing the potential relationships among variables and then to divide the variables
according to relationships (network structure).
The remainder of this paper is organized as follows: Sect. 2 gives the formu-
lating process of S-fJSP and mathematical programming model; Sect. 3 describes
the BNPSO combined BOA Particle swarm algorithm (PSO) with multiple sub-
populations in detail; Sect. 4 gives kinds of detailed numerical experiments and
computational results; finally, Sect. 5 comes to the conclusion of the paper.
In order to solve a stochastic flexible job-shop scheduling problem (S-fJSP), it
assumes that the probability distribution of the processing time is known in
advance. In this paper, we use a pure integer programming model to transmute
the processing times in terms of stochastic variable. The S-JSP can be formu-
lated as an extended version of fJSP. Difference to the conventional fJSP, each
operation oij is carried out under uncertain random disturbance with pre-given
expected valued E [pij ] and variance vij , where pij is the processing time of Oij
on the machines. The distribution of the variance can be predicted from the
experimental data such as normal distribution, uniform distribution and expo-
nential distribution etc. It may also consist of several assumptions as follows:
A1. Each machine processes only one job at a time.

A2. Each job is processed on one machine at a time.
A3. For each operation, the probability distribution of processing time and
machine assignment are known in advance, but the processing time is sto-
chastic.
A4. There are no precedence constraints among operations of different jobs.
A5. Operations cannot be interrupted.
A6. Neither release times nor due dates are specified.
The S-fJSP has already been confirmed as one of the NP-hard combinatorial
problems. There are N jobs and M machines to be scheduled; furthermore, each
job is composed of a set of operations and the operation order on machines is
pre-specified. Each operation is characterized by the required machine and the

processing time [5].
Index:
i,h: indices of tasks, i, h = 1, 2, · · · , n;

j: indices of resources, index, j = 1, 2, · · · , m;
k,g: indices of operations, k, g = 1, 2, · · · , ni .
Parameter:
n: number of total tasks;

m: number of total resources;
ni : number of operations for task i;
nik : k th operation of task i;
tikj : processing time of Oik processed on j th resource;
Aik : the available resource set of Oik ;
cik : the completed time of Oik .
Decision variables:

1, if oik is performed on machine j
xikj =
0, otherwise.
The objective function is minimizing the makespan, as follows:

min E [CM ] = E max max max ξcikj . (1)
i k i
For the process of scheduling, the operations are not allowed to terminate
until it is completed and for each resource. The resource constraints are described
as follows in Eq. (2):
s.t. ξcik − ξci(k−1) ≥ ξtikj × xikj k = 2, · · · , ni , ∀i, j. (2)
For each task, it consists of several operations and each operation is allo-
cated different machine, although each operation can be processed on different
resources, the operation sequence within each task must be observed. Note that
the two constraints (3) and (4) only one need to be satisfied at the same time.
(ξchg − ξci(k−1) − ξtikj ) × xhgj ≥ 0 ∀j, g, h, (3)

(ξcik − ξchg − ξtikj )xikj ≥ 0 ∀i, j, g, h. (4)
Besides the constraints mentioned above, the constraint (5) guarantees

machine allocation that for each operation can only process on one machine
from machine set at one time.

xikj = 1 ∀i, k, j. (5)
xijk ∈Aik
1432 L. Sun et al.
For each resource, it contains only two cases: chosen or not chosen. And The
constraint (6) and the constraint (7) gives the restriction constraint of decision
variables.
xikj ∈ {0, 1} ∀i, k, j, (6)

ξcik > 0 ∀i, k. (7)
3 Algorithm Design
BNPSO starts by random grouping all decision variables encoded by real num-
bers into several subcomponents, each subcomponent contains s variables shown
in Fig. 1. Then, to optimize the global optimal solution based on the co-
evolutionary framework. In the process of evolution, to record down the data
set for BN learning according to the value variation of each decision variable.
Each 50 iterations, update the structure of BN follows BOA and adjust the
grouping scene. Meanwhile, self-adjust the parameters used in evolution algo-
rithms. Several key parts of BNPSO listed in the following subsections in detail.
Fig. 1. The network based for decoding the operations
3.1 Encoding and Decoding
PSO is a typical met-heuristics algorithm in EAs, so the encoding and decoding

methods will effect the performance partly. We choose two-section real priority
encoding method in order to explore larger solution space. An example of rep-
resentation and decoding network for one particle with 5 operations is shown
Fig. 2.
The searching space is created in a 2 × n × m search space for n jobs on
m machines. Each particle consists of 2 × n × m variables and is represented
by real number. As mentioned above, S-fJSP can be considered as two sub-
problems: operation sequences and machine allocation. So the first section of
particle presents the operations; the second part presents the machines allocation
Fig. 2. An example of representation and decoding network for one particle
corresponding to each operation, we use the network based decoding method

(shown in Fig. 3(b)) to get S-fJSP solution from the particle which can confirm
the solutions are all feasible solutions.
PSO uses the following formulas to update velocity and position in each
generation t, respectively. The initial values of velocity and position are set
randomly with the lower bound and upon bound which are given in advance. The
ω presents inertia weight, rand1 and rand2 are random value within [0,1]. The
bigger ω means new velocity will mainly be effected by search history of particle
itself; the smaller presents new velocity will mainly be effected by search history
of the whole population. The parameters used in PSO is adapted according to
the methods studied in the paper [10].
vt (t + 1) = ωvi (t) + c1 rand1 [pbest (t) − xi∂ (t)] + c2 rand2 [lbest (t) − xi (t)], (8)
xt (t + 1) = xt (t) + vt (t + 1). (9)
In order to explore larger solution space, the velocity and position update
model with both Cauchy and Gaussian distributions was proposed in [15]. This
model updates the position directly through personal best: pbest , local best: lbest .
pbest means the best particle in the searching history of each particle, lbest means
the best one among the ith particle, (i − 1)th particle and (i + 1)th particle each
iteration. rand and p are used for deciding which formula to be selected, rand is
a random value within [0,1] and p is a given value at initial stage. C(1) means
the Cauchy distribution value with mean value 1; N (0, 1) means the normal
distribution value with mean value 0 and the standard variance 1. The position
updating formula is shown as follows, each time, generate rand, and compare
rand with p to decide which equation to be chosen. If rand ≤ p, the position will
update follow the first formula, otherwise, follow the second formula.

pbest (t) + C(10 |pbest (t) − lbest (t)| , if rand ≤ p
x(t + 1) = (10)
lbest (t) + N (0, 1) |pbest (t) − lbest (t)| , otherwise.
3.2 Bayesian-Based Grouping Mechanism

(1) Bayesian network structure learning
As mentioned above, BN structure learning is also a NP hard problem, if we use
the traditional method to learn its structure, it will make our algorithm be too
1434 L. Sun et al.
heavy, so we use another evolutionary algorithm. Inspired by K2 which is the

most classical and typical algorithm for learning BN structure, learning structure
can be divided into two parts: the node sequence and the edges. Hence, we still
use two-section encoding, in which the first section represents node sequence, and
the second section represents the edge. We give an example with 4 variables in
one group in Fig. 3(a). We can get variable sequence (descending order): x4 − >
x2 − > x1 − > x3 , then get the edges by transferring the real values in the second
part particle into binary value for the first 4 (equals to the number of variables
s in current group) values, then to randomly select from the got binary values
for other 2 (equals to s × (s − 1)/2 − s) binary values, shown in Fig. 3(b).
Fig. 3. The particle used for learning BN structure
Then, we decode particle into a BN by a strictly upper triangular matrix

which means to put the binary sequence in the shadow position successively
from the bottom to the end in Fig. 4. In which, 1 represents the variable in
column is the parent of the variable in row. Each generation, we can get one
network structure, after m generation, we can m candidate networks. A simple
BN is shown in Fig. 4, it presents that the network constructed by particle in
Fig. 3(a). Then, we adjust the grouping situation depends on BN structure, x3 is
a dependent variable, hence, we put it to another random group, after adjusting,
this group only contain 3 variables temporary until other variables are put in.
Fig. 4. A simple BN structure example get from a particle

(2) Sampling train data

We record down the variance of each decision variable of particle between (i+1)th
generation and the ith generation. Because, the variance of each variable of global
best particle shows the relationship when global best particle evaluates.
Fig. 5. Sampling the training data
If the value of variable becomes bigger, then to record down 1, otherwise,

if the value becomes smaller or keeps invariant, to record down 0. Then we
can get the training data set for learning BN structure, shown in Fig. 5. After
evaluating the network structure, we can find a optimal network with BDe score.
Then, we adjust the group by putting the discrete variables into the other groups
randomly. The whole process of BNPSO is shown as follows.
4 Experiment and Results

To prove the efficiency of our proposed algorithm logically, we do two kinds
of experiments for S-fJSP on 4 scales of data: 5 × 5 (50 dims), 10 × 10 (200
dims), 15 × 15(450 dims) and 20 × 20 (800 dims). The dimensions are decided by
encoding method mentioned in Sect. 3. For example, 5×5 means this benchmark
has 5 jobs, 5 machines and 5 (the same to the number of machines) operations.
These data instances can be accessed on OR-Library by Dirk et al. For the
uncertainty processing time, the processing time of operation oij may equally
take any real value from the uniform distribution U (lower, upon) where lower is
80% of static processing time and upon is 120% of static processing time. In order
to ensure the reliability of the experiments and avoid the shock of results, all
experiments are repeated 30 times to get the mean value and calculate variance.
The machine environment of running experiments is Intel(R) Core(TM) i7-4770
CPU @ 2.2 GHz, 16 GB.
4.1 Experiment Settings

All compared algorithms contains: classical genetic algorithm (GA) [1], binary
genetic algorithm (bGA) [1], differential evolution (DE) [12], classical parti-
cle swarm optimization (PSO) [16], self-adaptive neighbourhood search DE
1436 L. Sun et al.
(SaNSDE) [16] and cooperative coevolutionary particle swarm optimization

(CCPSO) [8]. The values of the parameters c1 , c2 , and w used in PSO were
selected based on [8]. The crossover and mutation probability are set to 0.5 [1].
The initial parameters of DE and SaNSDE were set based on [16].
Besides above parameters, there still exist common parameters such as pop-
ulation size, initial group size and the max iterations for all algorithms, they are
shown in Table 1.
Table 1. Common parameters
Data Pop size Initial group size Max iteration

5×5 100 10 1000
10 × 10 1000 10 5000
15 × 15 1000 50 5000
20 × 20 5000 100 10000
4.2 Results and Discussion
(1) Bayesian-based grouping mechanism

In order to prove the effectiveness of Bayesian-based grouping mechanism, we did
experiments focus on fJSP and S-fJSP. And we used PSO as the basic algorithm,
doing experiment on classic PSO (classical), PSO with fixed grouping (fixed),
set-based grouping PSO (set) and Bayesian-based (Bayes) grouping PSO respec-
tively. The objective is to minimize the mean value of makespan, the result are
shown in Table 2. We can get the conclusion from the table that, Bayesian-based
grouping mechanism has better performance no matter for fJSP or S-fJSP on
mean value of makespan.
(2) Effectiveness of BNPSO
We compared BNPSO GA, bGA, PSO, DE, CCPSO and SaNSDE on the mean
value and variance of makespan, the results are shown in Table 3. Firstly we can
get the conclusion that, for encoding methods, real number encoding method
(GA) performances better than the binary encoding method (bGA). The reason
is that real number can use only one number to represent one variable, hence,
for equal-length particle, real number encoding particle has larger search space.
Secondly, for the basic algorithms which contains: GA, bGA, DE, PSO, we can
find that GA performs better than PSO in most time, however, due to the
characters of GA, co-evolution GA performs bad unfortunately. Compared to
improved algorithms: CCPSO and SaNSDE, we get the conclusion that BNPSO
performs better than them on most evaluation indexes.
Table 2. Results of fJSP and S-fJSP for grouping mechanism
FJSP Target Classical Fixed Set Bayes

5×5 Mean 245 282 212 157
Target 273.4 340.47 279.86 217.2
10 × 10 Mean 742 631 694 569
Target 748.9 686.47 766.86 696.1
15 × 15 Mean 1209 1080 1074 979
Target 1277.4 1174.37 1232.7 1132.63
20 × 20 Mean 1784 1715 1677 1372
Target 1794.83 418.9 1774.96 1555.06
S-fJSP Target Classical Fixed Set Bayes
5×5 Mean 273.43 261.53 279.86 217.2
Target 831.64 768.32 708.048 774.35
10 × 10 Mean 748.966 730.523 766.86 696.1
Target 468.4322 1489.65 2105.64 3148.75
15 × 15 Mean 1277.4 1210.4 1232.7 1132.63
Target 1886.37 2832.33 4389.41 5596.76
20 × 20 Mean 1794.83 1674.83 1774.96 1555.06
Target 216.672 1198.022 4211.4 6958.72
Table 3. Mean value and variance value of makespan for all algorithms
Scale Target GA BinaryGA DE PSO SaNSDE CCPSO BNPSO

5×5 Mean 291.5 305.76 314.1 273.43 310.66 279.86 217.2
Variance 573.25 1068.44 318.75 831.64 203.2 708.048 774.35
10 × 10 Mean 778.4 834.06 809.9 748.966 805.7 766.86 696.1
Variance 2028.1 1905.72 907.29 468.4322 1102.74 2105.64 3148.75
15 × 15 Mean 1301.5 1365.8 1316.06 1277.4 1238.166 1232.7 1132.63
Variance 1242.71 3612.49 304.46 1886.37 548.595 4389.41 5596.76
20 × 20 Mean 1832.46 1869.266 1816.96 1794.83 1778.533 1774.96 1555.06
Variance 2841.64 2370.02 705.09 216.672 1170.51 4211.4 6958.72
5 Conclusion
This paper presents an effective BNPSO, which solves the S-fJSP with the uncer-
tainty of processing time. It minimized the expected average makespan within
reasonable time. With the framework of the proposed BNPSO, we construct BN
structure according to the data showing the relationship among variables and
adjust the group scene based on the independence shown by structure. We first
proved the effectiveness of BN-based grouping mechanism, and then, we com-
pared proposed algorithm with other compared famous algorithms, the results
shown that, BNSPO performed better than other algorithms. In our future work,
1438 L. Sun et al.
we will extend BNPSO to adapt to real case study based on multiobjective sto-
chastic flexible job- shop models (moS-JSP).

Foundation of China under Grant 61572100 and in part by the Grant-in-Aid for Sci-
References
1. Gao J, Sun L, Gen M (2008) A hybrid genetic and variable neighborhood descent
algorithm for flexible job shop scheduling problems. Comput Oper Res 35(9):2892–
2907
2. Gen M, Cheng R (2000) Genetic algorithms and engineering optimization. John
Wiley, NewYork
3. Gu J, Gu X, Gu M (2009) A novel parallel quantum genetic algorithm for stochastic
job shop scheduling. J Math Anal Appl 355(1):63–81
4. Hao X, Chen X et al (2011) Cooperative Bayesian optimization algorithm: a novel
approach to simultaneous multiple resources scheduling problem. In: Second inter-
national conference on innovations in bio-inspired computing and applications, pp
212–217
5. Lawler EL, Lenstra JK, Kan AHGR, Shmoys DB (1989) Sequencing and schedul-
ing: algorithms and complexity. Handb Oper Res Manag Sci 4(2):445–522
6. Lei D (2010) A genetic algorithm for flexible job shop scheduling with fuzzy process-
ing time. Int J Prod Res 48(10):2995–3013
7. Li J, Zhang JQ et al (2014) Team of Bayesian optimization algorithms to solve task
assignment problems in heterogeneous computing systems. In: IEEE international
conference on systems, man and cybernetics, pp 127–132
8. Li X, Yao X (2012) Cooperatively coevolving particle swarms for large scale opti-
mization. IEEE Trans Evol Comput 16(99):1
9. Liu B, Wang L, Jin YH (2005) Hybrid particle swarm optimization for flow shop
scheduling with stochastic processing time. In: Proceedings of the Computational
Intelligence and Security, International Conference, CIS 2005, Xi’an, China, 15–19
December 2005, pp 630–637
10. Omidvar MN, Li X et al (2010) Cooperative co-evolution for large scale optimiza-
tion through more frequent random grouping. In: Evolutionary Computation, pp
1–8
11. Potter MA, Jong KAD (1994) A cooperative coevolutionary approach to function
optimization. Lecture Notes in Computer Science, vol 866, pp 249–257
12. Price K, Storn RM, Lampinen JA (2005) Differential evolution: a practical app-
roach to global optimization. Natural computing series, vol 141(2), pp 1–24
13. Tavakkoli-Moghaddam R, Jolai F et al (2005) A hybrid method for solving sto-
chastic job shop scheduling problems. Appl Math Comput 170(1):185–206
14. Van den Bergh F, Engelbrecht AP (2004) A cooperative approach to particle swarm
optimization. IEEE Trans Evol Comput 8(3):225–239
15. Yang J, Xu H et al (2011) Task scheduling using Bayesian optimization algorithm
for heterogeneous computing environments. Appl Soft Comput 11(4):3297–3310
16. Yang Z, Tang K, Yao X (2008) Self-adaptive differential evolution with neighbor-
hood search. In: Evolutionary Computation, pp 1110–1116
17. Zhou R, Nee A, Lee H (2009) Performance of an ant colony optimisation algorithm
in dynamic job shop scheduling problems. Int J Prod Res 47(11):2903–2920
Exchange Rate Movements,
Political Environment and Chinese Outward FDI
in Countries Along “One Belt One Road”
Wenjing Zu1 and Haiyue Liu2(B)

1
School of Economics, Zhejiang University,
Hangzhou 310013, People’s Republic of China
2
seamoon@vip.sina.com
Abstract. This paper explores the role of political environment (PE)

and exchange rate (ER) movements in determining China’s outward for-
eign direct investment (OFDI) for a period of 2003–2015. It includes 65
countries along “OBOR” (One Belt One Road) to build up a panel data
base. ER level, ER volatility and Political Environment are examined
in relation to Chinese OFDI in these “OBOR” countries. The empiri-
cal results show that RMB appreciation is associated with higher Chi-
nese OFDI flow but Chinese OFDI is not sensitive to ER volatility and
PE instability. Considering the growing importance of the “OBOR” ini-
tiative, we believe that our research results shed light on the trend of
Chinese OFDI.
Keywords: Chinese outward foreign direct investment · “One Belt One

Road” Initiative · Exchange rate · Political environment
1 Introduction
Since China adopted its “going out” policy in 2001, its OFDI flows have grown
rapidly, reaching more than $100 billion in 2013. As an upgraded strategy for
foreign investment, the “OBOR” was initiated in 2013 to better serve China’s
“going out” policy. In the area defined by this “OBOR” initiative, more than
60 countries have expressed their interest in cooperating with this program.
In addition, the China-led Asian Infrastructure Investment Bank (AIIB) was
created to provide professional, efficient financial support for the tremendous
infrastructure needs under the “OBOR” initiative. At present, a total of 57
countries covering five continents have been approved as founding members of
this international financial institution. As an Asian regional multilateral devel-
opment agency, AIIB will play a constructive role in strengthening economic
growth engine of infrastructure construction, improving the capital utilization
efficiency and raising the regional development level.

1440 W. Zu and H. Liu
However, it is becoming increasingly clear that the global market is full of

uncertainties that can undermine even carefully laid plans. The risks that Chi-
nese firms are facing when they are encouraged to invest abroad should not
be overlooked. According to Wang [26], the major risks that Chinese firms deal
with during outward investment are political and financial risks, which accounted
more than 60% of the total risks. The risks are mainly influenced by the political
environment and exchange rate movements which in turn affect the volume and
regional allocation of Chinese OFDI. Therefore, this paper examines these two
major risks and their influences on Chinese OFDI. The panel dataset includes
65 countries along the “OBOR”. This paper uses an index that is composed of 6
weighted indicators covering both political and social attributes to measure PE
risk. The ER risks are calculated by measuring ER level and volatility.
The contribution of this paper is twofold. Firstly, little paper discussed
China’s OFDI in “OBOR” countries, let alone analyzed the impact of finan-
cial and political risks. Unlike the previous studies which are mainly focused on
the Chinese FDI flows to developed countries or emerging economies, our study
looks at the countries along the “OBOR” diversified in South and Central Asia,
Central and Eastern Europe and Africa. Therefore, this paper contributes to
the existing empirical literature on Chinese OFDI in a new region. Secondly,
our result reveals that Chinese OFDI seems not to be affected by PE and ER
volatility, which is not in line with the previous literature.
The remainder of this paper is organized as follows: Sect. 2 reviews the
relevant theoretical and empirical literature. Section 3 provides a descriptive
overview of Chinese OFDI in the “OBOR” countries. Section 4 describes the
data. Section 5 presents the empirical specification and results. Section 6 con-
cludes and relates policy implications.
2 Literature Review
Theoretical and empirical studies have looked at the foreign investment behavior
of multinationals and have identified exchange rate movements and political
environment as important determinants of outward FDI. We conclude these
studies as follows.
2.1 The Impact of Exchange Rate Movements on FDI
The interest in the impacts of exchange rate and its volatility on international
capital flows such as FDI is growing among policy makers. As for empirical
research, quite a lot of scholars try to investigate the relationship between
exchange rate and FDI by taking developed countries and newly emerging coun-
tries as sample, which is critical to the formulation of FDI polices.
With regard to the exchange rate level, one strand of the literature empha-
sizes the positive correlation between the appreciation of source country currency
and FDI outflows based on the relative wealth effect, the expectation of future
profitability and capital market imperfection. Schmidt and Broll [23] showed
Political Environment and Chinese Outward FDI 1441
that the real exchange rate level has a positive effect on outward FDI flows in
nine industries from the US. Takagi and Shi [24] used the panel data of Japanese
FDI flows to nine dynamic Asian economics during 1978–2008 and found the cur-
rent depreciation of host country currencies significantly increased FDI inflows
from Japan. However, another strand emphasizes the ambiguous impact of cur-
rency appreciation. Pain and Welsum [21] provided no clear conclusions as to the
impact of exchange rate movements on FDI due to the types of multinational
activities undertaken in different countries. Lee and Min [17] took Korea’s eight
major FDI source countries in three different regions in the world as samples and
tried to identify the changing behavior of foreign investors in Korea following the
1997 crisis. The change in FDI in response to exchange rate level is quite mixed,
which is consistent with recently developed real option-based FDI theory.
There are contradictory conclusions of exchange rate volatility influences on
FDI flows in theoretical and empirical literature. The risk averse multinationals
consider the information-searching cost and may decide to put off investing over-
seas when faced with dramatic volatility of exchange rate. Udomkerdmongkol
and Morrissey et al. [25] found evidence for a negative impact of exchange rate
variation on FDI of US in emerging countries. However, another group of studies
highlight the positive impact of exchange rate volatility on FDI. Cushman [6], in
his theoretical models, concluded that the exchange rate uncertainty may posi-
tively affect FDI. In response to risk, the multinational firm reduces exports to
the foreign country but offsets this somewhat by increasing foreign capital and
stimulating direct investment. Goldberg and Kolstad [11] also came to a similar
conclusion, namely, an increase of the uncertainty stimulates FDI. Deseatnicov
and Akiba [8] employed a panel data analysis of 56 developed and developing
countries (country and industry level) and found that exchange rate volatility
positively affect Japanese FDI activities for all industry. The results showed that
Japanese MNCs could tolerate a slight increase in the exchange rate volatility
because the level of it may be far enough what is necessary.
Although Chinese OFDI has become a more interesting topic, relatively few
empirical studies have been conducted. Some research has made contribution to
the literature by examining the major determinants of Chinese OFDI including
ER risks as well. But the results are controversial and apparently they did not
include samples from enough counties. For example, Jin [14] found that the
appreciation of RMB promotes FDI after the reforms in the ER regime in 2005.
Hu [13] used a panel data including 49 countries from 2003 to 2010. They found
that Chinese OFDI was positively related to ER level and negatively related
to ER volatility. Liu and Deseatnicov [19] found out that Chinese MNCs tend
to invest in locations with higher financial uncertainty because exchange rate
volatility increases the competitive advantage of Chinese MNCs in developing
countries with respect to MNCs from developed countries motivating an increase
in Chinese OFDI.
2.2 The Impact of Political Risk on FDI

When entering foreign market, multinationals have to adapt their optimal invest-
ment strategy to the requirements of local political environment that may give
rise to varying political risks. Different political environment and its impacts
on economic activity have received substantial attention in the recent literature
on FDI. There are a lot of scholars trying to prove that any form of political
risk negatively affects the FDI profitability and incentive of an MNE as a whole.
Busse and Hefeker [3] took 83 developing countries as samples and tried to prove
the negative linkages among political risk, institutions, and FDI inflows. They
also found out that political risk and institutional indicators matter the most
when multinational corporations confront decisions about where to invest in
developing countries. Hayakawa et al. [12] used the overall FDI inflows for 89
countries to empirically investigate the effects on inward FDI of various com-
ponents of political risk. The results showed that internal conflict, corruption,
military in politics, and bureaucracy quality are inversely related to inward FDI
flows. Cezar and Escobar [4] tested their model using inward and outward FDI
data on OECD countries and found that institutional distance reduces both the
likelihood that a firm will invest in a foreign country and the volume of invest-
ment it will undertake. And they concluded that firms from developed economies
adapt more easily to institutional distance than firms from developing economies.
Nevertheless, some previous studies implied that better political environ-
ment may not attract more FDI inflows. Kurul [16] constructed a composite
institutional quality index to identify the overall impact of institutions on FDI
inflows based on 126 countries over the period 2002–2012. The empirical results
revealed that institutional quality affects FDI positively only after this measure
exceeds a certain threshold value. In particular, Buckley et al. [2] found that
Chinese foreign investment may be led by China’s political and ideological her-
itage and be directed to ideologically similar countries which have higher records
of political risk. Yeung and Liu [27] also supported this conclusion. They hold
the view that Chinese state-owned transnational firms, to a great extent, played
a crucial role in foreign investment. Partly due to China’s unique and changing
political-economic context, many state-owned transnationals are often involved
in infrastructure projects and resource extraction industries in some developing
economies with poor government institutions as a means of economic diplomacy.
3 The Background of Chinese OFDI in the Countries

Along the “OBOR”
3.1 The Regional Distribution of the “OBOR” Initiative Area
The “OBOR” initiative focuses on bringing together China, Central Asia, Russia
and Europe (the Baltic); linking China with the Persian Gulf and the Mediter-
ranean Sea through Central Asia and West Asia; and connecting China with
Southeast Asia, South Asia and the Indian Ocean (Table 1). In its largest def-
inition, the “OBOR” initiative involves more than 60 countries, with a total
Table 1. The geographic distribution of countries along the “OBOR”
Region Major countries

Northeast Asia Mongolia, Russian Federation
Central and Eastern Europe Bulgaria, Belarus, Austria, Belgium, Czech
Republic, Romania, Poland, Ukraine, Hungary
Western Asia and the Middle East Georgia, Turkey, United Arab Emirates, Israel,
Egypt, Oman, Qatar, Iraq, Iran, Syrian Arab
Republic, Saudi Arabia, Yemen, Rep.
Central Asia Uzbekistan, Turkmenistan, Tajikistan,
Kazakhstan, Kyrgyzstan
Southeast Asia Vietnam, Malaysia, Myanmar, Cambodia,
Thailand, Brunei Darussalam, Indonesia, Lao
PDR, Philippines, Singapore
South Asia Pakistan, India, Jordan, Bangladesh, Sri Lanka,
Afghanistan
population of 3.08 billion, accounting for about 44% of the world’s population.
The total GDP of this area reaches $12.8 trillion, accounting for 17% of the
global economy.
3.2 Chinese OFDI Along the “OBOR”
(1) The pattern of Chinese OFDI

In recent years, Chinese OFDI along the “OBOR” countries has increased rapidly
as economic ties strengthened. Figure 1 describes the Chinese direct investment
stock trends in the countries involved in the “OBOR” initiative from the period
of 2003 to 2015. By the end of 2015, total flows of OFDI by Chinese enterprises
to the countries covered by “OBOR” initiative reached $18.93 billion, with an
increase of 38.6%, accounting for 13% of the total flows.
China’s OFDI varies in different economic entities. Chinese direct investment
along the OBOR concentrated on the developing countries from 2003 to 2015,
accounting for 75% of the total share. Among these developing countries, ASEAN
countries have absorbed more than 40% of China’s OFDI and have become some
of China’s most important economic partners. In 2015, China’s OFDI flows to
ASEAN countries experienced rapid growth and reached US$ 14.604 billion. By
the end of 2015, there had been six ASEAN countries included in China’s top
ten OFDI targets along the OBOR (Table 2). ASEAN economies have been at
the forefront of the emerging markets success story of the past two decades. It
is obvious that ASEAN countries occupy important positions in Chinese foreign
investment. Additionally, Chinese overseas investing activities in the transition
region have increased by 20% over the last couple of years. Rich natural resources
and the expansion of markets in this region are and will continue to be a con-
siderable attraction for foreign multinationals. Recent growth for this transition
Fig. 1. Chinese OFDI Stock into the “OBOR” countries (2003–2015) Sources: Statis-
tical Bulletin of China’s Outward Foreign Direct Investment.
Table 2. Top ten countries of Chinese OFDI stock along the “OBOR” by the end of
2015
Rank Country Chinese OFDI stock (million dollar) Percentage (%)

1 Singapore 31984.91 27.65
2 Russia 14019.63 12.12
3 Indonesia 8125.14 7.02
4 Kazakhstan 5095.46 4.41
5 Laos 4841.71 4.19
6 United Arab Emirates 4602.84 3.98
7 Myanmar 4258.73 3.68
8 Pakistan 4035.93 3.49
9 India 3770.47 3.26
10 Cambodia 3675.86 3.18
Total 84410.68 72.98
a
Sources: Statistical Bulletin of China’s Outward Foreign Direct Investment (2015)
region is highest among all regional groups. Kazakhstan, Mongolia and Russia
are currently among the largest FDI recipients in the region. It is worth noting
that the transitional region might become a new spotlight for Chinese foreign
investment, especially motivated by the “OBOR” initiative.
(2) The industry structure and investment project of Chinese OFDI
There has been a trend of diversification in the industrial structure of Chinese
OFDI along the “OBOR”. In 2005, the Chinese large-scale project investment
along the “OBOR” only involved energy industry that was mainly dominated by
oil and supplemented by natural gas and coal. Between 2006 and 2008, China’s
large-scale project investment extended to the metal ore industry, real estate,
transportation and other industries. The Chinese enterprises further expanded
their investment to high-tech, agriculture, finance, and chemical industries from
2009 to 2015. These changes in China’s OFDI along the “OBOR” show it has
experienced a steady process of ascension. In general, the focused industry is
energy. Metal ore, real estate and transportation rank second to fourth. Agri-
culture, chemical and high-tech industries account for a smaller part of total
shares.
The “OBOR” initiative makes the interconnection of infrastructure as a
breakthrough in cooperation. Driven by this initiative, China actively devel-
ops high-speed railway networks, expressway networks, and regional aviation
networks. In 2015, China plans to invest infrastructure construction along the
“OBOR” for 1.04 trillion Yuan, especially in railways, water conservancy and
port engineering, and airports, in that order (Fig. 2).
Fig. 2. China planed investment projects in “OBOR” countries (2015) (Sources: Min-
sheng Securities)
(3) The investment subject of Chinese OFDI

Under the “OBOR” initiative, government participation makes up the most
important part while social capital participation remains a relatively small part.
From the point of investment scale, central state-owned enterprises are the main
force of China’s OFDI along the “OBOR” while local private enterprises only
play a supplementary role. By the end of the first half of 2014, the central
enterprises’ national large-scale projects investment stock in this “OBOR” area
reached US $86.45 billion, accounting for 67.4% of the total share. Especially, the
OFDI stock from state-owned enterprises belonging to SASAC was US $78.22
billion, making up 90.5% of the central enterprises’ investment. In addition, local
enterprises large-scale projects investment stock in this area was US $41.9 billion,
accounting for 32.6% of the total share.
Fig. 3. Chinese Investment Subjects in “OBOR” Countries (By the End of the First
Half of 2014, Sources: Minsheng Securities)
4 Data
The aim of this paper is to examine the impact of exchange rate movements and
political environment on Chinese OFDI along the “OBOR”. Since the Ministry
of Commerce, National Bureau of Statistics, and State Administration of Foreign
Exchange started to publish a detailed Statistical Bulletin of China’s Outward
Foreign Direct Investment from 2003, our dataset contains the Chinese OFDI
net flows to selected countries for a period of 2003–2015.
We realize that a large number of country samples are critical to our analysis.
The Land and Maritime Silk Roads pass through over 60 countries with transmis-
sion of cultures, religion and trade. However, these vast networks cover so many
developing and small countries, the detailed official FDI data of which cannot
be found in Statistical Bulletin of China’s Outward Foreign Direct Investment.
As a result, our study met difficulties in collecting enough samples of countries
along the “OBOR”. So we include countries that are not geographically located
on the “OBOR” but also belong to its coverage such as European countries,
Africa and Far East. In this paper, the total sample consists of 65 countries on
the basis of all available information (See Table 4 in Appendix).
4.2 Data Description

The dependent, independent and control variables are as follows. The summary
statistics of the empirical variables is reported in Tables 5 and 6 in Appendix
shows the sources of the data.
(1) Dependent variable
The dependent variable for our analysis is the outward FDI flows to the host
country by Chinese multinationals during the sample period. We use LRFDI,
the natural logarithm of real FDI flows from China to each recipient country
expressed in ten thousand dollars (the CPI is used to deflate the nominal values,
with 2010 as the base year), as a measure of FDI. To understand the details
of China’s OFDI, we downloaded an annual country-level data from Statistical

Bulletin of China’s Outward Foreign Direct Investment. This official government
data repository for all Chinese firms (including Limited companies, Private enter-
prises, State-owned enterprises, Foreign-invested or Joint-Venture firms) and has
rich information including the destination of OFDI, the various sectors, the name
of the investor in China and regional origin of the investor. This is so far the
most reliable information at country-level in long time-series.
(2) Independent variables
This paper mainly investigates the effect of the political environment and
exchange rate movements on Chinese OFDI in the “OBOR” countries. The mea-
surement of independent variables proceeded as follows.
(a) Exchange Rate variables
We also collect a set of observable monthly nominal exchange rates (RMB/local
currency, using US dollars as the transformation instrument) from selected coun-
tries which can be obtained from the IMF International Financial Statistics
database. In order to obtain the real exchange rates, we use host country’s CPI
index to deflate the nominal values, with 2010 as the base year. From this data,
we study the level and volatility of exchange rate to explore their impacts on
Chinese OFDI.
– LMEANR is the natural logarithm of the level of real exchange rates. We

calculate the first moment (average) of monthly real exchange rates around
year t (defined to include all monthly observations during year t − 1 and year
t) for host country.

24
LMEANR = log xi /24 .
2
i=1
– LVARR is the natural logarithm of the volatility of real exchange rates. We

calculate the second moment (standard deviation) of real exchange rate level
around year t (defined to include all monthly observations during year t − 1
and year t) for host country.
(b) Political Environment

Political environment, however, varies across countries and regions, which pro-
vides us an opportunity to test the impact of the severity of PE risks on out-
ward FDI by Chinese multinationals. In this paper, we use PE to represent the
evaluation of host country’s political environment. As one of the most widely
accepted system of completely independent political risk forecasting, the Politi-
cal Risk Services Group (PRS) offers us a general risk model to translate political
risk assessments into numbers and calculate risk ratings for its database of 140
developed, emerging, and frontier markets. The Political Risk Rating comprises
6 same weighted variables covering both political and social attributes in order
to focus on the needs of the particular investing firm (see Appendix 1d). These
actions can result in threats or do harm to the business climate. Thus, we calcu-
late the sum of PRS 6 indicators to capture the features of political environment
in host country. An increasing value (from 0 to 6) represents lower political risk
and better political environment.
(3) Control variables
We look at the macroeconomic determinants of outward FDI. The host country’s
CPI index is used to deflate the nominal values, with 2010 as the base year.
Following previous research, we tested main factors as follows:
• LRGDP is the natural logarithm of host country’s real GDP (Deflated by
CPI) based on current US dollars in a given year, which reflects the host
country’s future market potential or absorptive capacity.
• NATURAL is the natural logarithm of total natural resources rents (% of
GDP). The total natural resource rents are defined as the sum of oil rents, nat-
ural gas rents, coal rents (hard and soft), mineral rents, and forest rents. The
estimates of natural resources rents are calculated as the difference between
the price of a commodity and the average cost of producing it. This indica-
tor could be used to measure the abundance of natural resource in the host
country.
• LRW is the natural logarithm of real wage in host country. GDP per capita
based on purchasing power parity (current US $) serves as a proxy for real
wage due to the difficulties of getting average wage data for most of the
“OBOR” countries. It weighs the labor cost in host country and is widely
used to measure the advantage of human capital in attracting foreign manu-
facturing firms.
• LOPEN is the natural logarithm of the sum of annual imports and exports
volume of goods and services (% of GDP). It stands for the degree of host
country’s openness in the FDI recipient economy, which is considered to be
an important factor affecting FDI-related barriers.
• LTECH is the natural logarithm of host country’s high-technology exports (%
of real GDP). High-technology exports are products with high R&D intensity,
such as in aerospace, computers, pharmaceuticals, scientific instruments, and
electrical machinery. Data are in current U.S. dollars. This variable could
measure the level of technology in host countries, which becomes the one of
consideration in Chinese MNCs foreign investment activities.
• PLCY is a dummy variable that can be interpreted as the effect of the
“OBOR” initiative. The PLCY dummy takes a value of zero if the time
period is 2003–2013 and one otherwise. We splits the sample around 2013
because it is believed that Chinese OFDI activities in the countries along the
OBOR are strongly driven by the “OBOR” initiative that was first launched
by Chinese government in 2013.
5 Empirical Specification and Results

A linear dynamic panel-data model is used in this paper taking into consideration
the lagged FDI effect on current FDI. We use the System GMM (SYS-GMM)
estimator due to its capacity to deal with lagged values of the endogenous vari-
ables as poor instruments for first differences. With characteristics of missing
observations for some countries and years, our dataset represents as an unbal-
anced panel data. Thus, we use “forward orthogonal deviations” variables trans-
formation proposed by Arellano and Bond [7] in order to increase the number
of observations used in the analysis.
As for instruments choosing, the independent variables were treated as
strictly exogenous. So the system estimators use the first difference of all the
exogenous variables as standard instruments, and the lags of the endogenous
variables to generate GMM-type instruments. Finally, we added time dummies
in order to increase the likelihood of no correlation across individuals in the
idiosyncratic disturbances assumption to hold as suggested by Roodman [22].
As robustness check to our SYS-GMM regressions we report feasible gener-
alized least squares (FGLS) estimation method with heteroscedastic error term.
We include time dummies to the FGLS estimations as well and we exclude lagged
FDI variable in order to avoid the autocorrelation problem in the FGLS method.
We estimate the equation:

yit = δ × yit − 1 + Xit β + εit , (1)
where yit is the logarithm of annual outward FDI from China into a host ‘coun-

try i’ at time t and Xit denote an (1xk) vector of exogenous variables which
vary in the cross-section and in the time dimension. δ is a scalar. yit−1 is a
lagged dependent variable. εit is a stochastic error term, which is assumed to be
uncorrelated over all i and t. For FGLS estimation we omit δ yit−1 term.
All system GMM estimation results for Eq. (1) are shown in Table 3. In our
specifications, the Hansen test of overidentifying restrictions indicates that the
joint null hypothesis of valid instruments is not rejected. Besides, the FGLS
results displayed in Table 3 show that the results of system GMM estimation are
consistent and robust. Some interesting findings have been disclosed through the
regression and we give some interpretations and evaluations for them.
5.1 RMB Appreciation Stimulates Chinese OFDI in “OBOR”

Countries
ER level is found to be positive and statistically significant for the 65 selected
countries. It is expected correctly signed and positively associated with Chinese
OFDI. This suggests that Chinese multinationals are willing to invest abroad in
these countries and areas when the Yuan appreciated with the host’s currency.
More specifically, ER level coefficient indicates that a 1% increase in the value of
Yuan would promote an about 0.3% increase in Chinese OFDI. A considerable
amount of OFDI is conducted through M&A, Yuan appreciation can systemati-
cally lower the cost of acquisitions of certain foreign assets [10] and increase the
profitability of firms’ foreign branches and subsidiaries [1]. Thus, an appreciation
of home currency stimulates OFDI. Moreover, evidence found from the statistics
suggests that there is a stable uprising trend for Yuan during the sample year.
Especially after 2005, Yuan appreciated over 30% [28]. In this case, the yuan’s
bias toward relatively large appreciation shocks is associated with expectation
of the appreciation that leads to a lower cost of future reinvestment in the host
country and stimulates the foreign investment.
5.2 Chinese OFDI Is Not Sensitive to the Exchange Rate Volatility

in “OBOR” Countries
ER volatility coefficient shows a negative but insignificant relationship with

OFDI, indicating that the influences of ER volatility on Chinese OFDI are
ambiguous (Table 3). The possible explanation for this insignificant result is that
firstly, the Chinese government is very cautious with respect to the pace of the
Table 3. Panel regression by System GMM and FGLS for the 65 “OBOR” Countries
(Dependent variable: LRFDI)
VARIABLES SYS-GMM (Two Dif.) SYS-GMM (Two Ort.) FGLS

LRFDI LRFDI LRFDI
L.LRFDI 0.373 0.172
∗∗∗
(−0.031) (−0.061)∗∗
LMEANR 0.273 0.295 0.319
(−0.087)∗∗ (−0.124)∗ (−0.082) ∗∗∗
LVARR −0.11 −0.15 −0.12

(−0.061) (−0.104) (−0.071)
LRGDP 0.361 0.710 0.744
(−0.092)∗∗∗ (−0.172)∗∗∗ (−0.07)∗∗∗
NATURAL 0.012 0.038 0.045
(−0.004)∗∗ (−0.011)∗∗∗ (−0.007)∗∗∗
LRW 0.109 −0.177 −0.01
(−0.143) (−0.194) (−0.105)
LOPEN 0.207 0.343 0.435
(−0.057)∗∗∗ (−0.062)∗∗∗ (−0.087)∗∗∗
PE 0.07 0.161 −0.346
(−0.183) (−0.252) (−0.157)∗
LTECH 0.039 0.03 0.116
(−0.044) (−0.057) (−0.035)∗∗∗
PLCY 0.352 2.074 3.720
(−0.089)∗∗∗ (−0.278)∗∗∗ (−0.38)∗∗∗
Constant −7.106 −15.398 −15.454
(−2.079)∗∗∗ (−3.755)∗∗∗ (−1.646)∗∗∗
No. of observations 383 383 453
No. of groups 48 48 48
No. of instruments 38 38
Group instrument ratio 1.263 1.263
Hansen test of overrid. (P value) 0.169 0.241
Note: Standard errors in parentheses. ∗∗∗ , ∗∗
, ∗
indicate that the null hypothesis is rejected at 1, 5
and 10 percent, respectively.
RMB appreciation. During our sample year, China’s Central Bank intervened
and managed the unexpected RMB exchange rate fluctuation. Also RMB is
usually expected to appreciate gradually. Thus, Chinese multinationals pay less
attention to the variance of ER than to the level of ER in managing financial
risk. Secondly, unlike the multinationals in developed countries, the motivation
for Chinese MNCs is not to substitute exports but to obtain higher values of
assets or access natural resources in a host country. ER volatility is a sign of
instability but at the same time it discourages potential investors from devel-
oped countries to involve in competition. Thus, Chinese MNCs may be indifferent
to exchange rate volatility because of their motivation. Thirdly, less hedge tools
and RMB investment instruments designed for safeguard the exchange risks are
used in the international settlement practices for Chinese MNCs. They suffer
large exchange loss due to a lack of awareness towards exchange rate volatility.
This is a rather dangerous trait as China is going to make the RMB more flexible
and the Chinese Central Bank will gradually quit the normal intervention.
5.3 Chinese OFDI Seems to Be Indifferent to Political Instability

in “OBOR” Countries
The PE variable achieves a statistically insignificant result with a positive sign,

which is not in line with previous studies from the perspective of risk premium
and information asymmetries. This suggests that Chinese multinationals are not
sensitive to the political instability when investing in “OBOR” countries. This
result may be explained by three possible reasons. First, multinationals that are
predominantly oriented toward seeking natural resource have a greater tolerance
of political risk. Our empirical result shows that the abundance of fuel and ore
resource in a host country has a positive impact on attracting Chinese OFDI.
It indicates that Chinese OFDI in “OBOR” countries is motivated by the host’s
natural resources endowments. It is well known in the literature that countries
rich in natural resources, on average, grow more slowly and have ineffective
political institution than countries without such resources due to the “resource
curse” [20]. The more Chinese OFDI a country attracts, the worse its government
institution [15]. The “resource curse” effect can be confirmed by examples with
extremely high political risk in our sample. There are some countries (such as
Afghanistan, Iraq and Libya) in the Middle East, Central Asia and Africa, with
large oil and other critical energy production. Therefore, it is inevitable that
Chinese investors have to face and endure more political risk when conducting
resource- seeking FDI in “OBOR” countries.
Second, in China, State-owned enterprises are the main force in foreign invest-
ment, whose investment behavior is affected by the government’s willingness.
Economics activities carried out by Chinese national firms are not only for their
profit motive and economic efficiency, but also for distinctive political and diplo-
matic overtones in “economic diplomacy”. Consequently, strongly supported by
Chinese government, Chinese SOEs tend to be less averse to invest in countries
with higher political risk compared to non-SOEs [9]. Government support may
effectively compensate for the loss of the assets value and expected return due
to the changes of political environment.
Third, China’s political relations with potential hosts have significant influ-
ences on firms overseas investment decisions and patterns. Countries keeping bet-
ter interstate relations with China tend to get more foreign investment from Chi-
nese multinationals [18]. The “OBOR” initiative covers many countries that have
developed good long-term relationships, supporting each other politically and
co-operating economically. For instance, the members of SCO (Kazakhstan and
Russia), some ASEAN countries and African countries have long-term strategic
partnerships with China.
5.4 Natural Resources Abundance and Market Potential Attract

Chinese OFDI in “OBOR” Countries
In preliminary regressions, we see that natural resource is the positive significant

determinant of Chinese OFDI within the “OBOR” countries: that is, Chinese
investors preferentially seek out the ownership advantage of natural resource in
the “OBOR” countries. In the long term, Chinese resource-seeking foreign invest-
ment has been aimed at getting access to the scarce natural resource that China
lacks to ensure the supply of domestic production and continue development.
Especially, a great proportion of large acquisitions carried out by Chinese multi-
nationals concentrates on the resource sector. Purchases of a 96.9% stake in the
Russian oil company Udmurtneft by Sinopec, the equity M & A of Kazakhstan
PK Petroleum Company by China National Petroleum Corporation (CNPC)
and the strategic alliance between National Grid Corporation of the Philippines
and State Grid Corporation of China are examples.
Moreover, according to our results, the two alternative measures of host
market potential (real GDP and openness) attained significance. This finding
reinforces the view that market potential and FDI outflow are associated posi-
tively [5]. Especially, new emerging economies or transition economies along the
“OBOR” present more opportunities for generating profits than before. As an
emerging hotspot, ASEAN is working towards regional integration aimed at one
single market and free movement of goods, capital and services. It is becoming
a rapidly developing destination for foreign investment. Besides, the transition
economies have enjoyed the development dividend after their significant reforms
of institutional and economic structures. Their increase has been well beyond
that in Latin America and Africa in recent years.
5.5 A New Wave of Chinese OFDI in “OBOR” Countries Benefited

from Favorable Policies
The estimated policy variable coefficient is positive and statistically signifi-

cant in our specifications, which is intuitively plausible. The “OBOR” initiative
proposed in 2013 is a centerpiece of China’s foreign policy and domestic eco-
nomic strategies, substantially giving boost to multilateral ties and economic
co-operation that promote regional integration. Considering that “OBOR” ini-

tiative serves as a new engine to speed up investment facilitation, eliminate
investment barriers, and push forward negotiations on bilateral investment pro-
tection agreements, it is not difficult to understand that the sizable increase
of Chinese OFDI in “OBOR” countries. As the “OBOR” initiative has been
put on the agenda, the spotlight falls on infrastructure construction, energy
cooperation and advanced manufacturing sectors between China and the coun-
tries along the routes. For example, China has signed memorandum with 20
“OBOR” countries for institutional capacity cooperation, especially Indonesia,
Kazakhstan, and Mongolia. There have been 52 projects with a total investment
of $27 billion are being implemented under the “OBOR” initiative. Therefore,
the increasing policy bonuses are encouraging Chinese investors to expand their
presence overseas and explore new business opportunities (Table 7).
6 Conclusion and Policy Implications
The findings of our study have some policy implications for both the government
and firms under the “OBOR” initiative. The empirical results reveal that ER
level is a highly statistically significant determinant of Chinese OFDI while PE
and ER volatility are not. Based on this result, the Chinese government should
execute an active and strategy-focused diplomatic policy in the “OBOR” coun-
tries in order to better support and promote the Chinese multinationals’ foreign
investment. Research agencies and the Department of Commerce should coop-
erate to issue PE risk reference and precaution for the firms that have already
invested or have the potential to invest. Industry associations should also play
a role in guiding the firms to invest rationally at a steady pace.
Exchange rate risks should be taken into consideration as China is moving
towards a more flexible exchange rate regime. Especially, China’s Central Bank
has taken further action on the RMB internationalization, by widening the RMB
exchange rate’s floating band within the inter-bank market from 0.5% to 1.0%.
Becoming effective on April 16, 2012, this policy has been seen as a strong sig-
nal to the world that the RMB exchange rate regime is going to be driven more
and more by market forces, than by government intervention. Thus, exchange
rate volatility will increase as RMB exchange rate reforms are made. Finan-
cial institutions, like commercial banks, should provide more hedging tools for
Chinese-outgoing-firms to reduce exchange rate risks. As for Chinese firms, it
is vital to observe the ER indicators because these largely determine the safety
of investment and the overall profitability calculated in home currency. In con-
clusion, a trend that direct investment in the “OBOR” countries will steadily
increase can be foreseen. It is vital for Chinese Multinationals to be more aware
of risks to conduct efficient investment abroad.
Acknowledgements. This paper is supported by Sichuan Science and Technology

Department under Grant 2016ZR0017. The Usual disclaimer applies.
Appendix
Table 4. List of 65 countries used in the study
Middle East and Africa (15) Europe (27) East, South and Central Asia
(23)
Afghanistan, Angola, United Bulgaria, Belarus, Austria, Azerbaijan, Brunei
Arab Emirates, Israel, Algeria, Belgium, Switzerland, Czech Darussalam, Hong Kong,
Egypt, Liberia, Libya, Oman, Republic, Germany, Denmark, Indonesia, India, Jordan,
Qatar, Iraq, Iran, Syrian Arab Spain, Finland, France, United Kazakhstan, Lao PDR,
Republic, Saudi Arabia, Kingdom, Ireland, Romania, Pakistan, Philippines,
Yemen, Rep. Russian Federation, Georgia, Bangladesh, Sri Lanka,
Italy, Portugal, Luxembourg, Singapore, Uzbekistan,
Netherlands, Norway, Poland, Vietnam, Turkey, Thailand,
Ukraine, Sweden, Iceland, Turkmenistan, Macao,
Greece, Hungary Cambodia, Myanmar,
Mongolia Malaysia
Table 5. Summary statistics
Variable Obs Mean Std. Dev. Min Max

LRFDI 657 3.367761 2.630155 −4.60303 11.19835
LMEANR 729 −3.72812 2.920762 −7.48804 3.590258
LVARR 716 −6.82429 3.337829 −17.5878 3.547741
LRGDP 764 21.15035 1.648647 15.90119 24.36234
NATURAL 746 14.18228 19.19367 0.000372 92.01895
LRW 764 4.60327 1.531155 0.946498 7.054828
LOPEN 806 4.083142 1.054777 −1.03116 6.120905
PE 754 4.005506 1.051376 1.414773 5.920455
LTECH 838 4.229373 9.253167 0.000009 75.041053
Table 6. Data Sources
Variables Sources
LRFDI Statistical bulletin of China’s outward foreign direct investment
LMEANR International financial statistics
LVARR International financial statistics
PE Political risk services group
LRGDP World Bank database NATURAL World Bank database
LRW World Bank database
LOPEN World Bank database
LTECH United Nations, Comtrade Database
Table 7. Variables incorporated into Political Risk Services (PRS) index
Voice and accountability Military in politics

Political stability and absence of violence Democratic accountability
Government stability
Internal conflict
External conflict
Ethnic tensions
Government effectiveness Bureaucratic quality
Regulatory quality Investment profile
Rule of law Law and order
Control of corruption Corruption
References
1. Blonigen BA (1997) Firm-specific assets and the link between exchange rates and
foreign direct investment. Am Econ Rev 87(3):447–465
2. Buckley PJ, Zheng P (2009) The determinants of chinese outward foreign direct
investment. J Int Bus Stud 40(2):353–354
3. Busse M, Hefeker C (2007) Political risk, institutions and foreign direct investment.
Eur J Polit Econ 23(2):397–415
4. Cezar R, Escobar OR (2015) Institutional distance and foreign direct investment.
Rev World Econ 151(4):713–733
5. Chakrabarti R, Scholnick B (2002) Exchange rate expectations and foreign direct
investment flows. Rev World Econ 138(1):1–21
6. Cushman DO (1985) Real exchange rate risk, expectation, and the level of direct
investment’. Rev Econ Stat 67(2):297–308
7. Dan B (1991) Some tests of specification for panel data: Monte carlo evidence and
an application to employment equations: Monte Carlo evidence and an application
to employment equations. Rev Econ Stud 58(2):277–297
8. Deseatnicov I, Akiba H (2016) Exchange rate, political environment and FDI deci-
sion. Int Econ 148:16–30
9. Duanmu JL (2011) The effect of corruption distance and market orientation on
the ownership choice of MNEs: evidence from China. J Int Manage 17(2):162–174
10. Froot KA, Stein JC (1991) Exchange rates and foreign direct investment: an imper-
fect capital markets approach. Q J Econ 106(4):1191–1217
11. Goldberg LS, Kolstad CD (1995) Foreign direct investment, exchange rate vari-
ability and demand uncertainty. Int Econ Rev 36(4):855–873
12. Hayakawa K, Kimura F, Lee HH (2013) How does country risk matter for foreign
direct investment? Dev Economies 51(1):60–78
13. Hu B (2012) RMB exchange rate and Chinese outward foreign investment-a cross-
country panel analysis. Contemp Econ Res 11:77–82 (in Chinese)
14. Jin W, Zang Q (2013) Impact of change in exchange rate on foreign direct invest-
ment: evidence from China. Lingnan J Bank Finan Econ 4(1):1
15. Kolstad I, Wiig A (2009) What determines Chinese outward FDI? J World Bus
47(1):26–34
16. Kurul Z (2016) Nonlinear relationship between institutional factors and FDI flows:
dynamic panel threshold analysis. Int Rev Econ Finan 48:148–160
17. Lee BS, Min BS (2011) Exchange rates and FDI strategies of multinational enter-
prises. Pacific-Basin Finan J 19(19):586–603
18. Li Q, Sr GL (2012) Political relations and Chinese outbound direct investment:
evidence from firm-and dyadic-level tests. SSRN Electron J. doi:http://dx.doi.org/
10.2139/ssrn.2169805
19. Liu HY, Deseatnicov I (2016) Exchange rate and Chinese outward FDI. Appl Econ
51:1–16
20. Mehlum H, Moene K, Torvik R (2006) Institutions and the resource curse. Econ
J 116(508):1–20
21. Pain N, Welsum DV (2003) Untying the gordian knot: the multiple links between
exchange rates and foreign direct investment. JCMS J Common Mark Stud
41(5):823–846
22. Roodman D (2009) A note on the theme of too many instruments. Oxford Bull
Econ Stat 71(1):135–158
23. Schmidt CW, Broll U (2009) Real exchange-rate uncertainty and US foreign direct
investment: an empirical analysis. Rev World Econ 145(3):513–530
24. Takagi S, Shi Z (2011) Exchange rate movements and foreign direct investment
(FDI): Japanese investment in Asia, 1987–008. Jpn World Econ 23(4):265–272
25. Udomkerdmongkol M, Morrissey O, Görg H (2009) Exchange rates and outward
foreign direct investment: US FDI in emerging economies. Rev Dev Econ 13(4):754–
764
26. Wang Y (2015) China’s evaluation of investment risks in countries along ‘one belt,
one road’ initiative. China Opening J 181(4):15–21
27. Chung Yeung HW, Liu W (2008) Globalizing China: the rise of mainland firms in
the global economy. Eurasian Geogr Econ 49(1):57–86
28. Yue LH, Qiang J, Kai TY (2016) Determination of renminbi equilibrium exchange
rate, misalignment, and official intervention. Emerg Mark Finan Trade 52(2):420–
433
Post-Traumatic Stress Disorder Among
Survivors in Hard-Hit Areas of the Lushan
Earthquake: Prevalence and Risk Factors
Zongtang Xie1(B) and Hongxia Xiu2

1
School of Management, Northwest Minzu University,
Gansu 730030, People’s Republic of China
zongtangx171@126.com
2
School of Economics, Northwest Minzu University,
Gansu 730030, People’s Republic of China
Abstract. To estimate the prevalence of post-traumatic stress disorder

(PTSD) and assess determinants related to PTSD among survivors in
areas after the 2013 Lushan earthquake in China. This study used the sur-
vey instruments included the PTSD Checklist-Civilian Version (PCL-C)
and the Social Support Rating Scale (SSRS). We found that the preva-
lence rates of suspected post-traumatic stress disorder (PTSD) in hard-hit
areas were 35.6% (n = 500). The incidence of PTSD was more prevalent
among women (43.41%), compared with men (27.27%). Based on multi-
variate logistic regression analyses, female sex, Tibetan ethnicity, being
injured, having family members injured and social support were signifi-
cant risk factors in hard-hit areas.
Keywords: Prevalence · Risk factors · Post-traumatic stress disorder

(PTSD) · Lushan earthquake
1 Introduction
On April 20, 2013 at 8:02 am, an earthquake measuring 7.0 on the Richter scale
occurred in Lushan County, Sichuan Province, China. Its epicenter and focal
depth were at latitude 30◦ 308’N, longitude 102◦ 888’E, and 14.0 km, according
to the location of U.S. Geological Survey. Lushan earthquake occurred in the
southwestern part of the Longmen Shan fault zone [5]. It was the strongest
earthquake after the 12 May 2008, Ms 8.0 (China Earthquake Data Center) or
Mw 7.9 (U.S. Geological Survey), Wenchuan earthquake in Sichuan province,
China. Unlike other natural disasters, earthquakes usually occur without any
warning and can cause widespread devastation and expose thousands of people
to sudden bereavement, injury, loss of property, homelessness, and displacement
[8]. Lushan earthquake still caused strong shaking for human society around its
1458 Z. Xie and H. Xiu
epicenter. Since the source regions of the Lushan earthquake are highly pop-
ulated, and with numerous mountainous, it resulted in more than 200 deaths
or missing persons, more than 10,000 injuries, and huge economic losses accord-
ing to the local government’s official report. Besides deaths, physical injuries and
economic losses, earthquakes can have serious psychological impacts for survivors
as well as physical injuries and various mental health problems [1].
Many survivors of natural disasters experience post-traumatic stress disorder
(PTSD) in their adjustment to the loss of resources (e.g. housing, belongings)
or loved ones (e.g. family members) [1]. PTSD is the most frequently reported
psychological sequelae among victims of natural disasters [15]. Numerous stud-
ies have documented the estimated the PTSD among other earthquake sur-
vivors. Through a longitudinal survey on the onset and development of DSM-IV
Post-traumatic Stress Disorder (PTSD) after the Zhangbei earthquake in North
China, Wang et al. reported the rate of onset of PTSD within 9 months in
severely affected village (30.3%) was higher than that in lightly affected villages
(19.8%) [13]. Another survey two and a half months following the earthquake
in Beichuan and Langzhong Counties in Sichuan Province found that the preva-
lence rates of suspected PTSD in heavily (Beichuan County) and moderately
(Langzhong County) damaged counties reached 45.5% and 9.4%, respectively [6].
To summarize the above studies, previous assessments among survivors of
natural disasters have shown that PTSD are common, and the prevalence rates
of PTSD among the hard-hit survivors were far higher than those of lightly hit
survivors. Therefore, understanding PTSD is essential for identifying vulnerable
populations and developing culturally specific mental health interventions [11].
For the purpose of public health emergency response, we conducted a rapid
assessment of the prevalence of symptoms of PTSD and associated factors among
random samples of survivors in the 5 countries of Lushan, Baoxing, Mingshan,
Qionglai, and Tianquan, which were the most severely affected by the Lushan
earthquake.
2 Methods
2.1 Subjects
According to the hard-hit counties list published by the Central People’s Gov-
ernment of the People’s Republic of China, we conducted a cross-sectional sur-
vey in heavily damaged counties. The hard-hit counties were Lushan, Baoxing,
Mingshan, Tianquan, and Qionglai. These counties were selected because they
had suffered more extensive damage than other counties in China. The inclusion
criteria were as follows-having a high degree of exposure to the earthquake and
experiencing the complete process of the earthquake-with a fair distribution of
sex, age, and place.
The survey teams were temporarily established, consisted of well-trained
master’s level psychology students. They participated in a 6-day training pro-
gram that included lectures describing the study protocol and instruments, role-
play interviews, and mutual discussion. The survey comprised all assessments
Post-Traumatic Stress Disorder Prevalence and Risk Factors 1459
and was administered by senior staff psychiatrists and psychologists from Sichuan
University’s Medical School.
Before conducting the formal investigation, a pilot test was carried out in
August and September 2013, with a group of randomly selected survivors par-
ticipating. Minor modifications and adjustments were made according to the
feedback from the pilot test. The final version of the questionnaire was used in
the formal investigation. All assessment forms were translated from English to
Chinese and back-translated by a bilingual team of professionals.
From December 2013 to January 2014, master’s level psychology students
working as research assistants approached the participants in their own homes or
in temporary accommodation. To ensure privacy, interviewers and participants
were encouraged to complete the questionnaires in private places. Supervision
came out on a day-to-day basis throughout the survey.
2.2 Instruments
(1) PTSD Checklist-Civilian version

PTSD Checklist-Civilian version (PCL-C) is one of the most commonly used
instruments for PTSD symptoms that result from disaster related psychiatric dis-
orders [2]. The 17 items of the checklist is corresponding to each symptom in the
Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-
IV) PDSD criteria B (re-experiencing), C (avoidance/numbing), and D (hyper-
arousal). With regard to the Lushan earthquake attacks, participants indicated
the extent of each symptom correlated with the earthquake in a scale from 1 (not
at all bothered) to 5 (extremely severely). Criteria B was re-experience of the
earthquake (5 symptoms), criteria C was avoidance and emotional numbing
(7 symptoms), and criteria D was increased arousal (5 symptoms). The score
in each PTSD domain was simply added up, generating a total PTSD score
ranging from 17 to 85. A score of 50 was found to be the optimally efficient
cut-off score to predict the clinical diagnosis of PTSD [14]. The Chinese ver-
sion of PCL-C has shown good diagnostic agreement and sound reliability and
validity [7]. The coefficient of Cronbach’s alphas for PTSD was 0.88.
(2) The Social Support Rating Scale
The Social Support Rating Scale (SSRS) evaluates subjective social supports
from friends and family [12]. Numerous empirical studies have demonstrated the
protective function of social support for post-disaster mental health [12]. The
SSRS has been shown to have high reliability and validity on a wide range
of Chinese populations. It contains 10 items, measuring three dimensions of
social support: subjective support (SS, 4 items), objective support (OS, 3 items),
and support-seeking behavior (SA, 3 items). The item scores of the SSRS were
simply added up, generating a total support score ranging from 12 to 66 [12].
The subjective support score ranging from 8 to 32, the objective support score
ranging from 1 to 22, and the support availability score ranging from 3 to 12,
respectively. Higher scores indicate stronger social support. The Cronbach alpha
coefficient for social support was 0.87.
2.3 Data Analysis

In the present study, statistical analyses were performed using SPSS Version 16.0
(SPSS Inc., Chicago, IL, USA). All presented data were adjusted for clustering
and assigned a population-based weighting factor, based on the population size
of each cluster in the final listing of all villages and settlements. Demographic
data were analyzed using frequency. Chi-square test was used to compare PTSD
total scores among male and female participants. One-way ANOVA was used to
compare PTSD, total scores among different age groups. Multivariate logistic
regression analyses were used to investigate the factors that affect PTSD.
3 Results
Characteristics of the entire study group in hard-hit areas of the Lushan earth-
quake were summarized. The demographic data of the 500 participants shows
in Table 1. Males comprised 48.4% of the participant population compared
with 51.6% for females. Of the participants, 489 (97.8%) were of Han national-
ity, and 11 (2.2%) were of Tibetan nationality. The average age was 38.0 years,
with 44.6% of the participants in the 31–50 year age group; 71.8% of the par-
ticipants were married. Only 4 participants (0.8%) had a graduate education,
144 (28.8%) had a bachelor education and 352 (70.4%) had an elementary school
education or were illiterate. Among our subjects, 18.4% had a monthly income
of over 2000 yuan, 38.2% had an income of 1000–2000 yuan and 38.4% earned
less than 1000 yuan. With regard to the survivors’ fear during the earthquake,
65.2% reported “yes”. Of the survivors, 51.7% reported that their social sup-
port was high in the past year, whereas 28.6% reported low. One hundred and
eight participants (21.6%) reported that they suffered from at least one physical
disease, and no participants reported suffering from psychic diseases.
Percentages of PTSD status, as well as their scores, among males and females
are shown in Table 2. Based on the PCL-C total scores, participants were clas-
sified as suffering from PTSD if the score was equal to or higher than 50. The
overall percentage of PTSD was 35.6% (n = 178). The percentage of PTSD was
significantly higher among females (43.4%) than among males (27.3%), as were
PCL-C total scores (32.78 vs. 26.64, respectively).
Table 3 compares PTSD status among different age groups. Overall, there was
a significant difference between the age groups with regard to PCL-C total scores
and rates. The PCL-C total scores were significantly higher in the 41 to 50 year
age group, and the prevalence rate was higher than other age group. The lowest
prevalence rates and PCL-C total scores were found in the 15 to 30 year age
group.
Table 4 shows the factors that affected post-traumatic stress disorder (PTSD)
symptoms by multivariate logistic regression. Various demographic and loss vari-
ables were entered into the models. The prevalence rate of probable PTSD
was 35.6% (n = 178) based on the PCL-C cut-off score of 50 (Table 4).
Results of the multivariate logistic regression analyses indicated that the
prevalence of probable PTSD was significantly higher among individuals with
Table 1. Demographic data of the study population
Variables n %
Location Rural 381 76.2
Urban 119 23.8
Gender Male 242 48.4
Female 258 51.6
Age (years) 15–30 191 38.2
31–40 88 17.6
41–50 135 27
≥ 51 86 17.2
Mean ±S.D.: 38.0 ± 13.7 < 1000 RMB 192 38.4
Monthly income 1000–2000 RMB 191 38.2
2000–3000 RMB 85 17
> 3000 RMB 32 6.4
Ethnic group Han 489 97.8
Tibetan 11 2.2
Education level
Graduate 4 0.8
Bachelor 144 28.8
No degree 352 70.4
Lossa No 417 83.4
Yes 83 16.6
Injury of body 39 7.8
Family member injured 42 8.4
Family member died 7 1.4
Loss was defined as death of family member, or injury to family
member or self as s result of the quake.
Table 2. Comparison of PTSD status for males and females
Items Male Female P-value

n(%) n(%)
PTSD
Normal (< 50) 176 (72.73) 146 (56.59) < 0.01
Abnormal (≥ 50) 66 (27.27) 112 (43.41)
PCL-C (total score) 26.64 ± 3.19 32.78 ± 5.6 < 0.05
females, Tibetan people, injury of body, family member injured, middle-aged

(the 31 to 40 year and 41 to 50 year age group) and elderly group (the over 51 year
age group) than the control group. Comparing the 15 to 30 year age group with
the other age group, the 31 to 40 year age group, the 41 to 50 year age group,
Table 3. Comparison of PTSD status among different age groups
15–30 years 31–40 years 41–50 years 51 years P-value

N (%) N (%) N (%) N (%)
PTSD
Normal (< 50) 141 (73.82) 49 (55.68) 74 (54.81) 58 (67.44) < 0.01
Abnormal (≥ 50) 50 (26.18) 39 (44.31) 61 (45.19) 28 (32.56)
PCL-C (total score) 23.52 ± 3.98 31.25 ± 4.47 33.97 ± 4.63 26.67 ± 2.03 < 0.05
and over-51-year age group had higher PTSD symptoms and odds ratios (OR).
PTSD was significantly higher among victims who sustained an injury or whose
family member injured during the quake compared with noninjured survivors.
OR for risk of PTSD among victims who sustained an injury or whose family
member injured during the quake were 1.88 and 1.53 (compared with that of the
noninjured survivors) (Table 4).
4 Discussion
4.1 Prevalence of Probable PTSD
PTSD is still common eight months after the earthquake. Compared with the
prevalence of probable PTSD among hard-hit survivors soon after the earth-
quake, the prevalence after eight months has declined but remains significant.
The elevated prevalence rates of psychological symptoms showed that PTSD is
common mental health problems in the hard-hit areas after exposure to this
natural disaster and remained alarmingly high.
The present study found that approximately more than one third of partici-
pants suffered from probable PTSD (35.6%) based on the PCL-C cut-off score of
50. PTSD among survivors in hard-hit areas of the Lushan earthquake were com-
mon in our subjects. Our findings indicated a steady decline in the prevalence
of PTSD over time [4]. As time went by, the reduced mental health problems
among survivors in hard-hit areas may be associated with relatively good living
conditions and substantial social support given by the government and other
aid organizations. The prevalence of PTSD in our sample were high, compared
with rates of PTSD in previous studies after earthquakes. The prevalence rates
of probable PTSD among the Wenchuan earthquake victims in hard-hit areas
were 26.3% [15]. The prevalence rates of probable PTSD, among the survivors
in the hard-hit areas of the Yushu earthquake were 33.7% [16].
4.2 Risk Factors Associated with Probable PTSD
Many studies found numerous associations between the demographics and

PTSD. Among the demographic variables of interest, this study found that
women were at a higher risk of suffering from PTSD than men. When exposed to
Table 4. Multivariate logistic regression analysis of probable post-traumatic stress

disorder (PTSD) symptoms among the survivors of the Lushan earthquake
Variable Total PTSD

OR (95% CI) OR (95% CI)
Overall 500 178(35.6)
Gender Male 242 66(27.3) 1
Female 258 112(43.4) 1.84(1.45 − 2.34)∗∗
Age (years) 15–30 191 50(26.2) 1
31–40 88 39(44.3) 1.71(1.13 − 2.58)∗∗
41–50 135 61(45.2) 2.11(1.09 − 4.06)∗∗
≥ 51 86 28(32.6) 1.52(1.13 − 2.05)∗
Monthly income < 1000 RMB 192 89(46.4) 1
1000–2000 RMB 191 63(33.0) 0.79(0.29-2.17)
2000–3000 RMB 85 21(24.7) 0.61(0.36-1.05)
3000 RMB 32 5(15.6) 0.50(0.16-1.54)
Ethnic group Han 489 173(35.4) 1
Tibetan 11 5(45.5) 1.87(1.45 − 2.42)∗∗
Education level Graduate 4 2(50) 1
Bachelor 144 47(32.6) 0.52(0.15-1.75)
No degree 352 130(36.9) 0.67(0.37-1.22)
Injury of body No 461 161(34.9) 1
Yes 39 17(43.6) 1.88(1.47 − 2.41)∗∗
Family member injured No 458 160(34.9) 1
Yes 42 18(42.9) 1.53(1.13 − 2.07)∗
Family member died No 493 176(35.7) 1
Yes 7 2(28.6) 0.84(0.27-2.06)
Social support 500 1.06(0.87 − 1.29)∗
Based on single main effect logistic regression for each variable-OR is eβ . Significant
at 95% confidence interval does not include 1.
∗p < 0.05, ∗ ∗ p < 0.01.
earthquake trauma, females were nearly twice as likely as males to develop PTSD
(OR = 1.84). This may be partly due to cultural factors, as women in traditional
societies such as China tend to be more dependent, with such disasters causing
their dependence object to encounter heavy loss. Initiatives should give priority
to female, as they were more likely to develop mental health problems than men.
In our study, ethnicity was another significant predictor for PTSD in hard-hit
areas, which agrees with the conclusions of many previous studies [3]. China has
a myriad of ethnic minority groups, including the Tibetan. This study found that
the odds of PTSD (OR = 1.87), in members of the Tibetan ethnic group were
higher compared with the Han ethnic group. After the earthquake, the ethnic
minorities should receive more support, care, and help from the government,
aid organizations, and volunteers. However, a few studies were opposite to our
study [10].
In present study, being injured or having family members injured was also a
significant risk factor for the incidence of probable PTSD in the hard-hit areas.
The findings are consistent with the conclusion of many empirical studies that
the intensity of exposure to a disaster or loss is among the most robust predictive
factors for mental disorders [9].
Many studies have documented the significant relationship between social
support and PTSD in the aftermath of disaster [4]. Consistent with previous
studies on post-traumatic psychological health, the present study confirmed that
social support has a protective function [1].
4.3 Limitations
This study has some limitations. First, these participants belonged to an

extremely hard-hit region. Thus, it may not be appropriate to generalize con-
clusions in the present study to other earthquake-affected regions. Second, self-
report instrument is used, so the participants may be over reported or underre-
ported. Third, in response to the question about history of physical or psychotic
diseases, participants generally reported a history of physical disease, and few
participants reported a history of psychotic disease.
5 Conclusion
Despite the limitations, to the authors’ knowledge, this study has played an
exploratory role in revealing the prevalence and risk factors of probable PTSD
in the hard-hit areas among Lushan earthquake survivors. The findings revealed
that PTSD (35.6%) are common mental health problems in hard-hit areas even
eight months after the earthquake. Female sex, Tibetan ethnicity, being injured,
having family members injured and social support were significant risk factors
in heavily damaged areas. The findings were also one of a handful of studies on
the psychological sequelae of catastrophic natural disasters among non-Western
populations.
Acknowledgements. This research was supported by the National Social Science

Foundation of People’s Republic of China (Grant NO. 11XMZ069), Scientific and Tech-
nological Project of Gansu Province of People’s Republic of China (NO. 148ZCRA012)
and the Special Fund of Fundamental Scientific Research Business Expense for Higher
School of Central Government (Northwest Minzu University).
References
1. Andrews B, Brewin CR et al (2007) Delayed-onset posttraumatic stress disorder:
a systematic review of the evidence. Am J Psychiatry 164(9):1319–1326
2. Elhai JD, Gray MJ et al (2005) Which instruments are most commonly used to
assess traumatic event exposure and posttraumatic effects?: a survey of traumatic
stress professionals. J Trauma Stress 18(5):541–545
3. Galea S, Nandi A, Vlahov D (2005) The epidemiology of post-traumatic stress
disorder after disasters. Epidemiol Rev 27(1):78–91
4. Hafezalkotob A, Alavi A, Makui A (2015) Government financial intervention in
green and regular supply chains: multi-level game theory approach. Int J Manag
Sci Eng Manag 11:167–177
5. Jia K, Zhou S et al (2014) Possibility of the independence between the 2013 Lushan
Earthquake and the 2008 wenchuan earthquake on Longmen Shan Fault, Sichuan,
China. Seismol Res Lett 85(1):60–67
6. Kun P, Chen X et al (2009) Prevalence of post-traumatic stress disorder in Sichuan
Province, China after the 2008 Wenchuan earthquake. Public Health 123(11):703–
707
7. Li H, Wang L et al (2010) Diagnostic utility of the PTSD checklist in detecting
PTSD in Chinese earthquake victims. Psychol Rep 107(3):733–739
8. Mcmillen JC, North CS, Smith EM (2000) What parts of ptsd are normal: intru-
sion, avoidance, or arousal? data from the Northridge, California, earthquake. J
Trauma Stress 13(1):57–75
9. Panda S, Modak NM, Pradhan D (2014) Corporate social responsibility, channel
coordination and profit division in a two-echelon supply chain. Int J Manag Sci
Eng Manag 11(1):22–33
10. Sharan P, Chaudhary G et al (1996) Preliminary report of psychiatric disorder in
survivors of a severe earthquake. Am J Psychiatry 153(4):556–558
11. Van Griensven F, Chakkraband MLS et al (2006) Mental health problems among
adults in tsunami-affected areas in Southern Thailand. JAMA J Am Med Assoc
296(5):537–548
12. Wang L, Zhang Y et al (2009) Symptoms of posttraumatic stress disorder among
adult survivors three months after the Sichuan earthquake in China. J Trauma
Stress 105(5):879–885
13. Wang X, Gao L et al (2000) Longitudinal study of earthquake-related PTSD
in a randomly selected community sample in north China. Am J Psychiatry
157(8):1260–1266
14. Xu J, Song X (2011) A cross-sectional study among survivors of the 2008 Sichuan
Earthquake: prevalence and risk factors of posttraumatic stress disorder. Gen Hosp
Psychiatry 33(4):386–392
15. Zhang Z, Shi Z et al (2011) One year later: mental health problems among survivors
in hard-hit areas of the Wenchuan earthquake. Public Health 125(5):293–300
16. Zhang Z, Wang W et al (2012) Mental health problems among the survivors in the
hard-hit areas of the Yushu earthquake. PLoS ONE 7(10):e46449
How Corruption Affects Economic Growth:
Perception of Religious Powers
for Anti-corruption in Iraq
Marwah Abdulkareem Mahmood1,2(B) , Yizhuang Tian1 ,

and Karrar Abdulelah Azeez1,2
1
Business Administration, School of Management, Harbin Institute of Technology,
Harbin 150001, People’s Republic of China
2
Business administration, Collage of Economics and Administrative,
University of Kufa, Najaf, Iraq
marwahabdulkareem@yahoo.com
Abstract. Iraq is one of the most corrupt countries, as since it was part
of TI’s statistics occupy the last positions. The paper aims to analyze
and understand the perceptions of corruption in Iraq and provide a gen-
eral framework for anti-corruption by social tools. The education society
has selected as study sample by questionnaires randomly at the Univer-
sity students reached (600) student and economics data from 1979 to
2015. The multiple regression analysis, percentages and charts are used
to present the results. The findings indicate that corruption is a factor
behind the deterioration of the country. The regression test show nega-
tive relationship between corruption and economic growth then, religious
powers. In additional, the embezzlement and bribery are more common
forms in state institutions appearances. The study value suggests answers
out scope of government procedures in the fight against corruption that
provide social tools relate with self-behavior for corruption problems.
Keywords: Corruption · Economic growth · Religious power · Educa-

tion society
1 Introduction
The most common definition of corruption is the misuse of public office for
private gain [15]. It has serious consequences in reducing the economic level and
development [5]. As well as a reduction of per capita income [22,24], It hurts
the poor more than the rich [10], to load the economic costs of institutions and
social Large [18] and thus increases the levels poverty [3,9]. Corruption is not
a problem of theft only but includes many forms such as bribery that may be
in payments under the table and fraud, such as falsification of contract, assault,
How Corruption Affects Economic Growth 1467
for example, examination papers, or wasting time or selling state secrets and
nepotism seeking the benefit of friends and relatives for, misappropriation of
funds the public. What makes matters worse is the inability of the beneficiaries
of recording information about the services received [21]. It is worth mentioning
that there is corrupt behavior that considers a rate of return, including bribery
of more than 11–15 times [12]. Some divides corruption into the first two types
[8,20,21], a large corruption, which is intended to official corruption in the upper
levels. Small corrupt practices which routinely incorrect practiced by the staff
at the lower levels. Countries are trying, especially in developing countries to
build effective strategies to fight corruption, in spite of the existence of other
regulatory tools to take the fight against corruption in Iraq, such as the Integrity
Commission and the Financial Inspection Office and the offices of the inspectors
and the media, but discontent and still more from the Iraqi state institutions
reached levels alarming. Show us that since the beginning of the Iraq enters
Corruption Perception-wide and in 2003 for occupies ranked No. 133 and in
spite of increased measures to combat it, but it kept growing, because occupies
ranked No. 167 in 2015. As shown in Table 1. Also, the private sector is weak and
the level of economic growth and income individuals wobbling for these reasons
we focused on Iraq in our study. Here, the two puzzles: first, do you that the
Iraqi economy negatively affected by corruption, or is it similar to the Chinese
economy, which is growing with the increasing levels of corruption? Second, what
motives corruption in the country under study?
In this paper, we investigate into the impact of corruption on both eco-
nomic growth and capita income in Iraq. Second, provide estimates that relative
Table 1. Classification of Iraq, according to the Corruption Perceptions Index
Years Iraq sequence Total

2003 113 133
2004 129 145
2005 137 158
2006 160 163
2007 178 179
2008 178 180
2009 176 180
2010 175 178
2011 175 182
2012 169 174
2013 171 175
2014 170 174
2015 161 167
Source: Transparency Inter-
national publications
1468 M.A. Mahmood et al.
efficiency of the procedures with the anti-corruption organizations in Iraq, the

most important motives common forms of corrupt behavior. Moreover, we found
that corruption is one of the most important reasons impeding economic growth
in the country, economic hardship and poverty of the most important factors
that drive an individual to exercise corrupt behavior. As well as the current gov-
ernment measures are ineffective in the fight against corruption, it may be the
reason the ambiguity that is characteristic of the rules and procedures governing
the functioning of the organizations. [16] refers to resort to innovation may con-
tribute may raise overall performance. We believe that the focus on the ethical
aspect may be the golden key that will end the problem or at least contributes
to the minimization.
This paper has four parts, the Sect. 2 is research methodology. The Sect. 3
includes data analysis and discussion. Section 4 is conclusions.
2.1 Subscribers Properties
The study analyzes the concept of citizens for corruption in Iraq for determina-
tion the actual behavior of society. A questionnaire was distributed to a sample
of university students very (600), considering that young students are the heart
and the leaders of the future. Distributing the questionnaire manually on (714)
students returned them (637) have neglected them (37) to identify because of
their health or for other reasons. It was the response rate (89%). We made this
questioner basic on [20]. In addition, as well as data obtained (37) observa-
tions about corruption and economic growth variables from the organization of
integrity and the Iraqi ministry of planning from 1979 to 2015. The method was
used Multiple regression analysis, percentages, tables, and diagrams in the data
analysis.
2.2 The Main Hypotheses
Bryant and Javalgi [4] found corruption affects economic growth and investment
in each country. Particularly the impact on investment and promise a tax on
profits in developing countries. But in South Korea found that more corruption
leads to more economic growth, as it works on the development of economic
strength [11]. In the same way, Ajao et al. [2] explained the impact of corruption
on economics growth which emphasizes that corruption is the biggest inhibitor
of growth in most countries of the world. The corrupt system gives contracts to
dealers who paid the highest amount of bribes, regardless of the level of efficiency,
as well as a tax in their pockets and not in the state treasury [13]. We summarize
business literature about corruption by three hypotheses:
Hypothesis 1 (H1): The corruption variables negatively can decrease the

growth rate of a gross domestic product.
Hypothesis 2 (H2): The corruption variables negatively can decrease the

growth rate of capita income.
Hypothesis 3 (H3): The religious powers negatively can decrease corruption
level.
2.3 Model Development

The Business literature suggests that corruption has a deleterious effect on eco-
nomic growth through two main reasons: by directly decreasing the efficiency
of administrative systems And generated an atmosphere of frustration, resent-
ment, mistrust and Corrupt behavior carries additional financial costs for society
[6,7,10]. In this section, we set up a statistical model of the relationships among
corruption variables, economic growth and religious powers based on the model
of Gyimah-Brempong [10] and Shadabi [23] as follow:
g = γ0 + γ1 gc + γ2 ec + γ3 tsc + γ4 grc + μ, (1)
where g is growth rate of gross domestic product, gc is general corruption index,

es is employees corruption index, tsc is Transaction speeding corruption index,
grc is Governmental Unities regulations corruption index, γ0 is a coefficient that
estimates the growth rate of national income unassociated with corruption vari-
ables, γ1 is a coefficient that estimates the association between growth rate of
national income and general corruption, γ2 is a coefficient that estimates the
association between growth rate of gross domestic product and employees cor-
ruption, γ3 is a coefficient that estimates the association between growth rate of
gross domestic product and Transaction speeding corruption, γ4 is a coefficient
that estimates the association between growth rate of gross domestic product
and Governmental Unities regulations corruption, μ is a GDP growth error term.
gcapi = γ0 + γ1 gc + γ2 ec + γ3 tsc + γ4 grc + ω, (2)
where gcapi is the growth rate of capita income, and all other variables are as
defined in the equation above, ω is a growth rate of capita income error term.
c = β0 + β1 drp + β2 irp + σ, (3)
where c is corruption index, irp is development religious power, irp is individual

religion power, β0 is a coefficient that estimates the corruption unassociated
with religious powers, β1 is a coefficient that estimates the association between
corruption and development religious power, β2 is a coefficient that estimates the
association between corruption and individual religious power, σ is a corruption
error term.
There is evidence that corruption negatively has affected both the growth
rate of real income and income distribution [10]. We expect the coefficients of
γ1 , γ2 , γ3 and γ4 to be positive on both the growth rate of gross domestic product
and capita income, while in the third model, we expect the coefficients of β1 and
β2 to be negative on the corruption.
3 Data Analysis and Discussion

The first and second models estimate of GDP and capita income growth rates
presented in Table 2. Test statistics is consistent with our hypotheses. The coeffi-
cients of γ1 , γ2 , γ3 and γ4 are negative as expected and are statistically significant
at α < 0.1. The adjusted explanation coefficients are 0.57 and 0.47 respectively,
where f-value greater than scheduled value. This indicates that the growth rate
of GDP and capita income are negatively correlated with corruption variables.
These results support Hypotheses 1 and 2 and agreement with [1,3,9,24].
The last model estimates of corruption level presented in Table 2 panel B.
The coefficients of β1 and β2 are negative as expected, and are statistically sig-
nificant at α < 0.1. The adjusted explanation coefficients is 0.27. This indicates
that the corruption level is negatively associated with development religious
power and individual religious power by −0.39% and −0.43% respectively. This
result is consistent with H3 and means the religious powers have affected cor-
ruption decrease that pushes to find new tools relate to ethics and spiritual
behaviors these results fit with [26]. Shadabi [23] suggested but Contrasted with
Paldam [17] Who pointed out that religious variables are not enough to predict
corruption.
Table 2. Significant testing results of the study models–
Panel A. The regression analysis of corruption variables

Variable Coefficient Model 1 Model 2
Estimate SE (t-statistics) Estimate SE (t-statistics)
Intercept γ0 3.18∗∗∗ 0.29 (11.14) 2.82∗∗∗ 0.29 (9.64)
General corruption γ1 −0.51∗∗ 0.22 (−2.29) −0.24* 0.05 (−5.45)
Employees corruption γ2 −0.10* 0.07 (−1.37) −0.14* 0.10 (−1.42)
Transaction corruption γ3 −0.75∗∗∗ 0.18 (−4.15) −0.90∗∗∗ 0.004 (−3.29)
regulations corruption γ4 −0.18* 0.13 (−1.35) −0.38∗∗ 0.22 (−1.70)
R2 0.62 0.53
Adjusted R2 0.57 0.47
F-value 12.93 8.95
Significant level 0.000 0.000
N 37 37
Panel B. The regression analysis of religious powers - model 3
Intercept β0 3.89∗∗∗ 0.14 (27.43)
Development power β1 −0.39∗∗∗ 0.04 (−11.61)
Individual power β2 −0.43∗∗∗ 0.03 (−12.46)
R2 0.27
Adjusted R2 0.27
F-value 112.42
Significant level 0.000
N 600
Note: ∗ p < 0.1; ∗∗ p < 0.05; ∗∗∗ p < 0.01 standard error in parentheses.
In Fig. 1, We noted the higher percentage of the sample about (90%) of their
responses agree with the concept that argues corruption is a major problem in
the country by selected always scale about (50%) and often scale about (40%).
This is a very high percentage confirms the extent of realization of the corruption
Appearances in the society.
Fig. 1. The basic problem of corruption in Iraq
In comparison the impact of Corruption on economic growth. Figure 2

presents that (82%) of the participants stressed that corruption alters the level
of the country and its development by selected always scale (50%) and often
scale (32%). We prepared some questions to determine the level of corruption
in country and to make evidence whether it contributes to change the coun-
try economics. The analytic result reports that most of the participants agreed
that corruption and contributes significantly to the deterioration of the country.
Lisciandra and Millemaci [14] provided evidence that corruption has negative
impact on economic growth in Italian region and the relation between them is
nonlinear when economic growth fall, corruption increase.
Fig. 2. The Corruption and country economics
In this study, Fig. 3 presents that the embezzlement is more famous common
form to corruption phenomenon in Iraq by determined about (41.3%) of the
sample. While the second form of corruption is bribery by 29.8% of their pointed,
(16%) considered the third forms is favoritism, (8%) selected the extortion, and
the last percentage about (4.7%) considered the information misuse is type of
corruption there. Ajao et al. [2] confided that corruption forms of multiple and
different gradient and conflict of interest, embezzlement, extortion, bribery and
many others, which is a transfer or grab valuables and property to someone
Fig. 3. The most common forms of corruption
who has no right to corruption, but used to do so by virtue of his position, and
resorting to the use of misleading and erroneous information [21].
Wedeman [27] showed that widespread corruption at all levels and in the
center of the large levels. This case is in our study sample consistent with studied
confirmed the higher levels in the state were the most corrupt. Tabish and Jha
[25] used anti-corruption strategies in Indian public sector, they are four latent
constricts leadership, rules & regulations, training, and fear of punishment. The
finding showed all of them are highlighted to help institutions for understanding
the role of anti-corruption. Figure 4 describes the reasons of corruption based on
education society. A large number of participants of the sample about) 30.5%
(believes that the bribes paid based on the employee’s request. Whereas about
(25.2%) underlines the lack of receipt of services in the event of non-payment of
a bribe, (17%) push for speeding up transactions, and (16.1%) in order to avoid
problems, and finally (11.2%) pay because the treatment is not a fundamentalist.
Fig. 4. Reasons for paying bribes to society situation
This study tried to know the rules and capabilities of government for anti-
corruption by education society with five parts because that will make overview
about the problem size and corruption effect. In Fig. 5, the higher percentage is
confirmed that governmental anti-corruption measures have a low efficiency by
(44.2) of the respondents, (31.7%) confirms that the measures originally ineffi-
cient and ineffective, (17.5%) considered the effect is mild, (5.2%) agreed with
Fig. 5. The effectiveness of the procedures currently approved by government
Fig. 6. The main motivations for corruption in Iraqi society
efficient and effective, and (1.3%) were a very effective and influential. Sampford,
et al. [21] Delays in the implementation of judicial proceedings on the corrupt or
non-implementation makes the work of that anti-corruption a few effectiveness.
Lack of public confidence in the judiciary and the court and the police not to
push them to give information about corruption, because they believed that the
judge deal bribery or staff in court and this may be the result of the individual’s
personal experience in this regard. Corruption increases the uncertainty and
skepticism in the judicial and legal system, the presence of authority has indis-
putable sovereignty opportunistic supports corruption. Also, the performance of
organizations will decrease high levels of corruption, because corruption gives
individuals a competitive advantage unfairly [19]. The reasons may include the
failure of recent technology advances being on lower pollution emission intensity
[28] (Fig. 6).
4 Conclusions
This analytic investigating aimed to expand the understanding of the impact of
corruption on economic growth. Iraqi education society considers cultural focus-
ing that has simulation key for anti-corruption. The major problem and barely
missing from the country is Corruption. It also is inhibiting the development of
the nation and the inhibitor of economic development and growth. Poverty or
economic hardship of the individual may be the most important reasons driving
for corruption for the average employee, which paid to bribe demands, on the
other hand, the high cost of luxury, greed, panting behind excellence are paying
the highest levels in the country to engage in corruption. We found evidence that
the corruption levels are detrimental to long-run growth. This evidence is robust
and persistent throughout different studies. The impact of corruption is negative

with respect to economic growth at a high level for each Iraqi society.
FAQ mysterious and government actions ineffective staff gives a chance for
corruption and thus the clear wording of government and striving may prevent
the penetration of corruption in the institutions as well as provide the employee
with a copy of the instructions when receiving the functional tasks. Accordingly,
we can suggest a few considerations. A possible interpretation could be that anti-
corruption by adoption non-financial measures such as moral beliefs, attitudes
of safety and reliability. Finally, Corruption is not only economic but also far
back in history far back in morality and must aim to change the functional ethics
prevailing, and the laying of a plan put moral values purposeful organizations
to clean up corruption and make it fairer, and cannot omission on the role of
education society.
Acknowledgements. The authors would like to thank Kufa University in Iraq for
supporting a research about data collection.
References
1. Aidt TS (2009) Corruption, institutions, and economic development. Oxf Rev Econ
Policy 25(2):271–291
2. Ajao OS, Samuel D, Samuel O (2013) Application of forensic accounting technique
in effective investigation and detection of embezzlement to combat corruption in
Nigeria. Unique J Bus Manage Res 1:65–70
3. Beekman G, Bulte EH, Nillesen EEM (2013) Corruption and economic activity:
micro level evidence from rural Liberia. Eur J Polit Econ 30(30):70–79
4. Bryant CE, Javalgi RG (2016) Global economic integration in developing countries:
the role of corruption and human capital investment. J Bus Ethics 136(3):1–14
5. Chen J, Jie X, Han S (2017) Tourism economic effect divergence analysis-panel
data analysis of Hunan Province. Springer, Singapore
6. Daskalopoulou I (2016) Rent seeking or corruption? an analysis of income satis-
faction and perceptions of institutions in Greece. Soc Sci J 53(4):477–485
7. Gould DJ, Reyes JAA (1983) The effects of corruption on administrative perfor-
mance. World Bank Staff Working Pap 580:2514
8. Graycar A, Monaghan O (2015) Rich country corruption. Int J Publ Adm 38(8):
1–10
9. Gupta S, Davoodi H, Alonso-Terme R (2002) Does corruption affect income
inequality and poverty? Econ Gov 3(1):23–45
10. Gyimah-Brempong K (2002) Corruption, economic growth, and income inequality
in Africa. Econ Gov 3(3):183–209
11. Huang CJ (2016) Is corruption bad for economic growth? evidence from Asia-
Pacific countries. North Am J Econ Finance 35:247–256
12. Jain PK, Kuvvet E, Pagano MS (2016) Corruption’s impact on foreign portfolio
investment. Int Bus Rev 26:25–35
13. Kunieda T, Okada K, Shibata A (2014) Corruption, capital account liberalization,
and economic growth: theory and evidence. Int Econ 139(139):80–108
14. Lisciandra M, Millemaci E (2016) The economic effect of corruption in Italy: a
regional panel analysis. Regional Studies, pp 1–12
15. Lučić D, Radišić M, Dobromirov D (2016) Causality between corruption and the
level of GDP. Econ Res-Ekonomska Istraživanja 29(1):360–379
16. Nazim M, Saeed R et al. (2017) Parametric analysis of leadership styles on orga-
nizational performance and the mediating role of organizational innovativeness.
Springer, Singapore
17. Paldam M (2001) Corruption and religion adding to the economic model. Kyklos
54(2–3):383–413
18. U. S. S. i. S. t. U. t. o. Peace (2013) Governance, corruption, and conflict. In: A
study guide series on PEACE And conflict for independent learners and classroom
instructors, pp 1–54
19. Rocca ML, Cambrea DR, Cariola A (2017) The role of corruption in shaping the
value of holding cash. Finance Res Lett 20:104–108
20. Sahu SK, Gahlot R (2014) Perception about corruption in public servicies: a case
of BRICS countries. J Soc Sci Policy Implic 2:109–124
21. Sampford CJ, Shacklock AH, Connors C (2006) Measuring corruption. Ashgate,
Farnham
22. Sanchez JI, Gomez C, Wated G (2008) A value-based framework for understanding
managerial tolerance of bribery in Latin America. J Bus Ethics 83(2):341–352
23. Shadabi L (2013) The impact of religion on corruption. J Bus Inq 12:102–117
24. Stevens A (2016) Configurations of corruption: a cross-national qualitative com-
parative analysis of levels of perceived corruption. Int J Comp Sociol 57(4):183–206
25. Tabish S, Jha KN (2012) The impact of anti-corruption strategies on corruption
free performance in public construction projects. Constr Manage Econ 30(1):21–35
26. Treisman D (2000) The causes of corruption: a cross-national study. J Public Econ
76(3):399–457
27. Wedeman A (2005) Anticorruption campaigns and the intensification of corruption
in China. J Contemp China 14(42):93–116
28. Yang A, Lan X et al. (2017) An empirical study on the prisoners’ dilemma of
management decision using big data. Springer, Singapore
Procurement Risk Mitigation for Rebar Using
Commodity Futures
Jian Ni, Wei Zhou(B) , and Dan Yang
School of Financial, Southwestern University of Finance and Economics,

zhouwei@swufe.edu.cn
Abstract. This paper presents a model to mitigate the risk of budget

overruns during procurement of raw materials. A case example in rebar
procurement fora typical metallurgical manufacturing company in China
is used. The demand data was supplied by the company and the rebar
price data was obtained from Shanghai Futures Exchanges. The nov-
elty of this study lies in the incorporation of budgetary control into the
objective function of the model. A multistage rebalancing strategy of the
futures position is used to reduce the impact of unexpected changes in
rebar supply price and finished product demand. The complex objec-
tive function is approximated by a quadratic expression, and suboptimal
solutions are obtained for cases where raw the material price is indepen-
dent of the customer demand. When compared with the company’s spot
buying approach without hedging, the results from the model show the
proposed multistage futures balancing model can help reduce the budget
overspending by as much as 32%. Further experiments are carried out
to study the effects of changing price and demand volatilities, and the
results confirm the usefulness of model in mitigating the underlying risk.
Keywords: Raw material procurement · Risk management · Budged

control · Hedging with futures
1 Introduction and Background Literature
This paper is based on a case study in rebar procurement for a Chinese met-
allurgical machinery company that supplies iron and steel mills in China with
drill pipe, drill head, steel making and refining equipment, and blast furnaces.
The commodities markets in China are now mature and well established with a
plentiful supply and fast delivery, and spot procurement is now widely used by
many companies to try and avoid unnecessary inventory built-up. In recent years,
however, the spot prices of steel materials have become much more volatile, and
companies, if they want to remain competitive, cannot safeguard against such
price volatility by raising the final product price or by imposing a contractual
Risk Mitigation for Rebar Using Commodity Futures 1477
surcharge on its customers. Added to this uncertainty of procurement price is

the perennial problem of uncertain rebar requirements, resulting form uncertain
customer demand for the finished products, especially in the prevailing world-
wide economic uncertainties. This problem faced by the company in this study
is not uncommon; it is faced by all manufacturers whose material procurement
is exposed to severe risk arising from fluctuations in both material prices and
demand for the final products.
In response to increasing raw materials price volatility caused by speculation,
and to a lesser extent by volatility in material demand, many manufacturing
companies are now looking at the use of financial hedging policies. Financial
hedging is, of course, not an entirely new concept; much historical discussion
can be found in Myers and Thompson [10], Mathews and Holthausen [8], Lence
[7], McMillan [9], and Alizadeh, et al. [1]. However as suggested by Doege, et al.
[5], financial hedging strategies developed for use in the world of finance are not
directly applicable to risk mitigation in a manufacturing environment. With this
in mind, this study is aimed at developing a dynamic multistage approach that
can be used for budgetary control in material procurement under the specific
business requirements of a manufacturing company. Starting with the futures
contract as the instrument, the multistage hedging can be divided into two steps.
First, a budget is set for a single procurement. Second, the multistage hedging is
initialized and rebalanced at the end of several interim stages. The rebalancing
makes use of updated information on material price and demand. With this type
of dynamic rebalancing of the futures position, the financial hedge can respond
fairly quickly to any unexpected changes in price and demand. At the end of the
final stage, physical delivery takes place and the futures position is settled.
Another risk mitigation approach in material procurement, as reported in
several studies, is the use of flexible supply contracts, such as back-up agree-
ments [6], option contracts for physical delivery [19], and capacity reservation
contracts [16,17]. It is also possible to use the flexible contracts in a synergic
manner (e.g. by building portfolios) to achieve higher performance in procure-
ment risk mitigation [18]. Generally, flexible supply contracts can alleviate the
risk, but such contracts are difficult to price and may require the payment of a
premium to the spot price. However, there are many situations where flexible
supply contracts are not available, as is the case for the company in this study.
In fact, Chod, et al. [3] employed an exponential utility to explore the relation-
ship between operational flexibility and financial hedging opportunity. Using a
similar exponential utility, Ni, et al. [12] then revealed that there can be a weak
separation between the operations decision and financial hedging strategy of a
firm. Other related studies on the interaction between financial hedging and
operations management can be found in Okyay, et al. [20], Sayin, et al. [14], and
Zhao and Huchzermeier [21].
It is only recently that researchers began to study the use of financial hedg-
ing as an approach to mitigate risk in procurement for manufacturing. In this
context, Xu [20] obtained economical hedging strategies by developing an inte-
grated framework model for procurement and inventory risk management using
1478 J. Ni et al.
a financial hedge using the London Metal Exchange (LME) copper futures con-
tracts. Another analytical framework was developed by Caldentey and Haugh [2]
to investigate the appropriate hedging strategies for hedging against a possible
drop in profit of a company when profit is correlated with certain returns in the
financial markets. These two studies suggest that a financial hedging strategy,
if carefully designed, can effectively mitigate the procurement risk. This was
followed by Ni, et al. [11], who develop a multistage hedging strategy to miti-
gate procurement risk using a dynamic price and demand information updating
process, and concluded that using appropriate information to update the hedging
strategy can substantially improve the effectiveness of the hedge. Such updat-
ing makes an interim multistage rebalancing of the futures position possible,
thereby providing a timely adjustment using up-to-date information. This study
differs from that of Ni, et al. [11] in that it develops a hedging strategy that
focuses on budgetary control. The management of the machinery company used
in this study are very concerned with budgetary control and they emphasized the
need to mitigate the risk of budget overruns, especially during the current eco-
nomic uncertainties. With this in mind, a procurement budgetary control model
(PBCM) is developed that aims at controlling the risk of excessive over bud-
get. The PCBM model developed does not seem amenable to a direct solution,
and therefore a heuristic approach had to be used to obtain efficient suboptimal
solutions. Using the heuristic procedure, numerical experiments are conducted
to determine the effectiveness of the multistage hedging strategy in controlling
procurement spending.
In summary, this paper develops a financial hedging strategy for procurement
risk mitigation in the form of an interim multistage rebalancing and dynamic
information updating process, with the general objective to address the need to
maintain adequate budgetary control.
2 The Dynamic Futures Rebalancing Model

2.1 Budget Control and Financial Hedging
Since the global financial turmoil that started towards the middle 2008, the
management of the machinery company has been under increasing pressure to
control the procurement risk arising from both volatile prices and uncertain
demand. As considerable sums are being spent in procuring rebar steel required
for production, the company management is also under pressure to maintain
good budgetary control, and strives as much as possible to keep procurement
spending in line with predetermined budgets. However, budgets can often be
severely exceeded if both steel prices and product demand increase unexpect-
edly. When this occurs, the company may resort to banks loans to finance the
overspending, but the availability of loans at reasonable interest rates is lim-
ited, mainly as a result of the current tight regulatory supervision of banks and
higher Basel capital adequacy requirements. This is partly due to the economic
slowdown in China over the past year or so, and perhaps also to the general
downgrading of creditworthiness of private manufacturing companies as a result
of this slowdown. In light of the scarcity of loans at reasonable or even unreason-

able rates, the company and its peers in China need to maintain tight financial
control over their procurement spending so as to reduce the type of risk that
may lead to bankruptcy.
One existing approach to control procurement spending of a company is to
sign long-term supply contracts with their suppliers and so to lock up the pro-
curement price. However, this type of contract is restricted to fixed quantity
deliveries, and limits the company’s ability to quickly adjust to changing cus-
tomer demand and will generally involve paying a premium to the spot price.
China’s commodity market is now quite mature and sophisticated, and com-
modity futures contracts are available on Shanghai Futures Exchange (SHFE)
that can be used to hedge against the financial risk and to control procurement
spending. As rebar is one of the raw materials that the company uses in very
large quantities, the management is prepared to allow the authors to explore
the use of rebar futures contracts in controlling the procurement risk that it
faces. This case study is not unique, and highlights a general problem associated
with commodity procurement risk management, viz., easy access to the com-
modity markets has encouraged many manufacturers to gradually shift to spot
procurement to avoid excessive inventory built-up and achieve high flexibility
in supply management. But excessive world-wide speculative activity in modern
commodity markets has at the same time significantly increased price volatility.
This is evidenced by the fact that only a small proportion (some 3% for copper)
of futures contracts are actually settled through physical delivery, the rest can
only be regarded as speculative and/or risk hedging and are settled in cash.
2.2 Risk Mitigation Model Development

The design of financial hedging strategies for commodity procurement is distinct
in nature from their counterparts in the financial industry. It is not only the
uncertainties in the underlying volume that arises from the changes or updates
of customer orders that can extensively complicate the problem, but also the dif-
ferent hedging objectives (budgetary control in this paper) which require tailor-
made hedging strategies to be developed. In addition, the transaction cost is
usually negligible in hedging against the risk in procurement for a manufacturer
[19], while this cost may be highly significant for hedging in the financial indus-
try. This low transaction cost is mainly due to the fact that transaction costs
involved in the futures contracts are very low (less than 0.02% in SHFE) in
comparison to the value of the trade.
In the context of our model, the company that has to procure a certain
commodity for future production (rebar for the company used in this study),
where both commodity price and demand are expected to fluctuate. A multistage
hedging approach is used with the aim of achieving a good overall hedge that
can effectively reduce the possibility of budget overruns in procurement. Such a
multistage hedging procedure is illustrated in Fig. 1. The time period or planning
horizon in which we want to implement the financial hedge is T, which is divided
into T stages t = 0, 1, · · · , T − 1, T . For the general class of problems that we are
1480 J. Ni et al.
Fig. 1. The multistage hedging procedure
dealing with one stage can be a period of one week to one month, depending on
the availability of updated information for both commodity price and customer
demand. As shown in Fig. 1, at the beginning a budget plan is produced by the
management that serves as a benchmark that the subsequent financial hedging
strategy aims to meet. Using the price and demand information at stage 0, the
multistage financial hedge is initialised by entering into a long position of the
commodity futures that expire at the end of the final stage. At the subsequent
stages, the futures position will be rebalanced to achieve the best hedge using
the latest information available at each stage. At the end of the final stage,
the actual procurement of the physical commodity takes place and the futures
position is settled in cash. The cash settlement will completely or partially offset
any overspending incurred because of the variability in both commodity price
and customer demand.
The multistage hedging strategy is aimed at balancing the uncertainty of
the unhedged procurement spending, C, which is defined as the product of the
volatile commodity price P and the uncertain procurement quantity X, or C =
P X. The hedge can generate a payoff from the cash settlement of the futures
position, which leads to the hedged procurement spending as follows:
C̄ = C − H. (1)
For the multistage hedging strategy, the payoff can be calculated by analyzing
the stage-by-stage initialisation/rebalancing volumes yt and the futures position
zt at each stage t (1 ≤ t ≤ T − 1). The variables yt and zt are related as follows:

yt , t=0
zt = (2)
zt−1 + yt , ∀ 1 ≤ t < T.
Given the price of the commodity futures Ft (1 ≤ t ≤ T − 1) at the beginning

of stage t that will mature at the end of stage T − 1, the payoff H can be
calculated as follows:
T
−1
H = P × zT −1 − Ft × y t . (3)
t=0
Note that we do not consider the transaction cost. This is because of the low
rebalancing frequency (weekly at most) and the low transaction cost relative to
the settlement amount ratio (less than 0.02% in SHFE). From Eqs. 1 and 3, the
hedged procurement spending will be:
T
−1
C̄ = P (X − zT −1 ) + Ft × y t . (4)
t=0
A hedging strategy is said to be effective when the hedged procurement

expense C̄ is less exposed to the volatilities of the commodity price and demand
than the unhedged spending C. The design of an effective multistage hedging
strategy needs to incorporate information updating processes for the material
price and for the demand. These processes are used in deciding the initialisa-
tion/rebalancing volumes (yt ). In this study, the information updating process
is modelled by filtration [2], a method often used to adequately model the timely
T −1
availability of information. Generally, a filtration is an increasing family {Ft }t=0
of σ-algebra Ft [13]:
0 ≤ s < t < T ⇒ Fs ⊂ Ft ⊂ F. (5)
T −1
These σ-algebras {Ft }t=0 can be regarded as the available information set
ateach stage. Let Mt (Ω, Ft )be the function space that is composed of all the
Ft -measurable random variables, i.e.,
Mt (Ω, Ft ) = {ξ |ξ is a random variable on Ω, and Ft − measurable } .
The hedging decisions are based on the available information at each stage,
and any admissible adjusting volume yt must satisfy the following constraint:
yt ∈ Mt (Ω, Ft ) .
The multistage hedging strategy formulated below aims at keeping the pro-
curement spending in line with an allocated budgeted. To achieve this aim, a pro-
curement budgetary control model (PBCM) is developed which attempts min-
imise the expected budget overrun (EBO). The model can be stated as follows:
+
minY EBO = E C̄ − K . (6)
T
−1
s. t. C̄ = P (X − zT −1 ) + Ft · y t , (7)
t=0
zt = zt−1 + yt , ∀ 1 ≤ t < T, (8)
z0 = y 0 , (9)
yt ∈ Mt (Ω, Ft ) , (10)
+
here the objective function E C̄ − K is the EBO, which measures the aver-
age overspending with respect to the given budget K.
1482 J. Ni et al.
3 The Multistage Financial Hedging Strategy

3.1 Heuristic Solution
+
The discontinuity of E C̄ − K in the objective function of the PBCM model
developed above makes it difficult, if not impossible, to find a close-form solution
that can be easily implemented. This inherent complexity in the PBCM model
requires the use of some heuristic to generate suboptimal, though still effective
and potentially useful, solutions. In this section, a heuristic is developed by first
replacing the PBCM model by an α-PBCM model, where the value of α is set
heuristically. The approximate model, α-PBCM, can then be solved using the
closed-form results of Ni, et al. [11]; an exact solution to the approximate model.
It is decided to approximate the objective function as:
+ 1
E C̄ − K ∼
2
= E C̄ + α · E C̄ − K −K , (11)
2
where

E C̄ − K
α = 2 . (12)
E C̄ − K
α needs not to be constant. However, it is expected that α will be relatively

insensitive to changes in the hedging strategy used, suggesting that it should
be possible to use a constant value
for α in Eq. (11) and still produce a good
+
quadratic approximation for E C̄ − K . In this study, α is determined heuris-

tically using the following expression:

E C̄(Ŷ ) − K
α̂ = 2 , (13)

E C̄(Ŷ ) − K
T
where Ŷ = {ŷi }i=0 , ŷ0 = E0 (X), and ∀ 1 ≤ t ≤ T , ŷt = 0.
In the heuristic used, the value of α̂ is calculated on the assumption of a static
hedge. It is in fact a simplified way to take into account the effect of financial
hedging on the value of α. The effectiveness of α̂ is examined in Sect. 4.
When α is constant, the objective of the PBCM model can be approximated
by an α-PBCM model as follows:

2
min E C̄ + α C̄ − K , (14)
Y
where Y is constrained by Eqs. (7), (8), (9), and (10).

In general, and also in the company used in this study, customer demand
and material price are independent. Under this assumption the α-PBCM model
can be further simplified as:
Assumption: if for every stage t(0 ≤ t < T ) X and Ft+1 are independent,
then:
Et [Ft+1 X] = Et [Ft+1 ] Et [X] , ∀ 0 ≤ t < T. (15)
Under this Assumption, a close-form solution of the α-PBCM model can be

obtained using the analytical results of Ni, et al. [11], because α-PBCM model
has a similar quadratic objective function. While the α-PBCM and the model
of Ni, et al. [11] do share a group of constraints in the multistage hedge, some
modifications to two auxiliary variables are still required as follows:

Et+1 [X] − zt−1 , 0 < t < T
X̂t = (16)
E1 [X] , t = 0,
⎧
⎨
s−1
K− Ft · y t , 0 < s < T
K̂s = t=0 (17)
⎩
K, s = 0.
Using the analytical results (specifically Theorem 3) in Ni, et al. [11], a

closed-form solution to the α-PBCM can be obtained as:

E (Ft − Ft+1 )(Ft+1 X̂t − K̂t + 2α
1
) |Ft
yt∗ = − . (18)
2
E (Ft − Ft+1 ) |Ft
3.2 Information Updating of Price and Demand

The proposed multistage hedging strategy reacts to updates in the material price
and the demand. The dynamic initialisation/rebalancing of the futures position
from stage 0 to stage T − 1 requires the predictions of the future material prices
and demand levels, as well as some measure of their volatility. Such predictions
are made at each stage of the planning horizon using the updating stochastic
processes.
For the stochastic material price model, the long term/short term model of
Schwartz and Smith [15] is employed to simulate the dynamics. When is the
commodity spot price at time t, the process can be stated as:
ln (St ) = χt + ξt , (19)
dχt = −κχt dt + σχ dWt , (20)
dξt = μξ dt + σξ dBt , (21)
where dWt and dBt are correlated as dWt dBt = ρdt. Equations (19), (20) and
(21) produce the estimates, see Ni, et al. [11], as follows:
Et (Ft+1 ) = Ft A(t), (22)

2 2
Et Ft+1 = Ft B(t), (23)
1484 J. Ni et al.

t+1 λχ −κ(T −s)
where A(t) = exp t
λξ + κ e ds
t+1
λχ −κ(T −s)
B(t) = exp 2 λξ + e + k(T − s) ds . (24)
t κ
In connection with the stochastic product demand, Choi, et al. [4] describe
a Bayesian information update used in a two-stage application. However, for
the machinery company used in this study, the Bayesian method needs to be
extended to a multistage process. The raw material (rebar) required in this
company for physical procurement is an aggregation of customer orders over τ
stages. It is not unreasonable to assume that customer orders arrive randomly
as a Poisson process. If λ is the average number of customer orders that are
received in one stage, it follows that the number of orders Ñ arriving in τ stages
is a random number from the Poisson distribution:
−λτ
ke
P {Ñ = k} = (λτ ) . (25)
k!
The predictions of Ñ can be improved using the information on customer

orders received at each stage. The expected value (Et [Ñ ] for 0 ≤ t ≤ T − τ ) can
be calculated using Eq. (25) as follows:
Et [Ñ ] = λτ. (26)
For the case when T ≥ t > T − τ , Et [Ñ ] can be calculated as follows:
Et [Ñ ] = q + λ(T − t), (27)
where q is the number of orders received in stage T − τ + 1 to t.

Customer orders are fulfilled by successful production of the specified end-
products, which can only be done if the required amounts of various raw materials
are in stock. The raw material required can be computed almost instantly (2
days) after customer orders are received. However customer orders often vary in
size and due dates, and although the order size in the near future can be closely
predicted by managers with good relationships with the customers, accurate
long-term predictions are still difficult to make. So for certain raw materials,
rebar being a good example, the amount needed to serve customer orders at a
future point in time is variable, but the variability will reduce as that point in
time approaches. If we let d˜ be the expected (average) amount of rebar needed
to serve the orders of a single customer in stages T − τ + 1 to T, the reduction
in variability can then be expressed as:
˜ ∼ N (dt , Σt )
Et [d] (28)
and
Σt1 ≥ Σt2 , when t1 ≤ t2 (29)

Since the material demand can be computed almost instantly upon receiving
customer orders, the planned rebar supply quantities, D̃, required in stages T −
τ + 1 to T can be estimated as follows:
˜
D̃ = Ñ · d. (30)
The value of Et [D̃] at stage t can then be calculated as:

dt λτ, 0≤t≤T −τ
Et [D̃] = (31)
dt (q + λ(T − t)) , T − τ < t ≤ T.
4 Case Study Example

After analyzing the historical price data (weekly observations of prices for rebar
futures contracts in Shanghai Futures Exchange from 03/2009 to 03/2011), the
parameters of the Kalman filter [13] are estimated as shown in Table 1.
Table 1. Estimation of the price model parameters
Parameter Description Estimate Standard Error

κ Short-term mean-reversion rate 1.0837 0.0132
σχ Short-term volatility 0.3241 0.0056
λχ Short-term risk premium 0.121 0.0126
μ Equilibrium drift rate 0.0095 0.0005
μ Equilibrium risk-neutral drift rate 0.0213 0.0024
σω Equilibrium volatility 0.2976 0.0112
ρ Correlation in increment 0.6132 0.0466
There are 6 months in the planning horizon, corresponding to 6 stages in

the model. The initial price of rebar is set at 4,500 RMB per tonne and the
budgeted unit price is set at 5,000 RMB per tonne, which corresponds to a total
procurement budget of 5000×E(D̃) Yuan. According to the purchasing manager
of the company, a customer order arrives every two weeks and raw material is
procured every month, so λ = 4.3 per month and τ = 1. The expected demand is
assumed to take unit values, i.e., d0 = 1. This assumption is acceptable because
the purpose of the Monte Carlo simulation experiments is to analyse risk and
volatility in relative terms, and no generality is lost by such an assumption as long
as the volatility Σt applies equally to the two strategies that are being compared,
i.e. hedged and unhedged. From the historical data for rebar quantities required
by the company, it is estimated that Σ0 = 0.532 . It then follows that:
T −t
Σt = Σ0 , ∀ 0 ≤ t ≤ T − 1. (32)
T
1486 J. Ni et al.
To obtain the heuristic solution, the parameter also needs to be estimated

using Eq. (13); resulting in α = 1.21 × 10−3 .
Finally, with Eq. (18) together with the price and stochastic demand updating
processes, the heuristic multistage hedging model is completely defined for the
specific rebar case example. Monte Carlo simulation is used to generate a large
sample (5,000) of the rebar price and of the demand. The performance of the
proposed multistage hedging strategy is then evaluated in terms of the expected
budget overrun (EBO, see Eq. (6) and the standard deviation (SD). The results
are summarised in Table 2.
Table 2. Performance evaluation of different strategies (RMB)
SD EBO
No hedge 1,482.70 1,038.20
Multistage strategy 1,176.30 701.5
Table 2 shows that both the SD and EBO are significantly reduced by
the multistage financial hedge. The significant reduction of SD (from 1,482.7
to 1,176.3) indicates that the stability of the overall procurement spending is
improved by the multistage hedge. Further, the substantial reduction in EBO
(from 1,038.2 to 701.5) suggests that the procurement budget is much less vio-
lated under the multistage hedge than that with no hedging. Therefore, it can
be safely concluded that the multistage hedging strategy is effective in procure-
ment risk mitigation from the procurement budget perspective. Moreover, the
percentage reduction in EBO for the multistage hedge is larger than SD; EBO is
reduced by 32.4% while SD is reduced by 20.6%, see Table 2. These reductions
suggest that the proposed multistage hedging strategy is more effective in reduc-
ing overspending of the procurement budget than in reducing the uncertainty of
the overall total procurement spending.
A sensitivity analysis is conducted to judge the effectiveness of the multistage
hedging strategy when the price and demand become highly volatile, and also
when their volatility can be regarded as low. The sensitivity analysis first exam-
ines the influence of changing volatility in demand. This is done by assessing
the performance of the hedge with different levels of demand volatility, i.e., by
increasing/decreasing the standard deviation Σ0 through a changing ratio, such
that:
SDchanged = SDinitial × (1 + changing ratio). (33)
The influence of changing price volatility, represented by both the short-term
volatility σχ and the long-term equilibrium volatility σχ , is also investigated.
Here these two volatilities are simultaneously increased or decreased by a ratio
percentage. With either demand or price volatility changing by a ratio ranging
from −70% to 70%, the effects of such changes on the hedging performance are
shown in Fig. 2 (demand volatility) and Fig. 3 (price volatility).
Fig. 2. Sensitivity analysis of demand volatility
Fig. 3. Sensitivity analysis of price volatility
Figure 2 shows that changing the volatility of demand will affect the perfor-
mance of the multistage strategy, an obvious result. However, the hedged EBO
remains below the unhedged EBO for all changes in Σ0 , suggesting that the
effectiveness of the multistage hedging strategy is always superior. When Σ0
decreases, the hedged EBO is reduced by a larger ratio than the unhedged EBO,
indicating that the procurement budget is less likely to be violated. This sug-
gests that the multistage hedging strategy becomes relatively more powerful as
Σ0 decreases. On the other hand, the reduction ratio of the hedged EBO when
compared to the unhedged EBO will decrease as Σ0 increases. So, the multi-
stage hedging strategy will be relatively less powerful as increases, an obvious
result, since any procurement policy can be expected to deteriorate as volatility
increases.
1488 J. Ni et al.
Figure 3 shows that the performance of the multistage strategy is affected by

the price volatility, also an obvious result. On one hand, the multistage hedge will
become relatively more powerful as σχ and σω increase, as shown by the widening
difference between the hedge and the unhedged EBO. While on the other, the
multistage hedge becomes less powerful when σχ and σω decreases, as shown by
the narrowing difference between the hedged and unhedged EBO. Furthermore,
Fig. 3 shows that the difference between the hedged and unhedged EBO will
reduce sharply when the price volatility decreases significantly. This indicates
that price volatility plays a major role in the performance of the multistage
hedge model.
5 Concluding Comments
The financial hedging problem presented in this paper is modelled as a mul-

tistage strategy, and data from a machinery company is used to analyse the
expected budget overrun risk in procuring rebar. A closed-form heuristic solu-
tion is obtained for the model when the raw material price is independent of
customer demand. The novelty of the proposed multistage hedging strategy is
that it addresses the importance of budget control, and hence the risk of pos-
sible overspending in commodity procurement is mitigated. Another distinctive
feature of the model is the rebalancing of commodity futures position at the
intermediate stages of the planning horizon. This rebalancing allows appropri-
ate adjustments of the commodity futures position to be made at various stages,
by taking into account new trends in the stochastic price movement, and up-to-
date changes in the prediction of amount required.
The effectiveness of the model is tested via a case study in the supply of rebar
for a machinery company. The results confirmed the usefulness of the proposed
multistage hedging strategy in reducing procurement risk arising from variable
finished product demand and raw material prices when using a predetermined
level of budgetary control. This model developed in this study can be extended
in several directions. First, it could be beneficial to investigate the use of other
types of financial derivatives, such as options contracts, as the instrument of
the hedge. Second, the multistage hedging approach could also be expanded
to mitigate other types of operational risks, such as inventory risk and currency
risk. Another possible extension is to use a penalty (cost) function for budgetary
control, as under-spending by a certain amount is generally less serious than
overspending by the same amount.
Acknowledgement. This research was supported by the National Natural Science

Foundation of China (Grant No. 71601159; 71203185 and 71620107005) and the project
of Research Center for System Sciences and Enterprise Development (Grant No.
Xq16C03).
References
1. Alizadeh AH, Nomikos NK, Pouliasis PK (2008) A markov regime switching app-
roach for hedging energy commodities. J Bank Financ 32(9):1970–1983
2. Caldentey R, Haugh M (2006) Optimal control and hedging of operations in the
presence of financial markets. Math Oper Res 31(2):285–304
3. Chod J, Rudi N, Van Mieghem JA (2010) Operational flexibility and financial
hedging: complements or substitutes? Manage Sci 56(6):1030–1045
4. Choi TM, Li D, Yan H (2003) Optimal two-stage ordering policy with bayesian
information updating. J Oper Res Soc 54(8):846–859
5. Doege J, Schiltknecht P, Lüthi HJ (2006) Risk management of power portfolios
and valuation of flexibility. OR Spectr 28(2):267–287
6. Eppen GD, Iyer AV (1997) Backup agreements in fashion buying-the value of
upstream flexibility. Manage Sci 43(11):1469–1484
7. Lence SH (1995) On the optimal hedge under unbiased futures prices. Econ Lett
47(3–4):385–388
8. Mathews KH, Holthausen DM (1991) A simple multiperiod minimum risk hedge
model. Am J Agric Econ 73(4):1020–1026
9. Mcmillan DG (2005) Time-varying hedge ratios for non-ferrous metals prices. Res
Policy 30(3):186–193
10. Myers RJ, Thompson SR (1989) Generalized optimal hedge ratio estimation. Am
J Agric Econ 71(5):858–868
11. Ni J, Chu LK et al (2012) A multi-stage financial hedging approach for the pro-
curement of manufacturing materials. Eur J Oper Res 221(2):424–431
12. Ni J, Chu LK, Yen BPC (2015) Coordinating operational policy with financial
hedging for risk-averse firms. Omega 59:279–289
13. Øksendal B (2006) Stochastic differential equations - an introduction with appli-
cations. Universitext 51(10):1731–1732
14. Sayın F, Karaesmen F, Özekici S (2014) Newsvendor model with random supply
and financial hedging: utility-based approach. Int J Prod Econ 154(4):178–189
15. Schwartz ES (1997) The stochastic behavior of commodity prices: implications for
valuation and hedging. J Financ 52(3):923–973
16. Serel D (2007) Capacity reservation under supply uncertainty. Comput Oper Res
34(4):1192–1220
17. Serel DA, Dada M, Moskowitz H (2001) Sourcing decisions with capacity reserva-
tion contracts. Eur J Oper Res 131(3):635–648
18. Shi Y, Wu F et al (2011) A portfolio approach to managing procurement risk using
multi-stage stochastic programming. J Oper Res Soc 62(11):1958–1970
19. Spinler S, Huchzermeier A, Kleindorfer P (2003) Risk hedging via options contracts
for physical delivery. OR Spectr 25(3):379–395
20. Xu Y (2006) Procurement risk management using commodity futures: a multistage
stochastic programming approach/. Int J Prod Res
21. Zhao L, Huchzermeier A (2015) Operations–finance interface models: a literature
review and framework. Eur J Oper Res 244(3):905–917
Ubiquitous Healthcare and Ubiquitousness
of Chronic Disease Prevention and Control:
Theory and Design
Zhihan Liu(B)
School of Public Administration, Central South University,

Changsha 410083, People’s Republic of China
liuzhihan@csu.edu.cn
Abstract. With the gradual maturing of technologies such as cloud

computing, high-speed mobile Internet and Internet of things, to realize
the Ubiquitous Healthcare in the health field is not only an irresistible
trend, but also a realistic possibility. According to the chronic disease
management theory and the service-oriented architecture, this study pro-
posed the designing scheme and implementation methods of a mobile ser-
vice platform for the big data collection and analysis of chronic disease
based on cloud computing. With this platform, community residents can
enjoy the convenient and quick health self-management services, while
community healthcare staff can provide timely and efficient services to
the patients with chronic diseases or those potentially high-risk individu-
als precisely. This paper is aimed to provide theoretical reference for the
ubiquitous healthcare research specifically in chronic disease prevention
and control, and also help in the practice of upgrading community health
service and promoting a supply-side structural reform in chronic disease
management.
Keywords: Big data · Chronic disease · Cloud computing · Health

management · Ubiquitous healthcare
1 Introduction
Community healthcare service was identified by the World Health Organiza-
tion (WHO) as an effective measure for chronic disease prevention and control.
However, since China started the reform of urban healthcare system and the
development of community healthcare service in 1997, the condition of suffer-
ing from chronic diseases has not been improved as much as expected and the
medical expenses have also risen rapidly. Why is that the community healthcare
service in China did not have satisfactory effects on residents’ health? In addi-
tion to the service object factors (e.g., people’s health consciousness, lifestyle
and stereotype of treatment selection) and the policy factors (e.g., government
investment and allocation of health resources), the traditional form of commu-
nity healthcare service, the content homogenization with hospitalization service,
Chronic Disease Prevention and Control 1491
and the maladaptation to the needs of patients with chronic diseases are also
worthy of consideration.
In fact, it is difficult to achieve the goal of building an integrated system
of prevention and control towards the growing outbreak of chronic diseases just
through the traditional medical information management. In 2002, Kirn [6] for-
mally put forward the concept of Ubiquitous Healthcare (u-Healthcare or u-
Health). It is a kind of information system configuration enabling individual con-
sumers to access any types of health services through mobile computing devices
whenever and wherever [1–3,5,8–13]. This concept actually originated from the
idea of Ubiquitous Computing (ubicomp or Pervasive Computing) in the field of
information science and technology. Transcending the traditional desktop com-
puting, ubiquitous computing is a bran-new computing mode aimed at fusion
of cyberspace and physical space, under which people could acquire the digital
services freely and transparently [7].
For being humanistic, ubiquitous computing was doomed to accomplish great
deeds in the fields of medicine and healthcare (especially in the health manage-
ment area) which are closely related to human health and well-being, since the
day of its birth. As the 10th International Conference on Ubiquitous Healthcare
held in Yokohama, Japan in 2013 declared: “Enhancement in the welfare for
future requires the change to the current healthcare system. Our concern for
the healthcare is shifting from ‘recovery from illness’ to ‘maintaining wellness
and improving quality of life’. For the care of daily health level we need special
kinds of methods and technologies that we can be applied into our daily life
smoothly” [10]. Many developed countries have established relatively complete
u-Health systems with the following typical architecture [11] (see Fig. 1). Based
on the development of Internet of Things (IoT) and cloud computing, the med-
ical industry in Taiwan, China, has also witnessed the transition from “e-Health”
to “m-Health” and eventually to “u-Health”, in a short span of ten years. How-
ever, due to the late start of ubiquitous computing research in mainland China,
previous work on u-Health (especially for the chronic diseases management) is
quite limited at this time [12,13]. To addresses the problem, this study is the
first one in mainland China to explore the application of u-Health in the chronic
disease scenario. Not only the theory and the functions were introduced, but
also the technical route and the implementation methods were presented, aim-
ing at helping the Chinese government implement its comprehensive strategy for
chronic disease prevention and control.
2 Functional Requirements of Mobile Community

Health Management Service Platform for Chronic
Diseases Based on Cloud Computing
Chronic Disease Management (CDM) includes health information collection,
health data storage, health risk assessment, chronic disease monitoring, health
interventions, etc. All these tasks provides natural basis for the functional unit
division of a mobile community health management service platform for chronic
1492 Z. Liu
Fig. 1. Typical architecture of u-Health system
diseases. It is important to note that this study adopted the theory of Yang [14]
which insists that the receptors of community prevention and control should not
only include the patients with chronic diseases, but also the high-risk groups and
even the healthy ones, from the perspective of community health management.
Thus, the application system of this platform can be divided into two subsys-
tems: community health service center and community residents (not just the
patients). The main goal of the system is moving forward the “strategic pass” of
CDM to help in cost control under the guidance of health data real-time acqui-
sition, and carrying out a systematic, coordinated and integrated strategy on
chronic disease prevention and control by in-depth health data exploring. The
specific steps include: (a) collecting, managing and analyzing the personal health
information used in health risk factors assessment, tracking, and health behavior
guidance, for dynamical monitoring of community residents’ state of health; (b)
applying health interventions to the community residents based on the evalu-
ation results; and (c) tracking and evaluating the effects of interventions (see
Fig. 2).
3 Implementation Methods of Mobile Community

Health Management Service Platform for Chronic
Diseases Based on Cloud Computing
3.1 Technical Route
The system adopted the MySQL database for data storage, and used Java lan-
guage for realizing the background management function and the apps develop-
ment based on Android and IOS. Given the complexity of apps on the community
health management service platform, the platform was designed in according to
Service-Oriented Architecture (SOA), providing functions such as service defin-
ition, development, deployment, and running, etc. Logic reuse was implemented
Fig. 2. System functional requirements of the service platform
through the loose coupling between the layers, while the loose coupling between
modules on the same layer was implemented through the form of services. All of
these equipped the platform with good ability of extension.
3.2 Overall Architecture

The technical implementation of the service platform function involved various
kinds of data acquisition sensors, data transferring networks, storage databases,
cloud computing services and apps for data processing, interaction with users,
and many other complex factors. In order to improve the robustness of the
program and make the design more reasonable, this study adopted a five-layer
architecture (see Fig. 3).
Each layer has a clear job assignment as follows:
(1) Perception layer (PL): To be in charge of the perception, identification, col-
lection and processing of residents’ health data (under automatic control),
and connecting data acquisition devices with network layer and application
layer through communication module;
(2) Network layer (NL): To realize the function of information transmission
(i.e., uploading collected data and downloading information from applica-
tion layer), routing, and control. This layer is made up with core network
and access network, and depend on the public telecommunication networks
and the Internet;
(3) Support services layer (SSL): To set up a basic support service platform for
the processing of CDM related information and operations. Specifically, it at
least covers a SOA platform, an integration service platform for mass data,
a cloud computing service platform, a communication service platform, and
an IoT service platform;
(4) Application layer (AL): To carry and realize the various function of commu-
nity health management, i.e., health risk assessing, chronic disease monitor-
ing, health intervention and outcome evaluation;
(5) Interaction layer (IL): To provide users with a variety of interactive terminals
(such as PC, mobile phones, etc.), realize the interaction of information and
instructions between people and system.
1494 Z. Liu
Fig. 3. Overall architecture of the mobile service platform for community health man-
agement of chronic disease
3.3 Public Service Modules

Four public service modules were designed uniformly as the system implemen-
tation process may require cloud computing, massive data, IoT, communication
and other public services on many aspects.
(1) Big data integration services: According to the characteristics of massive
structured data and unstructured data in the vast distributed network envi-
ronment, the platform provided big data integration services based on SOA,
including organization, storage, sharing, release, analysis and mining, etc.
(2) Cloud computing services: Considering ubiquitous healthcare are higher than
current Internet several orders of magnitude on the amounts of data stor-
age and operation, a cloud computing solution was needed to be integrated
into the platform design on the basis of traditional IT system, to improve

the computing ability. For the community healthcare information centers,
cloud computing could make full use of the great deal of idle computing and
storage capacity, and provide important technical means for adaption to the
flexibility of business growth and reduction of application deployment cost.
(3) Services based on IoT capacity: By accessing to a variety of IoT terminals
and applications through standardized protocols and various adapter, the
platform was provided with data acquisition and transmission capacity of
IoT terminals, and finally realized the interaction and cooperation between
machine and machine, and between machine and people.
(4) Services based on communication capacity: This module included the com-
munication capacity required, such as SNS, WeChat, positioning, and call
centers, etc. The platform may also need some other third-party services and
resources.
4 Conclusion
Just as Dr. Robert M. Kaplan, who was Associate Director for Behavioral and
Social Sciences in the National Institutes of Health of US at that time said in his
paper: “Health-related information collected in psychological laboratories may
not be representative of people’s everyday health. For at least 70 years, there
has been a call for methods that sample experiences from everyday environ-
ments and circumstances” [4]. Ubiquitous healthcare not only provides us with
this method, but also with the development direction of medical information con-
struction. Under the u-Health theory and aiming at the key requirements and
typical applications of CDM, this study is the first in mainland China to present
a mobile service platform for the big data collection and analysis of chronic
disease, on the comprehensive basis of cloud computing, intelligent sensing and
perception technologies of IoT, data acquisition, storage and processing tech-
nologies, and computer decision support technology. The platform can be used
in the health data acquisition, supervision and forecast for community residents,
contributing to improve the community health services and lift the population
health level. It is expected to help in realizing the humanistic self-health manage-
ment of community residents, and providing community health centers with the
ubiquitous, comprehensive prevention and control services of chronic diseases,
so as to accelerate the supply-side structural health reform in the field of CDM.
In consideration of the rapid progress in the international development of infor-
mation technology, direction for future research would be getting the platform
put into practice as soon as possible, and adjusting and improving the system in
a timely manner according to the process of Chinese healthcare system reform.
Acknowledgement. This study was supported by National Natural Science Foun-

dation of China (grant No. 71603289), China Postdoctoral Science Foundation (grant
No. 2016M590759), Social Science Fund of Hunan Province (grant No. 15JD61), and
Postdoctoral Science Foundation of Central South University (No. 153656). The author
declares no conflicts of interest.
1496 Z. Liu
References
1. Arnrich B, Mayora O, Bardram J (2010) Pervasive or ubiquitous healthcare? Meth-
ods Inf Med 49:65–66
2. Conejar RJ, Kim HK (2015) Designing u-healthcare web services system. Int J
Softw Eng Appl 9(3):209–216
3. He C, Fan X, Li Y (2013) Toward ubiquitous healthcare services with a novel
efficient cloud platform. IEEE Trans Biomed Eng 60(1):230–234
4. Kaplan RM, Stone AA (2013) Bringing the laboratory and clinic to the community:
mobile technologies for health promotion and disease prevention. Ann Rev Psychol
64:471–498
5. Kim J, Ahn CW (2011) Diabetes management system based on ubiquitous health-
care. J Paediatr Child Health 12(3):133
6. Kirn S (2002) Ubiquitous healthcare: The OnkoNet mobile agents architecture. In:
Objects, Components, Architectures, Services, and Applications for a Networked
World. Springer, Heidelberg, pp 265–277
7. Krumm J (2010) Ubiquitous computing fundamentals. Ergonomics 53(5):724–725
8. Panagiotakopoulos T, Fengou MA et al (2009) Ubiquitous healthcare. In: Biocom-
putation and Biomedical Informatics: Case Studies and Applications, p 254
9. Shin D, Shin D, Shin D (2016) Ubiquitous healthcare platform for chronic patients.
In: International Conference on Platform Technology and Service, pp 1–6
10. Tamura T, Park KS (2016) Invitation to the 10th International Conference on
Ubiquitous Healthcare (u-healthcare 2013). http://u-healthcare2013.l-bmi.org/
11. Touati F, Tabish R (2013) U-Healthcare system: state-of-the-art review and chal-
lenges. J Med Syst 37(3):1–20
12. Wu X, Ye M et al (2012) Pervasive medical information management and services:
key techniques and challenges. Chin J Comput 35(5):827–845
13. Xie Y (2016) Ethical perspectives of ubiquitous healthcare. J Jishou Univ (Social
Science Edition) 37(3):56–62 (in Chinese)
14. Yang JX (2010) Thoughts about strategies for chronic disease integrated control
on the basis of community health management. Chin Health Econ 29(7):67–69 (in
Chinese)
Can Media Supervision Improve the Quality
of the Firms Internal Control?
Tian Yang1,2 , Qianwei Ying3(B) , and Xingyao Li3

1
The College of Literature and Journalism, Sichuan University,
2
School of Overseas Education, Sichuan University,
3
Business School, Sichuan University, Chengdu, People’s Republic of China
yingqw@scu.edu.cn
Abstract. Using the sample of firms listed in China Growth Enterprise

Market (GEM) from 2011 to 2015 we empirically test the effect of media
coverage on the corporate internal control quality and the quality of inter-
nal control information disclosure. We found that the higher the degree
of media coverage would lead to higher quality of internal control and
information disclosure. The empirical results of this paper indicate that
media can act as an important external supervision mechanism besides of
state laws and regulations to regulate the behavior of corporate internal
control, to protect the interests of investors, and to enhance investors’
confidence.
Keywords: Media coverage · Internal control · Information disclosure
1 Introduction
With the development of information technology and the spread of internet,
media’s functions as supervision and governance are taking more and more sig-
nificant roles in economy. Media, regarded as a forth power being independent of
legislation, administration and judiciary, plays an essential role in disseminating
information, restraining incompliant actions of government and businesses, and
enhancing market’s efficiency. Increasing number of scholars in economics and
management have focused on studying the corporate governance role of media
as well as its influence mechanism. Dyck and Zingales [8] found that the pressure
from media supervision contributes drastically to reducing the personal interests
of the controlling shareholders from their controlling rights. Dyck and Zingales
[9] discovered that media supervision is able to effectively suppress or mod-
ify the corporate decisions that may harm external investors’ interests. Miller
[20] made a conclusion that through rebroadcasting investigation of information
from information intermediary like securities analysts or original ones, media
can disclose accounting frauds in advance. Joe et al. [15] claimed that an ineffi-
cient board would take actively measures to improve efficiency after exposed by
1498 T. Yang et al.
media. Maistriau and Bonardi [19] reported that media visibility has a positive
effect on CSR performance of a sample of British firms, and Zyglidopoulos et al.
[22] found that media coverage is positively associated with heightened levels of
CSR performance of a sample of S&P 500 firms.
Using a large cross-country sample of firms, Ghoul et al. [10] further found
strong evidence that firms engage in more CSR activities if located in coun-
tries where the media has more freedom. Dai et al. [5] investigated whether the
media plays a role in corporate governance by disseminating news. Using a com-
prehensive data set of corporate and insider news coverage for the 2001–2012
period, they showed that the media reduces insiders’ future trading profits by
disseminating news on prior insiders’ trades available from regulatory filings.
Based on a sample of over two million newspaper articles, Kim et al. [17] found
the media in China has an incremental impact on stock price efficiency, and a
market-driven media can play the role of compensating for the underdeveloped
governance institutions in transitional economies such as China.
Although prior studies accumulate evidence that supports the monitoring role
of the media, they ignore an important issue: Can media supervision encourage
firms to take precautions and initiate to improve the quality of internal con-
trols and information disclosures? Previous literature regards media supervision
as a substitute of internal controls in corporate governance, focuses on corpo-
rate inherent characteristics only when investigating determinants of corporate
internal controls, and neglects the potential interaction between media’s exter-
nal supervision and corporate internal controls. Internal controls mean a series
of restrictive organizing, planning, procedures and methods executed in entities,
aiming to boost operating efficiency, acquire and use resources in an effective
manner, and achieve given objectives. The setting-up, perfection and valid oper-
ation of business internal controls have direct impact on disclosures of financial
information and investors’ interests. Therefore, launching a research on the way
media supervision influences corporate governance has theoretical and practical
meanings. This article selects data of GEM firms in China’s stock market from
2011 to 2015, and studies impacts of media coverage on internal controls and
related information disclosures. The contributions of this paper mainly lie in the
following two aspects.
Firstly, this paper enriches the literature on the corporate governance role of
media. Existing studies on the corporate governance role of media mainly focus
on the effect of media coverage on firms’ socially beneficial behaviors such as
CSR performance, or on firms’ frauds and misconducts hurting the interests of
the investors. This paper extends the scope of this research area by investigating
the effect of media coverage on firms’ internal control quality and internal control
information disclosure.
Secondly, this paper also enriches the literature on the determinants of inter-
nal control quality and disclosure. Most prior studies on the determinants of
internal control quality and disclosure mainly focus on corporate internal charac-
teristics (e.g., [4,7,11]), while very few researches investigated the role of external
factors, such as auditor expertise (e.g., [12,21]). In this paper, we find that media
Can Media Supervision Improve the Quality of the Firms Internal Control? 1499
coverage, as another important external governance mechanism, also contributes

significantly to the firms’ internal control quality and disclosure.
2 Prior Literature and Hypotheses Development
As an important segment in corporate governance, the quality of internal con-

trols plays a key role in business operation and development, as well as the
protection of the interests of investors. Existing research indicated that good
internal controls would suppress earning management [1], and reduce financing
cost of firms effectively [16].
Most previous literature about the determinants of internal control quality
mainly focused on corporate internal characteristics. Doyle et al. [7] examined
determinants of weaknesses in internal control for 779 firms disclosing material
weaknesses from August 2002 to 2005. They found that these firms tend to be
smaller, younger, financially weaker, more complex, growing rapidly, or under-
going restructuring. Firms with more serious entity-wide control problems are
smaller, younger and weaker financially, while firms with less severe, account-
specific problems are healthy financially but have complex, diversified, and
rapidly changing operations. Guo et al. [11] investigated the role of employment
policies in reducing internal control ineffectiveness and financial restatements.
They found that employee-friendly policies significantly reduce the propensity for
employee-related material weaknesses. Moreover, they illustrated that financial
restatements caused by unintentional errors are less likely to arise in firms that
invest more in employee benefits. Chen et al. [4] documented a negative relation
between board independence and the disclosure of internal control weaknesses.
They also documented that the negative relation is stronger for firms with uni-
tary leadership (combined positions of CEO and Chairman) than for firms with
dual leadership.
Only a few studies studied the effects external factors on internal control,
and these studies mainly focus on the role of audit expertise. Haislip et al. [12]
argued that greater audit firm IT expertise promotes improved internal controls
for their clients, especially those controls that are dependent on IT. Schroeder
and Shepardson [21] found that internal control audits initially provided inter-
nal control quality benefits. However, after the 2007 auditing standards change,
internal control quality deteriorated for ICFR audited versus unaudited firms.
An important external governance mechanism, media supervision, has been
ignored by prior studies on the determinant of internal control quality. Most
countries, particularly the transitional countries, are facing problems of imper-
fect legal system and inadequate protection of the investors’ interests. Among the
alternative governance mechanisms besides of legal system, media supervision is
one of the most important. Most scholars hold an optimistic opinion towards the
role of media supervision playing in corporate governance. Dyck and Zingales [8]
found that media supervision could reduce the personal interests of the control-
ling shareholders on the sacrifice of other shareholders. Miller [20] showed that
the media plays a positive role in the process of revealing accounting scandals.
1500 T. Yang et al.
Dyck and Zingales [9] further found that the media exposure could encourage
enterprises to correct violations of the rights and interests of external investors.
Zyglidopoulos et al. [22] found that media coverage is positively associated with
heightened levels of CSR performance. Dai et al. [5] showed that media cover-
age can reduce insiders’ future trading profits by disseminating news on prior
insiders’ trades available from regulatory filings.
We can conclude that existing researches have provided sufficient evidences
for the corporate governance role of media supervising. Therefore, we ascertain
that media coverage can urge firms to improve internal control quality, and
reduce adverse actions to investors as far as possible, to avoid greater negative
market response after exposed by media. At the same time, media coverage also
encourages firms to build a good image in external stakeholders and enhance
business values through improving internal controls. Hence, we can obtain the
following hypothesis:
Hypothesis 1: Media coverage has a positive effect on the firms’ quality of internal
controls.
One important characteristic of internal controls is that they consist of internal
managerial processes that are not directly observable for investors. Investors’
judgment of the effectiveness of these internal controls is based on the informa-
tion disclosed by managers.
From shareholders’ perspective, higher levels of internal control disclosure
reduce investors’ information risk, lower down their required rate of return (e.g.,
[3,6]). At the same time, higher levels of disclosure enable shareholders to mon-
itor managers more closely, resulting in fewer agency problems.
From the manager’s perspective, the decision to voluntarily disclose informa-
tion is based on a tradeoff between the expected benefits and costs of disclosing
[14]. The main benefit of voluntarily disclosing information on the firm’s inter-
nal controls is that it adds to managerial reputation building. Not having a
reputation for credible reporting not only reduces the effectiveness of the man-
ager’s communication efforts, but also adversely affects her reputation in the
managerial labor market [18]. Establishing a reputation for credible reporting
requires disclosure of accurate and timely information as well as of information
that is complete [14,20]. This means that career concerns incentivize a manager
to voluntarily disclose information, even if the information is not favorable [3].
Hammersley et al. [13] find a significantly negative abnormal return following
a firm’s announcement that the firm’s internal controls were not effective, but
more importantly also document that the adverse effects on returns are more
pronounced when the firm’s managers claim that the internal controls are effec-
tive but the independent auditor report indicates that they were not.
An important cost of voluntarily disclosing information on the firm’s internal
controls is that it may have potential legal consequences. If a manager discloses
inaccurate information or if a manager discloses incomplete information man-
agers could be sued, face legal liability, and owe damages [3]. Moreover, once this
matter is known to the public, it will also adversely influence the manager’s rep-
utation on the managerial labor market. To guarantee the accuracy and quality
of the information disclosures on internal controls, the manager might need to

pay extra efforts and related costs.
The above analysis and discussion on the benefits and costs of disclosing
internal controls information implicate the importation role of the managers’
reputation concerns. We can expect that the reputation mechanism will even be
more important when a firm is in the spotlight of media coverage, since media
can spread information to magnify the influence of signals and scope of audience,
drawing more investors’ attention [2]. When the media coverage is higher, the
reputational benefit of disclosing high quality information of internal control
will be larger, and the reputational cost of concealing information or disclosing
poor information on internal controls will be greater. Therefore, higher media
coverage can lead to higher quality of internal control information disclosure.
We can conclude the second hypothesis as follows:
Hypothesis 2: Media coverage has a positive effect on the quality of internal
control information disclosure.
3 Research Design
3.1 Data and Variables
This article selects GEM Growth Enterprise Market firms from 2011 to 2015 as
initial sample, and eliminates observations with missing data on the key vari-
ables of this paper. The reason why we choose GEM firms is that these firms,
which have small sizes and young ages in most cases, are opaque in information
disclosure and have higher business risks in operation. In this case, the quality
of internal controls and information disclosure of GEM firms are of great impor-
tance for investors to understand their business situations. Besides, this article
uses balanced panel data, keeping only the firms with public financial data in
each year from 2011 to 2015. The final sample of this paper includes 151 GEM
firms, 755 observations in the period ranging from 2011 to 2015.
The corporate internal control quality index and its information disclosure
quality index derive from the DIB Internal Control and Risk Management Data-
base. Higher values of the two indexes represent higher qualities of the internal
control and information disclosure respectively. The data of media coverage is
from the News and Reports Database of WIND Info, which records the Chinese
listed firms’ daily news reported by more than 100 major financial newspapers
and websites in China, covering nearly all the relevant news about public firms
from major financial media in China. Using this database, we count the annual
number of news reports for each GEM listed firm, and construct a media coverage
indicator by taking the natural logarithm of one plus the number of new reports.
Referring to existing research results about determinants of internal control
quality and information disclosure quality (e.g., [7]), our paper includes a set of
basic financial characteristics and corporate governance variables as control vari-
ables. The basic financial characteristics variables include the firm size (natural
logarithm of total assets), asset-liability ratio, return on assets, and sales growth.
1502 T. Yang et al.
The corporate governance variables include ownership concentration (the total

share ratio of the top 10 major shareholders), executives’ ratio of shareholding,
equity separation degree (the quadratic sum of the share ratio of the top five
major shareholders), and the separation of ownership and control. In addition,
we also control the industry dummies. The data of these control variables is
from CSMAR database. Continuous variables are censored at the 1st and 99th
percentiles to mitigate the effect of extreme values. The descriptive statistics of
variables are shown in Table 1.
Table 1. Descriptive statistics of variables
Na Description Mean S.D. Median Min Max

IC Internal control quality index 6.833 0.405 6.891 5.165 7.524
Disclosure Information disclosure of Internal control index 0.382 0.098 0.37 0.18 0.58
Media Media concern extent 2.58 0.837 2.477 1.097 4.628
Size Corporate size 20.785 0.528 20.77 19.65 22.25
LEV Asset-liability ratio 0.171 0.123 0.142 0.028 0.619
ROA Return on Assets 0.07 0.037 0.065 −0.003 0.164
Growth Sales growth 0.317 0.307 0.263 −0.21 1.512
Ten Ownership concentration 0.678 0.089 0.691 0.448 0.857
Mshare Executives’ shareholding 0.305 0.239 0.358 0 0.708
Herf 5 Equity separation degree 0.16 0.087 0.142 0.022 0.392
Separation Separation of control and ownership 3.662 6.873 3.224 0 31.216
3.2 Empirical Modelling

Based on the preceding theoretical analysis, this paper established the following
model (1) and (2) to analyze the effects of the media coverage on the qualities
of internal control and information disclosure of internal control respectively:
ICi,t =β0 + β1 Mediait + β2 LEVit + β3 ROAit + β5 Growthit + β6 Tenit
+ β7 Herl 5it + β8 Msharei,t + β9 Separationi,t (1)
+ Industry and year fixed effects + εit ,
Disclosurei,t =β0 + β1 Mediait + β2 LEVit + β3 ROAit + β5 Growthit
+ β6 Tenit + β7 Herl 5it + β8 Msharei,t + β9 Separationi,t (2)
+ Industry and year fixed effects + εit ,
where IC represents the internal control quality index. Disclosure represents the
quality of internal control information disclosure, and Media represents the extent
of media coverage. The definitions and descriptions of other control variables are
as shown in Table 1. Meanwhile, Industry and year fixed effects are also controlled
in the above regressions. This article uses the OLS regression model, and clusters
standard errors of the coefficients at firm level for more accurate statistical infer-
ence [18]. β0 is the constant term in regression, while β1 to β9 represent the regres-
sion coefficients of the explanatory and control variables. We are mainly interested
in the coefficient β1 , which indicates the effect of media coverage on internal con-
trol quality in model (1) or internal control disclosure in model (2).
4 Empirical Results and Analyses

4.1 Regression Results
Following the above regression models (1) and (2), this paper conducts regres-
sion analyses respectively with internal control quality (IC) and its information
disclosure quality (Disclosure) as dependent variables. The regression results are
shown in Table 2.
In the column 1 of Table 2, we used media coverage as the sole explana-
tory variable, without controlling other firm characteristic variables. The result
shows that the regression coefficient of media coverage is positive and significant
at 10% significance level, with a value of 0.099. This regression result indicates
Table 2. The media coverage’s effects on the quality of internal control and internal
control information disclosure
−1 −2 −3 −4 −5 −6
IC IC IC Disclosure Disclosure Disclosure
Media 0.099∗ 0.161∗∗ 0.169∗∗∗ 0.146∗∗ 0.168∗∗∗ 0.153∗∗
− 1.89 − 2.15 − 2.66 − 2.13 − 2.72 − 2.49
Size 0.297∗∗∗ 0.137∗∗ 0.135∗∗ 0.131∗∗
− 3.86 − 2.27 − 2.15 − 1.98
LEV − 0.03 − 0.029 − 0.196∗ − 0.17
(− 0.31) (− 0.30) (− 1.94) (− 1.49)
ROA 0.256∗∗ 0.180∗∗ 0.245∗∗ 0.255∗∗
− 2.39 − 1.99 − 2.11 − 2.36
Growth 0.069 0.081 0.195∗ 0.181
− 0.89 − 1.11 − 1.71 − 1.5
Ten 0.004 0.111
− 0.04 − 0.96
Mshare 0.104 0.06
− 1.14 − 0.65
Herf 5 − 0.086 0.039
(− 0.91) − 0.34
Separation − 0.05 0.07
(− 0.56) − 0.89
constant 0.935∗∗∗ − 0.755∗∗∗ 0.549 1.116∗∗∗ − 0.645∗∗∗ 1.211∗∗∗
− 18.4 (− 3.21) − 1.39 − 23.32 (− 2.89) − 2.71
Industry dummy YES YES YES YES YES YES
N 755 755 755 755 755 755
R2 0.158 0.188 0.236 0.133 0.207 0.212
∗∗∗ ∗∗ ∗
Note: , , represents significance at 1%, 5% and 10% level respectively.
The t-statistics reported in parentheses are based on standard errors clustered by firm.
1504 T. Yang et al.
that media coverage has a positive and significant effect on internal control qual-
ity. In column 2, we controlled the basic financial characteristics of the firms,
including firm Size (Size), asset-liability ratio (LEV), return on assets (ROA),
and the firms’ sales growth (Growth). The positive regression coefficient of media
coverage, with a value of 0.161, is significant at 5% level, indicating that the pos-
itive relationship between media coverage and internal control quality becomes
more apparent after controlling the influences of these variables of firms’ finan-
cial characteristics. In column 3, we further controlled the corporate governance
characteristics, including ownership concentration (the total share ratio of the
top 10 major shareholders), executives’ share holding, equity separation degree,
and the separation of ownership and control. The results show that after the
control of the corporate governance, media coverage illustrates an even greater
positive effect (β1 = 0.169, p < 0.1) on the internal control quality. Thus, we
can conclude that media coverage does help improve the quality of the firm’s
internal control, which is consistent with Hypothesis 1 above.
Column 4 to 6 use quality of internal control information disclosure as depen-
dent variable. It could be found that whether using media coverage as an explana-
tory variable alone, or controlling the firm’s basic financial characteristics and
characteristics of corporate governance, there remains a positive effect of media
coverage on the quality of internal control information disclosure. Thus, the
media supervision not only helps to improve the quality of internal control of the
company, but also help to improve the company’s internal control information
disclosure, thereby enhancing the company’s transparency and self-restraint.
4.2 Robustness Test
In the above regressions, the measure of media coverage includes the news reports
from the mainstream financial media channels in China, but excludes the Wind
Info’s own news reports. Although Wind Info is not a common source of infor-
mation for retail investors, it is widely regarded by institutional investors and
analysts as one of the important sources of information access. In order to test
the reliability of the previous empirical results, this paper uses an alternative
measure of media coverage, which includes Wind Info’s own news reports, to
do the robustness test. The regression results in Table 3 show that the main
conclusions of the paper have not changed because of the replacement of media
coverage measure. With the alternative measure of media coverage, denoted as
M edia A, the regression results still show that media coverage has a positive
effect on the quality of internal control and the quality of internal control infor-
mation disclosure, and this conclusion remains the same regardless of whether
we control the company’s basic financial characteristics or corporate governance
variables.
Table 3. The robustness test using alternative measure of media coverage
−1 −3 −4 −1 −3 −4
IC IC IC Disclosure Disclosure Disclosure
Media A 0.142∗ 0.186∗∗ 0.205∗∗ 0.129∗ 0.180∗∗ 0.193∗∗
− 1.81 − 2.08 − 2.39 − 1.71 − 2.1 − 2.46
Size 0.207∗∗∗ 0.133∗∗ 0.132∗∗ 0.135∗∗
− 3.01 − 2.15 − 2.03 − 2.18
LEV 0.042 − 0.012 − 0.190∗ − 0.154
− 0.43 (− 0.12) (− 1.81) (− 1.29)
ROA 0.232∗∗ 0.182∗∗ 0.248∗∗ 0.260∗∗
− 2.18 − 2.07 − 2.16 − 2.48
Growth 0.035 0.017 0.188 0.175
− 0.35 − 0.17 − 1.62 − 1.43
Ten 0.003 0.105
− 0.03 − 0.9
Mshare 0.117 0.072
− 1.28 − 0.78
Herf 5 − 0.055 0.068
(− 0.59) − 0.58
Separation − 0.085 0.093
(− 0.91) −1
cons 0.658∗∗∗ − 0.767∗∗∗ 0.728∗ 1.726∗∗∗ − 0.649∗∗∗ 1.375∗∗∗
− 8.86 (− 3.26) − 1.75 − 19.84 (− 2.80) − 2.93
N 755 755 755 755 755 755
R2 0.168 0.194 0.234 0.121 0.194 0.21
Note: ∗∗∗ , ∗∗ , ∗ represents significance at 1%, 5% and 10% level respectively.
The t-statistics reported in parentheses are based on standard errors clustered
by firm.
5 Conclusion
Based on the financial data of 151 listed companies in GEM from 2011 to 2013,
this paper analyzes the impact of media coverage on the internal control qual-
ity and its information disclosure quality. The results show that media coverage
positively relates to the quality of internal control and the quality of internal
control information disclosure, and the conclusion is still valid after control-
ling corporate financial characteristics and corporate governance characteristics.
This article provides new evidence for the corporate governance role of media
supervision from the perspective of corporate internal control. The conclusion of
this paper implicates that, under the imperfect Chinese legal system and market
institutional environment, media, as an alternative form of supervision mecha-
1506 T. Yang et al.
nism other than legal system, plays an important role in regulating the behavior
of listed companies and protecting the interests of small and medium sharehold-
ers. In the current stage of China’s market-oriented reform, strengthening the
external supervision role of the media is of great significance to improving cor-
porate governance of listed companies and protecting the interests of investors.
Acknowledgements. The authors would like to acknowledge the financial support

from the National Science Foundation of China (NSFC71373167), the social sci-
ence research fund of Sichuan University (skqx201608), and Excellent Young Scholars
Research program of Sichuan University, (2013SCU04A32).
References
1. Ashbaugh-Skaife H, Collins DW et al (2008) The effect of sox internal control
deficiencies and their remediation on accrual quality. Acc Rev 83(1):217–250
2. Barber BM, Odean T (2008) All that glitters: the effect of attention and news
on the buying behavior of individual and institutional investors. Rev Finan Stud
21(2):785–818
3. Campbell JL, Chen H et al (2014) The information content of mandatory risk
factor disclosures in corporate filings. Rev Acc Stud 19(1):396–455
4. Chen Y, Knechel WR et al (2016) Board independence and internal control weak-
ness: evidence from sox 404 disclosures
5. Dai L, Parwada JT, Zhang B (2015) The governance effect of the media’s news
dissemination role: evidence from insider trading. J Acc Res 53(2):331–366
6. Dan D, Hogan C et al (2011) Internal control disclosures, monitoring, and the cost
of debt. Acc Rev 86(4):1131–1156
7. Doyle J, Ge W, Mcvay S (2007) Determinants of weaknesses in internal control
over financial reporting. J Acc Econ 44(1–2):193–223
8. Dyck A, Zingales L (2004) Private benefits of control: an international comparison.
J Finan 63(2):537–600
9. Dyck A, Zingales L (2008) The corporate governance role of the media: evidence
from Russia. J Finan 63(3):1093–1135
10. Ghoul SE, Guedhami O et al (2016) New evidence on the role of the media in
corporate social responsibility. J Bus Ethics 62:1–29
11. Guo J, Huang P et al (2015) The effect of employee treatment policies on inter-
nal control weaknesses and financial restatements. Soc Sci Electron Publishing
91(4):1167–1194
12. Haislip JZ, Peters GF, Richardson VJ (2016) The effect of auditor it expertise on
internal controls. Int J Acc Inf Syst 20:1–15
13. Hammersley JS, Myers LA, Shakespeare C (2008) Market reactions to the disclo-
sure of internal control weaknesses and to the characteristics of those weaknesses
under section 302 of the sarbanes oxley act of 2002. Rev Acc Stud 13(1):141–165
14. Healy PM, Palepu KG (2001) Information asymmetry, corporate disclosure, and
the capital markets: a review of the empirical disclosure literature. J Acc Econ
31(1–3):405–440
15. Joe JR, Louis H, Robinson D (2009) Managers’ and investors’ responses to media
exposure of board ineffectiveness. J Finan Quant Anal 44(3):579–605
16. Kim JB, Song BY, Zhang L (2011) Internal control weakness and bank loan con-
tracting: evidence from sox section 404 disclosures. Acc Rev 86(4):1157–1188
17. Kim JB, Yu Z, Zhang H (2016) Can media exposure improve stock price efficiency
in China and why? China J Acc Res 9(2):83–114
18. Kothari SP, Shu S, Wysocki PD (2009) Do managers withhold bad news? J Acc
Res 47(1):241–276
19. Maistriau EA, Bonardi JP (2014) How much does negative public exposure on envi-
ronmental issues increase environmental performance? Academy of Management
Annual Meeting Proceedings 2014(1):11328
20. Miller GS (2006) The press as a watchdog for accounting fraud. J Acc Res
44(5):1001–1033
21. Schroeder JH, Shepardson ML (2015) Do sox 404 control audits and management
assessments improve overall internal control system quality? Acc Rev A Q J Am
Acc Assoc 91:1513–1541
22. Zyglidopoulos SC, Georgiadis AP et al (2012) Does media attention drive corporate
social responsibility? J Bus Res 65(11):1622–1627
Fuzzy Chance Constrained Twin Support Vector
Machine for Uncertain Classification
Qilin Cao, Xiaodan Fu(B) , and Yaomin Guo

lovesens@163.com
Abstract. In this paper, we propose a new fuzzy chance constrained

twin support vector machine (FCC-TWSVM) via chance constrained
programming formulation and fuzzy membership, which can deal data
with measurement noise efficiently. This paper studies twin support vec-
tor machine classification when data points are uncertain statistically.
With some properties known for the distribution, the FCC-TWSVM
model is used to ensure the small probability of misclassification for
the uncertain data. The FCC-TWSVM model can be transformed to
second-order cone programming (SOCP) by the properties of moment
information of uncertain data and the dual problem of SOCP model is
introduced. In addition, we also show the performance of FCC-TWSVM
model in artificial data and real data by numerical experiments.
Keywords: Fuzzy data mining · Support vector machine · Robust opti-

mization · Chance constraints · Venture capital
1 Introduction
Nowadays, support vector machines (SVMs) are considered as one of the most
effective learning methods for classification, which emerged from research of sta-
tistical learning theory [10,14]. The main idea of this classification technique is
by mapping the data to the higher dimensional space with some kernel meth-
ods and then determining a hyperplane separating binary classes with maximal
margin [9,19,26].
In recent years, SVM classification methods have made breakthrough progress
and enjoyed great success in many fields. Binary data classification methods have
made breakthrough progress in recent years. Mangasarian et al. [15] proposed gen-
eralized eigenvalue proximal support vector machine (GEPSVM). Jayadeva et al.
[12] proposed a twin support vector machine (TWSVM) to solve the classification
of binary data, motivated by GEPSVM. The main idea of TWSVM is generating
two nonparallel planes, one of which is closest to one class and another plane is
as far as possible from the other. At the same time, the ν-TWSVM [18] was pro-
posed for handling outliers as an extension of TWSVM. Some extensions to the
TWSVM can be founded in [8,13,20].
Fuzzy Chance Constrained Twin Support Vector Machine 1509
For the methods mentioned above, the parameters in the training sets are
implicitly assumed to be known exactly. However, in real world application,
the parameters have perturbations since they are estimated from the data sub-
ject to measurement and statistic errors [11]. When the data points are uncer-
tain, different have been proposed to formulate the traditional SVM with uncer-
tainties. Bi et al. [5] assumed the data points are subject to an additive noise
which is bounded by the norm and proposed a very direct model. However, this
model cannot guarantee a generally good performance on the uncertainty set.
To guarantee an optimal performance when the worst case scenario constraints
are still satisfied, robust optimization is utilized. Trafalis et al. [22,23,25] pro-
posed a robust optimization model when the perturbation of the uncertain data
is bounded by norm. Robust optimization [4,21] is also used when the constraint
is chance constraint which is to ensure the small probability of misclassification
for the uncertain data. Based on different bounding inequalities, Ben Tal et al.
[1,2] employed moment information of uncertain training points to developing
different chance-constrained SVM model. However, to our best knowledge, there
is no researcher considering the chance-constrained optimization in TWSVM
problem. Therefore, it is interesting and important to study the TWSVM with
chance constraints for the uncertain data classification problem. The main pur-
pose of this paper is to make an attempt in this direction.
Combining the capability of processing the uncertainty of chance constraints
and the benefits of TWSVM, in this paper, we propose a fuzzy chance con-
strained twin support vector machine (FCC-TWSVM). The main method of
this paper is to apply the moment information of uncertain data to transform
into second order cone programming (SOCP). The rest of this paper is organized
as follows. Section 2 recalls SVM and TWSVM briefly. In Sect. 3, we introduce
the model of FCC-TWSVM. Experimental results on the uncertain data sets are
presented in Sect. 4. Conclusions are provided in Sect. 5.
2 Preliminaries
In this section, we briefly recall some concepts of TWSVM and CC-TWSVM for
binary classification problem.
2.1 TWSVM
Consider a binary classification problem of l1 positive points and l2 negative

points (l1 +l2 = l). Suppose that data points belong to positive class are denoted
by A ∈ Rl1 ×n , where each row Ai ∈ Rn (i = 1, · · · , l) represents a data point
with label +1. Similarly, B ∈ Rl2 ×n represents all the data points with the label
−1. The TWSVM determines two nonparallel hyperplanes:
T T
f+ (x) = w+ x + b+ = 0 and f− (x) = w− x + b− = 0, (1)
1510 Q. Cao et al.
where w+ , w− ∈ Rn , b+ , b− ∈ R. Here, each hyperplane is close to one of the two

classes and is at least one distance from the other class points. The formulation
of TWSVMis as follows:
min 1 Aw+ + e+ b+ 22 + C1 eT− ξ
w+ ,b+ 2 (2)
s.t. −(Bw+ + e− b+ ) + ξ ≥ e− , ξ ≥ 0
and
min 1 Bw− + e− b− 22 + C2 eT+ η
w− ,b− 2 (3)
s.t. (Aw+ + e+ b+ ) + η ≥ e+ , η ≥ 0,
where C1 , C2 are positive numbers, e+ , e− are vectors of ones of corresponding
dimensions. The nonparallel hyperplanes Eq. (1) can be obtained by solving Eqs.
(2) and (3). Then the new point is classified by following decision function
xT wr + br = min | xT wr + br | . (4)
r=+,−
2.2 Chance Constrained Twin Support Vector Machine
When uncertainties exists in the data points, the TWSVM model need to be
modified to contain the uncertain information. Suppose there are l1 and l2 train-
ing data points in Rn , use A i = [Ai1 , · · · , Ain ], i = 1, · · · , l1 to denote the
uncertain data points and the label is positive +1. And let B i = [Bi1 , · · · , Bin ],
i = 1, · · · , l2 to denote the uncertain data points and the label is negative −1
respectively. Then A = [A1 , · · · , A
l ]T and B = [B 1 , · · · , B
l2 ]
T
represent two
1
data sets. The chance-constrained program is to ensure the small probability of
misclassification for the uncertain data. The chance-constrained TWSVM(CC-
TWSVM) formulation is

l1
2 E{Aw+ + e+ b+ 22 } + C1
1
min ξi
w+ ,b+ i=1
i w+ + b+ ) ≤ 1 − ξi } ≤ ε (5)
s.t. P{−(B
ξi ≥ 0, i = 1, · · · , l1
and

l2
2 E{Bw− + e− b− 22 } + C2
1
min ηi
w− ,b− i=1
i w− + b− ) ≤ 1 − ηi } ≤ ε (6)
s.t. P{(A
ηi ≥ 0, i = 1, · · · , l2 .
where E{·} denote the expectation under corresponding distribution, C1 , C2 are
positive numbers, e+ , e− are vectors of ones of corresponding dimensions, 0 <
ε < 1 is a parameter close to 0 and P{·} is the probability distribution. The
model ensures an upper bound on the misclassification probability.
3 Fuzzy Chance Constrained Twin Support

Vector Machine
In this section, we propose a chance constrained twin support vector machine to
process uncertain data points. When uncertainties exists in the data points, the
TWSVM model need to be modified to contain the uncertain information. Like
CC-TWSVM, suppose there are l1 and l2 training data points in Rn , use A i =
[Ai1 , · · · , Ain ], i = 1, · · · , l1 to denote the uncertain data points and the label is
positive +1. And let B i = [Bi1 , · · · , Bin ], i = 1, · · · , l2 to denote the uncertain
data points and the label is negative −1 respectively. Then A = [A l ]T
1 , · · · , A
1
and B = [B 1 , · · · , B
T
l2 ] represent two data sets. By considering the contribution
of different misclassified points, the chance-constrained program is to ensure the
minimum misclassification rate for the uncertain data. The chance-constrained
TWSVM(FCC-TWSVM) formulation is

l1
2 E{Aw+ + e+ b+ 22 } + C1
1
min ti ξi
w+ ,b+ i=1
i w+ + b+ ) ≤ 1 − ξi } ≤ ε (7)
s.t. P{−(B
ξi ≥ 0, i = 1, · · · , l1
and

l2
2 E{Bw− + e− b− 22 } + C2
1
min ti ηi
w− ,b− i=1
i w− + b− ) ≤ 1 − ηi } ≤ ε (8)
s.t. P{(A
ηi ≥ 0, i = 1, · · · , l2 .
where E{·} denote the expectation under corresponding distribution, C1 , C2 are
positive numbers, ti ∈ (1, 1] denotes the fuzzy membership of positive and neg-
ative samples, 0 < ε < 1 is a parameter close to 0 and P{·} is the probabil-
ity distribution. The model ensures an upper bound on the misclassification
probability, but two quadratic optimization problems (7) and (8) with chance
constrained are typically non-convex, so the model is very hard to solve.
The work so far to deal with the chance constraint is to transfer them by dif-
ferent bounding inequalities. When the mean and covariance matrix of uncertain
data points are known, then multivariate bound [3,16,17] by robust optimization
can be used to express the chance constraints in special condition [4,21].
Lemma 1 [3,16]. Let X ∼ (μ, Σ) denote random vector X with mean μ and
covariance matrix Σ, the multivariate Chebyshev inequality states that for any
closed convex set S, the supremum of the probability that X take a value in S is
sup P{X ∈ S} = 1
1+d2
X∼(μ,Σ)
−1 (9)
d2 = inf (X − μ)T (X − μ).
X∈S
Theorem 1. Assume the first and second moment information of random vari-
ables A i are known. Let μ+ = E[A
i and B i ] and μ− = E[Bi ] be the mean
i i
1512 Q. Cao et al.
+ i − μ+ )] and − = E[(B
i − μ+ )T (A i −
vector seperately. And let i = E[(A i i i
− T −
μi ) (Bi − μi )] be the covariance matrix of the two data set uncertain points
respectively. Then the problems (7) and (8) could be reformulated respectively as:
1 T + T
l1
min 2 w+ G w+ + w+ μ b+ + 12 l1 b2+ + C1
T +
ti ξi
w+ ,b+ i=1 (10)
1
−2
s.t. −(μ−
i w+ + b+ ) ≥ 1 − ξi + k i w+ , ξi ≥ 0
and
1 T − T −T

l2
min 2 w− G w− + w− μ b− + 12 l2 b2− + C2 ti ηi
w− ,b− i=1 (11)
1
+2
s.t. μ+
i w− + b− ≥ 1 − ηi + k i w− , ηi ≥ 0,

1−ε
where k = ε and

l1
T
l1
G+ = (μ+ + +
i μi + Σi ), μ+ = μ+
i
i=1 i=1
with

l2
T
l2
G− = (μ− − −
i μi + Σi ), μ− = μ−
i .
i=1 i=1
Proof. Now we prove that problem (7) can be reformulated to Eq. (10). In fact,
it follows from Eq. (7) that
l1
1
min + + e+ b+ 22 } + C1
E{Aw ti ξi
w+ ,b+ 2 i=1
l1 l1 l1
1
i w+ b+ + l1 b2+ } + C1

= min E{ (Ai w+ )2 + 2 A ti ξi
w+ ,b+ 2 i=1 i=1 i=1
l1 l1 l1
1 T T
i }w+ b+ + 1 l1 b2+ + C1

= min w+ E{ Ai Ai }w+ + E{A ti ξi
w+ ,b+ 2 i=1 i=1
2 i=1
l1 l1 l1
1 T T
i }w+ b+ + 1 l1 b2+ + C1

= min w+ E{Ai Ai }w+ + E{A ti ξi
w+ ,b+ 2 i=1 i=1
2 i=1
l1 l1 l1
1 T T i } + Σi+ )w+ +

i }w+ b+ + 1 l1 b2+ + C1

= min w+ (E {Ai }E{A E{A ti ξi
w+ ,b+ 2 i=1 i=1
2 i=1
l1 l1 l1
1 T +T + + 1
= min w+ (μi μi + Σi+ )w+ + μi w+ b+ + l1 b2+ + C1 ti ξi
w+ ,b+ 2 i=1 i=1
2 i=1
l1
1 T + T T 1
= min w+ G w+ + w+ μ b+ + l1 b2+ + C1 ti ξi ,
w+ ,b+ 2 2 i=1
where

l1
T
l1
G+ = (μ+ + +
i μi + Σi ), μ+ = μ+
i .
i=1 i=1
Moreover, for the constraint of (7), we know that the set {−(Bw + + e+ b+ ) ≤
1 − ξ} is a half-space produced by a hyperplane and so it is a closed convex set.
Using Lemma 1, we obtain
sup i w+ + b+ ) ≤ 1 − ξi } =
P{−(B 1
1+d2
i ∼(μ− ,Σ − )
B i i
−1
d2 = inf (X − μ− T +
i ) Σi (X − μ−
i ).
−(Xw+ +b+ )≤1−ξi
Now we claim that −(μ+ i w+ + b+ ) > 1 − ξi . Suppose that −(μi w+ + b+ ) ≤

+
−
1 − ξi . Then by taking X = μi , we can get d = 0 and
2
sup i w+ + b+ ) ≤ 1 − ξi } = 1,
P{−(B
i ∼(μ− ,Σ − )
B i i
which is a contradiction to 0 < ε < 1. Let

−1 1
ui = Σi− vi = −Σi− w+ ,
2 2
(X − μi ), i w+ + b+ + 1 − ξi .
γi = μ+
Then γi < 0 and
d2 = inf uTi ui .
viT ui ≤γi
Consider the Lagrangian

L(ui , λi ) = uTi ui + λi (viT ui − γi ), λi ≥ 0.
By taking the derivative to be zero, we have 2ui = λi vi and viT ui = γi . Thus,
−1
d2 = inf (X − μ− T −
i ) Σi (X − μ−
i )
−(Xw+ +b+ )≤1−ξi
= uTi ui
γ2
= Ti
vi vi
(μ−
i w+ + b+ − 1 + ξi )
2
= T − .
w+ Σi w+
For the constraint (7), by
sup P{−(Xw+ + b+ ) ≤ 1 − ξi } < ε,
X∼(μ− −
i ,Σi )
2 ≤ ε and so d ≥ ε . Since −(μi w+ + b+ ) − 1 + ξi > 0,

1 2 1−ε +
it is easy to see that 1+d
one has
− 12
−(μ−i w+ + b+ ) ≥ 1 − ξi + k i w+ , ξi ≥ 0.
Similarly, we know that problem (8) can be expressed as (11). This completes
the proof.

1514 Q. Cao et al.
For any non-zero vector (w+ , b+ ), we have

1 T + T +T 1
w+ G w+ + w+ μ b+ + l1 b2+
2 2
1
+ 1 +T
G μ
= (w+ , b+ )T 21 + 2 1 (w+ , b+ )
2μ 2 l1
1 + + e+ b+ 2 }
= E{Aw 2
2
≥ 0.
On the other hand, it is easy to see that the constraint of problem (10) is a form
of second cone.
Let T
1 G+ μ+
H+ = . (12)
2 μ+ l1
Then the matrix H + is positive semi-define. To ensure the strict convexity of
problem (10), we can always append a perturbation I ( > 0, I is the iden-
tity matrix) such that the matrix H + + I is positive define. Without loss of
generality, suppose that H + is positive define.
Theorem 2. The dual problems of chance-constrained TWSVM models (10)
and (11) can be formulated as the following models

l1 T T T T
max λi − 12 s+
i H1 G H1 si − 2 l1 si H2 H2 si − μi H1 si H2 si
+ + + + 1 + + + + + + + + +
λi ,ν i=1

l1 T 1
l1 (13)
λi μ− + kΣi− ν ,
2
s.t. − i λi = s+
i
i=1 i=1
0 ≤ λi ≤ C1 ti , ν ≤ 1
and

l2 T T T T
max γi − 12 s− − − − − − − − − + + − + −
1
γi ,υ i=1

l2 T 1
l2 (14)
= s−
2
s.t. − γi μ+
i − kΣi+ υ , γi i
i=1 i=1
0 ≤ γi ≤ C2 ti , υ ≤ 1,
where −1 −1
H+ = [H1+ , H2+ ], H− = [H1− , H2− ].
Proof. We only need to prove that the dual problem of Eq. (10) can be formu-
lated as Eq. (13). The Lagrangian is given by
1 T + 1 l1
T +T
L(w+ , b+ , ξ, λ, β) = w+ G w+ + w+ μ b+ + l1 b2+ + C1 ti ξi (15)
2 2 i=1

l1 1
l1
λi −(μ− − 2
− i w+ +b+ )−1 + ξi −kΣi w+ − βi ξi ,
i=1 i=1
where λi , βi ≥ 0.
Recall that for any x ∈ Rn , we have the relationship x = max xT y. Then
y≤1
the equivalent model of Eq. (15) are given as follows:
l1
1 T + T +T 1
L1 (w+ , b+ , ξ, λ, β, ν) = w+ G w+ + w+ μ b+ + l1 b2+ + C1 ti ξi (16)
2 2 i=1
l1 1
l1
− λi −(μ−
i w+ + b+ ) − 1 + ξi − k(Σi
−2
w+ )T ν − βi ξi ,
i=1 i=1
where λi , βi ≥ 0, ν ≤ 1. The Lagrangian L1 has the same optimal value

as L when maximized with respect to ν’s subject to the constraint ν ≤ 1.
Therefore,
L(w+ , b+ , ξ, λ, β) = max L1 (w+ , b+ , ξ, λ, β, ν).
ν≤1
Similar to the discussion of Sect. 5 of [6], we know that solving Eq. (10) is equiv-
alent to finding the saddle-point of the Lagrangian L1 . This fact combining the
convexity implies
min max L(w+ , b+ , ξ, λ, β)
w+ ,b+ ,ξ λi ,βi
= min max L1 (w+ , b+ , ξ, λ, β, ν)

w+ ,b+ ,ξ λi ,βi ,ν≤1
= max min L1 (w+ , b+ , ξ, λ, β, ν). (17)

λi ,βi ,ν≤1 w+ ,b+ ,ξ
By eliminating the primal variables in Eq. (17), we can obtain the dual problem.
Taking partial derivatives of L1 with respect to w+ , b+ , and ξ, respectively,
one has
⎧
⎪
⎪ T l1 T l1 1
⎪ ∂L1 (w+ ,b+ ,ξ,λ,β)
λi μ− λi Σi− ν,
2
⎪ = G+ w+ + μ+ b+ + +k
⎪
⎨ ∂w+
i=1
i
i=1
∂L1 (w+ ,b+ ,ξ,λ,β) T +T
l1
(18)
⎪
⎪ = w+ μ + l 1 b+ + λi ,
⎪
⎪
∂b+
⎪
⎩ ∂L1 (w+ ,b+ ,ξ,λ,β) = C t − λ − β .
i=1
∂ξi 1 i i i
By equating the equations in Eq. (18) to zero, we can obtain following

formulations
⎧
⎪ l1
−T l1 1
⎪
⎪ G+
w + μ +T
b + λ μ + k λ Σ −2
ν = 0,
⎪
⎨
+ + i i i i
i=1 i=1
T +T

l 1 (19)
⎪
⎪ w+ μ + l 1 b+ + λi = 0,
⎪
⎪
⎩ i=1
C1 ti − λi − βi = 0.
Thus, by solving Eq. (19) and considering (12), we can infer that
l
1
T 1 l1
2
H + [w+ , b+ ] = − i λi = s+ i . (20)
i=1 i=1
1516 Q. Cao et al.
Since H + is positive define, it is easy to see the solution of Eq. (20) can be
obtained as
⎧
⎪ l1 T 1
l1
⎪
⎨ w+ = −H1
+
λi μ− i + kΣi− ν ,
2
λi = H1+ s+i ,
i=1 i=1
(21)
⎪
⎪ l1
−T −2
1
l1
⎩ b+ = −H2+ λi μi + kΣi ν , + +
λi = H2 si ,
i=1 i=1
+−1
where H = [H1+ , H2+ ].
According to Eqs. (16), (19) and (21), we can obtain the dual problem of Eq.
(10) as follows

l1 T T T T
min λi − 12 s+
+ + + + 1 + + + + + + + + +
λi ,u i=1

l1 T 1
l1
2
s.t. − i λi = s+
i
i=1 i=1
0 ≤ λi ≤ C1 ti , ν ≤ 1.
Similarly, we know that the dual problem of Eq. (11) can be expressed as
Eq. (14). This completes the proof.

In this section, our FCC-TWSVM model is illustrated by numerical test based on
two types of data sets. The first test is implemented to certify the performance
of our FCC-TWSVM by artificial data. And in second test, we also test the
performance of FCC-TWSVM model on real-word classifying data sets from
UCI Machine Learning Repository. All results were averaged on 10 train-test
experiments and carried out by Matlab R2012a with 2.5GHz CPU, 2.5G usable
RAM. The SeDuMi 1 software is employed to solve the SOCP problems of FCC-
TWSVM.
4.1 Fuzzy Membership Function

The design of fuzzy membership function is the key to the fuzzy algorithm using
fuzzy technology. In this paper, we use class center method to generate fuzzy
membership. Firstly, we denote the mean of class +1 as class-center x+ and the
mean of class −1 as class center x− , respectively. The radius of each class r+ and
r- are the farthest distance between the each class training points and its class-
center, respectively, namely r+ = max xi − x+ and r− = max xi − x− .
xi ,yi =1 xi ,yi =−1
Fuzzy membership ti is a function of the mean and radius of each class

1 − xri+−x+
+σ if yi = 1
ti =
1 − xri−−x−
+σ if yi = −1,
where σ > 0 is used to avoid the case ti = 0.

1
http://sedumi.ie.lehigh.edu/.
(a) ε = 0.1 (b) ε = 0.01
Fig. 1. The performance of FCC-TWSVM in the first data set.
4.2 Artificial Data
To give an intuitive performance of FCC-TWSVM, we construct uncertain sets

of 2-dimension data generated randomly from two normal distribution. The con-
tribution of binary classes of the first data is

+ 0 + 10 − −1 − 70
μ = , Σ = , μ = , Σ = .
2 04 0 03
Figures 1 shows that the performance of FCC-TWSVM to two uncertain

data set points. In numerical experiments, different data points are generated
by respective distribution. In each data set, +1 class is generated by normal
distribution (μ+ , Σ + ) and −1 class is generated by normal distribution (μ− , Σ − ).
Each class has 50 points, and 20 points are randomly picked as the training
points, the other points are the test points. In Fig. 1, the blue stars are the
points of +1 class, while −1 class with the red circles. The blue and red lines are
the separating hyperplane (2-D) that we look for. In fact, the value of parameter
ε also affects the determination of two hyperplanes. Figure 1(a) and (b) perform
the effect of various parameters.
4.3 Application
China’s VC [7] market is now in a substantial expansion of scale, while of the

decline in return. The industry system is between preliminary cooperation and
in-depth cooperation. The former is reflected in the VC firms, and the latter is
reflected in the investment groups.
We call such structure a community which is closely related to each other and
has far more links than other nodes in social network analysis method. From the
sociology level, a community is a social network with strong relationships. The
preference-dependent phenomenon in random network model has been widely
accepted, but the condition of new VC firm’s attachment is still unknown.
1518 Q. Cao et al.
The formation of communities is determined by the attributes of VC firms.

We select five basic attributes: capital scale, registration place, institution type,
number of investments, age. Capital scale stands for the strength of VC firms.
More capital controlled means stronger bargain power, which means the cooper-
ative VC firms could make the advantage position in the equity pricing. If capital
scale plays an important role in the decision tree, it means that big capital trends
to cooperate with big capital as well as small capital, where there is a polariza-
tion. Registration place can be deemed to be the place where the headquarters
or the institution with high administrative authority is, at least. It shows one
of the most active areas in social and economic activities. Institution type indi-
cates the main business type of VC firms. In addition to VC, PE, there are PIPE,
strategic investment, angel investment, broker direct investment, which are clas-
sified as other categories. The number of investments represents the active level
of VC firms. Although there is a certain correlation between capital scale, age,
and the number of investments, taking these three indicators into consideration
simultaneously could reduce the impact of relevance, to a certain extent.
In practice, the distribution properties are often unknown but need to be
estimated from data. For example, if an uncertain data point xi = [x i1 , · · · , x
in ]
T
N
has N samples xik , k = 1, · · · , N , then the sample mean xi = N1 i=1 xik is used
to estimate the mean vector μi = E[xi ], and the sample covariance
1
N
Si = (xi − xi )(xik − xi )T
N − 1 i=1 k
is used to estimate the covariance matrix
Σi = E[(xi − μi )(xi − μi )T ].
However, these could cause possible estimation errors. Some special cases
were proposed when the mean vector μi and covariance matrix Σi may not
exactly known. Panos M. Pardalos et al. [24] has discussed the way to process-
ing these special cases. In our practical experiments, similar to Pardalos, we
employ mentioned methods to modify the estimation and make the result easier
to interpret.
Since the data sets are uncertain, the performance measures are worth dis-
cussed. Ben-Tal et al. [2] proposed using nominal error and optimal error to
evaluate the performance. In our experiment, we choose these index to calculate
the accuracy of our model.
The expression for NomErr is

i 1yi =yi
pre
NomErr = × 100%.
the amount of training data
The optimal error is based on the probability of misclassification. The chance
constraints in the model (7), (8) can be reformulated to Eqs. (10) and (11), then
we can derive the least value of ε called εopt . Then the OptErr of data point xi is

1 if yipre = yi
OptErr =
εopt if yipre = yi .
And the OptErr of whole test set is

i OptErri
OptErr = × 100%.
the amount of training data
We tested the from Qingke data set. Now each test data point xi has 50
replicates xik . The class label yipre is decided by the majority label of the repli-
cates based on Eq. 4. For the OptErr, the sample mean and sample covariance
εopt . The real data sets have too many of the features to test the performance of
FCC-TWSVM model. The results over 10 times experiments are shown in Fig. 2
(a) NomErr (b) OptErr (c) Training time
Fig. 2. The performance of FCC-TWSVM in the VC data set
The average results for VC data set and Ionosphere set are shown in Fig. 2.
The boxplots of the results for two different sets show that the misclassification
rate decreases when ε reduces. In addition, the OptErr is always bigger than
NomErr. The experiment time is stable for different parameters ε. We can draw
a conclusion that regional disparity is the biggest factor in the difference of
community. There is a high information asymmetry between enterprises and VC
firms, and the latter usually give more attention to local projects.
5 Conclusions
A new chance constrained twin support vector machine (FCC-TWSVM) via
chance constrained programming formulation for classification was proposed,
which can deal data with measurement noise efficiently. This paper studied twin
support vector machine classification when data points are uncertain statistically.
With some properties known for the distribution, the FCC-TWSVM model was
used to ensure the small probability of misclassification for the uncertain data.
The FCC-TWSVM model could be transformed to second-order cone program-
ming (SOCP) by the properties of moment information of uncertain data and
1520 Q. Cao et al.
the dual problem of SOCP model was also introduced. Then we obtained the
twin hyperplanes by calculating the dual problem. In addition, we also showed
the performance of FCC-TWSVM model in artificial data and real data by
numerical experiments. In the future work, how to further make the model more
robust is under our consideration. In addition, dealing the situation of nonlinear
classification with chance constrained is also interesting.
Acknowledgements. This paper is supported by the Fundamental Research Funds

for the Central Universities of Sichuan University (skqy201646, skgt201602).
References
1. Ben-Tal A, Nemirovski A (2008) Selected topics in robust convex optimization.
Math Progam 112(1):125–158
2. Ben-Tal A, Bhadra S et al (2011) Chance constrained uncertain classification via
robust optimization. Math Program 127(1):145–173
3. Bertsimas D, Popescu I (2005) Optimal inequalities in probability theory: a convex
optimization approach. SIAM J Optim 15(3):780–804
4. Bhattacharyya C, Grate LR et al (2004) Robust sparse hyperplane classifiers: appli-
cation to uncertain molecular profiling data. J Comput Biol 11(6):1073–1089
5. Bi J, Zhang T (2004) Support vector classification with input data uncertainty.
Procof Neural Infprocsyst 17:161–168
6. Boyd S, Vandenberghe L (2004) Convex optimization. Cambridge University Press,
Cambridge
7. Cao Q, Lu Y et al (2013) The roles of bridging and bonding in social media
communities. J Assoc Inf Sci Technol 64(8):1671–1681
8. Carrasco M, López J, Maldonado S (2016) A second-order cone programming for-
mulation for nonparallel hyperplane support vector machine. Expert Syst Appl
54(C):95–104
9. Chang CC, Lin CJ (2007) Libsvm: a library for support vector machines. ACM
Trans Intell Syst Technol 2(3):389–396 Article 27
10. Cortes C, Vapnik V (1995) Support-vector networks. Mach Learn 20(3):273–297
11. Goldfarb D, Iyengar G (2003) Robust convex quadratically constrained programs.
Math Program 97(3):495–515
12. Jayadeva Khemchandani R, Chandra S (2007) Twin support vector machines for
pattern classification. IEEE Trans Pattern Anal Mach Intell 29(5):905–910
13. Lee YJ, Mangasarian OL (2001) SSVM: a smooth support vector machine for
classification. Comput Optim Appl 20(1):5–22
14. Lin CF, Wang SD (2002) Fuzzy support vector machines. IEEE Trans Neural Netw
13(2):464–471
15. Mangasarian OL, Wild EW (2006) Multisurface proximal support vector machine
classification via generalized eigenvalues. IEEE Trans Pattern Anal Mach Intell
28(1):69–74
16. Marshall AW, Olkin I (1960) Multivariate Chebyshev inequalities. Ann Math Stat
31(4):1001–1014
17. Nemirovski A, Shapiro A (2006) Convex approximations of chance constrained
programs. SIAM J Optim 17(4):969–996
18. Peng X (2010) A v-twin support vector machine (v-TWSVM) classifier and its
geometric algorithms. Inf Sci 20(180):3863–3875
19. Scholkopf B, Smola AJ (2003) Learning with kernels: support vector machines,
regularization, optimization, and beyond. MIT Press, Cambridge
20. Shao YH, Deng NY (2012) A coordinate descent margin based-twin support vector
machine for classification. Neural Netw Official J Int Neural Netw Soc 25(1):114–
121
21. Shivaswamy PK, Bhattacharyya C, Smola AJ (2006) Second order cone program-
ming approaches for handling missing and uncertain data. J Mach Learn Res
7(7):1283–1314
22. Trafalis T, Gilbert R (2007) Robust support vector machine for classification and
computatioanla issues. Optim Meth Soft 1(22):187–198
23. Trafalis TB, Gilbert RC (2006) Robust classification and regression using support
vector machines. Eur J Oper Res 173(3):893–909
24. Wang X, Fan N, Pardalos PM (2015) Robust chance-constrained support vector
machines with second-order moment information. Ann Oper Res 253:1–24
25. Xanthopoulos P, Pardalos PM et al (2013) Robust data mining. Springer briefs in
optimization
26. Yang B, Wang MH et al (2016) Ramp loss quadratic support vector machine for
classification. Nonlinear Anal Forum 1(21):101–115
A Projection Pursuit Combined Method
for PPP Risk Evaluation
Xinli Zhang(B) , Tianjin Wang, and Sihan Li

zhangxinli1231@163.com
Abstract. Risk assessment is very crucial to the building and operation

of PPP (Public-Private-Partnership) project in its long-term cooperation
during the long life circle. Based on the risk factors and their influences,
the traditional risk assessment method of PPP project is achieved by the
experts experience, Therefore, the evaluation result is subjective. Under
the background of Big Data, the study proposed a projection pursuit
combined method to model the PPP risk, which was called projection
pursuit risk assessing model. Firstly, we use Monte Carlo simulation to
get the data of the risks factors and their influence. Secondly, estab-
lish a multivariate regression function with projection pursuit method
to research the relationship between the multivariable risk factors and
the increment of internal return rate of PPP project. Thirdly, we offer
the project risk level through the analysis of the regression results. The
application in the specific project of this new method shows that the new
model can accurately estimate the risk level of PPP project, and explain
the trend of project risk.
Keywords: PPP (Public-Private-Partnership) · Risk assessment · Pro-

jection pursuit regression · Monte Carlo simulation
1 Introduction
In recent years, PPP (Private-Public-Partnership) pattern has been introduced
to China as a new mode, aiming at attracting foreign capital to participate in
infrastructure construction projects to make up for the financial lack of public
sectors as well as improving the operational efficiency. Thus the mode is widely
regarded as a new infrastructure project financing tool.
However, due to the long cooperation process, the wide investment scale,
and the large number of participants, PPP projects have more potential and
complex risk factors during the whole life cycle compared with average engineer-
ing project. Especially, risk assessment is a crucial link in the risk management
process. However, currently few domestic cases have involved the implementa-
tion of the PPP project, thus the PPP risk assessment mostly depends on the
expert’s subjective judgment, such as the fuzzy mathematical evaluation method
and decision tree method often used in related research and field [3].
Combined Method for PPP Risk Evaluation 1523
In recent research, s the methods for PPP assessment are mainly quantitative
assessment methods. For example, fuzzy-AHP-based risk assessment and fuzzy
synthetic evaluation approach to PPP projects were established to assess risk
factors in the PPP expressway project in China [8]. Another research identify
the risk factors in PPP projects through a comprehensive literature review, and
then introduce fuzzy logic in the pairwise comparison [12]. A few studies deal
with uncertainty by presenting several estimates based on different values for the
exogenous inputs using sensitivity analysis. Sensitivity analysis is a typical way
to address uncertainty via variations in key inputs. It hence provides an insight
into what happens if some variables values differ from the basic case [7].
As research shows, system dynamics (SD) modelling aims at developing a
dynamic model to assess demand risk by evaluating how different variables
jointly affect demand for services provided by PPP infrastructure projects. The
research objectives thus involve identifying, understanding, mapping and mea-
suring these variables and their complex interrelations [1].
FSE is a branch of fuzzy set theory (FST), a number of researchers have
attempted to exploit FSE in the construction project risk management field.
Yeung et al. present a model to assess the level of risk of PPP toll road projects
in which experience knowledge based on linguistic variables were incorporated
into the analysis using FSE [2].
In summary, the current PPP project risk assessment has a series of complex
problems, for example, the original project data involves various risk factors
which share no obvious function relationship between these factors. It is difficult
to precisely describe the PPP risks with accurate mathematical model since most
risk assessment methods are limited to subjective factors. The risk assessment
is a high-dimensional problem, which the current assessing method is unable to
solve. Thus the good risk management needs to establish a risk estimation model
which can convert high-dimensional problem into low-dimensional problem.
Considering the deficiencies of existing evaluation methods, we present the
projection pursuit regression method (PPR) based on the data analysis as
an exploring method of risk assessment. The method can convert the high-
dimensional risk variables into low-dimensional projection variables, meanwhile
building up the relationship between the projection variables and the economic
evaluation index variables [13].
The realization of the new method is further elaborated in four parts. Section 2
introduces a PPP risk assessment method with PPR; Sect. 3 offers the building
and parameter optimization process of the new PPP risk assessing model, followed
by a case study applying the model into assessing the risks of an ecological land-
scape PPP project in Sect. 4. We present our conclusions in Sect. 5.
2 Projection Pursuit Risk Assessing Method

2.1 Traditional PPP Risk Assessment Procedure
Among the existing PPP risks assessing methods, the fuzzy synthetic evaluation
(FSE) method is the most prevalent, which is capable of effectively dealing with
1524 X. Zhang et al.
the imprecise non-numerical (linguistic) terms and accounting for the fuzziness
in expert knowledge that typifies risk assessment [11]. As Fig. 1 shows, the hier-
archical analysis approach utilized for assessing the risk level of principal risk
factors.
b1
P b2
Construction risk Operating risk

R1 R2
c1 c2 c3
Policy risk Financial risk Technology risk

r1 r2 r3
w1 w2 w3 w4 w5 w6 w7 w8 wn
x1 x2 x3 x4 x5 x6 x7 x8 xn
Fig. 1. FSE assessing procedure
where xi represents the ith risk factor of the risk factor layer (i = 1, 2 · · · , n),
ri represents the sub factor layer, r1 , r2 , r3 respectively represents the second-
class risk factor layer, R1 ,R2 respectively represents the first-class risk fac-
tor layer. P represents the comprehensive risk membership of the project. wi
(i = 1, 2, · · · , n) respectively represents the ith risk factor’s weight in the second-
class risk factor layer. c1 ,c2 ,c3 respectively represents the second-class risk fac-
tor’s weight in the first-class factor layer. b1 , b2 respectively represents the first-
class factor’s weight in the comprehensive risk membership of the project. The
relationships among these parameters can be presented as
X = {x1 , x2 , x3 , · · · , xn } , (1)
R = {r1 , r2 , r3 } , (2)
R = {R1 , R2 } , (3)
C = {c1 , c2 , c3 } , (4)
B = {b1 , b2 } , (5)
W = {w1 , w2 , · · · , wn } . (6)
And then,
R = W × X, (7)

R = R × C, (8)
P = R × B. (9)
After several transformations of matrices, we get the value of the evaluation
index. Though the FSE method has been widely used in the risk assessing field,
it still has its limitations:
(1) The mechanism of how a risk factors exerts impact on the project objectives
is unclear, so the actual process might be fuzzier.
(2) The interaction of various risk factors is ignored.

(3) Difficulty in accurately determining the membership.
Thus we need to discover a new method to optimize the traditional PPP
assessing methodology.
2.2 Projection Pursuit Regression

Based on the analysis above, we need to find a method which can not only
simplify the PPP risk assessment process compared with the traditional risk
assessment process, but also improve the accuracy of risk assessment, to fulfill
the purpose of redressing the defects of subjective evaluation. Thus the projection
pursuit regression is introduced.
The projection pursuit regression is based on a regression method derived
from projection pursuit thought [10]. Compared with the traditional methods
focusing on exploring the relationship between the independent and depen-
dent variables, this method projects multidimensional independent variables and
determines the projection variables after obtaining a low dimensional projection
of variables. The projection can reduce the number of variables which are used
to establish the regression model, meanwhile transform the regression relation-
ship between multiple variables and single variables into one solely between the
single variables.
As for the PPP project risk assessment, it can be considered as a multi-input
and single-output variable prediction problem. The ideology is, under different
risk situations, we have data of set of risk values of 1 groups Xk (k = 1, 2, · · · , l)
and these corresponding economic evaluation results(Yk ), then we can set up
the relation between Xk and Yk . Since Xk is a high-dimensional set of variables,
we need to project these variables to low-dimensional space and find proper
projecting directions zj (j = 1, 2, · · · , m). After each projection, we’ll get its
corresponding residual value Δyj (Δyj = yj − yj where yi is the actual value
yj is the fitting value) until the residual sum of squares meet certain accuracy
requirement, then the parameter j(j = 1, 2, · · · , m) is optimal number of pro-
jection direction. And by minimizing a projection, finding out the projection
can reflect the structural characteristics of high-dimensional data. The projec-
tion values of each direction are fitting by one-dimensional function. Then the
regression function is approximated by one-dimensional fitting function [4]. The
procedure of the projection pursuit method of risk assessment is shown in Fig. 2.
The non-parametric regression f (x) can be summed up as:
m
n
T
Y = g(j) ( αji X), (10)
j=1 i=1
n

2
αji = 1, (11)
i=1
where, X is n-dimensional independent variable xi , Y is the response variable,

gj represents the j th smooth ridge function. m is the number of approximation
function, aji is the ith component projection direction of the j th smooth ridge
function. The obtaining of big data needed in the function and algorithm of the
calculation of PPR will be presented in Sect. 3.
Fig. 2. The procedure of the projection pursuit method of risk assessment
3 PPP Risk Assessing Model Process and Its Parameter

Optimization
3.1 The PPP Risk Assessing Model
Based on the analysis of the previous chapter, we are able to establish the
relationship between the independent variable (risk variables) and the depen-
dent variable (economic evaluation index) by the projection pursuit regression
method. For every PPP project, multiple groups of risk values are not read-
ily available to obtain. Instead they call for establishing and calculating after
a large amount of data conversion and multiple model procedures. After that,
the projection pursuit regression models can be summed up by Fig. 3 whose
left column refers to the methods used to obtain the data, the middle column
representing the specific form of the relevant data in the model, and the right
column indicating the main mathematical model in the processes of data input
and output.
In Fig. 3, Iki (i = 1, 2 · · · , n, k = 1, 2 · · · , s) represent the ith risk’s impact
value according to s groups of historical data. σ(Ii ) , μ(Ii ) refer to the stan-
dard deviation and the mean value of the calculated values of the Iki . Di rep-
resents the distribution of the ith risk, Iki , EEIk respectively represents the
k th (k = 1, 2 · · · , l). Group of data of the project risk value and project eco-
nomic evaluation index. αj (j = 1, 2, · · · , m) represents the parameter in the j th
projection direction, REk (k = 1, 2 · · · , l) refers to residual error. We can divide
the model construction procedure into four specific steps as follows:
Scene Analysis
Modeling
calculation
External environmental Data flow
process
and historical data
information processing system
Input variable:
Monte Carlo Simulation Output variable:

start
Y-axis
precision of
the test
X-axis
Let
generate the mathematic

random numbers models
Bring into
statistical
analysis
Probability distribution of economic

indicators
Back model
Projection Pursuit Regression
Fig. 3. The PPR risk assessing model for PPP project
Step 1. Determine the independent variables of the statistics, and fit distributions

From historical data, we can obtain numerous data of different risk’s
impact value Iki of different tasks. After calculation, we can get the
average value μi and standard deviation σi of the ith risk’s statistic
value. Meanwhile, combining with experts’ judgement, we can get the
distributions of various risks.
s i i
j=1 pk Ik
μi = , (12)
s
s

σi =
2
(pik − μi ) × Iki . (13)
k=1
Step 2. Extract random number through Monte Carlo simulation and get the
sample observation value.
We establish the project’s economic evaluation model, and set the known
variable as constant parameters variable of fixed value model. We set
Ii as input random variable, which is followed by selecting economic
evaluation index (EEI) according to the demand of project economic
evaluation and setting it as an output variable. In order to obtain more
groups of sample data, we extract random number through Monte Carlo

Simulation and transform the probability distribution of random num-
ber as sampling values of input variables while repeating the experiment
to meet a certain level of accuracy. At this time, we set the simula-
tion iteration number is set as l, then we get l groups of input vari-
ables Iki and output variables EEIk of the sample data. Now we let
xik = Iki , yk = EEIk , and establish a functional relationship between
yk and xik .
Step 3. Take the sample data to the projection pursuit regression model
m
n

T
yk = g(j) αji Xk , (14)
j=1 i=1
X k = x1k , x2k , · · · , xik (k = 1, 2, · · · , l) (15)
We send l groups of data to the PPR model, where Xk represents the
k th set of the input variable xik (i = 1, 2, · · · , n), yk is the correspond-
ing output variable. Other parameters mentioned in the Eq. (15) can
refer to Sect. 2. Meanwhile, There are many forms to fit a variety of
cell functions, such as the numerical function [13] and the polynomial
function [11], a number of numerical functions, etc.
Step 4. Output the results
After the algorithm realization process, we get αj and REk of the
regression model. Generally, the smaller the REk , the better the fit-
ting effect. Then according to the distribution of EEI, we can arrive at
some conclusions about the project’s risk rank.
3.2 The Algorithm Realization Process
Since the numerical function involves an enormous numerical tables, it is not

convenient to predict during the interpolation and extension [5]. Thus the paper
uses multinomial to fit the unit function.
Take l groups of risk variables as input variables, and select economic evalu-
ation index as output variables brought into Eq. (14) to determine the optimal
output variables. Then optimize the number of unit functions and its corre-
sponding projection direction parameters. In the model, the genetic algorithm
is adopted to optimize the parameters as follows:
(1) Generate m projection directions randomly in the range of [−1, 1], which is
followed by selection, crossover and mutation of genetic algorithm to generate
3 × n new projection directions [14].
(2) Respectively project m construction parameters of l sample projection on
the projection direction, obtaining l projection variables for each projection
direction.
(3) The polynomial is used to fit the corresponding relation between projection

variables and the internal rate of return, resulting in 3 × n fitting polyno-
mials, then we can evaluate each fitting results according to the correlation
coefficient d calculated through Eq. (16).
2
(y − y)
d=1− 2 (0 ≤ d ≤ 1). (16)
(y − y)
In the Eq. (16), represents the actual value of the economical evaluation
index. y is the corresponding fitting value, and y stands for the average
value of actual internal rate of return.
(4) In the 3 × n projection direction, the coefficient d is calculated, and n pro-
jection directions with superior deterministic coefficient are chosen to enter
the optimization process in the next round of genetic algorithm.
(5) After several times of optimization, when the deterministic coefficient’s
absolute value of difference between two calculations from start to finish
come out less than some arbitrary positive number, the optimization of first
unit function finishes.
(6) Output polynomial and its corresponding projection direction m, which lead
to calculate the relative error between the fitting value and the actual value.
According to the risk assessment, if the relative error is less than 10%, the
model parameters are determined [6]. Otherwise, it is indicated that an arbi-
trary unit function can not meet the fitting requirements, and we should
move to the next step.
(7) In the this step, we follow the same principle, which infers that the residual
Δy(y − y ) completed in the fitting of the first step is to replace y to repeat
the first step’s work. We start the optimization aimed at second element
function, until we have obtained fitting unit function and projection direction
requirements which meet the requirements [9].
4 Case Study
4.1 Data Origin
The PPP risk assessment model based on projection pursuit regression method
was validated through a case study, an ecological landscape PPP project case
from Sichuan province. In this case, we identified 10 risk factors basing on the
historical data, which respectively refers to cost overruns risk, construction dura-
tion extended risk, design risk, construction safety risk, operation risk, quality
risk, policy changing risk, force majeure risk, tax increasing risk, and laws risk.
In the PPP risk assessing model, we select (the impact value of the ith risk) as
independent variable, meanwhile ΔIRR as dependent variable, which can rep-
resent the increment of the internal rate of return, thus reflecting the economic
viability of the project.
4.2 Input and Output Variables

We define xi as input variable, y as output variable. The distributions of each
risk’s impact value above can be seen in the Figs. 4 and 5, which are determined
by numerous historical data Then we send the input variables and output vari-
ables into the Monte Carlo simulation model, resulting in 461 groups of sample
data. The input and output variables’ distributions are respectively shown in
Figs. 4, 5 and 6.
4.3 Forcasting Results

With the 461 groups of sample data simulated above (436 groups are used for
fitting, and 30 groups are used for testing), we can build up PPR. After expe-
riencing algorithm realization process, the model’s parameters are achieved. We
can get two figures reflecting the fitting results. Figures 7 and 8 respectively
shows the comparing results between the actual values and simulation values,
the comparing results between the actual values and tested values, Figs. 9 and
10 respectively shows the percentage of residual error of the simulation variables
and tested variables.
Fig. 4. The distribution of the first five risks
Fig. 5. The distribution of the post five risks

70
Δ IRR
60
50
40
30
20
10
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Fig. 6. The distribution of ΔIRR
fitting value real value

0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
1
39
77
115
153
191
229
267
305
343
381
419
-0.05
Fig. 7. The comparing results between the actual values and fitting values
fitting value real value

0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29
Fig. 8. The comparing results between the actual values and tested values
Fig. 9. The percentage of residual error of the fitting variables
Fig. 10. The percentage of residual error of the tested variables
According to the figures above, we can calculate some performance indexes

in Tables 1 and 2.
Table 1. The fitting performance of PPR
The percentage of relative error The sample size The percentage of all samples
More than 20% 25 6%
Table 2. The tested performance of PPR
The percentage of relative error The sample size The percentage of all samples
More than 20% 6 20%
The results above show that PPR can get relatively accurate fitting per-
formance for 88% sample. What’s more, the majority of the relative errors are
less than 10%, which indicates the variation range of the IRR is small. Namely,
the anti-risk of this PPP project is good. Meanwhile, considering the risks, the
investors can compare the IRR with the IRR of the benchmark yield to decide
whether to invest the project.
5 Conclusion
This study introduces the projection pursuit regression method as a new risk
assessment method, establishing a new risk assessment model related to the tra-
ditional risk assessment methods: Firstly we fit out the probability distribution
of risk value based on historical data and expert’s subjective opinions. Secondly
we take the random numbers through the Monte Carlo simulation as input vari-
ables, and other economic indicators as output variables to obtain the sample
data. Thirdly, we set up the interrelation between the two variables based on
the data above. And then we establish a risk assessment model by projection
pursuit regression method combined with the actual case, which is used to deal
with the sample data regression to achieve a reasonable calculation effect. The
results above shows that the model can reveal the change regulation of risk
factors affecting PPP project economic evaluation index. The new assessment
model can catch the risk features of PPP project influenced by many factors, to
give an integrated results through dimension reduction.
The proposal of the evaluation model establishes direct interrelations between
the risk factors and economic evaluation of PPP projects, reducing the intrica-
cies of traditional risk assessment and making the risk assessment process more
effective and rapid. At the same time, the evaluation model is also more accu-
rate compared to its predecessors. However, some shortcomings also exist in this
study: (1) the existing data is not sufficient enough to completely fit the proba-
bility distributions of all risks, which implies that the existing assessment of the
risk still partly relies on expert assessment; (2) the evaluation model in normal
circumstances is relatively stable, but part of the deviation still persists. Thus
the accuracy of the algorithm calls to be improved in the optimization model in
the future to improve the accuracy of risk assessment.
Acknowledgements. This paper is founded by Sichuan University (skqy201652).
References
1. Alasad R, Motawa I (2015) Dynamic demand risk assessment for toll road projects.
Constr Manag Econ 33:1–19
2. Ameyaw EE, Chan APC (2015) Evaluation and ranking of risk factors in public-
private partnership water supply projects in developing countries using fuzzy syn-
thetic evaluation approach. Expert Syst Appl 42(12):5102–5116
3. Chan APC, Yeung JFY et al (2011) Empirical study of risk assessment and alloca-
tion of public-private partnership projects in China. J Manag Eng 27(3):136–148
4. Du H, Wang J et al (2008) Prediction of retention times of peptides in rplc by using
radial basis function neural networks and projection pursuit regression. Chemometr
Intell Lab Syst 92(1):92–99
5. Durocher M, Chebana F, Ouarda TBMJ (2016) Delineation of homogenous regions
using hydrological variables predicted by projection pursuit regression. Hydrol
Earth Syst Sci Discuss 20(12):4717–4729
6. Friedman JH, Stuetzle W (1981) Projection pursuit regression. J Am Stat Assoc
76(376):817–823
7. Kriger D, Shiu S, Naylor S (2006) Estimating toll road demand and revenue.
NCHRP Synthesis of Highway Practice
8. Li J, Zou PXW (2011) Fuzzy AHP-based risk assessment methodology for PPP
projects. J Constr Eng Manag 137(12):1205–1209
9. Mas A, Ruymgaart F (2015) High-dimensional principal projections. Complex Anal
Oper Theor 9(1):35–63
10. Samarov AM (1993) Exploring regression structure using nonparametric functional
estimation. J Am Stat Assoc 88(423):836–847
11. Taheriyoun M, Karamouz M, Baghvand A (2010) Development of an entropy-based
fuzzy eutrophication index for reservoir water quality evaluation. Iran J Environ
Health Sci Eng 7(1):1–14
12. Valipour A, Yahaya N et al (2015) A fuzzy analytic network process method for
risk prioritization in freeway PPP projects: an Iranian case study. J Civ Eng Manag
21(7):933–947
13. Yamout G, Jamali D (2007) A critical assessment of a proposed public private
partnership (PPP) for the management of water services in lebanon. Water Resour
Manag 21(3):611–634
14. Zhang X, Wang S, Ding J (2002) Application of projection pursuit regression in
estimating building’s material consumption. J Sichuan Univ 29:17–23
Optimal Ownership Pattern to Control Agency
Conflict in Manufacturing Industry of Pakistan
Muhammad Kaleem Khan1 , He Ying1 , Umair Akram1(B) ,

Muhammad Hashim2 , Xiaoyue Yuan1 , and Lv Gaoyu1
1
leo umairinfo@yahoo.com
2
Faculty of Management Sciences, National Textile University,
Faisalabad 37610, Pakistan
Abstract. Asymmetric information always creates conflict of interest

between stakeholders. Purpose of this study is to discuss agency con-
flict from perspective of financiers, and suggesting its remedies as con-
trolling mechanism. Monitoring role of different types of ownerships is
studied. Study has investigated the impact of four different types of own-
ership structures namely; managerial ownership, family ownership, insti-
tutional ownership and ownership of block-holders on agency issue. Role
of financial institutions is also discussed. To explore the agency problem
from view point of both equity-holders and debt-holders, Agency cost
is divided into agency cost of equity (ACE) and agency cost of debt
(ACD). Study has also given empirical evidences to check the robustness
of each monitoring mechanism. Using financial data of top capitalized
100 companies of Pakistan from 2010 to 2015; results are drawn by fix
and random effect (GLS) model. Ownership of insiders, block-holders
and institutions were found to be most effective monitoring devices for
firms’ growth. Study also contains recommendations for investors and
debtors.
Keywords: Monitoring mechanism · Agency cost · Ownership struc-

tures · Manufacturing industry
1 Introduction
After the research stream on agency cost developed by Jensen [22] extensive
empirical results can be found on evidence claiming that presence of agency
conflicts has an adverse effect on business sustainability. Two types of agency
cost are found: Agency cost of equity (ACE) and agency cost of debt (ACD).
If alignment of interest of owners and managers is missing, this is hazardous
1536 M.K. Khan et al.
for the company by causing agency cost of equity. If debt-holders’ interest is

overlooked (due to conflict of interest between managers/shareholders and debt-
holder), then agency problem of debt is created. To monitor the agent’s action,
incurs cost that is called the agency cost.
ACE arises when principal hires agent and delegates his authority. Problem
originates when agent could not meet up with principal’s expectations and don’t
utilize firm’s resources efficiently to maximize shareholders’ interest. ACD is
created when more risk is transferred towards debt-holder. If capital structure
has higher proportion of debt, managers have strong incentives to invest on
the assertion of higher return in the extremely risky projects. If the project is
successful, managers and shareholders will get benefit; if not, cost will be bear
by debt-holders. Managers may expropriate wealth from debt-holders in order to
increase shareholders wealth or increase their own luxuries. Similarly managers
can expropriate wealth from debt-holders for shareholders by following actions:
getting additional debt in the case where new issued debt has equal or higher
claim than the previous debt (existing debt-holders hurt), Engaging in asset
substitution (switch from relatively low risky assets to highly risky assets) and
by underinvestment (managers may reject profitable projects if debt-holders get
benefit from such projects) and by overinvestment (when success only benefits
to shareholders and failure leads to financial distress).
Present study is focusing how both types of agency cost can be reduced by
decisions regarding ownership structures. Present study is addressing four differ-
ent ownership structures that are most common in Pakistan: family ownership,
managerial ownership, institutional ownership and ownership of block-holders.
Convergence of interest hypothesis (presented by Jensen) is also going to be
checked in Pakistan through the ownership of managers and family (insiders).
We will be also in a position to verify whether block-holders and institutions
employ their greater monitoring ability in Pakistan.
Financial institutions may alleviate not only managers’ moral hazards but
also agency problems between managers and creditors through a grading function
ex ante and a monitoring role for their firms ex post [28]. Banks grant loan only
if they are satisfied with management performance and firm’s financial strength.
Concept of institutional debt financing has been driven from theory of agency
cost [21]. Thus institutional debt financing is indirectly involved in reducing the
agency costs due to efficient monitoring of firms’ decisions. Banks monitor the
activities of the firms to fulfil loans extension criteria. Monitoring forces the firm
to utilize assets efficiently to repay loans. Present study is focusing to check the
monitoring capability of financial institutions in case of Pakistan.
Literature does not have much more about agency problem in Pakistan,
making our study a quite new topic for literature. Moreover, agency issues are
particularly severe issues in emerging economies [20]. Secondly ownership struc-
ture is still a new topic for researchers and there is enough room to work on this
topic. So combination of these two is expected to be very valuable for Pakistan
and other developing countries. We hope that our working will prove a valuable
initiative for researchers of our region. Our findings and recommendations will
Optimal Ownership Pattern to Control Agency Conflict 1537
help Pakistani manufacturing sector to enhance its performance by identifying

and reducing both types of agency problem.
2 Overview of Previous Studies

Literature does not answer that which type of ownership or set of different
ownerships can be optimal ownership structure. It is because researchers have
not included all equity ownership styles in their study. This section compiles
prominent researches regarding agency cost with different ownership structure.
However, some researchers have also concluded that agency cost is irrelevant of
any type of ownership structure [11,29].
Convergence of interest hypothesis advocates that when ownership and man-
agement is in the same hands, then owner-agent interests are converged [21].
Agency conflict arises due to difference of interest between owners and agents
and inability of the owners to monitor agents’ action. If managers become part of
the ownership then this might reduce agency cost [6]. These studies also report
that there is better asset utilization with managerial ownership in the firms
[32,44]. Proportion of insider ownership was found very low in firms with high
agency cost [2,31,46]. Leverage, dividends, managerial ownership, the percent-
age of outside directors and executive compensation are effective in controlling
the agency cost of insider trading [5]. But there is also reporting of contradic-
tory results of convergence of interest hypothesis [9,50]. Managerial ownership
cannot reduce the agency costs of ownership within family owned firms. Private
ownership sets the firm free from external governance and capital market moni-
toring [42]. But majority of researchers are convincing that managerial ownership
acts as an effective device in controlling agency cost. For example see [1,2,4–
6,13,19,24,30–32,36,38,39,44,49].
Institutional ownership is considered to be an effective monitoring mecha-
nism due to many reasons. Qualified analysts work for institutions to analyse a
firm and its financial performance [13]. Institutions also have blocks of shares
that strengthen its power and can control managers efficiently to work for the
attainment of financial goals [48]. Institutional investors are found successful in
reducing managerial overspending in the case of firms with low growth opportu-
nities [49], family and institutional ownerships are found to be complementary
devices in monitoring the CEOs [46]. There are also findings of some studies,
which don’t confirm the effective monitoring role of institutions to reduce agency
problem [24,36]. As previous studies found significant evidence that institutional
ownership reduce agency cost [1,4,5,13,41,46,49], our research also theorizes
that there is inverse relationship between institutional ownership and agency
cost. Study finds this argument more convincing that institutions have greater
monitoring abilities and strong influential power on the managers to direct them
to do actions that increase owners’ wealth. Although some studies found no influ-
ential monitoring power of institutions in reducing ACD [24,36] but this result
might be due to specific environment of their countries.
Almost all the previous studies are advocating the influential monitoring
power of family ownership to overcome agency conflict [6,22]. Family ownership
is more committed towards stability and continued existence of the firm due
to their reputations [10]. Family firms perform better than nonfamily firms [4].
Due to higher proportion of insider ownership, founding family firms have lowest
agency cost [41]. Founding family ownership is also associated with a significantly
lower cost of debt financing [4]. Family firms control agency problem by delivering
more amount of dividends to shareholders and by employing higher amount of
debt in capital structure [43]. But in family owned firms, managers act for the
controlling family, but not for shareholders in general [37]. Following [3,4,41],
study also hypothesizes the inverse relationship between agency cost and family
ownership.
Managers of a firm, whose ownership is held by blocks of shares, are less
involved in utilization of firm resources for their discretionary activities [9,14,17].
More the outsider block holding of the firm, more is the monitoring of the man-
agerial activities [44]. Recent research has proposed a signaling theory related
with block-holding ownership, according to which an entrepreneur chooses to
attract an outside blockholder in order to signal his low “propensity to expro-
priate [45]. Large board and small non-management block-holder ownership
face severe agency problems and poor corporate governance [30]. But there is
also evidence that high level of block holding is responsible for higher level of
agency cost in firms [41]. If ownership is intense, it can also create agency con-
flict [16]. Some researchers found that block-holders are not expert enough in
monitoring [24].
Bank debt is an effective device to control the agency problem. Bank credit
shows the credit worthiness of borrower, reducing the information asymmetry,
which affects ACD [19]. Agency cost decreases with the increase in the monitor-
ing by the banks [6]. Bank debt reduces cash holdings [17]. Institutions monitor
and evaluate the company regularly for extending credit [22].
3 Empirical Implementation
In this section, we present model used to examine the impact of ownership

structures on agency cost of equity and agency cost of debt. We also describe
the data used and firm level variables used in this study.
ACEit = ∂i + β1 M Oit + β2 F Oit + β3 IOit + β4 BOit + β5 Dit (1)

+ β6 DIVit + β7 Sizeit + β9 Qit + εit ,
where ACEit is Agency cost of equity measured in terms of assets turnover

ratio: M Oit is managerial ownership: F Oit is family ownership, IOit is institu-
tional ownership, BOit is ownership of block-holders: Dit is debt ratio: DIVit
is dividend payout ratio: Sizeit is size of firm: Qit represents growth opportu-
nities for the firms. Asset utilization ratio determines management aptitude to
grasp growth opportunities by efficiently employing firm resources. If this ratio
is significantly good as compared to industry average, it means that high level
of sales and cash flows are generated from firm assets. Whereas a low ratio of
turnover represents that the managers are using assets for unproductive pur-
pose rather than utilizing them in the activities that generate cash flows [27].
However firm with low assets turnover ratio bears more agency conflicts relative
to the firm with high turnover ratio. See also [6,19,36,44]. In addition, we also
assess the effect of different types of ownership structures on agency cost of debt.
This notion is examined substituting the agency cost of equity variable ACE by
Agency cost of debt ACD variable. The new model is described in Eq. (2). Debt
agency cost is defined as conflict between shareholders and bondholders [13].
Manso took the ACD as the difference between the total value of the all-equity
and levered firms [34]. We obtained the proxy of agency cost of debt (ACD)
from previous studies [13,26,40]. If major proportion of capital is invested in
fixed assets, then managers have no more liquid resources to use it on their own
luxuries. If firms has more liquid assets, managers may found it very easy to use
these assets for their own requisites or transfer cash to shareholders in form of
dividends, rather than keeping it for future payments of interest payments or
principal amount to debt holders [25].
ACDit = ∂i + β1 M Oit + β2 F Oit + β3 IOit + β4 BOit + β5 Dit (2)

+ β6 DIVit + β7 Sizeit + β9 Qit + εit .
Inclusion of variables of ownerships of management, family, institutions and

block-holders in the model is to assess the optimal ownership pattern. Addi-
tion of debt variable in the model is to test Jensen’s proposition of “debt as
the remedy” of agency problem [22]. In case of agency cost of equity, when firm
employs more debt, then it improves the efficiency of the firm to utilize assets
to produce more sales and minimize its cost [18]. It is because mangers have
responsibility to pay out installment and premium regularly. If they don’t do so,
then creditors may take firm to court to file bankruptcy. Mangers may lose their
jobs in such scenarios. So debt disciplines the mangers. In case of ACD, managers
are constrained to pay fix installments against debt. This constraint directs the
managers to act in such a way that they don’t waste firm resources. Debt serves
as a monitoring device to mitigate agency problem between owner and principal
[38,44]. Firms with more debt tend to have lower agency costs [9,12,36]. Family
firms take up higher debt levels to resolve agency problem [43]. Firm’s capital
structure affects agency cost [47]. Some studies did not find relationship of debt
with agency problem [23]. The amount of free cash flow was found increased
with increased level of debt [17,35,49]. Addition of Dividend payout ratio in the
model is to test whether Large incremental disbursements mitigate the agency
problem associated with excess funds or not [33]. Money that is not given back
to outside shareholders in the form of dividend that money can be invested in
negative NPV projects or can be used for the perks of insiders [7,15]. But Byrd
[9] found that Dividend is not influential in defining agency cost. ACD is more
influential on dividend policy as compared to that of agency cost of equity [8].
Size variable is added in model to test the notion: “Greater the size of firm,
more agency cost is expected” [17,44]. Larger firms bears higher agency costs
because of greater complexity in their operations [13]. That’s why present study
also hypothesize that there is positive relationship between firm size and agency
cost. Model also has variable of growth opportunities because literature directs
that if there are growth opportunities for the firms, then more cash generation
in the future is expected. When firms expect incremental growth opportunities,
they hold extra cash in hands to invest in future projects. It leads towards the
extra cash available to the manager to be used for wasteful expenditures. Thus
chances of agency problem increases with increased level of growth opportuni-
ties [17,19]. Conflicts of interest between shareholders and managers vary with
growth opportunities and free cash flow [49] (Fig. 1).
Out of 819 listed companies in Karachi Stock Exchange, which is largest
stock exchange of Pakistan, we selected 100 top capitalized firms. Our sample
contains only non financial firms. So we ignored 353 firms related with finan-
cial services in Pakistan. By this, study got 22% representation of all the non
financial firms in Pakistan, present during the period of 2010 to 2015. Our sam-
ple got representation of all the manufacturing sectors listed in KSE for the
stated time. Manufacturing sector is the largest sector of the Pakistan economy.
Fig. 1. Effect of ownership pattern on agency cost
Table 1. Descriptive statistics of variables
Variables N Minimum Maximum Mean Standard dev

ATR 500 0 5.849 1.31676 1.005924
ACD 500 0.005 0.97 0.54933 0.232779
MNG 500 0 0.73 0.02428 0.078006
FML 500 0 0.921 0.15391 0.225166
INST 500 0.008 1 0.59789 0.309082
GEN 500 0 0.831 0.21549 0.195952
BLOCK 500 0 1 0.51472 0.302945
IDF 500 0 0.926 0.18757 0.21483
DR 500 0 0.934 0.19038 0.176863
DP 500 −14.47 22.726 0.4671 1.828188
SIZE 500 19.055 26.156 2.27E+01 1.48701
GROWTH 500 0.038 15.65 1.34697 1.647557
Table 2. Estimation results of model 1 and 2 by fixed effect and random effect models
Variables Model
1 2
MO 2.612∗ −0.227
FO −0.914 −0.169
IO 1.053∗ .153∗
BO .617∗ −0.047∗∗∗
∗∗
D .166 −0.058∗
DIV −0.007 0
Size 0.071∗∗ .021∗
∗
Q 0.160 .033∗
R 0.481 0.083
Model type Fix effect model Random effect model
Estimation by FE, RE and Hausman test. Dependent
variables are ACE and ACD in models 1&2 respec-
tively. ∗ , ∗∗ , ∗∗∗ show significance at 1%, 5% and 10%
respectively.
The contribution of manufacturing sector of Pakistan in GDP growth is slowly

build up during the last three years from 18 percent to 19.2 percent. Data sources
of Karachi Stock Exchange, State Bank of Pakistan, Bulletins of Statistical
Bureau of Pakistan were used to collect the data (Table 1).
We checked our results by generalized least square method and Hausman
Test. In models 1&2 relationship between agency cost and variables of our inter-
est is explored. The purpose of this model is to find out the joint effect of all the
variables (Table 2).
4 Results and Discussions

Managerial ownership was found influential in lowering agency cost of equity.
Our results are supported by various researches [1,2,6,19,31,32,36,38,39,44,46].
Results tell that MNG is negative and significantly related with ACD. This
means when managers are also owners of the firm, then they don’t overlook debt-
holders interests. Our results are convincing that in Pakistani firms, insiders
don’t create asset substitution or risk shifting problems. They don’t add to
the assets of owners on the expense of debt-holders. And they don’t produce
incentives to expropriate debt-holders’ wealth as they maintain lower proportion
of such assets which is not tied in fix plant and equipment or liquid assets. The
motive behind this deed from managers might be because of the reason that
when insiders own the firm, they want long-term performance and return from
the firm, so they avoid increasing the price of debt.
Relationship between ownership of family and asset turnover of that firm is

negative and significant. This means family has weak monitoring over managers.
Our findings are also consistent with [37] which narrates that in family owned
business, managers work for the welfare of controlling family, rather than for gen-
eral shareholders. This is because Pakistan is among those emerging economies
where minority rights are not well protected under any corporate law author-
ity and due to these reason insiders with more concentration of shareholdings
in companies have the capacity to misuse their rights for their personal bene-
fits. Results are indicating that FML is negatively and significantly related with
ACD i.e. increase in the ownership of family causes reduction in debt agency
cost. Family owned firms have lesser proportion of such assets which are difficult
to monitor by debt-holders [4].
Evidence supported that with the increase in ownership of institutions in
a firm, managers make efficient utilization of assets and agency cost of equity
is reduced. It can be inferred that institutions usually have strong monitoring
power. Due to their monitoring and controlling power, managers are forced to
make efficient utilization of assets and by this equity agency cost is reduced.
Our results are consistent with majority of researchers like [1,5,6,13,46,49].
Regression results illustrate that INST is positive and significantly related with
ACD. A firm which is more owned by institutions, have larger the proportion
of such assets which are not tied in fixed plant and equipment, so debt-holders
cannot monitor such assets easily.
With the increase in the ownership of block-holding, equity agency cost
reduces. Block-holders have incentives to actively monitor agents because they
have larger stake in the firm. Results tell that Pakistani block-holders have influ-
ential monitoring power. When firm is greater subjected to the monitoring of
block-holders, then managers make efficient utilization of assets. So it mitigates
the equity agency cost [6,42]. BLOCK is negative and significant related with
debt agency cost. Higher cost of debt leads to additional cost of financing. Higher
price of debt may result in the financial distress of the firm. As block-holders
have large stake associated with the firm, they want to avoid such situations. So
they try that there should be no debt agency cost [42,44].
Debt ratio (DR) was found to be effective in reducing both ACE and ACD.
But Dividend pay out ratio came out to be insignificant in determining any
type of agency cost. Size reduces ACE but on the other hand it increases the
intensity of ACD. It is due to the reason that larger firms are difficult to be
monitored by debt-holders. Assets of larger firms are diversified. Larger firms
usually have enough free cash flows to cater future growth opportunities [13,17].
Firms having more growth, have better utilization of its assets. GROWTH is
positive and highly significantly related with debt agency cost. A firm with good
growth and investment opportunities is likely to have a relatively high quantity
of cash and other short-term assets for future investment opportunities. The
availability of more liquid assets increase the discretionary power of the manager
hence increase the agency cost of debt (Table 3).
5 Conclusion
This research examines and provides further evidence on agency theory. Present
study looks into the impact of ownership structures over equity agency cost
(ACE) and debt agency cost (ACD). Equity agency cost is the disagreement
of interest between shareholders and mangers whereas debt agency cost is the
divergence of interest between shareholders and debt-holders or debt-holders
and managers. Equity agency cost incurs when managers don’t act in the best
favor of shareholders and acquire perquisites for themselves. Agency cost of
debt arises when managers, themselves or on behalf of shareholders, expropriate
wealth from debt-holders and increase wealth of shareholders or do discretionary
actions on the expense of debt-holders interest. Study incorporates four types of
structures of ownership namely: managerial, general public, family, institutional
and ownership of block-holders. We used different proxies for ACE and ACD.
ACE is measured by asset turnover ratio. ACD is measured by amount of firm
assets not invested in fixed plant and equipment and by liquidity of firm assets.
Our sample is companies from KSE 100 index and study window is 2010–2015.
We found no empirical study that has explored agency cost with set of hypotheses
which present study has used.
Our research tested the convergence of interest hypothesis. In Pakistan, con-
vergence of interest hypothesis is followed because ownership of managers’ result
in the reduction of agency cost. Ownership of family is causing reduction in
efficient utilization of assets. However firms having managerial ownership and
family ownership exhibit lower agency cost of debt. This conveys that insiders
in Pakistan protect the interest of debt-holders. Due to greater monitoring abili-
ties, institutions and block-holders have positive impact on firm performance by
reducing agency cost of equity. In firms, which are more owned by institutions
and block-holders, there is better protection of shareholders’ interests in case
of Pakistan. But institutions only take care of shareholders wealth and don’t
safeguard the interest of debt-holders; agency cost of debt is present in firms,
which are more owned by institutions. But ownership of block-holders, not only
reduces equity agency cost, but also mitigates debt agency cost by protecting
debt-holders wealth and interests. So ownership of block-holders and managers
are the structures of equity ownership, which reduce both types of agency costs.
Pakistani firms which has greater debt ratio, also has efficient asset turnover and
lower proportion of such assets which are difficult to monitor by debt-holders.
Size worsens the both types of agency problems whereas growth opportunities
for firm reduce agency cost of equity but exacerbates agency cost of debt.
Present study classified block-holders as all the shareholders that have more
than 10% ownership of equity. Future researchers may expand this concept by
classification of block-holders. For example, there are two major types of block-
holders in Pakistan, family block-holders (more than 10% equity is owned by
certain family) and institutional block-holders (more than 10% equity is owned
by institutions). There are many types of institutions in Pakistan which have
ownership in companies like NIT, pension fund managers, insurance companies
etc. Future research should check individual monitoring impact of each type of
institution in mitigating agency costs. The study employs five year data (2010–
2015) from Pakistani economy. This period is recognized as unchecked inflation
but stable or declining interest rate. Present study did not control variables for
industry classification. No doubt asset utilization is different for different types of
industries. Similarly other variables are different too for different industries. We
were unfortunately enough that Pakistan has not codes for specific industries like
SIC codes in other countries. If we use codes for different industries for control
for industry specification, results might be better.
Acknowledgements. Authors of this study thank Professor Ahmad Kaleem, Dean

Investment Banking Institute Business School, Melbourne for valuable suggestions.
Appendix
Table 3. Variables and definitions
Variables Formula
Asset Turnover Ratio Net Income divided by Total Assets
Agency Cost of Debt Proportion of firm assets not invested in fixed plant &
equipment = One minus ratio of fixed assets to total
assets
Managerial Ownership The number of common shares own by insiders divided
by total number of common shares outstanding. Insid-
ers consist of officers, affiliated directors, beneficial
owners and principal shareholders [26]
Institutional Ownership Number of shares held by institutions as a proportion
from total ordinary shares of the company [36]
Family Ownership Ratio of shares held by family as a group to total
shares [43]
Ownership of Block-holders Total number of shares held by block-holders divided
by total number of shares outstanding. Where block-
holders are shareholders having more than 10% own-
ership in total equity outstanding
Size Natural Logarithm of total assets
Debt Ratio Book value of contractual long term debt/Book value
of total assets
Dividend Dividend per share divided by earnings per share
Growth Ratio of (market value of equity + book value of debt)
to Book value of total assets
References
1. Agrawal A, Knoeber CR (1996) Firm performance and mechanisms to con-
trol agency problems between managers and shareholders. J Financ Quant Anal
31(3):377–397
2. Ahmed HJA (2008) Managerial ownership concentration and agency conflict using
logistic regression approach: evidence from Bursa Malaysia. J Manage Res 1(1):1–
10
3. Anderson RC, Reeb DM (2003) Founding-family ownership, corporate diversifica-
tion, and firm leverage. J Law Econ 46(2):653–684
4. Anderson RC, Mansi SA, Reeb DM (2003) Founding family ownership and the
agency cost of debt. J Financ Econ 68(2):263–285
5. Ang JS, Cox DR (1997) Controlling the agency cost of insider trading. J Financ
Strateg Decis 10(1):15–26
6. Ang JS, Cole RA, Lin JW (2000) Agency costs and ownership structure. J Finance
55(1):81C106
7. Borokhovich KA, Brunarski KR et al (2005) Dividends, corporate monitors and
agency costs. Financ Rev 40(1):37–65
8. Brockman P, Unlu E (2009) Dividend policy, creditor rights, and the agency costs
of debt. J Financ Econ 92(2):276–299
9. Byrd JW (2010) Financial policies and the agency costs of free cash flow: evidence
from the oil industry. SSRN Electron J https://ssrn.com/abstract=1664654
10. Casson M (1999) The economics of the family firm. Scand Econ Hist Rev 47(1):10–
23
11. Demsetz H, Villalonga B (2001) Ownership structure and corporate performance.
J Corp Finance 7(3):209–233
12. Dharwadkar B, George G, Brandes P (2000) Privatization in emerging economies:
an agency theory perspective. Acad Manag Rev 25(3):650–669
13. Doukas JA, Pantzalis C (2000) Security analysis, agency costs, and company char-
acteristics. Int Rev Financ Anal 14(5):493–507
14. Ducassy I, Guyot A (2017) Complex ownership structures, corporate governance
and firm performance: the french context. Res Int Bus Finance 39:291–306
15. Faccio M, Lang LHP, Young L (2000) Dividends and expropriation. Am Econ Rev
91(1):54–78
16. Fan JPH, Wong TJ (2001) Corporate ownership structure and the informativeness
of accounting earnings in East Asia. CEI Working Pap 33(3):401–425
17. Ferreira MA, Vilela AS (2004) Why do firms hold cash? evidence from EMU coun-
tries. Eur Financ Manage 10(2):295–319
18. Fleming G, Heaney R, Mccosker R (2005) Agency costs and ownership structure
in Australia. Soc Sci Electron Publ 13(1):29–52
19. Florackis C (2008) Agency costs and corporate governance mechanisms: evidence
for UK firms. Int J Manag Finance 4(1):37–59
20. Guariglia A, Yang J (2015) A balancing act: managing financial constraints and
agency costs to minimize investment inefficiency in the Chinese market. J Corp
Finance 36:111–130
21. Jensen MC (1986) Agency costs of free cash flow, corporate finance, and takeovers.
Am Econ Rev 76(2):323–29
22. Jensen MC, Meckling WH (2000) The theory of the firm: managerial behavior,
agency costs and ownership structure. Theory Firm, Bd 1:248–306
23. Jong AD, Dijk RV (2007) Determinants of leverage and agency problems: a regres-
sion approach with survey data. Eur J Finance 13(6):565–593
24. Jung K, Kwon SY (2002) Ownership structure and earnings informativeness: evi-
dence from Korea. Int J Account 37(3):301–325
25. Kalcheva I, Lins KV (2007) International evidence on cash holdings and expected
managerial agency problems. Rev Financ Stud 20(20):1087–1112
26. Kayakachoian G (2000) On agency costs and firms’ decisions
27. Kim KA, Kitsabunnarat P, Nofsinger JR (2004) Ownership and operating per-
formance in an emerging market: evidence from thai IPO firms. J Corp Finance
10(3):355–381
28. Kim S, Lee H, Kim J (2016) Divergent effects of external financing on technology
innovation activity: Korean evidence. Technol Forecast Soc Change 106:22–30
29. Kumar J (2003) Ownership structure and corporate firm performance. Finance
https://core.ac.uk/download/pdf/9315361.pdf
30. Kusnadi Y, Kusnadi Y (2004) Corporate cash holdings and corporate governance
mechanisms. SSRN Electron J https://ssrn.com/abstract=479401
31. Lafond R, Roychowdhury S (2008) Managerial ownership and accounting conser-
vatism. J Account Res 46(1):101–135
32. Lasfer M (2002) Board structure and agency costs. SSRN Electron J https://ssrn.
com/abstract=314619
33. Lie E (2000) Excess funds and agency problems: an empirical study of incremental
cash disbursements. Rev Financ Stud 13(1):219–247
34. Manso G (2008) Investment reversibility and agency cost of debt. Econometrica
76(2):437–442
35. Mao CX (2003) Interaction of debt agency problems and optimal capital structure:
theory and evidence. J Financ Quant Anal 38(2):399–423
36. Mcknight PJ, Weir C (2009) Agency costs, corporate governance mechanisms and
ownership structure in large UK publicly quoted companies: a panel data analysis.
Q Rev Econ Finance 49(2):139–158
37. Morck R, Shleifer A, Vishny RW (1988) Management ownership and market val-
uation: an empirical analysis. J Financ Econ 20(88):293–315
38. Moussa FB, Chichti J (2011) Interactions between free cash flow, debt policy and
structure of governance: 3SLS simultaneous model. J Manage Res 3(2):1–34
39. Mustapha M, Ahmad AC (2011) Agency theory and managerial ownership: evi-
dence from Malaysia. Manag Audit J 26(5):419–436
40. Prowse SD (1990) Institutional investment patterns and corporate financial behav-
ior in the United States and Japan. Brain Res 61(1):267–78
41. Randøy T, Goel S (2003) Ownership structure, founder leadership, and perfor-
mance in norwegian SMEs: implications for financing entrepreneurial opportuni-
ties. J Bus Ventur 18(5):619–637
42. Schulze WS, Lubatkin MH, Dino RN, Buchholtz AK (2001) Agency relationship
in family firms: theory and evidence. Organ Sci 12(2):99–116
43. Setia-Atmaja L, Tanewski GA, Skully M (2007) How do family ownership and con-
trol affect board structure, dividends and debt? australian evidence. In: Proceeding
of The 16th European Financial Management Association Conference. Citeseer,
Vienna, Austria, pp 27–30
44. Singh M, Iii WND (2003) Agency costs, ownership structure and corporate gover-
nance mechanisms. J Bank Finance 27(5):793–816
45. Stepanov S, Suvorov A (2009) Agency problem and ownership structure: Outside
blockholder as a signal. SSRN Electron J 133:87–107
46. Ugurlu M (2000) Agency costs and corporate control devices in the Turkish man-
ufacturing industry. J Econ Stud 27(6):566–599
47. Vilasuso J, Minkler A (2001) Agency costs, asset specificity, and the capital struc-
ture of the firm. J Econ Behav Organ 44(1):55–69
48. Woidtke T (2002) Agents watching agents?: evidence from pension fund ownership
and firm value. J Financ Econ 63(1):99–131
49. Wu L (2004) The impact of ownership structure on debt financing of Japanese
firms with the agency cost of free cash flow. SSRN Electron J
50. Yeo GHH, Tan PMS, Ho KW, Chen S (2002) Corporate ownership structure and
the informativeness of earnings. J Bus Finance Account 29(7–8):1023–1046
The Impact of Institutional Investors on Stock
Price Synchronicity: Evidence from the
Shanghai Stock Market
Li Kun, Lei Yu(B) , and Xiaoxue Hu

245025176@qq.com
Abstract. The article mainly consists of four sections, the first section
is the basic introduction and literature review, the second is the back-
ground review of institutional investors and stock synchronization, the
third is the empirical research and the last is the Conclusion. This paper
use the basic literature review method, descriptive statistics and OLS
models analysis to explore the Impact of institutional Investors on Stock
Price Synchronicity. Using the data of 569 listed firms at Shanghai stock
market from 2009 to 2012, this paper investigates the impact of insti-
tutional investors’ behaviors on stock price synchronicity. We find that
the institutional investors’ long-term investments can reduce stock price
synchronicity and increase the information efficiency of the stock market.
However, the short-term trading of institutional investors does not nec-
essarily reduce stock price synchronicity. Therefore, the Chinese police-
makers should regulate institutional investors and encourage construc-
tive long-term investment behaviors, to make it really rise to enhance the
effectiveness of capital market information, and improve the efficiency of
capital market resource allocation. The main contributions of this paper
is that we investigate the stock price synchronicity from the perspective
osf institutional investors’ behaviors and A–share market.
Keywords: Stock price synchronicity · Institutional investors · Long-

term investment · Investment activity
1 Introduction
The term “stock price synchronicity” refers to a phenomenon in which stock

prices rise and fall in tandem. As a proxy for stock price informativeness, a
higher degree of synchronization reflects more commonality regarding informa-
tion on markets and industry; however, heightened synchronization suggests less
detailed firm-specific information contained in stock prices, weakening the capac-
ity to transfer signals of corporate value. In this circumstance, the pricing mech-
anism fails, thus reducing the efficiency of resource allocation in capital markets
through prices.
The Impact of Institutional Investors on Stock Price Synchronicity 1549
Morck and Yeung [12] examined the stock price synchronicity in the major
stock markets around the world during the first 26 weeks of 1995. They found
that the average stock price synchronicity was higher in less developed coun-
tries. In the Chinese stock markets, approximately 79% of stocks move in the
same direction, only lower than Poland where all the stocks moved in the same
direction. After a lapse of nearly two decades, what can be observed regarding
stock price synchronicity in the Chinese stock markets? Figure 1 illustrates the
price synchronicity of 1,218 stocks on the Shanghai Stock Market during the
first 26 weeks of 2013. As Fig. 1 shows, averagely more than 70% of stocks were
moving in the same direction, suggesting that price synchronicity still remained
at a high level in China.
Fig. 1. The changing synchronicity of Chinese stock prices
Stock price movement is driven by investors, especially those who domi-

nate the market. The Chinese stock market is dominated by individual investors
with short-term vision. They tend to blindly follow trends without investigating
the basic information of firms, thus exacerbating price synchronicity. In con-
trast, institutional investors are usually more professional and rational, and they
engage in deeper and more fundamental research on firm-specific information.
Just as Hasan and Song [6] argued, stock price synchronicity is lower when
institutional investors dominate the market. However, Boubaker et al. [2] dis-
agreed with this conclusion, they believe that the relation between institutional
behavior and stock price synchronicity is not that straightforward. In recent
years, institutional investors in China have been encouraged to expand for three
main reasons: to return the markets to reasonable levels; to improve the pricing
mechanism; and to heighten the effectiveness of market information. Therefore,
1550 L. Kun et al.
it is important to clarify the influence of institutional investors on stock price

synchronicity.
Since Roll, Richard [4] proposed as a proxy for stock price synchronicity, a
growing body of literature has addressed this issue. Morck [12] found that the
systematic component of returns variation is large in emerging markets, and
lower firm-specific returns variation is associated with poorer corporate gover-
nance. Jin and Myers [3] noted that lack of transparency increases by shifting
firm-specific risk to managers, and opaque stocks with high are also more likely
to crash. Heng [1] found that both stock price synchronicity and crash risk are
negatively related to the firm’s ownership by dedicated institutional investors,
which have strong incentive to monitor due to their large stake holdings and
long investment horizons. However, Skaife et al. [9] argued that the zero-return
metric is a better measure of firm-specific information impounded into share
prices than the synchronicity measure internationally. In this paper, we use as
the proxy for stock price synchronicity, because of its broad recognition and
extensive applicability.
The stock price synchronicity in the Chinese stock market is significantly
higher than in many other markets. Some researchers believe that this is related
to the institutional development in China, e.g. Iftekhar [6] found that better
institutions–respect for property rights and the rule of law - increase the value
of private information which reduced synchronicity in the Chinese stock mar-
kets. However, most domestic researches tend to conclude that the high degree of
speculation in the Chinese stock markets lead to high price synchronicity. This
speculation is mainly caused by the composition of Chinese investors, which con-
sists of many individuals but few institutions. This reinforces the view that insti-
tutional investors can reduce the price synchronicity. The institutional holdings
can increase the quantity of firm-specific information, thus reducing stock price
synchronicity and enhancing the efficiency of company signal-transmission mech-
anisms. However, some researchers disagree with this conclusion. Conversely,
they claim that the institutional investors display “herd behavior” that can
exacerbate stock price synchronicity [7].
In summary, scholars have not reached a consensus regarding the impact
of institutional investors on stock price synchronicity, and more researches are
required to confirm related findings, especially in the Chinese stock markets
where stock price synchronicity is high. The contribution of this paper is reflected
in two aspects. Firstly, this paper adds an index to demonstrate the investment
activities of the institutional investors under examination. We refine institutional
investor behaviors into shareholding level and examine both the effects of their
long-term and short-term behaviors. This deepens the analysis of institutional
investors’ effect on stock price synchronicity. Secondly, this paper focuses on the
study of Shanghai Stock Market, taking 569 listed companies from 2009 to 2012
as samples. The recent panel data illuminates the latest situation of stock price
synchronicity in China, thus enriching the research on the effects of institutional
investors on stock price synchronicity.
The remainder of this paper is organized as follows. Section 2 provides the the-
oretical background. Section 3 presents the research design and the main empir-
ical results. Section 4 concludes the paper.
2 The Theoretical Background

2.1 Stock Price Synchronicity
In capital markets, the fluctuation of stock prices is driven by information shocks

(unexpected or unpredictable events that affect perception of stock values).
According to Roll [4], this information can be categorized into market-level and
firm-level sources. The information at the market level refers to the external
macro-level environment, including policies, regulations and important economic
events. Firm-level information refers to the unique information of firms, including
their changing financial indicators, executive turnover and other events. If the
fluctuation of stock prices is caused by market-level information, then all stocks
are vulnerable to the same systematic risk, and thus a significant correlation
among different stock prices indicates a high degree of stock price synchronicity.
As Song [10] finds, firms’ stocks are less synchronized with the entire market and
have less crash risk if firms have superior accounting disclosure policies. However,
if the price fluctuation is based on a firms’ specific information, then the entire
market displays heterogeneity in terms of its stock price fluctuation. Therefore,
significant stock price synchronicity in a specific market is due more to market
information and due less to the firm-specific information in stock prices. On this
occasion, a company’s stock price cannot reflect the true quality of a company,
which lowers the reference value of corporate information for investors. This
damages the company signal transmission mechanism and lowers the efficiency
of price-based market resource allocation.
2.2 Institutional Investors’ Effects on Stock Price Synchronicity
Based on transaction incentives, traders can be classified into three categories:

informed traders, noise traders, and liquidity traders. Informed traders conduct
transactions based on a firm’s value, and thus the related information can be
conveyed into stock prices by their transactions.
Institutional investors are generally regarded as informed investors. They
have access to information resources that individual investors do not have, and
their teams include professional research staff. These advantages can make their
evaluations of firms more accurate than those of individual investors, increasing
the quantity of firm-level information contained in stock prices. Related studies
have shown that the behaviors of institutional investors can affect the process
of stock pricing and their information environment, and their transactions add
firm-specific information into stock prices. This can help to reduce stock price
synchronicity and improve the efficiency of capital markets.
1552 L. Kun et al.
The influence of institutional investor behaviors on stock price synchronicity

also depends on their specific investment behaviors. On the one hand, if the insti-
tutional investors focus on short-term benefits and have high portfolio turnover
rates, they will not have a strong incentive to collect firm-specific information
related to corporate long-term benefits. In that case, they will not be able to
decrease stock price synchronicity. On the contrary, the short-termism of insti-
tutional investors will make them more prone to herd behaviors, make them
follow the trends to buy or sell stock and intensify stock price synchronicity. On
the other hand, if institutional investors maintain a large stake in the firm for
a long period, they may have a greater incentive to closely monitor their port-
folios. To secure their investment, they may even take large positions with their
shares and involve themselves in corporate governance. Therefore, the long-term
investment of institutional investors is logically more likely to decrease the stock
price synchronicity.
3 Empirical Design and Results
3.1 The Sample and Data
The sample consists of listed firms on the Shanghai Stock Market from 2009
to 2012, and we exclude the following firms from the sample: (1) firms without
complete data; (2) financial institutions; (3) utility firms; (4) firms that con-
ducted IPO or major assets restructuring in the current year; (5) and firms
whose stocks were delisted or suspended. Finally we obtain 569 eligible compa-
nies, a total of 2276 sample points of panel data. All data are from China Stock
Market & Accounting Research Database (CSMAR) and RESSET database.
Table 2 shows the descriptive statistics.
3.2 Variables Construction
(1) The dependent variable: stock price synchronicity

To measure the synchronicity of stock price, we first calculate the for firm i
in fiscal year t from the expanded market model regression in Eq. (1).
ri,w = α + βi × rm,w + εi,w , (1)
where ri,w is the return of firm i in week w, rm,w is the market return in week w.
Since R2 is highly skewed and bounded between 0 and 1, we apply a logistic
transformation to obtain a near formally distributed variable, SYNCH. A higher
value of SYNCH indicates that the stock price is more synchronized.

Ri,t
2
SYNCHi,t = ln . (2)
1 − Ri,t
2
(2) The explanatory variables: institutional investor behavior

(a) The measurement of institutional investors’ overall behaviors
We add up all the proportions of shares held by institutional investors in a
listed firm to construct a proxy for the total stake of institutional investors.
n
j=1 INSj,i,t
LIi,t = . (3)
TOTALi,t
The is the institutional investors’ total proportion of firm i in the year t,
INSj,i,t is the shares of firm i held by institutional investor j in the year t, and
TOTALi,t is the total shares of firm i in the year t.
(b) The measurement of institutional investors’ investment liveness
According to Liu [8], 90% QFII (the Qualified Foreign Institutional Investors)
hold the stocks for 1–4 quarters, and the QFII whose holding period are more
than 5 quarters accounts for less than 10%. Therefore, we can deduce that the
institutional investors are not necessarily long-term investors. In order to mea-
sure the short-termism of institutional investors, we construct an indicator.
ABS(IAi,t − IAi,t−1 )
SIi,t = , (4)
VOLi,t
where IAi,t is the total shares of firm i held by institutional investors in the
year t, and VOLi,t is the individual shareholders’ trading volume of firm i in the
year t. So SIi,t is the ratio of the absolute value of institutional investors’ shares
variation and stocks volume for firm i in the year t. The higher SIi,t means more
active investment behavior of institutional investors, and short-termism is more
significant.
(3) The control variables
In order to better examine the influence of institutional investors on stock
price synchronicity in China, our control variables include the following: ROE,
BM, SIZE, LEV, and CEN.
ROE is the contemporaneous income before extraordinary items divided by
the book value of equity. ROE is a core indicator of corporate management
performance, and when related to stock price synchronicity, ROE differs across
models. Thus, we presently avoid making calculations on the regression coeffi-
cient.
BM is the ratio of the market value of equity to the book value of equity, and
it is often used to measure the growth of a company. A lower BM ratio means
stronger growth for a company. Many scholars agree that a company’s growth
and expanded demand for financing leads to more information disclosure; con-
sequently more information will be contained in its stock price. In contrast, a
blue-chip company with less demand for financing is reluctant to release informa-
tion. Therefore, we expect that the company with a lower BM ratio will exhibit
a less synchronous stock price.
SIZE is the natural log of the firms’ market value of equity at the end of
the fiscal year. Roll [3] held that a larger company is more vulnerable to the
influence of macroeconomic or industrial events. Meanwhile, a larger company
1554 L. Kun et al.
has a stronger ability to avoid risk, leading to less stock price volatility. Thus
it can be concluded that the larger size of a company would bring about more
significant stock price synchronicity. Others, however, hold that due to its greater
scale, the company can garner more attention, which can add more information
into its stock price and attenuate price synchronicity. Because of this controversy,
we do not calculate a regression coefficient of the relation between SIZE and stock
LEV is the book value of all liabilities scaled by total assets at the end of the
fiscal year. LEV can reflect the level of a company’s financial risk, and a higher
LEV indicates a higher financial risk in a company. In this case, the company is
more sensitive to credit risk, and thus the price fluctuation is mainly affected by
the firm-specific financial risk. Thus, we expect a negative correlation between
LEV and stock price synchronicity.
CEN is the concentration of corporate ownership, and a higher CEN means a
stronger control held by large shareholders and more rights to internal informa-
tion. These large shareholders can transfer firm-specific information into stock
prices through their transactions, accordingly raising the firm-level information
content in stock price [5]. However, some researchers found a concave functional
relationship between stock price synchronization and CEN: CEN increases at a
decreasing speed with a rising CEN until the maximum point, and then it starts
to fall [11]. This suggests an unclear relationship, and any clarification would
require further research.
3.3 Model Setup

We estimate the relation between institutional investors and the firms’ stock
price synchronicity in Shanghai A-stock market by the following model:
Model (1): SYNCHi,t = αt + γi + β1 LIi,t + β3 ROEi,t + β4 BMi,t + β5 SIZEi,t

+ β6 LEVi,t + β7 CENi,t + εi,t ,
Model (2): SYNCHi,t = αt αt + γi + β2 SIi,t + β3 ROEi,t + β4 BMi,t + β5 SIZEi,t
Model (3): SYNCHi,t = αt + γi + β1 LIi,t + β2 SIi,t + β3 ROEi,t + β4 BMi,t + β5 SIZEi,t
where the explained variable SYNCHi,t is the measurement of stock price syn-
chronicity of firm i in the year t, LIi,t is the shares proportion held by institu-
tional investors of firm i in the year t, and SIi,t is the transaction activity of
firm i conducted by institutional investors during the fiscal year t, as a proxy
for the short-term behavior of institutional investors.

Using the data of 569 listed firms at Shanghai stock market from 2009 to 2012,
Table 1 provides the descriptive statistics of the sample. The mean and the max
of Stock Price Synchronicity is 0.363 and 0.840 respectively. After applying a

logistic transformation to obtain a near formally distributed variable, SYNCH,
the mean, max and min of the SYNCH is −0.673, 1.650 and −5.790 respectively.
The mean of LI which is the proportion of shares held by institutional investors in
a listed firm is 0.155. The mean of SI which is the transaction activity conducted
by institutional investors is 0.025. ROE, return on equity, show a large variation,
with a mean of 4.815 and standard deviation of 17.517.
N Min Max Mean Std.

Statistic Statistic Statistic Statistic SE Statistic
R2 2276 .000 .840 .363 .0034 .162
SYNCH 2276 −5.790 1.650 −.673 .0188 .896
LI 2276 .000 .920 .155 .004 .169
SI 2276 .000 2.020 .025 .001 .0667
ROE 2276 −289.460 84.830 4.815 .367 17.517
BM 2276 .010 2.060 .429 .006 .280
SIZE 2276 −4.850 28.410 4.749 .215 10.262
LEV 2276 .030 .970 .539 .004 .186
CEN 2276 .000 .980 .491 .003 .165
N 2276
In order to decide whether we should use fixed effect panel data model or random
effects model on the panel data, we conduct a Hausman Test on the foregoing
model.
Table 2. Hausman test result of investment behaviour
Test Summary Chi-Sq. Statistic Chi-Sq.d.f Prob.

Cross-section random (long term) 38.597 6 0.000
Cross-section random (short term) 49.000 6 0.000
Cross-section random (overall) 46.736 7 0.000
This table shows the Hausman Test result of the institutional investors’ over-
all investment behavior. The significant result encourages us to apply a fixed-
effects model in the regression process.
1556 L. Kun et al.
Table 3. The empirical result
Variable Mod (1) Mod (2) Mod (3)

C −1.135 (0.000) −1.156 (0.000) −1.129 (0.000)
LI −0.487∗∗ (0.037) −0.527∗∗ (0.033)
SI −0.055 (0.846) 0.149 (0.617)
ROE −0.005∗∗∗ (0.001) −0.005∗∗∗ (0.001) −0.005∗∗∗ (0.001)
BM 0.871∗∗∗ (0.000) 0.888∗∗∗ (0.000) 0.868∗∗∗ (0.000)
∗∗ ∗∗
SIZE 0.004 (0.013) 0.004 (0.023) 0.004∗∗ (0.015)
∗ ∗
LEV −0.502 (0.070) −0.474 (0.087) −0.498∗ (0.072)
CEN 0.889∗ (0.051) 0.742∗ (0.099) 0.882∗ (0.053)
Adjusted R-squared 0.311 0.310 0.311
Note: ∗∗∗ p < 0.01, ∗∗ p < 0.05, ∗ p < 0.1.
Table 3 reports the estimated coefficients and associated p values. From this
table we find the estimated coefficient of LI is −0.487 in Model (1) and −0.527 in
Model (3), and both are significant at the 5% level. This demonstrates a signifi-
cantly negative correlation between the proportion of institutional shareholding
and stock price synchronicity, namely, the stock with a higher proportion of
holding produces a less synchronous stock price. This supports the argument
that institutional investors’ behavior helps to convey firm-specific information
to the market and facilitate a company’s signal transmission mechanism, which
can improve the efficiency of resource allocation with prices.
The other explained variable, SI, is a proxy variable for institutional short-
term transactions. The empirical results show a negative correlation (−0.055) in
Model (2) and a positive correlation (0.149) in Model (3), neither is significant,
which is consistent with expectations earlier in this paper. This indicates the
insignificant correlation between the institutional investor activity and stock
In this paper, we believe this insignificant correlation is caused by the double
effects of institutional investors’ speculative behavior. On the one hand, frequent
short-term trading transmits corporate-specific information to prices, which can
raise the information content in stock prices and help relieve high stock price
synchronicity. On the other hand, there herding is visible in the behavior of
these informed investors, causing stock price to fluctuate in the same direction.
Therefore, the offsetting dual functions lead to an insignificant effect on stock
Empirical result shows that ROE and BM have significant influences on stock
price synchronicity, and both of them are significant at the 1% level in all the
three models. Firstly, the higher ROE, the less there is price synchronicity. The
firm with a higher ROE suggests better corporate management; consequently,
the firm gains more attention by institutional investors and the public, and thus
the firm-specific information can be readily conveyed in the price. However, the
absolute value of its estimated coefficient is small, which illustrates a limited
influence on price synchronicity. Secondly, the estimated coefficient of BM is

significantly positive, meaning that stronger growth companies are more likely
to be tracked by investors, resulting in more information in the stock price.
The coefficient of SIZE is significant at the level of 5% in all three models, and
this positive correlation is consistent with our former expectation. Thus a larger
company suffers from more synchronous prices. The estimated coefficient of LEV
and CEN are significant at the 10% level in all three models, and the stock price
is less synchronous in a company with higher LEV or lower CEN.
Overall, the adjusted of the model test is approximately 31%, which indicates
our model’s high explanatory capability. Therefore, our conclusion is highly cred-
ible.
4 Conclusion
The empirical results of this paper suggest that institutional investment can sig-
nificantly decrease stock price synchronicity on the Shanghai Stock Market, and
that this influence is achieved by long-term shareholding and not by frequent
transactions. This paper provides empirical grounds for encouraging institutional
investors to engage in long-term value investment, which can increase the effi-
ciency of price-based resource allocation in our capital markets. For the policy
- makers, it is necessary to improve the regulations of China’s capital market in
order to create a favorable external environment for institutional investors. They
also should guide institutional investors to long-term value investment, enhance
the confidence of long-term value investment and improve the quality of listed
companies.
References
1. An H, Zhang T (2013) Stock price synchronicity, crash risk, and institutional
investors. J Corp Financ 21(1):1–15
2. Boubaker S, Mansali H, Rjiba H (2014) Large controlling shareholders and stock
price synchronicity. J Bank Financ 40(1):80–96
3. Cfa JHG (2006) R2 around the world: new theory and new tests. J Financ Econ
79(2):257–292
4. Epps TW, Epps ML (1976) The stochastic dependence of security price changes
and transaction volumes: implications for the mixture-of-distributions hypothesis.
Econometrica 44(2):305–321
5. Han JF, Wang ZY (2012) Can the major shareholders trading improve market
efficiency-from the perspective of the information content of stock price. J Shanxi
Financ Econ Univ 7:38–45 (in Chinese)
6. Hasan I, Song L, Wachtel P (2014) Institutional development and stock price syn-
chronicity: evidence from China. J Comp Econ 42(1):92–108
7. Jin Y, Yan M et al (2016) Stock price synchronicity and stock price crash risk:
based on the mediating effect of herding behavior of qfii. China Financ Rev Int
6(3):230–244
8. Rao Y, Min L (2013) The influence of qfii holding on stock price synchronicity in
China. J Manage Eng 02:202–208 (in Chinese)
1558 L. Kun et al.
9. Skaife HA, Gassen J, LaFond R (2006) Does stock price synchronicity represent
firm-specific information? The international evidence. MIT Sloan Research Paper
10. Song L (2015) Accounting disclosure, stock price synchronicity and stock crash
risk: an emerging-market perspective. Int J Acc Inf Manage 23(4):851–854
11. Wang X (2013) Ownership concentration and stock price synchronicity: the evi-
dence from Chinese listed companies. Contemp Econ 14:94–95 (in Chinese)
12. Yeung B, Wu W (1999) The information content of stock markets: why do emerging
markets have synchronous stock price movements? J Financ Econ 58(12):215–260
Research on Risk Allocation Model in PPP
Projects: Perspectives from the SOEs
Chuan Chen, Tengteng Yan(B) , and Ping Chen

ytt19941128@foxmail.com
Abstract. PPP (Public-Private-Partnership), which is the cooperation

between the government and private sector, is a major innovation in
public service mechanism. Effective risk identification and reasonable
risk allocation is not only a major challenge to each party, but also the
key to manage projects efficiently, reduce the whole life cost and improve
benefit. Under the influence of special economy, politics and culture, the
state-owned enterprise accounts for absolute proportion of private sector
in China. However, the systematic analysis of the state-owned enterprise
participating in the PPP project is deficient. Given that the Chinese cen-
tral governments have sped up the development of the PPP application
in recent years, this paper provides a risk allocation model through the
fuzzy comprehensive evaluation method, which is aiming at optimizing
the allocation of risk, reducing transaction costs and enhancing compre-
hensive benefits from the perspective of state-owned enterprises (SOEs).
Keywords: PPP · SOEs · Risk allocation · Fuzzy comprehensive eval-

uation
1 Introduction
PPP (Public-Private-Partnership) model is an innovative financing tool for

infrastructure projects procurement in China. After 2013, the 18th national
congress of CPC proposed that the market should play a decisive role in resource
allocation. In September 2014, the ministry of finance issued Notice about pro-
moting PPP Model and devoted continuously to promoting PPP Model nation-
wide. In September 2015, The State Council issued Guideline to Urge SOE Mod-
ernization through Mixed Ownership Reform, which had further defined the over-
all requirements of the state-owned enterprises developing the mixed ownership
economy [15]. According to the statistics of National PPP Integrated Informa-
tion Platform, there were 9,285 projects that would be procured by PPP up to
the end of June 20, 2016, and their total investment surged to 10.6 trillion [3].
Because of the national condition, the state-owned enterprises with large scale,
strong capital power and rich experience have inherent advantage to participate
in the PPP projects. As a result, the state-owned enterprises accounted for 80%
1560 C. Chen et al.
while the private enterprises only accounts for 20% of the private sector [14].
With a large amount of investment, long term cooperation, technical complex-
ity, and uncertainty factors, appropriate risk allocation is not only a critical
success factor (CSF) of PPP but also an important driving factor of meeting
the Value for Money [4,6]. While the country issued relevant legal policies to
standardize the operation and execution of PPP projects, the guidance of risk
allocation is still relatively broad and vague [7].
At present, there is weak control force between government and SOEs, and
government always has absolute competitive advantage. Initially, some risks
would be allocated to SOEs with underperforming ability to control risk and no
willingness to take risks. Once the risks occur, SOEs are unable to control the
consequences of risk, thus the government need to bear losses, which inevitably
causes most risks retransferred to the government at a higher cost. In addition,
the complex contractual arrangements in PPP projects would lead to risk expo-
sure, and inappropriate risk allocation only let the infrastructure debt simply
transfer from the local government balance sheets to SOEs. However, it can’t cut
overall leverage utility of the public sector effectively. In terms of risk allocation,
the government and SOEs always face severe challenge, strict requirements and
high standards. However, there are few systematic analysis of the state-owned
enterprises participating in the PPP project in Chinese journals. Therefore, this
paper will establish risk allocation model combined with fuzzy comprehensive
evaluation method and entropy coefficient method to optimize the allocation of
risk perspectives from the SOEs.
This paper is organized as the following: Sect. 1 begins with an introduction,
Sect. 2 makes a literature review about PPP risk allocation, Sect. 3 describes
the modeling ideas and methods combined with fuzzy comprehensive evalua-
tion method and entropy coefficient method, which optimizes multi-objective
decision of risk allocation in PPP projects perspectives from the SOEs, Sect. 4
takes the case of Mianzhu Integration of Water Supply and Drainage Project to
demonstrate the effectiveness of the model, Sect. 5 presents a brief conclusion .
2 Literature Review
Zhang put forward that the research topics of PPP papers in Chinese and Inter-
national Journals mainly include PPP model’s application, risk management,
financing and economic issues, legal and procurement issues and government
regulation [16]. The foundation of financing risk-sharing in PPP projects is the
accurate identification of risk factors. Grimsey identified nine risks: financial risk,
political risk, environment risk, construction risk, technology risk, operational
risk of infrastructure, infrastructure recovering risks, force majeure and project
default risk, introduced the risk evaluation index and evaluation method from
the perspective of different project stakeholders [9]. Bing creatively carried out
the dividing analysis of risk from the macro, medium and micro three levels
[2]. Zou identified twenty-seven critical risk factors influencing the success of
PPP projects from five dimensions: the owner, design, contractors, government
institutions and the external environment [17].
Risk Allocation Model in PPP Projects 1561
The academic research for the risk-sharing mechanism in PPP project mainly
concentrated on the risk-sharing principle, method and model. Abednego put for-
ward the connotative meaning and conditions of risk sharing [1]. Deng analyzed
and summarized nine principles in risk-sharing of PPP projects, such as the
principle of responsibility, justice principle [8]. Bing put forward that the risk-
bearer should be divided into the private sector, public sector, together under-
taking, which is depended on the specific conditions of project [2]. Jin and Chang
explained the risk allocation of PPP projects from the perspective of transaction
cost theory (TCE) and resource-based view (RBV) [5,11]. At present, most stud-
ies tend to adopt quantitative methods to optimize the PPP project risk sharing
mechanism and process. The questionnaire survey was the typical quantitative
research method in risk allocation [16], such as Ke investigated the experienced
practitioners to determine the preferences of PPP project risk allocation [13],
Guo adopted the real option method to research the risk allocation strategy of
highway PPP projects under the delay investment decisions [10], Jin selected
the artificial neural network to realize the allocation [12].
In conclusion, with the perfection and development of theory and practice of
risk-sharing in PPP projects, the scholars at home and abroad had a thorough
study in risk allocation, which mostly confined to some specific risks that should be
shared to specific stakeholders, government or private sector. However, there are
few studies involved specific allocation strategies and proportions. Moreover, there
is almost no study about the risk allocation whose perspective is from the SOEs.
Identify risk factors of PPP project
Determine the risk evaluation index

system for PPP project
Determine the index evaluation Determine the index weight based on

standard and membership vector entropy coefficient method
Calculate the comprehensive

evaluation vector
Determine the sharing subject and

sharing proportion for specific risk
Fig. 1. The procedure of risk allocation whose perspective from the SOEs in PPP
project
3 Method
The quantitative method adopted in optimizing risk allocation mainly include
game theory, real option theory, artificial neural networks and fuzzy compre-
hensive evaluation theory. The game theory, option theory and artificial neural
1562 C. Chen et al.
networks objectively reflect the actual negotiation process with complex model,
which are difficult to apply in practice. In consideration of subjectivity and uncer-
tainty in risk-decision mechanism, the fuzzy comprehensive evaluation (FCE)
based on fuzzy mathematics will be reasonable in optimizing risk allocation. It
can turn the qualitative questions into quantitative analysis based on member-
ship degree theory. In addition, the method has the advantages of clear result,
quite strong systematicness and can better solute the problems that are difficult
to quantify, which has been widely applied in the field of engineering and applied
science, such as water quality evaluation, clinical diagnosis, information technol-
ogy, air quality assessment and product design. Some scholars also gradually
attempt to apply fuzzy mathematics in the construction risk management. The
paper establishes the PPP project risk allocation model to optimize the multi-
objective decision-making in PPP project risk allocation perspectives from the
SOEs. Figure 1 summarizes the procedure of PPP project risk allocation.
3.1 Identify Risk Factors of PPP Project

Risk identification further defined the risk factors and risk category, which is
the foundation of risk allocation. In the practice of risk management in PPP
project, there are several applicable methods, such as the Delphi method, Check
list method and Brainstorming method. There are not many differences in risk
factors and risk category between the SOEs and private enterprise, but there
are considerable differences in the specific risk assessment. Moody’s Investors
Service said that China’s state-owned enterprises were heavily involved in the
PPP project, and it would reduce government leverage and debt at a certain
extent. With the promoting of supply-side structural reform, Moody’s said that
the proportion of the debt in GDP in SOEs reaches to approximate 115%. Mostly
SOEs faced systemic downgrade in credit rating, these factors would affect the
risk assessment of PPP project.
3.2 Determine the Risk Evaluation Index System for PPP Project
This section discusses the risk evaluation index system based on allocation prin-
ciple and influence factors. There are many factors and principles affecting the
risk allocation of state-owned enterprises, the paper selects the basics of qual-
itative research to comb the influencing factors and principles. The influenc-
ing factors include project system properties, stakeholders’ comprehending mis-
take for the PPP model, the risk-taking attitude and intention of stakeholders.
Combined with the theoretical research and case study, the guiding principle of
PPP project could be summarized as follows: equitable principle, criterion of
liability, effective control principle, risk premium peer, upper limit principle and
dynamic principle. Setting up the evaluation index system should be carried out
in accordance with the principle: quantifiable, critical, objectivity, systematic,
operability, independence, comparability and representative, which can full and
accurately reflect the comprehensive management ability to control risk of the
specific subject. Table 1 shows the evaluation index system.
Table 1. The evaluation index system in risk allocation
The evaluation index system

Ability to control U1 Foresee Risk It can predict the probability and
risk evaluate the consequences of risk
occurrence (Degree of information
symmetry)
U2 Ability to monitor The ability to avoid, reduce,
risk supervise and control the likelihood
of occurrence for the specific Risk
U3 Ability to improve Construct management information
the risk system, improve the measures of
management system dealing with risks
Ability to U4 The flexibility of The ability to stabilize and control
undertake risks dealing with risk risk consequences, don’t make it
worse, take risks in the lowest price
U5 Solid financial Private capital (asset-liability ratio,
performance risk reserve rate, profitability, etc.)
the government (fiscal balances,
GDP growth, debt levels)
U6 The experience to Rich experience and reputation of
cope with the risk honoring agreements makes better
coordinate various stakeholders, and
then take risks at the lowest cost
Risk premium U7 Obtain direct The ability to get a reasonable and
level premium level acceptable level for risk premium
U8 Obtain indirect The ability to improve the
premium level credibility, reputation and efficiency
of stakeholders
Risk preference U9 Risk attitude The tendency of the subject facing
the specific risks, such as risk-averse,
risk-neutral, risk-taking
U10 Inclination of Transfer inclination based on the
transfer risk ownership and control power
between the SOEs and the
government
3.3 Determine the Index Evaluation Standard and Membership

Vector
The fuzzy comprehensive evaluation model is based on the franchise agreement

between the government and private sector. Combined with the risk evaluation
index system in Table 1 and triangular fuzzy set theory, this paper set up the
fuzzy reasoning rules. V includes five evaluation standards, V = (Very low, Low,
Medium, High, Very high).
1564 C. Chen et al.
The triangular fuzzy function T = (f, g, h), Very low (0%, 15%, 30%), Low
(15%, 25%, 35%), Medium (35%, 45%, 55%), High (55%, 65%, 75%), Very high
1
(75%, 85%, 95%), Through the formula V (T ) = (f + g + h) to calculate the
3
fuzzy evaluation of value V = (15%, 25%, 45%, 65%, 85%).
⎛ ⎞
r11 r12 · · · r1m
⎜ r21 r22 · · · r2m ⎟
⎜ ⎟
R=⎜ . . . . ⎟.
⎝ .. .. . . .. ⎠
rn1 rn2 · · · rnm
We use Delphi method to determine the evaluation value. This paper establishes
the subordinated matrix R based on the index system and evaluation standard
to calculate the subordinated vector ri = (ri1 , ri2 , · · · , rim ) indicates the value
of index i in the evaluation standard of j.
3.4 Determine the Index Weight Based on Entropy Coefficient

Method
Considering the subjectivity of expert evaluation system, this paper adopts
entropy coefficient method to determine the index weight, thus enhancing the
reliability of the model and improving the accuracy and rationality of the evalu-
ation results. The entropy concept derives from thermodynamics, which initially
described the system disorder. Information entropy value is directly propor-
tional to the disorder, which is inversely proportional to the information util-
ity value. Entropy coefficient method is mainly according to the principle of
V arianceDrive, which can largely avoid human factors to determine the index
weights. The specific steps are as follows:
(1) Use the Stirling Formula to calculate the entropy value of each index utilizing
Stirling Formula.
m
1
ei = − Rij InRij .
Inm j=1
(2) Determine the weight value of each index.

• Utility value hi is the difference between the entropy ei and 1. hi = 1 − ei ;
• The weight value of each index can be indicated as W =
(w1 , w2 , w3 · · · wi );
hi
wi = n .
i=1 hi
3.5 Calculate the Comprehensive Evaluation Vector

According to subordinated vector and the index weight, we need to calculate
the comprehensive evaluation vector B = W ◦ R, ◦ represents fuzzy operator,
Table 2. Alternative fuzzy operators
Characteristic Alternative fuzzy operators

M (∧, ∨) M (•, ∨) M (∧, ⊕) M (•, ⊕)

n
n

Formula max min(wi , rij ) max min(wi , rij ) min 1, min(wi , rij ) min 1, wi rij )
1≤i≤n 1≤i≤n i=1 i=1
Manifests the Not obvious obvious Not obvious obvious
weighting
function
Comprehensive Weak Weak Strong Strong
degree
Using R Inadequate Inadequate More adequate Adequate
information
Type Dominant-factor Dominant-factor Weighted average model Weighted average
Highlight Highlight model
the commonly used fuzzy operators include M (∧, ∨), M (•, ∨), M (∧, ⊕), M (•, ⊕),
M (•, ⊕) comprehensively consider the subordinated vector and weight of the
index, and other fuzzy operator would lack some information M (∧, ∨), M (•, ∨)
are suitable for single constraints model, which only take the main indicators
into account and ignore the other secondary index. The fuzzy operators’ char-
acteristics are shown in Table 2.
Based on the above analysis, this paper will choose M (•, ⊕)to calculate the
comprehensive evaluation vector B = (b1 , b2 , b3 · · · bm ). The methods widely
adopted to deal with the evaluation vector are weighted average method and
the principle of maximum membership degree. However, the latter take the
concept of optimal sorting in the evaluation scheme, and sometimes it would
miss key information. So this section select the weighted average method to
normalize the index system, which makes the comprehensive evaluation vector
B = (b1 , b2 , b3 , · · · , bm ) converted to comprehensive score S. S = BK × V , V
represent the weight of e index evaluation standard, V = (0.150.250.450.650.85).
3.6 Determine the Sharing Subject and Sharing Proportion for

Specific Risk
The government and SOEs respectively undertake a comprehensive assessment
of specific risk, the government should bear the proportion Q1 = S1S+S
1
2
× 100%,
S2
SOEs should bear the proportion Q2 = S1 +S2 × 100%.
4 Case Study
To further integrate the existing system of water supply and drainage, improve
the systematicness, harmonization, sharing and economy of water supply and
drainage, the government of Mianzhu, Sichuan province proposed to launch the
Mianzhu Integration of Water Supply and Drainage project in November 2015.
The project’s investment is 680 million yuan, and its concession period is about
30 years, which includes the construction and reform period of 3 years. Mianzhu
1566 C. Chen et al.
Mianzhu goverment
Authorize
Platform company of Housing and Urban-Rural

Equity ratio 10%
local government Construction Agency
Franchise Agreements
Asset transfer
agreement
Special Purpose Vehicle
Organize Loan Financing Institution
(SPV)
Joint venture
agreement
BOT+TOT
Pochuan Water Investment
Equity ratio 90%
Co. Ltd
Mianzhu Integration of Water

Supply and Drainage Project
Fig. 2. The investment and financing structure
government authorized Housing and Urban-Rural Construction Agency as imple-

menting agency to supervise and manage the implementation of project. The
project intended to adopt BOT & TOT mode and complete government pro-
curement through competitive consultation on February 5, 2016. Eventually,
Pochuan Water (SOEs) was selected, and then the two sides formally signed on
June 14, 2016. Pochuan Water is a joint-stock subsidiary companies of Sichuan
Development Equity Investment and Fund Management Co. Ltd. and Poten
Environment Group Co. Ltd. The investment and financing structure is shown
in Fig. 2.
4.1 Risk Identification List Based on the VFM Evaluation
According to Guidelines for VFM Evaluations of PPP (for Trial Implementa-

tion) (Cai Jin [2015] No. 167), the VFM evaluation meeting was convened. The
risk list identified by the participating experts is shown in Table 3.
Take the tax regulation change for example, this section will verify the appli-
cation of fuzzy comprehensive evaluation method in the risk allocation decisions
between Mianzhu government and Pochuan Water.
4.2 Determine the Subordinated Vectors of Mianzhu Government

and Pochuan Water
According to the evaluation score from 30 experts (project management, engi-

neering, finance, assets evaluation, industry, water industry, law, registered
accounting field), the subordinated vectors of Mianzhu government and Pochuan
Water are R1 , R2 . It shows that the comprehensive assessment about the ability
to control and undertake risk, incentive mechanism, risk preference for Mianzhu
government and Pochuan Water according to Mianzhu Integration of Water
Table 3. Risk identification list of Mianzhu Integration of Water Supply and Drainage
project
Risk category Major risks Developing stage

Financial risk Capital availability Construction
Credit risk in financial Construction
institutions
Exchange rate risks Construction and Operation
Interest rate risk Construction and Operation
Inflation Risk Operation
Refinancing risk Operation
liquidity risk Construction and Operation
Design risk Design changeThe Construction
government put forward
Design changeThe SOEs Construction
put forward
Construction risk Geological exploration Construction
risk
Low process/technology Construction
Duration delay Construction
Cost overrun Construction
Not timely access to Construction
labor/equipment
Acquire land Construction
Project quality Construction
Construction safety Construction
Damaged environment Construction
and cultural relics
Operational risk Operation quality Operation
Operation safety Operation
Payment delay Operation
Water quantity and Operation
water quality change risk
Law and policy National law/policy Full life-cycle
changes
Tax regulation change National general tax or Operation
tax law applicability
Political risks Approval delay Construction
Transfer risk Not meet the Transfer
requirements of handover
Natural environmental risk Force majeure Full life-cycle
Residual risk Not considering the Full life-cycle
unknown risk
1568 C. Chen et al.
Supply and Drainage project. R1 , R2 are shown below.

⎡ ⎤ ⎡ ⎤
0.50 0.23 0.20 0.04 0.03 0.29 0.34 0.25 0.12 0.01
⎢ 0.15 0.42 0.24 0.13 0.06 ⎥ ⎢ 0.02 0.10 0.37 0.34 0.17 ⎥
⎢ ⎥ ⎢ ⎥
⎢ 0.01 0.03 0.26 0.40 0.31 ⎥ ⎢ 0.00 0.05 0.44 0.34 0.17 ⎥
⎢ ⎥ ⎢ ⎥
⎢ 0.59 0.32 0.03 0.03 0.03 ⎥ ⎢ 0.05 0.20 0.27 0.24 0.24 ⎥
⎢ ⎥ ⎢ ⎥
⎢ 0.01 0.01 0.22 0.50 0.26 ⎥ ⎢ ⎥
R1 = ⎢ ⎥ , R2 = ⎢ 0.07 0.12 0.49 0.25 0.07 ⎥ .
⎢ 0.08 0.04 0.53 0.26 0.09 ⎥ ⎢ 0.01 0.05 0.49 0.32 0.14 ⎥
⎢ ⎥ ⎢ ⎥
⎢ 0.02 0.14 0.37 0.34 0.14 ⎥ ⎢ 0.07 0.10 0.51 0.25 0.07 ⎥
⎢ ⎥ ⎢ ⎥
⎢ 0.02 0.01 0.53 0.40 0.04 ⎥ ⎢ 0.01 0.07 0.59 0.24 0.10 ⎥
⎢ ⎥ ⎢ ⎥
⎣ 0.06 0.13 0.52 0.23 0.06 ⎦ ⎣ 0.01 0.05 0.65 0.22 0.07 ⎦
0.33 0.27 0.14 0.17 0.09 0.08 0.21 0.37 0.21 0.13
4.3 Determine the Index Weight Based on Entropy Coefficient

Method
According to the entropy coefficient method, the index weight is calculated fol-
lowing the step 4 in the above chapters.
(1) For Mianzhu government
⎡ ⎤
−0.69 −1.47 −1.61 −3.22 −3.51
⎢ −1.90 −0.87 −1.43 −2.04 −2.81 ⎥
⎢ ⎥
⎢ −5.30 −3.51 −1.37 −0.92 −1.17 ⎥
⎢ ⎥
⎢ −0.53 −1.14 −3.51 −3.51 −3.51 ⎥
⎢ ⎥
⎢ −4.31 −4.27 −1.54 −0.69 −1.36 ⎥
InRij = ⎢⎢ ⎥,
⎥
⎢ −2.53 −3.22 −0.63 −1.35 −2.41 ⎥
⎢ −4.20 −2.00 −0.99 −1.08 −1.97 ⎥
⎢ ⎥
⎢ −3.91 −4.61 −0.63 −0.92 −3.22 ⎥
⎢ ⎥
⎣ −2.81 −2.04 −0.65 −1.47 −2.81 ⎦
−1.11 −1.31 −1.97 −1.77 −2.41
⎡ ⎤
−0.35 −0.34 −0.32 −0.13 −0.11
⎢ −0.28 −0.36 −0.34 −0.27 −0.17 ⎥
⎢ ⎥
⎢ −0.03 −0.11 −0.35 −0.37 −0.36 ⎥
⎢ ⎥
⎢ −0.31 −0.36 −0.11 −0.11 −0.11 ⎥
⎢ ⎥
⎢ −0.06 −0.06 −0.33 −0.35 −0.35 ⎥
Rij • InRij = ⎢ ⎢ ⎥,
⎥
⎢ −0.20 −0.13 −0.34 −0.35 −0.22 ⎥
⎢ −0.06 −0.27 −0.37 −0.37 −0.28 ⎥
⎢ ⎥
⎢ −0.08 −0.05 −0.34 −0.37 −0.13 ⎥
⎢ ⎥
⎣ −0.17 −0.27 −0.34 −0.34 −0.17 ⎦
−0.37 −0.35 −0.28 −0.30 −0.22
m
1
ei = − Rij InRij = 0.77 0.89 0.75 0.62 0.71 0.77 0.83 0.59 0.80 0.94 ,
Inm j=1

hi = 1 − ei = 0.23 0.11 0.25 0.38 0.29 0.23 0.17 0.41 0.20 0.06 ,
hi
wi = n = 0.10 0.05 0.11 0.16 0.12 0.10 0.07 0.17 0.09 0.03 .
i=1 hi
(2) For Pochuan

⎡ Water ⎤
−1.24 −1.08 −1.39 −2.16 −5.30
⎢ −3.91 −2.30 −0.99 −1.08 −1.77 ⎥
⎢ ⎥
⎢ −6.91 −3.02 −0.82 −1.08 −1.77 ⎥
⎢ ⎥
⎢ −3.00 −1.61 −1.31 −1.43 −1.43 ⎥
⎢ ⎥
⎢ −2.66 −2.12 −0.71 −1.39 −2.66 ⎥
InRij = ⎢⎢ ⎥,
⎥
⎢ −5.30 −3.10 −0.71 −1.14 −1.97 ⎥
⎢ −2.66 −2.30 −0.67 −1.39 −2.66 ⎥
⎢ ⎥
⎢ −5.30 −2.66 −0.54 −1.43 −2.30 ⎥
⎢ ⎥
⎣ −4.61 −3.00 −0.43 −1.51 −2.66 ⎦
−2.53 −1.56 −0.99 −1.56 −2.04
⎡ ⎤
−0.36 −0.37 −0.35 −0.25 −0.03
⎢ −0.08 −0.23 −0.37 −0.37 −0.30 ⎥
⎢ ⎥
⎢ −0.01 −0.15 −0.36 −0.37 −0.30 ⎥
⎢ ⎥
⎢ −0.15 −0.32 −0.35 −0.34 −0.34 ⎥
⎢ ⎥
⎢ −0.19 −0.25 −0.35 −0.35 −0.19 ⎥
Rij InRij = ⎢⎢ −0.03 −0.14 −0.35 −0.36 −0.28 ⎥ ,
⎥
⎢ ⎥
⎢ −0.19 −0.23 −0.34 −0.35 −0.19 ⎥
⎢ ⎥
⎢ −0.03 −0.19 −0.31 −0.34 −0.23 ⎥
⎢ ⎥
⎣ −0.05 −0.15 −0.28 −0.33 −0.19 ⎦
−0.20 −0.33 −0.37 −0.33 −0.27
m
1
ei = − Rij InRij = 0.84 0.84 0.74 0.94 0.82 0.72 0.80 0.68 0.62 0.93 ,
Inm j=1

hi = 1 − ei = 0.16 0.16 0.26 0.06 0.18 0.28 0.20 0.32 0.38 0.07 ,
hi
wi = n = 0.08 0.08 0.13 0.03 0.09 0.14 0.09 0.15 0.18 0.04 .
i=1 hi
Comprehensively considering the Mianzhu government and Pochuan Waters’

information entropy evaluated by 30 experts, finally it is concluded that the
weight of each index value are calculated. W = [0.09 0.06 0.12 0.10
0.10 0.12 0.08 0.16 0.14 0.03].
4.4 Calculate the Comprehensive Evaluation Vector

(Comprehensive Membership)
Based on the above calculation, using fuzzy operator can calculate the compre-
hensive evaluation vector for the Mianzhu government and Pochuan Water, and
the results are shown below.
1570 C. Chen et al.
B1 = W ◦ R1
⎡ ⎤
0.50 0.23 0.20 0.04 0.03
⎢ 0.15 0.42 0.24 0.13 0.06 ⎥
⎢ ⎥
⎢ 0.01 0.03 0.26 0.40 0.31 ⎥
⎢ ⎥
⎢ 0.59 0.32 0.03 0.03 0.03 ⎥
⎢ ⎥
⎢ 0.01 0.01 0.22 0.50 0.26 ⎥
= 0.09 0.06 0.12 0.10 0.10 0.12 0.08 0.16 0.14 0.03 ◦ ⎢
⎢ 0.08
⎥
⎢ 0.04 0.53 0.26 0.09 ⎥
⎥
⎢ 0.02 0.14 0.37 0.34 0.14 ⎥
⎢ ⎥
⎢ 0.02 0.01 0.53 0.40 0.04 ⎥
⎢ ⎥
⎣ 0.06 0.13 0.52 0.23 0.06 ⎦
0.33 0.27 0.14 0.17 0.09

= 0.15 0.13 0.34 0.27 0.11 ,
B 2 = W ◦ R2
⎡ ⎤
0.29 0.34 0.25 0.12 0.01
⎢ 0.02 0.10 0.37 0.34 0.17 ⎥
⎢ ⎥
⎢ 0.00 0.05 0.44 0.34 0.17 ⎥
⎢ ⎥
⎢ 0.05 0.20 0.27 0.24 0.24 ⎥
⎢ ⎥
⎢ 0.07 0.12 0.49 0.25 0.07 ⎥
= 0.09 0.06 0.12 0.10 0.10 0.12 0.08 0.16 0.14 0.03 ◦ ⎢
⎢ 0.01
⎥
⎢ 0.05 0.49 0.32 0.14 ⎥
⎥
⎢ 0.07 0.10 0.51 0.25 0.07 ⎥
⎢ ⎥
⎢ 0.01 0.07 0.59 0.24 0.10 ⎥
⎢ ⎥
⎣ 0.01 0.05 0.65 0.22 0.07 ⎦
0.08 0.21 0.37 0.21 0.13

= 0.05 0.11 0.47 0.25 0.11 .
4.5 Determine the Sharing Subject and Sharing Proportion for Tax
Regulation Change Risk
Then the section normalize the index system, which makes the comprehensive
evaluation vector B = (b1 , b2 , b3 , · · · , bm ) converted to comprehensive score S.
T
S1 = B1 × V = 0.15 0.13 0.34 0.27 0.11 × 0.15 0.25 0.45 0.65 0.85 = 0.48080,
T
S2 = B2 × V = 0.05 0.11 0.47 0.25 0.11 × 0.15 0.25 0.45 0.65 0.85 = 0.50897.
The Mianzhu government should bear the proportion:

S1
Q1 = × 100% = 48.577%.
S1 + S2
The Pochuan Water should bear the proportion:
S2
Q2 = × 100% = 51.423%.
S1 + S2
The example shows that the risk of tax regulation change should be shared
by the Mianzhu government and Pochuan Water. The shared proportions were
48.577% and 51.423%. The Tax regulation change is belong to macroeconomic
level risk, and the local government or superior government, central government
can influence macroeconomic conditions. But the allocation is involved in local

government and social capital, the local government is unable to fully control the
occurrence and adverse consequences adverse consequences of tax risk, therefore
the tax regulation risk should be shared.
5 Conclusion
Risk allocation is the key factor in the success of infrastructure PPP projects,
and the reasonable risk allocation can effectively reduce the risk level. This
paper explores the effective approach of risk allocation form the perspectives
of the SOEs, and builds the allocation model based on fuzzy comprehensive
evaluation and entropy coefficient method. In addition, the paper demonstrates
the objectivity, rationality and feasibility of the allocation model, which provides
the scientific methods and guidance for SOEs to participate in PPP projects,
further deepening the reform of state-owned enterprises and gradually making
the economic grows change from the factor-driven into the innovation-driven,
promoting the transformation of public ownership economy.
Acknowledgements. This paper is attributed to the financial support from the

National Natural Science Foundation of China (NSFC) under the project No. 71371131.
References
1. Abednego MP, Ogunlana SO (2006) Good project governance for proper risk allo-
cation in public-private partnerships in Indonesia. Int J Project Manage 24(7):622–
634
2. Bing L, Akintoye A et al (2005) The allocation of risk in PPP/PFI construction
projects in the UK. Int J Project Manage 23(1):25–35
3. China Public Private Partnerships Center (2016) Quarterly report on the project
library of the national PPP integrated information platform. http://www.cpppc.
org/en/Quarterly/4007.jhtml
4. Chan APC, Lam PTI (2010) Critical success factors for PPPs in infrastructure
developments: Chinese perspective. J Constr Eng Manage 136(5):484–494
5. Chang C (2013) A critical review of the application of TCE in the interpretation
of risk allocation in PPP contracts. Constr Manage Econ 31(2):99–103
6. Chen C, Chen P, Wang Q (2017) Comparing the efficiency of public-private part-
nerships with the traditional procurement: based on the Chengdu No. 6 water plant
B. In: Proceedings of the tenth international conference on management science
and engineering management, pp 487–501
7. Chou JS, Pramudawardhani D (2015) Cross-country comparisons of key drivers,
critical success factors and risk allocation for public-private partnership projects.
Int J Project Manage 33(5):1136–1150
8. Deng X, Qiming LI et al (2008) Summary and application of the principles of risk
allocation in PPP model. Constr Econ 09:32–35 (in Chinese)
9. Grimsey D, Lewis MK (2002) Evaluating the risks of public private partnerships
for infrastructure projects. Int J Project Manage 20(2):107–118
1572 C. Chen et al.
10. Jian G (2013) The risk allocation strategy research of PPP project in highway
traffic infrastructure. Manage Rev 25(7):11–19
11. Jin XH (2009) Determinants of efficient risk allocation in privately financed public
infrastructure projects in Australia. J Constr Eng Manage 136(2):138–150
12. Jin XH, Zhang G (2011) Modelling optimal risk allocation in PPP projects using
artificial neural networks. Int J Project Manage 29(5):591–603
13. Ke Y, Wang SQ et al (2010) Preferred risk allocation in China’s public-private
partnership (PPP) projects. Int J Project Manage 28(5):482–492
14. Liu S (2016) Promote the reform of state-owned enterprises in the PPP way. Mod
SOE Res 9:78–81 (in Chinses)
15. The State Council (2015) China urges SOE modernization through mixed own-
ership reform. http://english.gov.cn/policies/latest releases/2015/09/24/content
281475197422388.htm
16. Zhang S, Chan APC et al (2016) Critical review on PPP research-a search from
the Chinese and international journals. Int J Project Manage 34(4):597–612
17. Zou PXW, Zhang G, Wang J (2007) Understanding the key risks in construction
projects in China. Int J Project Manage 25(6):601–614
An Exploratory Case Study of the Mature
Enterprise’s Corporate Brand Building
Based on Strategic Perspective
Yuan Yuan, Jing Xu, Liming Zhang(B) , and Rui Zhou

zhangliming@scu.edu.cn
Abstract. Corporate brand plays an increasingly important role in

intensive competition. Therefore, further research needs to be done on the
relationship between strategic action and corporate brand. Particularly
for mature enterprises, how to build corporate brand through strategic
action is an important issue in this research field. Based on the case stud-
ies of Jinjiang Hotel Industrial Co., Ltd. and Luzhou Laojiao Co., Ltd.,
this paper analyzes the corporate brand management of mature man-
ufacturing and service-oriented enterprises, giving the conclusion that
strategic action will contribute to the corporate brand management and
the brand construction of a mature enterprise is the consequence of its
strategic action, since enterprises can manage corporate brand building
by improving brand awareness, brand structure and the building path to
construct their own corporate brands.
Keywords: Corporate brand · Strategic action · Mature enterprise
1 Introduction
Over the past two decades, corporate branding has been advocated by scholars
as an effective way for companies to build competitive advantage [4], it has
become a key strategic resource of enterprise [9]. In the contemporary global
competitive environment, products and services typically become more similar,
which makes it difficult for consumers to differentiate offerings coming from
different enterprises. Consequently, promoting entire companies as a brand has
become an efficient approach to create differentiation [8]. Corporate brand has
significant value in consumers concern, product supporting, core communication,
employee incentive [13]. And it is capable of helping enterprise to reduce cost,
giving customers a sense of belonging, building consensus in stakeholders [11].
Therefore, the study on corporate brand from the perspective of strategy is
significant.
To our knowledge, existing research about corporate brand mainly focuses on
corporate brand itself [3,5,10,12,14], or on the marketing area, which explores

1574 Y. Yuan et al.
how corporate brand affects the consumers’ reviews, attitudes or purchasing ten-
dency [3,5,15,16]. Besides, the relationships among corporate brand and orga-
nization identity, image, reputation, etc. [7,12], the influence factors of corpo-
rate brand, and the effect of corporate brand on other aspects [7,8,16,17] are
also research hotspots now. But there is little research on the inner relation-
ship between corporate brand and enterprise strategy. Especially for mature
enterprises, as to promote the corporate brand, the ways of facing the competi-
tion from start-ups and releasing from past traces become essential issues to be
resolved.
So, this research plans to carry on an exploratory case study of two mature
company cases, to analyze the role and impact of strategic actions taken in the
enterprise development process on corporate brand rebuilding.
2 Literature Review
The research on corporate brand sprang up in 1930s’. To some extent, corporate
brand is defined as an organization from the whole, which is the platform of all
corporate brands and carrier used for promoting various differential services [1].
However, Knox and Bickerton [12] thought corporate brand is a manifestation
of unique business mode of an organization in vision, diffusion and behavior.
Harch and Schultz [11] defined corporate brand as brand umbrella covering all
corporate products and plate brilliance for product.
2.1 Elements of Corporate Brand
Aaker [1] concluded the elements of corporate brand are historical foundation,
corporate performance, employee performance, value view and priority item,
local and global orientation, social resonance. Hatch and Schultz [9] proposed
that corporate brand is made up of vision, culture and image. Vision comes from
leadership consciousness of senior leaders, culture comes from employees, forming
internal cognition and cohesion, and image is the evaluation of external public.
Balmer [6,10] pointed out the components of corporate brand include brand
vision, culture, location, personal feature, public relation and information.
2.2 Brand Building and Maintenance
Brand building and maintenance work need to design systematic and clear guid-
ance plan from organizational strategic level. Philip Kolter [2] thought that enter-
prise should recognize advantages and disadvantages of brand regularly, maintain
the development of brand constantly and propose four aspects of brand construc-
tion which are positioning, name selection, holder’s decision and development
strategy. With the gradual increasing of competition, corporation pays more
attention to seek for differential competition of brand through unique emotional
experience rather than function or characteristic in brand construction [10].
Mature Enterprise’s Corporate Brand Building 1575
2.3 Management and Application of Corporate Brand
It’s clear that theoretical research on definition, composition, role and feature,
etc. of corporate brand at present has made some achievement, but corporate
brand management research is somewhat lacking in guiding practice. Mukher-
jee and Balmer [14] clearly pointed out that the current articles about corpo-
rate brand are too ideal, they lack empirical data support and related mea-
sure description for possible conditions. The news shows that many scholars are
developing related research on corporate branding, for example, Knox and Bick-
erton [12] pointed out corporate brand management is suggested to consider
four aspects as vision management, culture management, image management
and competition management. Hatch and Schultz [11] proposed corporate brand
management is the process of eliminating the difference of corporate vision, cul-
ture and image.
In conclusion, not only does corporate brand has profound relationship with
corporate strategic action, but it has close relation with corporate develop-
ment stage and business type. Strategic action is strategic operation affecting
future development direction of corporation and the result of executing corpo-
rate strategic plan, which reflects corporate strategic choice. However, strategic
choice will promote the formation and change of corporate brand directly. Cor-
porate brand can concentrate and abstract corporate vision, culture, value view,
behavior and the expectation of stakeholders, etc. as a kind of contract relation-
ship through certain strategic action so that corporation and stakeholders can
establish stable ‘emotion connection’. For many emerging enterprises, in order
to obtain competitive advantage through corporate brand management in grow-
ing period, strategic actions with great influence on corporate brand must be
adopted.
3 Research Design
Based on the current corporate brand theoretical achievement, this research

adopts cross-case comparison research method to carry out exploratory research
on brand management method of mature enterprises. After analyzing corpo-
rate brand management activities for mature enterprises with different business
scope, the development circumstances of mature enterprises’ corporate brand
presented with related strategic actions can be drawn, and then exploring ways
of mature enterprises’ corporate brand.
To make sure the representativeness and comparability of selected sample, sam-

ple enterprises should be in different industries with differentiated products and
management operation mode, as a result, the mature manufacturing enterprise
and service-oriented enterprise are selected as samples for exploratory cross-case
comparison study.
1576 Y. Yuan et al.
One case is Luzhou Laojiao, a mature manufacturing enterprise founded in

1982, which is a large brewing backbone group that developed from 36 ancient
workshops in Ming and Qing dynasties, has long been history of working on the
brewing of organic liquor and commits to advocate healthy lifestyle. Luzhou Lao-
jiao received various prizes from its establishment, and its ‘Luzhou’ registered
trademark is one of ten most famous trademarks in first Chinese trademarks
election. Past brilliant achievement usually makes the company lose the direc-
tion, however Luzhou Laojiao, through successful corporate brand management
and continuous strategic revolution, broke through the dilemma and returned as
the leader in Chinese wine industry.
Another case is Jinjiang Hotel, a mature Service-oriented enterprise founded
in 1958, which was assigned personally “Jinjiang” as the name by Marshal Zhu
De and Marshal Chen Yi when it was completed, honored as the first five-star
business and tourist hotel in Southwest China. The enterprise’s age sometimes is
the constraint affecting its development. Jinjiang hotel adopted series of strategic
actions like strategic alliance to build and cultivate corporate brand when facing
the dilemma. And brand of Jinjiang Hotel eventually becomes the industry leader
in Southwestern China.
Basic situation of sample enterprises is as shown in Table 1.
Table 1. Basic information of sample enterprises
Name of enterprises Luzhou Laojiao Co., Ltd. Jinjiang Hotel Co.,

Ltd.
Time of foundation In 1982 In 1958
Main business Wine products Hotel services
Nature of ownership State-owned enterprises State-owned
enterprises
Scale of enterprise Owns more than 5000 employees and A five-star hotel,
total assets of 3 billion RMB, annual belongs to the
output of Luzhou series of liquor Sichuan Tourism
50,000 tons, brand valuation is 345.84 Development Group,
billion RMB owns more than 700
rooms
Business scope Mainland China Mainland China
3.2 Data Collection

This research explores dynamic interaction of corporate strategic action and
corporate brand through longitudinal case study. Main data collection starts
from 2013, and ends in October 2015. During this period, various mixed data
collection methods were adopted by starting from the purpose of explorative
research, including Personnel interview, secondary data collection and on-site
observation.
1. Personnel interview. Adopting half structural interview manner; interviews

are enterprise management.
2. Documents collection. Collection object are sample enterprises’ brief report,
internal management handbook, product menu etc.
3. On-site observation. Observation objects are sample enterprises’ headquar-
ters, manufacturing base, sales and service places etc.
4 Case Analysis and Discussion

4.1 Case Analysis
(1) Luzhou Laojiao Co., Ltd.

In the 1990s, rapid development of China’s economy and improvement of
people’s material life level stimulated the consumer’s brand awareness. Espe-
cially in liquor-making industry, the high-end brand is more profitable. Maotai,
Wuliangye, Jiannanchun have Stopped selling goods in nationwide and raised
product prices to develop in the high-end market, while Luzhou Laojiao per-
sisted in the road of mass brand and missed the development opportunities.
When Luzhou Laojiao realized failures in the high-end market, it tried to
sale its high-end liquor products, but the brand unlimited expansion followed let
Luzhou Laojiao face bigger difficulties. For making up previous wrong doings,
Luzhou Laojiao adopted series of strategic actions to manage its corporate brand.
It cut down 49 sub-brands and reserved core brand through sliming brand, clean-
ing up the market sorting out product structure. After 2 years, Luzhou Laojiao
finished long-term brand development plans, and formulated reasonable brand
strategy architecture including two brands as well as three series, shown in Fig. 1.
It improved brand identification and consumer recognition. After that, to estab-
lish brand image and improve brand value, Luzhou Laojiao decided to stop selling
goods nationwide and raise price. Simultaneously, Luzhou Laojiao started stock
reforming and coordinated the relationship between the stakeholders. Addition-
ally, Luzhou Laojiao integrated channel strengths through rewarding retailers
and established brand image through actively participating social campaign.
Luzhou Laojiao rebecame a leading enterprise in Chinese liquor industry with
series of strategic planning and actions.
Fig. 1. Brand strategy architecture of Luzhou Laojiao

1578 Y. Yuan et al.
Corporate brand development process based on strategic actions for Luzhou

Laojiao is as shown in Table 2.
Table 2. Corporate brand development process based on strategic actions for Luzhou
Laojiao
Period and enterprise Important strategic Corporate brand

status actions affecting the development situation
corporate brand
development
1980s Glorious history, Establishment and 1. Initially establishing
high brand value and development under the corporate brand
market acceptance government’s recognition
and protection
2. Strengthening brand
value through old pits
1980s–1990s Adhering to the mass Brand image was damaged
brand Indulging in the
past
1997–2003 Inappropriate 1. Overstretched 1. Brand excessive
action under fear sub-brand extension to destroy
high-grade brand image of
“Luzhou Laojiao”
2. Taking the brand of 2. Marketing model of
“Luzhou Laojiao” as the paying money to OEM
intangible assets damaged the brand image
2002 - Brand renewal, 1. Sliming brand, cleaning 1. Corporate brand
value return up the market, sorting out achieving value return,
product structure Luzhou Laojiao embarking
on the road of corporate
brand management
2. Stopping selling goods 2. Rebuilding brand image
in nationwide in the minds of dealers
and consumers
3. Government approving
listed companies to carry
out tradable share reform
4. Non-public offering of
no more than 3000 million
shares
(2) Jinjiang Hotel Co., Ltd.

1961–1979, as a hostel owned by government, Jinjiang Hotel didn’t have author-
ity or consciousness to manage its corporate brand. 1979–2001, because of the
reform and opening-up policy, Jinjiang Hotel went into modern enterprise admin-
istration phase of operating independence and self-financing. Since then, the time
gave Jinjiang Hotel a great new identity. In 1995, Jinjiang Hotel, given the title
of five-star hotel by National Tourism Administration, became the first five-star
hotel in southeast of China. But with flourishing of Chengdu tourist market and
increasing number of new hotels entering the market, fierce competition caused
Jinjiang Hotel the development to receive the restriction again.
From 2001, the management team of Jinjiang Hotel decided to use advantage
of carrying out the reshaping of the brand rebuilding work. It used “building a
national top brand, being the best one of hotel industry” as vision, also offering
services and products for high-end business market. And it redesigned brand
logo, as well as inviting famous designer to redesign and redecoration the whole
hotel. Then Jinjiang Hotel started building corporate culture and educational
training system matched to corporate brand, and joined hands with strong brand
enterprise. Through more than ten years of development, Jinjiang Hotel has
successfully rebuilt its corporate brand through step-by-step strategic action.
Corporate brand development process based on strategic actions for Jinjiang
Hotel is as shown in Table 3.
4.2 Case Discussion

According to the exploration of corporate brand management process regarding
sample enterprises above, it can be found that the difference of business types
can substantially make enterprises chose different strategic actions to build and
manage corporate brand.
(1) Different extent of demand of corporate brand. Service-oriented enter-
prises possess more demand than manufacturing enterprises, since corporate
brand is essential even only component of brand equity for service-oriented enter-
prises however, manufacturing enterprises have to concern product brand and
related issues between product brand and corporate brand. Sample enterprises
Jinjiang hotel relies on corporate brand to get involved in the competition, thus
ways of adopting strategic actions to promote corporate brand are essential.
Luzhou Laojiao has various product brand, it is significant to consider how to
deal with the relationship between product brand and corporate brand through
strategic actions.
(2) The difference of the essence of strategic choice. Service-oriented enter-
prises are keen on providing services with high quality to make success. However,
manufacturing enterprises have to rely on innovation and R&D, high quality
product and extraordinary manufacture to form strong corporate brand, which
are implementation significance of strategic actions. Jinjiang hotel as sample
enterprise has more concerns about services, and Luzhou Laojiao focuses more
on product.
(3) Difference of implementation focus regarding organizational capability.
Service-oriented enterprises tend to utilize intangible assets management to real-
ize management goals of corporate brand. Manufacturing enterprises are sug-
gested to improve tangible assets’ operational capability to realize management
1580 Y. Yuan et al.
Table 3. Corporate brand development process based on strategic actions for Jinjiang
Hotel
Period and Important strategic actions affecting Corporate brand the corporate brand
enterprise development situation development
status
1961–1979 Building and opening led by government 1. Initially establishing corporate brand
Proud of
“Yellow
Vest”
2. Political overtonesbuilding strong
brand image and basis
3. Official hostel, lack of awareness of
business operations
1979–2001 1. Getting rid of government plans, 1. Reforming corporate brand, laying the
Develop- becoming self-managed modern foundation of systematic brand
ment of enterprise management project
new
identity
2. Establishing enterprise management, 2. Learning operation and management
system and reforming official human resource system
3. Initiatingunion and establishing hotel 3. Becoming top1 hotel in southwestern
management company China
4. Carrying out hardware upgrading 4. Having some management
infrastructure
5. Regulating management 5. Difficult to resist external competition
6. Awarded as the title of five-star hotel
7. Intense competition and inner defect
2001 - 1. Selecting brand development strategy 1. Carrying out series of ways to rebuild
Vigorous corporate brand
of brand
develop-
ment
2. Establishing vision, defining position, 2. Discovering history heritage,
designing corporate logo increasing the essence of corporate brand
3. Upgrading hardware facilities, 3. Initially realizing the vision of
improving product and service quality “Nation’s top and Industry leaders’
brand”
4. Embedding ideas through culture to 4. Becoming the leading enterprise in
prove service quality Sichuan tourism and hotel industry
5. Starting service personnel training 5. Having ability to compete with
system international brands
6. Focusing on opinions of leaders
7. Group expansion
8. Extending the brand network
9. Associating with strong brand
enterprises
10. Rechanging management system
goals of corporate brand. Sample enterprises Jinjiang hotel awakes consumers’

memory through digging enterprises historical heritage and reforming image of
corporate brand. Luzhou Laojiao improved the integrated value of corporate
brand through re-positioning and forming product brand.
In the meantime, we found that from the points of strategic views, mature
enterprises are suggested to consider tactics below to break through the dilemma
of corporate brand management.
(1) Building the innovation expectation of corporate brand. And the con-
struction of corporate brand has to be adjusted according to the changes of
market conditions. At the first place, two sample enterprises immersed into the
past distinctive achievement without improvement, which led corporate brand
was lack of appropriate adjustment and revolution. Afterwards, in order to break
through management delimma, Luzhou Laojiao conducted series of strategic
actions to maintain its industry leading position.
(2) Knowing the methods of corporate brand system construction. Although
sample enterprises were able to comprehend the significance of building corpo-
rate brand, but they still came across the crisis without adopted proper strategic
actions. Luzhou Laojiao ignored the intrinsic connection between product brand
and corporate brand in the process of exploring corporate brand management,
so that the abuse of product brand weakened the value of corporate brand. How-
ever, through series of implementation, the corporate brand system was rebuilt.
Therefore mature enterprises should clarify the position and role of corporate
brand in the enterprise brand construction at the prior place. Designing the net-
work structure of corporate brand, and adopting strategic actions, like putting
the corporate brand as the core, to adjust brand structure is helpful for the
strategy planning of corporate brand construction, which can the relationship
and influence manner between corporate brand and product brand.
(3) Exploring the path of corporate brand rule development. Firstly, enter-
prises in different period of lifecycle, have different conditions of resources and
capability and face different strategic choice, therefore management ways suit-
ing corporate brand are different. And the difference of enterprise scale, the
competition position, business type and operation ways can lead changes of cor-
porate brand management ways. Enterprises are considered to constantly adjust
corporate brand development path in accordance with development phases and
strategic planning.
5 Conclusions
The research is based on strategic views, and integrates the strategic manage-
ment and brand management theory. Besides, the research adopts the methods
of vertical case study, and analyses ways of building corporate brand between
mature manufacturing enterprises and service-oriented enterprises. From the
process of building corporate brand for sample enterprises, implementing series
of strategic actions which benefit prospects of future corporate brand develop-
ment circumstance and direction, can boost the corporate brand construction
of mature enterprises. Moreover, mature enterprises can adopt certain tactics
to break through the dilemma of corporate brand management, thus realizing
the corporate brand construction. It can be concluded that corporate brand
construction and enterprises’ strategic actions are correlated, mature enterprises
brand construction is the consequence of related strategic actions, in the con-
trast, corporate brand construction has the certain impact on enterprises’ strate-
gic actions. The further discussion is placed regarding detailed roles, influence
ways and path on enterprises’ strategic actions from corporate brand.
1582 Y. Yuan et al.
Acknowledgements. We thank the support of the Projects of the National Social

Science Foundation of China (16AGL003), Social Science Planning Annual Project of
Sichuan Province (SC16C011), Soft science research plan project of Sichuan province
(2017ZR0098, 2015ZR0091), Soft science research project of Chengdu city (2015-RK00-
of Sichuan University (sksz201603, skqy201623, skqy201209), Projects of the Center for
systems science and enterprise development (xq15B11).
References
1. Aaker DA (2006) Brand portfolio strategy. Strateg Dir 22(10):468–468
2. Armstrong G (2009) Marketing: an introduction. Pearson Education, New Jersey
3. Balmer J (2017) Advances in corporate brand, corporate heritage, corporate iden-
tity and corporate marketing scholarship. Emerald
4. Balmer JMT (2012) Strategic corporate brand alignment: perspectives from iden-
tity based views of corporate brands. Eur J Mark 46(7/8):1064–1092
5. Balmer JMT, Abratt R, Kleyn N (2016) Corporate brands and corporate market-
ing: emerging trends in the big five eco-system. J Brand Manag 23(1):3–7
6. Balmer JMT, Powell SM et al (2016) Advances in corporate branding. Palgrave
Macmillan, UK
7. Buil I, Catalán S, Martı́nez E (2016) The importance of corporate brand identity
in business management: an application to the uk banking sector. Bus Res Q
19(1):3–12
8. Chang A, Chiang HH, Han TS (2015) Investigating the dual-route effects of cor-
porate branding on brand equity. Asia Pac Manag Rev 5(3):120–129
9. Dinnie K (2009) Taking brand initiative: how companies can align strategy, culture,
and identity through corporate branding. J Brand Manag 16(7):496–498
10. Harris F, Chernatony LD (2001) Corporate branding and corporate brand perfor-
mance. Eur J Mark 35(3/4):441–456
11. Hatch MJ, Schultz M (2001) Are the strategic stars aligned for your corporate
brand? Harvard Bus Rev 79(2):128
12. Knox S, Bickerton D (2003) The six conventions of corporate branding. Eur J Mark
37(7/8):998–1016
13. Lewis S (2000) Let’s get this in perspective. Unpublished presentation given at the
Confederation of British Industry, Branding and Brand Identity Seminar, Bradford
University School of Management
14. Mukherjee A, Balmer JMT (2008) Preface: New frontiers and perspectives in cor-
porate brand management: in search of a theory. Int Stud Manag Organ 37(4):3–20
15. Srivastava K, Sharma NK (2013) Service quality, corporate brand image, and
switching behavior: the mediating role of customer satisfaction and repurchase
intention. Serv Mark Q 34(34):274–291
16. Tu YT, Li ML, Chih HC (2013) An empirical study of corporate brand image,
customer perceived value and satisfaction on loyalty in shoe industry. J Econ Behav
Stud 5:469
17. Voss KE, Mohan M (2016) Corporate brand effects in brand alliances. J Bus Res
69(10):4177–4184
Case Study: Packing and Distribution Logistics
Optimization of Fashion Goods
Young Jae Jang(B) and Shin Woong Sung
Industrial and Systems Engineering, KAIST, Daejeon, Korea

yjang@kaist.ac.kr
Abstract. In this short paper, we describe the development and imple-

mentation of the optimal packing and distribution of fashion products
applied to a leading outdoor fashion brand in South Korea, Kolon Sport
(K/S), through the academic-industry Business Analytics project with
KAIST. We explain the framework of Business Analytics defined for the
project and how modeling and analysis were performed under the frame-
work. The packing and distribution decision involves how to determine
a set of different box configurations and how to distribute these boxes to
stores to meet store demand. The packing and distribution problem was
formulated as a mixed-integer linear program, and a solution approach
was developed to efficiently solve the industry-size problem in a timely
manner. The modeling and analysis were performed with SAS packages.
In particular, the optimization model was implemented with SAS/OR.
The proposed approach was validated with simulation and on-site pilot
test. The simulation showed an improvement in the packing and distri-
bution process of around 15% in terms of meeting the demand in the
stores, and it projected a 5 ∼ 10% improvement in revenue. After the
successful pilot test, the approach was implemented in the internal IT
system of K/S in July 2015. Consequently, the mass-produced items for
the fall/winter season of 2015 were distributed by the system.
1 Introduction
1.1 Kolon Sport
Kolon Sport (K/S) is the signature brand of Kolon Industries, Inc., which was
founded in 1957 in South Korea and operates more than 10 different fashion
brands under its business unit. The total revenue of the Kolon Industries was
around US $5 billion in 2014 [3]. The K/S brand was first established in 1973
and is now one of the top three outdoor fashion brands in Korea, a country with
the world’s second-largest market share in the global outdoor fashion industry,
as of 2014, after the United States. K/S operates more than 250 stores over the
country and designs, manufactures, and distributes more than 4,000 different
types of products in all sales seasons. More information about the brand can be
found at its website (http://us.kolonsport.com/). Note here that K/S refers to
both the brand name and the department unit in Kolon Industries responsible
for K/S operations. We use the term K/S interchangeably for the two definitions.
1584 Y.J. Jang and S.W. Sung
1.2 Kolon-KAIST Business Analytics Project
The Kolon-KAIST Business Analytics project was first initiated in February

2014. The Big Data Analytics Team at Kolon Industries, an internal business
analysis unit in the company that was established in 2013 to analyze massive
amounts of data with advanced algorithms to support business operations and
strategic decisions, was primarily responsible for the operation and management
of the project. The team from the Korea Advanced Institute of Science and Tech-
nology (KAIST), the top science and technological university in South Korea,
was responsible for guidance and for providing the technology for the project.
The authors of this paper are the primary investigators of the project in the
KAIST team.
The project goal was to improve the efficiency of the packing and distribution
operation at K/S with Business Analytics. We defined Business Analytics as the
process of delivering a specific tangible effect on the business operation by ana-
lyzing the data, developing scientific decision-making methods, and implement-
ing methods with advanced IT solutions. Compared with conventional business
analysis in which the final deliverables are forms of reports or documents of
as-is analysis, the defined Business Analytics tried to provide a way to deliver
specific actions and results. In contrast to the conventional business analysis
or data-driven approach, which starts with collecting data and analyzing data
first, we tried a problem-centric top-down approach. We first ask, “What is the
problem?” and then ask, “Can this problem be solved with data and advanced
algorithms and computation?” If the answer is yes to the preceding question,
we try to collect data or create a system to collect the required data. More-
over, we emphasize an algorithm-based decision. To drive an action, “specific”
action should be made. Therefore, optimization modeling and algorithm devel-
opment are emphasized. We also include IT implementation in the definition of
Business Analytics. To deliver sustainable results rather than a one-time as-is
report, implementation of the analysis procedures and algorithm into the com-
pany’s existing IT system is mandatory. This definition of Business Analytics
clearly outlines the job scope of the project. It includes problem identification,
data analysis to find the insight, development of an optimization model and
algorithm to solve the complex problem, and then implementation of the system
to derive the action.
Figure 1 shows the framework of Business Analytics that we constructed for
the project. The framework describes the required techniques and corresponding
departments or organization necessary to support the techniques. As shown, the
framework consists of three different technical layers. The bottom layer com-
prises data collection and data management. The fact-based and data-oriented
approach is the core concept of Business Analytics. Identifying the required
data and effectively managing them were the fundamentals of the project.
The IT department at K/S was responsible for this layer. The next layer
addresses knowledge discovery through the analysis of data. Using data mining
and statistical analysis, we tried to find new insights and knowledge or vali-
date the existing knowledge. This function was done in cooperation with the
Case Study: Packing and Distribution Logistics Optimization 1585
Fig. 1. Business analytics framework for the project
marketing team, business operation team, and various analytics-centric teams

at K/S under the guidance of both the K/S Big Data Analytics and KAIST
teams. The top layer concerns decision making. With the knowledge and insight
gained from the underlying layers, this top layer attempts to support the deci-
sion for a complex problem. Mathematical optimization techniques were used in
this layer. Because K/S did not have an internal team or organization using such
optimization, this layer was the primary responsibility of the KAIST team.
2 Packing and Distribution Problem

To expedite the shipping process and reduce the operational costs of the dis-
tribution, K/S has been using direct delivery of their apparel products from
the manufacturers, which are the third-party OEM supplier to K/S, to the K/S
stores for the last several years. In the past before direct delivery was used, all
of the manufactured products were shipped to the K/S warehouse, unpacked,
sorted, and repacked for distribution to the stores. Given the fact that K/S dis-
tributes more than 1.6 million items in 1,000 different styles during each selling
season, this repacking process required significant time and effort. Moreover, due
to long delays in the repacking process, some stores did not receive products even
after a sales season had already started. Direct delivery provided a significant
benefit to the process. Each manufacturer packs the product with different sizes
based on the size configurations given by K/S and ships their goods direct to
each store. For example, K/S requests the manufacturers to provide two different
types of box configuration such that for product WJ0001, Box 1 should contain
10 small, 30 medium, and 20 large sizes, and Box 2 should contain 3 small, 5
medium, and 4 large sizes. Also, K/S informs the manufacturer of which stores
get how many boxes of each box configuration.
Although direct delivery significantly eliminated inefficiency in terms of deliv-
ery speed, it generated another problem. Because packing is done by third-party
manufacturers, it was not realistic to make different box configurations for every
store for each product, which would cause significant operational costs particu-
larly when the numbers of products and stores are large. Therefore, the number
of different box configurations was restricted to less than 10. Consequently, K/S
should decide how to design the box configurations and determine which con-
figuration and how many boxes are distributed to a store. These decisions are
collectively called packing and distribution in K/S. Tables 1 and 2 show an exam-
ple of packing and distribution decisions. The decisions on the configuration for
each box type and the number of boxes for each type should be made in the
packing and distribution process. In addition, which types and how many boxes
of each type should be sent to each store must also be determined.
Table 1. An example of the packing decision
Configuration type Configuration Number of boxes

Type 1 Box XS: 20, S: 25, M: 40, L: 15, XL: 10 10
Type 2 Box XS: 10, S: 15, M: 20, L: 12, XL: 5 50
··· ··· ···
Type n Box XS: 5, S: 10, M: 12, L: 8, XL: 3 30
Table 2. An example of the distribution decision
Store 1st box · · · mth box

Configuration type Number of boxes · · · Configuration type Number of boxes
Store 1 Type 3 Box 1 · · · Type 4 Box 1
Store 2 Type 4 Box 3 ··· - -
··· ··· ··· ··· ··· ···
Store s Type 2 Box 2 · · · Type 3 Box 1
Because the OEM manufacturers charge packing fees based on the number of
configuration types and the number of boxes shipped to the stores, appropriate
packing and distribution can significantly reduce costs. However, due to the
complexity of the configuration, decisions had been made on an ad-hoc basis
that had caused inefficiency - some stores received more items than needed and
some received less than needed.
3 Project Goal and Scope

The goal of the project was to develop a decision support system to optimally
determine packing and distribution. The project was broken down into the fol-
lowing tasks:
• Demand analysis for each product style and size using the past sales data;
• Optimization modeling and algorithm development;
• Validation of the model and algorithm with simulation;
• Validation of the overall approach with pilot tests;
• Implementation of the approach in K/S’s internal IT system.

The K/S Big Data team took the leading role of managing the overall project.
The KAIST team took a leadership role for the first, second, and third tasks.
The Sales and Operation Team at K/S was responsible for the fourth task, and
the IT Department at K/S was responsible for the final task. The KAIST team
also worked as the technical advisor for tasks four and five.
Note that demand forecasting or evaluation of the quantity of each product
type was not within the scope of the project because of the following reason.
The packing and distribution decisions were needed about one to two months
before each sales season starts. However, decisions on production quantity had
to be made one year before the target sales season. That is, by the time they
were ready for packing and distribution, all of the products would have been
ready for shipment at the manufacturers. Therefore, the product quantities for
the items were given as the input parameters for the box assortment.
Although the product quantities to be distributed were known, the num-
ber of each item to be distributed to each store still needed to be evaluated.
For instance, given that there are 100 medium-size winter jackets (item number
WJ000111) available for the season, the number to distribute to each store must
be determined. We evaluated this quantity based on the past sales data. Specif-
ically, for each product style, we evaluated the percent of sales made by each
store. This sales contribution value from the past data was used to estimate the
distribution quantity for each product. For statistical analysis, we used multiple
regression analysis and a standard method to estimate the future demand.
4 Packing and Distribution Optimization

The core part of the project was the development of the optimization model in
the second task. As mentioned above, in this model, we determine the number
and contents of the box configurations and their allocation to the stores to min-
imize the difference between store demand and the actual distribution quantity
as depicted in Fig. 2. As indicated in the figure, the decision variables in the
optimization model are the configuration of each box and the number of boxes
to be distributed to each store. The demand for each store is evaluated from
Task 1.
The constraints for the model are as follows. First, the types of configura-
tion are limited. The manufacturers do not want to have more than 7 different
types of configurations for each product. Second, the maximum number of boxes
is restricted, as is the capacity of the box. Of course, the maximum capacity
depends on the type of product. A 50 × 50 × 50 cm3 box can hold 10 puffy winter
down jackets at maximum, whereas it can hold 50 summer T-shirts. K/S has data
mapping on the size of each box type and its maximum capacity for each product
type. In addition, the maximum and minimum number of boxes a store receives
are constrained. The last constraint is the upper limit of overstocking for each
store. Note that due to the discrete nature of the problem, it may be impossible
to meet the exact demand for each store. Therefore, some levels of over-shipping
or under-shipping are unavoidable. However, K/S is more concerned with over-

shipping. As a result, we added an over-shipping constraint, which specifies the
upper limit of the allowable over-shipping to each store. The details of the gen-
eral mathematical model are provided in the Appendix. The solution algorithm
was also developed but is not presented in this document. For readers interested
in the specifics of the algorithm, please contact the corresponding author.
Fig. 2. Concept of the box assortment and allocation optimization problem
Similar formulations and optimization models are available in the academic

literature, including [1,2]. However, the details of the models and the scales of
the problem these optimization problems are trying to solve are different. To the
best knowledge of the authors, this is the first work to develop an optimization
model and solution algorithm at an industry-level scale and apply them to an
actual business operation. Note that this paper is not intended for the acad-
emic community, and therefore the academic value of the work and scholarly
implications are not discussed.
5 Validations
We conducted a simulation-based validation of the packing and distribution
process in one sales season to compare the performance of the proposed method
to that of the legacy procedure at K/S. The sales volume, i.e., number of sales,
from the proposed method was approximately 10.16% better than that with the
legacy method. We also conducted a pilot test of our approach by distributing a
number of actual products for the summer season of 2015, and the result shows
that compared to the legacy method of K/S, the inventory was distributed well
by our method in accordance with store demand. Finally, the proposed method
has now been completely implemented in the K/S’s internal system for use in the
upcoming packing and distribution season. The details of the validation results
can be found in [5].
6 Conclusion
In this document, we describe the Business Analytics project jointly conducted
by K/S and KAIST. The project has successfully developed a decision support
system to assist the packing and distribution procedure at K/S. Throughout
the project, the framework of the Business Analytics was first defined, and the
roles and responsibilities of each team were clearly identified based on the frame-
work. Then, multiple tasks were listed and performed by the joint team. The
proposed packing and distribution approach includes a statistical data analysis
that estimates the demand of stores for each product, the optimization of the
box configurations and their distribution, and the IT implementation for sustain-
able use. The core technologies developed in the project were the optimization
modeling and algorithm development. The box configuration of each box and
the distribution quantity for the boxes to the stores were optimized by mixed
integer programming. The proposed method was validated with simulation and
actual tests in the stores. As a result, the proposed method showed significant
effect, and K/S decided to implement it within their internal IT system.
The financial benefit delivered by the project is still under investigation.
Based on the simulation result and on-site tests, we project that the proposed
packing and distribution will improve sales by 5% 10% in terms of revenue. From
the cost-saving viewpoint, we estimate that costs related to the man-hour work-
ing time for determining the configuration, the boxes, and shipping and handling
will be significantly reduced. The unquantifiable contribution of this project is
also very significant. The project demonstrates how advanced data-driven meth-
ods and algorithms can be incorporated into a traditional retail fashion business.
It proves that in the modern data-driven society, a high-technology company is
not defined by its product but rather by how it handles its operations. Mr.
J.H. Jang, the Lead Manager of the Big Data Analytics Team, stated that “this
project showed how the scientific method can improve the operation and pro-
vided the direction of the Big Data Analytics Team”.
Moreover, this project has been an example of successful academic-industry
collaboration deriving actual tangible results. Often, academic researchers find
topics for research from other academic literature or come up with hypothetical
imaginary topics in the hope that these kinds of problems will be valuable to
industry. However, the KAIST researchers identify multiple topics that are not
known to academia but which are worthy of further investigation. The authors
are also currently working on an academic paper presenting the optimization
algorithm they developed for this project.
Finally, this packing and distribution problem is common across the retail
fashion industry. To our best knowledge, few retail fashion companies use logical
methods for the process. The process, model, and algorithm developed in this
project can be further developed as a service or software solution. We discovered
three patents (two applications and one awarded patent) for solution algorithms
of the optimization problem similar to what we constructed [4,6,7]. Among these
patents, two are from the Oracle Corporation [6,7] and one is from the SAS
Institute [4]. However, we found that there is still room for improvement in the
algorithms presented in these patents.
Appendix
Box assortment optimization model

The optimization model determines the number and contents of the box
configurations and their allocation to the stores to minimize overstocking and
understocking. The demand for items in each store is given by a multiple regres-
sion model. Because we know the maximum demand for each item among all
stores and the capacity of the box is finite, it is theoretically feasible to generate
a pool of all possible box configurations by selecting numbers between zero and
the maximum demand for each item and making a combination of them as a box
configuration, only if the sum of the number of items in a combination satisfies
the capacity. In this way, a set of possible box configuration is given.
When the box configurations and their assignments are established in K/S,
several managerial and operational requirements need to be met. First, a box
only contains items with an identical product type, style, and color, but different
sizes. For example, for such an item with five different sizes, a box can contain two
“Small”-sized items, five “Medium”-sized items, and three “Large”-sized items
because the manufacturer prefers to pack items of the same product type, style,
and color at one time unless the finished products are waiting until changes in
production setup and raw materials are needed. Additionally, the manufacturers
of the items might be different even if their product types and styles are the same
but their colors are different. Second, the following are constrained: the number
of different box configurations, the maximum and minimum quantities of items
in a box, i.e. capacity, the maximum and minimum number of boxes a store
receives, the maximum number of boxes, and the upper bound on overstocking.
Indices
S : set of stores (s ∈ S);
I : set of items (i ∈ I);
C : set of possible box configurations (c ∈ C)
Parameters
dis : demand of item i in store s;
cbi : integer number of item i in box configuration b;
αis : understocking penalty of item i in store s;
βis : overstocking penalty of item i in store s;
δis : upper bound on overstocking of item i in store s;
N B : number of different box configurations;
Ns : minimum number of boxes store s receives;
Ms : maximum number of boxes store s receives;
T : maximum number of boxes
Decision Variables
xbs : number of box configuration b allocated to store s;
yb : binary variable, which is 1 if box configuration b is used and 0 otherwise;
uis : understocking of item i in store s;
ois : overstocking of item i in store s
Objective

(αis uis + βis ois )
s∈S i∈I

Constraints
cbi xbs − ois + uis = dis ∀i ∈ I, s ∈ S

b∈B
xbs ≥ Ns ∀s ∈ S

b∈B
xbs ≤ Ms ∀s ∈ S

b∈B
yb ≤ N B

b∈B
xbs ≤ T
s∈S
b∈B
xbs ≤ Ms yb ∀b ∈ B

s∈S s∈S
xbs ≥ yb ∀b ∈ B
s∈S
xbs ≥ 0, integer ∀b ∈ B, s ∈ S
yb ∈ {0, 1} ∀b ∈ B.
References
1. Fischetti M, Monaci M, Salvagnin D (2015) Mixed-integer linear programming
heuristics for the prepack optimization problem. Discrete Optim 22:195–205
2. Hoskins M, Masson R et al (2014) The PrePack optimization problem. Springer
3. Industries K (2014) Financial information http://www.kolonindustries.com/Eng/
Service/service02.asp
4. Pratt RW (2014) Computer-implemented systems and methods for pack optimiza-
tion. U.S. Patent Application
5. Sung SW, Jang YJ (2015) Kolon-KAIST business analytics project. Internal report,
Kolon-KAIST Life Innovation Center
6. Vakhutinsky A (2012) Retail Pre-Pack optimization system. U.S. Patent Application
7. Vakhutinsky A, Subramanian S et al (2012) Retail pre-pack optimizer. U.S. Patent
Application
Integrated Project Management
Algorithmical and Program Functions
of Innovation Project Management
in Technoloji Park
Elchan Ghuseynov, Javanshir Mammadov, and Shafahat Rahim Rahimov(B)
Sumgait State University, Sumqayit, Azerbaijan

shafahat 61@mail.ru
Abstract. It was analyzed the present status of the innovation projects

management problem in many technology parks of European, Asian and
American countries. On the base of some of exists methods of option
and creation process of innovation projects, there were defined the basis
directions by innovation project management in technology park, also
scientific problems of subject area and option of the best projects by
means of flexible complex program system. In depend on the defined
scientific problems in the paper, the process of managing an innovation
project on the beginning stage of working in the departments of the tech-
nology park at higher educational school in Azerbaijan was considered.
On the basis of the priority scientific directions and the scientific pro-
files of scientific technology park in the Sumgait State University (SSU)
have been defined ways of decision a problem of an innovation project
management from its registration till its option and input to the data
base by an expert. On the base of the structural scheme of technology
park in SSU, its scientific profiles, infrastructure and specialists, a new
conception managing innovation projects in the technology park of SSU
in Azerbaijan by means of complex flexible interface was offered. For exe-
cuting all procedures of innovation project management on the stages its
creation, experimental research, manufacture and business, special algo-
rithm was worked out. By means of this algorithm, the block-scheme of
managing functions, the program tool for development of new innova-
tion project in the technology park beginning from the registration of
the project till its assessment was created.
Keywords: Technology park · Flexible manufacture · Innovation

project · Management · Higher educational school · Project assessment
1 Introduction
As known, creation of new innovation projects in the many science technology
parks, Zouain [13] and Lamperti [7] considered the problems of science technol-
ogy parks from its laboratory investigation, manufacture till economical process
1596 E. Ghuseynov et al.
in Brasilia organization is complex, difficult scientific-research, engineering, man-

ufacturing and economical managing process. Traditional technology parks in
of European, Asian and American countries solve own problem of innovation
project management by means of the corporative connections between different
departments of the technology park Volkonitskaia [12], Kirchberger and Pohl [6]
considered the ways of decision of functions innovation project development in
technology parks of different countries “without application of special software
systems what do not enough provide efficiency of planning process of creation
of innovation project. Application of such program will allow to realize some
science-technical, collection of necessary information about the new project, its
storage, choosing, expert analyzing and assessment. Analyze of the existed meth-
ods and ways of new innovation projects development Murzina [8] and Dı́ez-
Vial [4] considered international practice of innovation structure creation with
economical knowledge for technology parks” shown that, there is not identify
methodology for optimal choosing innovation project, its algorithmically and
program providing, which would be allowed to support high economical effi-
ciency and get great financial results for the budget of technology park.
Analyze of the problem by innovation projects option for them scientific
research, manufacture and commercial process shown that the date problem
solved without complex system application on a level of some experts solution
after presentation of projects. In this connection, for effective solution of the
best innovation project option under the complex system there were determined
problems of researching by the following directions:
• Determination of the basic scientific profiles in the technology park at Sumgait
State University (SSU) which can provide option of the necessary innovation
projects for development them in the parts of science, manufacture and busi-
ness of technology park;
• Creation of new conception of managing innovation projects in the technol-
ogy park of SSU in Azerbaijan which can provide realization of option the
best innovation project in the complex system for efficiency development of
technology park for short time;
• Development of algorithmically and program tools for providing management
of innovation projects in the scientific technology park of SSU which can pro-
vide executive menu procedures in corporative network system of technology
park “Rafael Avila Faraco [9] considered a problem of creation and using social
networks and knowledge transfer in technological park companies in Brazil,
what is solved only forming data base in the internal system and therefore
procedure of the best innovation project is done intuitive”.
In depended on the determined scientific problems, the purpose of the paper
defined as follows: it is necessary to work out universal algorithm and program
which would be allowed to create innovation projects in technology parks with
different scientific profiles.
This material was written by means of the executed sciences researches at
the department of “Information technology and programming” of Sumgait State
University, Azerbaijan Republic.
Algorithmical and Program Functions of Innovation Project Management 1597
2 Determination of the Basic Scientific Profiles in the

Technology Park at SSU
At present in Azerbaijan Republic, different innovations programs in many appli-

cation areas are realized. Specially, in many high educational schools of Azer-
baijan, as product of innovation technology, new technology parks are created.
Traditionally at first stage of technology park creation, a basic problem is option
of the necessary scientific profiles in the universities, as they have different spec-
ifies of education and different balance of scientific potentials-experts “Correia
and Gomes [3] considered a problem of forming potentialities and limits for the
local economic and innovative development of technology parks located in the
Northeast region of Brazil, where the problems of scientific directions are limited
owing to little application of intellectual information systems in different areas”.
As an object of investigation is SSU, where scientific technology park must
be created, then considering specifics of specialties of this universities and more
development scientific directions of SSU, the basics profiles of technology park
had been chosen:
• Application of intelligence information systems in different area of sciences

(practically in medicine, biology, physic and other areas);
• Investigation of physical and chemical process and theirs application in man-
ufacture;
• Application of modern commercial methods and systems for efficiency realiza-
tion of technology park.
The first scientific profile of SSU technology park is one of priorities direc-
tions in technology park, as application of progressive information technology,
intelligence systems and program tools provide computing functions of many
medical, biology, engineering and others important technical problems for citi-
zen. Option of this scientific profile was justified by means of the formed scientific
base which organized scientists of SSU, their scientific works which they conduct
about 50 years. Under scientific leader of Rafik Aliyev, a famous in world sci-
entific circles by development of fuzzy logic theory, was organized the scientific
school of scientists of SSU and theirs works had been implemented in flexible
manufacture systems in the metallurgy enterprise of Sumgait city. At present
the scientists as doctorates and professors work in SSU by own specialty. At
present 4 the laboratories of the departments of faculty “Engineering” provide
all education and sciences works of students and teachers.
The second scientific profile of SSU scientific technology park is multi-
scientific area which is connected theoretical and experimental investigations,
also implementation of results in manufacture process. Investigation on this
directions conduct scientists of physical and chemical faculties which have 4
the scientific laboratories and some special devices for experiments, are founded
in the scientific part of SSU.
The third scientific profile of SSU scientific technology park (SSU STP) is
one of the basis profile, as that profile must provide economical realization of
two profiles of SSU scientific technology park in whole. Application of modern

commercial methods and systems of management and marketing provide all way
economical control of innovation project beginning from an idea appearance, its
including to the technology park system, presentation, theoretical and practical
researching in laboratory, optimal option, embedding in the flexible manufacture
and sale an innovation product in the internal or external market.
On the base of the explained scientific profiles of the technology park, we
can mark that the existed laboratories for experimental researching in the first
and second profiles are science C research part of the technology park, and in
future they will have to be able to work by connection with business incubator
and flexible manufacture.
3 Managing Innovation Projects in the Technology Park

of SSU
Efficiency of working the experimental laboratories by 3 scientific profiles

depends on conception of managing innovation projects in the technology park,
as some of this projects have not the necessary text, graphical, animation and
video information for expert analyze, high technical and economical character-
istics, and also possibility of them making in the flexible manufacture system
owing to lack of some equipments and materials “Prencipe [10] considered a
problem of realization of innovation process in University Spin-offs” what is
shown only some innovation operation by economical management and market-
ing. In this connection, solution of the problem of new conception creation of
managing innovation projects in the technology park of SSU in Azerbaijan is
important, scientific-actuality researching problem.
In accordance with the offered conception of managing innovation project
in technology park, using principals of flexibleness and openness of working the
automation system, the way of managing innovation project is represented by
means of the following scheme:
• Input and presentation of scientist user idea with the necessary data of own
user;
• Expert analyze of project annotation;
• Experimental investigation in the laboratory and making a laboratory exem-
plar of the project;
• The project embedding in flexible manufacture and its output Elchan
Ghuseynov, Javanshir Mammadov, Gulnara Genjeliyeva [5] considered a prob-
lem of application of flexible industrial park in the scientific technology park
of Sumgait State University of Azerbaijan where the problems of innovation
project option and embedding in the manufacture are not considered.
4 Algorithmically and Program Tools for Providing

Management of Innovation Projects in the Technology
Park of SSU
Functions of management departments of SSU STP and development, commer-
cial way of innovation project are provided by means of the algorithm scheme
shown in the Fig. 1. The Procedures of innovation project management from the
beginning registration to its embedding into manufacture are represented on the
base of the following stages:
Fig. 1. The structural scheme of SSU STP with innovation project relocation
(1) The professors and teachers, also students of SSU present own ideas in the
global level of the system of registration. At first they input the registration
information to two the panels “Teacher registration” or “Student registra-
tion”. An User of the system inputs his data and information about himself
project (name, purpose of the project, short idea of the project-annotation)
which is saved in the “Annotation” panel for future familiarization of this
project by an expert. Also on the first stage for careful analysis in this part
of the system, the panel of “Prototipe of a project” is used.
(2) During a week, different experts in autonomy form control the registration
and annotation informations. On the base of standart form of the new project
receiving, a decision about acceptance or rejection of the project are given.
Procedures of an expert are executed and saved in the panel “Expert review”.
(3) For carefully presentation of the projects, the projects are represented to
the experts in the determined time. Actuality, newness, modernity, high
engineering decision and economical efficiency of the projects are the basis
of choosing the project for its first making as an experimental form in the
laboratory condition in accordance of the scientific profiles.
(4) In the experimental laboratory the best project is given for its constructor
design, material option, working out the first experience form of the project
and control of its mechanical, automation and others functions. At this stage
the basics technology characteristics of the project are defined and then
its quality level, the better data in difference of the prototypes ones. The
documents must be official formed by the status leader of the laboratory.
(5) All information of the designer is saved in his data base (in the part of
“Project data base”).
(6) The commercial department determines the basis of rules and the demanded
in the external and internal market and economical efficiency of the project is
computed. Čirjevskis [11] presented dynamically “signature business model”
which can provide durable competitive advantage at commercial procedures
of business incubator of technology park. “For presentation of the project to
the local and international markets Amit and Zott [1] worked out the model
for application in business process of innovation project by science and tech-
nology connection” in “Presentation of the new project” department, the
view of the 2, 3-measure pictures, animations, video and technical charac-
teristics are saved in the data base. The managers by the scientific profiles
chooses the clients and some information about them saves in the “Client
firm”. Between the scientific technology park and the firm, official meeting
is executed by means of this department. Therefore the official documents
are prepared.
(7) The procedures of project documents presentation are realized for the
experts in reality. On that stage all the documents registered by the expert
are sent to the flexible manufacture where a process of the project making
is executed.
The structural scheme of SSU STP with innovation project management is
given in Fig. 1. Also the functional connections between the jobs departments of
SSU STP are shown.
The represented text and block-scheme algorithms allow to develop a pro-

gram for checking, choosing a project way in the scientific technology park at
the Sumgait State University. In this connection, structural scheme of an inno-
vation project management is proffered on the base of this algorithms (Fig. 2).
The offered scheme consists of three blocks where each one works in different
rooms, beginning from the room of projects registration, the room of the saved
data base and the experts center.
between blocks in the rooms by means of

The module for information connecting
ROOM_1
The innovation projects

resiving module
corporative network
ROOM_3
The module of checking

the innovation projects
ROOM_2 by the experts
The module of forming

the scientific data base of
the innovation projects
Fig. 2. Structural scheme of an innovation project management under corporative net-

work of technology park
The innovation projects receiving module provides presentation of the basis

data about authors, title and aim of a project, receiving date and time of project
in the system, annotations of the project and its figure. The module of forming
the scientific data base of the innovation projects provides automatically saving
all data about a innovation project in the data base. Also in this database like
table form, the certification by application area is executed. For assessment of
every project which was marked in the system, an expert by corresponding appli-
cation area checks his innovation project and gives a mark by the development
project.
Checking one project can exist by some experts in this application area.
Information about assessment of a project is sent to an author of the project. In
this message answer can be by some versions: accepted; accepted, but must be
corrected; rejected.
5 Development of Program Tool for Expert Assessment

of Innovation Project in Scientific-Technoloji Park
In corresponding to functions of each the module, the program for expert assess-
ment of innovation project (EAIP) Carver et al. [2] considered the problem of
development of special Issue on software engineering technology and its applica-

tions corporative system of manufacture what is shown some program procedures
of designing innovation project in corporative network is worked out (Fig. 3). The
program for EAIP is worked by the following stages:
Fig. 3. The scheme of program functions in the EAIP
(1) At first stage a designer inputs his the data of authors names, title and aim
of a project, annotations of the project, the reciving date and time of project
into the system.
(2) On second stage all the data by the project are saved in the data base
management system as a table view.
(3) In third stage an expert checks information about the designer of a project
and begins to read and looking annotation of the project. After checking the
project, an expert assessment of this project gives by marks where a result
outputs in the system of EAIP and sends to the designer.
6 Conclusions
On the base of the made investigation by innovation project management, the
following results were got:
(1) In corresponding to the aim of the investigation problem, the algorithm pro-
cedure for management and assessment of an innovation project in scientific-
technopark was offered.
(2) The block-scheme of innovation management project, forming data base of

projects and assessment of a project by an expert was worked out.
(3) On the base of algorithm procedures of innovation project management, the
program operations with project registration; its author and text names; text
and graphical annotations; sending all information to expert by different sci-
entific areas; saving the neccessary data in the data base; expert assessment
of innovation project were developed.
References
1. Amit R, Zott C (2014) Business model design: a dynamic capability perspective.
Technical report. Citeseer
2. Carver D, Chan WK et al (2016) Special issue on software engineering technology
and applications. J Syst Softw 126:85–86
3. Correia AMM, Gomes MDLB (2014) Potentialities and limits for the local economic
and innovative development: a comparative analysis of technology parks located
in the northeast region of Brazil. Int J Innov Learn 15(3):274–298
4. Dı́ez-Vial I, Fernández-Olmos M (2015) Knowledge spillovers in science and tech-
nology parks: how can firms benefit most? J Technol Transf 40(1):1–15
5. Ghuseynov E, Mammadov J, Genjeliyeva G (2016) Application of flexible industrial
park in the scientific technology park of Sumgait state university of Azerbaijan.
Br J Appl Sci Technol 13(4):17–26
6. Kirchberger MA, Pohl L (2016) Technology commercialization: a literature review
of success factors and antecedents across different contexts. J Technol Transf
41(5):1–36
7. Lamperti F, Mavilia R, Castellini S (2015) The role of science parks: a puzzle of
growth, innovation and R&D investments
8. Murzina SE (2015) International practice of innovation infrastructure creation as a
mechanism for the innovative economy development and the improvement of land
use effectiveness. J Siberian Federal Univ Humanit Soc Sci 8:2535–2544
9. Mussi C, Angeloni MT, Faraco RA (2014) Social networks and knowledge transfer
in technological park companies in Brazil. J Technol Manage Innov 9(2):172–186
10. Prencipe A (2016) Board composition and innovation in university spin-offs: evi-
dence from the italian context. J Technol Manage Innov 11(3):33–39
11. Čirjevskis A (2016) Designing dynamically “signature business model” that sup-
port durable competitive advantage. J Open Innov Technol Mark Complexity
2(1):15
12. Volkonitskaia K (2015) Business models of technoparks in russia. In: Series: Science,
Technology and Innovation WP BRP, WP BRP 55/STI/2015
13. Zouain DM, Plonski GA (2015) Science and technology parks: laboratories of inno-
vation for urban development - an approach from Brazil. Triple Helix 2(1):1–22
Two-Stage Fuzzy DEA Models with Undesirable
Outputs for Banking System
Xiaoyang Zhou1,2 , Rui Luo1 , Benjamin Lev2 , and Yan Tu3(B)

1
International Business School, Shaanxi Normal University,
2
LeBow College of Business, Drexel University,
Philadelphia, PA 19104, USA
3
School of Management, Wuhan University of Technology,
Wuhan 430070, People’s Republic of China
tuyan belle@163.com
Abstract. In this paper, we propose two stage fuzzy DEA models with
undesirable outputs to evaluate banking system. The banking system
is divided to two subsystems: production stage and profit stage. In the
model, two kinds of assumptions (constant returns to scale and variable
returns to scale) are considered, and fuzzy parameters are adopted to
describe the uncertain factors. Chance constrained operator is used to
handle the proposed model and equivalent transformations are given to
make the models solvable. We illustrate and validate the proposed models
by evaluating 16 Chinese commercial banks. Some discussions are also
presented to show the differences and advantages of the models.
Keywords: Two-stage DEA · Banking system · Undesirable outputs ·

Constant returns to scale · Variable returns to scale
1 Introduction
Efficiency is particularly illustrated in the comprehensive competitiveness of
commercial banks. In terms of investment, supervision or independent opera-
tion, it is of great significance to evaluate the efficiency of commercial banks.
Data Envelopment Analysis (DEA), a non-parametric mathematical program-
ming approach to evaluate group of Decision Making Units (DMUs) with com-
parative efficiency, is widely applied to analyze the efficiency in the process of
banking operation [1,4,5,12,13,17].
It seems like the traditional DEA model treats DMU as “black box”, which
is unable to explain the internal structure inside the “black box”. In a banking
system, taking deposits and making loans are the main activities. Some banks may
have greater advantage in taking deposits, but perform not as well in making loans
to realize profit. Hence, it is better to consider the banking system as a two-stage
process. We should evaluate not only each stage, but also the whole system.
Undesirable Outputs for Banking System 1605
In the banking system, there are desirable outputs, like the total deposit
during the first stage, and the profit during the second stage. But in the oper-
ational process, banks want to get rid of the risk exposure and non-performing
loans/assets, which are undesirable outputs. Scheele [15] concluded that the
undesirable outputs could be handled by direct and indirect approaches. In
indirect approaches, the values of the undesirable outputs are converted to a
monotone decreasing function, and then we can treat the undesirable outputs
as the other normal desirable outputs. Direct approaches avoid data transfor-
mation and incorporate the undesirable outputs as the inputs directly into the
DEA models. In Hu’s work [20], undesirable outputs were treated as the inputs,
and the less of the undesirable outputs (inputs), the better. A new two-stage
DEA model with the undesirable outputs to measure the slacks-based efficiency
of Chinese commercial banks during years 2008–2012 was developed in [2], where
the banking operation process of each bank was divided into a deposit generation
stage and a deposit utilization stage. However, they neglected the undesirable
outputs of the deposit generation stage.
In recent years, scholars have already realized that some inputs and outputs
are imprecise in real life cases. Conventional DEA models used the accurate inputs
and outputs data, which may not fit in the real world cases. To deal with the impre-
cise data, the fuzzy theory was introduced in DEA to become fuzzy DEA.
According to Hatami-Marbini [7], the methods of handling the fuzzy DEA can
be divided into four types: (i) The tolerance approach [16]; (ii) The α-level based
approach [8,11]; (iii) The fuzzy ranking approach [6]; and (iv) The possibility
approach [9,10,19]. Puri [14] presented a fuzzy DEA model with undesirable
fuzzy outputs, giving a numerical illustration of the banking sector in India using
fuzzy input/output data for the period 2009–2011, but the internal structure of
banking operation system was not taken into account. Wanke [18] developed
new Fuzzy-DEA models to measure the impact of each model on the efficiency
scores and to identify the most relevant contextual variables of efficiency by using
bootstrap truncated regressions with fixed factors.
In this paper, we propose a two stage fuzzy DEA model with undesirable
outputs to evaluate 16 China’s commercial banks. The remainder of this study
is organized as follows. Section 2 presents problem statement. Section 3 develops
three kinds of two-stage DEA models with undesirable outputs. Section 4 dis-
cusses the solution method. Section 5 gives a case study with 16 banks in China.
Finally, conclusions are given in the last section.
2 Problem Statement
The efficiencies of several banking systems (DMUs) are under evaluation, and
each DMU, indexed by j = 1, · · · , J, has two sub-stages: production stage and
profit stage. During the production stage, funds are collected from depositors
while consuming resources such as number of employees xN j
E
, fixed assets xF A
j ,
operating expenses xj and number of institutions xj . Total deposits Zj
OE NI TD
is not only the desirable output of the production stage but also the input to
1606 X. Zhou et al.
x NE FA OE NI
j , xj , xj , xj Production z TD
j y RP
j
Profit Stage
Stage
∼z RWA ∼y NPA
j j
Fig. 1. Two-stage banking system
the profit stage. Total risk-weighted asset z̃jRWA is the undesirable output during
the production stage. In the profit stage, deposits ZjT D are used to invest into
other activities to get profits yjRP , but the undesirable output non-performing
loans/assets ỹjNPA will also be produced. Because deposits collected during the
production stage determines the investment decision in the profit stage, we model
this banking system as a two-stage problem, which is shown in Fig. 1.
In practice, some data could be imprecise. Total risk-weighted assets refer to
the sum of In-Balance-Sheet assets and Off-Balance-Sheet exposures, weighted
according to different risk coefficients, respectively. The risk coefficients are
affected by macroeconomic environment, current risk coefficient, exposure of
bank, etc., which is difficult to ascertain. The non-performing loans/assets are
loans with signs of difficulties to repay the loans principal and interest. It is
different to ascertain whether the borrowers can repay the loans on time, which
are affected by public policy, credit of the clients, actual situation of the clients,
etc. Due to the variability of risk coefficients and the repayment ability of bor-
rowers, we employ fuzzy numbers to describe the total risk-weighted assets and
non-performing loans/assets.
Definition 1. [21] Let L(·), R(·) be two reference functions. If the membership
function of fuzzy variable ξ has the following form:
m−x
L( α ), x ≤ m, α > 0
μξ (x) =
R( x−m
β ), x ≥ m, β > 0.
Then ξ is called an LR fuzzy variable denoted by (m, α, β)LR , where m is

the central value of ξ, α, β are the left and right spreads respectively of ξ. And
the γ − cut(0 ≤ γ ≤ 1) for the LR fuzzy variable is:
ξγ = (m − L−1 (γ) α, m + R−1 (γ) β),γ ∈ [0, 1].
3 Models
Charnes-Cooper-Rhodes (CCR) model assumes all the DMUs are Constant
Returns to Scale (CRS) [3]. In some cases we also have Variable Returns to
Scale (VRS). In this section, three kinds of DEA fuzzy models are developed for
commercial banking system evaluation.
3.1 Fuzzy CCR DEA Model with Undesirable Outputs
When the banking system is considered as a “black box”, then we formulate the
CCR model as follows.
ud y0RP
max hCCR0 = 4
i=1 vi xi0 + wu z̃0
RWA + u ỹ NPA
⎧ u 0
⎪
⎨ u d y RP
j (1)
4 ≤ 1, j = 1, · · · , J
s.t. v x + w z̃ RWA + u ỹ NPA
⎪ i=1 i ij
⎩ v , w , u , u ≥ ε > 0,
u j u j
i u d u
where vi , wu , ud , and uu are non-negative weights. ε is a small non-Archimedean

number imposed for avoiding ignorance of any factor. xij (i = 1, 2, 3, 4) represent
xN
j
E
, xF
j , xj , xj , respectively.
A OE NI
3.2 Two-Stage Fuzzy CRS DEA Model with Undesirable Outputs
Next we discuss the case when the banking system is divided into two stages.
Suppose all the DMUs are CRS, we develop models (2) and models (3) to cal-
culate the efficiencies of the first stage and the second stage.
1 wd z0T D
max hCRS
0 = 4
RWA
⎧
⎪
⎨ wd zjTD
(2)
4 ≤ 1, j = 1, · · · , J
s.t. i=1 vi xij + wu z̃j
RWA
⎪
⎩ v , w , w ≥ ε > 0,
i d u
and
2 ud y0RP
max hCRS
0 =
⎧ wd z0 + uu ỹ0NPA
T D
⎨ ud yjRP (3)
≤ 1, j = 1, · · · , J
s.t. wd zjT D + uu ỹjNPA
⎩
wd , ud , uu ≥ ε > 0.
Then we use ω1 and ω2 to represent the weights of the first stage and the
second stage, and the formulations are as follows.
4
i=1 vi xij + wu z̃j
RWA
ω1 = 4 , (4)
i=1 vi xij + wu z̃j
RWA + w z T D + u ỹ NPA
d j u j
and
wd zjT D + uu ỹjNPA
ω2 = 4 . (5)
i=1 vi xij + wu z̃jRWA + wd zjT D + uu ỹjNPA
1608 X. Zhou et al.
Then the two-stage fuzzy CRS DEA model with undesirable can be proposed
as Eq. (6).
1 2 wd z0T D + ud y0RP
max hCRS
0 = ω1 hCRS0 + ω2 hCRS
0 = 4
RWA + w z T D + u ỹ NPA
⎧ d 0 u 0
⎪
⎪ w z
d j
TD
⎪
⎪ 4 ≤ 1, j = 1, · · · , J
⎪
⎨ i=1 vi xij + wu z̃jRWA
s.t. ud yjRP
⎪
⎪ ≤ 1, j = 1, · · · , J
⎪
⎪ wd zj + uu ỹjNPA
T D
⎪
⎩
vi , wd , wu , ud , uu ≥ ε > 0.
(6)
3.3 Two-Stage Fuzzy VRS DEA Model with Undesirable Outputs

Often, we also face VRS. In order to consider the efficiency of the banks in the
case of VRS, we propose the two-stage fuzzy VRS model as follows.
wd z0T D + μ1 + ud y0RP + μ2
max hV0 RS = 4
RW A + w z T D + u ỹ N P A
⎧ d 0 u 0
⎪
⎪ w d z TD
j + μ 1
⎪ 4
⎪ ≤ 1, j = 1, · · · , J
⎪
⎨ i=1 vi xij + wu z̃jRW A (7)
s.t. ud yjRP + μ2
⎪
⎪ ≤ 1, j = 1, · · · , J
⎪
⎪ wd zjT D + uu ỹjN P A
⎪
⎩
vi , wd , wu , ud , uu ≥ ε > 0
If μ > 0, then the DMU has increasing returns to scale; if μ = 0, then the
DMU has constant returns to scale; if μ < 0, then the DMU has decreasing
returns to scale.
4 Solution Method
In this section, we propose a solution to the two-stage DEA models with fuzzy
parameters.
4.1 Solution Method Equivalent Two-Stage CCR DEA Model

First, applying the Charnes-Cooper transformation (C-C transformation), the
two-stage model (8) can be converted into the following linear program, and the
efficiency of DM U0 can be calculated. The optimal efficiency gjCCR is equal to
the optimal efficiency hCRS
j .
max⎧g0CCR =ud y0RP
⎪ 4
⎪
⎪ ud yjRP − vi xij − wu z̃jRW A − uu ỹjN P A ≤ 0, j = 1, · · · , J
⎪
⎨ i=1 (8)
s.t. 4
⎪
⎪ vi xi0 + wu z̃0RW A + uu ỹ0N P A = 1
⎪
⎪
⎩ i=1
vi , wu , ud , uu ≥ ε > 0.
Because of the fuzzy parameters, we use chance constrained operator to han-

dle model (8), and we get model. (9).
max⎧g0CCR = ud y0RP
⎪ 4
⎪
⎪ Pos u y RP
− v x − w z̃ RWA
− u ỹ NPA
≤ 0 ≥ γ, j = 1, · · · , J
⎪
⎪
d j i ij u j u j
⎪
⎪ i=1
⎪
⎨ 4
Pos vi xi0 + wu z̃0RWA + uu ỹ0NPA ≥ 1 ≥ γ
s.t.
⎪
⎪
i=1

⎪
⎪
4
⎪
⎪ Pos vi xi0 + wu z̃0RWA
+ uu ỹ0
NPA
≤1 ≥γ
⎪
⎪
⎩ i=1
vi , wu , ud , uu ≥ ε > 0
(9)
Theorem 1. [21] Given two LR fuzzy variables ξ1 = (m1 , α1 , β1 ), ξ2 =

(m2 , α2 , β2 ), where m are the central values; and α and β are the left and the
right spreads respectively. Then we have:
1 = (km1 , kα1 , kβ1 ) , k > 0

kξ
ξ1 + ξ2 = (m1 + m2 , α1 + α2 , β1 + β2 )
ξ1 − ξ2 = (m1 − m2 , α1 + β2 , β1 + α2 )
Pos {ξ1 ≥ ξ2 } ≥ γ ⇔ mR 1γ ≥ m2γ ,
L
where k is a real number, γ(0 ≤ γ ≤ 1) are the possibility values, mR 1γ is the

right end point of γ − cut of ξ1 , and mL
2γ is the left end point of γ − cut of ξ2 .
According to Theorem 1, model. (9) can be transformed into model. (10).
max ⎧g0CCR = ud y0RP
⎪
⎪ 4
⎪
⎪ u y RP − vi xij − (wu (zjRWA )m + uu (yjN P A )m ) − (1 − γ)(wu (zjRWA )β + uu (yjNPA )β ) ≤ 0,
⎪ d j
⎪
⎪
⎪
i=1
⎪
⎪ j = 1, · · · , J
⎪
⎨ 4
s.t. v x + wu (z0RWA )m + uu (y0NPA )m + (1 − γ)(wu (z0RWA )β + uu (y0NPA )β ) ≥ 1
⎪ i=1 i i0
⎪
⎪
⎪
⎪
⎪
4
⎪
⎪
RWA
)m + uu (y0NPA )m − (1 − γ)(wu (z0RWA )α + uu (y0NPA )α ) ≤ 1
⎪ i=1 vi xi0 + wu (z0
⎪
⎪
⎩
v i , wu , ud , uu ≥ ε > 0
(10)
After getting the optimal solutions vr∗ , wu∗ , u∗d and u∗u , the CCR DEA effi-
ciency g0CCR of the evaluated whole system can be obtained.
4.2 Equivalent Two-Stage CRS and VRS DEA Models

Similarly, models (6) and (7) can be converted into models (11) and (12):
1610 X. Zhou et al.
max ⎧g0CRS = wd z0T D + ud y0RP

⎪
⎪ 4
⎪
⎪ wd zjT D − vi xij − wu (zjRWA )m − (1 − γ)wu (zjRWA )β ≤ 0, j = 1, · · · , J
⎪
⎪
⎪
⎪
i=1
NPA
⎪
⎪ ud yj − wd zj − uu (yj
⎪
RP TD
)m − (1 − γ)uu (yjNPA )β ≤ 0, j = 1, · · · , J
⎨ 4
RWA
s.t. TD
v x + wd z0 + wu (z0 )m + uu (y0NPA )m + (1 − γ)(wu (z0RWA )β + uu (y0NPA )β ) ≥ 1
⎪ i=1 i i0
⎪
⎪
⎪
⎪
⎪
4
⎪
⎪ vi xi0 + wd z0T D + wu (z0RWA )m + uu (y0NPA )m − (1 − γ)(wu (z0RWA )α + uu (y0NPA )α ) ≤ 1
⎪
⎪ i=1
⎪
⎩
vi , wd , wu , ud , uu ≥ ε > 0,
(11)
g0V RS
max ⎧ wd z0T D
= ud y0RP
+ μ1 + + μ2
⎪
⎪ 4
RWA
⎪
⎪ w d z TD
− v i x − wu (z )m − (1 − γ)wu (zjRWA )β + μ1 ≤ 0, j = 1, · · · , J
⎪
⎪
j ij j
⎪
⎪
i=1
NPA
⎪
⎪ ud yj − wd zj − uu (yj
⎪
RP TD
)m − (1 − γ)uu (yjNPA )β + μ2 ≤ 0, j = 1, · · · , J
⎨ 4
RWA
s.t. TD
vi xi0 + wd z0 + wu (z0 )m + uu (y0NPA )m + (1 − γ)(wu (z0RWA )β + uu (y0NPA )β ) ≥ 1
⎪
⎪ i=1
⎪
⎪
⎪
⎪ 4
⎪
⎪ vi xi0 + wd z0T D + wu (z0RWA )m + uu (y0NPA )m − (1 − γ)(wu (z0RWA )α + uu (y0NPA )α ) ≤ 1
⎪
⎪ i=1
⎪
⎩
vi , wd , wu , ud , uu ≥ ε > 0.
(12)
By solving models (11) and (12), the CRS and VRS DEA efficiencies of the
evaluated whole system, g0CRS and g0VRS , can be obtained. Also the efficiencies
1 1 2 2
of each sub-stage g0CRS (g0VRS ) and g0CRS (g0VRS ) can be obtained.
5 Case Study
In this section, we evaluate the efficiencies of the top 16 Chinese banks: Bank Of
China (BOC), Agricultural Bank of China (ABC), Industrial and Commercial
Bank of China (ICBC), China Construction Bank (CCB), Bank of Communica-
tions (BCM), China Citic Bank (CNCB), Ping An Bank (PAB), Hua Xia Bank
(HXB), China Everbright Bank (CEB), China Industrial Bank (CIB), China Mer-
chants Bank (CMB), Bank Of Ningbo (NBB), Bank of Nanjing (NJB), Bank of Bei-
jing (BJB), Shanghai Pudong Development Bank (SPDB), and China Minsheng
Banking (CMSB).
5.1 Data
Table 1 presents the data of the inputs and outputs of the selected 16 Chinese
commercial banks in 2015.
5.2 Results
Based on the data in Table 1, first we set the decision-makers’ confidence levels
for the two stages as γ = 0.9, then the efficiencies of each bank in different
circumstances can be produced after solving two-stage fuzzy CCR, CRS, and
VRS DEA models by LINGO, respectively. The results are shown in Tables 2
and 3. And Fig. 2 compares the system efficiencies of 16 commercial banks based
on the proposed three kinds of models.
Table 1. Data
Bank Number of Fixed Operating Number of Total Total Retained Non-

employees assets (100 expenses institutions deposits risk-weighted profits Performing
(×100) million (100 (×10) (10 billion assets (10 (100 loans/Assets
RMB) million RMB) billion RMB) million (100 million
RMB) RMB) RMB)
BOC 3100 1820 2439 1069 1173 (1065,51,38) 1794 (1,309,181,120)
ABC 5031 1562 3033 2367 1354 (1099,13,54) 1808 (2,129,706,125)
ICBC 4663 2215 1024 1673 1628 (1322,55,88) 2777 (1,795,355,168)
CCB 3692 1595 3091 1492 1367 (1072,37,52) 2289 (1,660,372,160)
BCM 915 904 1081 314 448 (465,33,29) 668 (562,96,52)
CNCB 499 160 905 135 318 (347,41,27) 417 (361,66,25)
PAB 323 479 673 100 173 (166,18,25) 219 (176,45,41)
HXB 342 113 339 79 135 (133,9,14) 190 (163,43,24)
CEB 403 126 539 94 199 (219,23,30) 296 (244,65,11)
CIB 520 178 915 179 248 (343,37,16) 507 (260,61,67)
CMB 762 308 1272 171 357 (321,22,0) 580 (474,136,78)
NBB 95 502 115 29 36 (44,7,7) 66 (24,4,3)
NJB 74 450 139 15 50 (50,8,10) 71 (21,3,6)
BJB 138 713 230 44 102 (128,15,9) 169 (87,24,9)
SPDB 484 1906 4629 166 295 (337,36,24) 510 (351,104,73)
CMSB 595 3773 942 106 273 (335,35,13) 470 (328,90,53)
Fig. 2. Efficiencies based on CCR, CRS and VRS models
5.3 Comparison and Discussion
According to Table 2, ICBC, CCB, CNCB, CEB, CIB, CMB, NBB, BJB, SPDB,
and CMSB are at the frontier when using traditional CCR DEA model evalu-
ation, but the advantages are not maintained when using CRS DEA model.
At the same time, the efficiencies of BOC, BCM, HXB, CMSB also decline in
varying amounts. We conclude that often the CCR model may get high effi-
ciency scores when the system is considered as a “black box” by using four
inputs (Employees, Fixed assets, Operating expenses and Institutions) to pro-
duce three outputs (Total Risk-Weighted Assets, Retained profits and Non-
Performing loans/Assets). However, the efficiencies may be not that high any
more when the intermediates are taken into account. In other words, the tra-
ditional CCR model can only find weak effective system, but it is different to
distinguish which banks are truly more efficient.
1612 X. Zhou et al.
Table 2. Efficiencies based on two-stage CCR and CRS models
Bank CCR DEA CRS two-stage DEA efficiency Bank CCR DEA CRS two-stage DEA efficiency
efficiency efficiency
System Stage 1 Stage 2 System Stage 1 Stage 2
BOC 0.89 0.85 0.95 0.75 CEB 1 0.86 0.99 0.73
ABC 0.79 0.82 0.99 0.66 CIB 1 0.9 0.78 1
ICBC 1 0.92 1 0.84 CMB 1 0.9 1 0.8
CCB 1 0.91 1 0.82 NBB 1 0.94 0.75 1
BCM 0.89 0.85 0.96 0.73 NJB 1 1 1 1
CNCB 1 0.83 1 0.64 BJB 1 0.93 1 0.86
PAB 0.81 0.81 1 0.62 SPDB 1 0.89 0.92 0.85
HXB 0.9 0.84 1 0.69 CMSB 0.99 0.83 0.82 0.84
Table 3. Efficiencies under 2 return scale assumptions
Bank CRS two-stage VRS two-stage DEA efficiency Bank CRS two-stage VRS two-stage DEA efficiency
DEA efficiency System Stage 1 Stage 2 DEA efficiency System Stage 1 Stage 2
BOC 0.85 0.85 0.95 0.75 CEB 0.86 0.93 1 0.74
ABC 0.82 0.82 0.99 0.66 CIB 0.9 0.91 0.8 1
ICBC 0.923 0.924 1 0.84 CMB 0.9 0.91 1 0.8
CCB 0.914 0.915 1 0.82 NBB 0.94 1 1 1
BCM 0.85 0.85 0.96 0.73 NJB 1 1 1 1
CNCB 0.83 0.88 1 0.65 BJB 0.926 0.93 1 0.86
PAB 0.81 0.82 0.97 0.63 SPDB 0.89 0.89 0.92 0.85
HXB 0.84 1 1 1 CMSB 0.834 0.835 0.82 0.85
By using two stage DEA model, we can further find out more discriminating
DEA efficiencies for DMUs because of the stronger restriction on the production
possible set. Although ICBC, CCB, CNCB, PAB, HXB, CMB and BJB are DEA
efficient at the production stage, all the efficiencies of them at the profit stage
are below 1. Similarly, CIB and NBB are inefficient at the profit stage while
both of them are DEA efficient at the production stage. That is due to CCR
DEA model neglects the internal interaction between the two stages. It also
proves that model (6) reflects not only the efficiency of each stage, but also the
efficiency of the banking system considering the relationship of two sub-stages.
Interestingly, the efficiency of ABC is improved by using model (6), which
manifests ABC have a great advantage of the internal operational process. We
find that the ratio of risk-weighted asset and inputs of the NJB is relatively small,
the ratio of deposit and input is relatively high, and the input/output ratio of
the NJB is at a reasonable level. Further analysis shows that both efficiency
values of NJB in model (1) and model (6) are 1, i.e. they are both located in
the frontier. This is consistent with our analysis, and policy makers can consider
them as the industry benchmark.
If all the inputs of the bank have changed in the same proportion, this change
will have an impact on the total outputs of the system. That is, the banking
system will have VRS in the actual operation process. Based on above, we pro-
pose two kinds of models under the CRS and VRS assumptions. According to
the model (7), the corresponding DEA efficiency can be obtained in Table 3.
In Table 3, column 2 and column 3 are the two-stage banking system overall
efficiency values obtained under the two assumptions, respectively.
It is found that HXB, NBB and NJB are DEA efficient not only in the whole
bank system, but also in both of the two sub-stages. Compared with two-stage
CRS DEA model, the efficiencies of two-stage VRS DEA increase slightly, and
the number of points on the frontier also increases. Accordingly, we can draw
a conclusion that commercial banks will have better performance in the VRS
conditions. Therefore, in order to obtain a reasonable input/output combination
to improve the performance of the system, the decision-makers can improve
the scale of return by some actions (such as training staff to improve business
efficiency, upgrading of banking equipment, improve its business throughput, and
so on). In general, the difference between DEA efficiencies under the assumption
of different scale returns is in a small range, which proves that the models we
proposed have certain stability.
6 Conclusion
This paper explores a problem of efficiency evaluation of banking system under
an uncertain environment. We consider the bank operating process consists of
production stage and profit stage, and the two sub-stages of system are connected
with each other by intermediate factor. Funds are collected from depositors in
the production stage, which are used to invest into other activities to get profits
in the profit stage. The undesirable outputs of both two sub-stages are treated as
inputs. On this basis, we present two-stage fuzzy DEA models with undesirable
outputs based on two different returns to scale.
In this paper, we have three main contributions as follows: First, based on
two assumptions of returns to scale, we present the two-stage fuzzy CRS and
VRS DEA models with undesirable output. The efficiency of the whole system
is defined as the weighted sum of the efficiencies of two sub-stages, where the
weights depend on the importance of the two stages. Second, considering the
variability of risk coefficients and the repayment ability of borrowers, we apply
fuzzy numbers to describe the total risk-weighted assets and non-performing
loans/assets, which is closer to practice. Then we apply chance constrained
operator to handle the fuzzy factors, and the objective is optimized subject
to the chance constraints under certain confidence levels. Third, a case study
and detailed comparison discussion are given by examining the top 16 China’s
commercial banks in 2015.
This paper focuses on the actual operation of commercial banks, considering
the internal operational structure of commercial banks and its undesirable output
in the process of operation. Through the application of the model, it is found
that the model can better evaluate the input/output structure of the bank and
has the ability of pointing out the direction to be improved.
1614 X. Zhou et al.
Acknowledgements. We thank those that have given constructive comments and

feedback to help improve this paper. Support was provided by the National Nat-
ural Science Foundation of China (Grant No. 71401093), China Scholarship Council
(Grant No. 201606875006), Soft Science Research Project of Shaanxi Province (Grant
No. 2016KRM089), Research Center for Systems Science & Enterprise Development
(Grant No. Xq16B01), and the Fundamental Research Funds for the Central Universi-
ties (Grant Nos. 14SZYB08, WUT: 2016VI002).
References
1. Akther S, Fukuyama H, Weber WL (2013) Estimating two-stage network slacks-
based inefficiency: an application to bangladesh banking. Omega 41(1):88–96
2. An Q, Chen H et al (2015) Measuring slacks-based efficiency for commercial banks
in China by using a two-stage DEA model with undesirable output. Ann Oper Res
235(1):13–35
3. Charnes A, Cooper WW, Rhodes E (1978) Measuring the efficiency of decision
making units. Eur J Oper Res 2(6):429–444
4. Chen Y, Cook WD, Zhu J (2010) Deriving the DEA frontier for two-stage processes.
Eur J Oper Res 202(1):138–142
5. Cook WD, Liang L, Zhu J (2010) Measuring performance of two-stage network
structures by DEA: a review and future perspective. Omega 38(6):423–430
6. Hatami-Marbini A, Saati S, Tavana M (2010) An ideal-seeking fuzzy data envel-
opment analysis framework. Appl Soft Comput 10(4):1062–1070
7. Hatami-Marbini A, Emrouznejad A, Tavana M (2011) A taxonomy and review of
the fuzzy data envelopment analysis literature: two decades in the making. Eur J
Oper Res 214(3):457–472
8. Kao C, Liu ST (2011) Efficiencies of two-stage systems with fuzzy data. Fuzzy Sets
Syst 176(1):20–35
9. Khodabakhshi M, Gholami Y, Kheirollahi H (2010) An additive model approach
for estimating returns to scale in imprecise data envelopment analysis. Appl Math
Model 34(5):1247–1257
10. Lertworasirikul S, Fang SC et al (2003) Fuzzy data envelopment analysis (DEA):
a possibility approach. Fuzzy Sets Syst 139(2):379–394
11. Liu ST (2014) Restricting weight flexibility in fuzzy two-stage DEA. Comput Ind
Eng 74:149–160
12. Liu W, Zhou Z et al (2015) Two-stage DEA models with undesirable input-
intermediate-outputs. Omega 56:74–87
13. Puri J, Yadav SP (2013) A concept of fuzzy input mix-efficiency in fuzzy DEA and
its application in banking sector. Expert Syst Appl 40(5):1437–1450
14. Puri J, Yadav SP (2014) A fuzzy DEA model with undesirable fuzzy outputs and
its application to the banking sector in India. Expert Syst Appl 41(14):6419–6432
15. Scheel H (2001) Undesirable outputs in efficiency valuations. Eur J Oper Res
132(2):400–410
16. Sengupta JK (1992) A fuzzy systems approach in data envelopment analysis. Com-
put Math Appl 24(8):259–266
17. Tavana M, Khalili-Damghani K (2014) A new two-stage stackelberg fuzzy data
envelopment analysis model. Measurement 53:277–296
18. Wanke P, Barros C, Emrouznejad A (2016) Assessing productive efficiency of banks
using integrated fuzzy-DEA and bootstrapping: a case of mozambican banks. Eur
J Oper Res 249(1):378–389
19. Wu DD, Yang Z, Liang L (2006) Efficiency analysis of cross-region bank branches
using fuzzy data envelopment analysis. Appl Math Comput 181(1):271–281
20. Hu XY, Cheng XJ, Ma LJ (2013) Efficiency evaluation of commercial banks based
on two-stage DEA model considering undesirable outputs. J Univ Chin Acad Sci
30(4):462–471 (in Chinese)
21. Xu J, Zhou X (2011) Fuzzy-like multiple objective decision making, vol 263.
Springer
Increasing Effect in Lodger Number of Hot
Spring Hotel According to the Started
Operation of Hokuriku Shinkansen
Takashi Oyabu1(B) and Junko Nakamura2

1
Kokusai Business Gakuin College, Kanazawa, Japan
oyabu24@gmail.com
2
Kanazawa Seiryo University, Kanazawa, Japan
Abstract. In Hokuriku district, economic efficiency is improved accord-

ing to the stated operation of Hokuriku Shinkansen. However, there are
some variations in the effect each area. It is necessary to utilize the
effect into a future tourism strategy by indicating the effects. In this
experiment, the effect is evaluated using the characteristics of the lodger
number in nine spa areas of Hokuriku district. As a result, there is a
larger effect as an area is near to a Shinkansen station and the increas-
ing rate is over 30%. Other areas except Wajima area have the rates
of 15 to 20%. The one for Wajima area is 33%. It keeps at a distance
from Kanazawa which is the nearby station of Wajima area. Popular
NHK television-drama (title is “Mare”) was televised for end of March
to September 2015 (half-year). Wajima area was the filming site. It is
estimated that there is an effect over 10% as the filmed site. There is
a large effect by connecting Hokuriku with Tokyo area by Shinkansen
(high speed railway).
Keywords: Hokuriku Shinkansen · Hot spring · Lodger number · Eco-

nomic effect
1 Introduction
Japan’s population has entered into a decreasing phase due to the decreasing
birthrate and aging population. Productive population also decreases remark-
ably, therefore a measure to put on the brake for the phenomenon is required.
National budget is also tight due to the increasing of the cost of social security
including medical expense as the society ages [2]. It is essential to build a policy
which could create a new industry and put the brakes on the depopulation in
a rural area. One of the regional-revitalization measures is “tourism industry”.
Human exchange is a basic policy to develop the tourism industry. The following
measures are important in the exchange, namely information and transportation
infrastructures [4].

Increasing Effect in Lodger Number of Hot Spring Hotel 1617
Comfortable and express railway system (e.g. Shinkansen) had been desired
in Hokuriku district (Fukui, Ishikawa and Toyama Prefectures) for 50 years. The
newest section of the Hokuriku Shinkansen Line, between Nagano and Kanazawa,
opened on March 14, 2015. The line was opened about fifty years after Tokaido
Shinkansen Line started operations. The number of visitors was increased con-
siderably due to the operation in Kanazawa City (Shinkansen effect), which
was the prefectural capital of Ishikawa. The occupancy rate for hotels was over
85%. There are nine main spa areas mainly in Hokuriku district, namely Awara,
Yamanaka, Yamashiro, Katayamatsu, Awazu, Yuwaku, Wakura, Wajima and
Unazuki. The number of lodgers has increased nearly 15 to 30% in the areas.
The areas except for Unazuki (Toyama Prefecture) and Yuwaku (Kanazawa
City) take one hour and half by bus from each Shinkansen station. Unazuki
and Yuwaku take within thirty minutes by car or train. The nearby station of
Unazuki is Kurobe-Unazuki-Onsen and the one for Yuwaku is Kanazawa. The
number of visitors for both spa areas increased about 30% and the start of the
operation significantly affected the number for both areas. It is desirable for the
ripple effect to affect whole areas. The mass media reported that the number
increased in whole spa areas. It is necessary to investigate the effect and apply
the survey result to the future strategy.
Outline of ripple effect can be known by examining the monthly variability
of lodger number in each spa area and it is possible to determine the date of
various kinds of providing events [1,3,11]. Moreover, it is helpful to construct
the effective strategy that increases the guest number. The variability of lodger
number in nine spa areas of Hokuriku District for three years is summarized as
Shinkansen effect in this study. The data for December 2015 are estimated by
the authors.
2 Hokuriku Shinkansen
Hokuriku Shinkansen is a railway rout to connect Tokyo and Osaka via Hokuriku
District at the entire line available. When the entire rail line of Hokuriku
Shinkansen is opened, it can assume the bypath function of Tokaido Shinkansen.
A part of the line (from Tokyo to Nagano) opened in ahead of schedule in
1997 due to Nagano Olympics (1998 Winter Olympics). It took 18 years until
Kanazawa opening of business. It was fifty years late than the opening of Tokaido
Shinkansen (opened in 1964, Tokyo to Shin-Osaka). The line was the world’s
first high speed railway. The operation of Shinkansen line connected to Tokyo
was desired earnestly for the tourism promotion and economic development
in the district. The passenger number increased triple than the previous one.
However, airplane passenger decreased sharply (more than 30%) and the num-
ber of the flights (including number of seats) was decreasing after the start of
the Hokuriku Shinkansen Line. There are three airports in Hokuriku, namely
Komatsu (Ishikawa Pref.), Noto (Ishikawa) and Toyama (Toyama). The maxi-
mum speed of the line is restricted to 260 km/hour for noise reduction and it is
investigated to speed up (300 km/hour) in consideration of noise environment.
1618 T. Oyabu and J. Nakamura
The maximum speed of each Shinkansen Line is indicated in Table 1. The speed
of Tohoku Shinkansen is the fastest (320 km/hour). Speedup is important in
future. Photograph of Hokuriku Shinkansen is shown in Fig. 1. Sky blue and
copper color of the body provide a calm atmosphere.
Table 1. Maximum speed for each Shinkansen in Japan
Shinkansen km/hour
Tokaido 285
Sanyo 300
Tohoku 320
Jyoetsu 240
Kyushu 260
Hokuriku 260
Hokaido 260
Fig. 1. Hokuriku Shinkansen
The route of Hokuriku Shinkansen is indicated in Fig. 2. The line will

be expanded to Shin-Osaka and connected to Tokaido Shinkansen via Koto.
Tokyo Ueno Omiya Takasaki
Itoigawa Joetsu-Myoko Iiyama Nagano
Kurobe-Unazuki-Onsen Toyama Shin-Takaoka
Shin-Osaka Tsuruga Kanazawa

(Under Planning)
Fig. 2. The route of Hokuriku Shinkansen

It will take about twenty years [5]. The Linear Chuo Shinkansen will be able to
run between Tokyo and Nagoya in just 40 min in 2027. The line will be expanded
to Osaka from Nagoya in 2045 and it will take 67 min between Tokyo and Osaka.
Strategy to take the interchange of the person is necessary, but it is also nec-
essary to remark the comfortability, safety and security besides speed. Japan
developed firstly a rapid transit railway in the world. Moreover, noise and envi-
ronmental assessment should be remarked. Hokuriku Shinkansen is a pivot of
the regional revitalization.
About 8.15 million passengers used Hokuriku Shinkansen in total since
started operation (for Mar. 14, 2015 to Jan. 20, 2016; about 10 months). The
number was about 3 times than heretofore and the handling incomes increased
by 33%, which was released by JR West. The change of passengers for Mar.
14 to Sep. 13 (six months) is indicated in Fig. 3. The data was measured for
Jyoetsu-Myoko to Itoigawa (working area of JR west). The values in March and
September are very small because they are in half month. The average for April
to August is about 0.8 million and the number of passengers a day each station
is represented in Table 2. The one for Kanazawa is less than twice the one for
Toyama.
Fig. 3. Number of passenger of Hokuriku Shinkansen (Jyoetsu-Myoko to Itoigawa),

which is measured for Mar. 14 to Sep. 13, 2015)
Table 2. Number of passengers a day in five stations
Station Baoding number/day

Kanazama 8930
Toyama 4860
Shin-Takaoka 1600
Unazuki 840
Itoigawa 400
3 Number of Guest in Hokuriku Spa Areas

Hot spring is an important tourism resource and the economic effect by the
guests is especially large in Hokuriku District. There are nine spa areas as main
hot spring resorts in the district (Fukui, Ishikawa and Toyama Prefectures). In
this study, Shinkansen effect was evaluated using the number of hotel guest in
the areas. The number is introduced by the guests of the facilities which are
joining Ryokan association (inn association). Only a few facilities do not join
the association and the unjointed number is less than 10%. The data is not
correspond to the local overall hotel guests. The nine survey target areas are as
follows:
Awara (Awara, Fukui Pref.), Awazu (Komatsu, Ishikawa Pref.),
Yamanaka, Yamashiro and Katayamazu (Kaga, Ishikawa Pref.)
Wajima (Wajima, Ishikawa Pref.), Wakura (Nanao, Ishikawa Pref.)
Yuwaku (Kanazawa, Ishikawa Pref.), Unazuki (Kurobe, Toyama Pref.)
Yamanaka, Yamashiro and Katayamazu are collectively called Kaga Onsen-
Kyo (Kaga three major spas). Unazuki (15 km) and Yuwaku (20 km) are close
to a Shinkansen station and it takes less than half hour from each station to
these spa areas by bus or car. It takes over one and half hours form each near
station for other areas by train or bus.
3.1 Kaga Four Major Spa Areas

Kaga three major spas (Kaga Onsen-Kyo) and Awazu spa are collectively called
‘Kaga Four Onsen’. Awazu is located in Komatsu city which is neighbor of Kaga
city. There are many visitors from Kansai area in the four spa areas. Osaka is the
center of Kansai area in which the area population is about 20 million. There are
also many visitors from Chukyo area (Nagoya) in which the area population is
about 15 million. The number of visitor is decreasing gradually due to the slump
of the Japanese economy in the areas and some hot-spring hotels discontinue
their business. A cheap hotel chain is also expanding. Kaga three major spas
(Kaga Onsen-Kyo; Yamanaka, Yamashiro and Katayamazu) had over 4 million
visitors in about 1990. It was the maximum value. The number increased a
little in 2015 than the one in 2014. It had about 2 million in 2015 due to the
Shinkansen effect. Awazu had 0.17 million visitors in 2015 and the number was
about 10% of the total of Kaga three major spas. These four spa areas (Kaga
four major spa areas) are developing a tourism strategy comprehensively. The
capacity of the four areas is about 15 thousands and there is a loop bus (named
CANBUS) which connects the four areas. The area is unlimited ride by showing
the ticket for the tourist. The ratio of departure area of the visitor to the areas
was as follows before the opening of the Hokuriku Shinkansen Line (in February
2015), namely Kansai area is 32% and Kanto area is 4%. The one for Kanto
approximately quadrupled and the one for Kansai was 23.5% in June 2015.
There was the Shinkansen effects obviously [8]. The effect is high in Yamashiro
spa area especially.
The characteristics of lodger number for three years (2013–2015) in the area
are shown in Fig. 4. The values in 2013 and 2014 are 1.73 and 1.74 million.
They are nearly equal. The percentages in the areas are as follows, Yamanaka:
25, Yamashiro: 39, Katayamazu: 27 and Awazu: 9%. The value for Yamashiro
is maximum. The mean total visitors for April to November in 2013 and 2014
were derived and compared with the one for the same term in 2015 after the
Shinkansen opening of business. It increased by 16% after the opening. The
percentage is thought as the Shinkansen effect. The total lodger number for
2015 in the area is about 2 million. The characteristic of the lodger in number
in the spa areas of Hokuriku District has a peak in August. The tendency is
the same as the characteristic in Japan. However, there are also small peaks in
March and November and the characteristics resemble each other depending on
tourism resources. The visitors from Kanto area (area population: 42 million)
increased to double after the start of the Hokuriku Shinkansen Line but the
tendency was the same.
Fig. 4. Characteristic of lodger number for Kaga four major spa areas
3.2 Awara Spa Area
Awara spa area is located in Awara City (population: about 30 thousand) of

Fukui prefecture and is called as a spa-resort of Kansai region. The city adjoins
Kaga City. Therefore, a unified tourism strategy is sometimes built as the area
of Kaga Onsen-Kyo altogether (including Yamashiro, Ymanaka, Katayamazu,
Awazu and Awara). There are few tourism resources except the hot spring
in Awara. The characteristics of lodger number for three years (2013–2015)
in Awara spa area is shown in Fig. 5. The increasing percentage for April to
November 2015 was 18%. The same deriving process was mentioned in the last
section. The percentage is larger than the one for Kaga four major spa areas.
The numbers in 2013 and 2014 were 0.61 and 0.69 million. It took a peak value
(87 thousand) in August. The characteristics for Figs. 4 and 5 closely resemble
to each other and it is a feature in Kaga Onsen-Kyo. The correlation diagram for
both figures is shown in Fig. 6. The correlation coefficient R is 0.91 and the data
Fig. 5. Lodger characteristics in Awara spa area for 2013–2015
Fig. 6. Correlation diagram of lodger number between Kaga four major spa areas and
Awara spa area
has a very high correlational relationship. If the coefficient is small, it can sup-
plement the seasonal food and hotel guest mutually in the region. It is desirable
for each facilities to have a specific tourism resources. And it is desirable that
R is small. Lodger characteristics in five major spa areas of Kaga Onsen-Kyo
should be varied due to a proposed strategy.
3.3 Yuwaku Spa Area
Yuwaku spa is located in Kanazawa city (population; 0.46 million) and is called
as an inner room of Kanazawa. Business firms often give parties there to entertain
their customers. There is Yumeji Takehisa Kan (a kind of museum) and many
Yumeji-fans visit there. There is Lake Gyokusenko which is suited in the inner
part of the hot-spring street. There is Himuro House (icehouse) by the lake and
the winter snow is stored in the house. The stored operation is carried out in
June every year. The snow ice was offered to Tokugawa family by Kaga Domain
in the Edo Period. The many citizens have the habit of eating Himuro manju
(steamed bun) to pray for good health at July 1. The area was the location of
Hanasaku-iroha (an anime drama) broadcast on TV in 2011 and there were many
visitors as the place of pilgrimage. Characteristic of lodger number in Yuwaku

spa area is indicated in Fig. 7. The fluctuation for Yuwaku area is smaller than
the ones for Kaga Onsen-Kyo spa areas which have a marked peak in August.
The Shinkansen effect of 38% is derived, which is the biggest growth rate in this
investigation. It is thought that Yuwaku is relatively near Kanazawa station as
the reason for the high growth rate. The guests drift to the area because the
occupancy rate for hotels near Kanazawa station is sometimes over 90%.
Fig. 7. Characteristic of lodger number in Yuwaku spa area
3.4 Wakura Spa Area

Wakura spa area is an entrance of the Noto sightseeing. It was said that the
opening of the hot spring or bath house was about 1,200 years ago. A fisherman
found a damaged white heron healing the wound in the hot spring. It was the
origin. And there is a traditional Japanese inn “Kagaya” in the area. Kagaya was
receiving repeatedly the best award as the most evaluated inn in Japan for 36
years until 2016. The award “100 selections of the best Japanese hotels and inns”
is recognized by Ryoko-shinbun (a trip newspaper publisher). It is said that the
hospitality (Omotenashi in Japanese) of Kagaya is number one in Japan, and
the inn is also a place of corporate training by many companies. A picture of
sending guest off is shown in Fig. 8. Wakura spa area is developing centered on
Kagaya. Characteristic of lodger-number transition is presented in Fig. 9. It has
the same trend as the one for Kaga Onsen Kyo, which is indicated in the figure
and the increase of 20% is derived as Shinkansen effect. The decrease in number
of the guests is relatively small in winter.
3.5 Wajima Spa Area

There are some Japanese inns and Minshukus (private home that runs inn pro-
viding room and board) in Wajima area but the capacity of guests is small.
The number of guests was about 0.13 million in a year before the Hokuriku
Fig. 8. Hospitality of Kagaya (send-off)
Fig. 9. Lodger number as a function of month in Wakura spa area
Shinkansen Line started operation. It became 0.17 million in 2015 and the
increase of 30% is recognized. The increase rate of lodger number for April to
November in 2015 was 33%. The number of visitors for Wajima morning market
increased by 30%. The influence of the serial TV drama “Mare” produced by
NHK is also thought. Wajima was used as a location for making the drama.
It was broadcasted for March to September in 2015. There were really many
visitors to Osawa area in Wajima City. Magaki (board fence) which is made
up of bamboos, is famous in the area. Magaki protects a house from a strong
winter wind (sea breeze). Lodger-number characteristic for Wajima spa area is
presented in Fig. 10. The number increased after the broadcasted TV drama
(May). The details for the increased number (30%) are as follows. Shinkansen
effect for Wakura spa nearby Wajima is 20%. So Shinkansen effect is estimated
by 20%, TV drama effect is 10%. It is difficult to judge the effect clearly because
there is a synergistic effect.
3.6 Unazuku Spa Area

Unazuki is located to the east of Toyama prefecture and is the biggest hot spring
resort in the prefecture. The name of the closest Shinkansen station to Unazuki
is ‘Kurobe-Unazuki Onsen’. There are a private railway from the station to
Unazuki spa resort and it takes about 20 min by the train. There are a truck
Fig. 10. Lodger-number characteristic for Wajima area
train in Unazuki to visit Kurobekeikoku (canyon). Hundreds of thousands of

people visit the canyon annually. The scenery is wonderful especially for the
autumn leaves. As for the result, the increasing rate of lodger number for April
to November (after the Shinkansen opening of business) is over 30% and the
shortage of human resources for some inns are remarkable in the area due to
the increase of the room occupancy rate. The transition of lodger number is
presented in Fig. 11. The increase after April is remarkable because the area is
near from the Shinkansen station as well as Yuwaku. It is necessary to hold an
attractive event in June and equalize the lodger number over whole year.
Fig. 11. Lodger number for Unazuki spa resort
3.7 Trend of Lodger Number Due to Shinkansen Started Operation
There were various kinds of advantages and disadvantages due to the Hokuriku
Shinkansen started operation. The start of the operation spreads to over the
entire area of Hokuriku (three prefectures) but there are the light and shade of
the effect [8]. The movement of person is sucked into an area where an attrac-
tive event is held. This phenomenon is called “straw effect”. Some points to be
improved were listed by visitors to the area. The points are as follows.
• It is hard to make the reservation of the hotel.

• Hospitality is light due to the labor shortage.
• Rising of the meal price.
The increasing rates of lodger number for six hot spring areas as described
above, is indicated altogether in Table 3. The rates of three areas (Yuwaku,
Wajima and Unazuki) are over 30%. Yuwaku and Unazuki are near the Shinkansen
stations, and Wajima is the filmed place for the drama (Mare). Consequently, the
one for Wajima will decrease gradually. The correlation diagram for the lodger
number between Yuwaku and Unazuki is derived and indicated in Fig. 12. The
correlation coefficient R is 0.75 and the value is high. It is thought that there are
some different types of tourism resources for the visitors in the area, for example
autumn color of canyon, ski slope and cheap accommodation fee.
It was thought that the visitor number to Wajima increased by 10% due to the
place for the TV drama “Mare”. There was a synergistic effect with Shinkansen
began the operation (increased by about 20%) and many persons visited Wajima
to see the scene sites of the drama. Tokyo where the area population is 42 million,
is the starting station of the Hokuriku Shinkansen. There is high potential to
increase the number when the access becomes easy. The increasing rates of the
five spa resorts (Kaga three major spa areas, Awazu and Awara) were almost
20%. The one for Wakura was 21%. The access for Wakura is relatively good
because Wakura is connected directly to Kanazawa by JR railway and there is
also a bus route. The rate is higher than the ones for the above mentioned five
areas.
Coefficients of variation (cv: standard deviation/the mean value) for six spa
resorts are summarized in Table 4. The values in 2014 and 2015 are represented.
The ones for Wajima and Unazuki are exceed 0.3, namely the fluctuations for the
lodger number of every month are larger than the other areas. A variation of the
occupancy rate for each hotel is large, namely maintenance cost for the facilities
including labor cost becomes higher. The coefficient for Wajima becomes large.
Fig. 12. Correlation diagram for the lodger number between Yuwaku and Unazuki
Table 3. Increasing rates for six hot spring areas in Hokuriku district
Hot spring Kaga Awara Yuwaku Wakura Wajima Unazuki

Rate 0.16 0.18 0.38 0.21 0.33 0.30
Table 4. Coefficients of variation for six spa resorts in 2014 and 2015
Kaga Awara Yuwaku Wakura Wajima Unazuki

2014 0.20 0.26 0.20 0.19 0.23 0.36
2015 0.16 0.20 0.17 0.20 0.31 0.38
4 Visitor Number for Other Main Facilities
There was a tourism spot in which the visitor number increased remarkably. It
is Asakura remains in Ichijodani and located in Fukui prefecture. A CM movie
was filmed in the spot and about 1.1 million persons visited the spot [7]. It is nec-
essary to construct some facilities for visitors, namely parking place, rest house
and souvenir shop. The elderly interested in history are also visiting. The service
introducing the situation in ancient or olden times using a tablet is carried out.
Kenrokuen (one of the three largest gardens in Japan) is the spot which
has the most visitors in Ishikawa prefecture and the number is 2.9 million in
2015 (increased by 45.5% than the one in 2014). There was 2.25 million visitors
(increased by 81% compared with the previous year) in Kanazawa Castle Park
which is adjacent to Kenrokuen. The number of foreign people visited Kenrokuen
was 0.292 million (increased by 24.4% compared with the previous year). The
visitors from Taiwan occupied by 46.5%. The one from Hong Kong was 10% and
the one from USA was 5.3% (China was 5%). The rate of the one from Korea
was few [6].
There is the Noto Railway Corporation in Noto area where Wajima and
Wakura (in Nanao city) are located. It is a third sector railway company. The
service is operating between Wakura and Anamizu (about 40 min by train), and
JR operates a train between Nanao and Wakura. Many group tourists use the
Noto Railway Corporation after the Hokuriku Shinkansen line started its opera-
tion. Some events were hold in the train, for example famous sweets-service and
special guide by a conductor in the train. The events and TV drama “Mare”
affected to the increase of the visitor number. The number was 61 thousand in
2015 (increased by 73% compared with the previous year). The visitor number
form Kanto area were 42 thousand (70%). It is very large percentage compared
with the ones for other areas [10]. The number for Tateyama Kurobe Alpine
Route was 0.997 million in 2015 (increased by 10% compared with the previ-
ous year). There was a big snow wall (named Yuki-no-Otani in Japanese) in the
route. One million visitors were expected as a goal in 2015. It was slightly few. As
the details, the number of Japanese was 0.782 million and the one for foreign vis-
itors was 0.215 million (increased by 12%). The number of visitors from Taiwan
(132 thousand) is the biggest and the second is Hong Kong (25 thousand) fol-
lowed with Thailand (17 thousand) and Korea (16 thousand). Higher two coun-
tries are the same as Kenrokuen [9].
5 Conclusion
It is necessary to examine the visitor number after the Hokuriku Shinkansen
line started its operation and to apply the result to a future strategy in the
district. It is said that the number increases remarkably in the first year after
the opening. However, it is also said that the effect does not continue after the
second year. This phenomenon is called “Shinkansen jinx”. Some proper and
attractive events are required to break the jinx every season and to secure the
repeaters. Peculiar local hospitality (“Omotenashi” in Japanese) is also necessary
to secure the repeaters. Some travel agents say that the traffic infrastructure
and the accommodation cost (lodging and transportation expenses) are more
important for repeaters. Some regional inhabitants have the objection against the
increasing the number of visitors sharply because the community is in disorder.
It is necessary to constitute the area in which the habitants and visitors can live
together. The action of “Civic Pride” is also effective.
In this study, the fluctuation of the visitor number in spa resort areas is
examined as a Shinkansen effect. It is important to understand the fluctuation
and is also necessary for forming the social environment of the area in future.
The visitor number of nine spa areas in Hokuriku district were investigated.
As a result, it becomes obvious that the effect of visitors spreads to Hokuriku
district whole. Especially, the lodger number of the areas near the two Shinkansen
stations increased by about 30%. The Noto area was over 20% and Kaga area was
less than 20%. The lodger number of the spa area near the station remarkably
increased due to the beginning of the operation. It is necessary to review the
comfortable Omotenshi (hospitality) to the guests and the staffing management
in the high increasing areas. It is also necessary to level the fluctuation on the
characteristic of the lodger number by holding events. There are the dispersion
in the lodger number every month. The leveling leads to the effective utilization
of the resources, especially capacity and the meal supply.
In future, it needs to be examined on foreign visitors (inbounds) who have an
enhanced economic effect. It will also contribute Japanese globalization. Increas-
ing the inbounds is one of Japanese strategies.
Acknowledgements. A part of the data is provided by Mr. Shinya Inoue (Director

of The Yamanaka Onsen Inns Association). The authors gratefully acknowledge to him
for his kindness and effective discussions.
References
1. Bureau JT (2013) Viewpoint and practice of the management in tourism spot,
Maruzen
2. Insatsu N (2015) White paper on aged society. Cabinet Office (in Japanese)
3. Oyabu T, Nakajima M, Ohe Y, Hosono M (2013) Tourism and regional develop-
ment. Kaibun-do (in Japanese)
4. Kagaku Sha K (2015) Introduction to tourism informatics. Society for tourism
informatics in Japan (in Japanese)
5. Shinbun H (2016) The demand of Obama-Kyoto. Morning edition (in Japanese)
6. Shinbun HC (2015) Facilities user of Ishikawa prefecture becomes the most. Morn-
ing edition (in Japanese)
7. Shinbun HC (2015) Tourists exceeded one million. Morning edition (in Japanese)
8. Shinbun HC (2015) Tourists increased remarkably in each spot. Morning edition
(in Japanese)
9. Shinbun HC (2016) 0.29 million foreign visitors in Kenroku-en garden. Morning
edition (in Japanese)
10. Shinbun HC (2016) The advancement of the achievement (the aim 60,000 visitors).
Morning edition (in Japanese)
11. Suda H (2009) Tourism. Gakugei-shuppan (in Japanese)
Charging Infrastructure Allocation for Wireless
Charging Transportation System
Min Seok Lee and Young Jae Jang(B)
Korea Advanced Institute of Science and Technology,

Daejeon, Republic of Korea
yjang@kaist.edu
Abstract. The dynamic wireless charging electric vehicle (DWC-EV)

is the most developed form of EVs which use wireless power transfer
(WPT) technology. The vehicle can be charged using a wireless charger
embedded under the road, and an extremely competitive advantage of
DWC-EV is that it can be charged while on the move. Thus, this system
allows EVs to overcome the range limitation, long charging time, and
heavy battery weights. In DWC-EV system, the battery and charging
infrastructures called power tracks are major elements to operate vehi-
cles, and these factors occupy large portion of initial investment cost for
establishing the system. This paper suggests a mathematical model to
optimally find the battery capacity and the allocation of power tracks
in a multiple-route public transportation system. In addition, genetic
algorithm (GA), which is one of widely used metaheuristics, is proposed
as a solution approach to solve large and complex problems. Since chro-
mosome design determines the format and information of chromosomes,
it is crucial to the quality of the solutions and the performance of GA.
We suggest two chromosome design methods and the methods to deal
with infeasible chromosomes. A number of numerical experiments for
comparing solutions and computation time are conducted, and we use
the optimal solutions suggested by CPLEX solver as criteria. We also
show that the proposed GA design method can be useful and effective
compared to the exact solution approach based on the mixed integer
programming (MIP).
Keywords: Dynamic wireless charging · Electric vehicle · System opti-

mization · Genetic algorithm
1 Introduction
Electric vehicles (EVs) have attracted international attention as an alterna-
tive to internal combustion engine (ICE) vehicles, with the aim of reducing the
dependence on petroleum and adverse environmental effects. Many countries and
authorities have devoted much research and development to EVs, and numerous
policies have also been suggested with respect to electric transportation systems
Charging Infrastructure Allocation 1631
[15]. The number of EV users has grown in accordance with the global change,
and in particular, the market share of plug-in EVs (PEVs) becomes larger. In
addition, wireless power transfer (WPT) technology for EVs has recently been
introduced to overcome the disadvantages of the existing PEVs, for example
short travel distance and range anxiety.
The dynamic wireless charging electric vehicle (DWC-EV) is the most
advanced form of EV, and uses a wireless charger embedded under the road,
which charges the vehicle when it moves or stays on it. As it is possible to
charge during operation, the driver does not need to stop for a long time to
recharge the battery. The battery size can be reduced as the charging infrastruc-
ture can supplement the electric power throughout the operation. The online
electric vehicle (OLEV) is a DWC-EV developed by the Korea Advanced Insti-
tute of Science and Technology (KAIST). In DWC-EV system, however, the
charging infrastructure and the battery account for the largest proportion of the
initial investment cost. InGumi city, about 80% of the total cost was allocated to
these two key factors for setting the system [3,12]. In other words, the investment
cost is mainly determined by battery capacity and the allocation of the charging
infrastructure, so we focus on these factors in this study. To install a DWC-EV
system at a reasonable cost, an optimum level of charging infrastructure and an
estimation of battery capacity are required.
The purpose of this study is to identify the optimal allocation of charging
infrastructure and battery capacities for the stable and economic operation of
the DWC-EV system which is applied to public transportation buses. We focus
on the minimization of the initial investment cost, not the operational cost. In
this study, we suggest a mixed integer programming (MIP) model considering
a multiple-route public transportation system. To find optimal solutions for the
MIP model, specific algorithms would typically be used, but obtaining optimal
solutions to the complex problems with these algorithms is difficult. In this
case, genetic algorithm (GA), a widely used metaheuristic, can be useful for
finding optimal or near-optimal solutions within a short time. Therefore, we
design a GA procedure to solve the multiple-route problem. We analyze the
overall performance of the solutions suggested by GA and also compare it with
the optimal solutions obtained by solving MIP problem.
Several studies have been conducted on the system design for EVs, and some
focus specifically on infrastructure planning and economic analysis of hybrid
EVs or PEVs [4–6,9,14]. These studies generally focus on the deployment of
charging stations or determination for the capacity of the charging stations.
Also, recent studies related to the system design of DWC-EVs have dealt with
the minimization of total cost by determining the battery size and the positions of
power transmitters [7,8,8,10,11,13]. These studies have considered single route
cases or do not suggest practical methods to solve large and complex problem.
In actual, the optimization approach for these cases is not appropriate if the
DWC-EV application is planned at the city or multiple district scale.
The remainder of this paper is organized as follows. Section 2 presents the
basic characteristics of a DWC-EV system, optimization issues, and several main
1632 M.S. Lee and Y.J. Jang
assumptions used in the optimization model. Section 3 describes the optimization

modeling of the DWC-EV system, and Sect. 4 introduces the GA process and
chromosome design methods used to suggest solutions for our model. Section 5
solves example problems using GA, and the quality of the solutions are verified
by comparing them with the optimal solutions suggested by the CPLEX solver.
Section 6 concludes and suggests directions for future study.
2 Description of the DWC-EV System

2.1 The Characteristics of the DWC-EV System
The DWC-EV system has two main parts: the vehicles and the charging
infrastructure. The vehicle part mainly consists of a pick-up device for charging,
a battery, and a motor for moving vehicles. The charging infrastructure is com-
posed of an inverter and an inductive power cable, which is installed beneath the
road. In this paper, we refer to the charging infrastructure as the power track.
When a DWC-EV operates on the inductive cable, the power is delivered to the
motor or the battery through the regulator. Figure 1 illustrates the main compo-
nents of the DWC-EV system. The wireless power transfer system is described
in detail by Ko and Jang [10] and Ahn et al. [2].
The battery and power tracks are essential elements of the DWC-EV system
and constitute most of the total initial investment. It is important to find an
appropriate balance between the battery capacity and the allocation of power
tracks, but it is not economically viable to install power tracks under whole
routes or to equip vehicles with excessively large battery capacity. Two extreme
examples can help to explain the relationship between battery and power track. If
a vehicle has a large battery that is adequate for operation over the whole route,
it does not need to collect electric power during operation, and if it is charged at
the base station just before operation there is no problem. The battery will here
constitute most of the investment cost. If power tracks are installed under the
whole route, the vehicle can travel with a very small battery, and the cost of the
power track would be much higher than the battery cost. From the examples, we
know that the battery cost and the power track cost have a trade-off relation.
Therefore, determining the battery capacity and the allocation of the power
tracks should be verified precisely in the commercialization of the DWC-EV
system. The purpose of the optimization is to minimize the initial investment
cost of the DWC-EV system.
2.2 Assumptions of the DWC-EV System
In this study, we focus on the DWC-EV system applied to public transportation

buses. Buses must stop at several stations on each route during operation, and
begin and terminate their service at the base station. Before modeling the opti-
mization problems, we set several assumptions, which reduce the complexity of
the problems and reflect the practical issues.
Fig. 1. Major components of DWC-EV
First, the energy consumption between two stations is given. We approximate

the consumption is linearly proportional to the distance. The second assumption
is that power tracks can be installed only at station areas. In the DWC-EV
system, the level of battery charging depends on how long the bus stays on the
road where the power track is installed, so clearly installing power tracks at
locations where buses stay for a long time is advantageous. Another assumption
is that the amount of charging at stations is linearly proportional to the length
of power tracks.
The DWC-EV system also includes operational assumptions. First, if several
buses travel on an identical route their battery capacities will the same, but
battery capacity can differ according to the route. Buses are also assumed to
be fully charged at the base station during resting time. We do not consider
the installation cost of charging facilities at the base station in the optimization
modeling. We will consider only the cost of facilities used during operation.
3 Optimization Modeling
3.1 Decision Variables
We consider multiple routes for public transportation systems in our study. The
investment cost will change according to the number and the length of power
tracks, so we must identify the number, length, and location of power tracks.
Suppose that there are a total of N stations and M routes in the system. We
define the integer decision variable Xs as the length of the power track installed
at station s (s = 1, · · · , N ). Also, let ys be the binary variable for whether the
power track is installed at station s or not. The battery capacity of a vehicle in
route r (r = 1, · · · , M ) is denoted by Erc . Figure 2 describes decision variables
for a multiple-route optimization model.
Fig. 2. Decision variables for a multiple-route optimization model
3.2 Objective Function

The objective function is composed of the battery cost and the infrastructure
cost. Let cb be the battery cost per unit capacity. Thus, we can represent the
battery cost of a vehicle on route r as cb Erc . The power track related cost is com-
posed of a fixed cost and a cable cost. The fixed cost includes inverter installation
and setup costs. The length of power tracks has a direct influence on cable cost.
The longer the power track is, the more cost is incurred. We define cf and cv
as the fixed cost and the cable cost, Nrespectively. The power track installation
N
cost at station s is represented by s=1 cf × ys + s=1 cv × Xs . If there are kr
vehicles on route r, the objective function of the multi-route system is expressed
as follows:
N N

cb × kr × Erc + cf × ys + cv × Xs .
r∈R s=1 s=1
3.3 Constraints
In most cases, battery manufacturers set maximum and minimum energy levels
for battery stability, so the energy level is recommended to remain within a
fixed range. The maximum and minimum levels are represented by multiplying
a constant and the battery capacity. These equations are as follows:
Eru = nu × Erc ,
Erl = nl × Erc ,
0 ≤ nl ≤ nu ≤ 1.
The number of stations on route r is denoted by N (r), and k th station on

route r is represented by I(r, k). Let Er,k be the energy level of the battery at
the moment when the bus leaves k th station on the route r. The battery will
not have a higher energy level than Eru even if charged for a very long time.
Surplus energy will then result, which cannot be stored in the battery. We call
this remaining energy excess energy. The excess energy quantity at k th station
of route r is denoted by variable qr,k . The energy level Er,k is written in the
following form:
Er,k = Er,k−1 − dr,k−1 + p × XI(r,k) − qr,k r ∈ R, k = 1, · · · N (r),
where dr,k means the energy consumption between k th station and k + 1th sta-
tion on route r, and p is the amount of energy charge per unit length of the
power track.
The system must maintain the battery level within the upper and lower
bounds while in operation.
Er,k ≤ Eru r ∈ R, k = 1, · · · N (r),

Er,k − dI(r,k) ≥ Erl r ∈ R, k = 1, · · · N (r).
The DWC-EV bus is fully charged when it leaves a base station.
Er,1 = Erc r ∈ R.
The length of the power track cannot exceed the length of the station area.
Installing very short power tracks is not economically viable. The following
inequality constraint states the upper and lower bounds of the length of the
power track.
ys × X min ≤ Xs ≤ ys × X max s ∈ S,
where X min and X max are the minimum and the maximum length of the power
track.
We develop the following mathematical model for the optimization problem
using the objective function and constraints explained above.

min kr × cb × Erc + cv × Xs + cf × ys (1)
r∈R s∈S s∈S
s. t. Er,k = Er,k−1 − dr,k−1 + p × XI(r,k) − qr,k r ∈ R, k = 2, · · · N (r) (2)
Er,k ≤ Eru r ∈ R, k = 1, · · · N (r) (3)
Er,k − dI(r,k) ≥ Erl r ∈ R, k = 1, · · · N (r) (4)
Er,1 = Eru r∈R (5)
ys × X min
≤ Xs ≤ y s × X max
s∈S (6)
ys ∈ {0, 1} s ∈ S (7)
Xs and Erc are zeros or positive intergers s ∈ S. (8)
4 Genetic Algorithm
Several algorithms are used to solve MIP problems, such as the branch and
bound algorithms. However, these algorithms cannot suggest optimal solutions
for large-scale problems in a timely manner. In this case, heuristic methods or

metaheuristics are more useful for solving the problems. The genetic algorithm
(GA) is a widely used metaheuristic, and we use it as our main solution app-
roach in this study. Chromosome design determines the format of chromosomes,
and the chromosome’s information is crucial to the quality of the solutions and
the performance of GA. We suggest two chromosome design methods for our
problem.
4.1 Chromosome Design with Battery Capacity

(1) Format of chromosome
The first chromosome is divided into two parts. Figure 3 describes the format
of the chromosome, and the first part represents the battery capacities. The
value of each gene represents the battery capacity corresponding to each route.
The second part of the chromosome expresses the length of the power track at
each station. Therefore, this chromosome format contains information about the
decision variables Erc and Xs .
However, infeasible chromosomes can be produced with this format. Specifi-
cally, several chromosomes created during generation will not satisfy the energy
dynamics constraint introduced in Eqs. (3) and (4). The infeasibility of the chro-
mosome can be dealt with in two ways; either by giving a large penalty to the
chromosome, such as very bad fitness value, or by applying repair mechanism to
make it feasible [1]. We deal with these two methods in numerical experiments.
Fig. 3. The structure of a chromosome with genes of battery capacity
(2) Repair mechanism

From a generated chromosome, we calculate the energy level of the battery
using Eqs. (2) and (5) at all stations on multiple routes. We suppose that we start
the investigation from kth station of route r. We calculate Er,k − dr,k , which is
defined as remaining energy. From now on, we assume that the lower bound
of energy level Erl is zero for convenience. If the remaining energy is greater
than or equal to zero, we go to the k + 1th station and calculate the remaining
energy at the k + 2th station again. If k th station is the last station of route r,
the mechanism starts from the first station of route r + 1. The mechanism will
terminate after the investigation of the last station on the last route.
However, the remaining energy Er,k −dr,k may result, which is less than zero.
In this case, we calculate the additional battery capacity and the length of power
tracks, respectively, to supplement the shortage of energy. The additional battery
capacity or the length of the power track can replenish the energy amount dr,k −
Er,k , and then we calculate the corresponding cost for each additional element.
If the additional battery cost is cheaper than the additional power track cost, we
add the battery capacity to the existing battery capacity of route r instead of the
power track. In the opposite case, we attach the additional power track to the
existing power track. However, if the length after adding the additional power
track is longer than the upper bound, the rest of the additional power track is
converted to battery capacity. This generated battery capacity is then added to
the existing capacity. When the length after adding the additional power track
is shorter than the lower bound, we choose additional battery capacity instead
of power track.
Next, we modify the energy level Er,k , and go to the k + 1th station and
calculate the remaining energy at the station again. The mechanism should start
from the first station of the first route and finish after the process of the last
station on the last route. The new chromosome created by the repair mechanism
will satisfy all constraints in the MIP model.
4.2 Chromosome Design Without Battery Capacity
(1) Format of chromosome

In this format, we allocate only the power track part of the previous format
to a chromosome. Each gene of a chromosome expresses the length of the power
track at each station. Therefore, the decision variable Xi in the MIP model
corresponds to the ith gene of the chromosome. Figure 4 describes the structure
of the chromosome. Although the chromosome does not have information about
battery capacity, it is possible to calculate the minimum battery capacity to
operate the DWC-EV without depletion of energy based on the gene values
of power tracks. Here, there is no infeasible chromosome in terms of energy
dynamics constraints, so the repair mechanism is not needed in this format, and
we can reduce the waste of battery capacity as we calculate the minimum battery
capacity for each chromosome. However, we must calculate the battery capacities
for all chromosomes created until the last generation, so more calculation time
will be spent on the GA procedure than the previous format.
Fig. 4. The structure of a chromosome without genes of battery capacity
(2) Calculation of battery capacity

Let Wr,k be the required energy level of the battery at the moment when the
bus arrives at k th station on route r. We assume that Wr,N (r)+1 is zero, which is
the lower bound of the energy level, because station N(r)+1 represents the base
station of route r. When the bus departs from N (r)th station on route r, it needs
at least dr,N (r) + Wr,N (r)+1 . If a power track of XI(r,N (r)) in length is installed
at N (r)th station, Wr,N (r) would be max 0, dr,k + Wr,k+1 − p × XI(r,k) . In this
manner, it is possible to explain other Wr,k . Therefore, Wr,k can be expressed
as follows:

Wr,k = max 0, dr,k + Wr,k+1 − p × XI(r,k) , r ∈ R, k = 1, · · · N (r) (9)
To calculate the battery capacity Erc on route r, we define bat(k) as the

required battery capacity to move from k th station to k + 1th station on route
r. The battery will consume dr,k until arriving the next station, and the energy
state at k + 1th station should be Wr,k+1 . Thus, batr (k) is the sum of dr,k and
Wr,k+1 .
batr (k) = dr,k + Wr,k=1 , r ∈ R, k = 1, · · · N (r). (10)
The battery capacity Erc is the maximum value among all batr (k) to cover
all movements between two stations. Therefore, Erc is represented as follows:
Erc = max (batr (k)), r ∈ R. (11)

k=1,···N (r)
4.3 GA Settings and Procedures
Fitness function is an indicator of the quality of the solutions in GA. The objec-
tive function of the mathematical model is the fitness function of our problem.
The purpose of the GA procedure is to find a solution that minimizes the fit-
ness value. Before starting GA procedure, we generate the initial population to
enhance the GA performance. The chromosome in the initial population is gen-
erated using the optimization results of each single-route problem. Most single
route solutions are easily obtained using the algorithm for MIP model.
Selection function finds parents for the reproduction of new chromosomes
using the fitness value. In our case, the selection function is based on the roulette
wheel selection. After that, a portion of offspring is reproduced by the crossover
function, which generates a new chromosome by combining a part of each parent.
First, the crossover function creates a random binary vector that has a length
identical to a chromosome. From the first parent, genes placed at the location
1 in the binary vector are selected. Those placed at the location 0 are selected
from the second parent. The selected genes from the two parents generate a new
chromosome.
The mutation function prevents the solution from converging to a local opti-
mum, and is used to find various solutions in the searching space. In our case, it
generates a random vector with length the same as that of the chromosome, and
the random vector is added to a selected chromosome. If the new chromosome
satisfies the bounds and constraints, it enters the next generation, and we can
call it a mutated chromosome. We also set the elite reproduction which means
that some of outstanding chromosomes in terms of fitness value must survive to

the next generation.
In general, GA stops the procedure when several conditions satisfy the stop-
ping criteria. In our case, the number of generations is selected as a stopping
criterion. When the algorithm carries out the last generation, GA will find and
suggest the best solution in that final generation.
5 Numerical Experiments for a Multiple-Route

DWC-EV System
In this section, numerical experiments for a multiple-route model are handled,

and solution approaches introduced in the previous sections are used to solve
the problem. We consider 120 stations in Gumi city, which are used for three
real routes, 1-1, 8, and 17, in the public transportation system. We add three
hypothetical routes to these existing routes, which also use some parts of the 120
stations. Thus, there are 6 routes in the example problem. We assume that the
energy consumption between two stations is constant, and the data is already
given. Several numerical experiments are conducted as the number of the routes
increases from 1 to 6. From routes 1 to 6, there are 47, 32, 46, 36, 36, and 36
stations, respectively, and some of those routes share the same stations.
Table 1. The comparison of average computational time (in seconds)
Description Values
Unit battery cost (cb ) 800
Fixed cost (cf ) 20,000
Unit cable cost (cv ) 100
Charging rate (p) 0.5
The number of buses on each route (kr ) 5
The maximum length of the power track (X max ) 30
The minimum length of the power track (X min ) 10
The MIP model for the DWC-EV system is solved using CPLEX 12.5 embed-
ded in a JAVA environment, and the GA coded with the MATLAB Global Opti-
mization Toolbox is used to solve the problems. There are three approaches in
the GA. First, we apply the chromosome format with battery capacity, which
does not conduct a repair mechanism. This approach is defined as GA method 1.
In this method, infeasible chromosomes should have infinite fitness values. The
second approach is GA method 2, and carries out the repair mechanism to the
same format. The last approach, GA method 3, uses the chromosome format
without the battery capacity. Other GA procedures are applied identically to all
approaches. The constant values for the numerical experiments are summarized
Table 2. Parameter settings for GA
Description Values
Population size 1,000
The number of generations 1,000
Crossover fraction 0.84
Mutation fraction 0.01
Elite fraction 0.15
in Table 1. In this example, the maximum and the minimum energy level of the
battery are the battery capacity Erc and 0, respectively. Several parameter set-
tings for the GA procedures are defined in Table 2. These values are determined
from a number of test experiments to find the better solutions.
We select the routes in a regular sequence starting from route 1; for example,
if we select 3 routes, we choose route 1, 2, and 3. The total cost obtained from
CPLEX and three GA approaches is listed in Table 3 according to the number
of routes.
Table 3. The comparison of total cost between solution approaches
Algorithm CPLEX GA method 1 GA method 2 GA method 3

Description Obj. value Best Average Best Average Best Average
One route 325,100 325,100 357,139 325,100 345,759 325,100 326,934
Two routes 611,900 623,300 680,141 614,900 640,350 611,900 612,215
Three routes 967,400 1,031,200 1,098,603 982,900 1,032,737 975,100 977,649
Four routes 1,074,000 1,161,500 1,225,973 1,117,300 1,154,917 1,105,200 1,108,030
Five routes 1,266,600 1,476,900 1,554,957 1,373,500 1,446,741 1,314,500 1,321,851
Six routes 1,366,000 1,595,400 1,684,494 1,471,400 1,535,578 1,416,400 1,438,070
The GA results of Table 3 are obtained from 100 experiments for each case.
The results of CPLEX are objective values, which are calculated by optimal
solutions, so it is impossible to have costs lower than the results. These values
can be used as criteria to assess the performance of the GA results. In the case
of one route, the best value for each GA method is identical to the objective
value, but average values are different from each other. GA method 3 suggests
the smallest average value of all GA methods, while GA method 1 has the worst
average value. In the case of two routes, the best value of GA method 3 also has
the same value as the objective value, but GA methods 1 and 2 do not suggest
the objective value, and GA method 3 still draws the best values between the
three methods in both the best and average values. From the case of three to
six routes, no GA methods suggest optimal results, and the performance of
solutions in GA method 3 is always better than the others, in the same manner.
GA method 2 takes second place in every case, and GA method 1 proposes the
worst results.
Though methods 1 and 2 use the same format of chromosome in the GA
procedure, there are huge differences in the results, depending on whether the
repair mechanism is used or not. In GA method 1, many infeasible chromosomes
can be created during the 1,000 generations, so feasible chromosomes that do
not have good fitness value in comparison with the optimal results will easily
survive. In addition, the speed of improvement in the chromosomes slows, as the
population generally consists of poor chromosomes. The quality of the best chro-
mosome in the final generation will consequently be worse than in other meth-
ods. However, if the repair mechanism is applied to GA, it is impossible to have
infeasible chromosomes, so all chromosomes can have specific fitness values that
are not infinite. Unlike GA method 1, in which feasible chromosomes compete
with many infeasible chromosomes, GA method 2 creates competition between
comparable chromosomes, and the outstanding chromosomes will survive. Some
infeasible chromosomes may also be changed to remarkable chromosomes during
the repair mechanism. Therefore, we claim that GA method 2 can suggest better
solutions than GA method 1.
GA method 3 provides the best solutions among all GA methods because of
the format of chromosomes, and the fact that battery capacity plays an impor-
tant role in the total cost. Figure 5 represents the total battery cost of the best
chromosomes in all GA methods. GA method 3 suggests the smallest battery
cost in all cases except for the one route case. GA method 3 can have an advan-
tage in the battery cost as the battery capacities in the chromosome format are
calculated to have the minimum values without depletion when it is driving. In
contrast, other methods randomly allocate battery capacities, regardless of the
gene values in the power track part.
Fig. 5. A comparison of total battery cost in the GA methods

The power track costs of the best chromosomes in all GA methods are shown
in Fig. 6. In contrast with the battery cost, the power track cost has the biggest
value in GA method 3. The power track costs are identical in the different meth-
ods for the one route case, in the fourth case GA method 3 has a slightly smaller
value, but for the other cases method 3 has the largest. From the perspective of a
trade-off relationship, the results are reasonable as a small amount of power track
is allocated to the solution with large battery capacity, and a large amount of
power track is provided to the solution with small battery capacity. By compar-
ing battery and power track costs, we confirm that the difference in the battery
cost has a huge effect on the trend of the total cost.
Fig. 6. The comparison of total track cost in GA methods
A comparison of the average calculation time in CPLEX and all GA methods

is given in Table 4. The calculation time of GA method 1 and GA method 2 very
slightly increases as the number of routes increases. The results of GA method 2
are slightly higher than those of GA method 1 because of the repair mechanism.
The repair mechanism does not take up much time during the GA procedure as
not all chromosomes are infeasible, and the modification is carried out for only for
several genes in the infeasible chromosome. However, the results of GA method
3 indicate that the calculation time for each case is longer than for methods 1
and 2. The slope of the graph is also the steepest of the three methods. In the
case of six routes, GA method 3 takes 381.15 seconds to suggest a solution, a
calculation time about 5 times that of methods 1 and 2, as it must calculate
the minimum battery capacity for all chromosomes in every generation, and the
tasks for calculating the battery capacity also increase with the increase in the
number of routes.
In the first and second cases, CPLEX is faster than all GA methods, but the
time then increases exponentially from the third case, and for the six routes it
Table 4. The comparison of average computational time (in seconds)
Algorithm CPLEX GA method 1 GA method 2 GA method 3

One route 0.74 70.62 70.96 110.78
Two routes 1.26 71.4 72.14 156.05
Three routes 409.89 72.21 75.21 227.67
Four routes 619.68 77.13 80.36 274.96
Five routes 902.86 78.98 81.13 321.9
Six routes 40828.58 80.96 83.86 381.15
takes over 11 hours. This trend shows the characteristics of the exact algorithm,
such as the Branch and Bound algorithm, which means that the exact algorithm
is more sensitive to the size of problems compared to the GA. From the numer-
ical experiments, we identify that GA methods 1 and 2 have an advantage in
calculation time compared to GA method 3, but the quality of solutions from
method 3 outperforms the other two. The GA can suggest reasonable solutions
in a timely manner, but does not always guarantee optimal solutions. Therefore,
the GA can be useful for large scale problems.
6 Conclusion
This study has introduced a new kind of dynamic wireless charging electric
vehicle (DWC-EV) system. A mixed integer programming (MIP) model, which
reflects multiple-route public transportation systems, was suggested to identify
the economic allocations of the charging infrastructure and the battery capacity
of the vehicle. A GA was introduced as a solution approach, and two approaches
for chromosome design were suggested. Numerical experiments for an exact algo-
rithm and a GA were conducted, and we analyzed the performance of three GA
methods by comparing the optimal results of CPLEX. We checked that the
repair mechanism for infeasible chromosomes improved the algorithm perfor-
mance. A chromosome design without battery capacity was found to provide
better solutions than other GA methods, though its calculation time was slow.
These solutions are thus almost optimal solutions.
In this study, several assumptions were applied to develop the mathematical
model. The assumption that the energy consumption between two stations is
linearly proportional to the distance is particularly impractical, as this does not
reflect the uncertainties of traffic or unpredictable situations. The assumption
that the amount of charging is linearly proportional to the length of the power
track was used to simplify the model. These assumptions should be proved quan-
titatively, and robust optimization can be used to consider uncertainties. The
procedure of the GA must be established to solve the problem efficiently and
effectively. After extending the limits of this study, we plan to apply our app-
roach to broader multiple-route systems on a city or national scale. The GA is
a simple approach and can be useful for discovering solution properties, and the
solutions obtained can be reference points for heuristics and other algorithms,
which will be examined in future research.
References
1. Abdelmaguid TF, Dessouky MM (2006) A genetic algorithm approach to the inte-
grated inventory-distribution problem. Int J Prod Res 44(21):4445–4464
2. Ahn S, Suh NP, Cho DH (2013) Charging up the road. IEEE Spectr 50(4):48–54
3. Bulletin M (2013) Korean electric buses go wireless. Wall Street J 32:161–174
4. Chung SH, Kwon C (2015) Multi-period planning for electric car charging station
locations: a case of korean expressways. Eur J Oper Res 242(2):677–687
5. Dong J, Liu C, Lin Z (2014) Charging infrastructure planning for promoting bat-
tery electric vehicles: an activity-based approach using multiday travel data. Transp
Res Part C Emerg Technol 38(1):44–55
6. He F, Wu D et al (2013) Optimal deployment of public charging stations for plug-in
hybrid electric vehicles. Transp Res Part B Meth 47(1):87C101
7. Jang YJ, Jeong S, Ko YD (2015) System optimization of the on-line electric vehicle
operating in a closed environment. Comput Ind Eng 80(C):222–235
8. Jang YJ, Suh ES, Kim JW (2015) System architecture and mathematical models
of electric transit bus system utilizing wireless power transfer technology. IEEE
Syst J 99(2):1–12
9. Jia L, Hu Z et al (2012) Optimal siting and sizing of electric vehicle charging
stations. Dianli Xitong Zidonghua/Autom Electr Power Syst 36(3):54–59
10. Ko YD, Jang YJ (2013) The optimal system design of the online electric vehicle
utilizing wireless power transmission technology. IEEE Trans Intell Transp Syst
14(3):1255–1265
11. Ko YD, Jang YJ, Min SL (2015) The optimal economic design of the wireless pow-
ered intelligent transportation system using genetic algorithm considering nonlin-
ear cost function. Comput Ind Eng 89(C):67–79
12. Shin J, Shin S et al (2014) Design and implementation of shaped magnetic-
resonance-based wireless power transfer system for roadway-powered moving elec-
tric vehicles. IEEE Trans Ind Electron 61(3):1179–1192
13. Song YU, Park S et al (2013) A study on the validity of the infrastructure con-
struction cost for the commercialization of online electric vehicles. J Soc e-Bus
Stud 18(1):218–222
14. Sweda T, Klabjan D (2011) An agent-based decision support system for electric
vehicle charging infrastructure deployment. In: Vehicle Power and Propulsion Con-
ference, pp 1–5
15. Tali T, Telleen P (2013) Global EV outlook: understanding the electric vehicle
landscape to 2020. Int Energy Agency 87:22–30
The Research of Incentive Model Based on
Principal-Agent for R&D Personnel in
System-Transformed Institutes
Ying Wei(B) and Jun Gang
Sichuan Institute of Building Research, Chengdu, People’s Republic of China

452758107@qq.com
Abstract. The lack of innovation and vitality from R&D personnel is

a major problem to system -transformed institutes’ innovation activi-
ties, and the short of rational and effective incentive mechanism is the
source. Based on the analysis of present compensation and motivation,
this paper established an incentive model of income distribution and a
bonus incentive model by principal-agent theory, and then discussed the
advantages and disadvantages of the two incentive models. The result
shows that the incentive model of income distribution is more conducive
to stimulate the enthusiasm of R&D personnel for system-transformed
institutes.
Keywords: System-transformed institutes · R&D personnel · Incentive

mechanism · Principal-agent
1 Introduction
Over 10 years’ reform, some scientific research institutes have been transformed
into enterprises. In this process, the system-transformed institutes actively cul-
tivated the ability of technological innovation with excellent talent teams, and
obtained advanced scientific and technological innovation constantly, those are
inseparable with the creation and contribution of scientific and technological
innovation talents. R&D personnel as the key power of the system-transformed
institutes, is irreplaceable in the development and growth of enterprises. There-
fore, how to provide maximize innovative space and their creative enthusiasm
has become the most important issue for the institutes. But due to historical and
practical reasons, there are still some problems such as the inflexible incentive
mechanism, the single incentive means and so on, which restrict the innovation
ability and enthusiasm seriously. Consequently, whether from the development of
institutes itself or from the realization of the personnel value, it is very necessary
to study the motivation of R&D personnel in system-transformed institutes.
Although some institutes have already developed incentive policies, but most
of them are one-time awards, and usually pay attention to the number of achieve-
ments while ignore the quality and the efficiency, which are not conducive to
1646 Y. Wei and J. Gang
long-term enthusiasm for the R&D personnel. Many scholars at home and abroad
have carried out a lot of research work about the motivation. Ou [5] researched
on incentive of science and technology plan project by principal agent, ana-
lyzing the incentive factors and expound the design ideas of project incentive
mechanism. Dikolli and Kulp [1] studied the impact of interaction between per-
formance indicators of each task with multitask principal agent model. From the
perspective of creating core competence, Gao [3] focused on human resources
management problems and disadvantages about system-transformed institutes.
Guo and Zhang [4] summarized the policies and regulations of the income dis-
tribution of personnel achievements, and proposed suggestions to improve the
incentive mechanism. Ding and Chen [2] analyzed the Chinese S&T personnel
motivation legal laws and regulations, and put forward policy recommendations
to break the barriers to strengthen the transformation achievements motivation.
In summary, the current researches mainly concentrate on the demand char-
acteristic, incentive factors and other aspects of R&D personnel, having certain
referential significance to solve the motivation problem. On the basis of above
researches, combining with the actual situation of system-transformed institutes,
this paper took use of the principal-agent theory, optimizing the incentive mech-
anism design, discussed the incentive model of income distribution and bonus
incentive model respectively though the balance of benefits and risks among the
R&D personnel.
2 Background
2.1 The Connotation of the Collaborative Governance
A questionnaire survey for 18 system-transformed institutes in Sichuan has been
carried out. As we can see in Fig. 1 that R&D personnel are always with high
degree and different titles, showing that knowledge talents are the backbone of
personnel team. Speaking of the salary system and incentive mechanism, we
find that there are two main problems in current R&D personnel innovation.
One is how to combine the performance appraisal with income, the other is
the relationship between achievements and income. Each institute is actively
exploring or thinking to solve the innovation problems. About 63.6% of institutes
have a preliminary study on how to establish the performance evaluation and
incentive mechanism, while others do not give a complete management advice
on scientific and technological performance appraisal, still in a state of wait and
see on the issue of income distribution about achievements.
The composition of R&D personnel income includes intrinsic compensation
and extrinsic compensation, which can be seen in Fig. 2. Further, the intrinsic
compensation is closely related to incentives, containing basic salary, merit pay,
research prize and others. And the R&D personnel’s income is mainly dependent
on the basic salary and merit pay, taking 81.4% of the whole income, while less
incentive income in Fig. 3. Most system-transformed institutes have not estab-
lished a performance evaluation system different from the business production
personnel, especially the case of performance appraisal system is not well sound.
Incentive Model Based on Principal-Agent for R&D Personnel 1647
Fig. 1. The composition of R&D personnel in 18 institutes
Even though some institutes set up a merit pay, but not link to the performance
of scientific research, so the income systems are still egalitarianism with small
incentive income, especially lack of long-term incentive mechanism.
Invariable Basic salary

salary
the salary system of R & D personnel
Intrinsic Merit pay

compensation
Research
Variable pay
prize
Welfare
treatment
Others
Extrinsic Position
compensation promotion
Others
Fig. 2. The salary system of R&D personnel
3 Incentive Modeling
3.1 Income Distribution Incentive Model
Through the investigation and analysis above, it is not difficult to find that
even there are four parts of R&D personnel income, the first two are actually
relatively fixed, which can be seen as fixed income. And the other two are more
fluctuant and associate with the degree of effort, can be regarded as incentive
income. Assuming that the effort degree is proportional to the output, so the
output of R&D personnel can be shown as:
X = W0 + a + ε ε ∼ N (0, σ). (1)
Fig. 3. The composition of intrinsic compensation
At the same degree of effort, the output embodied in two levels, the single
value of scientific and technological achievement and the number of achievement.
For the labor of R&D personnel is the knowledge labor, long term on a small
number of achievements will get more value than those focus on numbers of
achievements. For example, a patent will be specialized or in 10 patents be
invented in one year. Obviously, the 10 patents tend not to produce too much
real value. Therefore, hypothesis that and satisfy the following relations:
a = qv 1/n . (2)
So, the actual benefits from R&D personnel to the institutes can be expressed
as:
n−1
S = vq = v n a. (3)
Assume that R&D personnel share the results of income in accordance with
a certain proportion k1 , so the personal income can be expressed as:
n−1
I = W0 + v n a. (4)
In addition, R&D personnel have to pay a certain cost as well as the effort.
Taking use of the common quadratic model in principal-agent theory, use to
represent the effort cost, which is the relationship between effort and cost, so
the cost can be expressed as:
1 2
c1 = ba . (5)
2
Otherwise, comparing with the scientific research institutes, R&D personnel
usually take risk aversion which needs to pay a certain risk costs. Set c2 on
behalf of the risk aversion, using Arrow-Pratt measure of absolute risk aversion
to construct the risk cost function as follows:
1 2 2n−2 2
c2 = ρk v n σ . (6)
2 1
So the actual expected income can be expressed as:

n−1 1 1 2n−2
E[I] = W0 + k1 v n a − ba2 − ρk12 v n σ 2 . (7)
2 2
Certainly, most of the institutes usually demand the minimum number of
achievements, which record as v. And set q as the lower limit of the single value,
so the value of a single scientific and technological achievement shall meet the
following constraints:
v ≤ υ ≤ a/q. (8)
In summary, the R&D personnel decision-making model is:
n−1 2n−2
max E[I] = W0 + k1 v n a − 12 ba2 − 12 ρk12 v n σ2
a,q (9)
s.t. v ≤ υ ≤ a/q.
n−1
Similarly, the income of institutes is (1−k1 )v n a−W0 . The risk of institutes
is often neutral relative to R&D personnel, pursue the maximization of expected
return usually. In addition, R&D personnel usually set a psychological bottom
line of personal income, which institutes must take it into consideration. So the
decision model of institutes is:
n−1
max E[V] = (1− k1 )v n a − W0
k (10)
n−1 2n−2
s.t. W0 + k1 v n a− 2 ba
1 2
− 12 ρk12 v n σ 2 ≥ Imin .
To sum up, the incentive model under the income distribution model for
R&D personnel, the decision model can be shown as:
n−1
max E[V] = (1− k1 )v n a − W0
k ⎧ n−1 2n−2
⎨ W0 + k1 v n a − 2 ban−1
⎪ − 12 ρk12 v n σ 2 ≥ Imin ,
1 2
2n−2 (11)
s.t. max E[I] = W0 + k1 v n a − 12 ba2 − 12 ρk12 v n σ 2 ,
⎪
⎩ a,q
s.t. v ≤ υ ≤ a/q.
For the income is directly linked to the transfer profit, so the R&D personnel
will try to create the greatest value of achievements under the premise of the
minimum completion number. The effort level of R&D personnel is positively
correlated with the income coefficient, the greater proportion of income distribu-
tion, the more initiative from R&D personnel. Strives for the partial derivatives
can get the following results:
∂k1 ∂a ∂E[V ] ∂k1 ∂a ∂E[V ] ∂k1 ∂a ∂E[V ]

< 0, < 0, < 0, < 0, < 0, < 0, < 0, < 0, < 0.
∂ρ ∂ρ ∂ρ ∂b ∂b ∂b ∂q ∂q ∂q
Thus it can be seen that, in the optimal state, the income coefficient, the effort
level and the income of institutes are negatively correlated with the risk aver-
sion, the effort cost and the minimum number of achievements. This suggests that
under the incentive model of income distribution, the system-transformed insti-
tutes should reduce the limit number of achievements can not only encourage R&D
personnel in a greater extent, but also improve the actual income of institutes.
3.2 Bonus Incentive Model
This part will discuss the incentive problem of R&D personnel under the bonus
incentive model. Hypothesis that the bonus is positively with the number of
achievements, and the achievement bonus of the institute is k2 . Still, the single
value υ and the number q satisfy the following relations at the same level of
effort:
a = qv 1/n . (12)
Then, the income of R&D personnel can be expressed as:

1
I = W0 + k2 q = W0 + k2 v n a. (13)
For the bonus is positively with the achievement number, so under the same
level of effort, R&D personnel must ignore the quality of the results, and pay
more attention to the number of results. Set v as the lower limit of the single value
of achievements, and the upper limit is a/q when meet the minimum number of
results, so taking the cost and risk into account, the expected return model of
R&D personnel is:
1 2
max E[I] = W0 + k2 v − n a − 12 ba2 − 12 ρk22 v− n σ 2
a,v (14)
s.t. v ≤ υ ≤ a/q.
The expected return of institute is:

1 n+1 1
E[V ] = S − W0 − k2 q = vq − k2 v n a − W0 = v n a − k2 v n a − W0 . (15)
So, under the bonus incentive model the incentive model can be shown as
follows:
n+1 1
max E[V] = v n a − k2 v n a − W0
k2 ⎧
−1 2 −2 2
⎨ W0 + k2 v n a − 2 ba −1 2 ρk2 v n σ ≥ Imin , 2
⎪ 1 2 1
(16)
s.t. max E[I] = W0 + k2 v − n a − 12 ba2 − 12 ρk22 v− n σ 2 ,
⎪
⎩
a,v
s.t. v ≤ υ ≤ a/q.
There we can get the following conclusions, (1) under the bonus incentive
model, in order to maximize the personal interests, R&D personnel will ignore
the actual value of the research results and blindly pursue the numbers. As a
principal, institutes can set the appropriate technology bonus by assessing the
lowest value of achievements. (2) For the bonus amount is proportional to the
lowest value, so the lower value of achievements, the more achievements R&D
personnel create in the same level of effort, so as to obtain more revenue. (3)
Stand in the perspective of the institutes, to avoid the emergence of low value
achievements, they should set a threshold of achievements, and strictly evaluate
them.
4 Comparison Analysis
Compare the income distribution incentive model with the bonus incentive model
(see in Table 1), we can find that: (1) R&D personnel have different pursuit in
the two models. R&D personnel pay more attention on the value of achieve-
ment under the income distribution incentive model, while pursue to create a
great number of achievements in the bonus incentive model. For the system-
transformed institutes take profit as the goal, thus using income distribution
incentive model more. And the colleges use bonus incentive model frequently
when facing with a large number of assessment indicators. (2) In the income
distribution incentive model, the revenue sharing factor of R&D personnel is
relative with effort cost, risk aversion coefficient and output random error, but
has nothing to do with the result itself. While the bonus factor in addition to
the impact above, also related to the lowest value of scientific and technological
achievements.
Table 1. The comparison of two models
Parameter Income distribution incentive model Bonus incentive model

n+2
1 v n
k 1+ρbσ 2 1+bρσ 2
1
v bn q n (1+ρbσ 2 )
v
1−n n+1
q v n
a bn (1+ρbσ 2 )n b+ρb2 σ 2
2n+2
q 2−2n v n
E[V ] 2b2n−1 (1+ρbσ 2 )2n−1
− Imin 2b(1+ρbσ 2 )
− Imin
5 Conclusion
Scientific incentives can attract and retain excellent researchers for scientific
research institutes, meeting the ultimate needs of beneficiaries constantly. And
salary is the prerequisite to ensure the material needs of R&D personnel, also is
the prerequisite and basis for the scientific and technological talents to pursue
higher level needs. Even though China’s policies and regulations for the allo-
cation of R&D personnel incentive requirements are put forward, most of the
institutes only pay fixed income when they can observe the effort level from
personnel. Normally, it is difficult to observe the effort level, still let researchers
share part of the output share, otherwise R&D personnel will not work hard
with basic salaries, which is detrimental to the development of enterprises, also
the country. From the analysis, R&D personnel focus on the value is applied
to the system-transformed institutes, while a great number of achievements will
be made to the colleges. The revenue sharing factor of R&D personnel is rela-
tive with effort cost, risk aversion coefficient and output random error, but has
nothing to do with the result itself. Therefore, the greater effort cost, the smaller
share of output, and the smaller risk for those who fear of hard work. Conversely,
the stronger ability, the lower effort cost.
Acknowledgements. This research was conducted with the support of science plan
program of Chengdu in 2016 (No. 2015-RK00-00062-ZF).
References
1. Dikolli SS, Hofmann C, Kulp SL et al (2009) Interrelated performance measures,
interactive effort, and incentive weights. J Manag Acc Res 21(1):125–149
2. Ding M, Chen B (2016) Research on the strengthening the S&T personnels trans-
formation achievements motivation based on the improvement of the reward mech-
anism determined by knowledge, technology, management, skills and other factors
market. Sci Manag Res 142(3):96–100
3. Gao J (2013) The incentive mechanism of human resources management of sci-
entific research institution after ownership transformation. Municipal Eng Technol
61(2):46–54
4. Guo Y, Zhang S (2015) The incentive mechanism of the income distribution of
scientific and technological achievements of technical personnel. Sci Sci Manag S&T
13:183–190
5. Ouyang J, Yan H, Huang C et al (2008) Research on incentive of science and tech-
nology plan project based on principal agent. Sci Technol Manag Res 28(8):246–248
Research on the Influencing Factors of Public
Sense of Security and Countermeasures in Major
Emergencies–An Empirical Analysis on the
Public Sense of Security in View Of the
Explosion in Binhai New Area in Tianjin
Jing Yang(B) and Wei Xu
School of Political Science and Public Administration,

University of Electronic Science and Technology of China,
Chengdu, The People’s Republic of China
yangjing8056@uestc.edu.cn
Abstract. In the crisis, the biggest psychological threat to the public

is the loss of public sense of security. Through the related research, it
is hypothesized that critical events, countermeasures of the government
and the media, emergency response capability of the individual and the
emergency psychological behavior of the group are the four main factors
influencing public sense of security in the emergencies. A total of 330 sam-
ples were randomly selected from those people who had experienced the
explosion in Tianjin. By exploratory factor analysis and the structural
equation model examination, the fourfactors all have different degrees of
explanation towards the loss of public sense of security. According to the
conclusion, we can see that it is the key for the government to construct
the governance mechanism by enhancing the emergency response capa-
bility of the group improve their certain sense of crisis and their sense of
control of the situation.
Keywords: Public sense of security · Critical events · Countermeasures

of the government and the media · Emergency response of the individual
and the group · Structural equation model
1 Introduction
A crisis is often caused by both the “objective crisis of the events” and “subjec-
tive crisis of the individual”, but the negative impact of the subjective crisis of
individual is often greater than that of the events. Just as the attack on Septem-
ber 11 has happened for 15 years while its impact and psychological fears left
to people have a long-term influence on American society. In 2011, Fukushima
nuclear accident caused “salt-buying panic” in China, the reason of which is
1654 J. Yang and W. Xu
people’s fear for the crisis. When the crisis comes, the biggest threat to the pub-
lic is the loss of their sense of security which is a prominent feature of public
emergency psychology after the crisis. But why a few events that are actually
less harmful are more influential on the public than those that have fatal con-
sequences? Why do crisis of the same types have different effects on the public
sense of security in a different time and place? This series of events have shown
that the formation of public sense of security requires both objective conditions
and also subjective factors.
Based on the existing research on the public sense of security, this paper
designs a model of the influencing factors of public sense of security and also
analyzes the data through questionnaire and the relationship between data and
their characteristics based on SPAA as well as the influencing factors of the public
sense of security based on the structural equation of AMOS. Besides, it measure
how these factors influence the public sense of security, which contributes to the
design of the emergency psychological governance mechanism in the emergency.
2 Conceptual Definition and Literature Review

2.1 Definition of the Public Sense of Security
According to scholars at home and abroad, sense of security is referred to the

sense of belonging in a community and the sense of self [3]. The “public sense
of security” in this article involves the sense of belonging, the certain sense
of control and the satisfaction of security needs that the public experience in
a certain public environment. Among them, the sense of belonging refers to
the warm and the reliable subjective experience of the cares and comfort from
other members in the community. The certain sense of control refers to the
certain cognitive and controlled experience for the causes, processes and trends
of the public crisis based on the timely, adequate and consistent information
attained from the main information channels. Satisfaction of security needs refers
to the effective sense attained from effectively coping with the crisis and safe
subjective experience, which is formed on the basis of the full understanding of
the development of the public crisis and the assessment of the social assistance
system and the public emergency capability.
2.2 Overview of Domestic and Foreign Research
Foreign research has gone through two stages: the first stage is the public sense
of security in the perspective of “crisis management”. “Bounded Rationality” of
Simon [7] thinks that due to the limitation of the individual in memory, think-
ing, computing power and other aspects, individual rationality is just bounded
rationality in constrained conditions. The second one is to introduce the the-
ory of “risk perception” with the public sense of security examined by “crisis
communication” model. Industrial pollution happening in 1984 in Union Car-
bide Corporation triggered panics due to slow release of information, which then
Research on the Influencing Factors of Public Sense of Security 1655
urged the scholars to reflect. Covello finds that the risk assessment by the author-
ities is very different from the risk perception by the public and also the public
show great distrust in the authority of risk management, which weakened their
sense of security and even lead to fear [1]. Sun Ding and Matthew, the American
scholars, thinks that the public sense of security is “the degree of anxiety and
concern of those who are becoming victims”, so the potential victims is selected
as the objects in their research. However, they also agree that the public sense
of security is a psychological phenomenon, which is characterized by anxiety or
concern. Lois Mok, an American scholar, put more focus on the power of control
in the sense of security and think it is the reflection on the normal order and the
weakened social control from of the public [9].
In recent years, we have conducted deeper research on the public sense of
security in the crisis, and the theoretical circle has shown a basic trend of mul-
tidisciplinary research and innovation. In terms of the social and psychological
impact from critical events, Zhang Yan, in her study of the critical events, divided
the psychological impact on the victims from critical events into three stages:
the first stage is the formation of risk experience; the second is psychological
cognition of risk; the third one is the implementation of mental decision and the
spread of psychological influences [10]. The most important empirical investiga-
tions of the public sense of security was the sample survey conducted by Institute
of public sense of security in the Ministry of public sense of security in 1988 in
15 provinces and urban areas with its results published in the book–“Do you feel
safe?”. The group defined the public sense of security as: the subjective feelings
and evaluation of the public for the social security are the comprehensive psy-
chological reflection of the infringement that the personal rights, property rights
and other lawful rights and interests of citizens have suffered or will suffer and
the protection they expect in a certain period of the social life [4]. Based on the
theory of prospect, Sun Doyong constructed the research model to study the
personal perception of fears in the crisis from three aspects: the features of the
event itself, the characteristics of the individual and the social factors [8].
3 The Hypothesis in the Research

3.1 Build a Model of the Influencing Factors of Public Sense of
Security
The critical events, the countermeasures of the government and the media, the
emergency response capacity of the individual and the psychological behavior
of the group in the natural disasters are described as four main factors that
influence the public sense of security in “Research on the Influencing Factors of
public sense of security and Countermeasures in Major Emergencies–An Empir-
ical Analysis Based on the public sense of security of Ya’an Earthquake on April
20”. Based on it, we assume that, in the disaster of security incidents, the critical
events, the countermeasures of the government and the media, the emergency
response capacity of the individual and the psychological behavior of the group
are still the most important factors that influence the public sense of security.
Critical events Countermeasures of the
Public sense of security in the emergencies
Emergency response capacity Emergency psychological behavior of

of the individual the group
Fig. 1. Model of the influencing factors the public sense of security in the emergent
events
Thus, we construct a conceptual model that affects public sense of security in

the disaster of security incidents, which is shown as in Fig. 1.
3.2 The Hypothesis in the Research

Caplan and Edward assumed that each person is basically in a state of dynamic
balance with the environment. When a person is facing a huge disaster in his life,
he will have a high degree of tension, anxiety, pessimism and disappointment and
a series of psychological problems if the disaster goes beyond his psychological
endurance [2]. Thus, we assume that:
Hypothesis 1 : In an emergency, the severity of the crisis, the extent of the
impact and the number of casualties are negatively related to the level of the
public sense of security. The more serious the crisis are, the more insecure the
public will feel.
“The media dependence theory” put forward by Melvin Devler, a prominent
American communication expert, points out that people are often anxious to get
to know the truth through the government and the media, and the dependence
will be enhanced obviously when the society is undergoing major changes but
the situation is not clear. Therefore, the timely warning and the objective and
fair reports from the government and the media is helpful to reassure the public
and prevent further deterioration of the situation [6].
Hypothesis 2 : In the emergencies, the countermeasures of government and
the media to the emergencies is positively related to the public sense of security.
The more timely and effective the countermeasures of the government and the
media are, the safer the public will feel.
Robert Hiss put forward the FPC model and thought that the individual’s
ability and familiarity constituted the core variables of effective response to the
crisis. The stronger the individual’ ability and the higher the familiarity with
the crisis are, the better the effect of dealing with the crisis will be [5].
Hypothesis 3 : In the emergencies, the individual’s knowledge of crisis, risk
preferences, psychological quality are positively correlated to the public sense of
security. The stronger the individual’ emergency capacity are, the safer he will
feel.
According to the theory of “Herd Behavior” which has been widely verified,
individual in the crisis, often affected by others’ behavioral strategies, will often
adopt the same behavioral strategy. That is to say, instead of being based on
their own information, their behavior choice is the imitation of others or over-
reliance on the public opinion, which tends to lead to the herd behavior. Thus,
we assume that:
Hypothesis 4 : In the emergencies, the psychology and behavior of the sur-
rounding people are positively related to the public sense of security. The more
stable the emergency psychological behavior of the group is, the safer the public
will feel [8].
4 Research Methods and Tools

4.1 Scale
The paper collects data from the victims suffering from in the “explosion that
happened in October 12 in Binhai New Area in Tianjin” by questionnaire survey
which is divided into three parts. The first part is about the measurement of
public sense of security and it makes the subjects score their sense of security;
the second part is the measurement of public sense of security in emergencies;
the third part is involved in the investigation of the five factors that influence
hypothesis to finally form the questionnaire for formal study which is consisted
of 25 items by the analysis of the factor and project.
4.2 Descriptive Statistical Analysis

A total of 330 questionnaires were distributed and 320 of them were collected
with a recovery rate at 96.9%. After removing the unqualified questionnaire, 305
valid questionnaire were collected with the effective rate at 95.3%.
4.3 Examination of the Internal Consistency

According to the third part of the questionnaire, SPSS was used to analyze the
internal consistency of the samples, which is shown as in Table 1. The results
show that the alpha coefficient is 0.839, which proves that the designed ques-
tionnaires have good internal consistency and also indicate that the design of
the questionnaire is scientific and reasonable.
Table 1. Analysis results of the internal consistency
Alpha of Cronbach Alpha of Cronbach that is based on Number of the projects

standardized project
0.839 0.837 20
5 Results
5.1 Analysis of the Impact of Gender on the Public Sense of
Security
In this questionnaire, we conducted survey on the demographic characteristics

of respondents from the perspective of gender.
Using Independent-samples T test of the SPSS, we analyzed the impact of
gender on sense of security with its results shown in Table 2. In the first part, we
tested Levene’ variance homogeneity with F = 0.041, p = 0.084 > 0.05, which
meant that the variance was homogeneous while in the second part, the tested
t = 2.585, p = 0.010 < 0.05, symbolized that the mean difference of the two
samples is significant.
Table 2. Results of T test about gender
Levene test in variance equation T test in mean equation

F Sig. T df Sig. (in Difference Standard Min. Max.
two between margin
sides) the means of error
Index of the 0.041 0.841 2.585 314 0.01 0.609 0.236 0.146 1.073
sense of
security
construct the
equal
variances
Construct the 0 0 2.569 298,915 0.011 0.609 0.237 0.143 1.076
unequal
variances
The results of T test show that the public sense of security is significantly
affected by the gender after the explosion and the sense of security of men is dis-
tinctly higher than that of women, which may be explained by the environment
where men and women grew up or their physical differences and also shows that
in the post-disaster relief, we need to put more focus on women to rescue them.
5.2 Exploratory Factor Analysis
SPSS was used to carry out exploratory factor analysis on the samples with
its results shown in Fig. 3. v13 to v32 are the corresponding items of the four
influencing factors in the questionnaire. Among them, v13 to v15 correspond to
critical events, v16 to v23 to the countermeasures of the government and the
media, v24 to v27 to the emergency response capabilities of the individual, and
v28 to v32 to the emergency psychological behavior of the group.
The results of exploratory factor analysis showed that the influencing factors
mainly concentrate on four aspects, which was consistent with the hypothesis
of the paper. Because of the dispersion of the concentration of the emergency
psychological ability of the group in exploratory analysis, after analyzing and
interpreting the corresponding items, we deleted two of them to conduct the

exploratory factor analysis once again, which showed that the polymerizing effect
of the four factors was good.
As shown in Table 3
Table 3. Results of exploratory factor analysis
Elements
1 2 3 4
v13 - - - 0.852
v14 - - - 0.777
v15 - - - 0.619
v16 0.778 - - -
v17 0.822 - - -
v18 0.775 - - -
v19 0.747 - - -
v20 0.674 - - -
v21 0.725 - - -
v22 0.687 - - -
v24 - - 0.646 -
v25 - - 0.479 -
v26 - - 0.630 -
v27 - - 0.781 -
v28 - 0.641 - -
v29 - 0.508 - -
v30 - 0.680 - -
v32 - 0.730 - -
Attainment Method Element
analysis of the subject. Pivot-
ing method Formal maximum
variation method with Kaiser.
5.3 Examination of the Model of the Influencing Factors of the

Public Sense of Security
Based on the above four factors, this paper uses Amos to test the structural
equation model, the indicators of which all meet expectation, as shown in Fig. 2.
It can be seen that the four factors have a better explanation towards their
impact on the public sense of security. Among them, only the impact of the crit-
ical events on the public sense of security is negative, indicating that they are
negatively correlated, while the other three factors have positive impacts. Com-
pared with the five hypotheses originally designed, the trust in the government
and the countermeasures of the media are combined as one influencing factor. At
the same time, according to the value of how these factors influence the public
sense of security, we can see that the emergency psychological behavior of the
group has the greatest impact, followed by the emergency capacity of the indi-
vidual, the countermeasures of the government and the media and the severity
of the critical events.
Fig. 2. Model of the influencing factors of the public sense of security. (ERCOG-
Emergency response capacity of the group, COGM-Countermeasures of the government
and the media, ERCOI-Emergency response capacity of the individual)
5.4 Examination on the Exploratory Factor Analysis of the

Structural Dimension of the Public Sense of Security
We divided the structural dimension of the public sense of security, examined

by SPSS, into sense of belonging, certain sense of control and sense of security
with its results shown in Table 4: We can see that 10 items of the questionnaire
are better aggregated into three categories, the coefficient of which statistically
meet the requirements. We can see that V3, V4, V5, V6 come together as the
first dimension while V7, V8, V9 as the second and V10, V11 and V12 as the
third, which is consistent with the idea of how questionnaire is designed, but
the structure of the questionnaire is more clear. According to the design of the
questionnaire, we reduced the first dimension to a sense of belonging and the
second dimension to a sense of security and the third to certain sense of control.
Table 4. Exploratory factor analysis of the structure dimension of the public sense of
security
Rotation matrix of the components

1 2 3
v3 0.708 - -
v4 0.841 - -
v5 0.777 - -
v6 0.557 - -
v7 - 0.608 -
v8 - 0.860 -
v9 - 0.815 -
v10 - - 0.752
v11 - - 0.730
v12 - - 0.800
5.5 How the Influencing Factors of the Public Sense of Security

Affect the Sense of Security
On the basis of determining the model of the influencing factors of the pub-
lic sense of security and how each factor affects the public sense of security,
we explored the mechanism of how the influencing factors affect the structural
factors of the sense of security.
(1) How the influencing factors of the public sense of security affect the sense of
belonging
First of all, we tested the structural equation model based on how the influ-
encing factors of the public sense of security affect its first structural factors–
“sense of belonging” with its results shown in Fig. 3. The results that may be
affected by the characteristics of the security incidents showed that the emer-
gency psychological behavior of the group had the strongest effect on the sense
of belonging of the individual while the emergency response capacity of the indi-
vidual comes second, followed by the countermeasures of the government and the
media. When accidents and disasters that human error should be to blame for
occurs, the credibility of the government and the media in the public mind will
somewhat decline. The public will be more dependent on how the surrounding
people deal with disasters, the rational behavior and effective countermeasures
of whom will give the public a “sense of belonging.” With the development of
the newly emerging net media, the knowledge of how to effectively deal with
the disaster has been more and more universal, at the same time the overall
emergency response capacity of the group is also improved, which is consistent
with the reality.
(2) How the influencing factors of the public sense of security affect the security
needs
Fig. 3. Model of how the influencing factors of the public sense of security affect
the sense of belonging. (ERCOG-Emergency response capacity of the group, COGM-
Countermeasures of the government and the media, SOB-Sense of belonging, ISOB-
Individual’s sense of belonging, ERCOI-Emergency response capacity of the individual)
First of all, we tested the structural equation model based on how the influ-
encing factors of the public sense of security affect its second structural factors–
“security needs” with its results shown in Fig. 4. The results showed that the
emergency response capacity of the individual was the most important factor
that affects the sense of security of life, followed by the emergency response
capability of the group and the countermeasures of the government and the
media. Although personal characteristics, knowledge of emergency response and
reactions on the spot are vastly different in the individuals, the security needs, for
us, is the most essential experience of our safe lives and our emergency response
capacity can directly affect our sense of security. In spite of the fact that the per-
sonal characteristics are difficult to control, it is possible to maximize the ratio-
nality by improving the personal crisis knowledge and his emergency response
capacity in the face of major emergencies and minimize the casualty through
effective emergency response. Once the individual has the clear understanding
of that he should take immediate effective emergency measures, he will maximize
the index of his sense of security. At the same time, the results also showed that
in the crisis the emergency psychological behavior of the surrounding people will
have great effect on the security experience of the individual.
(3) How the influencing factors of the public sense of security affect the certain
sense of control
First of all, we test the structural equation model based on how the influenc-
ing factors of the public sense of security affect its third structural factors–the
Fig. 4. Model of how the influencing factors of the public sense of security affect
the security needs. (ERCOG-Emergency response capacity of the group, COGM-
Countermeasures of the government and the media, SOS-Sense of security, SOSOL-
Sense of security of life, ERCOI-Emergency response capacity of the individual)
“certain sense of control” with its results shown in Fig. 5. The results showed
that the emergency psychological behavior of the group is the most important
factor which can affect the certain sense of control of the public with the coun-
termeasures of the government and the media coming second and the emergency
response capacity of the individual third. The emergency behavior of the group
can offer the demonstration and guidance to personal behavior, therefore, in the
security incidents, their rational and effective emergency behavior can enhance
the sense of control of the individual. In this mutual action model, the counter-
measures of the government and the media also have greater influence, because
it is a key to obtain the adequate, accurate and timely information for enhancing
the sense of control. In China, our government and media have the ability to
grasp the information of the emergencies and delivery them to the public ade-
quately, timely and accurately to help themata in the whole information and
improve their certain sense of control.
6 Countermeasures
In the major emergencies, the biggest problem that the public face is the loss of
their sense of security, which, as a concept of subjective cognition, is to a large
extent affected by self psychology. Thus, we can by certain means and ways affect
our sense of security so as to enhance the effectiveness of disaster aid.
Fig. 5. Model of how the influencing factors of the public sense of security affect the
certain sense of control. (ERCOG-Emergency response capacity of the group, COGM-
Countermeasures of the government and the media, SOC-Sense of control, CSOC-
Certain sense of control, ERCOI-Emergency response capacity of the individual)
When the crisis happens, it is a starting point of emergency psychological

governance to create the public sense of security and the measurement of its
influencing factors and how it is affected provide the evidence and direction for us
to create a countermeasure system of social emergency psychological governance,
which will be introduced in terms of the following three aspects.
6.1 Enhance the Countermeasures of the Emergency Response

Capacity of the Group
In the measurement of the influencing factors of the public sense of security,

the emergency response capacity of the group has the greatest impact. Personal
behavior, influenced both by the external factors and group action, is easy to
imitate the group action. In critical events, whether the emergency behavior
of the group is effective or not will directly affect the sense of security of the
individual. Therefore, it is necessary to enhance emergency response capacity of
the group by regularly relevant knowledge and disaster drills.
6.2 Improve the Countermeasures of Certain Sense of the Crisis
In light of the countermeasures of the government and media as the key influenc-
ing factors of the public sense of security, unscientific reports on critical events
from medias and the delay of the emergency measures from the government will
lead to a serious decrease of the public sense of security and also fuel their fears.
In the case of asymmetric information, government and the media should play
their role of “delivering information timely” to enhance the sense of certainty
of the public. The emergency measures taken by the government departments
should be combined with the information distributed by the media and the view
of the media to keep the information timely, adequate and accurate so that they
can enhance the certain sense of crisis of the public.
6.3 Put More Importance to the Countermeasure of the Sense of

Control of the Situation
This study also validate the assumptions that the nature and severity and other
aspects of the critical events have impact on the emergency psychology of the
public. Public fear in the face of crisis is an important reason for the loss of the
sense of security. A sense of control of the situation should be created to help
the public know how to deal with the crisis. They have become the important
topics of emergency psychological governance to strengthen crisis education of
the whole society so that people in the face of crisis can keep rational and orderly
in the risk society. We should create a four-in-one crisis education system of
the government, media, social organization and the public and design diagnostic
measures, publicity measures and mentoring approach of the effective emergency
measures in all kinds of situations for the public.
Acknowledgements. This paper is supported by the National Philosophy and Social

Science Fund Youth Project (14CZZ017), a Study on the Influencing Factors and Con-
struction Mechanism of Public Security in the Background of Major Incident, 2014/06–
2016/12 2 and the Sichuan Province Twelfth Five-year Plan Philosophy Social Science
Planning Fund Youth Project (SC11C002), Research on the Mobilization Mechanism
of Internet Virtual Community Organization and the Countermeasures of Government.
2010/06–2015/12.
References
1. Covello VT, Winterfeldt DV, Slovic P (1987) Communicating scientific information
about health and environmental risks: problems and opportunities from a social
and behavioral perspective. In: Covello VT, Lave LB, Moghissi A, Uppuluri VRR
(eds) Uncertainties in risk assessment and risk management, pp 221–239
2. Heath RL, Nathan K (1990) Public relations’ role in risk communication: informa-
tion, rhetoric and power. Public Relat Q 68(3):15
3. Lin Y (2007) Public sense of security and the construction of index system. Soc
Sci 7:67–68
4. Ren Y, Wei J (2015) Analysis on the influencing factors of public concern in public
crisis. Stat Decis 1:67–70
5. Robert M (2004) Heath, crisis management. Zhongxin Press, Beijing
6. Rocky B, Zheng Z, Wang B (2004) From “dependence on the media system” to
“communication infrastructure”-looking back to the development of “media system
dependency theory” and its new concept. Int Journalism 2:9–12
7. Simon HA (1978) On how to decide what to do. Bell J Econ 9(2):494–507

8. Yang J, Yang M (2016) Research on the influencing factors of public sense of
security and countermeasures in major emergencies-an empirical analysis based on
the public sense of security of ya’an earthquake on April 20. Exploration 1:172–179
9. Yang M (2015) Study on the influencing factors of public sense of security in major
natural disasters. Master’s thesis, University of Electronic Science and Technology
of China
10. Zhang Y (2011) Risk attitude, risk perception and trust: an analysis framework of
government information supplying mechanism in crisis situation based on prospect
theory. J Huazhong Univ Sci Technol 25(1):53–59
Effect Assessment of the Free Tissue Flap
Colorimetric Card in the Postoperative Flap
Management for Tumors of Oral and
Maxilla-Facial Region
Xiongtao Yang, Xueying Wang, and Qing Yang(B)
Department of Head and Neck Surgical Oncology,

Sichuan Cancer Hospital, Chengdu 610041, People’s Republic of China
57771050@qq.com
Abstract. The objective of this paper is to investigate the effect of free

tissue flap colorimetric card in the postoperative flap management for
tumors of oral and maxilla-facial region. Through observing 42 patients
who used free tissue flap colorimetric card after operation, it was found
that twelve patients had vascular disease, three had insufficient arter-
ial infusion, and nine had venous return obstruction. Among them, one
patient who had insufficient arterial infusion and two had venous return
obstruction received re-anastomosis of vascular surgery and hemostasis
surgery, and then all patients recovered well. The paper conclude that
the free tissue flap colorimetric card is a simple, timely, accurate and
effective tool for the observation of flaps, and it is suitable for clinical
application.
Keywords: Free tissue flap colorimetric card · Tumors of oral and

maxilla-facial region · Skin flap care · Crisis of blood-circulation
1 Introduction
Radical surgery for malignant tumors of oral and maxilla-facial region involves a
wide range of injury, resulting in soft and hard tissue defects, seriously affecting
the appearance and function of patients. Using their own tissue flap free graft to
the mouth for large area of tissue defects repair has become a common method to
improve the quality of life of patients [11]. It is the most important care content
after free skin flap transplantation surgery to timely detect and actively deal
with crisis of blood-circulation, as well as to promote the survival of the flap.
In order to improve the success rate of free tissue flap, reduce the occurrence of
crisis of blood-circulation and improve the success rate of the crisis tissue flap
rescue, the branch designed the colorimetric card for free tissue flap in 2011 and
applied the national utility model patent certificate (Certificate No. 3767860).
Whats more, it has been used in has been in clinical application for more than
1668 X. Yang et al.
3 years, and the reflecting is very good. In this paper, the effect of free tissue
flap colorimetric card in the postoperative flap management for tumors of oral
and maxilla-facial region was concluded.
2 Materials and Methods

2.1 Normal Information
From September 2013 to November 2016, 42 cases of free flap repair were per-
formed in our department, including 35 males and 7 females with an average age
of 46.2 years (range, 17–74 years), including 17 cases of tongue cancer, 9 cases
of buccal cancer, 7 cases of oral bottom cancer, 5 cases of gum cancer, 4 cases
of soft palate cancer.
The information of surgery: there were 27 forearm radial free skin flap in
repairing of oral cavity defect, 6 anterolateral thigh flap in repairing of oral
cavity defect, 4 free pectoralis major myocutaneous flap in repairing of oral
defect, 3 fibula myocutaneous flap in repairing of oral and maxillofacial defect,
2 osseo myocutaneous iliac flap in repairing of oral and maxillofacial defect.
2.2 Observation and Nursing of Skin Flap

Free flap is mainly observed blood flow of the skin flap, such as the color of the
flap, temperature, degree of swelling and Shiatsu back to the blood reaction so
as not to crisis of blood-circulation. Observer should avoid flashlight and other
light interference, while maintaining the correct posture, to prevent the flap
tension is too large or distorted [12]. The flap was observed by free tissue flap
colorimetric card immediately after the patient was returned to the ward, and
the results were recorded on the free tissue flap monitor. The observation was
recorded every 30 min within 24 h after surgery, every 2 h within 24–72 h after
surgery, every 4 h within 72 h–7 d after surgery. Abnormal flaps were observed
every 15 min to record, and observe the records as usual since removing crisis of
blood-circulation.
2.3 The Design of Free Tissue Flap Colorimetric Card

There are 10 colorimetric card for free tissue flap grafting made of silicone, its
four corners rounded. The color areas of the ten cards are selected according to
Table 1 and are different from each other, like Fig. 1.
The connecting shaft 6 has at least three elongated cards 1 whose one end is
provided with a connecting hole 2, and the other end is provided with a colored
area 3, and all the cards 1 are mounted on the connecting shaft through the
connecting hole 2 in a manner similar to the fan structure, which is rotatable
about the connecting shaft 6. The side edge of the card is also provided with
a graduation zone 4 (accurate to millimeters) adjacent to the color zone 3 to
accommodate various sizes within the oral cavity of the patient. In addition, the
card is also provided with a text description area 5, with the color area 3 (seen
Table 1).
Assessment of Effect About Free Tissue Flap Colorimetric Card 1669
Fig. 1. The design of free tissue flap colorimetric card
Table 1. Description of color about free tissue flap colorimetric card
Parameter Pale 1 Pale 2 Pale 3 Rosy 1 Rosy 2 Rosy 3 Violaceous 1 Violaceous 2 Violaceous 3 Atropurpureus
Tonal 20 20 20 20 9 9 234 235 225 225
Saturation 240 240 240 240 231 232 80 79 75 75
Brightness 226 216 204 195 166 152 169 146 126 45
Red R 225 225 205 255 252 252 205 188 172 63
Green G 240 230 217 207 136 113 155 122 96 33
Blue B 225 205 179 159 100 70 162 131 125 44
Note: the form of parameters are computer palette.
Pale1: arterial of flap flood inadequacy severally.

Pale2: arterial of flap flood inadequacy generally.
Pale3: arterial of flap flood inadequacy lightly.
Rosy1: the flaps grew well and the blood circulation was smooth relatively.
Rosy2: the flaps grew well and the blood circulation was smooth.
Rosy3: the flaps grew well and the blood circulation was very smooth.
Violaceous1: venous return is blocked lightly.
Violaceous2: venous return is blocked generally.
Violaceous2: venous return is blocked severally.
Atropurpureus: flaps is dead.
2.4 Detailed Description of Free Tissue Flap Colorimetric Card
The patient opened his mouth as fully as possible in order to expose the flaps
in the irradiation conditions of natural light or fluorescent white light. At first,
let the colorimetric card range by fan-out, compared and estimated flap color
belongs to which card 1 color area 3. Then in order to improve the accuracy,
the card 1 of the color region 3 is further inserted into the oral cavity of the
patient, and the contrast is confirmed in close proximity. If colorimetric card
was contaminated by the patient’s oral secretions, blood, etc. in the usage of
the card, it is necessary to be disinfected with alcohol or chlorine preparations
after cleaning with water, and to prepare a set of colorimetric cards for every
patients with specific infections. Colorimetric cards can also be set into a PVC
gloves, discarded after every used. Generally, each card 1 is provided with a
color region 3, and the number of the cards 1 can be determined according to
1670 X. Yang et al.
the circumstances. When more accurate conditions are required, more cards 1
are required to make the whole structure have more color areas 3, in order to
obtain more accurate flap discoloration.
2.5 Beneficial Effect of Free Tissue Flap Colorimetric Card

As the color areas were set in the card, the color of color area can be set by color
in different conditions of flap, then the medical staff can objectively compare
the transplanted flap color with the card on the color area in order to determine
t whether there is a problem in the growth of the flap and to solve timely.
This judgment of same standard reduces the judgment error caused by different
subjective consciousness or lack of experience, which makes the flap observation
simpler and more accurate.
3 Results
The 42 flaps were all live. Twelve patients had vascular disease, three arterial
insufficiency, and nine venous return obstruction. Among them, 2 cases of mild
arterial insufficiency maintain the correct position to release the pressure around
the wound and the symptom relief after strengthening the local negative pressure
drainage. 6 cases of mild venous return disorder were light by roasted lights.
Among them, 2 cases relief after plus acupuncture flap bleeding, heparin sodium
coated. 1 case of severe hypoperfusion of the arteries and 2 cases of severe venous
return obstructed patients received re-anastomosis underwent vascular surgery
and bleeding debridement surgery. 1 case of venous return to the middle and
severe blocked, skin flap dead. All the patients recovered well after phlebotomy,
and all the patients recovered well in phonation, chewing and swallowing (seen
Table 2).
Table 2. The status about 42 cases of flaps after surgery
Situation # of Vascular crisis time/h Interventions measures TFBBI

of flap cases
≤12 12–24 24–48 48–72 Lamp Needle Surgical ≤6 ≥6
illumina- bleeding debride-
tion of flap ment
Arterial 3 1 1 1 0 0 0 1 1 2
insufficiency
Venous 9 0 2 6 1 9 2 2 2 7
reflux
disorder
Good growth 30 0 0 0 0 0 0 0 0 0
of skin flap
Total 42 1 3 7 1 9 2 3 3 9
Note: TFBBI, Time of flap becoming better after intervention
4 Discuss
Large area of soft and hard tissue defects after the radical surgery for malignant
tumors of oral and maxilla-facial region, often resulting in severe postoperative
facial deformity and dysfunction. Although the survival rate of various free tissue
flap has reached more than 90% [1,2,7,9], there are still a small number of
patients whose tissue flap dead due to various reasons. The rate of crisis of blood-
circulation is about to 3.5%–9.1% [8,10,14,15] according to domestic and foreign
literature. Once the free tissue flap dead, it will give patients a disastrous blow,
not only to extend the length of stay, increasing medical costs, more likely to
make the mortality rate increased. The success of flap repair surgery, in addition
to surgical techniques and equipment, but also thanks to good observation and
nursing measures.
4.1 Routine Care of Flap After Surgery for Tumors of Oral

and Maxilla-Facial Region
(1) Respiratory management

It is easy to lead to edema of Tongue and mouth bottom after surgery for
tumors of oral and maxilla-facial region. The establishment and maintenance
of airway patency is an important link, if necessary, preventable tracheostomy.
There are a group of 19 patients with prophylactic tracheotomy. Strictly accord-
ing to tracheotomy care routine, the use of saline and chymotrypsin atomization
inhalation, 3 times a day, timely suction, no case of pulmonary infection occurred
in patients. 1 case of buccal cancer free forearm flap transplantation, body fat,
short neck, dyspnea occurred on postoperative day 3, and lift the respiratory
distress after emergency tracheotomy.
(2) Position of care
Tissue flap transplantation area position nursing: Maintain the patient’s
head in the median or slightly biased side of surgery, non-distortions, effectively
reducing the flap microvascular anastomosis tension. Raising bed 15◦ –30◦ after
patients wake up is conducive to wound drainage and tissue valve venous return
in order to reduce the surgical area and head and face edema. Attention to night
patrol, and timely correct the patient is not correct position, to avoid the flap
vascular pedicle compression, traction, discount.
Tissue flap for the position of care: Raising the donor limb elevation in order
to facilitate venous return, and pay attention to finger blood supply and swelling.
At the same time observe the wound bleeding situation, keep the dressing clean
and Bandage intact. 1 patient of the group complaint the tissue flap for side
limb numbness after 6 h, and the acral skin is pale by physical examination,
skin temperature decreased, then immediately notify the physician to relieve
decompression, symptom relief after rewrapping, to ensure the donor limb of the
blood circulation.
1672 X. Yang et al.
(3) Ward management

Temperature of ward room is in 22–26◦ C, humidity 50% to 70%, and ward
disinfect with ultraviolet every day 2 times. Attention to keep warm, to prevent
crisis of blood-circulation due to vascular spasm by cold stimulation. Publicize
non-smoking to family members and visiting staff, so that small artery spasm
due to nicotine in tobacco, affecting the flap blood supply and induced arterial
crisis.
(4) Medication care
It is important to attention to vital signs of observation, in particular changes
in blood pressure which should be in 130/70 mmHg above, to ensure adequate
blood volume. At the same time the usage of anticoagulation, anti-spasm, anti-
infective drugs, routine use of low molecular weight heparin sodium, papaverine,
dextran, Xiangdan should be as prescribed. In addition, observe the effect of
medication and anticoagulant treatment side effects.
Strengthening nutrition promote incision growth. All patients underwent PG-
SGA score, according to the results of the nutritionist to intervene, to provide
patients with high-calorie, high protein and vitamin-rich diet plan, to prepare
specialized nutrient solution if necessary. 42 patients retain the stomach tube
to contact vacuum suction device, fasting on the day of surgery, to be drained
liquid light green for nasal feeding which done 200 mL/time, interval 2 h 1 time.
After the incision healed, the patient was instructed to practice swallowing.
Nasogastric tube could be remove when oral feeding is without exception, instead
of gold-type tube feeding (Using 50 mL syringe, a section of about 15–20 cm of
the infusion tube, the tube set in the syringe, the injection of food when the
skin placed in the mouth between the second and third molar). Diet from the
liquid gradually transition to semi-liquid and solid food, avoiding cold, sharp,
hard, with bone, barbed food.
(5) Nursing care of the irradiation of roasted lamps
This method can improve the local temperature of the flap, reduce arteriove-
nous vasospasm, as well as maintain microvascular smooth blood supply. With
40–60 W roasted lamp irradiation anastomosis, the use of distance 30–45 cm
avoid baking lamp close-up flap, to prevent local temperature too high. There
were nine cases of vascular crisis in this group, in the emergence of flap color
darkening, flap tension increased to give roasted lights irradiation, with the use
of vasodilator drugs in 2 cases after 6 h, 7 cases the flap color improved after 6 h.
(6) Nursing of wound drainage
Maintain the drainage area smooth, to prevent blood clots oppression arteri-
ovenous effects of blood circulation and venous return. A one-time high vacuum
drainage bottle contacting the end of drainage tube continued to attract, prop-
erly fixed drainage tube, regular squeeze, to avoid pipe distortion, compression,
clogging, shedding, close observation of drainage fluid color, and measure the
amount of drainage fluid and make a record 1 time/1 h within 24 h after surgery.
If the drainage beyond 400 ml/d, bright red color, solidification, bleeding may be
considered. If the amount is small, dark red, then alert subcutaneous congestion,
and report to the doctor timely. If the drainage lower 10–15 ml/d, estuation can
be considered after 48–72 h.
(7) Oral care
It was necessary to do oral management and to keep the flap area clean, if
there were oral bleeding, more exudate after surgery. Routine oral care should be
three times a day. Cotton balls should be along the line of blood vessels and pay
attention to grasp the intensity when scrubbing. If necessary, to do oral cleaning.
With 20 ml syringe aspiration of 3% hydrogen peroxide and saline as well as
compound chlorine has been mixed, it was repeatedly washed, companying with
a negative pressure. Note that suction should be on the contralateral, so as not
to damage the flap and incision.
4.2 The Care of the Usage Free Tissue Flap Colorimetric Card
in the Postoperative Flap Management for Tumors of Oral
and Maxilla-Facial Region
After free flap anastomosis, a common complication is crisis of blood-circulation,
divided into venous crisis and arterial crisis, usually occurring after 24–48 h [4].
The main monitoring indicators is the color, acupres sure and flap temperature
difference [5]. The change of the flap color is the main basis for recognition
and treatment. Early period of flap reconstruction, targeted nursing monitoring
measures can help find vascular complications, providing first-hand information
for clinical treatment. Weng [13] reported that flaps generally tolerate up to 6 h of
ischemia without severe necrosis, 75% of the flaps would be necrosis after venous
thrombosis continued 6 h and Flap necrosis rate would be 100% if beyond 8 h.
Therefore, there some studies [6,13] pointed out that occurrence of flap ischemia
and the timely monitoring of care is to determine the success of the key to the
success of flap transplantation.
(1) Content of flap observation
Skin color changes and capillary reflow measurement, is the early minimum
two indicators of reflection of blood circulation status of the most direct and
rapid, subject to outside interference, and receiving repeat verification. Where
both the skin color and capillary reflow abnormalities or more than two abnormal
monitoring indicators, there is the crisis of blood-circulation. Venous crisis is
usually manifested as red flap color, capillary response, and gradually become
dark red, dark red, late black, acupuncturing out of dark red or black liquid
when swollen hardened. Arterial crisis showed that the flap color is pale, lack of
elasticity, acupuncturing without blood flow. However, the clinical description
of the flap color and elastic temperature difference is often subjective, and the
evaluation result is limited by the professional skill and nursing condition. First-
line duty personnel for flap blood flow status judgment may be insufficient. If
the flap color change with not knowing whether abnormal and the person do not
have enough judgment experience, it may delay the flap vascular crisis rescue.
We set up a free tissue flap management team, the department head deputy
head appointed director and deputy director and head nurse as deputy head.
1674 X. Yang et al.
The backbone experts in the organization department access to a large number

of domestic and foreign literature and review past patients pictures and case
information, referred to the skin Flap vascular crisis indicators, combined with
20 years of clinical experience, to design the free tissue flap colorimetric card.
In the flap color, degree of pale, ruddy, cyanotic were distinguished carefully
to help clinicians to determine the extent of vascular crisis. At the same time,
the observation of flap vision and abnormal changes in the scope of color flap
provide accurate information for clinical care analysis, diagnosis and treatment
according to the actual data. For 1 case of buccal cancer patients in the group
after surgery 3 h, at 1:00 at night the flap color is pale 1 and skin temperature
is slightly lower, as well as the examination body find no swelling of the neck,
with plasma drainage tube patency and drainage of 15 ml dark red liquid. Doctor
asked to record once every 15 min. The flap color transact pale 2 close to the
pale 3 after 1 h and the neck was swollen with the drainage of the blood of
the liquid drainage tube bright red liquid 50 ml. Then immediately report the
surgeon, decisively send the patient to the operating room emergency line of
neck incision debridement and vascular crisis exploration as well as small vessel
anastomosis. After that, no longer found bleeding by observing 1 h, with the
wound closed and a smooth period of 72 h risk period, the patients was back
to the ward and at the latter part, voice, chewing, swallowing function of the
patient recovered well.
(2) Observation field and time of flap
Defects in the postoperative flap management for tumors of oral and maxilla-
facial region are generally located in the mouth. Especially the flap near the
tongue observed inconveniently, so it need to increase the patient mouth opening,
so that the full exposure of the flap. For 1 patient of the group within surgery
12 h, the wound edge bleed too much, the edge was dark purple and the area
continue to expand to the center. It observed color change is not obvious within
2 s by sterile cotton swab compression capillary filling degree. And with 1ml
syringe needle piercing the center of the flap, there were 0.5 cm dark red liquid
outflow. Then we immediately notify the surgeon back to the operating room
emergency treatment within 6 h after the flap color improved. There was another
cases of oral bottom free radical repair, whose flap color was purple bark 1 and
slightly faster response of the capillaries when back to the ward. Nurses found
visible spots at mouth flap near the tongue 1/5 and increased speckle area with
time in 13 h after surgery. When roasted lamp irradiation treatment is invalid,
we immediately notify the doctor back to the operating room line exploration,
and it found to be venous cramps combined embolism. Finally, the flap survived
after the embolization vein re-anastomosis.
(3) The rule of crisis of blood-circulation
Flap crisis often occurs at night, which is linked with nocturnal high vagal
tone, vascular intimal injury caused by hemodynamic changes, the role of nerve
on the blood vessels and hormone secretion [3]. Nurses should strengthen the
night patrol and attention to patient complaints as well as strict shift. There
was 1 patients with left squamous cell carcinoma resection and left cervical
lymph node dissection and thoracic duct ligation as well as common carotid
artery epineurium and sublingual resection and free forearm flap excision repair
and small vascular anastomosis and abdominal free skin graft repair complained
neck pressure, pain, physical examination after 17 h. And physical examination
only saw the flap suture and tongue bruising. But shift officers have not attracted
great attention, so that the white skin flap color deepening, the area expanded
to 1/2 or more when the shift after 6 h. It found negative pressure drainage
tube shift, subcutaneous hemorrhage obvious, vascular anastomotic pressure and
venous return blocked when sent to the operating room exploration. However,
there was no obvious improvement in the flap condition at 6 h after operation,
and the capillary reflex was not obvious. The elasticity also was poor and the skin
temperature was low. Then the doctor surgery to trim the skin flap, to strengthen
drainage, with local iodoformize gauze dressing and intravenous anti-infection
as well as blood circulation and other treatment, so the flap survived.
5 Summary
Crisis of blood-circulation is one of the main complications of vascular anasto-
mosis after vascular tissue transplantation. It is an important factor to timely
find problems and take appropriate measures to ensure the survival of the flap.
By using free tissue flap colorimetric card, it describe the flap color, capillary
filling and the scope of flap color changes with the objective data as well as
text, which makes the observation of free tissue flap more Simple and accurate.
Therefore, it is worthy to promote in clinic.
Acknowledgements. This research was supported by the project from Sichuan

Provincial Philosophy and Social Science Base and The key Research Base of Human-
ities and Social Sciences in Sichuan (XQ16C09).
References
1. Corbitt C, Skoracki RJ et al (2014) Free flap failure in head and neck reconstruc-
tion. Head Neck 36(10):1440–1445
2. Eckardt A, Meyer A et al (2007) Reconstruction of defects in the head and neck
with free flaps: 20 years experience. Br J Oral Maxillofac Surg 45(1):11–15
3. Guo Q, Lu L, Xu M (2012) Observation of vascular crisis after transplantation of
12 cases of anterolateral flap. Chin J Nurs 47(3):215–217 (in Chinese)
4. Guo QY, Lu L et al (2012) Observation and nursing care of vascular crisis in 12
patients after anterolateral thigh flap transplantation. Chin J Nurs 47(3):215–216
(in Chinese)
5. Huang Z, Tao S, Hu L (2012) Nursing of 12 cases of complex tissue flap for large
area soft tissue defect. Chin J Nurs 47(3):213–214 (in Chinese)
6. Jallali N, Ridha H, Butler P (2005) Postoperative monitoring of free flaps in UK
plastic surgery units. Microsurgery 25(6):469–472
7. Kesting MR, Hölzle F et al (2011) Microsurgical reconstruction of the oral cavity
with free flaps from the anterolateral thigh and the radial forearm: a comparison
of perioperative data from 161 cases. Ann Surg Oncol 18(7):1988–1994
1676 X. Yang et al.
8. Lu M, Yang X, Wang Y (2010) Prevention and treatment of vascular crisis in free

tissue flap of head and neck. J Oral Sci Res 26(2):243–245 (in Chinese)
9. Mao C, Yu G, Peng X (2002) Clinical analysis of 50 cases of free forearm flap for
oral and maxillofacial. J Mod Stomatol. 16(2):138–140 (in Chinese)
10. Nakatsuka T, Harii K et al (2003) Analytic review of 2372 free flap transfers for
head and neck reconstruction following cancer resection. J Reconstr Microsurg
19(19):363–369
11. Qiu W (2006) Recent progress in reconstruction of oral and maxillofacial surgery.
Chin J Repar Reconstr Sur 20(4):339–343 (in Chinese)
12. Wang Q, Wang S et al (2010) Nursing care of 58 cases with microsurgical flaps for
repairing soft tissue defects of legs complicated with. Chin J Nurs 45(8):702–704
(in Chinese)
13. Weng X, Wang J, Ye M (2010) Prevention and treatment of complications after
maxillofacial defect repair. Chin J Stomatol Electron Ed 4(1):71–73 (in Chinese)
14. Ye WM, Zhang ZY et al (2004) clinical observation and treatment of failing flaps in
oral and maxillofacial surgery: a report of 44 cases. Shanghai J Stomatol 13(2):91
15. Zhang C, Zhang Z et al (2005) [Repair and reconstruction of oral and maxillofacial
defect-clinical analysis of 1973 cases]. Chin J Repar Reconstr Sur 19(10):773
The Impact of Intellectual Capital on High-Tech
Enterprise Performance: Applied Study
in China’s Second-Board Market
Hongchang Mei(B) and Kunling Wang
School of Administration, Chongqing Technology and Business University,

1426315339@qq.com
Abstract. This paper chooses high tech enterprises listed on China’s

Second-board Market as research objects. The aims of this thesis are to
find out the impact of intellectual capital on the performance of high-
tech enterprises. This study has carried out an empirical analysis with
the method of multiple linear regression. The empirical findings indicate
a trend that for high-tech industries on the GEM, the improvement of
their performance efficiency depends more on the human capital com-
pared to structural capital and physical capital. Both human capital
and structural capital significantly influence the performance efficiency
of enterprises in the high-tech industry, while physical capital practically
doesn’t impact it significantly. According to the empirical findings, this
paper has put forward some advices on capital allocation for Chinese
high-tech enterprises.
Keywords: Intellectual capital (IC) · Performance · High-tech

enterprises
1 Introduction
In the industrial age, human capital is the driving force of economic growth,
but in the age of knowledge economy, intellectual capital has become the power
source of economic growth, which is also changing the rules of business and
national competitiveness. Also, nowadays intangibles are almost universally con-
sidered as the main value drivers for companies [12].
China as a rapidly developing country, has strongly encouraged entrepreneur-
ial activities especially in high-tech fields. Similarly, intellectual capital seems
more important than any other factor in the fields of high-tech, that intellectual
capital management have become the domain of the so-called Chief Knowledge
Officer [1], there is a great need to study the impact of intellectual capital in
Chinese high-tech enterprises. The result will not only aid in selection for man-
aging intangibles, but also accelerate the pace of growing of high-tech companies.

1678 H. Mei and K. Wang
This research aims to investigate the impact of intellectual capital on the

performance of selected enterprises in the high-tech sector on the Second-board
market in China. The first section deals with the introduction of intellectual cap-
ital and related researches. This follows with research methodology and design.
In the third section, the research empirical findings are presented. Finally, the
conclusion outlines the limitations and implication of this study.
2 Literature Review
2.1 Intellectual Capital and Sub-Components of Intellectual Capital
The word Intellectual capital was first proposed by Senior as a synonym for
human capital in 1836, which he believes is a combination of knowledge and skills
of human beings. John Galbraith extended the concept of intellectual capital,
noting that intellectual capital was not just knowledge or pure intellect, but
also the process of making use of knowledge effectively. Guenther, Beyer [11]
argued that intellectual capital represents the sum of all the intangible assets of
the company. Also Intangible assets are defined as the company’s non-material
and non-financial resources, including technology, customer information, brand
name, honor and corporate culture. The intellectual capital of an enterprise can
improve the performance of an enterprise by relying on the dynamic interaction
of IC components [4].
IC is also considered as the sum of following three categories: (1) Human
capital, which refers to the people in an organization and describes their cumu-
lative tacit knowledge skills. (2) Structural capital, which consists of supportive
infrastructure, processes, and databases of corporation that enable human cap-
ital to function [8]. (3) Relational capital.
Another definition is that IC is derived insights about head value, future
earnings capabilities, based on human capital, as well as organizational, struc-
tural and relational capital [5]. Some scholars hold the view that intellectual
capital can be classified into human capital and structural capital, divides intel-
lectual capital into human capital and structural capital, which is accepted by
author of this article.
2.2 Relation Between IC and Firm Performance

Firer did an empirical research on the relationship between the three components
of value added intellectual coefficient (VAIC) in South Africa. He concluded that,
if the decision makers focus more on exploiting the physical capital reward, it
would lead to the neglect of the use of human capital.
The human capital is very important to the enterprise performance, that com-
panies should value it while managing their IC. Bontis [2] conducted a study on
the efficiency of IC to enterprise value in service industry and other industry
in Malaysian. The results show that human capital in various industries is an
important role as structural capital and is positively correlated with the per-
formance of enterprises. Fengetal [6] proposed that it was necessary to clarify
The Impact of Intellectual Capital on High-Tech Enterprise Performance 1679
the relationship between corporate performance and human capital with more
attention to human capital.
They also have mentioned that performance of enterprises should be con-
sidered by combining the characteristics and degree of competition of specific
industry. Zhu et al. [13] assumed that human capital is the key factor to deter-
mine enterprise performance.
2.3 IC Measurement Models
IC is difficult to measure as an intangible asset. So far there is no unified view

on the measurement method for intellectual capital. The current measurement
models are as follows.
(1) Difference between market value and book value, which assumes that the
intellectual capital of enterprises could be the difference between their market
value and book value.
(2) Tobin’sq, which refers to the ratio between a physical asset’s market value
and its replacement value. It was proposed in 1968 by James Tobin and
William Brainard [3]. If the replacement cost of the company’s assets is
lower than the company’s market value, the company’s investment behaviors
would have obtained excess profits, which are generated from IC [10].
(3) VAICAnte Public proposed Value Added Intellectual Coefficient (VAIC) in
1998 to measure the intellectual capital value of an enterprise. According
to this method corporate capital consists of physical capital and intellec-
tual capital, and corporate performance depends on the efficiency of value
added by a firm from its physical capital and intellectual capital resources.
This method assumed that the enterprise’s intellectual value-added efficiency
represented its performance.
Lentjušenkova and Lapina [7] proposed that, in a modern organization, its

managerial approaches and investment facilities are changing along with the
growing importance of IC and the changes of its components, thereby changing
the managerial approaches and focusing on value creation.
The paper chooses the VAIC method to analyze impacts of intellectual capital
on the performance of high-tech enterprises in China. In this sector, business in
the following specific industries are preferred, like high-growth, high-tech, high
value-added and new economy, new services, new agriculture, new energy and
new materials, that possess high growth potential with high risk [9]. In these
industries, the value and function of intellectual capital seems have exceeded
financial capital.
3 Research Methodology and Design
This study is a longitudinal research in a time-span 2013–2015 and refers to

China. Based on the developed VAIC model by Firer, this paper analyzes the
correlation and regression between value-added coefficients of IC components and

high-tech enterprise performance. And then explains the relationship between
intellectual capital components and enterprise performance based on the analysis
results. In this model, it is assumed that value creation of enterprise mainly
comes from physical capital resources and intellectual capital resources.
3.1 Research Hypotheses
Economic growth mainly depends on physical capital and intellectual capital.

For the composition of intellectual capital in this paper, it is assumed that IC
includes human capital and structural capital. Human capital is the sum of
knowledge, skills, experience of people in the corporate. IC can be performed
in the products and services, knowledge level and education level of corporate
members. This paper uses the index of human capital value-added coefficient
to represent its intensity. Structural capital is an organizational asset that is
embedded in organizational structure, management system standard and corpo-
rate culture. In this paper, the structural capital intensity will be represented
by its value-added coefficient.
Hypothesis 1 : The intellectual capital and physical capital of an enterprise

cause the performance of an enterprise.
Hypothesis 2 : Intellectual capital consists of human capital and structural
capital.
Hypothesis 3 : For high-tech enterprises, the impact of intellectual capital on
corporate performance is significant.
3.2 Sample Selection and Data Sources
The sample of this article comes from the financial statements of 80 companies
listed on the GEM from 2013 to 2015. Research samples in this paper are all
certified as high-tech by related government agency and listed on China’s second-
board market in Shenzhen stock exchange from 2013 to 2015. In order to ensure
the validity of the data, the criteria for screening the original samples are as
follows:
(1) Selecting the samples of the listed GEM companies, which are in continuous
operation during this period to ensure the necessary continuity of the sample
set.
(2) Financial services listed companies are eliminated because of their large total
share capital, which may cause deviation of the regression results.
(3) Excluding listed companies of which annual financial statement data are
incomplete, or not continuously operating in the time-span 2013–2015.
(4) Elimination of abnormal operating listed companies (e.g. Debt ratio surpass-
ing 100%) and companies with deteriorating financial condition to ensure the
validity of the sample data.
(5) Listed companies who have changed their main business because of mergers
and acquisitions or other acts will not be included in the sample set. The
final sample set includes 80 high-tech enterprises. This article used EXCEL
and EVIEWS9.0 in the processing of sample data.
3.3 Research Model

Enterprise’s ability to creating value by using physical capital and intellectual
capital is called “intellectual ability”, which is expressed by VAIC model. This
model consists of capital employed efficiency (CEE), human capital efficiency
(HCE) and structural capital efficiency (SCE). Relation between them is that,
VAICi = CEEi + HCEi + SCEi .
The main purpose of this article is to study the impact of intellectual capital
on the performance of high-tech enterprises. By establishing the multiple linear
regression model and doing a statistical hypothesis test, we obtained the esti-
mation for the regression coefficient. Based on all the above conditions, we will
interpret the impact of IC on performance of Chinese high-tech Enterprises.
The basic model used to estimate the binomial logistic regression models is:
ROA = α + β1 CEE + β2 HCE + β3 SCE + β4 Leverage + β5 LNSIZE + ε.
4 Research and Discussion

4.1 Variable Selection
(1) Selection of dependent variable
Interpretations of selected variables in this paper are as follows:
Table 1. Dependent variable
Variable symbol Variable explanation Variable expression

Dependent Variable ROA Return on Assets Net Income after Taxes/Assets
ROA shows the percentage of how profitable a company’s assets are in gener-
ating revenue. In the following analysis we use ROA to express the performance
of a company (Table 1).
(2) Selection of independent variables
(See Table 2)
1 CE is for the book value of total assets of an enterprise.

2 HC represents human capital of an enterprise which is measured by total
wage cost of an enterprise. The total wage cost in this article is indicated by
cash paid to employees as salary and other payments in cash flow statement
of a listed company.
Table 2. Independent variables
Variables symbol Variables explanation Variable expression

Independent variables CEE Capital Employed Efficiency CEE = VA/CE
HCE Human Capital Efficiency HCE = VA/HC
SCE Structural Capital Efficiency SCE = SC/VA
3 VA expresses the value added of enterprise, which is a sum of net income,

income tax and total wage cost.
4 Income taxes cost is used in place of income tax in this thesis.
5 With income tax cost as a approximation of income tax.
(3) Selection of control variables

(See Table 3)
Table 3. Control variables.
Variables symbol Variables explanation Variable expression

Control variables LNSIZE Logarithm of Main SIZE = LN (Total assets)
Business Income
Leverage Debt ratio Leverage = Total debt/Total assets
1 The size of the assets impacts on the value of enterprises. IT is usually

expressed by the total assets of enterprises. In order to reflect the size of
the assets of enterprises accurately, this paper selects the logarithm of main
business income as a control variable.
2 Debt Ratio is a financial ratio that indicates the percentage of a company’s
assets that are provided via debt. It can be used to measure the ability of
a firm to operate. Therefore, we use the Debt Ratio together as the second
control variable.
4.2 Model Construction
On the basis of previous theoretical analysis, this paper constructs a multiple

regression model between physical capital efficiency, human capital efficiency,
structural capital efficiency and firm performance. The model is set as follows:
ROA = α + β1 CEE + β2 HCE + β3 SCE + β4 Leverage + β5 LNSIZE + ε.
Here: α is for the constant, β1 − β5 are corresponding coefficients. ε represents

the random disturbance.
Table 4. Descriptive statistics of variables.
Observations 2013 2014 2015

Mean Std. dev Mean Std. dev Mean Std. dev
ROA 0.061608 0.070561 0.065428 0.097572 0.068864 0.063246
LNSIZE 20.20134 0.855858 20.38727 0.888847 20.59145 0.8757
Leverage 0.27982 0.329841 0.224613 0.227157 0.304221 0.320626
CEE 0.174146 0.155064 0.213168 0.232301 0.196046 0.250991
HCE 2.095241 1.147357 2.430141 3.53092 2.484896 4.281109
SCE 0.309643 0.658739 0.41756 0.340202 0.343141 0.790112

(1) Sample description statistics
We have described statistical relationships between variables in the model, that
empirical results are as Table 4.
From results of descriptive statistics, we can see that the three-year average of
ROA of high-tech enterprises on the China’s second-board market are 0.061608,
0.065428 and 0.068864, which show that the performance of the high-tech enter-
prises listed on the China’s second-board market is increasing year by year. The
means of CEE in the sample period respectively are: 0.174146, 0.213168 and
0.196046, which means the change of physical capital is not very stable. But
their standard deviations are: 0.155064, 0.232301 and 0.250991, which shows an
increasing trend with time. It indicates that the difference between impacts of
physical capital on each enterprise’s performance is widening per year. The mean
2.095241, 2.430141 and 2.4484896 of HCE over sample period, which is showing
an increasing trend, indicates that the influence of human capital on high-tech
firms’ performance listed on Chinese Second-board market is gradually increas-
ing. The mean 0.309643, 0.417560 and 0.3343141 of SCE in these three years is
showing an unstable trend without regularity.
We can also see from the above figure that mean of HCE is obviously greater
than the mean of CEE and SCE during the period from 2013 to 2015, which
tells that human capital of high-tech enterprises contribute more to the corporate
performance than physical capital and structural capital.
(2) Partial correlation analysis for variables
In order to make the multiple regression analysis meaningful, we have analyzed
the correlation between selected variables. The analysis results are as follows
(Table 5):
It can be seen from the above analysis, that, there is a positive correlation
between HCE, SCE, CEE and ROA. However, the correlation coefficient between
ROA and CEE in 2014 is 0.0638, with the concomitant probability of T test
0.5712, which hasn’t passed the confidence level of 1% test. This indicates that
the impact of physical capital on firm performance is only a little in this year, The
correlation coefficients of the other two years both correspond the 1% confidence
Table 5. Partial correlation analysis for selected variables.
ROA CEE HCE SCE Leverage LNSIZE

2013 ROA Correlation 1 0.478 0.6902 0.4722 0.4403 0.6734
Probability - 0 0 0 0 0
CEE Correlation 0.478 1 0.75 0.4914 0.712 0.7435
Probability 0 - 0 0 0 0
HCE Correlation 0.6902 0.75 1 0.634 0.5581 0.8792
Probability 0 0 - 0 0 0
SCE Correlation 0.4722 0.4914 0.634 1 0.3012 0.4356
Probability 0 0 0 - 0.0063 0
Leverage Correlation 0.4403 0.712 0.5581 0.3012 1 0.6431
Probability 0 0 0 0.0063 - 0
LNSIZE Correlation 0.6734 0.7435 0.8792 0.4356 0.643135 1
Probability 0 0 0 0 0 -
2014 ROA Correlation 1 0.0638 0.4471 0.4426 0.342 0.5711
Probability - 0.5712 0 0 0.0018 0
CEE Correlation 0.0638 1 0.427 0.6259 0.6802 0.6705
Probability 0.5712 - 0.0001 0 0 0
HCE Correlation 0.4471 0.427 1 0.6236 0.3954 0.5731
Probability 0 0.0001 - 0 0.0003 0
SCE Correlation 0.4426 0.6259 0.6236 1 0.5557 0.7725
Probability 0 0 0 - 0 0
Probability 0.0018 0 0.0003 0 - 0
Probability 0 0 0 0 0 -
2015 ROA Correlation 1 0.4335 0.4596 0.3742 0.4386 0.7515
Probability - 0.0001 0 0.0006 0 0
CEE Correlation 0.4335 1 0.418 0.3674 0.8234 0.6076
Probability 0.0001 - 0.0001 0.0007 0 0
HCE Correlation 0.4596 0.418 1 0.3517 0.456 0.508
Probability 0 0.0001 - 0.0013 0 0
SCE Correlation 0.3742 0.3674 0.3517 1 0.3422 0.3996
Probability 0.0006 0.0007 0.0013 - 0.0018 0.0002
Probability 0 0 0 0.0018 - 0
Probability 0 0 0 0.0002 0 -
level of T test. At the same time, there are positive correlations between HCE,
SCE and CEE, and all of them have passed T test with confidence level of
1%. This result tells that influencing factors of business performance are not
independent of each other, but interacting with each other.
(3) Multiple regression
The regression results have shown that the results are not very ideal. Only the
value of the intercept term in there gression equations passed the T test for linear
regression (here 5%), with CEE, HCE, SCE coefficients did not. This result tells
that the coefficients are not very credible. However, the F-test values 6.107650,
9.607723 and 5.317085 of each model have passed the F-test for linear regression
(here 5%), which indicates explained variables in the model could be explained
by explanatory variables (Table 6).
Table 6. Multiple regression model.
Time 2013 2014 2015

Dependent variable ROA ROA ROA
Coefficient t-Statistic Coefficient t-Statistic Coefficient t-Statistic
Constant −0.6637 −3.7726 −0.476358 −2.175358 −0.645427 −3.986521
(0.0003) (0.0328) (0.0002)
Independent CEE −0.0602 −1.1093 −0.237977 −5.567514 0.04047 1.134204
variables (0.2709) (0.0000) (0.2604)
HCE 0.0177 2.3107 0.004414 1.623358 0.001087 0.703479
(0.0236) (0.1088) (0.484)
SCE 0.005 0.3683 0.03222 1.119615 0.008388 1.020737
(0.7137) (0.2665) (0.3107)
Control Leverage 0.034184 1.427739 0.077395 1.812786 −0.025991 −0.927511
variables (0.1576) (0.0739) (0.3567)
LNSIZE 0.034031 3.966973 0.027024 2.563172 0.034416 4.429544
(0.0002) (0.0124) (0.0000)
Adjusted R-squared 0.244296 0.352664 0.214598
F-statistic 6.10765 9.607723 5.317085
Durbin-Watson stat 1.626685 2.100722 1.85182
Besides, we can see that the coefficients of CEE in 2013 and 2014 respectively
are −0.0602 and −0.237977, with 0.040470 in 2015, which indicates that CEE
has a negative effect on ROA. Coefficients of other explanatory variables to
ROA are positive, which tells that human capital and structural capital drive
performance of high-tech enterprises. DW 1.626685, 2.100722, 1.851820 in the
time-span are all close to 2, indicating that the regression equation does not have
strong autocorrelation.
5 Conclusion
5.1 Limitations
The results seem to be achieved, but there are some limitations of the research.
The first concerns the fact that the sample set limited to listed high-tech compa-
nies, because their financial data are available, while financial data of non-listed
high-tech companies are not. Accordingly, carrying out further research linking
IC and non-listed high-tech enterprises could develop this research field.
Another limitation concerns that some indicators of basic data used substi-
tutions due to the availability of fundamental data on each company’s financial
report. So it needs more in-depth, exhaustive analysis to be carried out. A fur-
ther limitation is, this analysis considers only the time-span 2013–2015 without
consideration of environmental differences.
5.2 Implications for High-Tech Enterprises in China

From the above empirical research, we find that human capital and structural
capital have significant positive correlation with the performance of listed high-
tech enterprises, while physical not significant. For the high-tech enterprises
listed on China’s GEM, human capital of intellectual capital has greater impact
on the corporate performance than structural capital and physical capital does.
Structural capital has relatively less effect compared to other human capital
on high-tech firms’ performance. However, high-tech enterprises still need to pay
attention to structural capital, and strengthen enterprise management. The aver-
age value-added efficiency of intellectual capital is bigger than that of physical
capital, which reflects that the dependence of enterprise performance on intel-
lectual capital is increasing in high-tech field. The author puts forward based on
this empirical research following suggestions:
(1) High-tech enterprises should devote more efforts to the development and
management of the intellectual resources of employees, and give full play to
the intelligence of them.
(2) The focus of high-tech enterprises should be transferred from physical capital
to intellectual capital.
(3) Designing enterprise size according to IC of a company will help to operate
structural capital and physical capital more reasonable, and thus enhance
their efficiency of creating corporate value.
References
1. Bontis N (2001) CKO wanted-evangelical skills necessary: a review of the chief
knowledge officer position. Knowl Process Manage 8(1):29–38
2. Bontis N, Keow W, Richardson S (2002) Intellectual capital and business perfor-
mance in malaysian industries. J Intell Capital 1(1):85–100
3. Brainard WC, Tobin J (1968) Pitfalls in financial model building. Cowles Found
Discuss Pap 58(2):99–122
4. Chen Y (2011) Formation and development of “new over-the-counter market”.

China Secur Futures China 11:15–24 (in Chinese)
5. Edvinsson L (2013) IC 21: reflections from 21 years of ic practice and theory. J
Intell Capital 14(1):163–172
6. Feng L (2007) An empirical analysis on the relationship between human capital
and firm performance. Commun Finance Account 2:66–69
7. Lentjušenkova O, Lapina I (2016) The transformation of the organization’s intel-
lectual capital: from resource to capital. J Intell Capital 17(4):610–631
8. Maddocks J, Beaney M (2002) See the invisible and intangible. Knowl Manage
3:16–17
9. Meiou L (2013) Research on China gem development. Master’s thesis, Fudan Uni-
versity, Shanghai (in Chinese)
10. M’Pherson PK, Pike S (2001) Accounting, empirical measurement and intellectual
capital. J Intell Capital 2(3):246–260
11. Subramaniam M, Youndt MA (2005) The influence of intellectual capital on the
types of innovative capabilities. Acad Manag J 48(3):450–463
12. Zambon S (2016) Ten years after: the past, the present and the future of scholarly
investigation on intangibles and intellectual capital (IC). J Intell Capital 17(1):42–
50
13. Zhu X (2003) Human capital and corporate performance. Guangxi Soc Sci 1:75–77
(in Chinese)
Site Selection of Public Fast Electric Vehicle
Charging Station by Using an Intuitionistic
Fuzzy Projection-Based Approach
Lin Zhong1 and Zhibin Wu2(B)

1
2
zhibinwu@scu.edu.cn
Abstract. As an important solution to reduce the dependency on crude

oil and minimize transportation-related carbon dioxide emissions along
with other pollutants, electric vehicles (EVs) are broadly promoted by
government. As the energy provider of EVs, electric vehicle charging
stations (EVCSs) play important roles in electric vehicle popularization.
The effective site selection of EVCS is in favour of promoting the willing
of consumer to use EVs. In this paper, we apply multi-criteria group
decision-making (MCGDM) method to site the public fast electric vehicle
charging stations. To avoid the information lost in aggregation process,
we use an intuitionistic fuzzy projection-based approach. The proposed
method is applied into a numerical example, and the results show that
the proposed MCGDM is feasible and efficient for EVCS site selection.
Keywords: Intuitionistic fuzzy set · Electric vehicle · Fast charging

station · Multi-criteria decision making algorithm
1 Introduction
Electric vehicles (EVs) are considered as a significant solution to reduce the
dependency on crude oil and minimize transportation-related carbon dioxide
emissions along with other pollutants [7]. As the energy provider of EVs, elec-
tric vehicle charging station (EVCS) is the foundation of electric vehicle indus-
try development [4]. Efficient, convenient and economic EVCS is in favour of
promoting willingness of consumers and also the industry development. As the
preliminary work of EVCSs constructions, the EVCS site selection is quite impor-
tant in the whole life cycle, which has significant impacts on the service quality
and operational efficiency of EVCS. Therefore, it is necessary to employ proper
methods to determine the optimal EVCS site.
Li et al. [6] considered dynamic origin-destination trip satisfaction and pub-
lic electric vehicle (EV) charging network expansion to develop a multi-period
Site Selection of Public Fast Electric Vehicle Charging Station 1689
refueling location model. Sheppard et al. [8] applied an agent-based simulation

modeling platform to site EVCSs in Delhi. Antunes et al. [3] used a maximal
coverage model to determine the optimal location, quantity and capacity for
EVCSs. Although many models have been proposed to site EVCS, few papers
have considered multi-criteria group decision making (MCGDM) techniques.
In classical decision-making problems, the preference value for each alterna-
tive expressed by each decision maker is precise [2]. Due to the fuzziness and
uncertainty of decision-making problems and the inherent vagueness of human
preferences, however, the best way for decision-makers to give their opinion is
using natural language [11]. For that reason, linguistic variables [18], the value
of which are expressed using words and sentences from natural language, have
been widely used to indicate the preference values for each attribute. Aggregating
linguistic information has generally been discussed under four main schemes [10].
The purpose of this paper is to apply the intuitionistic fuzzy projection-based
approach to determine the optimal site for public fast electric vehicle charging
station to promote the popularization of EVs in Chengdu. This framework differs
from the traditional method as the weights of decision makers are undetermined
and the complete process is conducted in an intuitionistic fuzzy environment.
The remainder of this paper is organized as follows. In Sect. 2, we briefly review
the basic concepts of intuitionistic fuzzy sets. The proposed fuzzy multi-criteria
group decision making framework is presented in Sect. 3. Then, in Sect. 4, the
proposed MCGDM is applied to determine the site for PFEVCSs. Finally, con-
clusions are drawn in Sect. 5.
2 Preliminaries
In this section, we briefly review the basic concepts of IFSs.
Definition 1. [17] Let X be a universe of discourse, then a fuzzy set A is defined

as
A = {x, μA (x)|x ∈ X, μA (x) ∈ [0, 1]}, (1)
where μA (x) denotes the membership degree of the element x to the set A.
Definition 2. [1] Let X be a universe of discourse, then an intuitionistic fuzzy

set A is shown as follows:
A = {x, μA (x), νA(x) |x ∈ X, μA (x) ∈ [0, 1], νA (x) ∈ [0, 1]}, (2)
where μA (x) and νA (x) respectively represent the membership degree and non
membership degree of the element x to the set A with the condition 0 ≤ μA (x) +
νA (x) ≤ 1.
Definition 3. [1] For each intuitionistic fuzzy set A, if
πA (x) = 1 − μA (x) − νA (x), ∀x ∈ X, (3)

1690 L. Zhong and Z. Wu
then, πA (x) is called the indeterminacy degree or hesitant degree of x to the set
A. The bigger the πA (x), the more the indeterminacy degree of knowledge about
x [15]. Especially, when μA (x) + νA (x) = 1, for ∀x ∈ X, the intuitionistic fuzzy
set A is reduced to an ordinary fuzzy set [9].
Definition 4. [14] Suppose there are two intuitionistic fuzzy numbers α =

(μα , να ) and β = (μβ , νβ ), in which μ ∈ [0, 1], ν ∈ [0, 1], μ + ν ≤ 1 and
π = 1 − μ − ν. Let λ be a real number, then the arithmetic operations are
defined as follows:
(1) Addition operation
α + β = (μα + μβ − μα μβ , να νβ ); (4)
(2) Multiplication by a real number operation
λα = (1 − (1 − μα )λ , νλα ), λ > 0. (5)
Definition 5. [15] Let α = (μα , να ) and β = (μβ , νβ ) be two intuitionistic fuzzy

numbers, then projection of the α on β is defined as follow:
α·β μα μβ + να νβ + πα πβ
P rojβ (α) = = . (6)
|β| μ2β + ν2β + πβ2
In general, the larger the value of P rojβ (α), the more the degree of the α
approaching the β.
Definition 6. [15] Let A = (αij )mn and B = (βij )mn are two intuitionistic
fuzzy matrices. Then the projection of A on B is defined as follows:
m n α β α β α β
i=1 j=1 (μij μij + νij νij + πij πij )
P rojB (A) = β . (7)
m n 2 + (νβ )2 + (π β )2
i=1 j=1 (μ ij ) ij ij
3 A Framework for MCGDM

The MCGDM problems have some common characteristics: multiple alterna-
tives, multiple criteria and multiple decision makers. The aim of this kind of
decision is to screen and rank the alternatives based on the preferences expressed
by decision makers. Due to different education background and subjective per-
ception, different decision makers have difference importance during decision-
making process. Besides, it is quite common that there is no decision makers’
weights information for the MCGDM, even decision makers’ weights are deci-
sive. Therefore, we construct a new framework to deal with MCGDM problems
without any decision makers’ weights information, as shown in Fig. 1.
Suppose that there are m alternatives Ai , i ∈ {1, 2, · · · , m} to be ranked. The
performances of n criteria Cj , j ∈ {1, 2, · · · , n} are defined in linguistic terms
which are obtained from t decision makers Dk , k ∈ {1, 2, · · · , t}. Each decision
Formulate the Identify the decision Select the criteria. Construct the
alternatives. makers. decision makers'
( A1, A2 , Am) (C1, C2 , , Cn ) evaluation matrices.
(D1, D2 , , Dt )
calculate the Normalize Translate

projection of each Determine the
the decision linguistic terms into
decision matrix on PID and NID. IFSs.
PID and NID. matrices.
Aggregate the Determine the Calculate the ranking

Determine the
decision matrices projection of each values for alternatives
weight of each
and detemine the alternative on PIA and determine the
decition maker.
PIA and NIA. and NIA. ranking order.
Fig. 1. The framework of proposed method
maker expresses his preference to each alternative regarding to each criterion

using linguistic terms. Then decision matrices are constructed based on the all
preferences, shown as follows:
⎛ k k k
⎞
r11 r12 · · · r1n
⎜ r21k k
r22 k ⎟
· · · r2n
⎜ ⎟
Dk = ⎜ . .
.. . . ... ⎟
. . (8)
⎝ .. ⎠
k k k
rm1 rm2 · · · rmn
3.1 Decision Makers’ Weights Determination
TOPSIS, which was first developed by Hwang and Yoon [5], is a technique for
establishing the order of preference by its similarity to the ideal point [13].
Inspired by [13,16], we present an extended TOPSIS method for the decision
makers’ weight determination under an intuitionistic fuzzy environment.
First, we construct the positive ideal decision (PID) matrix D+ based on the
average of all decision matrices, which is treated as a cursory group consensus
of decision-making group. The equation is as follows:
⎛ + + + ⎞
r11 r12 · · · r1n
⎜ r21+
r22
+ + ⎟
· · · r2n
⎜ ⎟
D+ = ⎜ . .. .. ⎟ , (9)
⎝ .. . . ⎠
rm1
+
rm2
+
· · · rmn
+
where
t
1 k
rij
+
= rij . (10)
t
k=1
Then, the negative ideal decision (NID) matrix D− is determined as the

most indeterminate decision matrix, which means that the indeterminacy degree
of worst decision matrix possesses is the highest in all decision matrices. The
equation is as follows:
⎛ − − − ⎞
r11 r12 · · · r1n
⎜ r21
− −
r22 − ⎟
· · · r2n
⎜ ⎟
D− = ⎜ . ,
.. . . . ... ⎟
. (11)
⎝ .. ⎠
− − −
where
−
rij = (min μkij , max νkij ). (12)
t t
Since PID and NID have been determined, we consider the projection of
each decision matrix on the ideal decision matrices. According to Eq. (7), the
projection of Dk on D+ can be calculated as follows:
m n Hk H+ k + k +
k i=1 + νH H H H
ij νij + πij πij )
j=1 (μij μij
ProjD+ (D ) = . (13)
m n H+ 2 H+ 2 H+ 2
i=1 j=1 (μij ) + (νij ) + (πij )
The projection of Dk on D− can be calculated as follows:

m n Hk H− Hk H− Hk H−
k i=1 j=1 (μij μij + νij νij + πij πij )
ProjD− (D ) = H− . (14)
m n 2 + (νH − )2 + (π H − )2
i=1 j=1 (μij ) ij ij
After that, the relative closeness of each individual decision matrix is deter-
mined as follows:
ProjD− (Dk )
RC k = . (15)
ProjD− (Dk ) + ProjD+ (Dk )
The relative closeness is used to determine the rank order for all decision
makers. Obviously, the closer to the best decision result, the larger the value of
RC k is. Therefore, based on the relative closeness, the weight of each decision
maker is obtained as follows:
RC k
k = t (16)
k
k=1 RC
where k represents the wight of the kth decision maker, and k ≥ 0,

t
k=1 k = 1. We use the vector = (1 , 2 , · · · , t ) to represent the decision
maker’s weight vector.
3.2 Alternatives Ranking Order Determination

As the weights for decision makers are determined, the aggregated decision
matrix can be calculated as follows:
D = (D1 , D2 , · · · , Dt ) · (1 , 2 , · · · , t )T
⎛ ⎞
r11 r12 · · · r1n
⎜ r21 r22 · · · r2n ⎟
⎜ ⎟
=⎜ . .. . . .. ⎟ , (17)
⎝ .. . . . ⎠
where rij is the group preference IFS of alternative Ai respect to attribute Cj

and
t
k
rij = k · rij . (18)
k=1
Then, the positive ideal alternative is determined as follows:
A+ = (r11
+
, r12
+
, · · · , r1n
+
), (19)
where
rij
+
= (max μij , min νij ). (20)
m m
The positive ideal alternative (PIA) should has the maximal separation from the
PIA A+ . It is a very natural idea that we consider following decision:
− − −
A− = (r11 , r12 , · · · , r1n ), (21)
where
−
rij = (min μij , max νij ). (22)
m m
Then, the projection of each alternative on the ideal alternatives are calcu-
lated as follows:
n i A+ Ai A+ Ai A+
(μA
ij μij +νij νij +πij πij )
ProjA+ (Ai ) = j=1
n , (23)
A+ 2 A+ 2 A+ 2
j=1 (μij ) +(νij ) +(πij )
n i A− Ai A− Ai A−
(μA
ij μij +νij νij +πij πij )
ProjA− (Ai ) = j=1
n . (24)
A− 2 A− 2 A− 2
j=1 (μij ) +(νij ) +(πij )
Finally, the relative closeness is defined to determine the rank order for all
alternatives. We have
ProjA− (Ai )
RVi = (25)
ProjA+ (Ai ) + ProjA− (Ai )
It is obvious that the bigger RVi , the better alternative Ai is.

4 Site Selection for PFEVCS

PFEVCS site selection, as the preliminary work of EVCS construction, is quite
important in the whole life cycle, which has significant impacts on the service
quality and operational efficiency of EVCS. In this section, the proposed method
is employed to select the optimal EVCS site from all the alternatives. In order
to determine the possible alternatives and obtain the linguistic preference rat-
ings, three groups of expert panels (k = 1, 2, 3) with expertise in the fields of
environment, economy, society were formed. After reviewing the project fea-
sibility research reports, the expert panels finally selected four EVCS sites as
alternatives A1 , A2 , A3 and A4 . EVCS sites alternatives A1 , A2 , A3 and A4 are
respectively located at Wuhou district, Qingyang district, Jinniu district and
Chenghua district in Chengdu.
The MCDM problem related to EVCS site selection includes three criteria
(environment, economy and society) and eight sub-criteria (j = 1, 2, · · · , 8).
After reviewing the general information of all the alternatives, each expert group
gave the linguistic ratings judgments for the sub-criteria performance of each
alternative, and the results are listed in Table 1. According to the transform
rules in Table 2, we transform the linguistic terms into intuitionistic fuzzy sets.
Step 1. We transform the decision matrices into intuitionistic fuzzy sets, and
then determine the PID and NID respectively based on Eqs. (9)–(12).
The projection of each decision matrix on PID and NID are calculated
using Eqs. (13), (14). Then, according to Eq. (15), (16), the weights for
decision makers are calculated as = (0.3351, 0.3317, 0.3332).
Table 1. The decision matrix of decision makers
Decision makers Alternatives Criteria

C1 C2 C3 C4 C5 C6 C7 C8
D1 A1 MH ML H VH H VH ML L
A2 ML VL M M MH M VL M
A3 M L M L M H L MH
A4 ML MH VL MH L VL MH VL
D2 A1 H L VH H H H L L
A2 M M ML L VH H L ML
A3 MH L M H M MH ML MH
A4 L VH L H ML L H VL
D3 A1 H L H MH VH H ML L
A2 ML ML MH ML MH VH M H
A3 M L L L M M M MH
A4 L VH ML H MH VL L VL
Table 2. The relationship between IFSs and linguistic terms [12]
Linguistic terms Corresponding IFNs

Very low (VL) [0.15,0.80]
Low (L) [0.25,0.65]
Medium low (ML) [0.35,0.55]
Medium (M) [0.5.0.4]
Medium high(MH) [0.65,0.25]
High (H) [0.75,0.15]
Very high (VH) [0.85,0.10]
Step 2. Based on the weights for decision makers and Eq. (17), the aggregated
decision matrix is calculated. Besides, the PIA and NIA are determined,
respectively, using Eqs. (19)–(22).
Step 3. The projection of each alternative on PIA and NIA are calculated based
on Eqs. (23), (24). Then the ranking value for each alternative is cal-
culated as RV = (0.9583, 0.9445, 0.9344, 0.8973). Therefore, the ranking
order for alternatives is A1 > A2 > A3 > A4 .
As a result, site A1 is the best site for public fast electric vehicle charging
station.
5 Conclusion
In MAGDM problems, it is quite common that decision makers’ weight infor-

mation is partly or completely unknown because of such problems as decision
makers lack of knowledge, the complexity of the decision-making environment,
and scientific development limitations. Therefore, it is important to research a
solution to MAGDM problems when no weight information is available. In this
paper, we applied the intuitionistic fuzzy projection-based approach to deter-
mine the optimal site for public fast electric vehicle charging station to promote
the popularization of EVs in Chengdu. In our model, the extended TOPSIS
method under an intuitionistic fuzzy environment was developed to determine
the attributes weights. With the weight determination method proposed in this
paper, we effectively reduced the subjective effects to a minimum, so, because
less information was required to make a decision using the proposed method,
we demonstrated how MCGDM problems may become easier based on our app-
roach. Finally, under an intuitionistic fuzzy environment, the proposed approach
was shown to be able to handle MCGDM problems effectively, thus guaranteeing
that there is less deviation in the decision making.

Foundation of China under the Grant No. 71671118.
References
1. Atanassov KT (1986) Intuitionistic fuzzy sets. Fuzzy Sets Syst 20(1):87–96
2. Dyer JS, Fishburn PC et al (1992) Multiple criteria decision making, multiattribute
utility theory: the next ten years. Manage Sci 38(5):645–654
3. Frade I, Ribeiro A et al (2011) Optimal location of charging stations for electric
vehicles in a neighborhood in Lisbon, Portugal. Transp Res Rec J Transp Res
Board 2252(2252):91–98
4. Guo S, Zhao H, Yan J (2015) Optimal site selection of electric vehicle charging
station by using fuzzy topsis based on sustainability perspective. Appl Energy
158:390–402
5. Hwang C, Yoon K (1981) Multiple attribute decision making. Springer, Heidelberg
6. Li S, Huang Y, Mason SJ (2016) A multi-period optimization model for the deploy-
ment of public electric vehicle charging stations on network. Transp Res Part C
Emerg Technol 65:128–143
7. Shareef H, Islam MM, Mohamed A (2016) A review of the stage-of-the-art charging
technologies, placement methodologies, and impacts of electric vehicles. Renew
Sustain Energy Rev 64:403–420
8. Sheppard CJR, Gopal AR et al (2016) Cost-effective electric vehicle charging
infrastructure siting for Delhi. Environ Res Lett 11(6):064,010
9. Shu MH, Cheng CH, Chang JR (2006) Using intuitionistic fuzzy sets for fault-tree
analysis on printed circuit board assembly. Microelectron Reliab 46(12):2139–2148
10. Wang JQ, Wang J et al (2014) An outranking approach for multi-criteria decision-
making with hesitant fuzzy linguistic term sets. Inf Sci 280:338–351
11. Wang JQ, Wang P et al (2015) Atanassov’s interval-valued intuitionistic linguistic
multicriteria group decision-making method based on the trapezium cloud model.
IEEE Trans Fuzzy Syst 23(3):542–554
12. Wang T, Liu J et al (2016) An integrating owactopsis framework in intuitionistic
fuzzy settings for multiple attribute decision making. Comput Ind Eng 98:185–194
13. Wu ZB, Zhong L (2016) Weight determination for magdm with linguistic informa-
tion based on it2 fuzzy sets. In: IEEE International Conference on Fuzzy Systems,
pp 880–887
14. Xu ZS (2007) Intuitionistic fuzzy aggregation operators. IEEE Trans Fuzzy Syst
15(6):1179–1187
15. Yue Z (2013) An intuitionistic fuzzy projection-based approach for partner selec-
tion. Appl Math Model 37(23):9538–9551
16. Yue ZL (2011) An extended topsis for determining weights of decision makers with
interval numbers. Knowl Based Syst 24(1):146–153
17. Zadeh LA (1965) Fuzzy sets. Inf Control 8(3):338–353
18. Zadeh LA (1975) The concept of a linguistic variable and its application to approx-
imate reasoning. Inf Sci 8(4):301–357
If the Medical Expenses Do Effect the Rural
Residents’ Consume
Songtao Jiang1(B) and Yuan Ji2

1
School of Economics, Sichuan University,
jst714@qq.com
2
Sichuan Technology and Business University,
Abstract. We exploit a fixed effect model to examine if the uncertain

medical expenses of rural residents reduce their total consume using the
provincial-level panel data from 2000–2013. To obtain a robust results,
we then add the model with the time fixed effect. The results show that,
rural residents’ medical expenses share raise will reduce the share of rural
residents’ total consume in GDP by more than 6%, and this effect is more
obvious than the rural residents’ education expenses which is less than
6%. We also find that if the government increase the funds proportion
of health care, the medical demand of rural residents will raise which
may reduce the percentage of rural residents’ total consume in GDP
by more than 6%. By the way, liquidity constrains may also enhanced
the rural residents’ precautionary saving motives based on the medical
expenses. So we can conclude that the medical expenses of rural residents
do effect the consume demand of rural households significantly. And the
government need to control the rural residents’ medical expenses if they
attempt to expand the consumption of rural residents.
Keywords: Medical expenses · Rural residents consume · Panel data ·

Saving motives
1 Introduction
As the uncertainty of global economic recovery increases, the major economic
growth engines of investment and export are facing the constraints of environ-
mental protection and trade barriers. In order to keep economic growth at a
reasonable rate China must pay more attention to the role of consumption in
stimulating the economy.
Figure 1 illustrates the change in consumption and investment rates between
1978 and 2013 in the gross domestic product of the expenditure approach. Where
S represents the consumption rate, and I represents the investment rate. It is
clearly that from 1978 to 2013 the rate of consumption in our country shows
a general downward trend, especially after 2000, it shows a steep downward
1698 S. Jiang and Y. Ji
70
60
50
40
30
1980 1990 2000 2010 2020

year
S I
Fig. 1. Trends of consumption and investment rates in 1978–2013
trend, and it does not start to stabilize until 2010, and rises slightly. In sharp
contrast to this is that since 1978, China’s economic growth into the fast-rising
channel, which indicated that China’s consumption rate is not with China’s rapid
economic growth and increase.
70
60
50
40
30
1990 1995 2000 2005 2010 2015

year
China India
Brazil Russian
USA Japan
Germany
Fig. 2. International comparison of final consumption proportion of residents in GDP
From the international comparison point of view, we can also get the same
conclusion. Figure 2 shows the trend of final consumption among main countries
in GDP. Can be seen from Fig. 2, China’s consumption rate is not only lower
than developed countries such as European countries and the United States,
with India, Brazil and other emerging countries, which were also significantly
lower. Developed countries from Europe and the United States, its consumption
rate is generally high, and long-term stable operation. From the Fig. 2, The US
consumption rate has maintained a high level of smooth operation (basically
remained above 64), and in a steady rise with a slight trend; Germany and
Japan, similar to the situation, although the final consumption rate ate not as
high as the United States, but can also maintained at a high level (basically
If the Medical Expenses Do Effect the Rural Residents’ Consume 1699
remained around 55) and with a more stable trend. While the emerging market
countries, the final consumption rate of residents compared with the developed
countries is characterized by the fluctuations are generally large, including China.
Whether China, India, Brazil or Russia, the consumption rate of its residents in
the 1990–2013 years have experienced a more vibration. It is obvious that the
final consumption rate of Chinese residents is still very low compared with that
of emerging market countries.
When looking at the household consumption which play a decisive role of con-
sumption, we can find that the rural residents’ consumption as a share of house-
hold consumption has not increased with the urban residents’, and it makes rural
consumption become a short board of China’s strategy of expanding domestic
demand.
80
60
40
20
1980 1990 2000 2010 2020

year
urbcomsu ruralconsu
Fig. 3. The share of urban and rural household consumption in 1978–2013
Figure 3 shows the proportion of urban residents’ consumption and rural resi-
dents’ consumption in China’s household consumption during 1978–1930, where
urbcomsu’ represents urban residents consumption in the proportion of con-
sumption, ‘ruralconsu’ represents rural residents consumption in the proportion
of consumption. As can be seen from the figure, before 1992, the proportion of
consumption of rural residents in China’s household consumption in a dominant
position, but the early 90s of last century, the proportion of rural residents began
to decline sharply and the urban residents began to dominate the proportion of
consumer demand. The consumption gap between urban and rural residents has
been widening year by year, until 2010, the gap widening trend has slowed down.
Generally, there are many factor which constrain China’ rural consumption,
but the imperfect medical security system and the uncertainty of medical expen-
diture of rural resident cannot be ignored. As the medical consumption has the
characteristics of uncertainty and urgency, the impact of medical consumption
on rural residents’ overall consumption may be significant. This paper using
the provincial-level panel data from 2000–2013 to test if the uncertain medical
expenses of rural residents reduce their total consume by adopting fixed effect
model. Through the empirical research, we expecting find out the relationship
between medical expenses of rural residents and rural residents’ total consump-
tion. And based on the conclusion, we may provide some reference for the policy
of rural residents’ consumption.
2 Literature Review
Scholars have studied the reasons for the downturn in rural consumption and
how to start rural consumption from different aspects: Wang’s research [14]
indicated that the uncertainty expectations of rural residents are one important
reason that restricts the consumption in rural areas in China. Similar views
also include Tan [11]. Qu [13] found that the pessimistic expectations on future
expenditure of rural residents will place greater emphasis on deposit but not
expenditure. Therefore to start rural consumption China’s policy makers should
improve the rural consumption environment and reduce the burden on farmers.
Sun and Zhu’ research [9] also emphasized the importance of optimizing the
rural consumption environment to improve the consumption of rural residents.
Cai [1] had point that in addition to income levels, rural social security system
construction is also an important factor affection rural consumption. Clearly, as
a important part of social security system, medical security system are closely
related to medical expenditure, which indicate that the medical expenditure of
rural residents may play an important role in consumer decision making.
There are also scholars’ studies focused on the consumer environment and
the expectations of rural residents. Liu [7] found that insufficient financial sup-
port limited the development of rural residents’ consumption. Yang and Chen’s
research [16] based on the view of rural public goods supply, proposed that
increasing rural public goods supply can reduce the uncertainty of expenditure
expectations of rural residents which may benefit the rural consumption. Zhou
and Yang analyzed the consumption both urban and rural areas and pointed
out that, precautionary saving motive may also effect rural residents as well as
urban residents.
Furthermore, some scholars studied about rural medical expenditure and the
existing literature mainly focuses on two aspects: first, the impact factors of
rural residents’ medical expenditure such as Tan and Zhang et al. [10] whose
study explores the influencing factors of rural households’ medical consumption
spending and its demand elasticity based on QUAIDS model and the cross-
section; And Huang and Liu [6] provided a similar study, which state that the
income level, government health investment and other factors do affect the med-
ical consumption of rural residents; Zhao [17] used the 1993–2012 panel data to
analysis the regional differences of rural residents’ health care consumption, and
state that the income, education, economic development and other factors on the
impact of rural resident’s medical spending are significant regional differences.
Second, the function of rural medical security system such as Bai and Li et al.
[3] found that the insurance coverage increases non-health care consumption by
more than 5% and the insurance effect exists even for households with medical
spending, by exploiting a quasi-natural experiment; Ding and Ying et al. [2]
also stated that the propensity to consume of rural households is universally low
in China, and the medical insurance do increased rural durables consumption
significantly, base on the analysis of China Health and Nutrition Survey panel
data from 1991 to 2009. Tan and Qin [12] used IV-TOBIT model to test the
crowding out effect of China’s households’ medical expenses and conclude that
the medical expenses do reduce the households’ expenses on food and clothing,
whose research also used the China Health and Nutrition Survey data (2013).
Although the above study about rural medical expenditure focused on difference
aspects, they provided us with the method of reference.
In general, although many studies focused on the rural residents’ consump-
tion expansion directly or indirectly mentioned factors such as rural residents’
expectation, consumption environment, and precautionary saving motive. But
there are few research put rural residents’ medical expenses as an independent
factor which effect the rural residents’ consumption behavior. For this reason,
we develop an empirical model that includes both residents’ medical expenses
and a set of factors that may affect the rural residents’ medical expenses. This
model also contain standard economic variables. Our study adds to the current
literature by utilizing pooled time series and provincial cross-sectional data over
a longer period than the household survey data cover. In addition, by utilizing
medical expenses variable, we are looking for if the medical expenses of rural
residents do effect the rural consume.
3 Investigation of if Medical Expenses Affect Rural

Residents’ Aggregate Consumption
The equation to be estimated for (i) provinces and (t) time periods is as follows:

C
= β0 + β1 medratioitit + β2 goeratioit + β3 loandepit + γ1 P1
Y it

+ · · · + γi−1 Pi−1 + λ1 D1 + · · · + λt−1 Dt−1 + σi Zit + μit ,
C
where
Y it
: rural residents’ consumption as ashare of provincial domestic
product;
medratioit : the share of rural household medical expense in rural household
consumption per capita;
goeratioit : local government expenditure on health as a share of total local
government expenditure;
loandepit : the share of loans plus deposits in provincial domestic product;
Pi , Dt : provincial and time dummies;
Zit : vectors of economic system variables;
growthit : the annual growth in provincial domestic product;
dincoit : tprovincial net income of rural residents per capita;
depratioit : dependency ratios;
eduratioit : the share of rural household education expense in rural household
consumption per capita.
This empirical model used here combines standard consumption determi-

nants relating to income, government expenditure and family structure which
expected to influence households’ consumption decisions. While this model also
concentrated to rural residents’ medical expenses and the medical expenses
expectation.
We use the share of rural residents’ consumption in domestic prod-
uct YC it as our dependent variable. This is the share that appears to behave
differently compared to urban. Our panel data set includes provincial data for
31 provinces between 2000 and 2013. The variation across provinces allows us
to explore possible factors in addition to the standard consumption variables.
The impact of medical expenses on household’s consumption is manifest in
two aspects: First, medical spending will be squeezed the share of other con-
sumption under the total budget constraints. Second, medical expenditure will
increase people’s cautious motivations, which will increase precautionary sav-
ings. In addition, as the medical treatment has high uncertainty, residents could
not predict the medical expenses in the time and amount, which may strengthen
residents’ saving motive. To test if this impact do affected China’s rural consump-
tion we applied the share of rural household medical expense in rural household
consumption per capita as a key variable (medratioit ). A reasonable hypothesis is
that the increase in the share of rural household medical expenses in rural house-
hold consumption per capita will lead rural residents to reduce consumption.
A second factor that might influence rural residents’ consumption relating
to medical expenses is government expenditure on health (goeratioit ). If the
government is taking care of residents’ health care needs, households would not
need to save for the possible expenditures or may save less. So we can hypothesis
that the government health expenditure variable is positively correlated with the
dependent variable.
The third factor we concerned relating to the household medical expenses
and the rural residents’ consumption is the ability of rural residents to use the
financial system to smooth consumption over time. So we adopt financial devel-
opment indicators as the third variable which can lessened liquidity constraints
of rural residents. The standard financial development indicators are the per-
centage of total financial assets to GDP [4] or money supply measures such as
M2 as a share of GDP [8]. But we do not have access to financial assets data for
all provinces for the period of time we study, and the money supply measures
do not vary by province but only vary by country. Therefore we follow [18] and
use the total deposits and loans issued by all financial institutions as a percent
of GDP reported by each province (loandepit ) as a proxy variable for liquidity
constraint.
For the control variables in the Z vector, we include a number of factors raised
by the literature: (growthit ) is the annual growth in provincial domestic product
to interpret the economic development of the provinces; (dincoit ) is provincial
net income of rural residents per capita; (depratioit ) is the provincial depen-
dency ratios, in principle, higher dependency ratio will lead to higher consumer
spending. In the empirical model we also adopt the share of rural household edu-
cation expense in rural household consumption per capita as a control variable
(eduratioit ) because it has been pointed out by the literature that education
expenses is an important factor that affected residents saving motions. By the
way, adopting the education expenses variable can make a comparison with the
medical expenses variable.
4 Estimation and Results
Table 1 reports our descriptive statistics as well as the between-province standard

deviation and within-province standard deviation (over the entire time period) of
our main variables in our model. This table confirms the considerable variations
of the dependent and independent variables of interest in our data set.
We begin our estimation with a pooled OLS model. The result reported in
Table 2 model 1. This model is estimated with pooled ordinary least squares using
Variable Mean Std. Dev Min Max Observations

C
Y it
Overall 0.1237 0.0665 0.0185 0.3669 N = 434
Between 0.0485 0.0285 0.2252 n = 31
Within 0.0463 0.034 0.2838 T = 14
medratio Overall 0.07 0.0219 0.0144 0.1399 N = 434
Between 0.0168 0.0252 0.1008 n = 31
Within 0.0144 0.0353 0.1376 T = 14
loandep Overall 2.5031 0.8581 1.28 6.66 N = 434
Between 0.8176 1.5621 5.9279 n = 31
Within 0.2966 1.3603 4.1188 T = 14
eduratio Overall 0.095 0.0297 0.0099 0.1841 N = 434
Between 0.0226 0.0216 0.1314 n = 31
Within 0.0197 0.0473 0.1541 T = 14
depratio Overall 38.0565 7.3399 19.27 57.58 N = 434
Between 6.3396 25.2482 51.7112 n = 31
Within 3.8588 28.8267 52.9967 T = 14
goeratio Overall 0.0215 0.0091 0.0019 0.0488 N = 434
Between 0.0078 0.0032 0.0403 n = 31
Within 0.0049 0.0071 0.0383 T = 14
dinco Overall 4073.121 2348.768 1332.142 14691.7 N = 434
Between 1790.959 2137.623 9200.569 n = 31
Within 1550.96 131.1782 9564.254 T = 14
The most data used in this paper are from China Statistical Yearbook. The
provincial net income of rural residents per capita data were adjusted based
on 1999 CPI. The government expenditure on health and education data are
budgetary data from both China Statistical Yearbook and China Fiscal Yearbook.
Table 2. Estimation results
Model 1 Model 2 Model 3 Model 4 Model 15

edratio −0.635∗∗∗ −0.643∗∗∗ −0.643∗∗∗ −0.390∗∗∗ −0.223∗ (0.117)
(0.094) (0.124) (0.169) (0.098)
goeratio 1.171∗∗∗ −0.674∗∗ −0.674∗ −0.786∗∗∗ -
(0.261) (0.290) (0.385) (0.187)
loandep −0.007∗∗∗ −0.017∗∗∗ −0.017∗ −0.026∗∗∗ -
(0.002) (0.005) (0.010) (0.008)
eduratio 0.479∗∗∗ 0.586∗∗∗ 0.586∗∗∗ 0.389∗∗∗ 0.321∗∗∗ (0.058)
(0.067) (0.103) (0.136) (0.068)
beds - - - - −0.000∗∗∗ (0.000)
- - - -
save - - - - −0.071∗∗∗ (0.014)
- - - -
yr00 04 - - - 0.055∗∗∗ 0.044∗∗∗ (0.009)
- - (0.011)
yr05 09 - - - 0.004 −0.006 (0.004)
- - - (0.006)
growth −0.008∗∗∗ −0.009∗∗∗ −0.009∗∗∗ −0.005∗∗∗ −0.005∗∗∗ (0.001)
(0.001) (0.001) (0.001) (0.001)
depratio 0.003∗∗∗ 0.002∗∗∗ 0.002∗∗∗ 0.001* 0.001∗∗∗ (0.000)
(0.000) (0.000) (0.000) (0.000)
dinco −0.000∗∗∗ −0.000∗∗∗ −0.000∗∗∗ 0.000* −0.000∗∗∗ (0.000)
(0.000) (0.000) (0.000) (0.000)
cons 0.905∗∗∗ 1.152∗∗∗ 1.152∗∗∗ 0.745∗∗∗ 0.698∗∗∗ (0.147)
(0.090) (0.083) (0.181) (0.158)
F test all u i = 0 P rob > F = 0.00
N 434 434 434 434 406
adj. R2 0.741 0.689 - - -
R2 0.745 0.716 - - -
Within-R2 - 0.716 0.716 0.808 0.823
Standard errors in parentheses
∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01
the full data set from 2000–2013. The share of rural household medical expense
variable (medratioit ) is significant and negative. Which met our hypothesis that
the increase in the share of rural household medical expenses in rural household
consumption will lead rural residents to reduce consumption. While the gov-
ernment expenditure on health variable (rugeoratioit ) is significant and positive
which met our hypothesis too that the government health expenditure variable
is positively correlated with the dependent variable.
There after we add cross-provincial fixed effect to controls for omitted vari-
ables that vary across provinces but do not change over time. Furthermore,
vast differences in economic, industrial and fiscal structures across provinces
are likely to influence household consumption. So the intercepts generated for
each province in this model should absorb the influences of these omitted
Table 3. Robustness checks
Tests Methods Hypothesis Results

Heteroskedasticity Modified Wald test H0: sigma(i)2 = sigma2 Prob > chi2 =0.0000
for all i
Cross-section Frees’ test H0: cross-sectional 5.828 > alpha = 0.01 : 0.36
correlation independence
Sequence Wooldridge test for H0: no first order P rob > F = 0.0000
correlation autocorrelation in autocorrelation
panel data
variables that differ for each province but are constant over time. Model 2 is
also estimated with ordinary least squares, but with provincial fixed effects
reported in Table 2. The F-statistics of the test on the significance of provin-
cial fixed effects show that there are significant variations among provinces
(F (30, 396) = 11.18, Pr ob > F = 0.0). In model 2 the government expendi-
ture on health variable (rugeoratioit ) is significant but negative, reflecting that
government expenditure on health increasing would not stimulate the rural res-
idents’ consumption. Although the share of rural household medical expense
variable (medratioit ) is also significant and negative.
Table 3 presents a series of robustness checks on model 2 including het-
eroskedasticity, cross-section correlation, and sequence correlation. As the test
results shown in Table 3. The result estimated in model 2 is not robustness. So
we re-estimates model 2 using Daniel Hoechle’s [5] method to get stronger results
in model 3. Compared with model 2, the estimate result in model 3 are fairly
similar, only the significant level of the coefficients has changed. But our main
independent variables are still significant.
Based on model 3, we then add time-fixed effects to control for variables that
are constant across provinces but evolve over time, such as the introduction of
new healthcare regime “Xinnong he” (New Cooperative Medical Scheme), the
nationwide economic reforms deepen in China rural area, etc. Model 4 replicates
model 3 except that these estimates include the time fixed effects as well as the
provincial fixed effects. We included dummy variables chosen for several time
periods to capture the trend of the introduction of new healthcare regime. As
the New Cooperative Medical Scheme start the pilot at 2003 and realized full
coverage at 2009, considering the lag of policy, our first time dummy starts from
2000 and ends in 2004. The next time dummy covers the period of 2005 to 2009.
The last dummy variable covers the time period between 2010 and 2013. In model
4, at least one newly generated time dummy variables is statistically significant
and produce coefficients with a trend: over time the coefficient increase each year.
The basic findings of model 3 remain robust in model 4, although some coefficient
changed in magnitude. So we can draw our conclusion based on model 4.
In model 4, our main dependency variable were all significant although we
added the time-fixed effects. The share of rural household medical expense is
negative, which is consistent with our assumption. The government expenditure
on health is negative, implying that government expenditure on health do not
release rural residents’ saving motive, which doesn’t met our hypothesis. The
financial development is negative in model 1–model 4, which indicated that the
financial development did not lessened rural households’ liquidity constraint. In
Table 2 we also found that the shares of rural household education expense in
rural household consumption, which we adopted to make a comparison with the
medical expenses as a control variable, were significant and positive in model
1–model 4. So we can conclude that the impacts of medical expenditure and
education expenditure on rural residents’ consumption are not consistent.
As the government expenditure on health and the financial development’s
estimated results did not consistent with our expectations and the general theory,
we re-estimated model 4 by alternative measures of these two variables as a
further robustness check, presented in model 5.
In model 5 we used the number of beds in township health institutions to
replace the government expenditure on health. Simultaneously, we adopted the
net saving rate of rural residents [15] per capita to replace the financial devel-
opments as the proxy variable of liquidity constraint of rural residents. Since
the statistical agency is no longer announced the number of beds in township
health institutions of Shanghai and Beijing after 2009, our estimation in model
5 did not include the two sections of Shanghai and Beijing. Also because we
adopted the panel data contain 31 sections, so there is no substantial impact on
the estimation from the elimination of the two sections.
In model 5, although there are some changes in the estimation results com-
pared with model 4, but the share of rural household medical expense remain
significant and negative, by the way the shares of rural household education
expense are also significant and positive, which are corresponded with model 4.
Furthermore, the number of beds in township health institutions is significant
and negative too, as a part of government expenditure on health, which indi-
cated that the result of government expenditure on health estimated in model 4
has quite robustness. As the net saving rate of rural residents is significant and
negative, we can conclude that the rural residents faced obvious liquidity con-
straints. In addition, as the model 5 adopt the net saving rate of rural residents,
the coefficient of the share of rural household medical expense becomes smaller,
indicated that medical expense always accompanied with liquidity constraint,
which is associated with the amount of medical expenditure relatively large.
We use various models and estimation techniques to examine if and how the
medical expenses affect the variation in rural residents’ consumption as a share
of GDP. As our expectation the medical expenses do lessened the rural resi-
dents’ consumption as a share of GDP. In our study, the most significant and
robust discovery is that reducing rural residents’ motivations for precautionary
savings through government’s expenditure at the local level have the opposite
effect in increasing rural residents’ consumption. While this result contrary to
our expectation and the general theory, the alternative measure that captures
the role of the government expenditure through the number of beds in town-
ship health institutions was also significant. From the direct meaning of this
estimation result, we can conclude that the increasing of government spending
on health has stimulated the rural household saving motive. But this is by no
means a justification for reducing the fiscal responsibility of the government. On
the contrary, the result implied, to some extent, that the government expendi-
ture on rural health is not enough, and there is a heavy responsibility on the
government to improve the rural medical and health environment. Because at
the pooled medical and health environment level, the rural household’s medical
consumption demand has been repressed.
The financial development and its alternative measurer always robust regres-
sors in all the models, while the measurements are small and negative, which
means the liquidity constraints is a factor that do not bode well for the rural
consumption. The shares of rural household education expense are always signif-
icant and positive, while the coefficient absolute value is large. As demonstrated
in this study, this means that, the impacts of medical expenditure and educa-
tion expenditure on rural residents’ consumption are not consistent, so we need
different options to develop the medical and education in rural areas.
For the control variables in the Z vector, although the annual growth in
provincial domestic product is significant influenced the independent variable but
the coefficient absolute value is too small, while the provincial net income of rural
residents and the provincial dependency ratios suffer from the same problem.
These results point to several suggestions that the increasing medical
expenses is the obstacle of boosting domestic consumption especially in rural
areas. So the policy makers need to formulate policies to optimize the consumer
environment in rural areas, just like increase investment in rural health care,
although the government expenditure on health is negative in this paper. The
cautious motive which lead rural household to increase their savings has been
proved indirectly by the estimated result in this paper as the share of rural house-
hold medical expense in rural household consumption is significant and negative.
In addition, as the liquidity constraint has also been found in the rural are as from
the estimated results. To enhance the rural consumption, the government needs
to not only lift the burden of rural households medical expenditure but lessen the
liquidity constraint also. In our opinion, to increasing the coverage and funding
levels of NRCMS is an effective method. Furthermore, increasing the accessibility
and convenience of rural health care facilities may also a effective way.
References
1. Cai YZ (2009) Economic stimulus package and start-up of rural consumption in
China: empirical study based on the decomposition of rural household income. J
Financ Econ 9:4–12
2. Ding J, Ying M, Du Z (2013) China’s rural household consumption behavior
research: based on the perspective of health risk and medical insurance. J Financ
Res 10:154–166
3. Ding JH, Ying ML, Du ZC (2013) Health insurance and consumption: evidence
from China’s new cooperative medical scheme. Consum Econ 10:154–166
4. Goldsmith RW (1969) Financial structure and development. Stud Comp Econ
70(4):31–45
5. Hoechle D (2007) Robust standard errors for panel regressions with cross-sectional
dependence. Stata J 7(3):281–312
6. Huang XP, Liu H (2011) An analysis on the influencing factors of medical con-
sumption of Chinese rural residents. Consum Econ 11:77–80
7. Liu GM (2011) Discussion and analysis about financial support to rural consumer
market’s exploitation. J Central Univ Financ Econ 6:35–40
8. Mckinnon RI (1973) Money and capital in economic development. Brookings insti-
tution
9. Sun HQ, Zhu C (2012) An empirical analysis of urbanization and rural consumption
growth in China. Stat Decis 5:90–93
10. Tan T, Zhang YY, Jun HE (2014) An analysis of influencing factors and elasticity
of rural households’ medical consumption spending in China. J Shanghai Univ
Financ Econ 3:63–69
11. Tan YH (2009) On fiscal policies of stimulating consumer demand in rural areas.
J Central Univ Financ Econ 8:5–9
12. Tang Q, Qin XZ (2016) An empirical study on the crowding-out effect of family
medical expenditure in China. Econ Sci 3:61–75
13. Tao Q (2009) Analysis on the consumption behavior and its constraint factors in
rural China. Economist 9:54–60
14. Wang B (2012) The deep reason and improving ways of rural consumption demand
in China. Consum Econ 1:29–32
15. Wang P (2014) Rural institutional change, liquidity constraint and consumption
growth of rural residents in China. J Shanxi Univ Financ Econ 10:1–10
16. Yang L, Chen C (2013) An analysis of the effect of education and health pub-
lic goods on rural residents’ consumption: from the perspective of human capital
promotion. J Agrotechnical Econ 9:4–12
17. Zhao ZR (2014) A study on reginal differences of influencing factors of rural resi-
dents’ health consumption in China. Consum Econ 3:24–29
18. Zhou L, Wang Z (2008) An empirical analysis of financial development and eco-
nomic growth in China. J Financ 10:1–13
A Discrete Time-Cost-Environment Trade-Off
Problem with Multiple Projects: The Jinping-I
Hydroelectric Station Project
Huan Zheng1,2(B)
1
Management School, Chongqing Technology and Business University,
290871340@qq.com
2
Engineering School, Widener University, Philadelphia 19013, USA
Abstract. Stronger environmental protection regulations and innov-

ative contracting methods focused on reducing environmental impact
through new incentives have placed increasing pressure on decision mak-
ers, especially those on large-scale construction projects. This increase
has led to a need to identify optimal decision and select the most efficient
construction modes for each activity to minimize not only the project’s
time and penalty costs but also its environmental impact. To address
this need, this paper applies a mathematical model based on a hybrid
genetic algorithm with fuzzy logic controller (flc-hGA) to the time-cost-
environmental trade-off problem with multiple projects under a fuzzy
environment. Applying the proposed method to the case study of the
Jinping-I Hydropower Station, a large-scale project in Southwest China,
clearly demonstrates its economic, technological, and social ecological
effectiveness.
Keywords: Time-cost-environment · Trade-off · Multiple project ·

Fuzzyt · Hybrid genetic algorithm
1 Introduction
Two of the most crucial aspects of any such construction project are time and
project cost, both of which have received considerable research attention [9,12].
One particularly important element of effective project scheduling theory and
applications is the discrete time-cost trade-off problem (DTCTP) introduced
by Harvey and Patterson [6]. Of more recent concern, however, are accusa-
tions that the construction industry is causing environmental problems that
range from excessive consumption of global resources-both in terms of construc-
tion and operation-to pollution of the surrounding environment. Hydroelectric
projects, particularly, contribute significantly to changes in river environments
1710 H. Zheng
[2,17]. Yet previous studies have paid more attention to evaluating the envi-
ronmental impact at later stages of reservoirs and hydraulic electricity rather
than considering it from the construction stage. Recently, however, a tighten-
ing of environmental protection regulations, especially for hydroelectric projects,
has increased pressure on project managers to reduce the environmental impact
through the selection of optimal construction modes. There is an imminent need,
therefore, to study management decisions on selection methods to ensure more
environmentally friendly means of construction during the project planning stage
when the environmental impact can best be incorporated into other project
objectives. For hydroelectric projects such as the JHS-I, this planning stage
should include the optimization of construction and subsequent operations for
environmental impact as well as time and cost.
The expanding scale of such construction projects worldwide, however, makes
effective project management extremely complex. To deal with this complexity
while still achieving management objectives, construction managers must employ
a project management decision system that effectively controls total project
duration, penalty costs, and environmental impact. One effective method for such
control is the discrete time-cost-environment trade-off problem with multiple
projects (DTCETP-mP), which is an extension of the original DTCTP. This
analysis thus applies DTCETP-mP to solve the three main objectives of the
JHS-I problem: (1) minimization of the total project duration, (2) minimization
of the total project penalty cost, and (3) minimization of the environmental
impact.
In non-routine projects such as the new construction at JHS-I [10], the dura-
tion of each activity and completion time may be uncertain, so the project man-
ager must handle multiple conflicting goals in an uncertain environment in which
information may be incomplete or unavailable. In this context, activity duration
uncertainty can be modeled using either probability-based methods [12] or fuzzy
set-based methods [18,19] depending on the situation and the project manager’s
preference. When a project manager has difficulty characterizing the random
variables, as in the current scenario of a new construction project with unique
activities and a lack of historical data, the fuzzy method is the most effective
approach.
Because the DTCETP-mP is an extension of the DTCTP, it is an NP-hard
problem that is difficult to solve [15]. Yet even the most exact of currently avail-
able methods can only solve small projects with under 60 activities, a far cry
from the numerous activities and modes per activity involved in the large-scale,
complex JHS-I project, whose optimal solutions are beyond the capabilities of
traditional production scheduling methods like PERT (Program Evaluation and
Review Technique) and CPM (Critical Path Method) [3]. More suitable heuris-
tic solution procedures for solving the DTCETP-mP are thus needed, several
of which have been suggested in the literature. For example, Franck et al. [4]
demonstrated that a genetic algorithm (GA) performed slightly better than a
tabu search (TS) procedure but required more computing. In earlier work, Wang
et al. [16] addressed larger, more complex problems by introducing an improved
hybrid genetic algorithm (hGA) that uses a fuzzy logic controller (flc) to adap-
A Discrete Time-Cost-Environment Trade-Off Problem 1711
tively regulate the GA parameters (including generation number, population

size, crossover ratio, mutation ratio) and automate mutation ratio fine tuning.
The effectiveness and efficiency of the proposed flc-hGA is evaluated through
comparison with other methods.
This present application of the DTCETP-mP to identify optimal solutions
for the JHS-I large-scale deeply buried tunnel group project amends the GA first
introduced by Holland [7] to an flc-hGA. The analytic results indicate that this
optimization method is both practical and highly efficient in solving the JHS-
II problem and achieving minimization of the project’s total duration, penalty
costs, and environmental impact.
2 Problem Description and Mathematical Model

2.1 Problem Description
Because the JHS-I project is unique, lacking in historical data, technically diffi-
cult to construct, structurally complex, and subject to strict requirements and
many construction uncertainties, it is difficult to estimate the exact duration of
each activity or assign each a strict function [8,11]. Duration is thus characterized
by fuzziness and subjective uncertainty [18,19]. In such a situation, not only the
duration but also the cost and environmental impact of each activity depends
on construction mode selection, which is in turn related to the assignment of
construction materials and machines, crew formation, and overtime policies. In
this case, therefore, a DTCETP-mP discrete mode selection for these variables
is more practical than a continuous mode selection.
In the DTCETP-mP analysis, each subproject activity is assigned certain dis-
crete well-defined construction methods so that the analysis can determine the
best combination(s) of duration, cost, and environmental impact for all activi-
ties. The start time for each subproject activity is optimized by the flc-hGA, and
all activities are executed in a certain order to achieve completion under bud-
getary, environmental, and cash flow constraints in any time period. The project
completion date is as estimated by project experts and lies between the earliest
start time for the first activity and the latest finish time for the last activity.
The total project penalty costs, also assessed by these experts, comprise penalty
costs for delays in all subprojects and for the environmental impact. Project
planning involves the selection of proper methods, crew sizes, equipment, and
technologies to ensure efficient activity completion.
Despite such careful planning, however, project completion inherently
involves a trade-off between time, cost, and environmental impact. For example,
using more productive resources or technologies (e.g., more efficient equipment,
more workers, increased overtime) may save time but raise costs. Conversely, a
reduction in time may lower environmental impact, while a reduction in cost
may increase both time and environmental impact. These three important con-
siderations of time, cost, and environmental impact are thus the criteria used to
determine the best scheduling and combination for the DTCETP-mP, thereby
pinpointing the optimal construction mode combination for all activities.
1712 H. Zheng
2.2 Assumptions
The DTCETP-mP model for the JHS-I makes the following assumptions:
(1) The DTCETP-mP comprises multiple projects, each containing several
activities;
(2) The start time of each activity is dependent upon the completion of its
predecessor;
(3) The capital used by all activities does not exceed the limited quantities in
any time period, and the total project budget is within a predetermined
limit;
(4) The environmental impact caused by the activities does not exceed the
limited quantities in any time period, and the total project environmen-
tal impact is within a predetermined limit;
(5) When an activity begins, it cannot be interrupted;
(6) The managerial objective is to minimize the total project time, total tardi-
ness penalty, and total environmental impact for all subprojects.

The problem is represented on an activity-on-node (AON) network with a single
starting and a single ending node, each of which corresponds to dummy activities.
The analysis uses the following notation:
Index:
i : index of subproject in a project, where i = 0, 1, 2, · · · , I, S and T are dummy
projects;
j : activity index in each subproject, j = 0, 1, 2, · · · , J , and are dummy activity.
Parameters:
d˜ij : processing time for activity j in subproject i;
t̃D
i : specified project completion time for subproject i;
lijc : cost of activity j in subproject i per unit of time;
lije : environmental impact of activity j in subproject i per unit of time;
cpi : total penalty cost of subproject i per unit of time;
bc : maximum cost allowable per unit of time for the subproject;
be : maximum environmental impact allowable per unit of time for the
subproject;
Bc : maximum cost allowable per unit of time for the project;
Be : maximum environmental impact allowable per unit of time for the
project;
t̃F
ij : finish time of activity j in subproject i;
Vij : environmental impact of activity j in subproject i;
wij : weight of the environmental impact of activity j in subproject i;
Sp : set of activities
in progress in period p
Sp = j tSij ≤ p ≤ tSij + tij , i = 1, 2, · · · I + 1; j = 1, 2, · · · J + 1 ;
Pr e(i) : set of the immediate predecessors of subproject i;
Pr e(j) : set of the immediate predecessors of activity j
Decision variables:
t̃Sij : start time for activity j in subproject i
Functions:
z1 : total duration of the project;
z2 : total penalty costs of the project;
z3 : total environmental impact of the project
2.4 Multiobjective Model

The solution proposed by the multiobjective optimization model is based on
managerial objectives and project constraints, which for clarity are discussed
separately in the subsections below.
(1) Objective Functions
The first objective is to minimize total project time; that is, the sum of the
completion times for all subprojects, which can be expressed as follows:

I
min z1 = E(t̃F
iJ ). (1)
i=1
The second objective is to measure and minimize the total cost by minimizing
the total penalty costs of multiple projects:

I
min z2 = cpi (E(t̃F
iJ ) − E(t̃i )).
D
(2)
i=1
The third and final objective is to minimize total environmental impact:

I J
j=1 Vij × wij (E(t̃iJ ) − E(t̃i ))
F D
min z3 = . (3)
i=1
E(t̃F
iJ )
(2) Constraints
Because a specific subproject must be completed before another subproject
can be initiated (the precedence constraint of multiple projects), the model
includes the following constraint:
E(t̃SeJ ) ≤ E(t̃Si,1 ) − E(d˜eJ ), e ∈ Pr e(i). (4)
Likewise, because the start time of each activity is dependent upon the com-
pletion of some other activities (the precedence constraint of activities), the next
activity must be started after a specific activity is completed:
E(t̃Sil ) ≤ E(t̃Sij ) − E(d˜il ), l ∈ Pr e(j). (5)
The project is also subject to a limitation on the total capital and capital
per time period,

lijc ≤ bc i = 1, 2, · · · , I, (6)
j∈Sp
1714 H. Zheng
as well as on the total environmental impact and the environmental impact per
time period:

lijc ≤ Bc , (7)
i∈Sp j∈Sp

lije × wij ≤ be , i = 1, 2, · · · , I, (8)
j∈Sp
Vij
lije × wij ≤ Be , lije = . (9)
dij
i∈Sp j∈Sp
The nonnegative variables are described in the model by the following equa-
tion:
˜
ij ), E(dij ) ≥ 0, ∀i ∈ I, ∀j ∈ J.
E(t̃Sij ), E(t̃F (10)
3 Hybrid Genetic Algorithm with Fuzzy Controller
Although mathematically, several Pareto optimal solutions are possible for the
multiobjective model formulated above, in real-world construction, only one opti-
mized solution is needed in each time-constrained decision-making situation.
Hence, the multiobjective model is transformed into a single-objective model
using a weighting method. Additionally, because accurately determining the GA
parameters is especially important in solving large-scale problems like the JHS-I
project, GA effectiveness is improved by adaptively regulating the crossover and
mutation rate during the genetic search process using the fuzzy logic controller
(flc) [16]. This regulation reduces CPU time and enhances optimization quality
and stability by regulating the increasing and decreasing crossover and mutation
rate ranges [1,5,13,14,20].
3.1 Overall Procedure for the Proposed Method
Solving the problem with the flc-hGA involves the following steps:
Step 1. Concentration of multiple objectives using the weight-sum procedure.

Step 2. Setting of the genetic algorithm parameters: population size, crossover
rate, mutation rate, and maximum generation.
Step 3. Generation of an initial set of individuals.
Step 4. Choosing of the selection and hybrid genetic operators: crossover and
mutation.
Step 5. Evaluation of the fitness value of the chromosome.
Step 6. Selection of the best total penalty for the minimized total project
time and minimized environmental impact and storage of an alterna-
tive schedule for the minimized total project time.
Step 7. Check of the termination: If one individual has achieved the predefined
fitness value, the process stops: otherwise, it goes on to step 8.
Step 8. Regulation of the mutation rate through adaptive use of the fuzzy logic
controller; otherwise, the process returns to step 4.
The model uses two hybrid genetic operators, a position-based crossover and
a swap mutation (SM) operator. The crossover operator randomly takes some
genes from one parent and fills any vacuum with genes from the other parent
by scanning from left to right, while the SM operator selects two projects at
random and swaps their contents.
3.2 Hybrid Genetic Operators

The model uses two hybrid genetic operators, a position-based crossover and
a swap mutation (SM) operator. The crossover operator randomly takes some
genes from one parent and fills any vacuum with genes from the other parent
by scanning from left to right, while the SM operator selects two projects at
random and swaps their contents.
3.3 Fuzzy Logic Controller

The fuzzy logic controller (flc) is used to automatically tune the GA parameters
whose determination is so important in large-scale problems. Here, only a muta-
tion flc is used because in the proposed hGA, the effects of the crossover and
crossover flc are almost the same. The main difference between the hGA and the
flc-hGA is that a mutation flc based on the flc is implemented independently
to adaptively regulate the mutation ratio during the genetic search process by
considering the changes in average fitness in each parent and offspring popu-
lation over two continuous generations. Mathematically, this procedure can be
expressed as follows: letting f (t) be the difference in the average fitness function
between the tth and (t − 1)th generation, ε is a small positive number near to
zero (in this paper, ε = 0.1), allowing the mutation ratio for the next generation
to be derived using an if-then procedure:
(1) If |f (t) − f (t − 1)| < ε, then the mutation ratio pm for the next generation
should be rapidly increased;
(2) If f (t) − f (t − 1) < ε, then the mutation ratio pm for the next generation
should be decreased;
(3) If f (t) − f (t − 1) > ε, then the mutation ratio pm is selected for the next
generation.
When f (t) and f (t − 1) are the flc’s inputs and the change in mutation
ratio m(t) is its outputs. m(t) Once the input values are assigned, the scaling
value Z(a, b) can be determined by setting λ ∈ [−1.0, 1.0] as the given values
for regulating an increasing and decreasing range for the mutation ratio. The
changes in the mutation ratios are then determined by Δm(t) = λZ(a, b) and
the mutation ratio values for the next generation by pm (t + 1) = pm (t) + Δm(t),
where pm (t) is the mutation ratio at generation t.
1716 H. Zheng
4 Case Study: The Jinping-I Hydropower Station

The Jinping-I (JHS-I) and Jinping-II Hydropower Stations, which epitomize the
large-scale constructions of China’s West-East Electric Transmission Project,
have a combined capacity of 8,400 MW and were planned for Jingping River
Bend, a major artery 150 Km in length whose downstream river section is sep-
arated from the opposite bank by only 16 km. Along that length, the elevation
drops 310 m, creating an excellent site for hydroelectric production. Whereas
JHS-I will rely on its high dam and reservoir to supply water, JHS-II, located
7.5 Km downstream (at coordinates 28◦ 07 42 N 101◦ 14 27 E) is diverted by a
much smaller dam and will rely on the world’s four largest and longest (16.7 km)
diversion tunnels. This latter has a total installed capacity of 4,800 MW (8 × 600
MW), for a multiyear average annual generation of 24.23 TWh.
Because such scale and complexity presents computational challenges, our
model includes only the elements typical of constructions, making it generaliz-
able to similar projects. As previously emphasized, the study goal is to iden-
tify the most effective strategies and thus the most effective and viable project
management methods for JHS-I, so that these may be implemented in future
large-scale projects (Table 1).
Table 1. Subprojects and activities
A1 Spillway project A11 Earth-rock excavation A12 Concreting A13 Gates hoist
equipment installation A14 Clearing up and finishing work
A2 River diversion A21 Import and export hole dug A22 Concrete lining and
during construction check-gate installation A23 Lockup A24 Gen set installation
A25 Substation construction and equipment installation
A3 Dam construc- A31 Concrete cut-off wall A32 Dam foundation lock cut A33
tion Dam filled A34 Asphalt concrete watertight diaphragm
A4 Power capacity of A41 Diversion opening A42 Air pressure system A43 Second-
stream stage cofferdam A44 Bore-hole A45 Concrete lining
A5 Transport and A51 Road clear A52 Warehouse and factory construction A53
power system Water supply A54 Power transmission project
4.1 Data Collection
The data for the JHS-I project, obtained primarily from the Ertan Hydropower
Development Company, includes observations of managerial practice and inter-
views with designers, consultants, contractors, subcontractors, and a city govern-
ment officer at the station. This data set is supplemented by information from prior
research. The construction manager’s project experience, in particular, was invalu-
able for researcher comprehension of the projects’ specific nature and configura-
tion. In addition to two dummy (start and end) projects, JHS-I has five subprojects:
a transport and power system, river diversion during construction, dam construc-
tion, a stream power capacity project, and a spillway project (see Table 2).
Table 2. Detailed information for each activity
Activity DA C EI and W PA DP PP
S Dummy project S < 1, 2, 3
1 s (dummy activity) s < 1,2 11 1<4,5
1 2 2 20.0,0.05 1 < 3,4
2 5 2 32.6,0.055 2 < 4,t
3 5 2 16.7,0.03 3<t
4 3 4 37.5,0.08 4<t
2 s (dummy activity) s < 1,2 11 2 < 4,5
1 4 2 23.5,0.037 1<3
2 2 3 24,2,0.048 2<3
3 4 2 23.8,0.038 3 < 4,5
4 3 1 19.8,0.053 4<t
5 2 3 15.7,0.027 5<t
3 s (dummy activity) s<1 11 3 < 4,5
1 2 3 25.0,0.027 1 < 2,3
2 5 3 23.6,0.019 2<4
3 3 3 21.3,0.033 3<4
4 3 1 21.6,0.041 4<t
4 s (dummy activity) s < 1,2 11 4<T
1 2 1 11.6,0.027 1 < 3,4
2 5 2 16.3,0.047 2 < 4,5
3 4 3 11.4,0.028 3<t
4 2 3 13.5,0.059 4<t
5 4 1 11.5,0.037 5<t
5 s (dummy activity) s < 1,2,3 11 5<T
1 1 1 21.6,0.031 1<4
2 2 2 26.7,0.0122 2<4
3 5 1 17.2,0.018 3<4
4 3 2 22.0,0.03 4<t
T Dummy project
Note: DA = expected value for activity duration (month),
C = cost (million RMB), PA = activity predecessors, DP =
expected value for project duration, PP = project predeces-
sors, EI and W = environmental impact and weight.
Each activity must be performed in one of mi possible modes, each with a

corresponding duration, cost, environmental impact, and budget and cash flow
limit (see Tables 3 and 4). Hence, each activity has a certain maximal cost and
environmental impact unit requirement. Other relevant data for the analysis
1718 H. Zheng
are as follows: The maximum capital (units: million RMB) and environmental
impact for each time period are 14 and 10 units, (Bc , Be ) = (12, 10), respectively,
while the maximum capital (units: million RMB) and environmental impact for
each subproject time period are 6 and 6 units, (bc , be ) = (8, 6). The penalty cost
for each subproject in each time period is cpi = 12 (units: million RMB). The
evolutionary parameters are a population size of 20, maximal generation of 200,
an optimistic-pessimistic index of λ = 0.5, and a weight for each objective of
η1 = 0.5, η2 = 0.2, η3 = 0.3. The remaining variables are activity duration, cost,
activity predecessors, project duration, and project predecessors.
4.2 Case Study Results

To achieve managerial objectives, all project aspects must be optimized through
the best possible arrangement of start times and construction modes for each
subproject activity. According to the proposed model, for optimal scheduling,
the JHS-I activities should be arranged in the order below and corresponding
construction modes chosen to fulfil the decision maker’s requirements:
S = {P2 (1, 2, 3, 5, 4) → P3 (1, 2, 3, 4) → P1 (1, 3, 2, 4) → P4 (2, 1, 3, 5, 4) → P5 (2, 1, 3, 4)}.
Table 3. Optimal solution for the JHS-I case study
S = {P2 (1, 2, 3, 5, 4) → P3 (1, 2, 3, 4) → P1 (1, 3, 2, 4)

Optimal schedule
→ P4 (2, 1, 3, 5, 4) → P5 (2, 1, 3, 4)}
Sum of completion times for 54 (month)
all subprojects
Total penalty cost −18 (million RMB)
Total environmental impact −6.88
Fitness value 0.871
Table 4. Contrast between the optimal solution and actual data
z1 z2 z3 Optimal order and executed mode

Actual project 56 32 7.76 S = {P1 (2, 1, 4, 3) → P2 (2, 1, 3, 4, 5) →
data P3 (1, 3, 2, 4) → P4 (2, 5, 1, 4, 3) →
P5 (3, 2, 1, 4)}
Optimal result 54 −18 14.64 S = {P2 (1, 2, 3, 5, 4) → P3 (1, 2, 3, 4) →
P1 (1, 3, 2, 4)
→ P4 (2, 1, 3, 5, 4) → P5 (2, 1, 3, 4)}
Net decrease in 2 50 15.5
project value
Rate of decrease 3.57% 156% 200%
It should also be noted, however, that, because of cost and environmental

impact constraints, certain noncritical activities in subproject 1 (2), executable
between the 4th and 9th month, are flexible but other noncritical activities in
subproject 3 (2), executable between the 3rd and 8th month, are not. The model
thus defines the critical path meaningfully enough to be used in practical (rather
than simply theoretical) scheduling. In particular, it enables the project manager
to schedule activities according to the situation, which can be affected by such
factors as available manpower, equipment, holidays and the need to harmonize
with other parallel projects or activities. Moreover, even though a comparison of
the actual project data with the optimal result generated by the flc-hGA reveals
differences, these differences move in a positive direction. For example, the net
decrease from plus to minus in each objective, when considered specifically for
the changes in penalty costs and environmental impact, signals a shift from
penalty to reward. Overall, the rate of this decrease for each objective, which
ranges from 9.56% to 200%, signals an improvement in construction efficiency
that could bring considerable economic benefit to any construction project, but
especially to one that is large scale.
Admittedly, because the assumptions on which the mathematical model is
based may generate certain modeling errors, the results do not represent a 100%
optimal DTCETP-mP solution. Nevertheless, they can still serve as a useful
optimal scheduling guideline for decision makers in current construction projects.
5 Conclusions and Implications for Future Research
The multiple objective optimization model developed in this paper extends a

traditional single project model to an advanced multiple project trade-off model
able to determine optimal scheduling and construction mode selection for the
project at Jinping-I Hydropower Station. The model is designed to not only
minimize duration and penalty costs but also reduce the DTCETP-mP’s envi-
ronmental impact. It thus controls for the project constraints of duration, cost,
and environmental impact.
Because the DTCETP-mP is an NP-hard problem, the optimization method
employs an flc-hGA to solve it. The analytical results for the JHS-I case study
indicate that the trade-off between competing incommensurate objectives is
actually dominated by the project manager’s determination of the weight for
each objective. Hence, decision makers can use the importance of project objec-
tives to determine the optimum trends for time, cost, and environmental impact.
One major advantage of the proposed method is that it enables decision makers
to systematically and feasibly control the schedule according to an optimistic-
pessimistic index. In addition, the flc-hGA developed for the study can be used
to enhance optimization quality and stability.
The findings reported here suggest three important areas for future research,
all of which warrant equal concern. Firstly, a more holistic measurement should
be developed for environmental impact to ensure a more reasonable and effective
model. Secondly, future studies should address more complex practical problems,
1720 H. Zheng
such as those involving resource constraints, and uncertainties. Lastly, more

efficient heuristic methods need to be developed to solve NP-hard problems with
a greater number of constraints. The method and model proposed here offer a
useful foundation for all such investigations.
Acknowledgements. This research was supported by the Chongqing advanced and

applied basic research project (cstc2016jcyjA0442).
References
1. Afruzi EN, Najafi AA et al (2014) A multi-objective imperialist competitive algo-
rithm for solving discrete time, cost and quality trade-off problems with mode-
identity and resource-constrained situations. Comput Oper Res 50(10):80–96
2. Chen S, Chen B, Su M (2011) The cumulative effects of dam project on river
ecosystem based on multi-scale ecological network analysis. Procedia Environ Sci
5:12–17
3. Eshtehardian E, Afshar A, Abbasnia R (2009) Fuzzy-based MOGA approach to
stochastic time-cost trade-off problem. Autom Constr 18(5):692–701
4. Franck B, Neumann K, Schwindt C (2001) Truncated branch-and-bound, schedule-
construction, and schedule-improvement procedures for resource-constrained
project scheduling. OR Spectr 23(3):297–324
6. Harvey RT, Patterson JH (1979) An implicit enumeration algorithm for the
time/cost tradeoff problem in project network analysis. Found Control Eng
4(3):107–117
7. Holland JH (1992) Adaptation in natural and artificial systems. MIT Press, Cam-
bridge
8. Jeang A (2015) Project management for uncertainty with multiple objectives opti-
misation of time, cost and reliability. Int J Prod Res 53(5):1503–1526
9. Ke H, Ma J (2014) Modeling project time-cost trade-off in fuzzy random environ-
ment. Appl Soft Comput 19(2):80–85
10. Long LD, Ohsato A (2008) Fuzzy critical chain method for project scheduling
under resource constraints and uncertainty. Int J Project Manage 26(6):688–698
11. Meier C, Yassine AA et al (2016) Optimizing time-cost trade-offs in product
development projects with a multi-objective evolutionary algorithm. Res Eng Des
27(4):1–20
12. Monghasemi S, Nikoo MR et al (2014) A novel multi criteria decision making
model for optimizing time-cost-quality trade-off problems in construction projects.
Expert Syst Appl 42(6):3089–3104
13. Mungle S, Benyoucef L et al (2013) A fuzzy clustering-based genetic algorithm
approach for time-cost-quality trade-off problems: a case study of highway con-
struction project. Eng Appl Artif Intell 26(8):1953–1966
14. Said SS, Haouari M (2015) A hybrid simulation-optimization approach for the
robust discrete time/cost trade-off problem. Appl Math Comput 259:628–636
15. Tareghian HR, Taheri SH (2007) A solution procedure for the discrete time, cost
and quality tradeoff problem using electromagnetic scatter search. Appl Math
Comput 190(2):1136–1145
16. Wang PT (1997) Speeding up the search process of genetic algorithm by fuzzy
logic. In: Proceedings of 5th European congress on intelligent techniques and soft
computing
17. Xu J, Zheng H et al (2012) Discrete time-cost-environment trade-off problem for
large-scale construction systems with multiple modes under fuzzy uncertainty and
its application to Jinping-II hydroelectric project. Int J Project Manage 30(8):950–
966
18. Zheng H (2014) The fuzzy time-cost-quality-environment trade-off analysis of
resource-constrained multi-mode construction systems for large-scale hydroelectric
projects
19. Zheng H (2014) The fuzzy time-cost-quality-environment trade-off analysis of
resource-constrained multi-mode construction systems for large-scale hydroelectric
projects. Lect Notes Electr Eng 242:1403–1415
20. Zheng H, Zhong L (2017) Discrete time-cost-environment trade-off problem and
its application to a large-scale construction project. Springer Singapore
Research on Equalization of Public Services
Based on Changes of Urban and Rural
Population Structure
Qian Fang(B) and Yi Sheng
Sichuan Academy of Social Sciences,

Chengdu 610072, Sichuan, People’s Republic of China
lily lily009@163.com
Abstract. This paper starts the discussion from five aspects population
distribution, age structure, employment characteristics, family charac-
teristics and flow population, and conducts qualitative and quantitative
analysis on the changes of urban and rural population of new period
of China, so as to master the basic situation and development trend of
urban and rural population. On this basis, the paper proposes sugges-
tion and countermeasures on equalization of public services in urban and
rural areas in our country, basically from three aspects-“increasing the
flexibility of the urban and rural public service supply mode”, “promot-
ing the urban and rural service supply balance and accessibility” and
“strengthening the effective supply to important groups and regions”.
Keywords: Equalization of public services · Urban and rural areas ·

Population structure · Public service supply
The problem of urban and rural areas is always a hot issue during the process of
China’s development. In November 2013, in the “Decision on Several Significant
Problems on Comprehensively Deepening Reform by the Central Committee of
the Communist Party of China” issued by the Third Plenary Session of the
18th PCP Central Committee, it is clearly proposed to “propel the equalization
of basic public services in urban and rural areas.” There is no doubt that in
the new period, the government has taken the perfection of urban and rural
development integration mechanism as the main path for solving the problem of
the unbalanced development between urban and rural areas, and has taken the
equalization of public services as the main measure for relieving the conflict on
development between urban and rural areas.
On the premise of the given orientation, it is urgent to conduct reasonable
allocation on the public service resources of urban and rural areas and ensure
the balance. In order to solve this problem, three aspects shall be taken into con-
sideration: Firstly, what is the current situation of urban and rural population
Research on Equalization of Public Services 1723
structure? Secondly, how to define the equalization of urban and rural public ser-
vices? Thirdly, how to realize the equalization of urban and rural public services?
Domestic and foreign scholars have given some answers to the above questions,
but not sufficient. Take the urban and rural population structure as an exam-
ple. In 2011, the proportion of urban residents in our country exceeded that of
the rural residents for the first time, which is called by media as “the reversal
of urban and rural population structure since thousands of years in China” [3];
China has turned to the urban civilization stage with characteristics of industry
and service industry from the giant agricultural country and agricultural civ-
ilization. It is noted in the “Social Blue Book” by Chinese Academy of Social
Sciences that: “this is not only a change in the urban population percentage, but
means the extremely profound change on the production method, occupational
structure, consumption behavior, living method and value system of people”
[12]. Is 2011 the time-point for the qualitative change of the urban and rural
structure of China? If we narrow the scope to four major economic regions east
region, middle region, west region and northeast region, we can find that such a
so-called reversal has already appeared in Jiangsu, Zhejiang and Fujian as early
as 2005. In the west region, the rural population in many regions is still more
than that of the urban population (such as Guangxi, Sichuan, Guizhou, Yunnan,
etc.), and no reversal has happened. The conclusion obtained from the average
value of the statistical data is both scientific and some bit of biased, which shall
not be taken too much consideration. Besides, the urban and rural population
structure often discussed by some scholars is actually the urban and rural reg-
istered population structure, i.e., the personal status relationship of urban and
rural residents, which is greatly different from the actual distribution of urban
and rural population. In 2014, the population of separation of registered and
actual residence in our country was 298 million, and the floating population was
253 million. It means that the era judging the population distribution accord-
ing to family register has passed in our country. In brief, the urban and rural
population structure in our country has come to a complex stage; from now on,
equalization of public service will step into a period of overall promotion, accu-
rate definition and courageous attempt. On the basis of accurate mastering of
the change of urban and rural population structure, we will propose reform ideas
and suggestion aiming at the new problem on urban and rural public service,
which is an interesting topic, and also the original intension of this research.
2 Research Review
When conducting retrieval by taking CNKI as the retrieval source, “public ser-
vice” and “public product” as the title and key word, and “urban and rural”
as the title, more than 1800 results are obtained, in which there are 189 master
and doctoral thesis, 149 conference papers, more than 570 papers on newspaper
and 940 journal literatures.
1724 Q. Fang and Y. Sheng
There are 344 literatures with high quality, which are published on SCI, EI,
Chinese core periodicals and CSSCI source periodicals. The following conclu-
sion is obtained after conducting analysis on literatures by utilizing Citespace
software.
Firstly, the high quality periodical achievements are concentrated during
2003–2016, mainly after 2010. Secondly, the top three authors in terms of the
number of publishing are Wu Yemiao, Liu Chengkui and Yu Yaguai. No prefer-
able cooperative relationship is formed between the authors, and most litera-
tures are researched independently. Thirdly, the top three research institutions
are the Institute of Public Administration of Sichuan University, the School
of Government of Nanjing University and the Financial and Tax Institute of
Shandong University of Finance. Fourthly, the research emphasis on urban and
rural public service field in our country is concentrated in basic public service,
equalization, urban and rural overall development, urban and rural gap, rural
public products, sports public service and public finance. The research emphasis
has been gradually transferred from the public product (2004), urban and rural
overall development (2005), urban and rural public service (2006–2007) of the
early stage to equalization of basic public service (2008–2010) and urban and
rural basic public service (2010). Since “the 11th five-year plan”, the domestic
research has tended to specific contents such as social insurance, sports public
service, infrastructure, etc. The research pays more attention to the evaluation
on public service, to discuss on public service in combination of the theme of
urbanization. Fifthly, there are two sudden nodes during the whole time period:
public product and urban and rural overall development. The little node quantity
indicates that the research on this field is not active enough, without significant
emerging trend. Of course, it may be due to the short term of the research on
this field in our country.
These researches mainly explore the equalization of public services from four
aspects. First, it is the research on the specific contents of the basic public ser-
vices. Secondly, it is the argument on the unbalanced situation of urban and
rural public service. Just as what is proposed by Wu [14] that the unbalanced
allocation of urban and rural public resources in our country is mainly expressed
on infrastructure, basic education, social insurance and public medical health.
According to Li [6], the unbalanced allocation of urban and rural public resource
in our country is embodied on social insurance resource, social welfare resource,
public health resource, basic education resource and infrastructure construction.
Thirdly, explore the reasons for unbalanced urban and rural public service. It
is regarded by some scholars that the subsidy on industry by agriculture and
the subsidy on city by village in early stages of our country is the fundamental
reason for unbalanced urban and rural public resource allocation [2]. It is con-
sidered by some scholars that the long term unbalanced urban and rural public
resource allocation is related to the thoughts of “inevitable primary accumula-
tion”, “insufficient national financial resources” and “obstruction on efficiency
by justice” in the society [15]; some scholars emphasize economic development
level [5], and that the finance is the initiator of the evil [7]. Fourthly, explore the
practice of promoting the equalization of urban and rural public service. Some
scholars propose on ideology aspect to formulate the development strategy in
accordance with the urban and rural integrated development, to increase the
investment intensity in rural areas [18], to break through the functional restraint
of “economic-oriented government”, and to create public service-oriented gov-
ernment [16]; some scholars propose on financial aspect to take the investment by
the Party Central Committee as the leading force, and to improve the efficiency
of implementation of policy [11]. Some scholars think that it is needed to stan-
dardize the government expenditure behavior, and clarify the functions of various
levels of governments, to increase the financial investment of the central and local
governments [2]. Some scholars advocate to promote the balanced allocation of
urban and rural public resources through marketization [1,19], but it is needed to
avoid the corruption generated thereby. Some scholars think that it is available
to conduct allocation on public resources by non-profit organizations [13]. Some
scholars think that the balanced allocation of public resources is actually a prob-
lem about institutional arrangement in essence; To change the administrative
philosophy of local government, to establish sustainable financial expenditure
mechanism of rural public resource, to establish the peasants-participating pub-
lic resource allocation decision mechanism, and to perfect the supervision and
restraint mechanism of balanced allocation of social public resource, are four
path choices for promoting the balanced allocation of public resources [6]; some
scholars propose that it is necessary to realize the integration and association
within the system, between the systems, between government and the society as
well as among the diversified subjects of the society [4]. Besides, some scholars
propose the method to improve the dual development on urban and rural areas
from the aspect of specific field of public resources such as infrastructure, public
education and medical health.
To sum up, the existing researches have laid firm foundation for us to continue
the exploration on the urban and rural public service field, but there are still
deficiencies on these researches. Firstly, compared with the discussion over the
country (for example, 31 provinces and cities), there is obviously insufficient
discussion on the equalization of urban and rural public services. Secondly, most
researches are only conducted superficially on the unbalanced allocation of public
service, resulting in superficial and impractical reform path. Thirdly, the existing
researches are static, taking no population mobility and demand variability into
consideration.
3 Changes on Urban and Rural Population Structure

Characteristics in China
In recent ten years, there have been great changes on urban and rural population
structure in China, which are embodied on aspects such as space, ages, family,
education, income and employment. As for the population, the urban population
in our country in 2011 exceeded more than 50% of the total population, forming
a “neck and neck” situation on urban and rural population. On space distribu-
tion, the trend of migration from rural to urban area still goes on, and weight
transition of urban and rural population and the adjustment of population den-
sity have become normalcy. On aspect of age groups, the fierce urban and rural
population aging trend still goes on, and the situation of economic increase
depending on “demographic dividend” has passed. On aspect of employment
selection, the urban and rural population rushes to the tertiary industry (ser-
vice industry), to make the employment transferred from “industry-oriented” to
“service-oriented” form. With the resultant force of decreased family population
and higher income, it has been the general situation for public service to lean to
fields such as medical health and pension.
3.1 Changes on Space Distribution of Urban and Rural Population
Since the establishment of the nation, China has had an increasingly grown
population, which increased to 1.36 billion in 2013 from 0.54 billion in 1949. The
urban population has been on the rise, which increased to 0.73 billion in 2013
from 0.0577 billion in the early years of the new nation. Compared to the “up
and down” situation of urban population, the rural population proportion was
on a slow and continuous downtrend. In the early years of the new nation, the
rural population in our country was near to 90%, which decreased to 46.27% in
2013.
Extend the analysis vision field to the four major economic regions, i.e., the
east region, the middle region, the west region and the northeast region. The
changes of urban and rural population show both generality and difference. On
one hand, the changing trends of urban and rural population proportion in the
east region, the middle region, the west region and the northeast region are
almost the same, i.e., the proportion of urban population increases year-by-year,
as shown in Table 1. On the other hand, there is significant difference on the
proportion of urban population among the four economic regions. At the end of
“the 10th five-year plan” (2005), the proportion of the urban population of the
east region had been near to 60%, and the proportion of the urban population of
the northeast had been near to 55%, while the proportion of the urban population
Table 1. Changes on proportion of urban population in the four major economic

regions in China (2005–2013)
Region 2005 2006 2007 2008 2009 2010 2011 2012 2013
China 42.99 44.34 45.89 46.99 48.34 49.95 51.27 52.57 53.73
The east region 59.23 60.15 60.87 61.65 62.48 64.43 65.19 66.11 66.92
The middle region 37.58 38.96 40.27 41.73 43.03 44.44 46.28 47.98 49.26
The west region 35.18 36.15 37.31 38.53 39.61 41.45 42.81 44.26 45.43
The northeast region 54.77 55.15 55.42 56.22 56.39 57.04 57.98 58.75 59.35
Data source: Obtained by arrangement of “China Statistical Yearbook 2014”
of the middle region and the west region was between 35%–38%. During 2006–
2013, the urbanization of population in east region slowed down, with only 6.77%
growth during 8 years; the middle region had the quickest growth speed, with
10.3% growth; the west region had a growth speed of 9.28%; the northeast region
had the slowest growth speed, with only 4.2% growth during 8 years. Moreover,
the active force for urbanization during the latest ten years in China was in the
middle region and the west region, which are also the regions with the largest
development potential in the future.
3.2 Changes on Age Structure of Urban and Rural Population
It is agreed by most domestic and foreign experts that “demographic dividend” is

the main active force for the rapid development of 30 years since the reform and
the opening-up policy. The young and middle-aged labor force plays a huge role
in the rapid economic growing period. However, as time goes by, the population
proportion of national labor force significantly decreases. During 2011–2014, the
proportion of the population of 16–59 years’ old decreased year-by-year, from
0.941 billion in 2011 to 0.916 billion in 2014. Correspondingly, the proportion of
the population above 60 and 65 years’ old increased year-by-year.
Up until the end of 2013, except for Tibet and Xinjiang regions, the 29 regions
of the whole country had entered into the aging society to varying degrees. In
which the aging degree of Tianjin, Shanghai, Jiangsu and Shandong of the east
region, Anhui and Hunan of the middle region as well as Sichuan and Chongqing
of the west region had exceeded than 10%. The top three regions of aging degree
were Chongqing, Sichuan and Jiangsu.
Viewing from the four major economic regions, the proportion of the labor
force population of 15–64 years’ old in the east region is higher than that of
the middle region and the west region, as shown in Table 2. In consideration
that the permanent resident population of the east region occupies 37% of the
31 provinces and cities, which is much higher than that of the west region, the
middle region and the northeast region, the development power in the east region
is more sufficient and higher.
Seen from the aspect of dependency ratio, the national aging population
dependency ratio increased from 8% in 1982 to 13% in 2013; the children depen-
dency ratio decreased from 54.6% in 1982 to 22.2% in 2013. Aging of population
Table 2. Comparison on population age structure of the four major economic regions
(2013)
Region 0–14(%) 15–64(%) Over 65 (%)

The east region 14.87 75.52 9.61
The middle region 18.34 72.09 9.57
The west region 18.2 71.89 9.9
The northeast region 11.3 79.07 9.63
has a considerable influence on family and the society. Compared with the east
region, the middle region and the northeast region, the aging population depen-
dency ratio in the middle region is on the high side, and that in the west region
and the northeast region is on the low side. The dependency ratio of children
in the west region is the most highest, and that in the northeast region and the
east region is on the low side. As a whole, the dependency ratio of families in
the west region is the highest (38.29%), and that of the northeast region is the
lowest (26.57%), and the middle region and the east region are between the two.
3.3 Changes on Employment Characteristics of Urban and Rural

Population
During the 30 years since the reform and opening-up policy, the changes on
employment structure of national urban and rural population are firstly embod-
ied on employment population distribution of three industries; it appears in a sit-
uation that the proportion of the employment population of the primary industry
decreases, and that of the secondary industry increases but with frequent fluctu-
ation, and that of the tertiary industry continuously increases. The national total
employment population in 2013 was 769.77 million, in which 382.4 million were
urban employment population and 387.37 million were rural employment popu-
lation. The employment population between urban and rural areas appeared in
a “fifty-fifty” situation.
There are totally different employment structures in urban and rural areas.
The urban employment population includes the people working in state-owned
units, collective units, joint stock partnership units and joint ownership units,
as well as in foreign-funded enterprises, private enterprises and individual enter-
prises. There are mainly three types of employment population in rural areas:
the employment population in private enterprises, individual employment and
faming population. Since 2004, the proportion of population working in private
enterprises of rural employment population has exceeded than that of individ-
ual enterprises, showing the agriculture industrialization and enterprise-oriented
development trend, as shown in Fig. 1.
Fig. 1. Changes on proportion of employment population in private and individual

enterprises in rural areas of China (2000–2013)
Fig. 2. Changes on proportion of urban employment population in China (2000–2013)
A large part of urban employment population “works in institutions as well as

enterprise and public institutions of various levels”, and please refer to Fig. 2 for
the urban employment population situation. If the urban employment popula-
tion working in state-owned enterprises, collective enterprises, joint-stock enter-
prises and associated enterprises is collectively called as “state enterprises”, we
can see two major changes of the national employment. Firstly, the employ-
ment population of “state units” appears in a slow rising trend. In the “state
units”, the employment population in state-owned units, collective units, joint
stock partnership units and joint units decreases, and that in limited liability
companies and joint-stock companies increases. These changes are correspond-
ing to the state-owned enterprise reform and the separation of ownership and
management, as well as the mixed ownership reform. Secondly, the employment
population in private enterprises, investment units (foreign investment as well as
investment made by businessmen of Hong Kong, Macau and Taiwan) and indi-
vidual employment increases. The increase speed of the employment population
of private enterprises is the fastest, followed by individual employment, and the
investment unit is the slowest.
3.4 Changes on Family Characteristics of Urban and Rural

Population
During 1949–2010, there was a small fluctuation in family scale in our coun-
try, decreasing from 4.43 persons for each household, which was the maximum
value, to 3.10 persons for each household. In 2013, the national per household
scale decreased to 2.98 persons for each household. When making comparison
among the families in the east region, the middle region, the west region and
the northeast region, the middle region has the largest family household scale,
achieving 3.19 persons for each household. The northeast region has the smallest
family household scale, with only 2.76 persons for each household. The middle
region and the west region have the family household scales of 2.98 persons and
3.04 persons respectively.
From the comparison on urban and rural residents income situations among
the east region, the middle region, the west region and the northeast region, we
can know that the per capital income of the urban residents in the east region
is the highest, and that in the west region is the lowest, and that in the middle
region and the northeast region is between the two. The urban resident income
in the east region is 1.44 times of that in the west region, and the per capita net
income of rural residents is 1.76 times of that in the west region. Viewing from
the data in the recent ten years, the proportion of income between residents in
the east region and the west region gradually decreases, from 1.52 in 2005 to
1.43 in 2013, and the proportion of rural resident income decreases from 1.98 in
2005 to 1.76 in 2013.
According to the data, the top four expenditures for urban residents are
“food”, “transportation and communication”, “culture, education and entertain-
ment” and “clothing”; the top four expenditures for rural residents are “food”,
“residence”, “transportation and communication” and “medical health”. There
are great difference on public service demands between rural residents and urban
residents. “Transportation and communication” is the common part for urban
and rural residents, but rural residents need more on the improvement of “res-
idence” and “basic medical service”, while urban residents pay more attention
on “culture, education and entertainment”.
It can be known from the comparison on urban residents expenditure in
the east region, the middle region, the northeast region and the west region
(as shown in Table 3) that the top four expenditures for urban residents in the
four regions are “food”, “transportation and communication”, “culture, educa-
tion and entertainment” and “residence”. In which the consumption on “food”
(37.07%) is the largest in the west region; the consumption on “transporta-
tion and communication” (16.36%) and “culture, education and entertainment”
(13.35%) are the largest in the east region; the consumption on “residence” is
the largest in the northeast region. It can be known according to the expendi-
ture of rural residents among the east region, the middle region, the northeast
region and the west region (as shown in Table 4) that the top four expenditures
for rural residents in the four regions are “food”, “residence”, “transportation
and communication” and “medical health”. In which the consumption on “food”
occupies the largest proportion (38.17%) in the west region; the consumption on
“residence” (20.74%) occupies the largest proportion in the northeast region; the
“transportation and communication” (13.22%) occupies the largest proportion
in the east region; the consumption on “medical health” (12.16%) occupies the
largest proportion in the middle region. Therefore, the urban and rural residents
in the four regions have different emphasis on public service supply, and the
quality and quantity of public service supply shall be improved according to the
emphasis of resident expenditures.
Table 3. Proportion of per capita consumption expenditure of urban residents in the

four major economic regions (2013)
Region Unit Food Clothing Residence Household Transportation Culture, Medical Others
articles equipment and and enter- education health
communication tainment
National % 35.02 10.55 9.68 6.74 15.19 12.73 6.2 3.88
The % 35.33 9.04 9.45 6.61 16.36 13.15 5.78 4.29
east
region
The % 35.68 11.36 10.02 6.97 13.28 12.87 6.38 3.41
middle
region
The % 32.28 12.19 11.25 5.92 13.44 11.62 9.08 4.23
north-
east
region
The % 37.07 11.51 9.57 6.67 12.99 11.93 6.69 3.57
west
region
Data source: obtained by calculation and arrangement of “China Statistical Yearbook 2014”
Table 4. Per capital consumption expenditure of rural residents in the four major
economic regions (2013)
Region Unit Food Clothing Residence Household Transportation Culture, Medical Others
articles equipment and enter- educa- health
and commu- tainment tion
nication
National % 37.67 6.62 18.62 5.84 12.01 7.33 9.27 2.64
The east % 37.35 6.66 16.98 5.66 13.22 8.12 8.86 3.16
region
The % 34.45 7.82 17.28 4.03 12.27 9.02 12.16 2.96
middle
region
The % 37.42 6.1 20.74 6.34 10.33 6.94 9.46 2.36
northeast
region
The west % 38.17 6.84 18.98 5.85 12.09 6.12 9.46 2.49
region
Data source: obtained by calculation and arrangement of “China Statistical Yearbook 2014”
3.5 Changes on Space and Characteristics of Floating Population
The floating population of China in 2014 was 0.253 billion, occupying 18.5% of
the total population of the whole country. Liu, et al. [8] conducted measurement
on the directivity and activeness of regional floating population in our country,
and divide the floating population regional type into four types. In which the
urban agglomeration regions such as the Pearl River Delta, the Yangtze River
Delta and Beijing are the active regions for net inflow. The rural areas in the
east region and the middle and west regions with concentrated population are
the active regions for net outflow. The periphery or rural areas of the developed
urban agglomeration in Zhejiang, Fujian, Guangdong in the east region are the
active balanced regions. The vast middle region, the west region, Shandong,
Hebei and north Jiangsu in the east region, Jilin and Liaoning in the northeast
region are the non-active regions. There is relatively stable distribution propor-
tion of floating population in the east region, the middle region and the west
region during 2000–2010, with the increase rate of 115%–120%. The floating
population in the east region occupies 2/3 of that of the whole country, and
with rapid growth speed. The floating population concentration area in coastal
regions gradually diffuses, with a continuous trend [9]. This trend is very sig-
nificant in the Yangtze River Delta area, and the diffusion in the two floating
population concentration areas of the Pearl River Delta and the Beijing-Tianjin-
Hebei region is relatively limited. Megalopolises such as provincial capitals of
inland regions attract a large amount of floating population; at the same time,
there is significant northward movement on the distribution center of floating
population (Table 5).
Table 5. Differentiation on zones of floating population in China [10]
Index Year The east region The middle region The west region
Total floating 2000 5110.4 1237.8 1552.6
population (10
thousand persons)
2010 10987.1 2661.9 3407.1
The growth rate during 115.00% 115.10% 119.40%
the 10 years
Proportion of floating 2000 64.7 15.7 19.7
population in the
whole country (%)
2010 64.4 15.6 20
Proportion of floating 2000 11.1 3 4.4
family resister (%)
2010 22.1 5.8 8.8
When making comparison on the population floating regional difference in

the east region, the middle region and the west region, we can find that the
population floating in the east region is mainly the floating to other provinces,
and that in the middle region is mainly the floating within county, and that in
the west region is mainly the floating between provinces. As a whole, the float-
ing between provinces occupies an important role; however, the within-province
floating has a more fierce development trend, with the scale not less than the
cross-province floating. Because of the lower cost and smaller obstruction, the
floating within province of population will become the mainstream, and local
urbanization will become the main method for the urbanization in our country.
Besides, the domestic population floating mode appears in significant
“domestication trend”. Most Chinese families adopt the “progressive floating
mode”, i.e., some family members enter into cities to work and live firstly. There
are differences on the characteristics of family floating in different regions. For
example, “the complete family type floating occupies the highest proportion in
the middle region and the cross-county floating, which occupies a relatively low
proportion in economic developed regions and under developed regions. This
phenomenon illustrates that the threshold or living cost for family reunion in
developed regions is relatively high, and the under developed regions fail to have
an attraction for family reunion to floating population because of the lack of
high quality resources” [17].
3.6 Countermeasures and Suggestions on Promoting of Equalization

of Urban and Rural Public Service
Based on the change features on urban and rural population structure character-
istics, the author proposes three suggestions and specific measures for promoting
the equalization of urban and rural public services in China.
(1) Increase the Flexibility of Urban and Rural Public Service Supply Modes
Strengthen the statistical investigation, and accurately master the urban and
rural population characteristics. It is suggested to constitute the yearly popula-
tion sample survey working group by the statistical bureau and related depart-
ments by relying on existing population census organizations in our country,
to develop yearly survey on population quantity, age, structure, distribution,
residence and public service willing. Hold the joint conference of the statistic
department, the public security department, the family planning department,
the civil affair department, the environmental-protection department, the finan-
cial department and the population and society department according to the
requirements on data collection, to clarify the rights and obligations of vari-
ous departments on coordination for collection index. Establish urban and rural
public service evaluation index system, to conduct analysis on data collected by
each year by multiple forms including service outsourcing, cooperation between
governments, franchising, subsidy and voluntary services, and the government
releases the “Survey Report on Urban and Rural Public Services Situations in
China” to the whole country each year. Change the method of allocating financial
capital according to registered population, to promote the coordination between
the urban permanent residence population and the central fiscal appropriation.
Eliminate the urban-oriented policies, and deepen the household registration
system reform. Explore the way to take the ID card as the national use and
unchanged for lifelong social guarantee number, and implement the social insur-
ance one-card system for the national social insurance unities. On the premise
of giving respect to the autonomous right of peasants to live and stay in cities,
encourage them to go to the town. Prohibit taking the giving up of contractual
right of land, right to the use of curtilage and the collective benefit distribution
right as the conditions for peasants to settle down in cities. Broaden the resi-
dence migration policy, and take the legal stable residence or the legal stable
employment registered residence as the basic conditions, to allow rural residents
to apply for the registration of permanent residence household, and broaden the
settling conditions for purchasing, building and inheriting of houses in cities,
and simplify the population registration system with the principle of “conduct-
ing management on population through house, to achieving between population
and household”.
Promote the interaction of population, and construct integrated labor force
market. It is suggested to drive the equalization of urban and rural public ser-
vice by the urban and rural human resource integration. Establish the urban-
rural unified employment standard, to desalt the localization and census register
conditions of employment field, determine the subjective position of the labor
force market, and realize the allocation of labor force resources by the market.
Establish the labor force market with “online + offline” dual platforms, and
improve the construction of information network, to establish the information
system with the joint participation of enterprises, job seekers and occupation
introduction institutions. Perfect the employee training and promotion system,
and perfect the incentive mechanism, and encourage enterprises to maintain the
employee training rights, and enlarge the promotion space for peasant workers,
to ensure them to acquire the living capital in cities in an effective way.
(2) Promote the Balance and Accessibility of Urban and Rural Service Supply
Through giving directional difference to the rural area, relieve the unbal-
anced urban and rural structure. Further perfect the infrastructure construction
of transportation in rural areas, to realize the smooth roads, and finish the con-
struction of outward roads of poor villages. Improve the management and main-
tenance level of rural roads, and stable the road situation of rural roads, and
extend the service life, and solidify the rural roads construction achievements.
Promote the urban-rural integration, and perfect the urban system planning and
village and town building plan for all the urban and rural areas. Pay more atten-
tion on aging population and school children on rural public service resources.
Promote development on small and middle-sized cities, and find out support-
ing point for city back seeding countryside. It is suggested to conduct compre-
hensive consideration on terrain and economic conditions on the basis of building
the center cities in the middle region and the west region, to select a batch of
type I and type small-sized cities, to cultivate and encourage them to grow larger
and stronger, to form different layers of radiation circles, to provide substantial
support for the development of rural areas.
Formulate differential equalized strategy according to regional characteris-
tics. Conduct consideration by linking the infrastructure and the accessibility of
public services with the space behaviors of residents, to strengthen the accessibil-
ity of public space, employment, administrative center and other comprehensive
public services. Conduct different environment governance measures between
urban and rural areas, and lay the emphasis on environment improvement (such
as on aspects of drinking water, toilet, rubbish and waste water processing) in
rural areas, and lay the emphasis on air quality and traffic congestion improve-
ment in urban areas. According to the public service levels in the east regions,
the middle region, the west region and the northeast region, formulate different
propelling schemes. Highlight the characteristics of urban and rural areas, and
strengthen the advantages of landscapes in urban and rural areas. Avoid rural
areas copying the urban community mode, or substituting rural landscape with
urban elements, and replacing greenbelts with firm grounds. It is not advocated
to promoting the unified allocation standard between rural and urban areas, to
avoid the loss of rural characteristics and folk culture in rural areas.
(3) Strengthen the Effective Supply to Important Groups and Key Areas
Improve the public service supply level and guide the peasant workers to
integrate in cities in an orderly way. Perfect the infrastructure construction of
the labor force market, and set up and integrate the “urban and rural labor
force market information platform” integrated with enterprise employment infor-
mation, labor force employment hunting information, industry employee infor-
mation and employment policy. Increase the employment stability of peasant
workers, and improve the vocational skill quality of peasant workers by means
such as releasing free training tickets and implementing enterprise training fee
deduction taxes and duties, to increase opportunities for employment. Increase
the participation of the peasant worker group to urban public affairs, and set
up peasant worker service platform in communities, to help them adapt to the
city life as soon as possible by providing services such as formality agency, living
guidance, employment consultation and children entrance information, etc.
Construct service network, so as to improve the medical and pension services
for senior citizens in urban and rural areas. Conduct general investigation on
the data, residence and family conditions of the senior citizens in national urban
and rural areas, to establish documents for the senior citizens. Arrange medical
health institutions or nursing institutions for the aged in urban and rural areas,
to build the pension community integrated with medical and pension services,
to form the urban-rural covering and reasonably functional medical and pension
service networks with sharing resources between pension service and medical
health service. Improve the ability of door-to-door service for senior citizens
by the basic level medical health institutions, and encourage medical institu-
tions to provide green channel for senior citizens to seek for medical service,
and encourage nursing institutions for the aged to provide medical service for
senior citizens. Guide the enterprises to develop integrated pension communi-
ties integrated with residence, health care, pension, recovery and entertainment;
guide the community health service institutions to conduct transformation and
upgrading, and encourage the comprehensive medical institutions to cooperate
with nursing institutions for the aged, to give play to the radiation and leading
functions of medical alliance.
References
1. Chen J (2008) Market-oriented allocation of public resources: theoretical basis, risk
analysis and route selection. J Municipal Party Committee Party Sch Yinchuan
2. Chen J (2010) Empirical analysis on different influences of fiscal expenditure behav-
ior on infrastructures of urban and rural areas. Xinjiang Acad Soc Sci 3:24–29 (in
Chinese)
3. Daily H (2011) Reversal of urban and rural population structure. Hubei Daily
4. Jing T (2015) Base line justice: welfare basis point of balanced development
between justice and development. J Beijing Technol Univ (Soc Sci Ed) 1:1–7 (in
Chinese)
5. Lei X, Zhang N (2012) Research overview on urban and rural overall development
of social insurance. Stud Soc Secur 1:91–96 (in Chinese)
6. Li J, Zhou Q (2015) Supervision mechanism and route selection on balanced allo-

cation of urban and rural public resources. Shanghai City Manage 1:34–37 (in
Chinese)
7. Li Y, Peng Y (2016) Research on fiscal policy of the equal access to basic public
services between urban and rural areas. Finan Theory Pract 3:56–61 (in Chinese)
8. Liu S, Deng Y, Hu Z (2010) Division method and space distribution characteristics
for regional types of floating population in China. J Geogr Sci 10:1187–1197 (in
Chinese)
9. Liu T, Qin Y, Cao G (2015) Space pattern evolution mechanism of floating popu-
lation in China and urbanization effect-analysis based on population census county
basis of 2000 and 2010. J Geogr Sci 4:567–581 (in Chinese)
10. Liu Y, Qi Y, Cao GJ (2015) Evaluation mechanism of space pattern of floating
population in China and urbanization effect-analysis based on the county data of
population census of 2000 and 2010. J Geogr Sci 4:2–13 (in Chinese)
11. Luo Y (2010) Analysis on economic effect of difference on infrastructure of urban
and rural areas in China-based on space panel measurement model. Chin Rural
12. Social Sciences Academic Press (2011) Social blue book: Chinese social situation
analysis and prediction. Social Sciences Academic Press
13. Qin B, Xiao S (2013) Public resource allocation on non-profit orgnizations. Finan
14. Wu L, Xu C (2014) Discussion and analysis on balanced allocation on urban and
rural public resources. Acad Commun 5:131–134 (in Chinese)
15. Wu Z (2008) The chinese society stepping to justice. Shandong People’s Publishing
House, Jinan
16. Wu Z (2013) The institutional constraint and solution on urban and rural education
integration. J South China Normal Univ (Soc Sci Ed) 1:29–32 (in Chinese)
17. Yang J, Chen C (2013) Current state and characteristics of floating population
family-oriented pattern: analysis on characteristics of floating process. Popul Dev
18. Zheng Y, Liu Z, Wang Y (2012) The infrastructure construction of urban and
rural overall development under the new rural background. Soc Sci Res 6:25–29
(in Chinese)
19. Zhou C (2013) Strategic thinking on promotion of public resource allocation
market-oriented reform. Forum People 11:66–67 (in Chinese)
Economic Cycle, Accounting Conservatism
and Financial Constraints
Hong Wang1(B) , Pan Liang1 , and Juqiu Deng2

1
979166074@qq.com
2
Economics School of Sichuan University, Chengdu 610065,
Abstract. This paper adopts cashflow sensitivity of cash holding model

to investigate the influence of economic cycles and accounting conser-
vatism on corporate financial constraints, and to investigate the different
impact of the two mechanisms from the property right perspective. Our
results show that the recession of economic cycles aggravates the degree
of corporate financial constraints, and this negative influence is more
apparent on non-SOEs than SOEs. Moreover, the increase of accounting
conservatism can release corporate financial constraints, and the allevi-
ating effect of accounting conservatism is more significant on non-SOEs.
Keywords: Economic cycle · Accounting conservatism · Financial

constraints
1 Introduction
Financial constraints can effect the corporate innovation [9] and investment effi-
ciency [10], which seriously restrict the further development of corporates and
ultimately affect economic growth. The World Bank report [7] shows that 75% of
non-financial listed companies in China choose financial constraints as a major
obstacle to business development, the highest percentage in 80 surveyed coun-
tries. Chinese corporates are facing the test of financial constraints.
China’s economy characteristics has been markedly different from the past
30 years. It mainly expresses in the following aspects: (i) economic growth tailed
off to around 7%, or lower; (ii) a large sum of money flows into the real estate and
infrastructure construction instead of the real economy; (iii) the real economy
such as manufacturing achieve consistently poor performance. The lack of capital
in the real economy is detrimental to the long-term sustainable economic growth.
And capital flowing into the real estate sector may not be able to promote
economic growth and add to economic stability. Instead, it may boost the housing
prices and make the real economy more difficult. Against this background, it is of
practical significance to study the causes and solutions of the financial constraints
of traditional enterprises.
1738 H. Wang et al.
Theories have studied the causes of financial constraints from information

asymmetry, agency costs, firm size and political connection [6,19,23]. However,
all these studies are carried out from the microperspective. After taking these
factors into account, the phenomenon that the degree of financial constraints of
Chinese corporates fluctuates with the macroeconomic can’t be explained. Trying
to find out the reason, this article raises the first question: since the fluctuation
of economic cycle is one of the main characteristics of market economy, do the
economic cycles affect the corporate financial constraints? If so, what is the
influence mechanism?
Accounting conservatism is an important index to measure the quality of
accounting information [3,4]. Fair value has been questioned after 2008 financial
crisis, while theories begin to rethink the positive effect of accounting conser-
vatism [17,24]. The increases in conservatism improve the firm information envi-
ronment and lead to subsequent decreases in information asymmetries between
firm insiders and outsiders [15]. And accounting conservatism reduces under
investment in the presence of information frictions [2]. So this article raises the
second question: is it easier for the enterprises with high accounting conservatism
to get financing in the capital markets? Can accounting conservatism ease the
financial constraints, especially in the economic recession?
Using the financial data of Chinese listed A - share listed manufacturing com-
panies, we test whether and how economic cycle and accounting conservatism
affect corporate financial constraints to obtain the empirical evidence. Our first
contribution is our explicit focus on the impact of economic cycles on corpo-
rate financial constraints. Typically, existing studies tend to focus on studying
the cause of financial constraints from the enterprise’s perspective. Macroeco-
nomics and enterprise behaviors are disjointed in existing research [11]. There
is a strong evidence, however, that the economic cycles can influence corpo-
rate financial constraints. As financial constraints may be specific to different
property rights, we investigate the different impact from the property right per-
spective. It can broaden the research horizon on influencing factors of financial
constraints from purely concerning about the internal factors of corporates to
the external macroeconomic environment.
The second main contribution of this paper is that we pertain to provid-
ing an approach to ease financial constraints suffered by Chinese traditional
enterprises. We suggest that it’s beneficial for corporates to concern about the
external business environment, especially the economic cycles. As for the impact
of economic cycles-there are two representative views: the first is that the effect
of the economic cycles on each firm is exactly the same, so changes in the eco-
nomic cycles do not lead to additional losses for individual firm [16]; Another
view is that the profitability of a firm is linked to the economic cycles, and
its investment and business strategies are effected by economic cycles [18]. Our
research confirms that the recession of the economic cycle aggravates corporate
financial constraints. Enterprises should take advance measures to cope with the
lough business environment. Moreover, we study the positive effect of account-
ing conservatism on decreasing corporate financial constraints and make a useful
Economic, Accounting and Financial Constraints 1739
supplement to the economic consequences of accounting conservatism. It will

inspire the traditional enterprises to improve accounting information quality for
decreasing financial constraints.
The remainder of the paper is organized as follows. Section 2 explains the
relevant theoretical analyses and our hypotheses. Section 3 presents our research
sample and model. Section 4 shows our main empirical findings, and Sect. 5
presents our conclusion.
2 Theoretical Analyses and Hypothesis
In the economic recession, reduced external demand leads to production downsiz-

ing, which inevitably damage enterprises profitability. Further, there is a reduc-
tion in internal funds. So, enterprises must rely on more external financing to
meet funding needs. Bernanke and Gertler [6] point out that enterprises suf-
fer higher external financing costs and more serious cash-flow problems during
the downturn. This is mainly caused by the tough external business environ-
ment. Besides that, the deterioration of corporate operating situation leads to
the increase of information asymmetry. At the same time, investors become more
cautious with big confidence attack. The stock market is in depression, so enter-
prises are difficult to raise enough funds by issuing shares. In addition, financial
institutions such as banks may reduce the loan amount and raise the loan interest
rate to control the default risk, increasing the difficulty of obtaining loans from
financial institutions. According to financial accelerator theory, enterprises suffer
more serious financial constraints because the external capital suppliers require
a higher risk premium to compensate for possible loss risk in a recession. On
the contrary, the optimistic expectations help to ease the financial constraints
in the macroeconomic boom. It can be seen that there is a significant negative
correlation between the degree of financial constraints and the economic cycles.
Therefore, this paper gives the first hypothesis:
Hypothesis 1: The economic cycles can influence the degree of corporate financial
constraints. During the boom, the degree of financial constraints is relatively low,
while in the economic recession, the degree of financial constraints is high.
Based on the soft budget constraint theory, the influence of economic cycle
on financial constraints can be different due to property rights. On the one,
SOEs take more political tasks than non-SOEs. The government is more likely
to give stealth or explicit support to help SOEs out of troubles when SOEs
involve in financial risks [21]. On the other, the long-term mutually beneficial
contract between the state-owned banks and the SOEs can alleviate the infor-
mation asymmetry between the two. So SOEs enjoy preferential treatment in
bank loans. The Property rights, to a certain extent, can weaken the influence
of economic cycle on corporate financial constraints. So, this paper gives the
second hypothesis:
Hypothesis 2: Compared with SOEs, non-SOEs suffer more serious financial con-
straints.
1740 H. Wang et al.
Enterprises with poor financial status may carry out earnings management
to meet the financing requirements to get funds. Creditors and shareholders are
in a weak situation with less information and can’t effectively supervise the use
of funds. So enterprises are more prone to moral hazard and adverse selection.
Creditors and shareholders will require a higher risk premium for taking a greater
default risk. Therefore, under the same conditions, the higher degree of informa-
tion asymmetry, the more serious financial constraints enterprises suffer. Some
studies show that accounting conservatism can help to ease the agency conflict.
Beaver and Ryan [5] find that accounting conservatism mitigates agency conflicts
between management and creditors by undervaluing net assets and discouraging
management from transferring creditor interests to shareholders through asset
substitution. Lafond and Roychowdhury [14] argue that management share hold-
ing ratio is negatively correlated with accounting conservatism because account-
ing conservatism eases agency conflicts between management and shareholders
by suppressing earnings management and over-investment. As a signal trans-
mission mechanism, accounting conservatism can effectively reduce the agency
costs and avoid the moral hazard and adverse selection caused by information
asymmetry. It shows that accounting conservatism plays a positive role in easing
financial constraints. Based on this, Hypothesis 3 is proposed.
Hypothesis 3: Under the same conditions, the improvement of accounting con-
servatism can alleviate the corporate financial constraints.
Due to the special national situation of our country, the SOEs still occupy the
dominant position in the market economy. SOEs not only have more resources
and policy advantages, but also enjoy the implicit guarantees from government.
When the SOEs fall into financial troubles, the government is more willing to pro-
vide financial assistance and banks are more willing to provide loans. It reduces
investors’ concern of accounting conservatism, that is, accounting conservatism
has little effect on alleviating the financial constraints of SOEs. On the contrary,
accounting information quality of non-SOEs is one of the most important factors
considered by the external capital suppliers. The conservative accounting policies
can help enterprises obtain financing from banks and investors. In the downturn,
non-SOEs suffer more serious financial constraints, so it is extremely important
for non-SOEs to obtain financing from external markets. Using accounting con-
servatism to show their stable profitability and good operating conditions can
alleviate the financial constraints. Viewed from the bank point, SOEs also have
the credit quotas granted by banks. So the financial constraints suffered by SOEs
aren’t serious even in the downtown and the financial constraints caused by eco-
nomic downturn will be more serious in non-SOEs. Based on the above analysis,
the hypothesis 4 is proposed.
Hypothesis 4: Under the same conditions, the increase of accounting conservatism
weakens the negative influence of economic recession on corporate financial con-
straints, and the interaction is more significant in non-SOEs.
3 Research Sample and Model Design

3.1 Definition of Variables
(1) Financial constraints

Both cash flow sensitivity of investing model and cash flow sensitivity of cash
holding model are widely used in present research about financial constraints.
However, due to the completely opposite conclusions in different studies using
cash flow sensitivity of investing model, this paper uses cash flow sensitivity
of cash holding model to measure financial constraints. And Almeida et al. [1]
also prove that the cash flow sensitivity of cash holding model can overcome the
defect of cash flow sensitivity of investing model.
(2) Accounting conservatism
Accounting conservatism is typically defined in the empirical accounting litera-
ture as the differential verifiability required for recognizing gains as compared to
losses [4,22]. More precisely, conservatism is the practice of reducing earnings in
response to “bad news”, but not increasing earnings in response to “good news”.
Basu [4] uses the positive stock returns to express good news and negative stock
returns to show bad news, and construct the following model to measure enter-
prises’ accounting conservatism.
EPSi,t /Pi,t−1 = α0 + α1 EPSi,t + α2 RETi,t + α3 DRi,t × RETi,t + εi,t , (1)
where i indicates the firm, EPSi,t is earnings, RETi,t is returns, and DRi,t is a
dummy variable that equals 1 if RETi,t is less than 0, and 0 otherwise. So α2 is
the good news timeliness measure and α3 is the incremental timeliness for bad
news over good news, or conservatism.
Khan and Watts [13] improved the Basumodel. They considered firm-year
specific coefficients α2 (timeliness of good news) and α3 (conservatism) can be
expressed by linear functions of firm-year characteristics that are correlated with
the timeliness of good news and conservatism:
G-SCORE = α2 = μ1 + μ2 SIZEit + μ3 LEVit + μ4 M Bit , (2)

C-SCORE = α3 = λ1 + λ2 SIZEi,t + λ3 LEVi,t + λ4 M Bi,t , (3)
where SIZEit is the natural log of the market value, M Bit is the market-to-book
ratio, and LEVit is the debt-to-equity ratio. Replacing α2 and α3 in Eq. (1) by
Eqs. (2) and (3), respectively, yields the following empirical regression model:
EPSi,t /Pi,t1 = α1 +α2 DRi,t + (μ1 + μ2 SIZEi,t + μ3 LEVi,t + μ4 M Bi,t )

× RETi,t + (λ1 + λ2 SIZEi,t + λ3 LEVi,t + λ4 M Bi,t ) (4)
× DRi,t × RETi,t + εi,t
1742 H. Wang et al.
(3) Economic cycle

Domestic scholars use different methods to divide the economic cycle. Jiang and
Liu [12] used GDP growth rate to distinguish different economic cycle. Shi and
Zhang [20] divided the economic cycle based on the relationship between the
real economic growth rate and the potential economic growth rate. In addi-
tion, Dong Jin (2006) [8] used a variety of measurement methods such as linear
trend method, H-P filter method, B-P filter method and the production function
method to divide China’s economic cycle. This paper references the research of
Jiang and Liu [11], etc., and use the annual GDP growth rate as the surrogate
variable of the economic cycle (Clycle GDP). First, GDP growth rate as a contin-
uous variable can effectively avoid the error of using subjective dummy variables;
And the use of continuous variable can also effectively avoid the systemic impact
of institutional change, such as the implementation of new accounting standards,
the split share structure reform.
(4) Property rights
This paper divides enterprises into SOEs and non-SOEs only, and researches the
influence of economic cycle and accounting conservatism on financial constraints
from these two respects. The property rights are defined as dummy variables,
which are judged according to the ultimate controller of the listed company. If
it’s a SOE, it equals 0 and 1 otherwise.
(5) Control variables
Referencing the model of the existing research, this paper selects the operat-
ing cash flow (Cfo), enterprise size (Size), asset liability ratio (Lev), growth
(growth), profitability (Roa), short-term borrowing difference (Debt) as the
control variables. The more sufficient operating cash flow, the more liquidity
the company has, indicating that the company is in a good condition. So the
expected sign is negative. Company size is defined as the natural logarithm of
the total assets at the end of the year. Large scale companies are more resistant
to risks. So, the default risk is smaller and the expected sign is positive. The
higher the debt-to-asset ratio, the weaker the solvency, so the expected sign is
positive. Enterprises will undertake a higher cost of capital if they have a higher
growth and more likely to fall into financial distress. But on the other hand,
they have better development potential and less defaultrisk. Profitability is the
enterprises’ operating efficiency. The better the enterprise’s profitability, the less
likely the default occurs. The sign is expected to be positive.
3.2 Model Design
As for the research model of financial constraints, Almeida’s cash flow sensitivity
of cash holding model has been recognized by many scholars. The model suggests
that if enterprises suffer higher financial constraints, more cash will be held
from the enterprise’s cash flows to prepare for investment opportunities, and the
cash flow sensitivity of the enterprise will be higher. In order to verify the four
hypotheses, this paper builds extension models based on Almeida’s cash flow
sensitivity of cash holding model.
(1) Test the influence of economic cycles on corporate financial constraints
In order to test the influence of the economic cycle on financial constraints
suffered by enterprises, this paper adds a cross-term between the operating cash
flow and the economic cycle to cash flow sensitivity of cash holding model, and
prove whether the economic cycle affect the financial constraints by checking
the coefficients of the cross terms. As for the Hypothesis 2, the group multiple
regression analysis is performed according to the different property rights.
Cash = β0 + β1 Cf o + β2 Cycle GDP + β3 Cycle GDP × Cfo + β4 Size

(5)
+ β5Lev + β6 Roa + β7 Growth + β8 Debt + ε
(2) Test the influence of accounting conservatism on corporate financial con-

straints
In order to test the influence of accounting conservatism on the degree of finan-
cial constraint, this paper adds the cross-term between accounting conservatism
and operating cash flow to cash flow sensitivity of cash holding model, and
confirm the role of accounting conservatism by observing the coefficients of the
cross terms. As for the Hypothesis 3, the group multiple regression analysis is
performed by property rights.
Cash = β0 + β1 Cfo + β2 Conserv + β3 Conserv × Cfo + β4 Size

(6)
+ β5 Lev + β6 Roa + β7 Growth + β8 Debt + ε.
(3) Test the influence of economic cycle and accounting conservatism on corpo-
rate financial constraints
In order to test the interaction of economic cycles and accounting conservatism
on the degree of financial constraints, this paper adds the cross-terms of economic
cycle, operating cash flow and accounting conservatism to cash flow sensitivity of
cash holding model. We observe the coefficients of the three cross-terms to verify
how the financial cycles and the accounting conservatism interact the corporate
financial constraints. And the group multiple regression analysis is performed by
property rights also.
Cash = β0 + β1 Cfo + β2 Conserv + β3 Cycle GDP × Cfo

+ β3 Cycle GDP × Cfo × Conserv + β3 Conserv × Cfo (7)
+ β4 Size + β5 Lev + β6 Roa + β7 Growth + β8 Debt + ε.
3.3 Sample Selection and Data Sources

The sample consists of all A-share listed Chinese manufacturing enterprises
between 2007 and 2015. The sample selection criteria are as follows: (1) Exclud-
ing enterprises with missing data; (2) Excluding companies with negative net
1744 H. Wang et al.
assets; (3) excluding enterprises with uncertain property rights. Finally, 6933
samples are obtained, including 3975 SOEs and 2958 non-SOEs. In addition,
extreme values of 1% and 99% of the relevant variables are Winsorized to elimi-
nate the effects of extreme values. All financial data comes from CSMAR data-
base and we use Stata11.0 to analysis the data (Table 1).
Table 1. Sample distribution
Year 2008 2009 2010 2011 2012 2013 2014 2015 Total
Number 593 626 675 853 989 1066 1066 1065 6933
4 Empirical Results
Table 2 provides descriptive statistics for the variables, while the whole sample is
grouped by property rights. Table 3 reports the descriptive statistics of the vari-
ables in SOEs and non-SOEs. It can be seen that the mean value of Cash is
−0.00048, the minimum value is −0.759 and the maximum value is 0.636, which
indicates that enterprises have large differences in cash holdings; And the mean
value of Cash of non-SOEs is significantly larger than that of SOEs, indicating
that non-SOEs are more cautious than SOEs; The mean of Cfo is 0.0472, the min-
imum value is −4.27 and the maximum is 0.549, indicating that cash flows of dif-
ferent enterprises is also a gap. The mean of Conserve is 0.0397, the minimum is
−0.00223 and the maximum is 0.276, which shows that accounting conservatism
are widely used in Chinese enterprises, but there are big differences among enter-
prises. The mean of Conserving SOEs is 0.0397, and the mean of the non-SOEs
is 0.0476, which indicates that non-SOEs have higher accounting conservatism
than SOEs. The mean of Lev is 0.432, indicating that debt financing is one of the
Table 2. Descriptive statistics of all samples
Variable N Mean Std Max Min Med

Cash 6933 −0.00048 0.0935 0.636 −0.759 0.000434
Cfo 6933 0.0472 0.0919 0.549 −4.27 0.0452
Cycle GDP 6933 0.0841 0.0129 0.104 0.0690 0.0770
Conserv 6933 0.0397 0.0203 0.276 −0.00220 0.0395
Size 6933 21.88 1.194 26.96 17.47 21.74
Lev 6933 0.432 0.213 2.992 0.00708 0.431
Roa 6933 0.0404 0.0873 4.837 −1.292 0.0353
Growth 6933 0.376 9.396 665.5 −1 0.104
Debt 6933 0.0108 0.0785 0.566 −3.04 0.000485
Table 3. Descriptive statistics of variables grouped by property rights
Variable N Mean Std Max Min Med Mean T test

SOEs
Cash 3975 −0.0107 0.102 0.548 −0.732 −0.007 10.6247∗∗∗
Cfo 3975 0.05 0.078 0.549 −0.656 0.0477 −2.9576∗∗∗
Cycle GDP 3975 0.0826 0.0126 0.104 0.069 0.077 11.4272∗∗∗
Conserv 3975 0.0338 0.0189 0.276 −0.0015 0.033 29.7379∗∗∗
Size 3975 21.58 0.983 25.88 17.47 21.49 25.7232∗∗∗
Lev 3975 0.371 0.199 2.992 0.0071 0.362 29.4408∗∗∗
Roa 3975 0.0484 0.0629 0.429 −1.164 0.0442 −8.8354∗∗∗
Growth 3975 0.445 10.94 665.5 −1 0.121 −0.7081∗∗∗
Debt 3975 0.0135 0.0701 0.566 −0.787 0 −3.3414∗∗∗
Non-SOEs
Cash 2958 0.0132 0.0781 0.636 −0.759 0.0088 10.6247∗∗∗
Cfo 2958 0.0434 0.108 0.484 −4.27 0.0403 −2.9576∗∗∗
Cycle GDP 2958 0.0861 0.013 0.104 0.069 0.077 11.4272∗∗∗
Conserv 2958 0.0476 0.0193 0.175 −0.0022 0.0491 29.7379∗∗∗
Size 2958 22.29 1.325 26.96 18.37 22.09 25.7232∗∗∗
Lev 2958 0.514 0.202 1.867 0.0224 0.529 29.4408∗∗∗
Roa 2958 0.0298 0.111 4.837 −1.292 0.0248 −8.8354∗∗∗
Growth 2958 0.283 6.793 363.1 −0.957 0.0831 −0.7081∗∗∗
Debt 2958 0.0071 0.0885 0.351 −3.039 0.0013 −3.3414∗∗∗
t statistics in parentheses, ∗ p< 0.1, ∗∗ p< 0.05, ∗∗∗ p< 0.01
important financing channels. The mean of the growth is 0.376, the minimum is
−1, and the maximum is 665.5. Assets scale and debt level of SOEs are higher
than those of non-SOEs, but the non-SOEs have higher growth and profitability.
4.2 Correlation Analysis
Table 4 reflects the results of the correlation analysis. It can be seen that the rela-
tionship between the dependent variable and independent variables is consistent
with the hypothesis. Cfo is positively correlated with Cash, indicating that
A-listed manufacturing enterprises generally have financial constraints. Cfo is
negatively correlated with Cycle GDP, which indicates that the economic cycle
has obvious inhibitory effect on the external financing ability of enterprises.
Conserv is negatively correlated with Cfo, which indicates that the improve-
ment of accounting conservatism can alleviate the financial constraints suffered
by enterprises; In the following empirical regression analysis, a more comprehen-
sive summary will be made. In addition, all the correlation coefficients are less
than 0.6, indicating that there is no multicollinearity between them.
1746 H. Wang et al.
Table 4. Correlation analysis results of major variables
Variable Cash Cfo Cycle GDP Conserv Size Lev Roa Growth Debt State
Cash 1.000
Cfo 0.135∗∗∗ 1.000
Cycle GDP 0.021∗∗∗ −0.007∗∗∗ 1.000
Conserv 0.164∗∗∗ −0.127∗∗∗ 0.098∗∗∗ 1.000
Size 0.161∗∗∗ 0.081∗∗∗ −0.107∗∗∗ 0.415∗∗∗ 1.000
Lev 0.165∗∗∗ −0.129∗∗∗ 0.096∗∗∗ 0.999∗∗∗ 0.402∗∗∗ 1.000
Roa 0.117∗∗∗ −0.140∗∗∗ 0.053∗∗∗ −0.323∗∗∗ 0.010 −0.319∗∗∗ 1.000
Growth 0.036∗∗∗ −0.006 −0.005 0.024∗ 0.021∗ 0.024
0.010 1.000
Debt 0.070∗∗∗ 0.053∗∗∗ 0.056∗∗∗ 0.124∗∗∗ 0.082∗∗∗ 0.121∗∗∗
−0.330∗∗∗ 0.059∗∗∗ 1.000
State −0.127∗∗∗ 0.036∗∗∗ −0.136∗∗∗ −0.336∗∗∗ −0.295∗∗∗ −0.333∗∗∗ 0.106∗∗∗ 0.009 0.041∗∗∗ 1.000
Table 5. Regression analysis of economic cycles and financial constraints
Independent variable All samples SOEs non-SOEs

Regression model (5)
Cfo 0.692∗∗∗ 0.200∗∗∗ 0.947∗∗∗
(4.412) (0.764) (5.417)
Cycle GDP 0.020∗∗∗ 0.009** 0.025∗∗∗
(6.640) (2.149) (5.986)
Cycle GDP×Cfo −2.253∗∗∗ −1.910∗∗∗ −2.254∗∗∗
(−5.006) (−3.069) (−3.575)
Size 0.003∗∗∗ 0.002 0.007∗∗∗
(2.933) (1.419) (3.684)
Lev 0.076∗∗∗ 0.020∗∗∗ 0.121∗∗∗
(12.554) (2.642) (12.622)
Roa 0.249∗∗∗ 0.215∗∗∗ 0.282∗∗∗
(17.808) (14.116) (9.995)
Growth 0.000* 0.000** 0.000
(1.888) (2.209) (0.857)
Debt 0.139∗∗∗ 0.096∗∗∗ 0.162∗∗∗
(9.171) (5.071) (6.978)
State −0.017∗∗∗
(−7.294)
cons −0.105∗∗∗ −0.043∗ −0.217∗∗∗
(−4.777) (−1.747) (−5.939)
N 6933 2958 3975
r2 0.093 0.078 0.100
r2 a 0.092 0.075 0.098
F 78.804 31.183 54.939
Table 6. Regression analysis of accounting conservatism and financial constraints

Cfo 0.396∗∗∗ 0.280∗∗∗ 0.510∗∗∗
−13.309 −6.297 −12.609
Conserv −3.088** −1.819 −4.774∗∗∗
(−2.431) (−0.929) (−2.880)
Conserv× Cfo −4.458∗∗∗ −2.913∗∗∗ −4.162∗∗∗
(−7.303) (−3.483) (−4.560)
Size 0.003∗∗∗ 0.001 0.009∗∗∗
−2.598 −0.784 −5.004
Lev 0.403∗∗∗ 0.225 0.570∗∗∗
−3.357 −1.218 −3.649
Roa 0.248∗∗∗ 0.252∗∗∗ 0.070∗∗
−16.884 −14.918 −2.379
Growth 0.000* 0.000∗∗ 0
−1.955 −2.169 −0.594
Debt 0.149∗∗∗ 0.100∗∗∗ 0.262∗∗∗
−10.333 −5.446 −11.399
State −0.016∗∗∗
(−7.007)
cons −0.127∗∗∗ −0.053** −0.281∗∗∗
(−5.399) (−1.966) (−7.379)
N 6933 2958 3975
r2 0.125 0.096 0.152
r2 a 0.124 0.093 0.15
F 109.992 39.098 89.005
t statistics in parentheses
∗
p< 0.1, ∗∗ p< 0.05, ∗∗∗ p< 0.01
4.3 Empirical Results and Analysis

(1) Economic cycle and financial constraints
The regression results of the regression model (5) are shown in Table 5. The
regression coefficients of Cfoare positive at the significance level of 1%, whether
in the regression of all samples or the sample of SOEs and non-SOEs, indicating
that financial constraints is a serious problem in both SOEs and non- SOEs. And
the coefficient of Cfo in non-SOEs is 0.947, which is significantly larger than 0.200
in SOEs, indicating that the financial constraints suffered by SOEs are more sig-
nificantly. Then the coefficients of the cycle-GDP ∗ Cfo are analyzed. The coef-
ficients are negative in the three regression at the 0.01significantly level, which
shows that the economic cycle has a s negative correlation with the corporate
1748 H. Wang et al.
Table 7. Regression analysis of economic cycles, accounting conservatism and financial

constraints

Cfo 0.318∗∗∗ 0.162∗∗∗ 0.465∗∗∗
−11.472 −3.823 −12.289
Cycle−GDP 0.021∗∗∗ 0.011** 0.022∗∗∗
−6.642 −2.464 −5.015
Conserv −0.388 −1.511 −1.228
(−0.291) (−0.753) (−0.689)
Conserv× Cfo −1.192∗∗∗ −0.659 −1.713∗∗∗
(−2.845) (−1.101) (−3.037)
Cycle-GDP∗ Cfo∗ Conserv −31.434∗∗∗ −5.870* −35.547∗∗∗
(−4.684) (−0.626) (−3.651)
Size 0 0 0.005∗∗∗
−0.351 −0.194 −2.811
Lev 0.149 0.191 0.239
−1.181 −1.011 −1.429
Roa 0.283∗∗∗ 0.268∗∗∗ 0.109∗∗∗
−19.459 −15.717 −3.676
Growth 0.000** 0.000** 0
−2.02 −2.179 −0.654
Debt 0.155∗∗∗ 0.102∗∗∗ 0.272∗∗∗
−10.773 −5.504 −11.839
State −0.018∗∗∗
(−7.741)
cons −0.075∗∗∗ −0.035 −0.205∗∗∗
(−3.078) (−1.275) (−5.122)
N 6933 2958 3975
r2 0.129 0.095 0.159
r2 a 0.128 0.092 0.157
F 93.165 30.905 75.076
t statistics in parentheses: ∗ p< 0.1, ∗∗ p< 0.05, ∗∗∗ p< 0.01
financial constraints. The first Hypothesis has been confirmed. The coefficient of
Cycle-GDP ∗ Cfo of SOEs is −1.910, while the coefficient of Cycle-GDP ∗ Cfo
of non-SOEs is −2.254, which shows that the financial constraints of non-SOEs
are more affected by the changes in the economic cycle. The Hypothesis 2 has
been confirmed.
(2) Accounting Conservatism and Financing Constraints

The regression results of the regression model 6 are shown in Table 6. It should
be noted that the coefficients of the cross-variable Conserv×Cfo are signifi-
cantly negative in the three regression results, indicating that the increase of
the accounting conservatism can alleviates the financial constraints suffered by
the enterprises. The coefficient in SOEs is −2.913 and the coefficient in non-
SOEs is −4.162, which shows that the degree of accounting conservatism of
non-SOEs has a greater influence on financial constraints, the hypothesis 3 has
been confirmed.
(3) Economic cycles, accounting conservatism and financial constraints
The regression results of model (7) are shown in Table 7. The coefficients of the
cycle-GDP × Cfo × Conserv are significantly negative in the three regression
results, indicating that the improvement of the accounting conservatism weakens
the negative influence of economic cycles. In other words, enterprises with higher
accounting conservatism can get the external financing easier during the down-
turn. And the coefficient of SOEs is −5.870, while the coefficient of non-SOEs is
−35.547, which shows that the positive role of accounting conservatism in eas-
ing financial constraints during the downturn can be more obvious in non-SOEs.
The Hypothesis 4 has been confirmed.
5 Conclusion
This paper examines the interaction of the economic cycle and accounting con-
servatism on corporate financial constraints using the data of A-share listed
manufacturing companies from 2008 to 2015 from the property right perspec-
tive. It is found that the macroeconomic recession aggravates the financing diffi-
culty of enterprises and cause more serious corporate financial constraints. And
this negative influence is more significant on non-SOEs. Besides, the improve-
ment of corporate accounting conservatism can help to alleviate the financial
constraints suffered by enterprises, and this positive influence is more significant
on non-SOEs as well. We test the dynamic response of enterprises in the face of
the economic cycle recession by combining the economic cycle with the micro-
enterprise behavior. The micro-enterprisescan weaken the negative influence of
the economic recession by increasing accounting conservatism, which not only
helps to understand micro-transmission mechanism of economics cycle, but also
helps enterprises to improve cash using efficiency.
The shortcomings of the study: As the national macro-data is annual
announcement, only the annual data of the sample companies can be chosen in
order to match with the macro data. It reduces the length of research period and
may not accurately reflect the influence of macroeconomic fluctuations monthly
on corporate financial constraints. In addition, the accurate measurement model
is the basis of empirical research. Although there are many measurement meth-
ods abroad, there is no conclusive conclusion. Domestic research in this area is
1750 H. Wang et al.
a reference to foreign research. Therefore, the focus of future research is create a

more accurate accounting conservatism measurement method based on China’s
institutional background.
Acknowledgements. This study is founded by Sichuan University (skqy201523) and

Social Science Annual Project of Sichuan Province (SC16TJ011). We are very grateful
for their support.
References
1. Almeida H, Campello M, Weisbach MS (2004) The cash flow sensitivity of cash. J
Finance 59(4):1777–1804
2. Balakrishnan K, Watts R, Zuo L (2016) The effect of accounting conservatism
on corporate investment during the global financial crisis. J Bus Finance Account
43(5–6):513–542
3. Ball R, Shivakumar L (2004) Earnings quality in uk private firms: comparative
loss recognition timeliness. J Account Econ 39(1):83–128
4. Basu S (2013) The conservatism principle and the asymmetric timeliness of earn-
ings. Contemp Account Res 24(1):3–37
5. Beaver WH, Ryan SG (2005) Conditional and unconditional conservatism: concepts
and modeling. Rev Acc Stud 10(2):269–309
6. Bernanke B, Gertler M (1989) Agency costs, net worth, and business fluctuations.
Am Econ Rev 79(79):14–31
7. Claessens S, Tzioumis K (2006) Measuring firms’ access to finance. World Bank
8. Dong J (2006) Measuring China’s business cycles. Econ Res 7:41–48 (in Chinese)
9. Efthyvoulou G, Vahter P (2016) Financial constraints, innovation performance and
sectoral disaggregation. Manchester Sch 84(2):125–158
10. Guariglia A, Yang J (2015) A balancing act: managing financial constraints and
agency costs to minimize investment inefficiency in the Chinese market. J Corp
Finance 36:111–130
11. Jiang GA, Rao P (2011) Macroeconomic policies and corporate behaviorllbroaden
accounting and corporate finance research horizon. Account Res 3:9–18
(in Chinese)
12. Jiang L, Liu X (2011) Economic cycle fluctuations and the cash holding behavior
of listed companies. Account Res 9:40–46 (in Chinese)
13. Khan M, Watts RL (2009) Estimation and empirical properties of a firm-year
measure of accounting conservatism. J Account Econ 48(2–3):132–150
14. Lafond R, Roychowdhury S (2008) Managerial ownership and accounting conser-
vatism. J Account Res 46(1):101–135
15. Lara JMG, Osma BG, Penalva F (2014) Information consequences of accounting
conservatism. Eur Account Rev 23(2):173–198
16. Lucas RE (1978) Asset prices in an exchange economy. Econometrica 46(46):1429–
1445
17. Macve RH (2015) Fair value vs conservatism? aspects of the history ofaccounting,
auditing, business and finance from ancientmesopotamia to modern China. SSRN
Electron J 47(2):124–141
18. Gregory MN (2009) Principle of economics. Tsinghua University Press, Beijing
19. Shen CH, Lin CY (2015) Political connections, financial constraints, and corporate
investment. Rev Quant Finance Account 47(2):1–26
20. Shi X, Zhang S (2010) Behavior of substitution between trade and bank borrowing
through economics cycles: evidence from China. J Manage Sci China 13(12):100–
122 (in Chinese)
21. Sun ZA, Liu F, Li ZC (2005) Market development, government influence and cor-
porate debt maturity structure. Econ Res J 5:52–63
22. Watts RL (2003) Conservatism in accounting part I: explanations and implications.
Soc Sci Electron Publ 17(3):207–221
23. Weiss A (1984) Information imperfection in the capital market and macroeconomic
fluctuation. Am Econ Rev 74(2):194–199
24. Wier HA (2009) Fair value or conservatism: the case of the gold industry. Contemp
Account Res 26(4):1207–1233
Traffic Lights Dynamic Timing Algorithm
Based on Reinforcement Learning
Chenqing Lu1 , Feng Wen1(B) , and Mitsuo Gen2

1
School of Information Science and Engineering, Shenyang Ligong University,
Shenyang, The People’s Republic of China
wenfeng mail@163.com
2
Abstract. Traffic light control is an issue that metropolitan cities are

confronted with because of ever-increasing vehicles growth. A proper tim-
ing strategy can release the congestion level and minimize vehicle delays
better than standard fixed signal timing (FST) strategy. In this paper,
a self-adaptive Q-learning based traffic light green time control strat-
egy (QTGCS) is designed to reduce the congestion. Q-Learning acts as
the learning mechanism for light at each intersection as an independent
agent to release itself by operating traffic light duration dynamically. Q-
Learning is based on discrete representation of state and action spaces. In
order to make QTGCS not independent of the environment and designer
criterion, Fuzzy Q-learning based traffic light green time control strat-
egy (FQTGCS) is proposed by applying the fuzzy logic control to opti-
mize timing strategy based on the number of vehicles receiving from the
adjacent intersections. We show through simulation that our algorithm
effectively over the real road networks.
Keywords: Light control · Reinforcement learning · Q-learning · Fuzzy

logic control
1 Introduction
Traffic congestion has been an urgent problem of metropolitan cities in the world.
It is generally recognized that traffic signal improvements offer the biggest payoff
for reducing congestion and increasing the effective capacity of existing road net-
works, and the adaptive traffic signal control systems hold the most promise for
improvement. Reinforcement learning approach implicitly models the dynamics
of complex systems by learning the control actions and the resulted changes of
traffic flow. Meanwhile, it seeks the (sub)optimal signal plan from the learned
input-output pairs. RL, usually formalized as a framework of Markov decision
process (MDP), assumed that an intersection behaves similar to an intelligent
agent learning to plan green times in each cycle using current traffic information.
Traffic Lights Dynamic Timing Algorithm 1753
Specifically, Q-learning obtains Q-function by learning an action-value function.

Thus, Q-learning is well-suited RL algorithm that obtains an optimal action-
selection policy for given traffic state in any road networks. If traffic light indi-
cates a suitable signal for vehicles, unnecessary waiting time of road users and
the level of congestion state will be reduced compared to standard fixed signal
timing (FST) strategy.
There are lots of researches aimed at reducing the congestion by the road
users through RL methods. The isolated traffic signal was defined as an agent of
RL, agent try to learn relationships between actions and their effect on the
environment [1]. Extensive studies tried to implement RL or integrate with
other algorithms for light timing, such as neuro-fuzzy [8] and kernel method
to approximate Q-function, which is applied to estimate the feature vector [6].
Multi-objective RL proposed for traffic signal control in [7], in this approach, con-
solidating all rewards in one Q-function to obtain a global objective instead a
multi-objective optimization problem. In [5], Q-learning algorithm acts as learn-
ing mechanism that react with environment for light intersections to release itself
and in [4,9] taking advantage of fuzzy logic, the state and actions is set by a
fuzzy algorithm which can be learned from an isolated intersection environment.
In consideration of the correlation with adjacent intersections, there are also few
researches aimed to coordinate light timing strategy together in small area sepa-
rately [2], but how to form specific cooperation region not mentioned especially
in large-scale road networks.
Fuzzy logic has the advantage of interpreting the linguistic values in terms
of logical variables. With the aid of this, the traffic congestion situation can
be clearly classified and comprehended [3]. Most of the fuzzy models regularly
query the traffic conditions in order to decide whether to extend or terminate a
current green phase based on Q value or reward value but not congestion relation
on the adjacent sections. In this paper, we formulate the traffic light timing
problem as an MDP, design a self-adaptive Q-learning based traffic light green
time control strategy(QTGCS). In order to make QTGCS not independent of the
environment and designer criterion, Fuzzy Q-learning based traffic light green
time control strategy (FQTGCS) is proposed based on the number of vehicles
receiving from the adjacent intersections. The indistinguishable boundary can
be used evaluate the state approached next step. In the development process, we
use SUMO simulator, an open-source software to help the real-world traffic light
simulation. We show that FQTGCS yields a policy that has better performance
than both the FST and QTGCS.
2 Traffic Lights Timing Strategy

2.1 Traffic Phase Model
The traffic light infrastructure consists of roads and nodes. A road connects two
nodes, and can have several lanes in each direction. Phase is the permission
which determines the duration of green time for traffic flows in one or some
1754 C. Lu et al.
directions at the intersections. It plays an important role in ITS especially in

light timing method. In this paper, we have defined two phase models: one for
three-arm intersection phase model is shown in Fig. 1.
Fig. 1. Three-arm intersection phase model
Another four-arm intersection phase model is shown in Fig. 2. We can see

easily from the figure that each model including four-phase consist of left-turning,
right-turning, straight-going, and U-turn.
Fig. 2. Four-arm intersection phase model
In Fig. 1, phase-one denotes for the east-west direction, traffic light turns
out to be green at a given time and vehicles get the pass permission, and for
the south direction, the light is red at an three-arm intersection; phase-two is
quite the reverse at that junction, the south direction for green light, and for
the east-west direction, the light is red. In Fig. 2, phase-one denotes light turns
green in the east-west direction and vehicles get the pass permission, and for the
south-north direction, the light is red at a four-arm intersection; phase-two also
denotes at that junction on the opposite way. In a FST strategy, the phases are
set in a round-robin (RR) manner and the duration of green time for each phase
is assigned in advance.
Due to the dynamic and uncertain nature of the traffic environment, we sup-
pose that the phase of release order is changing frequently for light timing. The
study will set phase selection strategy according to vehicles density. If the vehi-
cles density in phase-one direction is higher, then choose phase-one; otherwise
choose phase-two. In such way, which phase obtained the evacuation right can
be easily determined.

phase-one ρphase-one > ρphase-two
phase(t) = (1)
phase-two ρphase-one ≤ ρphase-two ,
where, phase(t) is the phase we have choose by phase selection strategy, ρphase-one
is vehicles density of phase-one direction, ρphase-two is the density of phase-two
direction.
2.2 MDP Framework
An MDP provide a coherent theoretical framework for which action is chosen in

state s at each time step, and finally find a sequential decision by controller. In
traffic light timing, we regard each traffic light as an control agent, and describe
the process as a four-tuple (S, A, P, R), where S and A are the state space and
action space, respectively, and every element s ∈ S is called a state. Every
element a ∈ Ais called an action; P : S × A × S is the transition probability
which is a finite set of probability distributions in which an element p(s, a, s )
denotes the probability given the current state-action pair (s, a) to next state
s ; and R : S × A × S is the reward function, and r(s, a, s ) denotes immediate
control performance received after transitioning from state s to state s , due to
action a, where s,s ∈ S, and a ∈ A. Hence, the whole traffic environment state
is Markov, and the goal is to choose the optimal policy π ∗ that maximizes the
total discounted reward, as:
π ∗ = arg maxV π (s). (2)

π
In Eq. (2), V π (s) is the value function accumulated under the policy, which
defined as follows:

V π (s) =
s ∈S P (s, a, s )[R(s, a, s ) + γV (s )], (3)
where γ ∈ (0, 1) is the discount factor, we use the value function denote the
discounted sum of the rewards.
In reinforcement learning, we can easily formulate light timing strategy
within the MDP framework described, and the light at a signalized intersec-
tion as a learning agent. With this, light agent observe the congestion situation
of the environment around, and take an action which based on exploration strat-
egy, act on the environment and obtain a reward to evaluate the action. The light
agent learning through the trial-error experience to achieve the optimal policy.
2.3 Q-Learning Algorithm
MDPs required having a good model in its implementation. Therefore, in this

research we use Q-learning which from reinforcement learning with a model-free
1756 C. Lu et al.
approach, and the traffic light at junctions as a learning agent. The light agent
are not necessary to know how the environment works. It works by estimating
the values of state-action pair, called Q-value, which represents the maximum
discounted sum of future rewards an agent except to receive if it starts in state
s, choose action a and then continues to follow an optimal policy.
This paper designed a distributed traffic light control strategy based on Q-
learning. In such traffic light timing control model, each traffic light is an agent
in Q-learning, agent chooses green time of traffic light as the action of agent,
the density of vehicles in traffic lane at the intersection as the state, vehicles’
average traveling time as the reward of Q-learning in traffic lane. Learning sys-
tem interact with the environment constantly to get feedback and adjust map
strategy of state to action.
The update of Q-learning formula is as follows:
Q(s, a) = (1 − a)Q(s, a) + a [R(s, a) + γQmax (s , a )] , (4)
where, α is the learning rate, and is the discount factor to decide convergence
rate, R(s, a) in function is the reward incurred by the light agent for taking an
action a when the current state s and end up in state s , Q(s, a) is the sum
of future rewards to evaluate the performance of action we have took. It is an
iterative method that modifies the estimates of Q-value. In the learning process,
according to the Q-learning update rule set so that, the estimates converge to
the optimal Q-value and the light agent’s knowledge tends to be precise.
In this paper, Q-values are updated by the Q-learning update rule using
Boltzmann based exploration strategy. The light agent decides its action based
on the current learnt Q-value, the next section with Boltzmann exploration
strategy that is used by the light agent to update Q-value and subsequently
for choosing actions, the formulate is defined as follows:
Q(s,a)
e −t
p[a |s ] = , (5)
Q(s,a)
a∈A e
−t
where A is action space and τ is temperature parameters, p[a |s ] is probability

that select a action in state s. The states represent the traffic density in the lanes
connected to the junction described in Sect. 2.4, the action which represents the
time for the highest congestion phase get the green light on.
Light timing control algorithm process based on Q-learning is shown below:
Step 1. initialize Q-learning algorithm parameters;

Step 2. analyzing the state of surrounding environment;
Step 3. according to the state, the light agent will select by action selection
strategy;
Step 4. action act on the environment, and make environment changed, mean-
while the light agent obtain the reward according to reward function,
and update the Q-value by this reward;
Step 5. if not satisfy the stop condition, return to step 2.
2.4 Fuzzy Logic Control Optimization Strategy

Consider the incoming traffic flows react on the congestion level of joint junc-
tion, and even if calculate the optimal state-action pair, there must be some
disparities between actual optimal solution. For the consideration above, in this
section, we propose a method to optimize action solution based on fuzzy logic
control strategy. With fuzzy logic strategy, the learning agents coordinate with
joint agents by information sharing, namely, the light agent allot green time
to the phase that maintained the highest congestion level and obtain the sta-
tus of approach phase flows by the vehicle information, these data are used to
decide whether to extend or shrink the current light through-phase to achieve
action solution optimization point. Furthermore, we use the density of vehicles
to measure phase congestion level, and defined model discussed above in Fig. 3.
Fig. 3. Phase flow model
In this method, a fuzzy description of the traffic density during each phase
at the intersection and its joint intersection, and the reward are implemented
respectively, means fuzzify. After that, we develop fuzzy rulebase and initialize
the reward values by the if-then relations.
(1) Fuzzy sets
A fuzzy set is a pair (U, A) where for each element x ∈ U , here U is a set, and A,
a membership function, is the degree of membership of x in (U, A). The closer
A(x) to 1, the higher likelihood of x ∈ A, and the closer A(x) to 0, the lower
would be. We use trigonometric membership function in this paper. The domain
of traffic density during the phase at the intersection is [0, 1], the density level will
be divided into four categories, density level below 0.25, density level between
0.25–0.5, density level 0.5–0.75, and the density level above 0.75, furthermore
the fuzzy sub-sets of traffic density is “zero” (Z), “low” (L), “medium” (M),
“high” (H), the membership function of each sub-sets is shown in Fig. 4:
The domain of the action reward is the real unit interval [−2, 2], the four
fuzzy sub-sets of reward is “zero” (Z), “small (S), “medium” (M), “medium big”
(MB), “big” (B), the membership function of subsets is shown in Fig. 5:
Traffic light optimize action selection strategy according to density situation
at current intersection (CI) and its joint intersection (JI), to maintain synergy
between traffic lights in a small scale, and improve the traffic capacity of the
total traffic road network efficiently.
1758 C. Lu et al.
Fig. 4. Membership function for the phase of vehicle density
Fig. 5. Membership function for the reward
(2) Fuzzy database

In this section, we develop the fuzzy database as follows:
• If the traffic density of the phase at current intersection is M, at the same

time, the approach density of traffic flows at joint intersection is L, then we
can define the action reward is M;
• If the traffic density of the phase at current intersection is H, the approach
density of traffic flows at joint intersection is Z at the same time, then the
action reward is MB.
Thus with linguistic descriptions of traffic density as input, so the phase of

vehicle density is determined by the fuzzy rules given in Table 1 described as:
If CI is T Di and JI is T Dj , then Tgreen is r,

(6)
(i, j = Z, L, M, H; r = Z, S, M, M B, B)
where CI is the current intersection and its phase density T Di , T Dj is the

density approach from the joint intersection JI, r describes the green time to
be extend or shrink. The fuzzy database is defined in Table 1
3 Experiments
In order to verify the validity and correctness of the traffic light control strategy.
In this paper, the vehicle induction system (SVIS) is based on the shortest path
Table 1. Fuzzy database
Action reward Corresponding intersection phase density

Z L M H
Current intersection phase Z Z S M MB
vehicle density
L S S M MB
M S M MB MB
H M MB MB B
algorithm. Simulation and experiment on the road network through the open-
source software SUMO simulator, which is wrote by java, the road network is
a part of the U.S. State of Vermont. In the formula about traffic light control
strategy is based on Q-learning, we set α is 0.7, and γ is 0.9. The road network
is shown in Fig. 6 (Table 2):
Table 2. Road network information
Attribute Quantity
The number of traffic lights (3-way or 4-way intersection) 51
Number of road sections 206
Starting point 8
Terminal point 8
Departure speed (v/h) 1440
Total simulation time 15000
The data of the algorithm is evaluated in this experiment is:
(1) Count the quantity of vehicles in traffic system per 100 s;

(2) Count intersection vehicles average travel time in each phase.
Fig. 6. Road network

1760 C. Lu et al.
In SVIS, in traffic light control strategy based on Q-learning algorithm

(QTGCS), light timing based on action information, and in the fixed-time green
light control strategy (FTGCS), traffic light phase update based on fixed-time
50 s set by system. The evaluation data of SVIS and QTGCS were compared
with the evaluation data obtained by SVIS and FTGCS. The quantity of vehi-
cles in traffic system is shown in Fig. 7, and intersection vehicles average travel
time in the lane is shown in Fig. 8 as follows:
Fig. 7. The quantity of vehicles in traf- Fig. 8. Intersection vehicles average

fic system travel time in the lane
From Figs. 7 and 8, we can observe that the evaluation data obtained in SVIS
and QTGCS collaboration is better than SVIS and FTGCS. The experimental
results show that the FTGCS can improve the efficiency of traffic system and
reduce vehicle travel time, compared with the traditional fixed-time strategy,
based on the QTGCS can using real-time information of the road network, deploy
green light of traffic light reasonably, and shorten vehicle travel time and delay
time.
In SVIS, vehicle path selection based on induction information, and through
the fuzzy logic control according to vehicles information to optimize timing
scheme (FQTGCS), the light timing based on fuzzy optimize action. The eval-
uation data of SVIS and FQTGCS were compared with the evaluation data
obtained by SVIS and QTGCS. The quantity of vehicles in traffic system is
shown in Fig. 9, and vehicle average travel time in the lane is shown in Fig. 10:
Fig. 9. The quantity of vehicles in traf- Fig. 10. Vehicle average travel time in
fic system the lane
From Figs. 9 and 10, we can observe that the evaluation data obtained the
FQTGCS performance better than QTGCS. The experimental results show that
the FQTGCS can improve the efficiency of traffic system and reduce vehicle
travel time, compared with QTGCS. And the FQTGCS can using real-time
information of the road network, deploy green light of traffic light reasonably,
and shorten vehicle travel time and delay time.
4 Conclusion
In this paper, we formulated the traffic light timing strategy as a MDP problem,
and applied Q-learning algorithm with Boltzmann based exploration strategy
and fuzzy logic control strategy. A light learning agent will represent an inter-
section, where an agent can control the traffic light as much as road connected
to the junction. The state will represent the congestion situation around the
junction, it is seen from the traffic density of each phase in the intersection. And
action represent which phase will get the green time to the light. Our algorithms
are adaptive in nature and do not require road network model information, it
can control traffic light in real-time by traffic information. Simulation results
show that the performance of traffic system has been improved by the proposed
traffic light timing strategy. In the current work, we only control traffic light
with the information from joint junction without the light around. Thus, in the
future, one could develop to coordinate the lights connected in a small area, and
we intend to do our assumption in the future.
References
1. Abdulhai B, Pringle R, Karakoulas GJ (2003) Reinforcement learning for true adap-
tive traffic signal control. J Transp Eng 129(3):278–285
2. Arel I, Liu C et al (2010) Reinforcement learning-based multi-agent system for
network traffic signal control. IET Intell Transp Syst 4(2):128–135
3. Azimirad E, Pariz N, Sistani MBN (2010) A novel fuzzy model and control of single
intersection at urban traffic network. IEEE Syst J 4(1):107–111
4. Bi Y, Srinivasan D et al (2014) Type-2 fuzzy multi-intersection traffic signal control
with differential evolution optimization. Expert Syst Appl 41(16):7338–7349
5. Chin YK, Wei YK et al. (2012) Q-learning traffic signal optimization within multiple
intersections traffic network, pp 343–348
6. Chu T, Wang J, Cao J (2015) Kernel-based reinforcement learning for traffic signal
control with adaptive feature selection, pp 1277–1282
7. Khamis MA, Gomaa W (2014) Adaptive multi-objective reinforcement learning with
hybrid exploration for traffic signal control based on cooperative multi-agent frame-
work. Eng Appl Artif Intell 29(3):134–151
8. Lu W, Zhang Y, Xie Y (2011) A multi-agent adaptive traffic signal control system
using swarm intelligence and neuro-fuzzy reinforcement learning. In: Integrated and
sustainable transportation system, pp 233–238
9. Moghaddam MJ, Hosseini M, Safabakhsh R (2015) Traffic light control based on
fuzzy q-leaming. In: Artificial intelligence and signal processing, pp 124–128
Research on the Collaborative Governance
of Innovation Network Based on the Extended
JM Model
Chiyan Zou1(B) , Changyi Zhao2,3 , Tao Wang1,4 , and Xin Gu1

1
Business School of Sichuan University, Chengdu 610064, People’s Republic of China
inzcysist@sina.com
2
Soft Science Institute of Sichuan University,
3
Science Institute of Sichuan University,
4
Institute of Innovation and Entrepreneurship Research,
Abstract. Orderly and efficient operation of the network security inno-

vation, and actively promote the coordination effect of innovation net-
work, is the key to further promote the economic development, is also
an important way to achieve an innovative country. Based on the con-
notation of the collaborative governance, run, and the network struc-
ture of governance logic thinking and further analysis of the preliminary
exploration to construct the basic framework of innovation network col-
laborative governance, in order to solve the existing problems such as
insufficient innovation network effects play.
Keywords: Collaborative innovation · Innovation network · Inter-

subjective relationship · Collaborative governance
1 Introduction
With the development of the economy, the model of single enterprise innovation
gradually evolved into cooperative innovation, further promoting the develop-
ment of innovation network. Collaborative innovation is the key impetus to pro-
mote the development of economy, it is complicated innovate organization. Coop-
erating the innovation network is a process that main bodies take the flowing of
knowledge as the carrier from communication, coordination and cooperation to
collaborative [15]. Therefore, the external formal of collaborative innovation is
the development of innovation network, to realize the collaborative innovation
is the ultimate purpose to form the innovation network.
Through the various innovative elements (micro and macro elements) inno-
vation network use different collaborative ways to achieve nonlinear interaction
Collaborative Governance of Innovation Network Based on the JM Model 1763
and coupling, coming into being effective knowledge flowing and optimized allo-
cation in the self-organizing mechanism, the coupling mechanism, the network
drive mechanism and coordination mechanism, to produce the overall effect is
greater than the sum of the parts. In innovation network resources knowledge
and behavior characteristic of subjects have great difference. There often appears
the phenomenon such as fuzzy relations and negative interaction in the operation
of the network. Perrow [8] proposed in a tight and complex system, the com-
plex and unpredictable interaction between non-standard parts (namely different
standards or specifications or character of the individual) will bring disastrous
threat. Some integrate both parts of the performance equation-process and pro-
ductivity from collaboration [6,7]. Throughout the domestic and international
related researches, these mostly focus on collaborative innovation performance,
operating carrier (mainly is knowledge), running mechanism and the role of sub-
jects. Although some scholars proposed to attaches great importance to the net-
work of relationships and interactions between subjects, but not for an in-depth
studying, and about the process of innovation network operation controlling and
synergies playing is not enough. In order to achieve the goals of innovation net-
work, orderly and efficient operation of network organizations is necessary. How
complicated and challenging collaboration can be, even though it may be needed
now more than ever [2]. So we need to do more. Collaborative governance can
effectively guarantee the network relationship, benign interaction and effective
synergy [21], more suitable for the process controlling of innovation network
operation.
The general governance focuses on network governance in a single point, line
or dimension, ignoring the design and planning of overall network governance.
However, collaborative governance emphasizes the integration of the network
structure, the relationship and interaction [29]. Finally, “synergy” includes mul-
tiple subjects participating and cooperating, and integrity, interactive and sus-
tainability; “Governance” is aimed at the network nodes, the relationship of
network and the overall network, also includes the control of the governance
target and the mechanism. Therefore, collaborative governance can guarantee
network efficient operating. In view of this, this research refers to the thought of
collaborative governance and try to build collaborative governance framework in
innovation network, to better solve the problems in the innovation network oper-
ation, providing a reference for the management and development of innovation
network.
2 The Connotation of the Collaborative Governance and

Implementation
2.1 The Connotation of the Collaborative Governance
Collaborative governance theory comes from the coupling of synergy theory and
governance theory. The collaborative governance study has got much more atten-
tion and is still in hot with different points, including highlight government admin-
istration [13,28], the multi-agent participating, and the multi-level cooperation
1764 C. Zou et al.
governance [17,22]. Recently, studying on politics and environment performance

by collaborative governance has blossomed, which concluded that different cor-
poration tensions came different effective [1,3,26]. The UN commission on global
governance in 1995 stressed that collaborative governance is a continuous process
of reconciling different conflicting stakeholders and taking united action [11]. Jin
[16] analyzed the collaborative governance of sublimation, and pointed out that
in the era of the Web2.0 collaborative governance focus on system innovation,
emphasizing the public power in the participation in the reasonable allocation
between subjects, to break the shackles of the traditional single center model, real-
ize the development in governance. On the basis of previous studies, this paper
argues that “collaborative governance” is the multiple stakeholders in the gover-
nance mechanism establish cooperation, and through the network members mul-
tilateral interacting and coordinating to solve conflicts, control of conflicts and
build cooperation in the advantageous network environment, finally, to guarantee
the orderly operation and realize the goal of innovation network.
2.2 Implementation of Collaborative Governance
A ‘Collaborative governance’ model should include mechanisms for contributing

more systematically, consistently and transparently [23]. About mechanism, stud-
ies include trusting and sharing mechanism, interactive process controlling and
innovation property allocation mechanism [20,27]. Zhang [33] put forward that
the collaborative model emphasizes to the idea of collaboration as the dominant
factor, advocated by the regulation and government guidance to construct multi-
center collaborative governance system; stressed the microeconomic foundation
building, and used cognitive guidance mechanism and the balance of interests dis-
tribution mechanism to fully implement the bodies’ self-management and from
two aspects of process and result to ensure synergy mechanism long-term operat-
ing. Deng et al. [14] considering with the whole network, discussed collaborative
governance mechanisms and analyzed the interaction between these mechanisms.
Han et al. [32] concluded that the governance of innovation network not only need
to pay attention to the subjects’ relationship, but also to the effective guidance and
governance of the communication and interaction between subjects, to provide
scientific and efficient operation, policy, economic and technical environment for
enterprise collaborative innovation network. Based with the previous researches,
this paper holds that the implementation of the collaborative governance sup-
ported and safeguarded by collaborative governance mechanism, through guid-
ing and controlling the relationship between network nodes, the network inter-
action, coordination effect and the interaction of the three tries to achieve self-
management and overall coordination effect optimization.
3 The Logical Framework of Innovation Network

Collaborative Governance Based on Expending JM
Model
The relationships, interaction and collaboration is namely the logic of collabo-
rative governance. Johanson and Mattsson [18] put forward to the relationship
and interaction model (JM model), it organically linked the coupling relation-
ship between network nodes and their interaction behavior. Sun [24] extended
JM model on this base. He thought the relationship and interaction also involves
the logic of network governance, and further to expand JM model from the two-
dimensional structure to three-dimensional structure which also included inter-
active results-the collaborative (governance performance). Extended JM model
provides a new perspective for innovation network governance. For this, we need
to define the connotation of relationships and interactions in innovation network,
as well as the possible synergies.
3.1 The Relationships Between Network Node

Granovetter [12] stressed that in the market, there are Strong relationship
(Strong Ties), Weak relationship (weak Ties) and Free trade (Arms-length Ties)
three forms among subjects, the Strong relationship means the subjects are more
passionate and interactive, have high-integrity, frequent-interaction resources
exchange. However, weak relationship is on the contrary. The degree of embed-
ding in actual innovation network (strength of relationship between different
subjects) and the network density (the ratio between network nodes connection
number and maximum connection number) can influence the flow of knowl-
edge in network, thus influence on the effect of collaborative innovation [30].
Michelfelder and Kratzer [19] analyzed that strong relationship with embedded
characteristics, deepen relations and strengthen interaction mainly through the
necessary commitment and contract. On the one hand, commitment encourage
special assets investment to curb opportunism behavior; through mutual under-
standing, on the other hand, as well as the basis of continuous interaction, to
promote trust based on emotional and relationship regulation emergence, to
enhance internal convergence. Fliervoet et al. [10] analyzed collaborative gover-
nance through social network and proved that the relatively high network density
improves the potential for collective action and collaboration. Strong relationship
emerging based on weak relationship, weak relationship make network maintain
high independence and flexibility, thus easier to expand relations connecting.
Innovation network continuously evolve in such main aspects as knowledge
creation, diffusion and use [9]. Essentially, it is a kind of social relation network.
Network relying on knowledge demands to establish relationship is weak rela-
tionship, which mainly is interests-oriented to make the network operation. With
the interaction is increasingly deep, the cooperation presents diversification, the
role of subjects shows the dynamic and diverse. Finally, weak relationship evolves
to strong relationship to achieve greater knowledge sharing. In terms of knowl-
edge’s acquisition and integration, strong relationship promotes tacit knowledge
1766 C. Zou et al.
flowing, and weak relationship promotes explicit knowledge flowing. Low cost
characteristics of weak ties can absorb new knowledge and resources more, but
generally causing the low-stable dominant network, insufficient knowledge flow-
ing. Strong ties guarantee stable network, the cooperation of high willingness
and sufficient knowledge flowing, strong relationship. Of course, too much strong
ties will restrict the development of the network and the spread of knowledge.
For these, we need to have different priorities. For most of China’s national high-
tech industry base, the high density of innovation network, for example, should
be properly promote weak relationship evolve to the strong relations, designing
and building complete industry chain and technology chain, to achieve higher
conversion rate and win-win among the members. Meanwhile, it should be paid
attention to the weak relationship developing to obtain more knowledge, espe-
cially the heterogeneous knowledge. In this way, in innovation network weak
relationship strengthen the heterogeneous knowledge acquisition, strong rela-
tionships improve the efficiency of knowledge sharing.
3.2 The Interaction in Network
Interaction is another important dimension to collaborate in innovation network.

It is mainly reflected in knowledge sharing and learning, the rational resources’
allocation and its optimization, and so on [4]. Knowledge sharing is one of the
process of knowledge flows, knowledge flow efficiency directly affects the synergy
of innovation network [25]. In innovation network, for the difference of knowledge
quality and quantity among subjects, and the difference of division, the comple-
mentary of knowledge, there is the relationship between knowledge supply and
demand. Interactions between the main body, causing continuous circulating
knowledge in the network between different subjects, to accelerate the process
of collaborative innovation. Knowledge flows is one of the most common form
in innovation activities, also is important activities to bring value-added. There-
fore, this article regards knowledge flows as interaction in innovation network. In
addition to the knowledge sharing, the knowledge flows also include knowledge
transfer and knowledge creation. At the early stage of the innovation network
operation, knowledge sharing is given priority, namely heterogeneity knowledge
exchange among the subjects, so as to realize sharing. Knowledge creation stage
is important process of creating new knowledge. Knowledge flows create knowl-
edge synergy, and bring greater knowledge advantage for the whole network.
Competition and cooperation are the two basic forms of interaction in social
[31]. In network, the cooperative interaction produces synergy effect, encour-
ages members to pursue common interests, and promotes the knowledge spread
effectively in network; benign competition interaction creates a new combination
of the knowledge and value, through mutual learning and using knowledge, to
reduce opportunism; Malignant competition interaction, however, causes contra-
dictions and conflicts. Finally, hinder the relations developing, further hinder the
synergistic effect. The two former increase the degree of familiarity and under-
standing between subject, reduce the communication cost, and avoid information
asymmetry, low knowledge matching degree and the bridging degree for unfa-
miliar subjects, especially in knowledge.
Interaction of collaborative governance logic stresses the main body of gov-
ernance is not a single subject, but all the members in network. They work
together to promote the development of the network. Based on stakeholder the-
ory [5], multiple participation not only promotes interaction, but also control
and guide malignant it. In this way, it deepens understanding and trust among
subjects and improves the efficiency of knowledge and resources flowing. Such
as regional economic competition forces the Yangtze river delta to seek deep
interactions with surrounding areas and other economic belt, and according to
the geographical relations and the level of economic set up different levels of
city circles. At the same time, improving city circle cooperation interaction, and
promoting the development of regional economy.
3.3 The Result of Relationship and Interaction-Collaboration
In the process of interaction, new weak relationship emerges, further to evaluate

strong relationship. Similarly, the benign interaction generates positive synergy,
and further to strengthen the relationship. The strong and the weak relationship
between network nodes, the benign interaction and positive synergy is in favor of
achieving the goal together. To be sure, because of the differences of resources and
capabilities between subjects directly leads to the result of interactive complex
and uncertain and the evolution of and relationship, the collaboration created
by relationship and interaction is diversity, dynamism and directivity. The col-
lective rationality of repeated game formed the positive collaboration improved
by Pareto, which is the balance of the relationship and benign interaction can
generate “social facilitation effect”, and add knowledge value, optimize innova-
tion network environment and realize the goal of innovation network. On the
contrary, due to the unbalance relationship and vicious interaction, there comes
negative synergy, and leads to “social-loafing”. Those cause the failure of innova-
tion network targets with the subjects’ negative behavior, deceptive interaction,
vices and low negative effects.
In innovation network, the demand of knowledge causes to produce rela-
tionship, relationship promotes interaction, interaction promotes synergy, in
Fig. 1. The collaboration governance logical framework based on expending JM model

1768 C. Zou et al.
turn, synergy impels interaction and make organizational relationships deep.

The whole process accelerates knowledge flow, promotes organizational power.
Further, Organization network is strengthened. Those promote more subjects
to participate in innovation, finally expanding the source of knowledge and the
synergy beneficiary, also improving the efficiency of innovation network. New
knowledge demands further to build new relationships and lead to a deeper inter-
action. It is a constant, circulating and spiral-rising process, as shown in Fig. 1,
constituting a logic framework of innovation network collaborative governance.
4 Safeguard Mechanism of Collaborative Governance

The another key to success collaborative governance innovation network, is col-
laborative governance mechanism can forward drive relationship, interaction and
coordination, ensure synchronous interaction and the orderly and efficient col-
laboration. as shown in Fig. 2, is the basic framework of innovation network
collaborative governance.
Fig. 2. The basic framework of innovation network collaborative governance
First, the configuration of power and responsibility, relational contract, the

trust mechanism, constraint mechanism and incentive mechanism promote weak
ties form, further to the strong relationship evolution, meanwhile, to control these
two relationships, to make up the malpractice and the insufficiency, to take their
advantages, to ensure maximum effect. Accrual configuration make the gov-
ernment, enterprises, universities, research institutes and intermediary service
organizations these subjects fully play function advantages, fully coordinate. At
the same time, overcome the limitations of self-regulation and government sin-
gle coordination. Relational contract and trust mechanism prompt relationship
formation and strengthen, so as to guarantee the stability of the network, to
promote knowledge and resources fully flow and exploit. Among them, the rela-
tional contract can product weak relations, to maintain the network openness
and the heterogeneity of knowledge, to develop the network value. Trust must
be embedded network, could it exist. To product and strengthen the relationship
can establish and strengthen trust; On the contrary, the trust is the necessary
condition to build strong relationships. On the one hand, Constraint mechanism
and incentive mechanism is in favor of reducing the cost of strong relationship,
and the contradiction caused by complicated relationship, On the other hand,
to encourage the network extension, improve network activity.
Second, establishing sharing mechanism for knowledge and other resources,
interest distribution mechanism and risk mechanism to avoid opportunism. Being
satisfied subjects with interests’ needs, and promoting cooperation and benign
interaction, promoting the knowledge’s efficient flowing in the network. As well
as principal partners in collaborative innovation network, subjects also is com-
petitors, benefit disputes and conflicts among them will affect the overall effect of
the collaborative governance. Meanwhile, their some characteristics can affect the
formation and development of innovation network collaborative governance. As a
result, interest distribution mechanism and risk mechanism reasonable guarantee
interest allocation effective and fair. The conclusion of the agricultural model,
when knowledge’s gap is very big, strong party interaction’s will is insufficient,
combined with the weak absorption ability is limited, leading to a lower efficiency
of interaction. Conversely, when the gap is small, the effective is better. Sharing
mechanism which can reduce the knowledge’s gap, and can avoid the information
asymmetry. In the process of knowledge flows, involving the subjects learn and
internalized knowledge. Therefore, organizational learning mechanism needs to
further strengthen, to reduce the knowledge differences between subjects, so as
to promote knowledge flow.
Third, through establishing supervision mechanism and evaluate mechanism,
evaluating subjects’ execution and the completion of target. Due to the diversifi-
cation of innovation network, supervision mechanism is hard to effective operate.
For effective supervising the whole process of collaborative governance, firstly,
Institutionalized problems such as the organization’s powers and the constitution
and the profit distribution, and implement the innovation network. Increasing
default cost and effectively reducing the coordination costs, to constraints sub-
jects’ behavior with institutional in collaborative governance. Second, strengthen
moral construction, forming public opinion and supervision mechanism. Super-
vising subjects’ behavior in the process of collaborative governance with moral.
Through the evaluate mechanism to clear synergy governance effect, clear and
control each problem source in every stage and every node.
Of course, the cooperative governance mechanism is not only for one link
in the relationship, interaction and synergy, but on the whole process gover-
nance. Constraint mechanism, for example, can also constraint subjects’ behav-
ior in the collaborative governance, control interaction. Supervision mechanism
and evaluating mechanism is also valid for each link and each movement. And
the characteristic of innovation network and collaborative innovation represents
the different phase characteristics, cooperative governance mechanism should
also be adjusted. In the process of the collaborative governance, collaborative
1770 C. Zou et al.
governance logic is the basis of collaborative governance, also is the junction

between innovation network and collaborative governance. Collaboration gov-
ernance mechanism is support and assurance of the collaborative governance.
All of them mutually dependent and interacting, forming the prerequisite and
guarantee of innovation network collaborative governance, driving the network
orderly and efficiently operating, finally to achieve maximum synergy innovation
effect.
5 Conclusion
This paper mainly discusses how to solve the problems in the operation of inno-
vation network through cooperative governance, according to the extension of
JM model further analysis of the collaborative governance logic, on this basis to
analyze the innovation network of cooperative governance, and put forward the
safeguard measures of innovation network collaborative governance, so as to sup-
port and guarantee for the operation of the collaborative governance, to promote
innovation network synergy. Effective cooperative governance mechanism is able
to efficiently governance the relationship, interaction and synergy in innovation
network. Finally, to solve the relationship which is not clear and soured between
the main body in innovation network, network is running off track, collaborative
innovation is unsustainable, coordination is inefficient and so on. In the future,
further research on collaborative governance mechanism and the strength of the
control network relations, and the definition and controlling for the relationship
intensity in governance model etc.
Acknowledgements. This paper is supported by the National Natural Science Foun-

dation of China the Research on Knowledge Chain Governance Based on the evolution
of inter organizational relationships: (No. 005020540109), and Study on informal gov-
ernance of enterprise innovation network (No. 71602132), and the Central University
Foundation of Sichuan University: the Research on the informal governance of Chinese
enterprises’ collaborative innovation network under the new normal condition (No.
kzx2015-sb49).
References
1. Ansell C, Torfing J (2015) How does collaborative governance scale? Policy Polit
43(3):315–329
2. Bryson JM, Crosby BC, Stone MM (2015) Designing and implementing cross-sector
collaborations: needed and challenging. Public Adm Rev 75(5):647–663
3. Charron N, Dijkstra L, Lapuente V (2014) Regional governance matters: quality
of government within european union member states. Reg Stud 48(1):68–90
4. Chen J, Yang YJ (2012) Theoretical basis and content for collaborative innovation.
Stud Sci Sci 62:270–277
5. Dugger WM (1996) The mechanisms of governance. J Econ Issues 44(4):261–281
6. Emerson K, Nabatchi T (2015) Collaborative governance regimes
7. Emerson K, Nabatchi T, Balogh S (2012) An integrative framework for collabora-
tive governance. J Public Adm Res Theor 22(1):1–29
8. Essay (1979) Complex organizations: a critical essay. McGraw-Hill companies,

Incorporated
9. Etzkowitz H, Leydesdorff LA (2001) Universities and the global knowledge econ-
omy: a triple helix of university-industry-government relations. Pinter
10. Fliervoet JM, Geerling GW et al (2016) Analyzing collaborative governance
through social network analysis: a case study of river management along the waal
river in the netherlands. Environ Manage 57(2):1–13
11. Commission on Global Governance (1995) Our global neighbourhood: the report
of the commission on global governance. Oxford University Press, Oxford
12. Granovetter M (2005) The impact of social structure on economic outcomes. J
Econ Perspect 19(1):33–50
13. Halachmi A (2005) Governance and risk management: challenges and public pro-
ductivity. Int J Public Sect Manage 18(4):300–317
14. Hong Q, Zhang H, Business SO (2016) The structure characteristic and self-
organizing evolution of collaborative knowledge innovation network. J Huaiyin Inst
Technol 317:41–49
15. Jin C (2011) Collaborative innovation and construction of national scientific
research ability. Sci Manage Res 12:1762–1763
16. Jin T (2016) Collaborative governance innovation web2.0 era: the system of gov-
ernance. Academics 324:5–14
17. Agranoff R, Mcguire M, Joaquin E (2005) Collaborative public management: new
strategies for local governments. J Federalism 35(4):637–639
18. Johanson J, Mattsson LG (1987) Interorganizational relations in industrial sys-
tems: a network approach compared with the transaction cost approach. Int Stud
Manage Organ 17(1):34–48
19. Michelfelder I, Kratzer J (2013) Why and how combining strong and weak ties
within a single interorganizational R&D collaboration outperforms other collabo-
ration structures. J Prod Innov Manage 30(6):1159–1177
20. Panico C (2011) On the contractual governance of research collaborations: allocat-
ing control and intellectual property rights in the shadow of potential termination.
Res Policy 40(10):1403–1411
21. Peng X (2010) Three-dimensional governance, relationship management, net-
work management and knowledge management, knowledge partner relationships
between network organization governance research. J Libr Intell 6:121–126
22. Powell WW (1996) Interorganizational collaboration and the locus of innovation:
networks of learning in biotechnology. Adm Sci Q 41(1):116–145
23. Prideaux M (2016) Considering collaborative governance: Nongovernmental organ-
isations, the convention on migratory species and the convention on wetlands of
international importance. Wyoming geological association
24. Sun G (2013) The relationship, interaction and collaboration, the logical of
network-organization. China Ind Econ 11:14–20
25. Tu ZZ, Gu X (2013) Study on process of industry-university-research institute
collaborative innovation based on knowledge flow. Stud Sci Sci 39:182–193
26. Ulibarri N (2015) Tracing process to performance of collaborative governance: a
comparative case study of federal hydropower licensing. Policy Stud J 43(2):283–
308
27. Uzzi B (1997) Errata: social structure and competition in interfirm networks: the
paradox of embeddedness. Adm Sci Q 42(2):35–67
28. Warren ME (2011) Foundations and frontiers of deliberative governance. Oxford
University Press, Oxford
1772 C. Zou et al.
29. Wu C, Zhuang Y (2013) Cooperative governance, the key variables, influencing

factors and implementation ways. J Theor Explor 3:73–77
30. Yubing HE, Zhang Y (2015) Research on coupling of network types and the collab-
oration modes of industry-university-research institution collaborative innovation.
Sci Sci Manage S&T 9:79–85
31. Zeng Z, Tu Z (1993) The heart of the collision, ethics-sociology of the virtual and
real. Hunan publishing, Changsha
32. Zhang S, Yuanjia HU et al (2011) Collaborative innovation model of traditional
Chinese medicinal pharmaceutical enterprises. Sci Technol Manage Res 27:42–58
33. Zhang Z (2015) On the generation logic and the construction path of collaborative
governance. Chinese public administration
Performance Evaluation of Housing Price
Regulation Policy in China: Based on ARIMA
Model and Intervention Analysis
Chang Liu1 , Yixiao Zhou2 , Wei Zhao3 , Qiang Jiang3(B) , Xuedong Liang3 ,
Hao Li3 , Hua Huang3 , and Shucen Fan4
1
School of Securities and Futures, Southwestern University of Finance
and Economics, Chengdu 610052, People’s Republic of China
2
School of Economics and Finance, Deakin Univesity, Perth, Australia
3
jiang.qiang@outlook.com
4
Department of Economics and Management,
Sichuan Technology and Business University,
Abstract. Chinese housing market is rapidly growing up. However, the

study on housing market regulation is inadequate. In this paper, housing
price indexes from 2006 to the present are analyzed based on the ARIMA
model and intervention analysis, and the change of this index can reflect
the effective date and valid date of country’s housing price regulation
policy. This paper established systematically the quantitative analysis
model, studied the timing effects of new regulations, and improved the
quality of the house price index by using the monthly data. Based on the
model results, we evaluated the performance of the regulation policy of
housing price on three criteria including anticipation, conductivity and
additivity. This can provide the field evidence for policy makers to design
better housing policy.
Keywords: ARIMA model · Intervention analysis · Quantitative analy-

sis · Housing price regulation policy · Performance evaluation
1 Introduction
Since the 30 years of reform and opening up, China’s real estate industry grad-
ually changed from the planned economy system to market economic system,
and real estate and related industries have been the most important industry in
China’s regional social economy and played an important role in the national
macroeconomic development. However, along with the gradual progress of the
marketization process of the housing market, we are facing some unexpected
1774 C. Liu et al.
negative effects of China’s housing market including high housing prices and low
housing affordability, housing prices increased too fast, the fever of land specu-
lation. These negative effects demonstrated that the housing market operation
mechanism has some limits. The direct consequence is that houses have become
unaffordable for the low-income groups being not satisfied and housing structure
becoming unbalanced.
To solve the current problems in China’s housing market, the government
introduced a series of monetary policy and fiscal policy. The policy makers tried
to change the performances of banks and real estate developers in the market and
cool the overheated housing market by standardizing housing market laws and
regulations. Especially after 2008, the rapidly rising housing price has become
the focus of public opinion, how to keep housing price away from rapid increases
has become a vital problem which the government needs to solve urgently. Under
this background, a series of price regulation policy were introduced at the historic
moment.
Since 2005, housing price regulation policy which issued by Chinese gov-
ernment can be roughly divided into two categories: First, policy instruments
are design to suppress housing prices, such as issuing purchase orders, increasing
bank interest rates, raising housing taxes, raising the down-payment ratio of real
estate and so on. Second, policy instruments are design to stimulate the housing
prices, such as: tax reductions, lower down-payments, lower interest rates, and
so on.
Since 2005, little evidence has been found in the efficiency of the existing
housing policy. As the housing market policy failure in 2005, China’s housing
market continued to deteriorate and housing prices in many cities increased
sharply in 2006. From 2005 to 2007, the national housing policy focused on
increasing bank interest rates and raising tax rates. For example, in May 2006,
the State Administration of Taxation issued a notice on strengthening housing
business tax levy management. This new tax regulation required that after June
1st, individuals housing purchased less than 5 years shall paid the business taxes
at the full rate; In 2007, the People’s bank of China increased interest rates by
6 times and the deposit reserve rate by 10 times, it is still failed to control rapid
increased property prices.
In 2008, to cope with the financial crisis, the Chinese government decided to
stimulate China’s housing market rather than suppress it. Therefore, the gov-
ernment launched a series of new policy to stimulate housing prices. These new
policies included the temporary exemption of stamp duty on personal sales or
purchase housing; the temporary exemption of land value-added tax on personal
sales housing; local government can introduce their own policies to encourage
housing consumption tax relief and so on. In 2009, the housing price regu-
lation policy was mainly on the inheritance, strengthening and deepening of
2008 real estate regulation policy, while highlighting the focus of strengthening
the housing security. Therefore, the early 2009 housing price regulation policy
mainly includes several aspects like tax, financial market and housing security.
In 2009, while most of the industry was facing a financial crisis, China’s real
Housing Price Regulation Policy in China 1775
estate industry had achieved a substantial development. However, these stimu-

lation policies since 2008 left a hidden danger to the subsequent regulation. After
2010, the government began to realize that the real estate market may be out
of control, so it was very difficult to curb its rising momentum again. In theory,
if a series of policies issued by the government can be effective, China’s housing
prices may change to the expected direction as the policy makers’ design. How-
ever, policy effects are not obvious in the China’s housing market performance.
In this paper, by using the ARIMA model and intervention analysis, we
analyze the effectiveness of price regulation in China. So researchers and policy
makers can understand the critical issues in China’s housing policies including
the effectiveness, failure, superposition and neutralization of housing policies. In
general, this study can provide policy recommendations for the design of a more
effective housing price regulation.
This paper has been organized in the following way. Section 2 begins by laying
out the theoretical dimensions of the ARIMA models and previous research in
this field, and looks at how ARIMA models can be applied in the China’s housing
policy research. In Sect. 3, we conduct the empirical model by exploring the
ARIMA model and establishing the intervention analysis model to see the results
of the housing price regulation. In Sect. 4, we applied the ARIMA model and
intervention analysis into China’ housing market. Section 5 presents the findings
of the research, focusing on the three key themes including the effectiveness of
policy, the validity of policy, the neutralization and superposition of policy. In
Sect. 6, we present the conclusions about the effectiveness of policies, the validity
of housing policies and the superposition between the policies.
2 Literature Review
There are very rich economic research results about national regulation policy
effect on housing prices. The existing studies mainly focus on four aspects:
First, it is the analysis of several factors affecting the housing prices to discuss
the effect of regulation policies, Such as Wang and Chen [14] in Wuhan City as
an example, the author thought that four indicators: sales accounted for below
90 square meters, the new supply area, the money supply (M2), the lending
rate, put a significant impact on Wuhan housing prices. This study focuses on
the qualitative analysis of the regulatory policy and its effect, but it is lack of
accurate interpretation to the effect of regulation policy.
Second, it is the single regulation policy effects on the real estate prices, for
example, monetary policy, land policy and tax policy. Nie and Liu [11] thought
that China’s monetary policy can affect real estate prices, and the money supply
was more significant than the interest rate policy regulation function by analyz-
ing the monetary policy from 2005 to 1994. At present, most of the domestic
literatures focus on qualitative analysis to study the effect of the policy on the
regulation of real estate market, and most research results show that monetary
policy has a significant impact on housing prices. This study did not consider the
regulation effects of a number of policies under the joint action comprehensively,
and be lack of a comprehensive research on overall coordination and interaction.
1776 C. Liu et al.
Third, it is to use a social science or econometric model to analyze, such as Wu

[17] based on PSR model to establish housing price regulation policy effectiveness
evaluation system, after the solution of pressure, state and response index weight
system, through the analysis of the relationship between the three indexes, found
that housing price regulation policies in recent years in 2005–2006, and 2011–
2012 were relatively effective, proposed when country introduced a more severe
structural adjustment policy can make effective control on housing prices.
Fourth, researchers are interesting in evaluating the policy performance. Chen
and Fang [3] explored VECM and DSGE model and found that policies on
restricting house purchase are mainly directed at speculative demands so that
they are suitable for house price control of first-tier cities, while the down pay-
ment ratio policy on mortgage is directed at rigid demands and they are fit for
non-first-tier cities. Chen W [2] applied a multiple-factor panel data model to
show that a high degree of market failure is associated with a high ratio of per-
sistent components in the gap between price and equilibrium. Chen found the
housing policies did improve the housing market’s efficiency. Feng and Wu [8]
studied whether there is an asset bubble in China’s residential housing market by
using a an equilibrium asset-pricing approach. They did not found the existence
of a house price bubble at the national level.
Fifth, non-economic factors such as education showed more impacts on hous-
ing prices. Agarwal et al. [1] studied the relation between school allocation rules
and housing prices in Singapore by using a quasi-experiment. They found that
private housing prices within 1-km zone and in 1-km to 2-km zone from the old
school zone decline by 2.9% and 6.0%, respectively, 6 months before the school
relocation events. Feng and Lu [7] studied if school quality will affect housing
prices in Sanghai, China. They found that a good school may increase housing
prices by 17.1%. Wen et al. [15] reported the similar results by using the data
from Hangzhou, China. The found that elementary and junior high schools have
a significant effect on local housing prices. Furthermore, they reported that the
housing price increases by 2.737% or 0.904% when the house is located less than
1 Km from the high school or college.
ARIMA has been an effective approach in the housing market study. Woo and
Sung-Suk [16] found ARIMA is a useful in exploring the relation between housing
prices and interest rates. Previous studies used logistic model [5,10], GARCH
model [4,21], canonical correlation analysis and other methods to analyze the
regulation effect of comprehensive policy, mainly through analyzing the change of
housing prices, land prices and other major indicators to judge the effectiveness
of regulation policy and evaluate the regulation effect of comprehensive policy.
Although the research analyzed the overall effect of the real estate regulation
policy, there was no analysis of the mutual influence between different time point
and type policies.
In recent years, there is a growing research interest from the management
science field. In the agent theory, Zhang et al. [19] studied Beijing’s housing
market by using a multi-agent model. Ge [9] applied another agent model in
American housing market. Neural Networks are another popular approach. Lim
et al. [13] applied Neural Networks in housing prices predications. Other studies
included index theories [6,12,18,20].
In summary, the current study mainly focused on the qualitative study of
single and comprehensive regulation policy effect, and there is less literature to
use quantitative analysis method to study the regulation policies the superpo-
sition and neutralizing effect between regulation policies. Therefore, this article
adopts the monthly price indexes between January 2006 and February 2015 to
establish housing price regulation policy effectiveness evaluation system based
on the ARIMA model and intervention analysis model. Furthermore, this study
analyzes the influence of policy factors on the real estate prices empirically.
According to the results of the model, we conducted a quantitative performance
evaluation on house price regulation policy by years. The quantitative study
focuses on three key themes including analyzing the anticipation, conductivity
of housing price regulation policy implementation effect in recent years, as well
as the mutual influence between different time point housing price regulation
policy implementation effect, their hedge and superposition, and reasonable pre-
dictions of the effects of policies. These results can provide a scientific basis for
the formulation of the new regulation policies.
3 Theoretical Basis
3.1 ARIMA Model

The basic idea of the ARIMA model is to regard the data sequence as a ran-
dom sequence which is formed by the time lapse of the forecast object, which
uses a certain mathematical model to describe the sequence. Once recognized,
the model can predict the future value from the past and present value of the
time series. The ARlMA model in economic forecasting process not only takes
into account the dependence of economic phenomenon in the time series, but
also considers the interference of the stochastic volatility, which has a higher
accurate rate to forecast the economic operation of short-term trend and is one
of the widely used methods. Because the economic data is often autocorrelation
and non-stationary time series, and the ARIMA model can effectively deal with
autocorrelation and non-stationary data, while the ARIMA model is stable that
can effectively combine with the intervention model, the ARIMA model is used
to predict the housing price index of our country.
Between January 2006 and February 2015, the monthly housing price index
can be seen as a random time series which is formed over time. By analyzing
the randomness, stationary and seasonal factors of the housing price index of
this time series, using ARIMA mathematical model to describe the correlation
or dependence between the values of these single month housing price index.
Thus taking advantage of the past and present housing price index to predict
the future index, it can achieve the purpose of predicting the influence of housing
price regulation policy on the future price.
1778 C. Liu et al.
3.2 Intervention Analysis
Time series are often influenced by special events and situations such as external
interventions. Intervention refers to the stand or fall of forecasting model fitting
degree, that is the pros or cons of fitting degree between the simulated value
generated by the prediction model and the historical actual value.
(1) Intervention variables’ form
The basic variables of intervention analysis model are intervening variables.
There are two kinds of common interference variables: the first is the sustainable
intervening variables, which indicates that there have been affected when T
moment happened, then the step function representation can be used; the second
is transient interference variables, which expresses that it can been affected just
for the moment when it occurs at a certain moment, using unit impulse function.
This article not only studied the regulation effect of housing policy at a certain
time, but also analyzed the impact of these policies on the future housing prices,
so we chose the first kind of intervention variables’ form, as Eq. (1):

0, before intervention events (t < T )
StT = (1)
1, after intervention events (t < T ).
(2) Intervention events’ form

In the process of policy implementation, the intervention events may have
an impact on the regulation effect. The influence of this kind of intervention
events may suddenly start, and last for a long time. Set intervention’s influence
on the dependent variable being fixed, starting from a time T , but the extent of
impact is unknown, that is, the size of the dependent variable is unknown. The
intervention model of this effect can be written as in Eq. (2):
Zt = ωStT (2)
It is the unknown parameter of the intervention impact strength. While the

rough stability series can be into stationary series by differencing, the interven-
tion model can be adjusted in Eq. (3):
(1 − B)Zt = ωStT , (3)
where: B is a backward shift operator. If the intervention event has to lag a

number of periods before impact, such as b periods, then the intervention model
can be further adjusted in Eq. (4):
Zt = ωB b StT (4)
where: StT = 1,
Zt (1 − δB) = ω, (5)
Zt − Zt−1 δ = ω. (6)
By deformation
ω
Zt = , (7)
1 − δB
where, B is a backward shift operator. According to the forecast results, if the
fit of the forecasting model is higher, and the forecast value is also in line with
the actual situation, then the prediction model has a certain application value.
On the contrary, the prediction model is invalid. So on the basis of the ARIMA
model, it is necessary to carry out the intervention analysis. The purpose of
our intervention analysis is to assess the specific impact of the other policy
intervention or emergency on the economic environment and economic process
by using the quantitative analysis.
4 Empirical Study
4.1 Evaluation Index Selection
This paper investigates how the housing market response to the change of the
housing price regulation. Thus housing price index is the most direct indicators
to reflect the housing price changes. Housing price index is the advantage of
“homogeneous”, which can exclude the price fluctuations which due to market
supply and demand relations and other reasons after the effect of these factors:
housing quality, construction structure, geographical location and others. The
monthly housing price indexes of 2006 to 2010 can be achieved in the iFinD
software directly, while the 2011 to February 2015 index can be calculated by
averaging monthly housing price indexes of 70 large and medium-sized cities
nationwide. Thus the random time series of housing price index is obtained.
4.2 Evaluation Index Structure

(1) Establishing a single variable ARIMA model
The following figure is a China commodity house actual price index time
curve graph (sampling frequency is once a month) from January 2006 to February
2015. It is sharply higher in 2007, 2010 and 2014, and housing price index fell to
trough in 2008 (Fig. 1).
Fig. 1. Real estate price index

1780 C. Liu et al.
Observing the housing price index time curve in Fig. 1, we find that the data
is not stable and contain a clear trend. Data is changed into a stationary random
sequence by first order difference, as shown in Fig. 2:
Fig. 2. Sales price index after first - order difference
After the normalization of the original data and repeated recognition of

ARIMA model, we determine p = 1, d = 1, q = 1. Carrying out parame-
ter estimation for ARIMA (1, 1, 1) model, we establish the univariate ARIMA
model:
Xt = 0.0190 + 0.8370Xt−1 + εt 0.2787εt−1 , Xt is difference, ε is residuals.
Predicting the residual, the residual series time curve of the model is not a
trend term, similar to random white noise, so it pass the model test (Fig. 3).
(2) Establishing intervention analysis model
We use the ARIMA (1, 1, 1) model to take extrapolating forecast for housing
price index from January 2006 to February 2015, then use the actual value
to minus the predicted value, the difference from which is the impact utility
generated by macroeconomic regulation policy, remember to Zt .
According to the form of selected intervention events, instability can be into
a stationary series by differencing, and using housing price index from 2006
January to 2015 February, ARIMA predictive value and economic policy conflict
utility, we can estimate the parameter of intervention model (Fig. 4).
By regression in Eq. (8),
Zt = 0.8357Zt−1 − 0.0044, (8)
where, Zt is the difference between the actual values and predicted values.
Namely in Eq. (9),
0.0044
Zt = . (9)
1 − 0.8357B
Through the t-test of the regression value of δ, it is found that the estimated
parameter δ is significant. So after the intervention analysis, the ARIMA model
is in Eq. (10):
0.0044
Xt = 0.0190 + 0.8370Xt−1 + εt − 0.2787εt−1 − . (10)
1 − 0.8357B
Fig. 3. Model testing
Fig. 4. Residual regression model

The study shows that the government’s housing price regulation policy is not
always effective, even if it is a utility policy, it may not take effect in time.
We also found that the different policies will affect each other. The following
paragraphs will analyze the effectiveness of policy, the validity of policy and the
superposition between the policies.
5.1 Effectiveness of Policy
By analyzing the housing price regulation policy since 2006, we can find that
there are three main kinds of policies which have a timely effect. First is involved
in the adjustment of the relevant taxes. Such as in May 2006, the state promul-
gated the country of six, and the State Administration of Taxation issued the
notice of the relevant issues on strengthening the management of housing busi-
ness tax collection. This policy promulgation is St2. Taking it into validity test,
we can see that the next month policy effectiveness is significant, which indi-
cates that China’s commercial housing price index has made the reflection in
that month.
Second is to adjust the bank loan interest rate and the first loan ratio. In
September 2006, the people’s bank of China issued a notice on strengthening
the management of the commercial real estate credit, which clearly stipulates
1782 C. Liu et al.
that the loans down payment ratio of the second set housing shall not be less
than 40%, lending rates may not be lower than 1.1 times of the benchmark
interest rate which announced by people’s Bank of China over same period and
grade, and the proportion of the loans down payment and interest rates will be
significantly improved with the increase of volume, the purpose of which is to
regulate trade order and rectify the real estate market.
Third is about to investigate and punish the violations of law and discipline
and the ownership disputes. However, the policy which can take effect imme-
diately is still relatively less, and most of the policies are lagging, such as the
state issued the regulations to strengthen land supply regulation and shorten
the period of land development in October 2007, and in the same month, prop-
erty tax “idling” ten spread to 10 cities. Housing price index began to make the
corresponding reflection in December.
5.2 Validity of Policy

Among these housing price regulation policies since 2006, raising lending rates,
introducing second-hand housing business tax policies, regulating loans down
payment ratio on second homes and other relevant policies played a profound
influence on housing price regulation, which has a long period of validity in a
period of time after the introduction of the policies. Such as in April 2006.
Interest rates on deposits and loans were raised eight times by the people’s
bank of China, mortgage interest rates were up-regulated again regulation staged
a comeback, the implementation of this policy had an indirect influence on the
real estate prices, which continued to 2008; In September 2007, the central bank
regulated that as a family unit, the loans down payment on second homes shall
not be less than 40%, the interest rate shall not be less than 1.1 times of the
benchmark interest rate, cracking down on speculation, which had played a cer-
tain influence on price regulation in later three years.
The property market regulation “new country of five” was introduced in the
state council executive meeting, which reiterated its adherence to the implemen-
tation of the regulation policy which took the purchase limit and credit limit as
the core, resolutely crack down on speculative investment, urged all localities to
publish the annual housing price regulation targets. “New country of five” also
stipulates that tax was levied on personal income of 20% in second-hand housing
transaction. This provision would discourage investment demand and affects the
volume, which is a positive effect for the real estate regulation, brought inhibitory
effect on the residents’ improved housing demand. The implementation of these
policies had a long-term impact on the housing price regulation in China. And
the introduction of the new policy in the recent two or three years, due to the
implementation time is not longer, we can’t see whether it has a long term.
5.3 Neutralization and Superposition of Policy

In the process of housing price regulation, in the face of different reflection of the
market, the later policies released by government will cause a certain influence to
the previous policies. This influence is two-sided, either playing a strengthening

effect on the previous policies, call superposition effect; or having a conflict to
previous policies, call neutralizing effect.
In July 2009, in order to strengthen the supervision of credit funds, banking
regulatory bureau released “Interim Measures for the administration of fixed
assets loans” and “guidelines for project financing”, to ensure that the fixed
assets loan fund is truly used for the needs of the real economy, to prevent
diversion of loans, and prevent the bank risk under the trend of rapid loan
growth. In the next month of document issuance, the market did not make
obvious reflection. We set this policy issuance as St23. In December, the State
Council discussed and passed “opinions on the trial of the social insurance fund
budget”. The views decided to make the period of the personal housing transfer
business tax exemption from 2 years be back to five years. And five ministries
and commissions such as the Ministry of finance, Ministry of land and resources,
the central bank, the supervision published “notice on the further strengthening
the management of land transfer revenue and expenditure”, improving the land
developer down-payment. We se this policy as St25.
As you can see, after the promulgation of ST23, St25 began to change signif-
icantly, indicating that the market began to be sensitive to the policy enacted
in July that year, that is, St25 had a superposition effect on St23.
However, in September 2010, with the introduction of the policy that the
national commercial banks suspended issuing residents and families to buy the
third and above housing loans (St27), the significance of the impact of the policy
enacted in July 2009 (St23) on the market disappeared, which shows that the
two policies have a neutralization. As shown in Fig. 5, the asterisk represents a
significant effect.
Fig. 5. Policy impact
6 Conclusion
This paper quantitatively analyzes the effectiveness of housing price regulation

policy. Empirical analysis results show that mostly policies do not began to take
effect in that month, and the policy regulation effect is much poorer in the second
half of 2007, the first half of 2008, 2009 and 2012. The effective time of different
1784 C. Liu et al.
policy is also different, where the policies have a certain impact on the next two
years housing prices in September 2007 and February 2013. Policies launched
by government may have overlapping or effect neutralization. These policies are
international.
Using ARIMA model and intervention analysis to analyze the effect of the
policy intervention, predicting the effect of this policy, the effect of superposition
or neutralize is analyzed by linking the policy and previous policy. It will provide
reference for the relevant functional departments to formulate effective housing
price regulation, so that the policies have been introduced together to achieve
the best regulation effect.
Acknowledgements. This paper is supported by the National Natural Science Foun-

dation of China (General Program Grant No. 71473204), Sichuan Soft science research
project (Grant No. 2015ZR0128), Chengdu Soft science research project (Grant No.
2015-RK00-00087-ZF), Chengdu Soft science research project (Grant No. 2015-RK00-
00257-ZF), Research Grant Foundation of Sichuan Education Department (Grant No.
xyzx1506) and Science Foundation of Ministry of Education of China (Grant No.
14YJC790053).
References
1. Agarwal S, Rengarajan S et al (2016) School allocation rules and housing prices: a
quasi-experiment with school relocation events in singapore. Reg Sci Urban Econ
58:42–56
2. Chen WD, Hall S, Pauly P (2016) Policy failure or success? Detecting market
failure in China’s housing market. Econ Model 56:109–121
3. Chen X, Fang Y (2016) Research on the implementation and exit effect of the
control policies on real estate based on VECM and DSGE model. Mod Econ Sci
38:31–43
4. Crawford GW, Fratantoni MC (2003) Assessing the forecasting performance of
regime-switching, arima [auto regressive integrated moving average] and garch [gen-
eralized autoregressive conditional heteroskedasticity] models of house prices. Real
Estate Econ 31:223–243
5. Dang Y, Liu Z, Zhang W (2014) Land-based interests and the spatial distribution
of affordable housing development: the case of Beijing, China. Habitat Int 44:137–
145
6. Dezhi L, Yanchao C et al (2016) Assessing the integrated sustainability of a public
rental housing project from the perspective of complex eco-system. M.E. Sharpe
7. Feng H, Lu M (2010) School quality and housing prices: empirical evidence from
a natural experiment in Shanghai, China. J Hous Econ 22(4):291–307
8. Feng Q, Wu GL (2015) Bubble or riddle? An asset-pricing approach evaluation on
China’s housing market. Econ Model 46:376–383
9. Ge J (2017) Endogenous rise and collapse of housing price: an agent-based model
of the housing market. Comput Environ Urban Syst 62:182–198
10. Hui ECM, Zhong J, Yu K (2016) Land use, housing preferences and income poverty:
in the context of a fast rising market. Land Use Policy 58:289–301
11. Nie X, Liu C (2005) An empirical study on the transmission of Chinese monetary
policy toward investment. Manage Rev 17(2):51–54
12. Shi J, Zou W (2009) Forecasting house price index for USA with autoregressive
integrated moving average models. In: The international conference on manage-
ment of technology, Taiyuan, pp 590–594
13. Wan TL, Wang L et al (2016) Housing price prediction using neural networks. In:
International conference on natural computation and fuzzy systems and knowledge
discovery, pp 518–522
14. Wang L (2013) The effect of recent real estate regulation policy on house price: a
case study of Wuhan city. Mod Prop Manage 221(2):306–315
15. Wen H, Zhang Y, Zhang L (2014) Do educational facilities affect housing price?
An empirical study in Hangzhou, China. Habitat Int 42(42):155–163
16. Woo K, Sung-Suk R (2015) Time series modeling for forecasting land price change
rate-focusing on the intervention ARIMA model. Korea Real Estate Acad Rev
60:142–154
17. Wu W (2014) Effectiveness evaluation of housing price regulation policy based on
PSR model.business information. Bus Inf 20:63–63
18. Xie X, Hu G (2007) A comparison of Shanghai housing price index forecasting. In:
International conference on natural computation, pp 221–225
19. Zhang H, Li Y, Li H (2013) Multi-agent simulation of the dynamic evolutionary
process in Chinese urban housing market based on the GIS: the case of Beijing.
Autom Constr 35(11):190–198
20. Zhang Z, Tang W (2016) Analysis of spatial patterns of public attention on housing
prices in Chinese cities: a web search engine approach. Appl Geogr 70:68–81
21. Zhou Z (2016) Overreaction to policy changes in the housing market: evidence
from Shanghai. Reg Sci Urban Econ 58:26–41
A Descriptive Analysis of the Impact of Air
Pollution on the Mortality of Urban and Rural
Residents in Mianyang
Jianping Xu1 , John Thomas Delaney2 , Xudong Chen3(B) , and Liming Yao4
1
Regional Economic Development and Enterprises Administration Center,
Sichuan University, Chengdu 610062, People’s Republic of China
2
Kogod School of Business, American University, 4400 Massachusetts Avenue,
Washington, DC 20016-8044, USA
3
School of Management Science, Chengdu University of Technology,
chengxudong198401@163.com
4
Abstract. The paper uses the air pollution data of Mianyang, a city
in Sichuan Province, and the death data of its urban and rural resi-
dents from 2008 to 2014 to investigate the relationship between air pol-
lution and human mortality. SPSS19.0 is utilized to conduct a descriptive
analysis of the correlation between the principal air pollutants (including
PM10 , NO2 , SO2 ) and the mortality (including gender, age, education,
respiratory diseases, non-respiratory diseases, and chronic obstructive
pulmonary diseases). Hence, the paper shows to what extent air pollu-
tion affects the mortality rate of the people in Mianyang, thereby further
evaluating the influence of air pollution on the health of the residents.
The results indicate that the rising trend of air pollution is consistent
with the expansion of industrial production. The paper estimates that the
mortality rate is correlated to air pollution during the normal develop-
ment of society. The government should emphasize air pollution control
and take measures to reduce air pollution, so as to improve the quality
of life of urban and rural residents and to reduce the mortality caused
by air pollution directly or indirectly.
Keywords: Air pollution · Mortality · Descriptive analysis
1 Introduction
Haze related pollution have been proved to be the reason for lung cancers and
some other diseases and also have serious impact on and threatening future
health consequences across the country [9]. Numerous Chinese and foreign stud-
ies have shown that PM10 and gaseous pollutants, such as sulfur dioxide (SO2 )
and nitrogen dioxide (NO2 ), have a significantly negative impact on human
A Descriptive Analysis of the Impact of Air Pollution 1787
health. Moreover, a positive correlation between mortality and air pollution has
been demonstrated in multiple researches [1,4,8,16]. The 2002 WHO report esti-
mated that the global urban air pollution caused at least 1 million deaths per
year and 7.4 million disability-adjusted life years (DALY) [3].
China is one of the countries with severe environmental problems. The con-
centration of particulate air pollution is far beyond the developed countries in
Europe and North America. In the past 30 years, the economy and industry
had been developing continuously. The rapid growth of energy consumption has
aggravated air pollution, which becomes one of the major risk factors for damag-
ing human health. The irrationality of the energy structure is that coal accounts
for nearly 70% of the total energy consumption in China. The PM and gaseous
pollutants emitted from coal combustion are the main sources of air pollution.
The cities in China face severe air pollution, and the PM10 is much higher than
the 20 µg/m3 (annual average) and 50 µg/m3 (24 h average)–the air quality
standards recommended by the WHO.
Mianyang City, which is located in Southwest China, is the second largest
city in Sichuan Province in terms of economic output. It has a dry climate and
is densely populated. The agriculture, industry, and service (transportation) are
well-developed there with a ratio of 15.3: 50.5: 34.2, indicating that the industry
represents half of the economic output. Therefore, it is of significance to examine
the impacts of air pollution on human health. In order to study the impact of
air pollution on the mortality rate of urban and rural residents in Mianyang,
the paper conducts a descriptive analysis, investigating the correlation between
the principal air pollution data (including PM10 , NO2 , SO2 ) and the human
mortality.
2 Data and Methods

Mianyang City, the main terrain of which is of hills, is located in the northwest
of the Sichuan Basin. It has a subtropical humid monsoon climate. Due to the
influence of the northerly air flow, Mianyang is dry and cold, with less rainfall
during the winter half year; however, controlled by the southerly air flow in
the summer half year, it is hot, rainy, and humid. The household population of
Mianyang was 5.488 million by the end of 2015.
The data used in this paper include mortality data and air pollution data
from 2008 to 2014.
Mortality data: All deaths from 2008 to 2014 are gathered from the Mianyang
municipal government. The paper clarifies the diseases into two groups accord-
ing to WHO International Classification of Diseases (ICD-10): deaths caused
by respiratory diseases (J00∼J98) and deaths caused by other diseases (or non-
respiratory diseases). It, then, conducts an individual analysis of the chronic
obstructive pulmonary disease (COPD). The human mortality data are clas-
sified into the following categories: sex-specific mortality, age-specific mortal-
ity, education-specific mortality, and mortality caused by respiratory and non-
respiratory diseases (the chronic obstructive pulmonary disease is singled out).
1788 J. Xu et al.
Principal air pollution data: The annual average concentration of PM10 ,

NO2 , SO2 from 2008 to 2014 are provided by the Mianyang Environmental
Protection Bureau.
Methods: SPSS19.0 is utilized to conduct a descriptive analysis of the air pol-
lution data of Mianyang and the mortality of its urban and rural residents from
2008 to 2014.
3 Results
3.1 A Descriptive Analysis of Air Pollution and Death Data

The statistical data of air pollution and death in Mianyang from 2008 to 2014
are shown in Tables 1, 2, and 3. The annual average PM10 of Mianyang from
2008 to 2014 was 0.059 mg/m3 , and the peak value was 0.082 mg/m3 in 2008.
The annual average SO2 concentration was 0.015 mg/m3 , and the peak value was
0.029 mg/m3 in 2009. The annual average NO2 concentration was 0.019 mg/m3 ,
and the peak value was 0.039 mg/m3 in 2009. The average death toll of men and
women in Mianyang was 14,140 and 9061, respectively from 2008 to 2014. The
average death toll of people with unknown education was 2098, illiterate or semi-
illiterate 7404, primary education 10354, secondary education 2963, university
and above 298. The average deaths were 2353 for people aging under 44, 6032
for residents from 45 to 64, and 14818 for those ageing over 65. The average
deaths caused by respiratory diseases and non-respiratory diseases were 6663 and
16541, respectively. The average deaths caused by chronic obstructive pulmonary
diseases were 767.
It can be seen from Table 1 and the corresponding line graph that the
Mianyang air pollution generally showed a declining trend from 2008 to 2014.
PM10 , SO2 and NO2 reached their peak from 2008 to 2009. Then, PM10 showed
a downward trend till 2014. SO2 and NO2 declined moderately from 2009 to
2012. After that, SO2 showed an upward trend from 2012 to 2014, while NO2
fluctuated repeatedly between 2012 and 2014.
The paper will discuss, from three aspects, why the overall trend of air pol-
lution in Mianyang declined from 2008 to 2014, even though it fluctuated some-
times during this period.
(1) The peak of air pollution between 2008 and 2009 was caused by the 2008
Wenchuan earthquake. The research shows that due to the earthquake, the col-
lapsed buildings, the leakage of toxic gases and radioactive materials lead to
the aggravation of air pollution [2]. Because of the great magnitude and severe
damage of the 2008 Wenchuan earthquake, Mianyang, located in Longmenshan
earthquake zone, was affected by the aftershocks. This earthquake caused heavy
casualties and severe property losses, and Mianyang became one of the hard-
est hit areas. The urban and rural housing, collapsed or damaged, reached
6,284,011and 58,441,355, respectively. In addition, 47% of the industrial enter-
prises in Mianyang were affected, with 74% of plant damages. Thus, the collapse
and damage of residential buildings and industrial plants caused by the earth-
quake, to a certain extent, lead to an increase in airborne dust particles and toxic
gas content, and exacerbate air pollution, thereby resulting in the Mianyang air
pollution peak between 2008 and 2009.
(2) During the restoration and reconstruction process starting from June,
2008, due to the industrial and agricultural losses and the impacts of the earth-
quake, the local workers tended to find employment outside Mianyang, whereas
the enthusiasm of the non-local job seekers, especially skilled workers and high-
tech talents, were seriously affected in a certain period of time. Second, because of
the severe capital losses caused by the damages to the plants and the equipment,
some enterprises need relocation and reconstruction, increasing the costs and
slowing down the process of the industrial and agricultural recovery in Mianyang.
Moreover, most of the enterprises need a longer recovery period, since the major
coal companies stop production due to the disaster; other kinds of electricity sup-
plies are affected; and the main roads and power facilities are severely damaged
[7]. During the recovery period, because of the less industrial and agricultural
emissions caused by the decreased productivity of the enterprises, and less traffic
pollution emissions, the air pollution generally showed a downward trend from
2009 to 2012. After a 3–4 year of recovery, the overall industrial and agricul-
tural industries were recovered in 2012, and the infrastructure was gradually
improved, resulting in a rise in air pollution.
(3) The reasons why the air pollution in Mianyang showed a decreasing
trend between 2013 and 2014 are twofold. First, according to Sichuan Air Pol-
lution Prevention and Control, compared to 2012, the annual average concen-
tration of PM10 should decrease by 10%. Mianyang, as one of the key cities
in Sichuan Province, takes strict emission reduction measures to control the
coal-fired elevated-source, industrial point source, urban non-point source, and
moving source (vehicle pollution), through strict assessment and accountabil-
ity. Therefore, air pollution emissions are reduced from a number of sources.
Second, because of the overall improvement in the environmental awareness of
the Mianyang residents, the development of tourism, the promotion of energy
conservation, as well as green travel, air pollution emissions are reduced.
Table 1. Air pollution in Mianyang from 2008 to 2014
Year PM10 (mg/m3 ) SO2 (mg/m3 ) NO2 (mg/m3 )

2008 0.082 0.015 0.038
2009 0.081 0.029 0.039
2010 0.05 0.021 0.02
2011 0.041 0.016 0.015
2012 0.051 0.001 0.003
2013 0.071 0.009 0.02
2014 0.035 0.012 0.0009
1790 J. Xu et al.
Table 2. The Death Toll in Mianyang from 2008 to 2014
Year/ Total Male death Female Unknown Illiterate/ Elementary Middle University
death toll population toll death toll degree of semi- school school and above
of education literate education education education
Mianyang death toll death death death death
2008 5407100 10360 6504 1867 6912 5245 2569 271
2009 5446500 10557 6698 2047 6781 5620 2625 280
2010 5418700 12129 7553 2543 7648 6194 3025 272
2011 5433600 14387 8999 2706 9449 7490 3453 284
2012 5454000 15395 10044 2980 10433 8197 3520 306
2013 5474000 15942 10382 3243 10605 8458 3725 290
2014 5488000 20212 13252 0 0 31275 1827 386
Table 3. The death toll in Mianyang from 2008 to 2014
Year/ 0–44 years 45–64 Over Respiratory Non- COPD

death toll old death years old 65 year old disease respiratory (chronic
toll death toll death toll death toll disease death pulmonary
toll disease)
death toll
2008 2274 4638 9952 4807 12057 153
2009 2107 4697 10451 4891 12364 143
2010 2415 5392 11874 5377 14305 269
2011 2503 6248 14633 6523 16861 280
2012 2417 6448 16573 7233 18205 519
2013 2343 6671 17309 7700 18623 638
2014 2412 8136 22939 10112 23375 3367
3.2 Description of the Relationship Between Air Pollution

and Human Mortality in Mianyang from 2008 to 2014
(Note: In order to better reflect the relationship between air pollution and human
mortality, the paper uses air pollution values * % to compare with the mortality
rate.)
Figure 1 illustrates that the PM10 fluctuated between 2008 and 2014. The
peak between 2008 and 2009 is partly because the collapsed buildings, the leak-
age of toxic gases and radioactive materials, due to the Wenchuan earthquake,
aggravated air pollution. PM10 , SO2 , and NO2 rose steadily since 2010, during
which the male and female mortality showed the same trend. Specially, the male
mortality rate increased greatly, indicating that men are more sensitive to the
change in air pollution concentration. Under the impacts of the national policies
and local laws and regulations, the air pollution concentration has decreased
since 2013. However, there is a greater increase in the mortality rate, which con-
tradicts the conclusion. By searching information and literature, the paper finds
that Mianyang faced a severe aging problem from 2012 to 2014. More precisely,
Mianyang has a large aged population which grows fast. By the end of 2013, the
residents aged 60 and above reached 1,07 million, accounting for 19.8% of the
total population, and the individuals who die of old age were 22,939 in 2014.
Therefore, the 2013–2014 mortality rate is influenced by the aging population
[10,14]. Because of the changes in the social demographic structure, the impact
of air pollution on human death is reduced.
Fig. 1. The relationship between air pollution and gender-specific mortality
Obviously, it can be seen from Fig. 2 that there was a dramatic change in
the death rate of residents with primary and secondary education. The reason is
because the Mianyang Statistic Bureau changed the classification standard of the
education-specific human mortality. That is, people with unknown education,
illiterate or semi-illiterate, primary, secondary education, and university and
above is changed into junior high and below, technical secondary education,
high schools, technical schools, universities and colleges. Hence, between 2013
and 2014, the dramatic changes in human mortality were due to the classification
criteria of the Mianyang Statistic Bureau, which was not related to air pollution.
The mortality rate of people with primary school education, illiterate or
semi-literate, grew steadily from 2008 to 2013, whereas that of residents with
secondary education, university and above showed a stable trend, maintaining
the same level. It indicates those in terms of the same air pollution level, the
higher the education level, the stronger the awareness of environmental protec-
tion. Therefore, these residents tend to take preventive measures more timely
and effectively when exposed to air pollution, and they are less affected [5,11].
The impact of air pollution on human death is positively correlated with the edu-
cational level of the residents. The more serious the air pollution is, the higher
the mortality rate is. However, due to the differences in people’s education level,
the awareness of air pollution prevention varies considerably. The people with
the lower education level may have the lower residents’ awareness of air pollu-
tion prevention, which results in a higher chance of being affected by the air
pollution.
1792 J. Xu et al.
Fig. 2. The relationship between air pollution and education-specific mortality
According to Fig. 3, between 2008 and 2014, the mortality rate of residents
aged 0–44 years and 45–64 years showed a stable trend, while that of people aged
65 years and over was consistent with the trend of air pollution between 2008
and 2013, indicating that air pollution impacted individuals aged 65 years and
above the most. In terms of residents aged 44 and over, their physical function
declines with age, and they are more easily affected by the external environment.
Air pollution, to a larger extent, impacts individuals aged 44 and over, the
death rate of whom may increase with a rise of the air pollution concentration
[13]. However, the mortality rate of people aged 65 years and over in 2013–2014
showed a reverse trend with air pollution, which contradicts the argument that
the senior residents are susceptible to air pollution. By searching literature, the
paper finds that Mianyang faced a severe aging problem from 2012 to 2014. By
the end of 2013, the residents aged 60 and above reached 1,07 million, accounting
for 19.8% of the total population. Hence, a high aging population impacts the
human mortality.
Fig. 3. The relationship between air and age-specific human mortality
It can be seen from Fig. 4 that the mortality rate caused by respiratory
diseases showed a steady upward trend from 2008 to 2013, but it grew rapidly
from 2013 to 2014. Due to the 2008 Wenchuan earthquake, Mianyang experienced
a recovery period of agriculture, industry, and transport infrastructure. As a
result, air pollution declined from 2009 to 2011. Nevertheless, the concentration
of air pollution increased from 2011 to 2013, which correspondingly saw a rising
trend of respiratory diseases-related deaths. It indicates that air pollution has a
significant impact on the residents who died of respiratory diseases. During the
period of 2013 and 2014, the human mortality rate caused by respiratory diseases
showed a reverse trend with air pollution, contradicting the argument that there
is, according to existing researches, a positive correlation between respiratory
diseases-related deaths and air pollution [6,12,15]. From the Mianyang 2013–
2014 government and other reports, the paper finds that this city faced a severe
aging problem. Due to the prevalence of respiratory diseases among the senior
residents, the 2013–2014 death rate was correlated considerably with the aging
population, less with air pollution.
Fig. 4. The relationship between air and mortality of disease
4 Discussion
The paper uses descriptive statistical analysis to investigate the correlation
between major air pollutants and residents’ mortality in Mianyang City. Given
the impacts of natural disasters, national policies, and socio-demographic
changes, the paper concludes that the urban and rural residents, who are male,
illiterate or semi-literate, or with primary education, the middle-aged or aged,
or with respiratory diseases, are more susceptible to the increase in the PM10 ,
NO2 and concentration.
The accumulation of pollutants and photochemical reactions in various cities
leads to complex regional air pollution issues. Mianyang, located in the northwest
of Sichuan Basin, has a subtropical humid monsoon climate. The hot, rainy, and
wet weather is not conducive to the diffusion of particulate matters, aggravating
the air pollution of Mianyang. Furthermore, factors, such a high level exploita-
tion of the rich mineral resources in Mianyang and its emphasis on industry
1794 J. Xu et al.
development, contribute to the severe air pollution. Therefore, the Mianyang

government should consider the health risk of air pollution, and take effective
and feasible measures to tackle this issue, thereby reducing the pollution-related
health risks of its urban and rural residents.
Here are four policy suggestions we want Mianyang government can accept
and conscientiously implement. First, government needs to improve the local
industrial structure by circulation type industrial building, which can reduce
air pollution and improve resource utilization efficiency, also can achieve the
transformation from primary industry into tertiary industry. Second, attaching
great importance to the city public transport priority, vigorously implement-
ing the bus priority development strategy, and increasing the city public traffic
investment and construction strength in order to reduce harmful automobile
exhaust emissions and protect residents’ health. Third, promotion of industrial-
ized and enterprise and market-oriented pollution treatment, encouragement of
fund raised from diverse channels and various quarters in society, and intensified
collection of charges on urban sewage treatment and garbage disposal, which
can effectively reduce every kind of contaminative source. Last but not least,
Mianyang government should make the enactment of relevant laws and regu-
lation and raise residents’ awareness about the health risks of air pollution to
reduce the risk of death for the people, who are male, illiterate or semi-literate,
or with primary education, the middle-aged or aged, or with respiratory diseases.
In addition, the study has certain limitations: First of all, due to the high
mobility of the Mianyang urban and rural household population, the reported
death information may be false or incorrect. Meanwhile, there may be a certain
degree of misclassification of death data, which might affect the accuracy of
residents’ mortality. Secondly, the air pollution concentration measured at the
fixed sites represents the exposure level of the urban and rural residents, which
may differ, to some extent, from the real exposure level. Moreover, due to damage
and impact of the Wenchuan earthquake, there might be a certain deviation of
air pollution statistical data gathered at each site, affecting the analysis results.
Acknowledgements. This research was supported by Sichuan Social Science Plan-

ning Program Key Project (Grant No. SC16A006), System Science and Enterprise
Development Research Center Project (Grant No. Xq16C12) and Sichuan University
Central University Research Fund project (Grant No. skqy201653), National Natural
Science Foundation for Young Scholars of China (Grant No. 71301109), Soft Science
Program of Sichuan Province (Grant No. 2017ZR0154), and Sichuan University (Grant
No. skqx201726).
References
1. Fischer PH, Marra M et al (2015) Air pollution and mortality in seven million
adults: the Dutch environmental longitudinal study (duels). Environ Health Per-
spect 123(7):697–704
2. Gotoh T, Nishimura T et al (2002) Air pollution by concrete dust from the Great
Hanshin Earthquake. J Environ Qual 31(31):718–723
3. Gulland A (2002) Air pollution responsible for 600 000 premature deaths world-
wide. BMJ (Clinical Res Ed) 325(7377):1380
4. Jerrett M, Burnett RT et al (2013) Spatial analysis of air pollution and mortality
in Los Angeles. Am J Respir Critical Care Med 188(5):727–736
5. Kan H, Chen B (2008) Season, sex, age, and education as modifiers of the effects
of outdoor air pollution on daily mortality in Shanghai, China: The public health
and air pollution in Asia (PAPA) study. Environ Health Persp 116(9):1183–1188
6. Katanoda K, Sobue T, Satoh H (2011) An association between long-term exposure
to ambient air pollution and mortality from lung cancer and respiratory diseases
in Japan. J Epidemiol 21(2):132–143
7. Kirby E (2008) Stress changes from the 2008 Wenchuan earthquake and increased
hazard in the Sichuan basin. Nature 454(7203):509–510
8. Lelieveld J, Evans JS et al (2015) The contribution of outdoor air pollution sources
to premature mortality on a global scale. Nature 525(7569):367–371
9. Li Y, Xu Y et al (2016) Haze-related air pollution and the tourism industry in
China. In: GAI international academic conferences proceedings, pp 122–131
10. Markides KS, Eschbach K (2005) Aging, migration, and mortality: current status
of research on the Hispanic paradox. J Gerontol 60(Spec No 2):68
11. O’Neill MS, Bell ML et al (2008) Air pollution and mortality in Latin America:
the role of education. Epidemiology 19(6):810–819
12. Saldiva PH, Lichtenfels AJ et al (1994) Association between air pollution and
mortality due to respiratory diseases in children in São Paulo, Brazil: a preliminary
report. Environ Res 65(2):218–225
13. Schwartz J, Dockery DW (1992) Increased mortality in Philadelphia associated
with daily air pollution concentrations. Am Rev Respir Dis 145(3):600–604
14. Strehler BL, Mildvan AS (1960) General theory of mortality and aging. Science
132(3418):14
15. Wong TW, Tam WS et al (2002) Associations between daily mortalities from
respiratory and cardiovascular diseases and air pollution in Hong Kong, China.
Occup Environ Med 59(1):30–35
16. Xie W, Li G et al (2015) Relationship between fine particulate air pollution and
ischaemic heart disease morbidity and mortality. Heart 101(4):257–263
Author Index
A Duarte, Susana, 1242

Abdullayeva, Narmina, 858 Duca, Gheorghe, 1283
AduAgyapong, Richard, 708
Ahmed, Shariq, 615 E
Ahmed, Syed Ejaz, 19, 891 Ebrahimnejad, Sadoullah, 511
Akram, Umair, 247, 680, 1535 Effati, Sohrab, 223
Ali, Sajid, 47
Amini, Alireza, 511 F
Aoki, Ritsuko, 1407 Fan, Lurong, 721
Arifjanov, Abdulla, 735, 1198 Fan, Shucen, 1773
Ash, John, 286 Fang, Qian, 1722
Ayupov, Shavkat, 735 Feng, Cuiying, 204
Azeez, Karrar Abdulelah, 1466 Fu, Xiaodan, 1508
Fujimoto, Yuki, 804
B
Benmessaoud, Tahar, 749 G
Gafar, Achmad Fanany Onnilita, 300
C Gan, Lu, 438
Cai, Dongliang, 499 Gang, Jun, 902, 1296, 1645
Cao, Qilin, 1508 Gannon, Patrick, 559
Chen, Chuan, 1559 Gaoyu, Lv, 1535
Chen, Jiaying, 1330 Gen, Mitsuo, 35, 106, 140, 212, 382, 768, 962,
Chen, Jingdong, 321 1030, 1752
Chen, Mo, 321 Ghuseynov, Elchan, 1595
Chen, Ping, 1559 Gonzalo, Alfredo Peinado, 1161
Chen, Xing, 962, 1030 Gu, Xin, 1762
Chen, Xudong, 1786 Guo, Chunxiang, 1122
Chen, Yao, 1019 Guo, Xinyan, 341
Cheng, Yong, 1318 Guo, Yaomin, 1508
Chuluunsukh, Anudari, 962, 1030 Guo, Yinpeng, 577
Cruz-Machado, Virgilio, 1147, 1242
H
D Hajiyev, Asaf, 858
Dai, Jingqi, 1111 Han, DongPing, 260
Delaney, John Thomas, 658, 1209, 1786 Hao, Xinchang, 35, 63
Deng, Juqiu, 1737 Hashim, Muhammad, 680, 1535
Deng, Yanfei, 538, 1220 Haviluddin, 300
Donata, Sobakinova, 260 Hayashi, Takahiro, 481
© Springer International Publishing AG 2018 1797

on Multidisciplinary Industrial Engineering, DOI 10.1007/978-3-319-59280-0
1798 Author Index
He, Changzheng, 878 L

He, Yue, 547, 878 Lan, Dao, 78
Hou, Ning, 1220 Lan, Hongxing, 129, 1007
Hu, Jiancheng, 331 Latif, Zahid, 490
Hu, Lin, 438 Lee, Min Seok, 1630
Hu, Xiaoxue, 1548 Lev, Benjamin, 902, 1604
Hu, Zhineng, 1318 Li, Fan, 1307
Hua, Wen, 975 Li, Hao, 1773
Huang, Hua, 1773 Li, Haojie, 1428
Huang, Huang, 600 Li, Lu, 588
Huang, Yeyun, 1230 Li, Luo, 1136
Huang, Yong, 421, 1089 Li, Luoji, 117
Hui, Peng, 680 Li, Qian, 791
Hussain, Zulfiqar, 490 Li, Qiulin, 117
Hussein, Abdulkadir A., 891 Li, Sihan, 1522
Li, Wenjie, 670
I Li, Xia, 814
Ida, Kenichi, 140 Li, Xiaofeng, 634
Inoue, Hisaki, 768 Li, Xiaoping, 1111
Isobe, Seiichiro, 937 Li, Xin, 986
Li, Xingyao, 1497
J Li, Ying, 658, 1209
Jan, Nasir, 490 Li, Yinghan, 194
Jang, Young Jae, 1583, 1630 Li, Yu, 1043
Ji, Yuan, 1697 Li, Zongmin, 721, 1186
Jiang, Dan, 274 Liang, Pan, 1737
Jiang, Jue, 947 Liang, Xuedong, 1773
Jiang, Lili, 1089 Lin, Lin, 63, 382, 1428
Jiang, Qiang, 1773 Lin, Shaojiang, 1330
Jiang, Songtao, 1697 Lin, Sihan, 247
Jiang, Wen, 1266 Lisawadi, Supranee, 19
Jiang, Wu, 499 Liu, Chang, 1773
Jie, Xiaowen, 1361 Liu, Haiyue, 1439
Jiménez, Alfredo Arcos, 1377 Liu, Hongxia, 522
Jo, Jung Bok, 768 Liu, Jian, 421
Liu, Jiawei, 398
K Liu, Kai, 1209
Kamran, Asif, 367 Liu, Ke, 452
Kang, Kai, 1103 Liu, Shuai, 998
Karrar, Abdulellah Azeez, 260 Liu, Tingting, 1351
Kasahara, Takehiro, 804 Liu, Xiaofeng, 1276
Katayama, Hiroshi, 937, 1054, 1407 Liu, Xin, 1318
Kawabe, Hiroyuki, 469 Liu, Xuhong, 1361
Kazemzadeh, Shahin, 1175 Liu, Yanhua, 1276
Kazmi, Syed Hasnain Alam, 615 Liu, Yu, 398
Khan, Abdullah, 615 Liu, Yue, 600
Khan, Faraz Ullah, 367 Liu, Yunqiang, 1230
Khan, Muhammad Kaleem, 680, 1535 Liu, Yuxi, 311
Kheirabadi, Akram, 223 Liu, Zhihan, 1490
Kimura, Haruhiko, 804 Liu, Zhusheng, 1089
Kojima, Koji, 804 Lu, Chenqing, 1752
Kun, Li, 1548 Lu, Li, 1078
Author Index 1799
Lu, Sun, 382 Qiu, Yixin, 247, 680

Lu, Yi, 194, 626 Qu, Minglei, 180
Luo, Min, 814
Luo, Ping, 658 R
Luo, Rui, 1604 Rafi, Usman, 47
Luo, Yong, 1136 Rahimov, Shafahat Rahim, 1595
Luo, Yuyan, 1019 Ramirez, Isaac Segovia, 1161
Razavi, Hamideh, 1417
M Razzaq, Abdul, 47
Ma, Hongsheng, 1136 Reangsephet, Orawan, 19
Ma, Jing, 1318 Rezaee, Babak, 1175, 1342
Ma, Ning, 835 Rubayat, Shibli, 1389
Ma, Sheng, 670 Ruiz-Hernández, Diego, 1395
Ma, Yanfang, 204, 1103
Ma, Yufeng, 878
Mahdi, Moudi, 1351 S
Mahmood, Marwah Abdulkareem, 1466 Saito, Masatoshi, 804
Mammadov, Javanshir, 1595 Salam, Shafaq, 490
Mancl, Karen, 538 Saleem, Muhammad Asim, 47
Márquez, Fausto Pedro García, 749, 912, 1161, Salman, Muhammad, 490
1377 Sarwar, Adnan, 708
Marugán, Alberto Pliego, 749, 912 Sato, Tetsuya, 1054
Mei, Hongchang, 1677 Sawada, Ayako, 644
Meng, Zhiyi, 129, 1007 Seino, Yoshio, 481
Mereuţa, Aliona, 1283 Seto, Shuichi, 469
Mohammedi, Kamal, 749 Shen, Charles, 351
Monirian, Masoud Amel, 1342, 1417 Shen, Feng, 78
Muñoz, Carlos Quiterio Gómez, 1161, 1377 Sheng, Yi, 1722
Murata, Koichi, 937 Shimomura, Yuko, 469
Simas, André, 1147
N Snowdon, Anne, 891
Nakamura, Junko, 1616 Song, Lingxi, 547
Nambo, Hidetaka, 151, 469, 804 Song, Tingting, 1276
Naseri, Nastaran, 1175 Song, Xiaoling, 351
Nazam, Muhammad, 247 Song, Xin, 522
Ni, Jian, 1476 Song, Yang, 421
Nomura, Kota, 804 Süer, Gürsel A., 559, 947
Nugroho, Hari, 658 Sun, Lu, 35, 1428
Sun, Yang, 194
O Sung, Shin Woong, 1583
Oshima, Sachiko, 758
Ou, Limei, 78 T
Oyabu, Takashi, 1616 Tahyudin, Imam, 151, 300
Takano, Daiki, 140
P Tang, Jing, 1330
Pakkar, Mohammad Sadegh, 695 Tang, Yingkai, 600
Paul, Sameer, 615 Tang, Yiwen, 891
Peña-Mora, Feniosky, 351 Tao, Zhimiao, 1067
Pinar-Pérez, Jesús María, 1395 Tareq, Mohammad, 1389
Purnawansyah, 300 Tavakkoli-Moghaddam, Reza, 511
ThiHoaiThuong, Nguyen, 708
Q Tian, Yizhuang, 1466
Qadeer, Talat, 708, 1186 Toga, Hirohisa, 804
Qian, Xiaoye, 791 Tu, Yan, 902, 1604
1800 Author Index
U Xu, Lei, 538, 1220

Ueda, Yoshihiro, 804 Xu, Qiuhua, 499
Xu, Wei, 1653
V Xu, Xiaojing, 530
Vaziri, Asadollah Mahmoudzadeh, 223 Xu, Xinxin, 286
Vaziri, Shabnam Mahmoudzadeh, 1342, 1417 Xu, Yang, 658
W Y
Wang, Aixin, 878 Yamamoto, Masahide, 411
Wang, Chunxiao, 106, 212 Yan, Fang, 204, 847
Wang, Fuzheng, 180 Yan, Jinjiang, 452
Wang, Hong, 1737 Yan, Sicheng, 670
Wang, Hongchun, 998 Yan, Tengteng, 1559
Wang, Hui, 600 Yang, Dan, 1476
Wang, Kunling, 1677 Yang, Jianchao, 1136
Wang, Lei, 490 Yang, Jing, 1653
Wang, Minxi, 986 Yang, Mengjia, 1253
Wang, Rui, 180 Yang, Qian, 438
Wang, Tao, 1762 Yang, Qing, 1667
Wang, Tianjin, 1522 Yang, Tian, 1497
Wang, Xian, 311 Yang, Xiongtao, 1667
Wang, Xueying, 1667 Yang, Ying, 311
Wang, Yahong, 1019 Yang, Zhen, 499
Wang, Yinhai, 286 Yao, Chenglin, 63
Wang, Yu, 106 Yao, Liming, 1351, 1786
Wang, Yuanyuan, 847 Ying, He, 1535
Wang, Yusheng, 1307 Ying, Qianwei, 1497
Wang, Zeming, 577 Yonezawa, Yuji, 804
Wang, Zhong, 1019 Yoshida, Taketoshi, 644
Wei, Qifeng, 166 You, Xiaoling, 986
Wei, Ying, 1296, 1645 Yu, Dongjing, 1122
Wen, Feng, 1752 Yu, Lei, 1548
Wu, Ke, 634 Yu, Weiping, 233
Wu, Mingcong, 421, 1089 Yuan, Xiaoyue, 1535
Wu, Pingwen, 721 Yuan, Yuan, 274, 538, 1573
Wu, Qiong, 791 Yun, YoungSu, 962, 1030
Wu, Yilun, 791
Wu, Zhibin, 1688 Z
Zakhidov, Romen, 1198
X Zeng, Jianqiu, 490
Xiao, Le, 106, 212 Zeng, Ziqiang, 286, 708
Xiao, Min, 547 Zhan, Qinglong, 1253
Xie, Jiming, 1019 Zhang, Dan, 878
Xie, Tao, 825 Zhang, Huanmei, 814
Xie, Zongtang, 522, 1457 Zhang, Jing, 204
Xing, Jiankai, 452 Zhang, Liangqing, 398
Xiong, Guoqiang, 311 Zhang, Liming, 1573
Xiu, Hongxia, 1457 Zhang, Linling, 91
Xu, Caiyang, 129, 1007 Zhang, Wenqiang, 106, 212
Xu, Dirong, 1296 Zhang, Xinli, 1522
Xu, Jianping, 1786 Zhang, Ye, 1043
Xu, Jing, 1067, 1573 Zhang, Yi, 868
Xu, Jinhua, 600 Zhang, Ying, 117, 680
Xu, Jiuping, 3, 923 Zhang, Yong, 814
Author Index 1801
Zhang, Yue, 547 Zhou, Guichuan, 825

Zhang, Zhaohui, 398 Zhou, Rui, 1573
Zhao, Changyi, 1762 Zhou, Wei, 1476
Zhao, Jing, 1103 Zhou, Xiaoyang, 902, 1604
Zhao, Jingsong, 382 Zhou, Yixiao, 1773
Zhao, Liang, 421, 1089 Zhu, Kai, 452
Zhao, Wei, 1773 Zia, Muhammad Azam, 47
Zhao, Xingchao, 78 Zomorodian, Sima, 626
Zheng, Huan, 1709 Zou, Chiyan, 1762
Zheng, Shuangyi, 825, 835 Zu, Wenjing, 1439
Zhong, Lin, 351, 1688 Zu, Xu, 233

Proceedings of The Eleventh International Conference On Management Science and Engineering Management

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Proceedings of The Eleventh International Conference On Management Science and Engineering Management

Uploaded by

Copyright:

Available Formats

Lecture Notes on Multidisciplinary Industrial Engineering

Jiuping Xu · Mitsuo Gen

Proceedings of the Eleventh

Asaf Hajiyev Fang Lee Cooke

Proceedings of the Eleventh

Lecture Notes on Multidisciplinary Industrial Engineering

© Springer International Publishing AG 2018

Printed on acid-free paper

This Springer imprint is published by Springer Nature

Welcome to the proceedings of the Eleventh International Conference on

research in the ﬁeld of Management Science and Engineering Management, the

May 2017 Jiuping Xu

ICMSEM 2017 was organized by International Society of Management Science and

Jiuping Xu Sichuan University, China

Program Committee Chairs

Mitsuo Gen Fuzzy Logic Systems Institute and Tokyo

Hayato Ohwada Tokyo University of Science, Japan

Organizing Committee Chair

Mitsuo Gen Fuzzy Logic Systems Institute and Tokyo

Local Arrangement Committee Chair

Takashi Oyabu Kokusai Business Gakuin College, Kanazawa,

Local Arrangement Committee

Ayako Sawada Hokuriku Gakuin Junior College, Kanazawa,

Mohammad Z. Abu-Sbeih King Fahd University of Petroleum and Minerals,

Christodoulos A. Floudas Princeton University, USA

Zhineng Hu Sichuan University, China

Lin Zhong Sichuan University, China

Zongmin Li Sichuan University, China

Ning Ma, Jingqi Dai, Lurong Fan, Yan Wang

Advances in Management Science of the Eleventh ICMSEM

Fast Multiobjective Hybrid Evolutionary Algorithm Based

A Priority-Based Genetic Representations for Bicriteria Network

Enterprise Operation Management

Disclosure Behavior of Annual Reports of Listed Companies

The Impact of Mixed Ownership Reform on Enterprise

Decision Support System

Integration of Sound and Image Data for Detection

Advances in Engineering Management of the Eleventh ICMSEM

Green Supply Chain

Bottleneck Management of Multi-stage Sorting-Packing Operations

Resource Optimization Management

Management of Technological Modes of System Distributed

An Empirical Study on the Impact of Niche Overlap of Tourism

Post-Traumatic Stress Disorder Among Survivors in Hard-Hit Areas

Integrated Project Management

Research on Equalization of Public Services Based on Changes

Uncertainty Decision-Making Laboratory, Sichuan University,

Abstract. Management is playing an increasingly more important role

Keywords: Computing methodology · Data analysis · Enterprise oper-

2.1 Computing Methodology

By providing the algorithms for management science and engineering manage-

Jardim and Silva presented a simple and computationally eﬃcient methodol-

2.2 Data Analysis

2.3 Enterprise Operation Management

2.4 Decision Support System

Decision support system (DSS) are computer-based information systems that

3 The Central Issues in the Proceedings Volume I

Fig. 1. Research topics in MS for the Eleventh ICMSEM

Fig. 2. Research topics in MS for the eleventh ICMSEM

4 Development Trends for MS and ICMSEM

4.1 CiteSpace Analysis Results

Fig. 3. Results of co-occurring keywords

4.2 Tendency in the ICMSEM Evaluation