ATLANTIS C OMPUTATIONAL I NTELLIGENCE S YSTEMS

VOLUME 7

S ERIES E DITORS : J IE L U, J AVIER M ONTERO

[email protected]
 Atlantis Computational Intelligence Systems

Series Editors:
Jie Lu
School of Software, Faculty of Engineering and Information Technology
University of Technology, Sydney, Australia

Javier Montero
Department of Statistics and Operational Research, Faculty of Mathematics
Complutense University of Madrid, Spain

(ISSN: 1875-7650)

Aims and scope of the series

The series ‘Atlantis Computational Intelligence Systems’ aims at covering state-of-the-

art research and development in all fields where computational intelligence is investigated
and applied. The series seeks to publish monographs and edited volumes on foundations
and new developments in the field of computational intelligence, including fundamental
and applied research as well as work describing new, emerging technologies originating
from computational intelligence research. Applied computational intelligence research
may range from computational intelligence applications in the industry to research projects
in the life sciences, including research in biology, physics, chemistry and the neurosciences.

All books in this series are co-published with Springer.

For more information on this series and our other book series, please visit our website at:

www.atlantis-press.com/publications/books

A MSTERDAM – PARIS – B EIJING


c ATLANTIS PRESS

[email protected]
Decision Aid Models for Disaster
Management and Emergencies
Begoña Vitoriano, Javier Montero
Universidad Complutense de Madrid
Facultad de Ciencias Matemáticas
Plaza de las Ciencias 3
28040 Madrid, Spain

Da Ruan †
Belgian Nuclear Research Centre (SCK • CEN)
Mol, Belgium

A MSTERDAM – PARIS – B EIJING

[email protected]
Atlantis Press
8, square des Bouleaux
75019 Paris, France

For information on all Atlantis Press publications, visit our website at: www.atlantis-press.com

Copyright
This book is published under the Creative Commons Attribution-Non-commercial license, meaning
that copying, distribution, transmitting and adapting the book is permitted, provided that this is done
for non-commercial purposes and that the book is attributed.
This book, or any parts thereof, may not be reproduced for commercial purposes in any form or by
any means, electronic or mechanical, including photocopying, recording or any information storage
and retrieval system known or to be invented, without prior permission from the Publisher.

Atlantis Computational Intelligence Systems

Volume 1: Linguistic Values Based Intelligent Information Processing: Theory, Methods, and
Applications - Da Ruan
Volume 2: Computational Intelligence in Complex Decision Systems - Da Ruan (Editor)
Volume 3: Intelligence for Nonlinear Dynamics and Synchronisation - K. Kyamakya, A.
Bouchachia, J.C. Chedjou
Volume 4: Trust Networks for Recommender Systems - P. Victor, C. Cornelis, M. De Cock
Volume 5: Answer Set Programming For Continuous Domains: A Fuzzy Logic Approach - J.
Janssen, S. Schockaert, D. Vermeir, M. De Cock
Volume 6: Computational Intelligence Systems in Industrial Engineering - C. Kahraman (Editor)

ISBNs
Print: 978-94-91216-73-2
E-Book: 978-94-91216-74-9
ISSN: 1875-7650


c 2013 ATLANTIS PRESS

[email protected]
†

Da Ruan unexpectedly passed away on July 31, 2011,

once this book was already launched. Born in Shang-
hai, September 10, 1960, Da Ruan got his Ph.D.
at the University of Ghent, Belgium, 1990, Etienne
Kerre being his Ph.D. advisor. He had been work-
ing at SCK • CEN, Belgium, since 1991, launching
in 1994 the FLINS conference series, and the ISKE
conference series in 2006. Da Ruan was the founder
of the International Journal of Computational In-
telligence Systems, currently the official journal of
European Society for Fuzzy Logic and Technologies
(EUSFLAT), and he has also served as Editor of this
Atlantis Press’ Computational Intelligence Systems
series. Da Ruan was the editor of 36 books and the
author of more than 270 papers in journals and con-
ferences. Leader of many fuzzy logic projects within
nuclear science, decision systems and risk analysis,
Da Ruan was a hard worker, helpful and inspira-
tive colleague, good friend and an honest and warm-
hearted person. His friends and his beloved family
will miss him forever.

[email protected]
Preface

Decision aid models for disaster management and

emergencies
Begoña Vitoriano 1,∗ , Javier Montero 1,2,∗ , and Da Ruan 3, †
1 Faculty of Mathematics Complutense University, Madrid 28040, Spain
2 Instituto de Geociencias IGEO (UCM,CSIC), Madrid 28040, Spain
3
Belgian Nuclear Research Centre (SCK • CEN), Mol, Belgium
∗ {bvitoriano,javier_montero}@mat.ucm.es

A disaster is understood as the disruption of the normal functioning of a system or commu-

nity, which causes a strong impact on people, structures and environment, and goes beyond
local capacity of response. In the global world of the 21st century, a disaster that occurs
anywhere of the planet affects to the whole human Society. All regions on The Earth are
interconnected and any serious problem in any region will not only have direct and terrible
consequences on that region, but it will also hit the social and/or economic structure of
other countries. The dimension and circumstances of each disaster will determine the relief
level needed (regional, national or international), but anyway they require a joint provision
strategy of many regions and the concourse of International Organizations, Governmental
and Non-Governmental.
Disaster management implies a number of actions to be developed before disaster happens,
during the disaster and after disaster takes place. Authorities at all levels as well many
businesses create some sort of preventive measures and plans to reduce the disaster effects
and return to normal function as quickly as possible, and from the experience learn how to
improve prevention and reaction.
† Prof. Da Ruan unexpectedly passed away on 31 of July 2011 after launching this book.

vii

[email protected]
viii Decision Aid Models for Disaster Management and Emergencies

Response to natural disasters (e.g., floods, earthquakes) or technological (e.g., nuclear,

chemical) is an extreme complex process that involves severe time pressure, various kind
of uncertainties, and many stakeholders. Disaster management often requires several au-
tonomous agencies to collaboratively mitigate, prepare, respond, and recover from hetero-
geneous and dynamic sets of hazards to Society. Almost all disasters involve high degrees
of novelty to deal with most unexpected various uncertainties and dynamic time pressures.
Existing studies and approaches within disaster management have mainly been focused on
some specific type of disasters with certain agency oriented. There is a lack of a general
framework to deal with similarities and synergies among different disasters by taking their
specific features into account.
This book provides with various decision analysis theories and support tools in complex
systems in general and in disaster management in particular. The book is also generated
during a long-term preparation of a European proposal among most leading experts in the
areas related to the book title.
After a general call among specialists and a standard peer review process, we sincerely
hope that this book becomes a reference in the field since it addresses not only a state-of-
the-art first section and third section devoted to original application oriented advances, but
also a second section that offers alternative approaches to deal with the different kinds of
uncertainties that appear in disaster management and emergencies.
This book starts with an introduction to the important role of information on disaster man-
agement and emergencies by A. Pedraza. In this introduction, some main challenges in
humanitarian logistics are described. Among other key issues, the author points out the
collaboration needed between different parties collaborating in disaster response and the
need of incorporating dynamic forecast of disasters in decision aid models. Three spe-
cific examples involving the use of information in humanitarian operations are discussed
more in detail, involving techniques from operations research, operations management, and
management science.
The first section of this book, devoted to an overview of the state-of-the-art, contains four
papers. In the first paper, M.T. Ortuño et al. state main definitions concerning disasters,
emergencies and humanitarian logistics, and focus the review on decision aid models and
systems developed until now in the area. In the second paper, F. Liberatore et al. offer a
general overview of the different problems that become a source for uncertainty in human-
itarian logistics in the context of disaster management, as well as the main techniques used
until now (risk mappings, probability, stochastic programming, optimization, simulation

[email protected]
Preface ix

and fuzzy models), offering a brief description of main references. In the third paper, B.
Öztaysi et al. focus on Fuzzy Inference Systems and their applications for disaster manage-
ment, showing an example in order to allow a quick and flexible solution for spontaneous
volunteer management. In the fourth paper of this section, D. Tang et al. present a litera-
ture review in security based operations, focussing in current and expected future research
in container line supply chains divided into three categories (general regulations, specific
security issues and risk analysis tools).
The second section presents four different alternatives to deal with uncertainties in disas-
ter management and emergencies. In the first paper of this section, L. Zou et al. present
two approaches for fuzzy risk analysis in order to handle both comparable and incompa-
rable linguistic information by means of information aggregation techniques and taking
into account 10 linguistic evaluation values. In a second paper, J. Liu et al. offer a belief
rule-based generic risk assessment framework for modelling, analyzing and synthesizing
risk-related information with various uncertainties, where risk factors are described using
linguistic variables in order to capture uncertain casual relationship between the risk factors
and the special risk estimate. In a third paper, M.A. Abchir et al. stresses again that re-
port and descriptions of the disaster that will support emergency management in their first
stages are usually expressed linguistically by witnesses, so the authors offer a procedure to
help to express diagnosis of the catastrophic event based upon such a natural language. In
the fourth paper, in order to deal the lack of information of decision makers, H. Bustince
et al. present a new method to amalgamate the opinions of several experts based upon
interval-valued information.
Finally, last section in devoted to applications, containing this time five papers. Since a
precise numerical evaluation of a potential disaster is unrealistic, the first paper by J.T.
Rodríguez et al. reformulates this problem as a severity classification procedure, offering
a bipolar classification methodology based upon the notions of semantic antagonism and
dissimilarity structure. In a second paper, the problem of transhipment multi-commodity
supply chain flow for humanitarian relief operations is addressed by A. Clark et al., of-
fering a model to assist policy makers and planners to design an effective supply chain so
that goods and service can reach those in need as quickly as possible, a model validated
with real life data from the South Asian Earthquake of October 2005. In the third paper
of this section, M. Leon et al. take advantage of Fuzzy Cognitive Maps to represent the
behaviour and operations of transport managing, showing how travellers base their deci-
sions on their perception of available information. A hierarchical assessment framework

[email protected]
x Decision Aid Models for Disaster Management and Emergencies

for evaluating safety against fire and explosion hazards in container line supply chains is
presented by Y.W. Chen et al., following a Evidential Reasoning approach which is tested
data collected from the Port of Liverpool. Finally, M. Naderpour et al. develop a Human
Situation Awareness Support System which has the ability to support the operators’ under-
standing and assessing the disaster, and by applying fuzzy risk assessment concepts allow
a better inference of the real situation and project its status in the near future.
We have to finish this preface acknowledging the key role that our missed friend Da Ruan
played in the conception of this book. In fact, the idea of this book emerged while he
was spending two months as Distinguished Visitor at Complutense of Madrid, invited by
B. Vitoriano within a program that pretends to bring to this University extremely qualified
researchers with leadership to launch new projects. Da Ruan perfectly fitted this descrip-
tion, and devoted a long time discussing with us ideas about common fields of interest. This
book is one of the projects we discussed, launched while he was still in Madrid, focussed on
researchers and practitioners demanding new approaches, particularly those coming from
Non Governmental Organizations. Along his short life, Da showed a tremendous capabil-
ity to launch new projects, putting together people from many different fields, either at the
academy or at the industry. Da was a taught worker, a kind person and a friend we could
always count on. We miss him.
We deeply appreciate the support of members of Spanish Red Cross, Action Against Hun-
gry and other Non Governmental Organizations that have been collaborating with us. We
hope that this book will be helpful for their objectives and bring more researchers into this
complex and multi facet field by means of a book that intends to become a reference text
for research and practice within disaster management and emergencies.

[email protected]
Contents

Preface vii
B. Vitoriano, J. Montero, and D. Ruan

1. On the Use of Information in Humanitarian Operations 1

A.J. Pedraza-Martinez

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Levels of Disaster Response . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Disaster Preparedness at the National Level: Centers for Environmental
Prediction (NCEP) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Disaster Preparedness and Response: Fleet Management in the Humani-
tarian Sector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.5 Disaster Response at the International Level: ReliefWeb . . . . . . . . . 11
1.6 Conclusions and Suggested Areas for Future Research . . . . . . . . . . 13
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2. Decision Aid Models and Systems for Humanitarian Logistics. A Survey 17

M.T. Ortuño, P. Cristóbal, J.M. Ferrer, F.J. Martín-Campo, S. Muñoz,

G. Tirado, B. Vitoriano

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 Decision Aid Models. Literature Review . . . . . . . . . . . . . . . . . . 23
2.3 Decision Aid Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

xi

[email protected]
xii Decision Aid Models for Disaster Management and Emergencies

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

3. Uncertainty in Humanitarian Logistics for Disaster Management.

A Review 45

F. Liberatore, C. Pizarro, C. Simón de Blas, M.T. Ortuño, and B. Vitoriano

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.2 Uncertainty in disaster management . . . . . . . . . . . . . . . . . . . . 48
3.3 Methodologies Description . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.4 Annotated Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

4. Fuzzy inference systems for disaster response 75

B. Öztayşi, H. Behret, Ö. Kabak, İ. Uçal Sarı, and C. Kahraman

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.2 Disaster Response System . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.3 Fuzzy Inference Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.4 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.5 A Fuzzy Inference System for Spontaneous Volunteer Management . . . 87
4.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

5. Security Based Operation in Container Line Supply Chain 95

D. Tang, D.-L. Xu, J.-B. Yang, and Y.-w. Chen

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.2 Basic Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.3 Current research on security issues in CLSC . . . . . . . . . . . . . . . . 98
5.4 Summary and limitations of current research and potential directions for
future research for CLSC security . . . . . . . . . . . . . . . . . . . . . 109
5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

[email protected]
Contents xiii

6. A Linguistic-Valued Information Processing Method for Fuzzy

Risk Analysis 117

L. Zou, Z. Pei, D. Ruan, and Y. Xu

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

6.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
6.3 Risk Analysis Based on Linguistic Aggregation . . . . . . . . . . . . . . 128
6.4 Risk Analysis Based on L10 . . . . . . . . . . . . . . . . . . . . . . . . . 137
6.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

7. A Belief Rule-Based Generic Risk Assessment Framework 145

J. Liu, S. Chen, L. Martinez, and H. Wang

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145

7.2 Hazard Associated Risk Specification Parameters . . . . . . . . . . . . . 149
7.3 A Belief-Rule Based Generic Hazard Risk Assessment Framework . . . . 151
7.4 A Belief-Rule Based Generic Hierarchical Risk Assessment Framework . 159
7.5 Conclusions and Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . 165
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166

8. Fuzzy semantics in closed domain question answering 171

M.-A. Abchir, I. Truck, and A. Pappa

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171

8.2 Disaster or Emergency Management . . . . . . . . . . . . . . . . . . . . 172
8.3 Using Natural Language Processing (NLP) Techniques . . . . . . . . . . 176
8.4 How to deal with imprecision in our context? . . . . . . . . . . . . . . . 178
8.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187

9. Decision Making with Extensions of Fuzzy Sets 189

H. Bustince, M. Pagola, E. Barrenechea, C. Lopez-Molina, J. Fernandez

9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189

9.2 Preliminary concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191

[email protected]
xiv Decision Aid Models for Disaster Management and Emergencies

9.3 Construction of Interval-valued Fuzzy Preference Relations from Fuzzy

Preference Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
9.4 Two generalizations of the weighted voting strategy . . . . . . . . . . . . 196
9.5 Two examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
9.6 Advantages of the method . . . . . . . . . . . . . . . . . . . . . . . . . 204
9.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206

10. Classification of Disasters and Emergencies under Bipolar

Knowledge Representation 209

J.T. Rodríguez, B. Vitoriano, D. Gómez, J. Montero

10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209

10.2 Problem description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
10.3 Strategic disaster severity assessment as an structured classification
problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
10.4 SEDD’s bipolar classification methodology . . . . . . . . . . . . . . . . 220
10.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231

11. A Network Transshipment Model for Planning Humanitarian

Relief Operations after a Natural Disaster 233

A. Clark and B. Culkin

11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233

11.2 Planning For Humanitarian Relief Operations . . . . . . . . . . . . . . . 235
11.3 Models Developed by other Researchers . . . . . . . . . . . . . . . . . . 237
11.4 A Network Transshipment model for Emergency Relief . . . . . . . . . . 238
11.5 Computational Analysis and Model Refinement . . . . . . . . . . . . . . 244
11.6 Recommendations and Conclusions . . . . . . . . . . . . . . . . . . . . 254
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255

[email protected]
Contents xv

12. A Travel Behaviour Study through Learning and Clustering of

Fuzzy Cognitive Maps 259

M. León, L. Mkrtchyan, B. Depaire, D. Ruan, and K. Vanhoof

12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259

12.2 Modelling with Fuzzy Cognitive Maps . . . . . . . . . . . . . . . . . . . 263
12.3 Capturing individuals’ mental representation: Automatic knowledge
engineering and FCM modelling . . . . . . . . . . . . . . . . . . . . . . 270
12.4 A learning method for FCMs inspired on Particle Swarm Optimization . . 273
12.5 Clustering of Fuzzy Cognitive Maps . . . . . . . . . . . . . . . . . . . . 279
12.6 Clustering and learning: digging deeper into individuals mental maps . . 285
12.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291

13. Fire and explosion safety assessment in container line supply chain 295

Y.-w. Chen, D.-L. Xu, J.-B. Yang, and D. Tang

13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295

13.2 Review of safety-related regulations and assessment approaches . . . . . 297
13.3 Fire and explosion safety assessment . . . . . . . . . . . . . . . . . . . . 298
13.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305

14. A Human Situation Awareness Support System to Avoid

Technological Disasters 307

M. Naderpour and J. Lu

14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307

14.2 Basic Concepts and Related Works . . . . . . . . . . . . . . . . . . . . . 309
14.3 A Human Situation Awareness Support System . . . . . . . . . . . . . . 313
14.4 HSASS Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 315
14.5 Conclusion and Future Study . . . . . . . . . . . . . . . . . . . . . . . . 324
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324

[email protected]
Chapter 1

Introduction
On the Use of Information in Humanitarian
Operations
Alfonso J. Pedraza-Martinez
Department of Operations and Decision Technologies, Kelley School of Business, Indiana
University, 1309 E. 10th Street, Bloomington IN 47405, USA

E-mail: [email protected]

1.1 Introduction

Obtaining reliable information for decision-making to prepare and respond to disasters

is difficult. Governments and humanitarian organizations often have to plan complex disas-
ter responses based on forecasts or without having reliable field assessments. Decision aid
models can help governments and humanitarian organizations by improving their forecasts
during disaster preparedness and by facilitating data collection during field assessments in
disaster response. Decision aid models can also facilitate collaboration and coordination
between different parties involved in disaster management. This book is an academic study
of decision aid models for disaster management.
The interest of academics in disaster management and humanitarian logistics has in-
creased significantly over the last decade. This is related to the fact that many of the
challenges humanitarian systems face during disaster response are related to operations
research, operations management and management science. For example, disaster man-
agement deals with scarce resources. This is because the impact of disasters on human life
seems to be increasing at a higher rate than the resources dedicated to disaster preparedness
and response. Also, high levels of uncertainty following disasters impose additional pres-
sure on the use of available resources. The complex challenges of disaster management
have motivated prominent academic institutions like Universidad Complutense in Spain,

B. Vitoriano et al. (eds.), Decision Aid Models for Disaster Management 1

and Emergencies, Atlantis Computational Intelligence Systems 7,
DOI: 10.2991/978-94-91216-74-9_1, Ó Atlantis Press 2013

[email protected]
2 Decision Aid Models for Disaster Management and Emergencies

INSEAD in France and Cranfield in England as well as Tufts University, Georgia Tech and
Rensselaer Polytechnic Institute in the United States, to create research groups aimed at
improving the capacity of humanitarian systems for disaster management.
Academic journals and societies have also exhibited increased interest to topics of dis-
aster management and humanitarian logistics. Journals like the European Journal of Oper-
ational Research and OR Spectrum have dedicated special issues to the topic. In addition,
The Journal of Humanitarian Logistics and Supply Chain Management published its first
issue in 2011. The journal is targeted to academics and practitioners in humanitarian,
public and private sector organizations working on all aspects of humanitarian logistics
and supply chain management (emeraldinsight.com). The remarkable increase of academic
production in humanitarian operations makes literature reviews like the ones included in the
first part of this book timely.
Additionally, the Production and Operations Management Society (POMS) founded
the College of Humanitarian Operations and Crisis Management (HOCM) in 2011. The
establishment of the HOCM College at POMS was followed by a call for papers for a
special issue on humanitarian operations in Production and Operations Management, still
under review when this article was written. Another important milestone of the HOCM
College was its first mini-conference, held in Chicago, USA in April 2012.
Plenary sessions at the HOCM mini-conference included high-level practitioners and
academics. Dr. Louis Uccellini, the President of the National Centers of Environmental
Prediction of the United States – and president of the American Meteorological society –,
spoke about the evolution of weather forecast and how accurate information on weather
prediction helps reducing disaster uncertainty. Professor Peter Walker, the president of
the International Humanitarian Studies Association, and Professor Luk Van Wassenhove,
President of POMS and INSEAD Humanitarian Research Group, also commented on the
value of information and pointed out that there is a need for inter-sector collaboration and
interdisciplinary research in the area of disaster management. We elaborate on the needs
identified by the HOCM mini-conference and discuss various examples involving the use of
information in humanitarian operations. The discussion is preceded by a short description
of different levels of disaster response used as a framework in the examples.

[email protected]
On the Use of Information in Humanitarian Operations 3

1.2 Levels of Disaster Response

The disaster management cycle includes at least four stages: preparedness, response,
rehabilitation and mitigation [1, 2]. Based on the objectives, uncertainty and duration, the
former two stages can be classified as ‘relief’ operations while the latter two stages can
be classified as ‘development’ operations. Pedraza-Martinez et al. [3] discuss that relief
and development operations result in different logistics challenges. On the one hand, relief
operations focus on life-saving objectives via first aid, evacuation, food and water distri-
bution, and shelter provision, subject to a budget constraint. Relief operations also face
high uncertainty regarding supply and demand. Demand uncertainty results from the lack
of information regarding the impact of the disaster in the affected population. Supply un-
certainty has different sources including: (i) destruction of local inventories and services
during the disaster, and (ii) difficulties to access the disaster area due to blockages. Fi-
nally, relief operations also have short duration that typically lasts up to three months. The
duration of operations can be extended up to six months like the one following the Haiti
earthquake in 2010.
On the other hand, development operations usually focus on minimizing a cost function,
subject to minimum acceptable service levels to meet demand. Development operations
also face demand uncertainty, but this uncertainty is lower than that faced by relief opera-
tions. Demand uncertainty may result from the mobility of potential beneficiaries, weather
conditions or localized security problems. Development operations may face some supply
uncertainty but it is typically negligible when compared to the one faced in relief opera-
tions. Supply uncertainty may result from lack of local inventory. Finally, development
operations typically have a longer duration, which can range from six months to several
years. Despite the differences in terms of objective, uncertainty and duration, relief and
development operations often share the same logistics networks. The multi-objective prob-
lem resulting from simultaneous relief and development operations is quite challenging for
the humanitarian sector [4]. Nevertheless, this problem has not been sufficiently studied by
academics.
Disaster response usually occurs on three levels: local, national, and international (Fig-
ure 1.1). Local response begins immediately after the disastrous event. Operational staff in
charge of the local response system coordinate search and rescue activities [5]. Sometimes
international humanitarian organizations have presence in the area of disaster and take part
in the local response. For example, Medecins Sans Frontiers set up an open-air hospital to
treat emergencies the day following the 2010 earthquake in Haiti.

[email protected]
4 Decision Aid Models for Disaster Management and Emergencies

If the magnitude of the disaster overwhelms local capacity then the national system
is activated. The national system usually includes donors of cash and basic items like
medicines, water, food, and shelter. It also includes logistics and heavy equipment for
debris removal. The national system often faces difficulties in delivering supplies in the
affected area, which causes a response delay. Depending on the magnitude of the disas-
ter, the national system may not have enough capacity to offer a complete response. In
such scenario the national government is responsible for requesting the activation of the
international system.

United Nations
International Agencies,
International
Initial response. Organizations,
National National Non-
Government, Governmental
Immediate response.
Local Surviving Local
Military Forces, Organizations,
National Healthcare International
Government, Police,
Officers, Civil Military Forces,
Healthcare Officers,
Defense, Red Cross Donors
Fire Department, Red
National Society,
Cross, NGO
NGO, Donors

Fig. 1.1 Different Operational Levels of Disaster Response

The international system is composed primarily of donors, international humanitarian

organizations and NGO—although in recent operations like the Haiti response in 2010 the
military forces of donor countries like the United States have also supported the system.
The first issue faced by the international system is access to information regarding the disas-
ter. When international humanitarian organizations have existing programs in the disaster
area, they may receive field information within minutes following the event. Otherwise,
they must rely on their forecasts, previous experience, the media and other organizations
until their own disaster response teams are able to reach the affected area and send infor-
mation back to the headquarters.
There is a number of coordination challenges within each level of disaster management.
Sometimes different parties involved in disaster management have different concerns. Take
the example of response operations involving the international level. As mentioned above,
these operations include donors and humanitarian organizations. While donors prefer to

[email protected]
On the Use of Information in Humanitarian Operations 5

earmark their funds in order to increase the accountability of their donations, interna-
tional humanitarian organizations would prefer no-earmarked funds to avoid exogenous
constraints—imposed by donors—to the use of resources. The coordination of donors and
humanitarian organizations is an interesting problem that has not been investigated suffi-
ciently.
There are also coordination challenges between different levels of disaster response.
Different response levels may have distinct opinions while participating in disaster response
operations. Take the example of the earthquake in Colombia, in 1999. The immediate
response to that disaster occurred at the local level. When the national level got involved
it intended to take control of the operation, but the local level was reluctant to give up
control to the national level arguing that they—the locals—had been directly affected by
the disaster and had better knowledge of the field [5]. Achieving coordination between
local and national levels consumed precious time that could have been used by staff to
reach the demand of humanitarian aid.
Additionally, there are third parties that do not participate but support disaster response
operations. Commercial suppliers and academics collaborate during different stages of the
disaster management cycle. Nevertheless, the objectives of these parties may not be aligned
with the ones of the parties directly involved in disaster response. For example, humanitar-
ian staff working on disaster response have very tight time constraints and must prioritize
different tasks. As expected, they allocate most of their time to life-saving activities and
sacrifice less important activities like data collection. While data collection may be sec-
ondary to humanitarian staff during disaster response, it is a critical input for decision aid
models developed by academics.
In the following sections we discuss three examples on the use of information in human-
itarian operations. These examples are presented using the framework introduced above.
We focus on examples that involve national and international levels of response instead
of those involving local response. The first example is on the National Centers for Envi-
ronmental Prediction (NCEP) in the United States and their remarkable work on weather
forecasting. This example refers to preparedness for disaster response at the national level
(Figure 1.2).

[email protected]
6 Decision Aid Models for Disaster Management and Emergencies

Level of response:

International
Fleet ReliefWeb
Management
National NCEP

Preparedness Response
Stage in the disaster management cycle

Fig. 1.2 Classification of examples according to level of response and stage in the disaster manage-
ment cycle

The second example is on collaboration between humanitarian organizations and aca-

demics to improve vehicle fleet management in humanitarian operations. This example
touches on preparedness at the international level and response at the national level. The
third example is on Reliefweb and its role in information dissemination. ReliefWeb is an
information-sharing platform primarily used by the international level of disaster response.
The ReliefWeb example describes an initiative to improve access to information for disaster
response at the international level.

1.3 Disaster Preparedness at the National Level: Centers for Environmental

Prediction (NCEP)1

Weather related events like floods, storms, cold waves and wildfires counted for 71% of
the total number of disasters covered by ReliefWeb in 2010. Fortunately, weather related
events can be predicted with enough accuracy to minimize their impact on human life [6].
The National Centers for Environmental Prediction (NCEP) are responsible for forecasting
weather related events in the United States. As part of their mission, NCEP “deliver na-
tional and global weather, water, climate and space weather guidance, forecasts, warnings
and analyses to its Partners and External User Communities.” (ncep.noaa.gov). NCEP
is composed of 9 centers: Aviation Weather Center (AWC), Climate Prediction Center
(CPC), Environmental Modeling Center (EMC), Hydrometeorological Prediction Center
(HPC), NCEP Central Operations (NCO), National Hurricane Center (NHC), Ocean Pre-
1 This example is based on the presentation given by Dr. Louis Uccellini, Director of the National Centers

for Environmental Prediction at the first Mini-Conference of the College of Humanitarian Operations and Crisis
Management, at the Production and Operations Management Society (POMS), 2012.

[email protected]
On the Use of Information in Humanitarian Operations 7

diction Center (OPC), Space Weather Prediction Center (SWPC), and Storm Prediction
Center (SPC). NCEP produce the majority of the weather related information in the United
States.
Weather forecasting is based on four components: Global Observing System, Com-
puters, Data Assimilation & Modeling/Science, and forecasters [7]. The global observing
system has evolved substantially during the last few decades. Kistler et al. [8] divide the
evolution of the global observing system into three phases: (i) early period, from the 1940s
to 1957, when the first upper-air observations were made and numerical models were in-
troduced; (ii) modern rawinsonde period, from 1958 to 1978; (iii) modern satellite period,
from 1979 to the present, when real time numerical prediction models were introduced. The
modern satellite based system allows forecasters to obtain accurate predictions of storms
several days in advance. Based on millions of data records, forecasters use mathematical
and numerical models to estimate possible duration, trajectories and intensity of weather
related events. The improvement in forecast models has positively impacted disaster re-
sponse in the United States. While in the 1970’s accurate forecasts were only available
hours before the event, in 2010 accurate forecasts allowed decision-making several days
before the potential disaster occurs.
NCEP participate actively in disaster preparedness and response at the national level
in the United States. From February 4 to 7, 2010 a major snowstorm affected the west
cost of the country. It produced up to 75 cm of snow in some areas of Maryland. From
February 9 to 11 the storm was followed by a blizzard, which produced up to 60 cm of
snow in northern Maryland. According to Reuters [9] it could have been the worst storm
in Washington in 90 years. President Barack Obama named the storm “snowmageddon”.
NCEP predicted the storm system more than seven days in advance. Accurate warnings
were emitted by NCEP three to five days in advance, depending on the geographic area.
As a result of NCEP early alarms, the affected States had time to prepare for the storm.
Following NCEP recommendations, these States implemented contingency plans for disas-
ter response. Furthermore, NCEP information helped airlines to decide the cancellation of
flights well in advance preventing thousands of passengers from getting stranded at airports.
Regarding supply chain management, NCEP weather forecast allowed retailers to preposi-
tion inventory to face the potential disruption. This was particularly important because the
storm coincided with the “Super Bowl”, the American football match to be played on Sun-
day, February 7, 2010. Retailers knew several days in advance that they had to advertise
“pre-storm” sales because mobility problems could deter consumers from purchasing dur-

[email protected]
8 Decision Aid Models for Disaster Management and Emergencies

ing the day of the football match. Due to the joint action of NCEP and State governments,
2010 snowmageddon did not have a major impact on human lives, travel or gross national
product, as opposed to other major snowstorms during the nineties [7].
NCEP do a remarkable job on predicting weather related events. This example shows
how the reliable and accurate early warnings produced by NCEP helped to prevent a major
disaster. Incorporating early warnings and forecasts into decision aid models is an interest-
ing research area that has not received enough attention from operations research, opera-
tions management and management science academic communities. Although Ragnier [6]
analyzed the impact of accurate disaster forecasts on evacuation decisions, the effect of
improving forecasts on areas of research like facility location, inventory management, and
transportation has not been widely explored yet.

1.4 Disaster Preparedness and Response: Fleet Management in the Humani-

tarian Sector

Fleet management in humanitarian operations is a complex set of tasks that involves

multiple parties at local, national, and international levels. Four-wheel drive (4DW) vehi-
cles are at the core of humanitarian fleets. Although it would be difficult to determine the
exact figure due to problems in data collection, Fleet Forum2 has estimated that the inter-
national fleet of 4WD vehicles has approximately 80,000 units. These vehicles are used to
support the transportation needs of humanitarian staff in disaster response, rehabilitation
and mitigation [10].
For almost 10 years Fleet Forum work has focused on developing solutions to fleet
management problems regarding efficiency, safety and environmental impact in the human-
itarian sector. In 2007 several Fleet Forum members agreed that some of the symptoms of
their problems were excessive fleet sizes, ageing fleets, lack of fleet standardization, delays
in disaster response and accidents. These organizations invited INSEAD Humanitarian Re-
search Group (HRG) to join them to collaborate in the search for solutions3. In 2007 HRG
began a long-term project with the aim of understanding how international humanitarian
organizations manage their vehicle fleets in field operations.
Due to the lack of academic references in the area of humanitarian fleet management,
the research design was built upon exploratory case studies. These cases included World
2 Fleet Forum is an organization of more than forty members, including NGOs, international organizations, the

UN, academic institutions, donors, and corporate partners (fleetforum.org, Tomasini et al. 2006) [11],
3 Professor Luk Van Wassenhove created HRG in 2001 as a platform for cross-disciplinary research and engage-

ment in humanitarian logistics

[email protected]
On the Use of Information in Humanitarian Operations 9

Vision International (WVI), World Food Programme (WFP), International Federation of

Red Cross and Red Crescent Societies (IFRC), and International Committee of the Red
Cross (ICRC). Interviews at local, national and international levels where conducted in
Africa, the Middle East and Europe with collaboration of the four organizations. Based on
the analysis of these interviews the research team proposed a series of hypotheses explain-
ing the main drivers of problems in humanitarian fleet management. These drivers included
the simultaneous objectives of disaster response (relief) and rehabilitation and mitigation
(development), difficult operating conditions, conflicting objectives at the different levels
of disaster response, decentralization, and earmarked funding [10].

Research Problem Dissemination to

practitioners:
ICRC, IFRC, WFP, WVI fleet management cases Fleet Forum Conference
Case based research: Exploratory cases (FFC) 2008

Vehicle replacement at the ICRC

Empirical research: Econometrics, Dynamic Programming FFC 2007, 2008, 2010;
managerial articles:
International Aid and
Fleet size Trade Review, Car Nation
Operations Research: Optimization

Fleet management structure

Simulation: System Dynamics FFC 2010

Defining Characteristics FFC 2012, Aid and

International
of Humanitarian Fleet Management Development Forum 2012

Fig. 1.3 Research Process on Fleet Management in Humanitarian Operations followed by INSEAD
Humanitarian Research Group (HRG)

The results of the case-based research were presented at the Fleet Forum Conference
in 2008 (Figure 1.3). At the conference humanitarian logisticians provided insightful feed-
back to the research team. Feedback from practitioners helped the researchers to refine the
preliminary findings. Practitioners also agreed to provide more information both qualita-
tive and quantitative to complement the initial findings. Equally important, humanitarian

[email protected]
10 Decision Aid Models for Disaster Management and Emergencies

logisticians agreed that the research added value to their job since academics were present-
ing a synthesis of fleet management challenges and their causes in a format that could be
clearly understood.
Parallel to the qualitative case studies, HRG team collected data on vehicle use from
ICRC fleet management system, FleetWave, a web-based enterprise fleet management so-
lution used by commercial companies like PaperCo and Coca Cola (chevinfleet.com). The
ICRC data was used to test a policy suggested by the ICRC international level to the na-
tional delegations in more than 80 countries. The ICRC policy was to replace vehicles every
150,000 km or 5 years, whichever came first. Vehicle manufacturers like Toyota and Nis-
san recommended this replacement policy to business using their products. Nevertheless,
ICRC had some special operating characteristics: difficult operating conditions and vehicle
purchasing discounts due to its humanitarian orientation. Using econometrics models and
dynamic programming, the research found that the optimal policy for the ICRC would be
to replace their vehicles every 100,000 km or its equivalent in age [12]. Difficult operating
conditions resulted in a relatively greater maintenance cost compared to commercial fleets
while purchasing discounts resulted in a lower cost of capital per vehicle.
The same research found that ICRC national delegations did not follow the 150,000
km or 5 years policy. This was explained by a misalignment of incentives between inter-
national headquarters and national delegations caused by earmarked funding for vehicle
procurement. The analysis on vehicle replacement was complemented with optimization
models to determine the optimal fleet size to support the transportation needs of programs
for rehabilitation and mitigation in cases of demand uncertainty. The results of quantitative
research boosted the interest of humanitarian logisticians in collecting and analyzing data
for decision-making purposes.
Next, the fleet management descriptions from case studies were combined with the
quantitative data on vehicle use from ICRC in a simulation-based research. The purpose
was to obtain further insights on decentralization and the effect of earmarked funding on
humanitarian fleet management. The research found that when decentralization and ear-
marked funding are present simultaneously humanitarian organizations using national ve-
hicle procurement might take longer to react to small and medium size disasters than those
using global procurement with long lead-times [4]. This is explained by the fact that ear-
marked funding creates resource allocation constraints that deter the deployment of the
fleet in a timely fashion.

[email protected]
On the Use of Information in Humanitarian Operations 11

Humanitarian logisticians have realized that there is value in carefully collecting oper-
ational data regarding the use of their fleets. They have also seen that better decisions can
be made when data is rigorously transformed in information. Due to their increasing in-
terest in gathering information regarding their fleets, some senior humanitarian logisticians
started pilot projects to install satellite-tracking technologies in their fleets. These tech-
nologies will allow logisticians to collect data on vehicle use in real time and react faster to
disruptions in transportation following disasters. Additionally, this data will facilitate more
sophisticated analyses on fleet management using optimization techniques.

1.5 Disaster Response at the International Level: ReliefWeb

Created in 1996 by the UN Office for the Coordination of Humanitarian Affairs

(OCHA), ReliefWeb aims “to be the world’s leading on-line gateway to information on
humanitarian emergencies and disasters” [13]. ReliefWeb was part of a humanitarian sec-
tor initiative in the early 1990s to improve the dissemination of information for decision-
making in the humanitarian sector. Another initiative that aimed the same objective was
the United Nations Joint Logistics Centre (UNJLC) Website [14, 15]. While the UNJLC
Website specialized to become the logistics cluster website, ReliefWeb consolidated as a
web platform offering multiple services beyond logistics to the humanitarian community.
Originally ReliefWeb had offices in Kobe, Geneva and New York to update content in
real time. Regarding cost, ReliefWeb’s total budget in 1999 was $1.14 million while in
2005 the budget was $2 million, supported entirely by donations. In 2011 the offices in
Kobe and Geneva were moved to the humanitarian hubs of Bangkok and Nairobi, which
allowed them to lower the operating cost without sacrificing the 24-hours a day coverage.
Every day the globally located ReliefWeb team contacts different sources of information
and searches the web to identify the most relevant humanitarian news, reports, maps, anal-
ysis and other content [16]. These documents are filtered to avoid duplication and posted
on the website. ReliefWeb services include posting vacancies, reports, emergency updates,
and a map center. Regarding emergency updates, ReliefWeb covered 182 disasters in 2011,
and covered 157 disasters in 2010 (Table 1.1). In 2011 the site received 9.3 million unique
visits [16], and published over 55,000 posts including documents, vacancies and maps.
Not surprisingly, the traffic of the website peaks in the aftermath of disasters. Following
the Asian tsunami in 2004 ReliefWeb received an average of 3 million hits per day.

[email protected]
12 Decision Aid Models for Disaster Management and Emergencies

Table 1.1 ReliefWeb coverage by disaster type 2008-2010.*

New disasters by type 2008 2009 2010
Flood 47 85 72
Storm* 33 30 29
Earthquake 13 21 17
Landslide & mudslide 11 16 11
Cold Wave 3 6 6
Epidemic 0 9 6
Volcano 6 6 6
Wildfire 0 0 5
Other 11 20 5
Total 124 193 157
* Storm includes local storm, tropical cyclone, hurricane and storm surge.
Source: ReliefWeb Annual Statistics [17–19].

ReliefWeb has global reach consisting of visits from Europe (29%), Africa (23.7%),
North America (19.4%) and Asia (18.3%). Geographic regions with lower percentages of
visits are Latin America and Oceania. The top five countries in number of visits during
2010 were: USA, UK, Kenya, Canada, and France. One of the main criticisms leveled
at ReliefWeb following an evaluation carried out in 2005 was the lack of access to the
local level of response [12], which was partially due to language problems. One of the
suggestions from users was to include content in other languages than English. In June
2012 content was available in English (399,799 posts), French (36,803 posts), Spanish
(21,363 posts), and Arabic (113 posts).
The sources of postings have evolved in the last few years. In 2005 the top five sources
were media outlets: Agence Fance-Presse, Integrated Regional Information Networks,
Reuters, Xinhua, and UN News Service. In 2011 the top five sources were: Agence France-
Pressse, Integrated Regional Information Networks, OCHA, Reuters, and the International
Federation of Red Cross and Red Crescent Societies (IFRC). The evolution of the top-five
source composition suggests that ReliefWeb has earned a high degree of trust from human-
itarian organizations as a platform for information sharing.
During the first month of the Haiti response in 2010 ReliefWeb posted daily situation
reports from organizations such as OCHA and IFRC. The first OCHA report was posted on
the day of the earthquake. The report included contact details of ReliefWeb regarding dona-
tions: “All humanitarian partners including donors and recipient agencies are encouraged

[email protected]
On the Use of Information in Humanitarian Operations 13

to inform FTS of cash and in-kind contributions by sending an email to: [email protected].”
The OCHA report also directed readers to ReliefWeb in order to get more information
about the disaster.
The example of ReliefWeb shows an effort to disseminate information in the aftermath
of disasters. ReliefWeb collects, filters, and publishes information related to disaster re-
sponse even the same day that disasters occur. That information is public and can be easily
accessed through their website. Nevertheless, ReliefWeb usually does not provide further
analysis to support decision-making. It would be very interesting to see alliances between
information-sharing platforms like ReliefWeb and academic researchers to incorporate real
time analysis based on decision aid models in web reports.

1.6 Conclusions and Suggested Areas for Future Research

This introduction discusses three examples of information on disaster management. In

doing so, it has also mentions some interesting research challenges in the areas of oper-
ations research, operations management, and management science. First, it mentions the
humanitarian logistics challenges resulting from simultaneous relief and development op-
erations. Second, it elaborates on the practical problem of conflicting objectives between
different parties collaborating in disaster response. Third, it comments on the need for coor-
dination between donors and international humanitarian organizations. Fourth, it discusses
the need of incorporating dynamic forecast of disasters in decision aid models. Fifth, it
comments on the possibility of producing real time analysis based on available informa-
tion using decision aid models. The topics mentioned in this introduction are only some
potentially interesting areas for further research.
The use of information and decision aid models to support disaster management is a
relevant research area. This book provides new insights on the use of information as well
valuable decision aid models to support disaster preparedness and response. The book
comes at a crucial moment considering the growing body of research in disaster manage-
ment and humanitarian logistics in the past few years. Apart from the literature surveys
included in the first part of the book, the second part presents a selection of decision aid
models with uncertainty. The third part of the book includes additional applications in
disaster management and emergencies.

[email protected]
14 Decision Aid Models for Disaster Management and Emergencies

Bibliography

[1] Norman, M. (2003). The role of local institutions and their interaction in disaster risk mitiga-
tion: a literature review. FAO Working Paper.
[2] Tomasini, R. and Van Wassenhove, L.N. (2009). Humanitarian logistics. Palgrave MacMillan,
Basingstoke.
[3] Pedraza-Martinez, A.J., Hasija, S. and Van Wassenhove, L.N. (2010). An Operational Mecha-
nism Design for Fleet Management Coordination in Humanitarian Operations. INSEAD Work-
ing Paper 2010/87/TOM/ISIC.
[4] Besiou, M., Pedraza-Martinez, A.J. and Van Wassenhove, L.N. (2012). Decentralization and
Earmarked Funding in Humanitarian Logistics for Relief and Development. INSEAD Working
Paper 2012/10/TOM/ISIC.
[5] Pedraza Martinez, A.J., Van Wassenhove, L.N., Estrada Mejia, C. and Stapleton, O. (2010).
The Armenia Earthquake: Grinding Out an Effective Disaster Response in Colombia’s Coffee
Region. Oper. Manag. Educ. Rev., 4, pp. 5–40.
[6] Regnier, E. (2008). Public Evacuation Decisions and Hurricane Track Uncertainty. Manag. Sci.,
54 (1) pp. 16–28.
[7] Uccellini, L. (2012). Building on the Ongoing Revolution in Weather Forecasting: Taking Pre-
diction Capabilities to Decision Support Services and Humanitarian Efforts. POMS HOCM
College Inaugural Mini-Conference. Chicago, IL, April 19, 2012.
[8] Kistler, R., Kalnay, E., Collins, W., Saha, S., White, G., Woollen, J., Chelliah, M., Ebisuzaki,
W., Kanamitsu, M., Kousky, V., van den Dool, H., Jenne, R. and Fiorino, M (2001). The NCEP–
NCAR 50-Year Reanalysis: Monthly Means CD-ROM and Documentation. Bul. of the Amer.
Meteor. Soc 82 (2) pp. 247–267 1
[9] Reuters. (2010). Powerful snowstorm hits U.S. East Coast. February 6, 2010. Retrieved June
30, 2012.
[10] Pedraza Martinez, A.J., Stapleton, O. and Van Wassenhove, L.N. (2011). Field Vehicle Fleet
Management in Humanitarian Operations: A Case-Study Based Approach. J. of Oper. Manag.,
29 (5) pp. 404–421.
[11] Tomasini, R. Van Wassenhove, L.N. and Khaw, G.L. (2006). Fleet Forum: Rethinking Human-
itarian Vehicle Management, INSEAD Case.
[12] Pedraza Martinez, A.J. and L.N. Van Wassenhove. (2010). Vehicle Replacement in
the International Committee of the Red Cross. Prod. Oper. Manag. doi:10.1111/j.1937-
5956.2011.01316.x.
[13] Wolz, C. and Park, N. (2006). Evaluation of ReliefWeb. Prepared for Office for the Coordination
of Humanitarian Affairs, United Nations. Retrieved from http://reliefweb.int
[14] Tomasini, R. and Van Wassenhove, L.N. (2005). Managing Information in Humanitarian Crises:
The UNJLC Website. INSEAD Case Study.
[15] Russell, T. (2005). The Humanitarian Relief Supply Chain: Analysis of the 2004 South East
Asia Earthquake and Tsunami. Master of Engineering in Logistics at the Massachusetts Insti-
tute of Technology.
[16] OCHA. (2011). Annual Report. Office for the Coordination of Humanitarian Affairs. United
Nations.
[17] ReliefWeb. (2010). 2010 Annual Statistics. Published online at reliefweb.int. Retrieved in June
1st , 2012.
[18] ReliefWeb. (2009). 2009 Annual Statistics. Published online at reliefweb.int. Retrieved in June
1st , 2012.
[19] ReliefWeb. (2008). 2008 Annual Statistics. Published online at reliefweb.int. Retrieved in June
1st , 2012.

[email protected]
Bibliography 15

Websites:

Emeraldinsights.com
fleetforum.org
chevinfleet.com
ncep.noaa.gov/mission/
reliefweb.int

[email protected]
Chapter 2

Decision Aid Models and Systems for

Humanitarian Logistics. A Survey

M.T. Ortuño1,∗ , P. Cristóbal2 , J.M. Ferrer1, F.J. Martín-Campo1, S. Muñoz1, G. Tirado1,

B. Vitoriano1
1Dpto. de Estadística e Investigación Operativa, Universidad Complutense de Madrid,
Spain
2 Dpto. de Matemática Aplicada a los Recursos Naturales, Universidad Politécnica de
Madrid, Spain

The number and impact of disasters seems to be increasing in the last decades, and their
consequences have to be managed in the best possible way. This paper introduces the main
concepts used in emergency and disaster management, and presents a literature review on
the decision aid models and systems applied to humanitarian logistics in this context.

2.1 Introduction

The number and impact of disasters seems to be increasing in the last decades, and
their consequences have to be managed in the best possible way. Recent years have seen
an explosion of literature regarding disaster and emergency management, as it is a topic
of high relevance in today’s world. Among this literature, there is an increasing amount
of research regarding mathematical models and systems which can help in the decision
aid processes developed when trying to respond to the consequences of a disaster. In this
introductory section the main definitions concerning disasters, emergencies and humani-
tarian logistics are stated, which will allow to classify the research into the main phases of
disaster management.
∗ Contact e-mail: [email protected]

B. Vitoriano et al. (eds.), Decision Aid Models for Disaster Management 17

and Emergencies, Atlantis Computational Intelligence Systems 7,
DOI: 10.2991/978-94-91216-74-9_2, Ó Atlantis Press 2013

[email protected]
18 Decision Aid Models for Disaster Management and Emergencies

2.1.1 Hazards and disasters

A hazard is a threatening event or probability of occurrence of a potentially damaging

phenomenon within a given time period and area. It can be both natural or human-made.

• Natural: naturally occurring physical phenomena caused either by rapid or slow on-
set events which can be geophysical, hydrological, climatological, meteorological or
biological (earthquakes, landslides, tsunamis, volcanic activity, avalanches, floods,
extreme temperatures, droughts, wildfires, cyclones, storm/wave surges, disease epi-
demics, animal plagues, etc.).
• Human-made or technological: events caused by humans and which occur in (or close
to) human settlements, such as complex emergencies/conflicts, famine, displaced pop-
ulations, industrial accidents (toxic dumps or radioactive escapes), catastrophic trans-
port accidents, etc.

An emergency is a situation that poses an immediate risk to health, life, property or

environment.
A disaster is the disruption of the normal functioning of a system or community, which
causes a strong impact on people, structures and environment, and goes beyond the local
capacity of response. Sometimes, to declare or not an emergency as a disaster is a political
decision, because it has consequences for the involvement of third parties in the interven-
tion or for insurance, for example.
Catastrophe is another term used in disaster management. There is also a discussion
in the literature about the difference between disaster and catastrophe. Usually a catas-
trophe is considered an extremely large-scale disaster. As stated in Quarantelli [74], just
as “disasters” are qualitatively different from everyday community emergencies, so are
“catastrophes” a qualitative jump over “disasters”. This qualitative jump is reflected in
several characteristics and results in important differences in the logistics of the interven-
tion, as discussed in what follows.

2.1.2 Disaster management and humanitarian logistics

Disaster response is a complex process that involves severe time pressure, high uncer-
tainty and many stakeholders. It also involves several autonomous agencies to collabo-
ratively mitigate, prepare, respond, and recover from heterogeneous and dynamic sets of
hazards to society.

[email protected]
Decision Aid Models and Systems for Humanitarian Logistics. A Survey 19

The agents involved differ depending on the type of disaster (civil protection and local
security agencies usually manage technological disasters, while natural disasters normally
involve also NGOs and international agencies), the disaster consequences and the place
where it strikes, due to vulnerability.
The agents involved in disaster response can be classified into three levels, as involve-
ment in the operations depends on the consequences:

• Local level: the first response level, usually addressed by local agencies, civil society
organizations and civil protection. Typically, this level of emergency is not declared as
a disaster.
• National level: the army and national civil protection, governmental organizations and
NGOs are usually involved when an emergency is defined as a disaster. Sometimes,
international organizations with local offices also participate at this level.
• International level: foreign governments and inter-governmental organizations, inter-
national NGOs for disaster response and the United Nations Agencies. Coordination
at this level is a crucial matter, usually performed by OCHA (Office for the Coordina-
tion of Humanitarian Affairs) of United Nations, and the IASC (Inter-Agency Standing
Committee), primary mechanism for inter-agency coordination, including key UN and
non-UN humanitarian partners. This level is reached when national capacity of re-
sponse is not enough (due to the scale of the disaster and/or the vulnerability of the
country) and the national government authorizes an international humanitarian opera-
tion.

The decision making processes in disaster management are thus extremely difficult,
due to the multiple actors (decision-makers) which are involved, and the complexity of
the tasks addressed. Among those tasks, all the process of planning, implementing and
controlling the efficient, cost-effective flow and storage of goods and materials as well as
related information, from the point of origin to the point of consumption for the purpose of
meeting the end beneficiary’s requirements and alleviate the suffering of vulnerable people
is called humanitarian logistics, as defined in the Humanitarian Logistics Conference,
2004 (Fritz Institute).
Based on this definition, humanitarian logistics also appears in contexts different from
disaster management; the World Food Programme (WFP) and the World Health Organiza-
tion (WHO), for instance, develop many operations which can be considered humanitarian
logistics without being a response to a specific disaster. However, it is in disaster man-
agement where the application of humanitarian logistics is more complex and difficult and

[email protected]
20 Decision Aid Models for Disaster Management and Emergencies

where more differences with business logistics appear. Therefore, our review is restricted
to this context. The main issues that differentiate humanitarian supply chains in the context
of disaster management from business supply chains are the following:

• Unpredictable demand in terms of timing, geographic location, type and quantity of

commodity.
• Short lead time and suddenness of demand for large amounts of a wide variety of
products and services.
• Lack of initial resources in terms of supply, human resources, technology, capacity and
funding (see Balcik and Beamon [9]).
• Presence of multiple decision makers that can be sometimes difficult to identify.

2.1.3 Phases, tasks and decisions of the disaster management cycle

In the decision making processes needed in humanitarian logistics for disaster manage-
ment, the context and the nature of the decisions to be made change over time as we move
from before to after the disaster event. Deciding about preventive actions to mitigate the
effects of a possible future earthquake is not the same as deciding about the precise actions
to undertake just after it strikes, or a month later. The context-related uncertainties and time
pressure may vary a lot from one situation to the other, as well as the nature of the decisions
and the criteria of the involved actors. This has led to distinguishing four successive phases
in the management of emergencies and disasters according to the main nature of the tasks
to be performed and their temporal location with respect to the disaster event:

• Mitigation: all the middle and long-term actions and decisions aimed to prevent and
mitigate the consequences of a future disaster, as long as it is not (known to be) immi-
nent. Typical tasks of this phase are the identification of risk groups and vulnerability
patterns and their treatment, or the development of prediction systems and emergency
plans and the allocation of resources for them.
• Preparedness: all the short-term interventions once the available prediction systems
have raised an alarm of an upcoming adverse phenomenon until it finally strikes. This
includes setting off the emergency systems and evacuation plans, the real-time track-
ing of the hazard, the analysis of the most probable scenarios, the reinforcement of
critical infrastructures, etc. This phase also includes some long-term decisions such as
inventory prepositioning and network design.

[email protected]
Decision Aid Models and Systems for Humanitarian Logistics. A Survey 21

• Response: this phase is focused on saving lives and it is characterized by a short du-
ration with high emergency and high uncertainty. It is usually divided into a first
response phase, devoted to the rescue and urgent medical assistance of injured and af-
fected people (depending on the disaster scenario, it may last around one week from
the moment of the disaster event), and a middle-term response phase, devoted to esti-
mate and mitigate the potentially unattended first needs of the affected population as
a result of possible damage to life-line infrastructures and resources (shelter, ordinary
medical assistance, water and food supply, etc.). This middle-term stage usually in-
volves the delivery of aid from outside of the affected zone and can last for weeks or
even months from the moment of the disaster, depending on its nature and magnitude
as well as on the economic and development circumstances of the affected country.
• Recovery: this phase is focused on achieving efficiency and it is characterized by
its long duration with low emergency and low uncertainty. It refers to all the long-
term actions and decisions aimed to recuperate the normal functioning of the affected
community and the reconstruction of the social fabric, including life-line resources,
services and infrastructure, and the necessary improvements in order not to repeat the
specific vulnerabilities shown by the affected groups and places. Sometimes, after cer-
tain disasters, a periodic flow of humanitarian aid will be needed to support particularly
vulnerable people, which is outside the scope of disaster management.

Though this division in phases is clear and well-founded from a temporal and concep-
tual point of view, it is important to remark that the process of disaster management has to
be understood as a whole, where phases are not independent of each other despite their dif-
ferent focus. In this way, the preparedness phase relies critically on the prediction systems
set up in the mitigation phase and the urgent decision-making of the former phase would
be impossible without the previous vulnerability analysis and emergency plans developed
in the latter phase. Similarly, the allocation of resources for first response operations must
be taken into account when designing the mitigation policy. Besides, the middle-term re-
sponse stage cannot be successful if it does not enable an adequate recovery process. In
this sense the delivery of external aid has to be carefully examined and carried out in or-
der not to destroy the local economy. Furthermore, notice that the recovery and mitigation
tasks partially overlap, in such a way that once an affected community has reached a new
equilibrium, it has to re-evaluate the possible occurrence of future disasters and thus a new
mitigation stage follows the recovery phase. Thus, the disaster management process is a

[email protected]
22 Decision Aid Models for Disaster Management and Emergencies

non-stop cycle (see Fig. 2.1, as in Tomasini and van Wassenhove [87]), in which each phase
is based on the previous ones, clearly showing the interdependence between all of them.

Fig. 2.1 Phases of the disaster management cycle

Some other factors appear that must be considered in all phases of the disaster man-
agement cycle. A key factor in humanitarian logistics is the management of funding and
donations. Donations are usually classified into two groups: earmarked donations, given by
donors to be used in specific operations, and non-earmarked donations, that have no con-
straints of use. The management of both kinds of donations is crucial to achieve efficiency,
and thus it is important to take them into account when making decisions.
In order to help in the decision making processes arising in each of the interrelated
previous phases, different mathematical models can be developed. In Altay and Green
[4], the literature in Operations Research models for disaster operations management up to
2005 is reviewed. In this work we present an up-to-date review of the Operations Research
literature, which is completed with a review on the decision aid systems already in use by
different organizations.
The paper is organized as follows. In Section 2.2 the literature on decision aid mod-
els is reviewed, classifying the bibliography according to the type of problem, the disaster
management phase and the solution technique. Section 2.3 introduces the decision aid sys-
tems most commonly used in practice, describing their main functionalities and providing
related references. Finally in Section 2.4 some conclusions are drawn.

[email protected]
Decision Aid Models and Systems for Humanitarian Logistics. A Survey 23

2.2 Decision Aid Models. Literature Review

This section is a humanitarian logistics literature review focused on papers presenting

mathematical models. They will first be classified according to the disaster management
phase considered and afterwards according to the type of problem approached. A special
subsection will be devoted to donations and funding. At the end of the section a brief
classification according to the solution technique used to solve the proposed models is also
provided.
We will restrict the review mostly to deterministic solution techniques, because, even
though uncertainty is a key factor in humanitarian logistics, the treatment of this uncertainty
in the literature is presented in a different chapter of this book.

2.2.1 Pre-disaster actions: mitigation and preparedness

In this section, the bibliography concerning mitigation and preparedness has been clas-
sified into the following six groups, according to the main problem tackled: assessment,
facility location, supply chain design, distribution planning, evacuation, and inventory plan-
ning.

Assessment

- Tovia [88] builds a simulation model for emergency response to be used by offices of
emergency preparedness. It is conceived to evaluate response capabilities and assess
the logistics resources required to evacuate, shelter and protect the population, in a
timely fashion in the event of a hurricane.
- Zobel and Khansa [100] present a multi-criteria approach providing a quantitative mea-
sure of resilience in the presence of multiple related disaster events, in order to be able
to compare several configurations and to assess their relative returns and risks for ef-
fective emergency management.
- Kung et al. [49] make use of multiple linear regression, multivariate analysis, back-
propagation networks and simulation for designing a real-time mobile system to pre-
dict and assess the occurrence of debris flows.
- Uchida [91] develops a simulation model for clarifying how disaster warning issuance
conditions affect “cry wolf syndrome” in the case of landslides caused by heavy rain-
fall.

[email protected]
24 Decision Aid Models for Disaster Management and Emergencies

Facility location

- Akkihal [3] considers a battery of mixed-integer linear programs for locating ware-
houses for non-consumable inventories required for the initial deployment of aid after
a disaster occurrence.
- Doerner et al. [32] provide a multi-criteria programming model for locating public
facilities as schools in areas close to the coast, taking into account risks of inundation
by tsunamis. It performs a comparison between two solution approaches: a genetic
algorithm and a decomposition technique.
- Eiselt and Marianov [33] propose a 0-1 linear programming model for locating or
strengthening cell phone towers in order to maximize service coverage and minimize
the loss of communications in case a natural disaster occurs.

Supply chain design

- Hale and Moberg [38] consider a set cover location model for establishing an efficient
network of secure storage facilities that can effectively support multiple supply chain
facilities.
- Van Wassenhove [93] shows the complexities of managing supply chains in humani-
tarian settings and outlines strategies for better preparedness.
- Nagurney et al. [62] state a variational inequality problem for designing a supply chain
network for critical needs, minimizing the total cost and considering demand uncer-
tainty.

Distribution planning

- Nolz et al. [66] provide a multi-criteria metaheuristic based on evolutionary concepts

for planning water distribution tours in disaster relief, determining the physical loca-
tion of water tanks and selecting roads to be used for the transportation of drinking
water.
- Rottkemper et al. [76] propose a bi-objective mixed-integer programming model for
integrated relocation and distribution planning of relief items to the affected regions in
the aftermath of a disaster.

[email protected]
Decision Aid Models and Systems for Humanitarian Logistics. A Survey 25

Evacuation

- Saadatseresht et al. [77] propose a three-step approach based on the use of multi-
criteria evolutionary algorithms and on a geographical information system to deter-
mine the distribution of evacuees into safe areas in an emergency situation.
- Bretschneider and Kimms [16] present a two-stage heuristic approach for a pattern-
based mixed integer dynamic network flow model to restructure the traffic routing for
evacuating an urban area affected by a disaster.

Inventory planning

- Lodree and Taskin [53] introduce some variants in the classic newsvendor model in
order to deal with demand uncertainty in an emergency response inventory planning
problem that is relevant to manufacturing, service, not-for-profit, and government or-
ganizations that provide supplies, equipment, and manpower to support disaster relief
operations.
- Qin et al. [73] propose an inter-temporal integrated single-period inventory model to
determine the optimal order quantity of emergency resources in a flood incident. It
also describes a genetic algorithm based simulation approach.

2.2.2 Disaster response

Most of the papers dealing with the disaster response phase are related to the distribu-
tion of humanitarian aid, in a global or a local context, the evacuation of people affected
by a disaster or the location of support centers. In this section, the bibliography is clas-
sified into the following four groups, according to the main problem tackled: last mile
distribution, evacuation, large scale distribution, and other problems.

Last mile distribution

- Tzeng et al. [90] use fuzzy multi-objective linear programming technics to approach
a multi-period aid distribution problem minimizing total cost and total travel time and
maximizing minimal satisfaction during the planning period.
- Balcik and Beamon [9] consider a two-phase multi-period planning model for relief
distribution. In the first phase, the possible routes for vehicles within the time horizon
are generated. In the second phase, the optimal routes and the amount of supplies to

[email protected]
26 Decision Aid Models for Disaster Management and Emergencies

be transported are selected for the coming periods by solving the underlying integer
programming problem.
- Campbell et al. [18] introduce and analyze two alternative objective functions for the
traveling salesman problem and the vehicle routing problem, considering arrival times
to destinations, in order to suit the humanitarian nature of the distribution. The new
problems are solved with an insertion and local search based heuristic approach.
- Gibbons and Samaddar [37] develop a simulation model for a problem of distribution
of vaccines in a pandemic situation and in a context of strong randomness.
- Blecken et al. [15] apply both exact and heuristic techniques to deal with a warehouse
provision and relief aid flow problem minimizing total cost.
- Nolz et al. [65] present a hybrid solution approach based on genetic algorithms, vari-
able neighborhood search and path relinking to tackle a multi-criteria problem of water
distribution to the population affected by a catastrophe.
- Vitoriano et al. [96] propose a multi-criteria double-flow model to address the problem
of distribution of goods to the affected population of a disaster. Criteria such as cost,
time of response, distribution equity, security and reliability are considered jointly
using a goal programming approach. Some preliminary results are also presented in
Vitoriano et al. [95] and Ortuño et al. [70].
- Berkoune et al. [13] develop a genetic algorithm for solving a multi-commodity and
multi-depot routing problem which aims to minimize the total duration of all trips.
- Huang et al. [42] propose three objective functions for a routing problem concerning
cost, speed and equity of the delivery. The three resulting problems are solved with
different GRASP based metaheuristics and the objective function values are compared.

Evacuation

- Kongsomsaksaku et al. [48] provide a flood evacuation planning model in two levels:
shelter location by the authority and shelter selection and evacuation routing by the
evacuees. The bi-level programming problem is solved using a genetic algorithm.
- Chiu and Zheng [26] present an evacuation and support planning model on a transport
network for real-time emergency response in no-notice disasters. The linear program-
ming model is solved through the matrix formulation.
- Yi and Özdamar [99] describe an integrated location-distribution-evacuation network
flow model for coordinating logistic support operations. The model determines the

[email protected]
Decision Aid Models and Systems for Humanitarian Logistics. A Survey 27

number of vehicles, the number of wounded people and the amount of commodities
traversing each arc via linear programming. Then, loads are distributed among vehicles
by solving a system of linear equations. Yi and Kumar [98] use ant colony optimization
to solve this problem, but disregarding facility location issues.
- Jotshi et al. [44] propose a simulation model with the support of data fusion for an
emergency problem of picking up and delivering patients to hospitals.
- Ben-Tal et al. [12] apply robust optimization for a multi-period transportation problem
consisting of dynamically assigning emergency response and evacuation traffic flow
with time dependent uncertainty on the demands.

Large scale distribution

- Angelis et al. [6] consider an integer linear programming model to solve a multi-period
weekly planning problem of aircraft routing to delivery food to Angola, in the context
of the United Nations World Food Program.
- Sheu [80] presents a dynamic aid plan to be used after an earthquake for several pur-
poses: grouping areas, estimation of affected people at each area, estimation of pri-
orities and distribution of aid. Different mathematical techniques are applied, such as
fuzzy clustering and multi-objective dynamic programming.
- Adıvar and Mert [1] propose a fuzzy linear programming model to design a plan for
transporting international aid from donor countries to the country in need.
- Charles [22] states a global multilevel facility location and distribution model. This
mixed integer programming model includes international suppliers, potential ware-
houses, affected areas and two types of means of transportation: boats and planes.

Other problems

- Beamon and Kotleba [11] introduce a multi-supplier inventory model with discrete
random demands and two types of orders: normal and emergency orders. Beamon and
Kotleba [10] develop and test three different inventory management strategies for the
previous model that are applied in the context of the emergency situation associated to
the civil war in South Sudan.
- Jin and Ekşioǧlu [43] present an integer programming model for determining alterna-
tive routes after a disaster in a road network. An algorithm is proposed to update the

[email protected]
28 Decision Aid Models for Disaster Management and Emergencies

parameters of the simplified linear program and it is tested using a simulated disaster
scenario.

2.2.3 Disaster recovery

In this section, the bibliography concerning the recovery stage has been classified into
the following seven groups, according to the activities involved: civil infrastructure sys-
tems, power system restoration, recovery planning, economic recovery, health care and
mental health recovery, urban disaster and housing recovery, and other activities.

Civil infrastructure systems: transportation networks, bridges, etc.

- Karlaftis et al. [47] present a methodology for optimally allocating funds for repair-
ing an urban infrastructure transportation network following natural disasters using a
three-stage genetic algorithm based approach.
- Permann [72] applies genetic algorithms to the problem of infrastructure networks
modeling and analysis in order to determine the optimum assets to restore or protect
from attack or other disaster.
- Mehlhorn [60] develop a method for prioritizing highway routes for reconstruction
after a natural disaster.
- Natarajarathinam et al. [63] review the management of supply chains in times of crisis,
including the recovery phase.
- Orabi et al. [68] present a model for damaged transportation networks to allocate lim-
ited reconstruction resources to competing recovery projects, generating optimal trade-
offs between minimizing the reconstruction duration and cost. In Orabi et al. [69] the
post–disaster recovery efforts of damaged civil infrastructure systems are optimized,
minimizing both the performance loss of the damaged transportation network and the
reconstruction costs.
- Matisziw et al. [56] present a multi-objective model to ensure that network restoration
during disaster recovery is prioritized so that system performance is optimized.
- Solano [81] works on preparedness, response, and recovery plans, selecting infrastruc-
tures for protection in presence of scarce resources.
- Van Hentenryck et al. [92] formalize the specification of the single commodity al-
location problem for disaster recovery and introduce a novel multi-stage hybrid-
optimization algorithm.

[email protected]
Decision Aid Models and Systems for Humanitarian Logistics. A Survey 29

- Maya and Sörensen [57] describe how to allocate scarce resources to repair a rural
road network after it has been damaged by a natural or man-made disaster.
- Mehlhorn et al. [59] present a plan for bridges’ repair to restore a highway network
that allows accessibility to key facilities in the damaged area.
- Losada et al. [54] incorporate the facility recovery time in a model that identifies
the optimal allocation of protection resources in an uncapacitated median network to
hedge against worst–case facility losses.

Power system restoration

- Crowther and Haimes [30] evaluate risk management options against multiple objec-
tives. This case study calculates an efficient frontier of solutions by integrating a sim-
plified model of the costs of recovery to the power sector derived from open-source
data with the Inoperability-Output Model.
- Ang [5] describes a model to plan the recovery of an electrical power transmission grid
that has been damaged by a natural disaster or terrorist attack.
- Xu et al. [97] determine how to schedule inspection, damage assessment, and repair
tasks so as to optimize the post-earthquake restoration of the electric power system.
- Coffrin et al. [28] model and solve the power system restoration planning problem for
disaster recovery.
- Coffrin et al. [29] present a model to decide how to store power system components
throughout a populated area to maximize the amount of power served after disaster
restoration.

Recovery planning

- Aftanas [2] develops an objective and repeatable process for optimizing the project
prioritization phase of the recovery effort.
- Buzna et al. [17] consider structures and organization in complex systems. The effec-
tiveness of recovery strategies is studied for a dynamic model of failure spreading in
networks.
- Nelson et al. [64] build a high level event and role based mobility model for simulating
disaster recovery networks.

[email protected]
30 Decision Aid Models for Disaster Management and Emergencies

- Chen et al. [23] propose a quantitative measurement to recoverability assessment, and

establish a model of recoverability process to minimize the recovery time and optimize
the allocation of resources.
- Balasubramanian et al. [8] implement an integrated disaster recovery plan based upon a
plurality of requirements for an application by receiving a first set of inputs identifying
one or more entity types for which the plan is to be formulated.

Economic recovery: small business, private enterprises, insurance, . . .

- Ghesquiere and Mahul [36] discuss the optimal level of sovereign insurance in the case
of developing countries.
- Rose [75] examines the consistency of the operational definitions of economic re-
silience in relation to antecedents from several disciplines, and evaluates the effec-
tiveness of economic resilience to reduce losses from disasters on the basis of recent
empirical studies.
- Saleem et al. [78] propose a model for pre-disaster preparation and post-disaster busi-
ness continuity–rapid recovery. The model is used to design and develop a web based
prototype of their business continuity information network system facilitating collab-
oration among local, state, federal agencies and the business community for rapid
disaster recovery.
- Luckey [55] presents a literature review from 2001 to 2008, which identifies key stages
for consideration when performing information technology disaster recovery planning
to ensure business viability if disasters occur.

Health care and mental health recovery

- de Mel et al. [31] measure mental health recovery and economic recovery for small
business owners affected by the 2004 Indian Ocean tsunami in Sri Lanka.
- Onyango [67] investigates humanitarian responses to complex emergencies.
- Cimellaro et al. [27] present a quantitative evaluation of disaster resilience which is
implemented for evaluation of health care facilities subject to earthquakes.

[email protected]
Decision Aid Models and Systems for Humanitarian Logistics. A Survey 31

Urban disaster and housing recovery

- Jung et al. [45] introduce some mathematical models for reconstruction planning in
urban areas and design of redevelopment strategies that minimize future risk of flood
and maximize net social benefit under spatial and equity constraints.
- Chang [20] presents a framework for assessing empirical patterns of urban disaster
recovery through the use of statistical indicators.
- El-Anwar et al. [34] study how to quantify and maximize the sustainability of inte-
grated housing recovery efforts after natural disasters.

Other activities

- Miles and Chang [61] present and operationalize a conceptual model of recovery to
create a numerical model of recovery which is intended for decision support.
- Fiedrich and Burghardt [35] show how applications of agent technology can be used to
support many processes throughout the phases of the disaster management cycle from
mitigation and preparation to actual response and recovery.
- Semsch et al. [79] develop an approach for increasing the robustness of plans in multi-
agent disaster, in which interactions between agents’ actions and plans may arise due
to simultaneous activity of the agents in a shared environment.
- Apte [7] discusses research issues and potential actions in disaster recovery.
- Chib and Komathi [25] implement information communication technologies in recov-
ery operations, drawing attention to vulnerability reducing potential of the initiatives.
- Kaklauskas et al. [46] describe the development of a knowledge model for post–
disaster management using multiple criteria decision making theory.
- Chao et al. [21] analyze the human–computer interaction of remote disaster recovery
systems.

2.2.4 Donations and funding

Several recent papers have approached the study of the role of donations in humanitar-
ian logistics:

- Besiou et al. [14] use system dynamics simulation to present studies supporting the
statement that earmarked donations decrease operational efficiency in decentralized
humanitarian operations.

[email protected]
32 Decision Aid Models for Disaster Management and Emergencies

- Pedraza Martinez and Van Wassenhove [71] use dynamic programming to find an op-
timal replacement policy for the International Committee of the Red Cross, showing
empirically that earmarked funding affects the implementation of replacement policies
in humanitarian contexts, increasing operational cost.
- Toyasaki and Wakolbinger [89] model the strategic interaction between donors and
humanitarian organizations in a deterministic setting using game theory.

The importance of donations is a fundamental difference of humanitarian logistics with

respect to commercial logistics; however, they have not yet been sufficiently studied by
academics and there is a wide field for future research in this context.

2.2.5 Technique classification

Now we categorize the bibliography based on the methodology utilized for solving the
formulated problems.

• Optimization techniques
– Mathematical programming: Aftanas [2], Akkihal [3], Ang [5], Angelis et al. [6],
Balcik and Beamon [9], Ben-Tal et al. [12], Blecken et al. [15], Bretschneider
and Kimms [16], Charles [22], Chen et al. [23], Chiu and Zheng [26], Coffrin
et al. [28], Coffrin et al. [29], Doerner et al. [32], Eiselt and Marianov [33],
Hale and Moberg [38], Jin and Ekşioǧlu [43], Jung et al. [45], Losada et al. [54],
Mehlhorn [60], Nagurney et al. [62], Pedraza Martinez and Van Wassenhove
[71], Rottkemper et al. [76], Van Hentenryck et al. [92], Xu et al. [97], Yi and
Özdamar [99].
– Metaheuristics: Berkoune et al. [13], Blecken et al. [15], Campbell et al. [18],
Coffrin et al. [28], Huang et al. [42], Jung et al. [45], Kongsomsaksaku et al.
[48], Maya and Sörensen [57], Nolz et al. [65], Nolz et al. [66], Saadatseresht
et al. [77], Van Hentenryck et al. [92], Yi and Kumar [98].
– Inventory control: Beamon and Kotleba [10], Beamon and Kotleba [11], Lodree
and Taskin [53], Qin et al. [73].
– Constraint programming: Coffrin et al. [28], Van Hentenryck et al. [92].
• Multi-criteria decision making: Aftanas [2], Crowther and Haimes [30], El-Anwar
et al. [34], Huang et al. [42], Jung et al. [45], Kaklauskas et al. [46], Matisziw et al.
[56], Nolz et al. [65], Orabi et al. [68], Orabi et al. [69], Ortuño et al. [70], Tzeng et al.
[90], Vitoriano et al. [95], Vitoriano et al. [96], Zobel and Khansa [100].

[email protected]
Decision Aid Models and Systems for Humanitarian Logistics. A Survey 33

• Simulation: Buzna et al. [17], Gibbons and Samaddar [37], Jotshi et al. [44], Kung
et al. [49], Miles and Chang [61], Nelson et al. [64], Tovia [88], Uchida [91].
• Expert systems and artificial intelligence: Chao et al. [21], Fiedrich and Burghardt
[35], Karlaftis et al. [47], Permann [72], Xu et al. [97].
• Probability and statistics: Chang [20], Cimellaro et al. [27], Ghesquiere and Mahul
[36], Solano [81].
• Fuzzy systems: Adıvar and Mert [1], Sheu [80], Tzeng et al. [90].
• Dynamic systems: Besiou et al. [14], Buzna et al. [17].
• Game theory: Semsch et al. [79], Toyasaki and Wakolbinger [89].

2.3 Decision Aid Systems

In the last two decades the development of decision aid systems for humanitarian logis-
tics has grown intensely. The most used systems have focused mainly on inventory control,
usually leaving aside the logistics related to the supply transportation and distribution. That
is the reason why the systems described below are mainly used in the phases of Prepared-
ness and Response, due to the importance that inventory control and fleet management
have in these two phases; however, depending on the particular situation, they could also
be helpful in the phases of Recovery and Mitigation.

In the following, some of the most important systems are briefly described.

SUMA and LSS. In 1992, the Pan American Health Organization, together with the
Regional Office of the World Health Organization under the support of different countries
(principally Holland), developed the Humanitarian Supply Management System (SUMA)
[85], being one of the most complete systems able to manage the information and resources
when a disaster happens. This system has been used as inventory support for other tasks in
employers’ organizations. The full system is composed by three modules:

• Suma Central: it is the core of the system and is normally situated in the emergency
operations center. The decision makers, situated at this level, collect and define every
parameter of the problem such as strategic recovery and delivery places, locations for
warehouses, etc.
• Field Unit: its aim is to manage the different operations in the crucial sites such as
airports, seaports and places defined in the above level. All the received supplies are
separated and classified in order to better tend to the needs of the affected population.

[email protected]
34 Decision Aid Models for Disaster Management and Emergencies

• Stock management: it registers all information concerned with the received and deliv-
ered supplies. This module balances the available inventory in a certain warehouse.

Once the strategic parameters have been defined by Suma Central, the supplies must
be managed properly. For this goal, Field Unit is used to sort and label the collected
supplies in the different reception sites. The classification is divided in three types of
items: “Urgent! Immediate Distribution”, “Non-Urgent Distribution”, and, “Non-Priority
items”. In a secondary way, the items are separated in different groups: food, medicines,
etc., and then the information is sent to Suma Central. Finally, the supplies are inventoried
and delivered to other warehouses or distribution hubs, arriving at the affected population
under the coordination with the operations center.
As shown above, this system requires a good tool for data transmission among the
three modules, which is not always possible in a place that has been affected by a disaster.
This system was used up to 2005, when its successor, called Logistics Support System
(LSS) [86], was implemented, taking SUMA as a support software. Agencies as PAHO
and OCHA, among others, have made the project possible. LSS makes the information
exchange among NGOs, donors and affected and contributed countries easier.
Like SUMA, LSS classifies and sorts the received supplies. However, the structure of
the system emphasizes the treatment of the supplies being included in a database:

• Entries: the different received supplies that have arrived in a certain local point are
registered in the system as well as classified and sorted.
• Deliveries: the supplies shipments are tracked by the system taking into account the
initial and destination points as well as the main features of the shipment.
• Pipeline: the expected supplies appear in this module of the system.
• Request: this part of the system manages the information about those places that need
specific supplies and is useful for determining how to allocate the required supplies.
• Stock basket: once the supplies have been classified in groups, this system tracks the
available stocks for better managing future requirements.
• Importing/exporting: in this part, the system is able to exchange information about the
supplies with other strategic places, external systems or agencies.

SUMA has been used since 1995, when it was applied in the Hurricane Louis in
Caribbean Islands, and afterwards deployed in other natural disasters such as the Hurri-
cane Mitch that hit Central America in 1998. LSS has been used since 2005, when it was
applied in the Tropical Storm Stan in El Salvador. Among many others, it was also de-

[email protected]
Decision Aid Models and Systems for Humanitarian Logistics. A Survey 35

ployed in the 2005 and 2007 earthquakes that hit Pakistan and Peru, respectively, and the
2008 volcano eruptions in Ecuador. For more details on SUMA and LSS deployments, see
[85, 86]. Many agencies have taken SUMA/LSS as support for humanitarian logistics, such
as UN, WHO, WFP, OCHA, UNICEF, UNHCR, PAHO, CTS, Red Cross and many NGOs
like ICRC, IFRC, MSF and OXFAM, among others.

HLS and HELIOS. Humanitarian Logistics Software (HLS) [82] is a centralized web
based supply chain system that was developed in 2003 by the Fritz Institute in collabora-
tion with the International Federation of the Red Cross (IFRC) and Red Crescent. It was
designed basically to track supplies and financing from donation to delivery, increasing
transparency of donations, speeding up the relief chain and providing more detailed infor-
mation to decision makers at headquarters and in the field. The system provides an online
overview of the relief pipeline and allows a quick management of resources by using web-
based supplier lists and catalogues, making it possible to make orders directly online. HLS
was first used in 2004 by IFRC in several relief operations, such as the Morocco earth-
quake, the South Asia tsunami or the Haiti hurricane. It was also used later in operations
such as the 2005 earthquake in Pakistan and the 2006 earthquake and tsunami in Indonesia,
among others.
Between 2007 and 2008, HLS was replaced by a more advanced software also devel-
oped by the Fritz Institute called HELIOS [83]. It is a specific humanitarian logistics deci-
sion support tool that provides organizations with online real-time data to improve supply
chain management in humanitarian relief. It consists of five modules:

• The Project Management module administrates the open projects, estimating the main
needs and managing requests. This is the main module, coordinating the execution of
the others when needed.
• The Request Processing module manages the information concerning issued requests,
such as the issuer data, required goods, pickup and delivery locations, etc.
• The Warehouse module deals with the management of stocks and it can be activated
when an issue is received or a delivery is performed.
• The Mobilization module is focused on donation management, from cash to in-kind
goods and services. It is used by organizations to monitor and report on donations,
highly improving the information provided to donors concerning the use of their do-
nations.

[email protected]
36 Decision Aid Models for Disaster Management and Emergencies

• The Procurement module is used to administrate and purchase goods, managing the
corresponding quotation requests and executing orders.

The operations to be performed are automated through a web based system in order to im-
prove the coordination between agents, but it could also be functional offline if an internet
connection is not available. In 2008, World Vision International and Oxfam GB imple-
mented HELIOS for the first time in two locations each. Later in 2009, Oxfam extended
the system to 20 different countries.

SAHANA. It is a disaster management system (see [52]) that was created by the Sri
Lankan IT community after the 2004 Indian Ocean earthquake and tsumani. It was con-
ceived as an internet based system to manage disaster relief efforts during the response and
recovery phases after the disaster. In fact, the word “Sahana” means “relief” in Sinhalese,
one of the national languages of Sri Lanka.
The first owner of the intellectual property of Sahana was the Lanka Software Founda-
tion making Sahana become a global open source software project supported by volunteer
contributors and several national authorities. In 2009 a non-profit organization called the
Sahana Software Foundation was created to continue with the development of this project,
employing experts in disaster management.
Sahana is a web based automated system mainly focused on coordination and planning
of humanitarian operations. From the first implementation of 2004 the software has been
further improved and at the moment it contains the following tools (see [51] for more
details):

• Sahana Vesuvius: led by the US National Library of Medicine, it is focused on dis-

aster preparedness and response needs of the medical community. It is used to assist
hospitals, medical facilities and jurisdictions to tie victim records with missing people
reports.
• Sahana Mayon: it is a personnel and resource management tool to manage large num-
bers of resources. It contains disaster scenario, facility and staff management solutions.
• Sahana Eden: it is a flexible tool with several modules that can be used in the different
phases of disaster management: mitigation, preparedness, response and recovery. The
core modules are the following: Organization Registry (track active organizations to
provide opportunities for collaboration and coordination), Project Tracking (organize
projects and identify where the greatest needs are), Human Resources (manage volun-
teers and people involved in the project, keeping a contact list and tracking where they

[email protected]
Decision Aid Models and Systems for Humanitarian Logistics. A Survey 37

are and what skills they have), Inventory (manage inventories of items and requests
of warehouses, recording and automating transactions for sending and receiving ship-
ments), Assets (manage assets such as vehicles, radio equipment or power generators,
tracking their position and condition and ensuring they are used efficiently), Assess-
ments (collect and analyze information from assessments of different organizations
through the use of custom reports, graphs and maps), Scenarios & Events (plan for
several scenarios according to human resources, facilities, etc.), Mapping (integrated
mapping functionality providing situational awareness and supporting standard for-
mats from other sources and GIS, such as natural hazard risks, population or weather)
and Shelter Management (track information about shelters and people arriving and
departing). All these modules, together with some additional optional ones, can be
configured and personalized by the users to adapt them to their particular needs.

Sahana was first deployed by the Center of National Operations (CNO) at the 2005 Sri
Lanka tsunami. Since then, it has been used by several organizations to provide relief as-
sistance, such as in the 2005 Kashmir Earthquake of Pakistan (by the National Database &
Registration Authority of the Government of Pakistan), the 2006 Yogjakarta Earthquake of
Indonesia (by the Indonesian Relief Source) or the 2010 Floods in Venezuela (by the Gov-
ernment of Venezuela). The Sahana Eden tool, recently implemented, was first used for
disaster response in the 2010 Haiti Earthquake, being self-deployed by the Sahana Soft-
ware Foundation, also for the first time. The system has been used in many other relief
operations, see [50] for more details.

HFOSS. The project HFOSS [39] is an open source software project that participates in
several other smaller projects of different kinds. In particular, in the area of disaster man-
agement HFOSS has contributed to the Sahana project, already mentioned, and developed
other open source tools for humanitarian relief. One of these tools is Collabbit [41], a web
software for information sharing between organizations during emergencies. It has been
used, for example, by the Salvation Army Emergency Disaster Services of New York, to
coordinate the 2008 Thanksgiving Day Dinner program to feed more than 10,000 people or
to monitor cooling centers throughout New York for people seeking refuge from the heat
in the summer. Another tool developed by HFOSS is Posit [40], an Android application to
be used by rescue workers to map the disaster area with a mobile phone and communicate
with a central server and with other workers. It has been used, for example, by the nonprofit

[email protected]
38 Decision Aid Models for Disaster Management and Emergencies

organization ACDI/VOCA during food and health service operations in remote rural areas
in Haiti, after the 2010 earthquake.

DMIS. The Disaster Management Information System (DMIS) [84] is a web based tool
developed by The International Federation of Red Cross allowing access to real time infor-
mation on disaster trends, available resources and databases, in order to support an efficient
disaster preparedness and response. It is only accessible to Red Cross and Red Crescent
staff, delegations and Geneva headquarters.

LOGISTIX. In 2006, Doctors without Borders developed an inventory tool for medical
needs so called LogistiX [58]. It focuses on medicines inventory and it is used in the
normal work of the organization. It is only accesible to Doctors without Borders, and the
organization gives some courses in order to teach their volunteers how to use this tool.

It is important to remark that, apart from the systems introduced before, there are also
some other systems that are not specifically designed for humanitarian logistics but are
also used by NGOs for humanitarian relief. In this group it is worth noting the system so
called FleetWave [24], that is a web-based enterprise information system focused on fleet
management that has been used, for example, by ICRC, IFRC and WFP.
Note that these systems are mainly focused on information management, making it eas-
ier to access data regarding active organizations, human resources and beneficiaries, orders
already placed or to be performed, stocks and available material, maps, etc., and improving
the communication between agencies and organizations involved. This information is es-
sential to provide decision support and it is very useful in practice, significantly improving
the performance of relief operations. However, to the best of our knowledge, the existing
systems have not implemented any decision model, such as the ones available in the liter-
ature and commented in the previous section, to provide automated detailed inventory or
distribution plans as output.

2.4 Conclusions

Recent years have seen an explosion of literature regarding disaster and emergency
management. Mathematical models have become an important tool to tackle disaster and
emergency humanitarian logistics, helping in the decision aid processes appearing when
trying to respond to the consequences. Even though more research is needed to incorporate
the additional uncertainties faced in humanitarian logistics and the inherent difficulties due

[email protected]
Bibliography 39

to communication and coordination problems to the models, a lot of work has been done in
this direction. The review presented in this work and the chapter devoted to uncertainty in
humanitarian logistics for disaster management show the extent of these efforts.
There are in the literature several studies regarding future research lines is the area.
Van Wassenhove and Pedraza [94] present an extensive study on Supply Chain Manage-
ment best practices that are still to be adapted to humanitarian logistics. Caunhye et al. [19]
present a study on difficulties and future research directions that can guide the development
of more accurate models, insisting on the need to develop comprehensive optimization
models.
Our purpose in this work has been to link the review on models with the review on
decision support systems. From this work, it can be noted that humanitarian logistics re-
search remains highly fragmented. It is not frequent to find models that combine several
operations, while in the situations faced by agents many coordinated decisions have to be
taken. Moreover, in general, these models are not incorporated into Decision Support Sys-
tems which can be used in practise by the involved organizations. The systems already in
use are focused on information management, but do not have optimization tools in them. It
is our belief that it is necessary to shorten the gap between academics and practitioners in
order to be useful to the final decision makers.

Acknowledgments

This work was supported in part by the grant TIN2009-07901 from Ministerio de Cien-
cia e Innovación, Spain.

Bibliography

[1] Adıvar, B. and Mert, A. (2010). International disaster relief planning with fuzzy credibility,
Fuzzy Optim. Decis. Making 9, 4, pp. 413–433.
[2] Aftanas, J. M. (2007). Optimizing the prioritization of natural disaster recovery projects, Mas-
ter’s thesis, Air Force Institute of Technology, Ohio.
[3] Akkihal, A. R. (2006). Inventory pre-positioning for humanitarian operations, Master’s thesis,
Massachusetts Institute of Technology, Massachusetts.
[4] Altay, N. and Green, W. G. (2006). OR/MS research in disaster operations management, Eur.
J. Oper. Res. 175, 1, pp. 475–493.
[5] Ang, C. C. (2006). Optimized recovery of damaged electrical power grids, Master’s thesis,
Naval Postgraduate School, Monterey.
[6] Angelis, V. D., Mecoli, M., Nikoi, C. and Storchi, G. (2007). Multiperiod integrated routing
and scheduling of World Food Programme cargo planes in Angola, Comput. Oper. Res. 34, 6,
pp. 1601–1615.

[email protected]
40 Decision Aid Models for Disaster Management and Emergencies

[7] Apte, A. (2009). Humanitarian logistics: a new field of research and action, Found. Trends
Technol. Inf. Oper. Manag. 3, 1, pp. 1–100.
[8] Balasubramanian, S., Mohan, T., Pineiro, R. C., Jain, R., Routray, R. R., Shah, G., Verma,
A. and Voruganti, K. (2011). Methods, systems, and computer program products for disaster
recovery planning, United States Patent 7945537.
[9] Balcik, B. and Beamon, B. M. (2008). Facility location in humanitarian relief, Int. J. Logist.
Res. Appl. 11, 2, pp. 101–121.
[10] Beamon, B. M. and Kotleba, S. A. (2006a). Inventory management support systems for emer-
gency humanitarian relief operations in South Sudan, Int. J. Logist. Manag. 17, 2, pp. 187–212.
[11] Beamon, B. M. and Kotleba, S. A. (2006b). Inventory modelling for complex emergencies in
humanitarian relief operations, Int. J. Logist. Res. Appl. 9, 1, pp. 1–18.
[12] Ben-Tal, A., Chung, B. D., Mandala, S. R. and Yao, T. (2011). Robust optimization for emer-
gency logistics planning: risk mitigation in humanitarian relief supply chains, Transp. Res.
Part B 45, 8, pp. 1177–1189.
[13] Berkoune, D., Renaud, J., Rekik, M. and Ruiz, A. (2012). Transportation in disaster response
operations, Socio-Econ. Plan. Sci. 46, 1, pp. 23–32.
[14] Besiou, M., Pedraza Martinez, A. and Van Wassenhove, L. (2012). The effect of earmarked
funding on fleet management for relief and development, Tech. Rep. 2012/10/TOM/ISIC, IN-
SEAD.
[15] Blecken, A., Danne, C., Dangelmaier, W., Rottkemper, B. and Hellingrath, B. (2010). Optimal
stock relocation under uncertainty in post-disaster humanitarian operations, in Proceedings of
the 43rd Hawaii International Conference on System Sciences (Koloa, Hawaii), pp. 1–10.
[16] Bretschneider, S. and Kimms, A. (2012). Pattern-based evacuation planning for urban areas,
Eur. J. Oper. Res. 216, 1, pp. 57–69.
[17] Buzna, L., Peters, K., Ammoser, H., Kühnert, C. and Helbing, D. (2007). Efficient response
to cascading disaster spreading, Phys. Rev. Sect. E 75, 5-2, pp. 56107–56108.
[18] Campbell, A. M., Vandenbussche, D. and Hermann, W. (2008). Routing for relief efforts,
Transp. Sci. 42, 2, pp. 127–145.
[19] Caunhye, A., Nie, X. and Pokharel, S. (2012). Optimization models in emergency logistics:
A literature review, Socio-Econ. Plan. Sci. 46, pp. 4–13.
[20] Chang, S. E. (2010). Urban disaster recovery: a measurement framework and its application
to the 1995 Kobe earthquake, Disasters 34, 2, pp. 303–327.
[21] Chao, W., Yuanbin, L. and Xuliang, D. (2011). The research on human-computer interaction
of remote disaster-recovery system, in M. Zhou and H. Tan (eds.), Advances in Computer
Science and Education Applications. Part II, Communications in Computer and Information
Science, Vol. 202 (Springer, Berlin Heidelberg), pp. 460–466.
[22] Charles, A. (2010). Improving the Design and Management of Agile Supply Chains: Feedback
and Application in the Context of Humanitarian Aid, Ph.D. thesis, Université de Toulouse,
Toulouse, France.
[23] Chen, A., Zhao, J. and Chen, N. (2009). A recoverability assessment model in emergency
management, Tech. Rep. RTO-MP-IST-086-09, NATO Research and Technology Organisa-
tion.
[24] Chevin Fleet Solutions (2001). FleetWave,
http://www.chevinfleet.com/us/fleet_software_fleetwave_web_based_
enterprise_fleet_management_solution.asp, last access: May 7th, 2012.
[25] Chib, A. and Komathi, A. L. E. (2009). Extending the technology-community-management
model to disaster recovery: Assessing vulnerability in rural Asia, in Proceedings of the 3rd
International Conference on Information and Communication Technologies and Development
(Doha, Qatar), pp. 328–336.

[email protected]
Bibliography 41

[26] Chiu, Y. C. and Zheng, H. (2007). Real-time mobilization decisions for multi-priority emer-
gency response resources and evacuation groups: model formulation and solution, Transp.
Res. Part E 43, 6, pp. 710–736.
[27] Cimellaro, G. P., Reinhorn, A. M. and Bruneau, M. (2010). Framework for analytical quan-
tification of disaster resilience, Eng. Struct. 32, 11, pp. 3639–3649.
[28] Coffrin, C., Van Hentenryck, P. and Bent, R. (2011b). Strategic planning for power system
restoration, in Proceedings of the International Conference on Vulnerability and Risk Analysis
and Management (Hyattsville, Maryland), pp. 1–8.
[29] Coffrin, C., Van Hentenryck, P. and Bent, R. (2011c). Strategic stockpiling of power system
supplies for disaster recovery, in Power and Energy Society General Meeting, 2011 IEEE (San
Diego, California), pp. 1–8.
[30] Crowther, K. G. and Haimes, Y. Y. (2005). Application of the inoperability input-output model
(IIM) for systemic risk assessment and management of interdependent infrastructures, Syst.
Eng. 8, 4, pp. 323–341.
[31] de Mel, S., D., M. and Woodruff, C. (2008). Mental health recovery and economic recov-
ery after the tsunami: high-frequency longitudinal evidence from Sri Lankan small business
owners, Soc. Sci. Med. 66, 3, pp. 582–595.
[32] Doerner, K. F., Gutjahr, W. J. and Nolz, P. C. (2009). Multi-criteria location planning for
public facilities in tsunami-prone coastal areas, OR Spectr. 31, 3, pp. 651–678.
[33] Eiselt, H. A. and Marianov, V. (2012). Mobile phone tower location for survival after natural
disasters, Eur. J. Oper. Res. 216, 3, pp. 563–572.
[34] El-Anwar, O., El-Rayes, K. and Elnashai, A. S. (2010). Maximizing the sustainability of
integrated housing recovery efforts, J. Constr. Eng. Manag. 136, 7, pp. 794–802.
[35] Fiedrich, F. and Burghardt, P. (2007). Agent-based systems for disaster management, Com-
mun. ACM - Emerg. Response Inf. Syst.: Emerg. Trends Technol. 50, 3, pp. 41–42.
[36] Ghesquiere, F. and Mahul, O. (2007). Sovereign natural disaster insurance for developing
countries: A paradigm shift in catastrophe risk financing, World Bank Policy Research 4345,
World Bank.
[37] Gibbons, D. E. and Samaddar, S. (2009). The maximal expected coverage relocation problem
for emergency vehicles, Decis. Sci. 40, 2, pp. 351–371.
[38] Hale, T. and Moberg, C. R. (2005). Improving supply chain disaster preparedness: a decision
process for secure site location, Int. J. Phys. Distrib. Logist. Manag. 35, 3/4, pp. 195–207.
[39] HFOSS Institute (2006). HFOSS, http://dev.hfoss.org/, last access: February 6th,
2012.
[40] HFOSS Institute (2008). Posit, http://posit.hfoss.org/, last access: February 6th,
2012.
[41] HFOSS Institute (2010). Collabbit, http://collabbit.org/about, last access: February
6th, 2012.
[42] Huang, M., Smilowitz, K. and Balcik, B. (2012). Models for relief routing: equity, efficiency
and efficacy, Transp. Res. Part E 48, 1, pp. 2–18.
[43] Jin, M. and Ekşioǧlu, B. (2010). Optimal routing of vehicles with communication capabilities
in disasters, Comput. Manag. Sci. 7, 2, pp. 121–137.
[44] Jotshi, A., Gong, Q. and Batta, R. (2009). Dispatching and routing of emergency vehicles in
disaster mitigation using data fusion, Socio-Econ. Plan. Sci. 43, 1, pp. 1–24.
[45] Jung, C., Johnson, M., Trick, M. and Williams, J. (2007). Mathematical models for recon-
struction planning in urban areas, Tech. rep., Heinz College, Carnegie Mellon University.
[46] Kaklauskas, A., Amaratunga, D. and Haigh, R. (2009). Knowledge model for post-disaster
management, Int. J. Strateg. Prop. Manag. 13, 2, pp. 117–128.
[47] Karlaftis, M. G., Kepaptsoglou, K. L. and Lambropoulos, S. (2007). Fund allocation for trans-
portation network recovery following natural disasters, J. Urban. Plan. Dev. 133, 1, pp. 82–89.

[email protected]
42 Decision Aid Models for Disaster Management and Emergencies

[48] Kongsomsaksaku, S., Yang, C. and Chen, A. (2005). Shelter location-allocation model for
flood evacuation planning, J. East. Asia. Soc. Transp. Stud. 6, pp. 4237–4252.
[49] Kung, H. Y., Chen, C. H. and Ku, H. H. (2012). Designing intelligent disaster prediction
models and systems for debris-flow disasters in Taiwan, Expert Syst. Appl. 39, 5, pp. 5838–
5856.
[50] Lanka Software Foundation (2004a). SAHANA deployments,
http://wiki.sahanafoundation.org/doku.php/deployments:start, last access:
February 6th, 2012.
[51] Lanka Software Foundation (2004b). SAHANA products,
http://sahanafoundation.org/products/, last access: February 6th, 2012.
[52] Lanka Software Foundation (2004c). SAHANA system,
http://sahanafoundation.org/, last access: February 6th, 2012.
[53] Lodree, E. J. and Taskin, S. (2008). An insurance risk management framework for disaster
relief and supply chain disruption inventory planning, J. Oper. Res. Soc. 59, 5, pp. 674–684.
[54] Losada, C., Scaparra, M. P. and O’Hanley, J. R. (2012). Optimizing system resilience: A
facility protection model with recovery time, Eur. J. Oper. Res. 217, 3, pp. 519 – 530.
[55] Luckey, T. S. (2009). Key stages of disaster recovery planning for time-critical business infor-
mation technology systems, Master’s thesis, University of Oregon, Oregon.
[56] Matisziw, T. C., Murray, A. T. and Grubesic, T. H. (2010). Strategic network restoration, Netw.
Spat. Econ. 10, pp. 345–361.
[57] Maya, P. and Sörensen, K. (2011). A GRASP metaheuristic to improve accessibility after a
disaster, Disaster Relief 33, 3, pp. 525–542.
[58] Médecins Sans Frontières (2006). LogistiX,
ftp://support.geneva.msf.org/permanent/LOG/SupplyChainManagement, last ac-
cess: February 6th, 2012.
[59] Mehlhorn, S., Racer, M., Ivey, S. and Lipinski, M. (2011). A single-objective recovery phase
model, Int. J. Inf. Technol. Proj. Manag. 2, 3, pp. 53–71.
[60] Mehlhorn, S. A. (2009). Method for Prioritizing Highway Routes for Reconstruction After a
Natural Disaster, Ph.D. thesis, University of Memphis.
[61] Miles, S. B. and Chang, S. E. (2006). Modeling community recovery from earthquakes,
Earthq. Spectr. 22, pp. 439–458.
[62] Nagurney, A., Yu, M. and Qiang, Q. (2011). Supply chain network design for critical needs
with outsourcing, Pap. Reg. Sci. 90, 1, pp. 123–142.
[63] Natarajarathinam, M., Capar, I. and Narayanan, A. (2009). Managing supply chains in times
of crisis: a review of literature and insights, Int. J. Phys. Distrib. Logist. Manag. 39, 7, pp.
535–573.
[64] Nelson, S. C., Harris III, A. F. and Kravets, R. (2007). Event-driven, role-based mobility
in disaster recovery networks, in Proceedings of the Second ACM Workshop on Challenged
Networks (Montreal, Canada), pp. 27–34.
[65] Nolz, P., Doerner, K., Gutjahr, W. and Hartl, R. (2010a). A bi-objective metaheuristic for
disaster relief operation planning, in C. Coello Coello, C. Dhaenens and L. Jourdan (eds.),
Advances in Multi-Objective Nature Inspired Computing, Studies in Computational Intelli-
gence, Vol. 272 (Springer, Berlin Heidelberg), pp. 167–187.
[66] Nolz, P. C., Doerner, K. F. and Hartl, R. F. (2010b). Water distribution in disaster relief, Int. J.
Phys. Distrib. Logist. Manag. 40, 8/9, pp. 693–708.
[67] Onyango, M. A. (2008). Humanitarian responses to complex emergencies, in Q. K. Heggen-
hougen (ed.), International Encyclopedia of Public Health (Academic Press, San Diego).
[68] Orabi, W., El-Rayes, K., Senouci, A. B. and Al-Derham, H. (2009). Optimizing postdisaster
reconstruction planning for damaged transportation networks, J. Constr. Eng. Manag. 135, 10,
pp. 1039–1048.

[email protected]
Bibliography 43

[69] Orabi, W., Senouci, A. B., El-Rayes, K. and Al-Derham, H. (2010). Optimizing resource
utilization during the recovery of civil infrastructure systems, J. Manag. Eng. 26, 4, pp. 237–
246.
[70] Ortuño, M., Tirado, G. and Vitoriano, B. (2011). A lexicographical goal programming based
decision support system for logistics of Humanitarian Aid, TOP 19, pp. 464–479.
[71] Pedraza Martinez, A. and Van Wassenhove, L. N. (2012). Vehicle replacement in the interna-
tional committee of the red cross, Prod. Oper. Manag. 10.1111/j.1937-5956.2011.01316.x.
[72] Permann, M. R. (2007). Genetic algorithms for agent-based infrastructure interdependency
modeling and analysis, in Proceedings of the 2007 Spring Simulation Multiconference, Vol. 2
(Norfolk, Virginia), pp. 169–177.
[73] Qin, J., Xing, Y., Wang, S., Wang, K. and Chaudhry, S. S. (2012). An inter-temporal resource
emergency management model, Comput. Oper. Res. 39, 8, pp. 1909–1918.
[74] Quarantelli, E. L. (2006). Catastrophes are different from disasters: some implications for cri-
sis planning and managing drawn from Katrina, http://understandingkatrina.ssrc.
org/Quarantelli/, last access: February 20th, 2012.
[75] Rose, A. (2007). Economic resilience to natural and man-made disasters: multidisciplinary
origins and contextual dimensions, Environ. Hazard. 7, 4, pp. 383–398.
[76] Rottkemper, B., Fischer, K. and Blecken, A. (2012). A transshipment model for distribution
and inventory relocation under uncertainty in humanitarian operations, Socio-Econ. Plan. Sci.
46, 1, pp. 98–109.
[77] Saadatseresht, M., Mansourian, A. and Taleai, M. (2009). Evacuation planning using multi-
objective evolutionary optimization approach, Eur. J. Oper. Res. 198, 1, pp. 305–314.
[78] Saleem, K., Luis, S., Deng, Y., Chen, S.-C., Hristidis, V. and Li, T. (2008). Towards a busi-
ness continuity information network for rapid disaster recovery, in Proceedings of the 2008
International Conference on Digital Government Research (Montreal, Canada), pp. 107–116.
[79] Semsch, E., Jakob, M., Doubek, J. and Pechoucek, M. (2008). Using player models to improve
robustness of HTN plans in multi-agent domains, in Proceedings of the 27th Workshop of the
UK Planning and Scheduling Special Interest Group (Edinburgh, UK), pp. 1–7.
[80] Sheu, J. B. (2007). An emergency logistics distribution approach for quick response to urgent
relief demand in disasters, Transp. Res. Part E 43, 6, pp. 687–709.
[81] Solano, E. (2010). Research brief: Methods for assessing vulnerability of critical infrastruc-
ture, Tech. rep., Institute for Homeland Security Solutions.
[82] The Fritz Institute (2003). Humanitarian Logistics Software (HLS),
http://www.fritzinstitute.org/prgTech-HLS.htm, last access: February 6th, 2012.
[83] The Fritz Institute (2007). HELIOS,
http://www.fritzinstitute.org/prgTech-HELIOS_Overview.htm, last access:
February 6th, 2012.
[84] The International Federation of Red Cross (2001). The Disaster Management Information
System (DMIS),
https://www-secure.ifrc.org/DMISII/Pages/00_Home/login.aspx, last access:
February 6th, 2012.
[85] The Pan American Health Organization (1992). The Humanitarian Supply Management Sys-
tem (SUMA),
http://www.disaster-info.net/SUMA/english/index.htm, last access: February 6th,
2012.
[86] The Pan American Health Organization (2005). Logistics Support System (LSS),
http://www.lssweb.net/, last access: February 6th, 2012.
[87] Tomasini, R. and van Wassenhove, L. (2009). Humanitarian Logistics (Palgrave Macmillan).
[88] Tovia, F. (2007). An emergency logistics response system for natural disasters, Int. J. Logist.:
Res. Appl. 10, 3, pp. 173–186.

[email protected]
44 Decision Aid Models for Disaster Management and Emergencies

[89] Toyasaki, F. and Wakolbinger, T. (2012). Impacts of earmarked private donations for disaster
fundraising, A. Oper. Res. 10.1007/s10479-011-1038-5.
[90] Tzeng, G. H., Cheng, H. J. and Huang, T. D. (2007). Multi-objective optimal planning for
designing relief delivery systems, Transp. Res. Part E 43, 6, pp. 673–686.
[91] Uchida, K. (2012). A model evaluating effect of disaster warning issuance conditions on “cry
wolf syndrome” in the case of a landslide, Eur. J. Oper. Res. 218, 2, pp. 530–537.
[92] Van Hentenryck, P., Bent, R. and Coffrin, C. (2010). Strategic planning for disaster recov-
ery with stochastic last mile distribution, in H. C. Heard, I. V. Borg, N. L. Carter and C. B.
Raleigh (eds.), Integration of AI and OR Techniques in Constraint Programming for Com-
binatorial Optimization Problems, Lecture Notes in Computer Science, Vol. 6140 (Springer,
Berlin Heidelberg), pp. 318–333.
[93] Van Wassenhove, L. N. (2006). Humanitarian aid logistics: supply chain management in high
gear, J. Oper. Res. Soc. 57, 5, pp. 475–489.
[94] Van Wassenhove, L. N. and Pedraza, A. J. (2012). Using or to adapt supply chain management
best practices to humanitarian logistics, Intl. Trans. in Op. Res. 19, 1–2, pp. 307–322.
[95] Vitoriano, B., Ortuño, M. T. and Tirado, G. (2009). HADS, a goal programming-based hu-
manitarian aid distribution system, J. Multi-Criteria Decis. Anal. 16, 1-2, pp. 55–64.
[96] Vitoriano, B., Ortuño, M. T., Tirado, G. and Montero, J. (2011). A multi-criteria optimization
model for humanitarian aid distribution, J. Glob. Optim. 51, 2, pp. 189–208.
[97] Xu, N., Guikema, S. D., Davidson, R. A., Nozick, L. K., Çag̃nan, Z. and Vaziri, K. (2007). Op-
timizing scheduling of post-earthquake electric power restoration tasks, Earthq. Eng. Struct.
Dyn. 36, pp. 265–284.
[98] Yi, W. and Kumar, A. (2007). Ant colony optimization for disaster relief operations, Transp.
Res. Part E 43, 6, pp. 660–672.
[99] Yi, W. and Özdamar, L. (2007). A dynamic logistics coordination model for evacuation and
support in disaster response activities, Eur. J. Oper. Res. 179, pp. 1177–1193.
[100] Zobel, C. W. and Khansa, L. (2011). Characterizing multi-event disaster resilience, Comput.
Oper. Res. 10.1016/j.cor.2011.09.024.

[email protected]
Chapter 3

Uncertainty in Humanitarian Logistics for

Disaster Management. A Review

F. Liberatore 1,∗ , C. Pizarro 1 , C. Simón de Blas 1 , M.T. Ortuño 2 , and B. Vitoriano 2

1 Universidad Rey Juan Carlos, Madrid, Spain
2 Universidad Complutense de Madrid, Madrid, Spain

Given their nature, disasters are generally characterized by a high level of uncertainty. In
fact, both their occurrence and their consequences are not easily anticipated. Thus, NGOs
and civil protection often have to take decisions and plan for their operations without having
the possibility of relying on exact or complete information on the magnitude of the disaster.
Over the years, a number of works and methodologies that address uncertainty in Disaster
Management have been presented in the literature. In this chapter we review different forms
of tackling uncertainty in Humanitarian Logistics for Disaster Management and propose a
classification of the advances in this research field.

3.1 Introduction

3.1.1 Humanitarian logistics and Disaster Management

Different definitions can be found in literature for Humanitarian Logistics, but probably
the most accurate one is the one provided in the Humanitarian Logistics Conference of
2004, by the Fritz Institute:
“Humanitarian Logistics is the process of planning, implementing and controlling the
efficient, cost-effective flow and storage of goods and materials as well as related informa-
tion, from the point of origin to the point of consumption for the purpose of meeting the
end beneficiary’s requirements and alleviate the suffering of vulnerable people.
It encompasses a set of activities, including preparation, planning, procurement, trans-
portation, storage, history and customs control.”
∗ Corresponding author. Contact: [email protected]

B. Vitoriano et al. (eds.), Decision Aid Models for Disaster Management 45

and Emergencies, Atlantis Computational Intelligence Systems 7,
DOI: 10.2991/978-94-91216-74-9_3, Ó Atlantis Press 2013

[email protected]
46 Decision Aid Models for Disaster Management and Emergencies

As it can be seen, the main difference between business or military logistics and human-
itarian logistics is based on the final aim of alleviating the suffering of vulnerable people.
Despite of the fact that this is not an exclusive point of disaster management (each human-
itarian aid operation as those performed by the World Food Program or the World Health
Organization into underdeveloped areas or many of the cooperation for development activ-
ities usually have this common final aim), the usual context where humanitarian logistics
has been studied is into Disaster Management. Although there are other studies in the field
of humanitarian logistics (e.g., Pedraza-Martinez et al. (2011), present a study of field fleet
management not under emergency), it is in this context where the main differences with
business logistics proposed in Balcik and Beamon (2008) fit better:

• unpredictable demand in terms of timing, geographic location, type of commodity,

quantity of commodity;
• short lead time and suddenness of demand for large amounts of a wide variety of
products and services;
• lack of initial resources in terms of supply, human resources, technology, capacity and
funding.

As it can be seen from these characteristics, the first one is related with the uncertainty.
Disaster management decision makers usually have to make their decisions under risk or
uncertainty coming from different sources.

3.1.2 Uncertainty and Risk

Uncertainty is defined by the Oxford Dictionary as the state of being uncertain; Un-
certain is defined as not able to be relied on; not known or definite. Risk is defined (among
other senses) as the possibility that something unpleasant or unwelcome will happen, and
Risk Management as the forecasting and evaluation of financial risks together with the
identification of procedures to avoid or minimize their impact. In our context this concept
will be extended to others than financial risks.
From these definitions, it can be seen that uncertainty and risk is not the same, and it is
more or less accepted in the literature that the term Decision under Risk refers to decision
making where the unknown states or consequences are represented through probabilities,
and Decision under Uncertainty is used when other characterizations appear to deal with
the not well-known or definite information. Nevertheless, through this paper no distinction
will be made in order to give a global picture of the topic.

[email protected]
Uncertainty in Humanitarian Logistics for Disaster Management. A Review 47

Typologies of uncertainties are available in the literature of different disciplines. Some

of the most typical ones are the following (see Moss and Schneider (2000)):

Problems with data

• Missing components or errors in the data.

• “Noise” in the data associated with biased or incomplete observations.
• Random sampling error and biases (non-representativeness) in a sample.

Problems with models

• Known processes but unknown functional relationships or errors in the structure of the
model.
• Known structure but unknown or erroneous values of some important parameters.
• Known historical data and model structure, but reasons to believe parameters or model
structure will change over time.
• Uncertainty regarding the predictability (e.g., chaotic or stochastic behaviour) of the
system or effect.
• Uncertainty introduced by approximation techniques used to solve a set of equations
that characterize the model.

Other sources of uncertainty

• Ambiguously defined concepts and terminology.

• Inappropriate spatial/temporal units.
• Inappropriateness of/lack of confidence in underlying assumptions.
• Uncertainty due to projections of human behaviour (e.g., future consumption patterns,
or technological change), which is distinct from uncertainty due to “natural” sources
(e.g., climate sensitivity, chaos).

In the context of disaster management, it is important to distinguish that some of the

sources of uncertainty come from the unknown future, because decisions are made before
the event happen. In this sense, we can distinguish two main phases:

• Pre-event phase: actions and decisions developed before the event has happened; they
are focused on mitigation and preparedness.

[email protected]
48 Decision Aid Models for Disaster Management and Emergencies

• Post-event phase: actions and decisions developed after the event has happened; they
include response and recovery activities.

The pre-event activities usually are developed assuming future scenarios, which should
be characterized, and decisions can be common for all of them, or defined for each of them
in case they would happen. This kind of studies includes risk management (e.g., natural
risks, social/vulnerability risks), where usually a probability distribution is associated to
the risk scenarios. But other possibilities as belief structures or evidential reasoning could
be also considered.
The post-event decisions usually include uncertainty coming from the lack of informa-
tion or some inherent ambiguity to it. Semantic uncertainty about the consequences is a
common issue, among others. For instance, when classifying disaster consequences, even
the frontiers between emergency, disaster or catastrophe are not crisply defined.
This Chapter is organized as follows: Section 3.2 focuses on the uncertain parameters
usually considered in Humanitarian Logistics, and the main methodologies used to deal
with the uncertainty are shown in Section 3.3. The specific literature review is compiled
and commented in Section 3.4, and, finally, some conclusions are drawn.

3.2 Uncertainty in disaster management

Disaster management includes activities performed to be prepared in case of a disaster

takes place, where the future is really unknown, so this preparedness is developed under an
uncertain future. Some kind of natural hazards, as tropical storms, usually are monitored
and as they approach the uncertainty about the future evolution of the event is disappear-
ing, but not about the consequences. When the event takes place, chaos and unpredictable
human behavior cause information flow transmissions breaks which result in information
lacks and uncertainty in structure, range and number of casualties or damages. Further-
more, when the affected population provides information, sources may be diverse causing
chaotic data without the aid of support tools and time for verification. Additionally, re-
porters, NGOs, and other emergency agencies usually act as the main information sources
leading to disaggregated data, instead of structured information as required by aid decision
management tools.
Figure 3.1 summarizes the different parameters under uncertainty considered in the
literature in the context of Humanitarian Logistics. The uncertain parameters are grouped
according to the associated system element.

[email protected]
Uncertainty in Humanitarian Logistics for Disaster Management. A Review 49

Fig. 3.1 Different parameters with uncertainty considered in the literature in the context of Human-
itarian Logistics

Most of the contributions in Humanitarian Logistics research have focused on demand

uncertainty. It can be divided in two groups: demand regarding the number of affected
population and demand of required relief goods. In the former, we may find contributions
dealing with the number of rescues (i.e., number of people to be evacuated), number of
affected by a disease, and number of injured. In the latter, the literature considers first
aid supplies, response equipments, potable water, food, medicines, disease treatments, and
required relief workers for commodities handling. De la Torre (2001) states that consider-
ing uncertainty in the supply of relief goods is more important as underestimating it might
result in substantial delays in the distribution.
Partial or completed supply losses at storage locations have also been considered. In
this category uncertainties regarding products quality, availability, and production capacity

[email protected]
50 Decision Aid Models for Disaster Management and Emergencies

in a disaster’s wake are included. A similar type of uncertainty is the one concerning
affected areas and demand locations, which includes magnitude of damages to supply
stores and occurrence of the disaster. Also, some contributions considered uncertainties
regarding product prices due to scarcity and survival rates.
Finally, transportation network uncertainties include links capacity (e.g., in terms of
number of evacuees and vehicles), reliability, availability (e.g., roads, bridges, crossovers),
and traversing time.
A conceptual map of the relations among the different elements presented is illustrated
in Figure 3.2. In the map, the arrows define the cause-effect relationship identified in the
real case data sets presented in the literature.

Fig. 3.2 Conceptual relationships among the uncertainty types

3.3 Methodologies Description

The main methodologies found in literature related with decision aid models with un-
certainty for disaster management, can be classified into those describing uncertainty based
on probabilities, and those based on other type of uncertainty characterizations such as
fuzzy logic.
Once the uncertainty is characterized, suitable techniques to develop decision aid mod-
els based on this information can be found in the literature. A first description about the
most used techniques when dealing with uncertainty is reported in the following.

[email protected]
Uncertainty in Humanitarian Logistics for Disaster Management. A Review 51

3.3.1 Risk Mapping and Probabilistic Values

Risk maps originated from models of disease transmission based on spatial and tem-
poral data. These models incorporate, to varying degrees, epidemiological, entomological,
climatic and environmental information.
As stressed in Huang et al. (2007), a risk map of natural disasters is an atlas of a com-
munity or geographical zone that identifies the places and the buildings that might sustain
heavy damages caused by natural disasters. The objective of a risk map is to identify the
probability of the occurrence of a specific hazard, as well as the contingent vulnerabilities
of exposed elements at risk, intensity and areas of impact, within a period of time. The
first risk maps were defined in 1800, when insurance companies where hit by fires in major
cities. Mapping was used to measure flammable exposures by placing pins on a map show-
ing the location of insurance buildings, limiting the number of insurances in a given city or
block. These pins maps where used until 1960’s and, because their creation was too time
consuming, they began to be replaced by more sophisticated computer generated maps that
took into account frequency and severity distribution to estimate potential losses.
Nowadays, risk maps are developed by international research groups, such as United
Nations Environment Program (UNEP), International Bank for Reconstruction and De-
velopment, The World Bank and Columbia University (in Dilley (2005)) and the Federal
Emergency Management Agency (FEMA1 ). Also, risk companies provide risk maps for
budget purposes (Insurance Center Associates2 ). Many countries have national regulations
forcing to develop risk maps, which are obtained by governmental entities (national or lo-
cal) of environment or safety. Risk maps are crucial for pre-event activities as mitigation
and preparedness, being usually inputs for decision aid models.
Other probabilistic values to model uncertainty, but not exactly considered risk map-
ping, are also included in this section, especially those focused on the uncertainty related
with the status of the infrastructure or reliability, after a disaster has happened. They are
used over the network, so a map is included, but they are useful in a post-event phase more
than for pre-event for risk management. Their treatment more as a quality of the elements
than as a probability distribution, suggests separating it from those stochastic programming
models, where probability distributions are considered themselves.
1 http://www.fema.gov/hazard/index.shtm
2 http://www.inscenter.com/info-center/disaster-planning/risk-profile

[email protected]
52 Decision Aid Models for Disaster Management and Emergencies

3.3.2 Probability Distributions and Stochastic Programming

Stochastic Programming is an approach for modeling optimization problems when the

parameters are uncertain, but assumed to lie in some given set of possible values following
a probability distribution. In such a case, one might seek a solution that is feasible for
all possible parameter choices and optimizes the expected value of a given function of the
decisions and the random variables.
Stochastic Programming models try to take advantage of the fact that probability distri-
butions governing the data are known or can be estimated. These probability distributions
can be estimated from data that have been collected over time, or in the absence of data
from future periods, these distributions will have to be elicited from some accompanying
model, which in its simplest form might derive solely from the prior beliefs of the decision
maker.
The most widely applied and studied stochastic programming models are two-stage
linear programs. Here the decision maker takes some action in the first stage, after which a
random event occurs affecting the outcome of the first-stage decision. A recourse decision
can then be made in the second stage that compensates for any bad effects that might have
been experienced as a result of the first-stage decision. The optimal policy from such a
model is a single first-stage policy and a collection of recourse decisions (a decision rule)
defining which second-stage action should be taken in response to each random outcome.
Stochastic programming models usually are large dimensions models, implying the use
of advanced mathematical programming techniques. Perhaps this is one of the reasons,
in addition to its good fit to the problem characteristics, to explain the large number of
academic articles devoted to this methodology in the last years into the area of disaster
management.

3.3.3 Robust Optimization

Robust Optimization is a methodology, combined with computational tools, to process

optimization problems in which the data are uncertain and only known to belong to some
uncertainty set. In contrast with Stochastic Optimization which models the uncertainty
through a probabilistic description, Robust Optimization models the possible set of values,
but nothing is said about their probabilities. Instead of seeking to immunize the solution in
some probabilistic sense to stochastic uncertainty, the decision-maker constructs a solution
that is admissible in some sense (e.g., quasi-optimal) for any realization of the uncertainty
in the given set. Robust optimization can be especially suitable in absence of data, or when

[email protected]
Uncertainty in Humanitarian Logistics for Disaster Management. A Review 53

there is no interest to give more importance to some values of the parameter than to others.
Computational tractability is also a primary motivation to choose this methodology. Cur-
rently, Evolutionary Computing introducing robust optimization with multiple objectives
could be seen as one of the most promising methodologies related, being the paper Dev
and Gupta (2006) a seminal paper of the area, very interesting and understandable written
in a very pedagogic way.

3.3.4 Simulation Models

A simulation model is a mathematical model of a system or process which tries to repli-

cate in the computer a system’s behaviour. The most extended use of simulation models
is to represent dynamic systems, but especially when including uncertainty, static systems
also are studied through simulation in order to obtain a simulated sample of the output.
Techniques to obtain values of random variables are known as Monte Carlo simulation. In
spite of having several shortcomings, simulation models are still widely used in modelling
some humanitarian logistic problems, as the egress problem. Simulation models have also
been widely used to incorporate evacuees’ behavioural effects on the evacuation process.
Compared to optimization models, simulation models have a major lack. In fact, they
do not provide the best strategy or decision to manage the system; they allow evaluating the
performance of a strategy. Choosing a best strategy or decision implies experimenting with
the model on a What if basis; i.e., by defining previously the decisions to be considered
and using the simulation to evaluate them.
Another limitation of simulation models with uncertainty is that, in order to obtain re-
sults of system performance measures within required confidence intervals, large amounts
of data and correspondingly large amounts of computation time and storage space are re-
quired. Further, it is not possible to generalize simulation results obtained from one system
to another.
Simulation models, however, allow representing systems with a higher degree of accu-
racy, and they can be also used to complement analytical models in approximating system
performance measures.

3.3.5 Fuzzy Sets

There are other concepts and notions of theories and models addressing imprecision,
uncertainty, ambiguity and partial knowledge with respect to probability concept.

[email protected]
54 Decision Aid Models for Disaster Management and Emergencies

For years, fuzzy modeling has been one of the most relevant issues in qualitative anal-
ysis, given the linguistic capabilities of fuzzy logic, but now it is the base of many decision
support models whose output is quantitative. In fact, fuzzy set theory, firstly introduce in
Zadeh (1965), enables the use of natural language, implements relative concepts like big,
old, strong, etc., and processes them on definite mathematical principles. Using fuzzy logic,
it is possible to process objects that partially belong to more than one category. There is no
need to generalize reality to fit it into classes; rather the degree of membership to the cat-
egory is given. Note that, due to linguistic and numeric requirements, the fuzzy-modeling
process has generally to deal with an important trade-off between the accuracy and the in-
terpretability of the model. In other words, the model is expected to provide high numeric
precision while incurring into a loss of linguistic descriptive power as little as possible.

3.4 Annotated Bibliography

In the last decades, Humanitarian Logistics has drawn the interest of both practitioners
and academics. As a result, the scientific contributions in this field have grown exponen-
tially over the last 20 years. On behalf of the extremely complexity of these problems due
to their inherit uncertainty, the major contributions in the literature deal in a deterministic
context with the aim of approximate to real problems. As Ben-Tal et al. (2010) stress we
can find high uncertainty in disaster consequences and human behaviour in catastrophic
situations. Therefore, humanitarian management is an important application area for opti-
mization with uncertainty. As they underline, the penalty cost for an infeasible solution is
loss of life or property and it is very difficult to estimate or forecast the uncertainty model
parameters due to unexpected human behaviour and nature of disaster.
We contribute with an annotated bibliography on Humanitarian Logistics for Disaster
Management under uncertainty. Our review focuses on papers dealing with Humanitarian
Logistics management where there exists a lack of data related to demand (in terms of
rescues, goods, etc.), locations (available suppliers, affected areas, etc.), supplies (hospitals,
assets, food, water, etc.) or any related issue.
Tables 3.1 to 3.5 presents a summary of the main contributions to the topic, briefly
described later. The first column of the tables shows the first author of the article and its
publishing year. The second column describes the problem type dealt by the authors. In
the next four columns we categorize the works depending on the emergency management
phases addressed. Following the emergency management phases definitions provided in

[email protected]
Uncertainty in Humanitarian Logistics for Disaster Management. A Review 55

Altay and Green (2006) we consider mitigation as the application of measures that will
either prevent the onset of a disaster or reduce the impacts should one occur; prepared-
ness as the set of activities that prepare the community to respond when a disaster occurs;
response as the employment of resources and emergency procedures as guided by plans to
preserve life, property, the environment, and the social, economic, and political structure
of the community; and recovery as the set of actions taken in the long term after the im-
mediate impact of the disaster has passed to stabilize the community and to restore some
semblance of normalcy.
The seventh column presents the uncertainty inherent to the problem considered. The
eighth column presents the objective or goal considered and, finally, last column describes
the methodology used by the authors to deal with the uncertainty. They are organized by
publication date.

[email protected]
 56

Table 3.1 Review of the scientific contributions to Humanitarian Logistics employing Risk Maps and Probabilistic Values
Pre- Post-
Author Problem type Disaster Disaster Uncertainty types Objective function Methodology
M P R RV
Location of oil spill
Belardo et
response   Event location Multiple objective Risk maps
al. 1984
equipment
Availability
ofemergency resources,
Huang et al. Maximal Coverage Maximize cover of critical Risk maps and Robust
 service and travel
2007 model network elements. optimization
times, emergency
frequency.
Network
Günneç and reliability and
Links status (available Risk maps and
Salman expected  Network reliability
or not) Stochastic optimization
2011 post-disaster

[email protected]
performance
Multi-objective model: risk
Nolz et al. Last-mile
  Arcs associated risk measure, coverage, total Probabilistic values
2011 distribution
travel time.
Multi-criteria model: time,
Vitoriano et Last-mile Arcs reliability and
  cost, equity, priority, Probabilistic values
al. 2011 distribution security.
reliability, security.
Decision Aid Models for Disaster Management and Emergencies
 Table 3.2 Review of the scientific contributions to Humanitarian Logistics employing Probability Distributions and Stochastic Programming
Pre- Post-
Author Problem type Disaster Disaster Uncertainty types Objective function Methodology
M P R RV
Initialsearch-and- Affected area, Stochastic optimization
Fiedrich et Minimize the total number
rescue period after  causalities, via probabilistic
al. 2000 of fatalities
strong earthquakes transportation times constraints
Multiple-depot, Minimize the sum of the Stochastic optimization
Chan et al.
multiple-vehicle,  Demand routing, depot and vehicle via probabilistic
2001
location-routing fixed-costs constraints
Multi-commodity, Affected areas,
Barbarosoglu Minimize total cost and Stochastic optimization
multi-modal, network  demand, transportation
Arda 2004 expected scenario cost via scenario analysis
flow routes
Determinationof a Minimize rescue
rescue resource equipment expected
Chang et al. Stochastic optimization
distribution system  Rescue demand shipping distance, local
2007 via scenario analysis
for urban flood bases setup and acquisition

[email protected]
disasters and transportation costs.
Uncertainty in Humanitarian Logistics for Disaster Management. A Review

Capacity Stochastic
Yazizi requirements and Minimize global travel optimization via
  Link capacity changes
Ozbay 2007 desirable shelter time probabilistic
locations constraints
Resourceallocation
Zhu et al. problem to Stochastic optimization
 Resource demand Minimize total rescue cost
2008 minimize the total via scenario analysis
rescue cost
57
 58
Pre- Post-
Author Problem type Disaster Disaster Uncertainty types Objective function Methodology
M P R RV
Medicamentsand
treatments pick-up
Stochastic optimization
Shen et al. and delivery in Supplies, demand and
  Minimize unmet demand via probabilistic
2009 large scale traffic conditions
constraints
bio-terrorism
emergency
Evacuationplan for Stochastic optimization
Song et al. Minimize total evacuation
transit-dependent  Demand via probabilistic
2009 time
residents constraints.
Hentenryck Singlecommodity Transportation routes, Stochastic optimization
 Minimize last delivery time
et al. 2010 allocation problem times, affected areas via scenario analysis
Locationand amount of
Storage and
Mete demand, available Minimize Transportation
distribution Stochastic optimization
Zabinsky   transportation routes duration and unsatisfied
problem of medical via scenario analysis
2010 and transportation demand
supplies

[email protected]
times
Minimize the expected
Location and
costs, commodity
Rawls and quantities of Demand and
acquisition and stocking, Stochastic optimization
Turnquist emergency supplies   transportation network
supplies shipments, unmet via scenario analysis
2010 to be availability
demand and unused
pre-positioned
material.
Allocationof
Salmerón budget to acquire Location and amount Minimize the expected Stochastic optimization
Apte 2010 and position relief
 via scenario analysis
of demand number of casualties
assets
Decision Aid Models for Disaster Management and Emergencies
 Pre- Post-
Author Problem type Disaster Disaster Uncertainty types Objective function Methodology
M P R RV
Networkreliability
Günneç and Risk maps and
and expected Links status (available
Salman  Network reliability Stochastic optimization
post-disaster or not)
2011 via scenario analysis
performance
Minimizesheltering,
Li et al. Sheltering network Location and amount resources holding, Stochastic optimization
planning
  transportation and via scenario analysis
2011 of demand
distribution costs.
Inventoryrelocation Minimize transportation,
Rottkemper problem with replenishment, inventory
  Demand Stochastic optimization
et al. 2011 overlapping holding, and unsatisfied
disasters demand costs.
Response(storage)

[email protected]
Demand, trans.
facility locations Stochastic optimization
Uncertainty in Humanitarian Logistics for Disaster Management. A Review

Minimize the expected

Noyan 2012  Capacities, and damage
and inventory total cost via scenario analysis
level of supplies
levels
Optimalnumber
Product quality, Stochastic optimization
Duran et al. and location of Minimize the average
  availability, and problem via scenario
2011 pre-positioning response time
production capacity analysis
warehouses
59
 60

Table 3.3 Review of the scientific contributions to Humanitarian Logistics employing Robust Optimization
Pre- Post-
Author Problem type Disaster Disaster Uncertainty types Objective function Methodology
M P R RV
Availabilityof
emergency resources,
Huang et al. Maximal Coverage Maximize cover of critical Risk maps and Robust
 service and travel
2007 model network elements. optimization
times, emergency
frequency.
Dynamicassignment
emergency
Ben-Tal et Demand (Time
response and   MinŰmax Robust optimization
al. 2011 dependent)
evacuation traffic
flow

[email protected]
Table 3.4 Review of the scientific contributions to Humanitarian Logistics employing Simulation Models
Pre- Post-
Author Problem type Disaster Disaster Uncertainty types Objective function Methodology
M P R RV
Capacityrequired to
Bakuli, Evacuation Maximize evacuees, delay
 provide an adequate Simulation
Smith 1996 network and circulation costs
service
Decision Aid Models for Disaster Management and Emergencies
 Table 3.5 Review of the scientific contributions to Humanitarian Logistics employing Fuzzy Sets
Pre- Post-
Author Problem type Disaster Disaster Uncertainty types Objective function Methodology
M P R RV
Internationaldisaster Availability and cost of Multi-objective model:
Adivar and
relief network   emergency resources, transportation and items Fuzzy optimization
Mert 2010 delivery delay time cost, max-min credibility.
optimization
Emergencylogistics
Relief-demand urgency
Sheu et al. operations in
 Relief demand under TOPSIS objective Fuzzy optimization
2010 large-scale natural
function.
disaster
DSSto predict the
effects of a disaster

[email protected]
in terms of
Uncertainty in Humanitarian Logistics for Disaster Management. A Review

Rodríguez
casualties, injures,    Demand Relief demand urgency Fuzzy optimization
et al. 2010
homeless, affected,
and total damage
cost (in US$)
61
62 Decision Aid Models for Disaster Management and Emergencies

Reviews in Humanitarian Logistics may be found in Altay and Green (2006), who fo-
cuses on operation research methodologies and potential research lines, Balcik and Beamon
(2010), who examines the coordination mechanisms and evaluates their adaptability to re-
lief environments, and Caunhye et al. (2011), who proposes a classification of pre-disaster
operations (i.e., evacuation, stock prepositioning and facility location) and a post-disaster
operations (i.e., evacuation, relief distribution, and casualty transportation). More recent
contributions are due to De la Torre et al. (2012) who present a survey of routing problems
applied to Disaster Relief. The authors refer in the survey some articles including uncer-
tainty in demand and supply, as well as in routes and vehicle fleets, always form the point
of view of the distribution aid model. Finally, this book includes other review chapters
referring to papers including some kind of uncertainty, or some chapters devoted to some
specific methodology to model uncertainty.
In the following, an annotated bibliography on Humanitarian Logistic for Disaster Man-
agement under uncertainty is presented. The papers are shown as presented in the previous
tables, by sections. Sometimes a paper appears in more than one table, so it is describes
on the section corresponding to the first table where it has been included. For each work, a
brief description of the main elements and contributions is reported.

3.4.1 Risk Mapping and Probabilistic Values

Belardo et al. (1984), deals with the problem of locating oil spill response equipment
with a partial set covering model, including both assessments of the relative probability of
occurrence and the impact after occurrence of various spill types by means of a multiple
objective approach. The purpose is to attain the best overall protection with existing re-
sources while minimizing the risk of being unprepared for politically and environmentally
sensitive events.
Huang and Fan (2007), compare different modeling approaches for the problem of
allocating multiple emergency service resources to protect critical transportation infras-
tructures. The objective of the paper is to test various risk preferences in decision making
under uncertain service availability and accessibility in case of disaster. A stochastic and
a robust formulation for the maximum coverage problem are proposed. The models are
tested on realistic data, and a sensitivity analysis is conducted to show the robustness of the
solutions found to changes in the models’ parameters. An extension of this work can be
found in Huang and Fan (2011).

[email protected]
Uncertainty in Humanitarian Logistics for Disaster Management. A Review 63

Günneç and Salman (2011), assess the problem of evaluating the reliability and the
expected post-disaster functionality of a transportation network under disaster risk. The au-
thors propose eight probabilistic measures of connectivity and expected travel time/distance
between critical elements of the network. Also, dependencies among the links are consid-
ered. The methodology is based on the generation of scenarios and a specifically tailored
algorithm is proposed. The applicability of the approach is illustrated by means of a case
study of the Istanbul highway system under earthquake threat. Also, the paper provides an
excellent review of the most relevant contributions on network risk assessment for disaster
preparedness.
Nolz et al. (2011), face the problem of designing a distribution system in a post-disaster
context. The authors propose a multi-objective problem considering a risk, a coverage, and
total distribution time objective function. In particular, risk measures represent the proba-
bility that a path might have become impassable as a result of the disaster. Different risk
measures, including correlated measures, are considered to model different disaster scenar-
ios. The rationale behind this work is that, in case of disaster, distribution systems should
be designed while taking into consideration the risk associated to the possible routes (e.g.,
it might be desirable to avoid bridges and tunnels). The problem is solved by an iterative al-
gorithm that generates potentially Pareto-optimal solutions. Finally, the algorithm is tested
on real-world data from the province of Manabí, Ecuador, and the different risk approaches
are analyzed in order to evaluate the most appropriate risk measure for delivering disaster
relief supplies.
Vitoriano et al. (2011), proposes a model for the optimization of last-mile distribution
operations. The formulation is based on a continuous aid flow model entwined with an
integer vehicle flow model. The authors recognize the limits of designing emergency op-
erations according to the classical cost minimizing measures. For this reason, they present
a multi-criteria model that optimizes with respect to attributes that are more relevant to a
disaster context, such as response time, equity in the distribution, reliability and security of
the routes. Specifically, reliability and security are represented by associating to every arc
a transit and ransack probability, respectively. These parameters represent respectively the
probability that the road conditions allow a vehicle to cross the link, and that the vehicle
will be ransacked during the journey. Various probability measures are calculated (e.g.,
global, maximum) and used as decision criteria. A preliminary version of this work can be
found in Ortuño et al. (2011).

[email protected]
64 Decision Aid Models for Disaster Management and Emergencies

3.4.2 Probability Distributions and Stochastic Programming

Fiedrich et al. (2000), minimizes the number of fatalities finding the best assignment of
available resources to operational areas taking into account the possibility of more succes-
sive natural disasters. Fatalities are calculated as result of primary and successive damages,
delays in rescues or lack in medical treatments. These authors propose a model taking into
account the possibility of secondary or even more successive natural disasters. Operational
areas are defined as places with trapped persons; locations endangered by secondary dis-
asters and immediate rehabilitation areas. Hospitals, crossroads and available depots are
considered as assets for refugee. Locations and tasks are linked together in an allocation
model solved by means of a dynamic optimization model. The main objective is the de-
crease in the time limit for search-and-rescue in the first few days conditional on the goal
to minimize the total number of fatalities. Uncertainty is model by means of survival rates
for trapped persons and rescues without medical treatments availability, probabilities of
successive disaster events, transportation times and time completion. Due to the problem
complexity, the authors implemented both Tabu Search and Simulated Annealing heuristic
algorithms and compare them by means of a set of fictitious data. The conclusions derived
from the analysis exhibit that approximately the first 3 days after an earthquake event are
essential for the performance of the relief efforts. Afterwards the probability for rescuing
trapped persons alive decreases radically.
Chan et al. (2001), considers a multiple-depot, multiple-vehicle, location-routing prob-
lem with stochastically processed demands and apply their methodologies to a medical-
evacuation case study of the U.S. Air Force. They solve the distribution center location
problem by means of the heuristic proposed by Clarke and Wright (1964) in a medical
evacuation framework where a plan for allocating aircraft was devised for both the begin-
ning of the conflict and as the conflict progresses. Uncertainty is tackle devoted to demand
by means of a priori or robust optimization, defined as procedures that one performs ahead
of time to anticipate future events. They consider stochastically processed demands as:
the requirement for raw materials to be stored at supply depots when surplus inventory
is strictly disallowed at the processing plants. A 90-day case-study is provided for future
researchers as a benchmark for other solution algorithms.
Barbarosoglu and Arda (2004), following the stochastic multiple-depot, multiple-
vehicle, location-routing problem, develops a two stage stochastic model for transportation
planning in earth quake response. In the first stage, decisions regarding the transportation
of goods from existing supplies for preposition are taken considering the objective of min-

[email protected]
Uncertainty in Humanitarian Logistics for Disaster Management. A Review 65

imizing transportation costs and scenarios expected costs. In the second stage, data related
to demand and supply is disclosed and supplies are transported to demand points minimiz-
ing total flow, shift and penalty for inventory holding and shortage costs. Uncertainty in the
event impact, damage and resource requirements is considered. The problem formulation
contemplates random arc capacities in order to represent the survivability of the routes and
the vulnerability of the supply nodes and availability and usability of commodities. A case
study with data provided by the Marmara earthquake in Turkey (M = 7.4) in 1999.
Chang et al. (2007), aims at developing a decision making tool for government agen-
cies to plan for flood emergencies logistics when the number of rescues in uncertain. They
determine the locations of rescue equipment storehouses and solve the model by using a
sample average approximation (SAA) scheme. The urban areas flooded are considered un-
certain, due to the difficulty of predicting where drainage overflows occur and, therefore,
the estimation of rescue demand is uncertain. The authors propose a preparation planning
for emergency logistics in order to make a prudent distribution of the rescue equipment.
Two stochastic programming models are proposed and solved by means of a heuristic al-
gorithm, which determines the rescue resource distribution plan for urban flood disasters,
including the structure of rescue organization, the location of rescue resource storehouses,
the allocation of rescue resources with respect to the capacity constraints, and their distri-
bution. Disaster areas are classified in groups by emergency level and the objective is to
minimize the expected shipping distance. In a second stage, selected local rescue bases that
need to be set up after the disaster, the quantity of rescue equipment in the storehouses, and
the transportation plans of rescue equipment are determined. The model is applied to Taipei
City, located in Northern Taiwan, with the additional constraint of total rescue operation
length being less than 36 hours.
Yazici and Ozbay (2007), centers in network reliability and treats the determination
of changes in capacity requirements and shelter locations taking into account capacity link
changes during evacuation procedures. They analyze their methodology with a simplified
multiple-origin, multiple-destination version of the Cape May, New Jersey network. They
seek for the desirable shelter locations as a result of link capacity changes during evacu-
ation being evaluated by letting a stochastic model assign flows generating the minimum
system wide travel time. The model formulation is an extension of a cell transmission—
based system optimal dynamic traffic assignment with probabilistic capacity constraints.
Uncertainty is considered in the unexpected change of link capacities due to the impact of
events such as hurricanes and earthquakes. The authors highlight the risk in considering

[email protected]
66 Decision Aid Models for Disaster Management and Emergencies

the predictions of the deterministic model in terms of food, medicine, and other emergency
supplies shortages in shelters.
Zhu et al. (2008), propose a resource allocation model to minimize the total rescue
cost subject to the local reserve depots’ capacity constraints. The aim is to determine the
location of reserve depots and the amount and type of resources to be stored. The type
of uncertainty considered is the disaster magnitude and resource needs, and is modeled by
means of discrete scenarios. Based on the impact of the disaster, scenarios are divided in
two sets: local government and national government responsibility recovery. Three objec-
tive functions are proposed for the problem formulation, based on commodities inventory
holding and transportation costs. The problem is solved by means of an LP relaxation
algorithm and a case study in Shandong Province (China) is considered.
Shen et al. (2009), considers an inventory routing problem where both routes and
delivered quantities are important for efficiency disaster management planning. The au-
thor develops a two stage model for routing vehicles to service in large scale bio-terrorism
emergency with the objective of minimizing unmet demand. The main difference with
other approaches is the short time limit for response. Demand in terms of required emer-
gency supplies and number of casualties exposures and traffic conditions are considered
to be uncertain. The first stage of the algorithm generates pre-planned routes in advance
of any possible emergencies. This problem is formulated as a mixed integer programming
model (MIP) and solved by means of a Tabu Search algorithm. In the second stage, in-
formation related to affected areas is disclosed and pre-planned routes are adjusted to the
given information by means of three different recourse strategies; i.e., LP, Knapsack, and
re-planning. The methodology performance is analyzed with random generated instances.
Song et al. (2009), deals with the problem of generating evacuation transportation
plans. The model optimizes shelter locations and evacuation routes to define bus routing
and passenger pick up points during hurricanes evacuations. A multi-graph street network
is considered for transit evacuation taking into account one way, two ways, prohibited turns
and different delay time for left-turn, right-turn and through running of transit vehicle at
intersections. The locations of the civilians to be evacuated are uncertain. The problem is
solved by means of a combination of genetic, artificial neural networks and hill climbing
heuristic strategies.
Hentenryck et al. (2010), proposes a three-stage stochastic optimization model for
single commodity allocation problem in disaster recovery management. The problem is
focused on potable water deliver for hurricane recovery with the objective of minimizing

[email protected]
Uncertainty in Humanitarian Logistics for Disaster Management. A Review 67

the amount of unsatisfied demands, transportation times from the commodity allocations
to demand points, and storage costs. As MIP approaches often do not scale to real world
instances, the authors propose a hybrid optimization method in three stages: in the first
stage, locations for commodity allocation are determined and solved by means of a MIP.
Then, in the second stage transportation routes from storage locations to demand points are
solved to optimality. Finally, in the third stage a neighborhood search determines the order
in which transportation routes obtained in the second stage must be schedule with the aim
of minimizing last delivery time.
Mete and Zabinsky (2010), follows the line of Belardo et al. (1984) and proposes a
two stage stochastic optimization approach for the storage and distribution problem. The
model is applied to medical supplies for disaster management when demands of goods are
uncertain. They develop a mixed integer programming model to select the storage locations
of medical supplies and required inventory levels for each type of medical supply. The ob-
jective is to minimize the transportation time weighted by the transported load. Uncertainty
is modeled through scenarios that define the location of affected areas, amount of demand,
network routing reliability, and transportation times. In the first stage, decisions related
to warehouse location and goods inventory level are taken to minimize the expected total
operating costs. In the second stage, decisions related to the delivery of goods to demand
points are taken minimizing the transportation duration (transportation times weighted by
the amount) and unsatisfied demand. Penalties incurred for unsatisfied demands are limited
by means of constraints. The methodology is applied to an urban medical supply storage
and distribution problem.
Rawls and Turnquist (2010), similarly to Mete and Zabinsky (2010), proposes a two-
stage stochastic optimization approach for the location and storage of emergency supplies
before a disaster and their distribution to multiple demand points after the catastrophe.
The model considers uncertainty in demand for the stocked supplies as well as uncertainty
regarding transportation network connectivity level after an event. In a first stage, the
response (storage) facility locations and the inventory levels to be pre-positioned are de-
termined minimizing the expected associated cost. In the second stage, the distribution
problem of available supplies to demand locations is solved. Through discrete scenarios
the possibility of loss (partial or completed) in supplies at storage locations, unusable links
connections in the transportation network, and demand uncertainty are represented. The
objective of the stochastic problem is to minimize the expected cost over all scenarios re-
sulting from the selection of the pre-positioning locations, stock and facility sizes costs,

[email protected]
68 Decision Aid Models for Disaster Management and Emergencies

transportation costs and unmet demand. The author proposes an adaptation of the Inte-
ger L-Shaped Method by means of Lagrangian relaxation to solve the resulting stochastic
mixed integer problem. A case study in the Gulf Coast area of the US focused on hurricanes
is considered. The network has 30 nodes and 55 links.
Salmeron and Apte (2010), proposes a two-stage stochastic optimization model to
guide the allocation of budget to acquire and position relief assets with the aim of minimiz-
ing the expected number of casualties. The first stage decisions involve the expansion of
resources (e.g.; warehouses, medical facilities with personnel, ramp spaces, shelters, and
access routes). As the author considers successive disaster events, priority is given to those
resources that allow for future expansions. The second stage (temporally located 3 days
after the event, approx) deals with the transportation of injured population to shelters and
hospitals, and the delivery of supply needs to the stay back population, with the objective
of minimizing the expected unmet transfer population that would result in casualties.
Günneç and Salman (2011), see on the section for Risk Mapping and Probabilistic
Values.
Li et al. (2011), devises a two stage stochastic model to determine shelters location, ca-
pacity, and resource level in a preparedness stage, and evacuees and resources distribution
in a response stage. Shelters that meet the FEMA standards in terms of mass and quality
sheltering are referred as Permanent Shelters; those that do not, are referred as Tempo-
rary Shelters. The objective function proposed gives priority to Permanent Shelters versus
Temporary Shelters and minimizes sheltering costs per evacuee, resources holding costs,
and the expected value of the response stage. The second stage objective function includes
transportation costs of evacuees, resource distribution costs, and surplus and shortage costs
for resources after an evacuation. Due to the binary variables employed in the formulation
for the first stage model, the author highlights the extreme computational complexity of
the model and proposes a classical L-Shaped decomposition method to solve the problem.
A case study of the Gulf Coast region of United Sates is presented for sheltering network
planning and operations against hurricanes.
Rottkemper et al. (2011), consider the inventory location problem in the context of
humanitarian logistics. In the inventory location model, flows between the depots must be
determined so that the total costs (i.e., the sum of transportation costs, replenishment costs,
inventory holding costs, and penalties for unsatisfied demand) is minimized. In their model,
the authors take into consideration overlapping disasters; i.e., subsequent emergencies or
incidents that can occur during the response to or the recovery from a previous disaster. As

[email protected]
Uncertainty in Humanitarian Logistics for Disaster Management. A Review 69

a consequence, the proposed model is multi-period in nature. Also, uncertain demands are
considered. The problem, formulated as a deterministic model where the unmet uncertain
demand is penalized in the objective function, is extensively tested on a Burundi case-study.
Noyan (2012) considers a risk-adverse stochastic model to determine the response
(storage) facility locations and the inventory levels of the relief supplies at each facility
in the presence of uncertainty in the demand and in the damage level of the disaster net-
work. Furthermore, their model allows damage at the links in the network and partial or
total loss of the supplies at storage location when the disaster event occurs. Compared to
Rawls and Turnquist (2010), the novelty of this approach is the incorporation of a condi-
tional value at risk measure. As integer variables are not considered in the second stage
formulations and uncertainty is characterized by a finite set of scenarios, the authors apply
a Benders’ decomposition algorithm with multicuts methodologies for problem solving. A
comparison with the case study proposed by Rawls and Turnquist (2010) is given in order
to draw attention to the solution differences with respect to risk preferences.
Duran et al. (2011), in a more general framework, explores a pre-positioning strategy
to improve the emergency response times in collaboration with CARE and Gatech. The au-
thor proposes a model that finds the optimal number, location, and inventory level of ware-
houses around the world in the face of a disaster or multiple disasters. The model considers
uncertainties regarding product quality, availability, and production capacity in a disaster
affected areas. The proposed mixed integer programming inventory location model con-
siders location costs to represent the possibility of using free or at cost warehouse storage
locations provided by some governments. The objective is the reduction in transportation
times for international aid in disaster management. Their data includes a set of 233 demand
instances and a given budget that determines the maximum number of warehouses and to-
tal inventory to allocate. The objective is to find the configuration of the supply network
that minimizes the weighted average response time; i.e., flying time from the supplier to
arrive at an entry port of the affected area plus one day for set-up and material handling.
Differently from previous contribution, this work considers the occurrence of catastrophes
at different locations which results in an adverse effect regarding to the replenishment lead
time.

3.4.3 Robust Optimization

Huang et al. (2007), see on the Risk Mapping and Probabilistic Values section.

[email protected]
70 Decision Aid Models for Disaster Management and Emergencies

Ben-Tal et al. (2011), follows the line introduced by Song and Zhang (2009) and deals
with the problem of generating dynamic (multi-period) evacuation transportation plans
which are robust to uncertainty in outgoing demand. Uncertainty is given by the difficulty
of estimating the number of civilians at source nodes due to unexpected human behavior
and nature of disaster. The problem is formulated as a linear program based on cell trans-
mission models. The role of hard constraint is emphasized since the penalty cost for an
infeasible solution is the loss of life or property. The authors develop a robust optimization
framework where actual decision variables (flows) are determined for each period t by in-
serting the revealed uncertainties from previous periods in the linear decision rule (“wait
and see” strategy). As the optimal solution may perform much worse in the worst case, the
aim is to provide a good solution for all possible scenarios. They apply the methodology to
two test instances (Chiu et al. (2007), and Yazici and Ozbay (2007)).

3.4.4 Simulation Models

Bakuli and Smith (1996), deals with the evacuation building problem in case there is
uncertainty in the capacity required for an adequate service. The authors found by computa-
tional experience the extremely complexity in using digital simulation for problem solving.
Therefore, the authors propose an analytical model to aid the network design process. The
problem was represented by means of a queuing model and solve with a mean value analy-
sis algorithm. The multi-objective approach of the problem was solved by maximizing the
number of evacuees and determining control variables for the critical capacity sections of
the network.

3.4.5 Fuzzy Sets

Adivar and Mert (2010), optimize a simple relief distribution system in which donor
countries provide relief items which are collected at collection points and, then, shipped
to points of delivery in disaster affected countries. The type of uncertainty considered is
related to imprecise information concerning the quantity of items provided by the donor
countries, the procurement items cost at donor country level, and the potential delay at col-
lection point level. The model, multi-objective in nature, includes the following objective
functions: minimizing the total procurement and transportation cost, and maximizing the
worst credibility (i.e., the highest delivery delay). The authors deal with the uncertainty by
fuzzifying both the uncertain parameters and the credibility. As a consequence, the prob-
lem is formulated as a multi-objective fuzzy mathematical model. In order to be solver,

[email protected]
Uncertainty in Humanitarian Logistics for Disaster Management. A Review 71

the model is reformulated as a parametric linear programming model. Finally, the devised
methodology is tested on real data collected from NATO-EADRCC reports (publicly avail-
able).
Sheu et al. (2010), presents a dynamic relief-demand management model in large-scale
natural disasters. Uncertainty is driven by the number of disaster affected people due to im-
perfect information. In fact, the relief demander is usually not the same as the information
provider. Also, information sources are diverse and hard to be promptly verified. There-
fore, the required information for logistics management in large scale events is desired to
be aggregated in urban areas. The methodology proposed is divided in three steps: data
aggregation in urban areas, area severity classification according to the expected number of
fatalities, and multi-criteria decision making to rank the order of priority. The time frame
considered is the 72 hours after the disaster. During this phase the most critical activities
are search and rescue of entrapped civilians. The case study replicates the massive Chichi
earthquake (7.3 on the Richter scale), which occurred in central Taiwan on 1999 (2455
deaths in total, more than 8000 injured, and the destruction of 38,935 homes). The results
where compared with the real severity level of 13 urban areas through the consideration of
9 possible scenarios. A previous and less sophisticated version of this work can be found
in Sheu (2007).
Rodriguez et al. (2010), design a Decision Support System (DSS) for aiding humani-
tarian organizations based on fuzzy logic. Such DSS, called SEDD, focuses on providing
an estimation of the effects of a disaster right after a disaster strikes; i.e., when there is a
lack of reliable knowledge on the real magnitude of the emergency. Given a disaster-type
and the affected area, SEDD makes use of the data stored in the EM-DAT [1] database to
predict the number of casualties, injured, homeless, affected, and the total damage value
(in US Dollars). The little data, technological, and infrastructure requirements make SEDD
particularly useful and accessible to NGOs.
In a subsequent paper, Rodriguez et al. (2012), the interpretability of the results pro-
vided by SEDD is improved by including three kinds of output; i.e., numerical, interval,
and class predictions.
Next, in Rodriguez et al. (2011), the authors compare the fuzzy DSS with classical
statistical analysis tools, such as multiple linear regression, linear discriminant analysis,
classification trees, and support vector machines. The conclusion of this work is that SEDD
outperforms the methods above in the task of simultaneously providing an accurate and
interpretable inference tool for the evaluation of the consequences of disasters.

[email protected]
72 Decision Aid Models for Disaster Management and Emergencies

3.5 Conclusions

Uncertainty, understood as the imprecision, ambiguity, lack of information or partial

knowledge, is a major topic in the literature, and especially when studying dynamic systems
including uncertain future. But it is especially relevant in Disaster Management, where the
high level of uncertainty inherit to most of the parameters involved in humanitarian logistics
makes a difference respect to other kind of logistics.
Different approaches to deal with uncertainty in humanitarian logistics can be found in
the literature. They differ in the way to model uncertainty, the way to manage it, the param-
eters considered to be under uncertainty. Decision aid models based on these approaches
have been developed with different aims and goals, but there is still a long way to go to put
these models into production, especially because many times they require deep analyses of
data, strong computation efforts and/or the required training of final users.
Our survey shows that the majority of the contributions in decision models for Human-
itarian Logistics make use of Stochastic Programming. Nevertheless, these studies usually
are not implemented in systems or tools applied by practitioners. In the real world, much
more is being applied of risk mapping for preparedness (emergency protocols developed
usually by national or local civil protection entities) or to image processing to obtain maps
of the affected area of a disaster (captured by satellites or planes). This information is the
base for making decisions on the humanitarian logistics related to disaster management but
usually is not integrated in decision making models. So, larger efforts should be done to
deal into the models with this kind of uncertainty, developing systems able to manage it
and training final users to understand it and capitalize on its potential.

Bibliography

[1] Adivar B., Mert A. (2010). International disaster relief planning with fuzzy credibility. Fuzzy
Optimization and Decision Making, Vol. 9, pp. 413–433.
[2] Altay N., Green W. G. (2006). OR/MS research in disaster operations management. European
Journal of Operational Research, Vol. 175, pp. 475–493.
[3] Bakuli D. L., MacGregor Smith J. (1996). Resource allocation in state-dependent emergency
evacuation networks. European Journal of Operational Research, Vol. 89, pp. 543–555.
[4] Balcik B., Beamon B. M. (2008). Facility location in humanitarian relief. International Journal
of Logistics: Research and Applications, Vol. 11 (2), pp. 101–121.
[5] Barbarosoglu G., Arda Y. (2004). A two-stage stochastic programming framework for trans-
portation planning in disaster response. Journal of the Operational Research Society, Vol. 55,
pp. 43–53.

[email protected]
Bibliography 73

[6] Belardo S., Harrald J., Wallace W.A., Ward J. (1984). A partial covering approach to sitting
response resources for major maritime oil spills. Management Science, Vol. 30 (10), pp. 1184–
1196.
[7] Ben-Tal A., Chung B.D., Mandala S.R., Yao T. (2011). Robust optimization for emergency lo-
gistics planning: Risk mitigation in humanitarian relief supply chains. Transportation Research
Part B, Vol. 45, pp. 1177–1189.
[8] Caunhye A.M., Niea X., Pokharelb S. (2011). Optimization models in emergency logistics: A
literature review. Socio-Economic Planning Sciences, Vol. 46 (1), pp. 4–13.
[9] Chan Y., Carter W. B., Burnes M. D. (2001). A multiple-depot, multiple-vehicle, location-
routing problem with stochastically processed demands. Computers & Operations Research,
Vol. 28, pp. 803–826.
[10] Chang MS., Tseng YL., Chen JW. (2007). A scenario planning approach for the flood emer-
gency logistics preparation problem under uncertainty. Transportation Research Part E: Logis-
tics and Transportation Review, Vol. 43 (6), pp. 737–754.
[11] Clarke G., Wright J.W. (1964). Scheduling of Vehicles from a Central Depot to a Number of
Delivery Points. Operations Research, Vol. 12, pp. 568–581.
[12] De la Torre L.E., Dolinskaya I.S., Smilowitz K.R. (2012). Disaster Relief Routing: Integrating
Research and Practice. Socio-Economic Planning Sciences, Vol. 46 (1), pp. 88–97.
[13] Deb, K., Gupta, H. (2006) Introducing robustness in multi-objective optimization, Evolutionary
Computation 14 (4) pp 463–494.
[14] Dilley, M., Chen, R. S., Deichmann, U., Lerner-Lam, A.L., Arnold, M., Agwe, J., Buys, P.,
Kjekstad, O., Lyon, B. and Yetman, G. 2005. Natural disaster hotspots: a global risk analysis.
Synthesis report.
[15] Duran S, Gutierrez MA, Keskinocak PN. (2011). Pre-positioning of emergency items world-
wide for CARE international. Interfaces, Vol. 41 (3), pp. 223–237.
[16] Fiedrich F., Gehbauer F., Rickers U. (2000). Optimized resource allocation for emergency re-
sponse after earthquake disasters. Safety Science, Vol. 35 (1-3), pp. 41–57.
[17] Günneç D., Salman F.S. (2011). Assessing the reliability and the expected performance of a
network under disaster risk, OR Spectrum, Vol. 33, pp. 499–523.
[18] Hentenryck P.V., Bent R., Coffrin C. (2010). Strategic Planning for Disaster Recovery with
Stochastic Last Mile Distribution. Proceedings of the Seventh International Conference on In-
tegration of Artificial Intelligence and Operations Research Techniques in Constraint Program-
ming (CPAIOR 2010), pp. 318–333.
[19] Huang Y., Fan Y., Cheu R.L. (2007). Optimal Allocation of Multiple Emergency Service Re-
sources for Protection of Critical Transportation Infrastructure. Transportation Research Board,
Vol. 2022, pp. 1–8.
[20] Huang Y., Fan Y. (2011). Modeling Uncertainties in Emergency Service Resource Allocation.
Journal of Infrastructure Systems, Vol. 17 (1), pp. 35–41.
[21] Li L., Jin M., Zhang L. (2011). Sheltering network planning and management with a case in
the Gulf Coast region. International Journal of Production Economics, Vol.131, pp.431–440.
[22] Mete HO, Zabinsky ZB. (2010). Stochastic optimization of medical supply location and distri-
bution in disaster management. International Journal of Production Economics, Vol. 126 (1),
pp. 76–84.
[23] Moss, R.H. and Schneider, S.H. (2000) Uncertainties in the IPCC TAR: Recommendations to
lead authors for more consistent assessment and reporting. In: Guidance Papers on the Cross
Cutting Issues of the Third Assessment Report of the IPCC [eds. R. Pachauri, T. Taniguchi and
K. Tanaka], World Meteorological Organization, Geneva, pp. 33–51.
[24] Nolz P.C., Semet F., Doerner K.F. (2011). Risk approaches for delivering disaster relief sup-
plies, OR Spectrum, Vol. 33, pp. 543–569.

[email protected]
74 Decision Aid Models for Disaster Management and Emergencies

[25] Noyan N. (2012). Risk-averse two-stage stochastic programming with an application to disaster
management. Computers & Operations Research, Vol. 39, pp. 541–559.
[26] Ortuño, M.T., Tirado, G., Vitoriano, B.: A lexicographical goal programming based decision
support system for logistics of Humanitarian Aid. Top, Vol. 19 (2), pp. 646–679.
[27] Pedraza-Martinez, A.J., Stapleton, O.,VanWassenhove, L.N. (2011) Field vehicle fleet manage-
ment in humanitarian operations: A case-based approach. Journal of Operations Management,
Vol. 29 (5), pp. 404–421.
[28] Rawls CG, Turnquist MA. (2010). Pre-positioning of emergency supplies for disaster response.
Transportation Research Part B: Methodological, Vol. 44 (4), pp. 521–34.
[29] Rodriguez J.T., Vitoriano B., Montero J. (2010). A natural-disaster management DSS for Hu-
manitarian Non-Governmental Organisations, Knowledge-Based Systems, Vol. 23, pp. 17–22.
[30] Rodriguez J.T., Vitoriano B., Montero J., Vojislav K. (2011). A disaster-severity assessment
DSS comparative analysis, OR Spectrum, Vol. 33, pp. 451–479.
[31] Rodriguez J.T., Vitoriano B., Montero J. (2012). A general methodology for data-based rule
building and its application to natural disaster management, Computers & Operations Research,
Vol. 39, pp. 863–873.
[32] Rottkemper B., Fischer K., Blecken A., Danne C. (2011). Inventory relocation for overlapping
disaster settings in humanitarian operations, OR Spectrum, Vol. 33, pp. 721–749.
[33] Salmerón J., Apte A. (2010). Stochastic Optimization for Natural Disaster Asset Prepositioning.
Production and operations management, Vol. 19 (5), pp. 561–574.
[34] Shen Z., Dessouky M.M., Ordoñez F. (2009). A two-stage vehicle routing model for large-scale
bioterrorism emergencies. Networks, Vol. 54 (4), pp. 255–269.
[35] Sheu J.B. (2007). An emergency logistics distribution approach for quick response to urgent
relief demand in disasters. Transportation Research Part E, Vol. 43 (6), pp. 687–709.
[36] Sheu J.B. (2010). Dynamic relief-demand management for emergency logistics operations un-
der large-scale disasters. Transportation Research Part E, Vol. 46, pp. 1–17.
[37] Song R., He S., Zhang L. (2009). Optimum transit operations during the emergency
evacuations. Journal of Transportation Systems Engineering and Information Technology,
Vol. 9 (6), pp. 154–160. United Nations International Strategy for Disaster Reduction
(UNISDR), (2009). Terminology on Disaster Risk Reduction. http://unisdr.org/files/
7817_UNISDRTerminologyEnglish.pdf
[38] Vitoriano B., Ortuño M.T., Tirado G., Montero J. (2011). A multi-criteria optimization model
for humanitarian aid distribution, Journal of Global Optimization, Vol. 51, pp. 189–208.
[39] Yazici M.A., Ozbay K. (2007). Impact of probabilistic road capacity constraints on the spatial
distribution of hurricane evacuation shelter capacities. Transportation Research Record: Journal
of the Transportation Research Board 2022, pp. 55–62.
[40] Zadeh, L.A. (1965). Fuzzy sets. Information and Control 8 (3), pp. 338–353.
[41] Zhu J., Huang J., Liu D., Han J. (2008). Resources Allocation Problem for Local Reserve
Depots in Disaster Management Based on Scenario Analysis. The 7th International Symposium
on Operations Research and Its Applications (ISORA’08). ORSC & APORC, pp. 395–407.

[email protected]
Chapter 4

Fuzzy inference systems for disaster response

Başar Öztayşi 1 , Hülya Behret 2 , Özgür Kabak 1 , İrem Uçal Sarı 1 , and Cengiz Kahraman 1
1 Istanbul Technical University, Department of Industrial Engineering, 34367 Macka
Istanbul Turkey
2 Erciyes University, Department of Industrial Engineering, 38039, Kayseri, Turkey

Disaster management is extremely important in today’s world, which is defined as the or-
ganization and management of resources and responsibilities for dealing with all human-
itarian aspects of emergencies, in particular preparedness, response and recovery in order
to lessen the impact of disasters. Disaster response is one of the critical stages of disas-
ter management, which necessitates spontaneous decision making when a disaster occurs.
Fuzzy inference systems are very suitable for such decision making environments since
the inputs and outputs of disaster events cannot be sharply defined. This chapter describes
potential applications of fuzzy inference systems in disaster response.

4.1 Introduction

The term disaster is derived from the Latin roots dis- and astro, meaning “away from
the stars” or, in other words, an event to be blamed on an unfortunate astrological config-
uration (Coppola, 2006). A disaster is defined as an overwhelming ecological disruption
occurring on a scale sufficient to require outside assistance (Syed, 2009). An event, con-
centrated in time and space, in which a society, or a relatively self-sufficient subdivision
of a society, undergoes severe danger and incurs such losses to its members and physical
appurtenances that the social structure is disrupted and the fulfillment of all or some of the
essential functions of the society is prevented. (Fritz, 1961). When a hazard has effect on
vulnerable people it is called a disaster. Disaster can be considered as the combination of
hazards, vulnerability and inability to reduce the potential negative consequences of risk
(URL-1).

B. Vitoriano et al. (eds.), Decision Aid Models for Disaster Management 75

and Emergencies, Atlantis Computational Intelligence Systems 7,
DOI: 10.2991/978-94-91216-74-9_4,  Atlantis Press 2013

[email protected]
76 Decision Aid Models for Disaster Management and Emergencies

According to International Federation of Red Cross and Red Crescent Societies (IFRC),
though often caused by nature, disasters can have human origins (URL-1). Natural haz-
ards are naturally-occurring physical phenomena caused either by rapid or slow onset
events. IFRC classifies the natural hazards into five main categories: geophysical (earth-
quakes, landslides, tsunamis and volcanic activity), hydrological (avalanches and floods),
climatological (extreme temperatures, drought and wildfires), meteorological (cyclones and
storms/wave surges), and biological (disease epidemics and insect/animal plagues) (URL-
1). Technological or man-made hazards, on the other hand, are human caused events that
occur in or close to human settlements. Examples are environmental degradation, pollution
and accidents (URL-1).
IFRC defines disaster management as follows:
“The organization and management of resources and responsibilities for dealing with
all humanitarian aspects of emergencies, in particular preparedness, response and recovery
in order to lessen the impact of disasters” (URL-1).
As shown in Fig. 4.1, comprehensive disaster management is based upon four distinct
components: mitigation, preparedness, response, and recovery (Coppola, 2006). The four-
phase classification is based on the Comprehensive Emergency Management concept intro-
duced in the 1978 report of the National Governors Association Emergency Preparedness
Project. Mitigation Involves reducing or eliminating the likelihood or impact on the society
of a hazard, or both. Preparedness involves equipping people who may be impacted by a
disaster or who may be able to help those impacted with the tools to increase their chance of
survival and to minimize their financial and other losses. Response involves taking action
to reduce or eliminate the impact of disasters that have occurred or are currently occur-
ring, in order to prevent further suffering, financial loss, or a combination of both. Finally
recovery involves returning victims’ lives back to a normal state following the impact of
disaster consequences. The recovery phase generally begins after the immediate response
has ended, and can persist for months or years thereafter.
Despite all activities accomplished in mitigation and preparedness phase, disasters oc-
cur and create damage. Disaster response is defined as the total sum of actions taken by
people and institutions in the face of disaster (Syed, 2009). Disaster response provides for
the immediate protection of life and property, reestablishing control and minimizing the
effects of a disaster (Rao, 2007). Response processes begin as soon as it becomes appar-
ent that a hazard event is imminent and lasts until the emergency is declared to be over
(Coppola, 2006). Disaster response includes the issuance of predictions and warnings, and

[email protected]
Fuzzy inference systems for disaster response 77

Fig. 4.1 Disaster Management Cycle (Alexander, 2002)

preparations immediately before an event occurs. Besides, disaster response contain other
forms of protective actions include actions aimed at limiting injuries, loss of life, and dam-
age to property and the environment that are taken prior to, during, and immediately after
a hazard event. Response includes not only those activities that directly address these im-
mediate needs—such as first aid, search and rescue, and shelter—but also includes systems
developed to coordinate and support such efforts.
In order to conduct an effective disaster management through decreasing the impact
of hazard and tackling vulnerability, it is required to deal with uncertainties. Especially
in the pre-event phase, risk management should deal with the uncertainties resulting from
unknown time and unknown consequence of disasters to prevent the onset of a disaster,
reduce the impacts, or prepare the community to respond when a disaster occurs. In the
post-event phase, complexity and uncertainty of the social, economic and political conse-
quences of disasters should be analyzed effectively for the employment of resources and
emergency procedures. Therefore uncertainty is one of the most challenging and important
problems in the disaster management.

[email protected]
78 Decision Aid Models for Disaster Management and Emergencies

In Altay and Green (2006) more than 50% of the analyzed papers take into account
uncertainty by means of operations research methodologies such as probability, simulation,
queueing theory, fuzzy sets and stochastic programming. Fuzzy sets are one of the least
preferred approaches in the literature for disaster management.
Fuzzy set theory, which was founded by Zadeh (1965), has emerged as a powerful
way of quantitatively representing and manipulating the imprecision in decision-making
problems. Fuzzy sets or fuzzy numbers can appropriately represent imprecise parameters,
and can be manipulated through different operations on fuzzy sets or fuzzy numbers. Since
imprecise parameters are treated as imprecise values instead of precise ones, the process
will be more powerful and its results more credible. Fuzzy set theory has been studied
extensively over the past 40 years (Kahraman et al., 2006).
Fuzzy set theory is being recognized as an important problem modeling and solution
technique. The use of fuzzy set theory as a methodology for modeling and analyzing
decision systems is of particular interest to researchers due to fuzzy set theory’s ability to
quantitatively and qualitatively model problems which involve vagueness and imprecision
(Kahraman et al., 2006).
Fuzzy set theory approaches to disaster management are scarce. Among the relevant
literature some noteworthy studies are as follows: Shen (2007) presents a hybrid fuzzy
clustering-optimization approach to the operation of emergency logistics co-distribution
responding to the urgent relief demands in the crucial rescue period. Tsai et al. (2008)
apply fuzzy set theory for decision making in a geographic information system to the al-
location of disaster shelters. Simonovic and Nirupama (2005) develop a new technique to
enhance the ability to address different uncertainties in spatial water resources decision-
making that provided a measurable improvement in the management of floods. Li et al.
(2005) use fuzzy comprehensive evaluation of risk in a decision support tool for typhoon
insurance pricing to support fuzziness and uncertainty. More literature survey is provided
in Section 4.3.
The aim of the chapter is to describe the potential application of fuzzy inference systems
(FISs) in disaster response systems. In this manner the chapter is organized as follows. In
Section 4.2 the disaster recovery systems are introduced in details. Section 4.3 describes
the FISs and provides a brief literature review about previous studies. Section 4.4 contains a
sample FIS for spontaneous volunteer management in disaster response systems and finally
further steps are discussed in conclusions.

[email protected]
Fuzzy inference systems for disaster response 79

4.2 Disaster Response System

As described in the previous section, disaster response activities are the actions taken
in a short term just after or just before a disaster occurs. According to Altay and Green
(2006) the typical activities involved in response phase are listed as follows:

(1) Activating the emergency operations plan

(2) Activating the emergency operations center
(3) Evacuation of threatened populations
(4) Opening of shelters and provision of mass care
(5) Emergency rescue and medical care
(6) Fire fighting
(7) Urban search and rescue
(8) Emergency infrastructure protection and recovery of lifeline services
(9) Fatality management

The aim of above-given activities is the employment of resources and emergency pro-
cedures as guided by plans to preserve life, property, the environment, and the social, eco-
nomic, and political structure of the community (Altay and Green, 2006).
The disaster response system functions may be classified into two groups as pre-disaster
and post disaster activities. Pre-disaster activities may be efficient if advanced warning
systems exist and appraise the upcoming disaster (Coppola, 2006).

(1) Warnings: refers to arrangements to rapidly disseminate information concerning im-

minent disaster threats to government officials and the population (Syed, 2009).
(2) Evacuation: involves the relocation of a population from zones at risk of imminent
disaster to a safer location.
(3) Logistics of resources and supplies: with the help of advanced warning of the disaster,
officials can find a chance to transport supplies into the affected site before hazard
conditions and consequences make such movement more difficult (Coppola, 2006).
(4) Last-minute mitigation and preparedness: Mitigation and preparedness are most effec-
tive when they are performed far in advance of a disaster. However, actions often may
be taken in the few hours or days before a disaster occurs.

As disaster response begins, the first priority is saving lives. The emergency activities
include search and rescue, first aid, and evacuation that are directly related to saving lives.

[email protected]
80 Decision Aid Models for Disaster Management and Emergencies

Besides, depending on the type of the disaster, other supporting activities may be needed.
The post-disaster activities are as follows:

(1) Search and Rescue: is the process of identifying the location of disaster victims that
may be trapped or isolated and bringing them to safety and medical attention.
(2) First Aid Medical Treatment: In a disaster the quantity of victims may be so great
that they completely overwhelm the capacity of local clinics or hospitals to care for
them all. In such situations, injured victims should be located and a first aid medical
treatment must be applied to stabilize their condition, and transport them to a facility
where they can receive the medical assistance necessary to save their lives (Coppola,
2006).
(3) Evacuation: Before, during, or after a disaster occurs, moving populations away from
the hazard and its consequences reduce the effect of many disasters.
(4) Disaster Assessment: As soon as possible after the disaster has begun, response offi-
cials must begin collecting data, which is then formulated into information to facilitate
the response. It is vital to know at any given time or at short intervals what is happen-
ing, where it is happening, what is required to address those needs, and what resources
are available.
(5) Logistics and Supply: The delivery of emergency relief will require logistical facilities
and capacity. A well-organized supply service is crucial for handling, storage and
dispatch of relief supplies.
(6) Safety and Security: In the response period of a disaster, the social order of the affected
area is disrupted. Local officials focus mostly on managing the hazard’s consequences.
However, safety and security needs of the survivors are superior to normal and must be
fulfilled.
(7) Health and Sanitation: In times of disaster, risk of injuries and illnesses is higher than
normal situations. However, health facilities may be full or damaged so facilities for
emergency healthcare operations must be prepared to accommodate the health needs
of the affected population. One of the most important issues is achieving proper sani-
tation.
(8) Donation Management: In disasters, individuals governments, private and religious
groups, and businesses are likely to donate disaster victims However, without an effec-
tive mechanism to accept, catalogue, store, and distribute those donations it is impos-
sible to transmit these donation to the victims.

[email protected]
Fuzzy inference systems for disaster response 81

Emergency operations management is a vital component of disaster response. Emer-

gency operations management of disaster response and recovery activities is best facilitated
from a centralized location that is normally established away from the disaster scene. De-
cision making is highly important for disaster response because the mentioned activities
must be accomplished with a limited resource and agencies that quickly descend upon the
impacted areas. For this reason disaster decision support systems have been the subject of
many studies in the literature (Lu and Yang, 2011; Arora et al., 2010; Rodríguez et al.,
2010; Yoon et al., 2008; Zografos et al., 2002; Wallace and Balogh 1985). Inference can
be defined as the act or process of deriving logical conclusions from premises known or
assumed to be true. In the disaster management area the inference system can be used as
a disaster response expert system. According to inference rules and special knowledge,
the inference system can analyze a situation and give feasible response to decision makers
(Jianshe, 1994).
In disaster response system, the decision makers deal with highly divergent conditions
and demands. Unique to each disaster, the participants, victims’ and community needs
differ and as a consequence the timing and order of events, actions and processes diverge.
Because of this dissimilarity in demands and actions, it is not easy to construct a single
model to describe the relationship of demand to capacity (Confort et al., 2004). Since
the disaster situation usually cannot be described precisely, and the information is often in-
complete decision makers do not have enough information to perform decision analysis. To
deal quantitatively with such an imprecision or uncertainty, the fuzzy set theory is adapted
to the studies about disaster management (Jianshe, 1994; Cret et al., 1993; Shiri and Kisi,
2010).

4.3 Fuzzy Inference Systems

Fuzzy inference systems (FISs) are rule based systems with concepts and operations
associated with fuzzy set theory and fuzzy logic (Ross, 2004; Mendel, 2001). FISs start
from highly formalized insights about the structure of categories found in real life and then
articulate fuzzy “IF-THEN” rules as a kind of expert knowledge. Fuzzy systems combine
fuzzy sets with fuzzy rules to produce overall complex nonlinear behaviors (Sen, 2010).
Fuzzy inference is the process of formulating the mapping from a given input to an
output using fuzzy logic. By this way, fuzzy inference allows constructing structures that
can be used to produce responses (outputs) to certain stimulations (inputs), based on stored

[email protected]
82 Decision Aid Models for Disaster Management and Emergencies

knowledge on how the responses and stimulations are related. In FISs, the knowledge is
stored in the form of a rule base. In some cases, this knowledge is obtained by extract-
ing information thoroughly from experts, which is usually known as fuzzy expert systems
(Hall and Kandel, 1991). In other cases, data driven approaches are used in which mem-
bership functions and rules are developed using a training data set where the parameters
for the membership functions and rules are consequently optimized to reduce training error
(Keshwani et al., 2008).
There are two systems for fuzzy inference, which are usually denoted as: Takagi–
Sugeno-Kang type (or just called Sugeno type) FIS and Mamdani type FIS. Sugeno type
FIS was first proposed by Takagi and Sugeno in 1985 and by Sugeno and Kang in 1988.
A sugeno type FIS has fuzzy inputs and a crisp output (linear combination of the inputs).
Sugeno type FIS is computationally effective and works well with optimization and adap-
tive techniques (such as ANFIS), which makes it very attractive in control problems, par-
ticularly for dynamic nonlinear systems. Mamdani type FIS is first proposed by Mamdani
in 1974 and Mamdani and Assilian in 1975. A Mandani FIS has fuzzy inputs and a fuzzy
output. The Mamdani type FIS is manually constructed on the basis of expert knowledge
and the final model is neither trained nor optimized. Mamdani type FIS uses a set of lin-
guistic rules obtained from experienced human operators (Mamdani and Assilian, 1975).
The main feature of such type of FIS is that both the antecedents and the consequents of the
rules are expressed as linguistic constraints (Zadeh, 1997). Mamdani type FIS can provide
a highly intuitive knowledge base that is easy to understand and maintain. Therefore, this
type of FIS is more widely used, mainly in decision making applications. Both models
are similar and consider fuzzy inputs but Mamdani returns fuzzy outputs while Takagi–
Sugeno returns crisp outputs (Mamdani, 1974; Takagi and Sugeno, 1985). Since Mamdani
approach is not exclusively dependent on a data set, with sufficient expertise on the system
involved, a generalized model for effective future predictions can be obtained (Keshwani et
al., 2008). As completely linguistic form of rules in Mamdani models has advantages in the
representation of expert knowledge and in the linguistic interpretation of dependencies. It
allows us to describe the expertise in more intuitive, more human-like manner. This study
would apply the Mamdani type inference to predict the parameters.
Fuzzy inference systems have been successfully applied in fields such as automatic
control, data classification, decision analysis, expert systems, and computer vision. Be-
cause of its multidisciplinary nature, FISs are associated with a number of names, such as

[email protected]
Fuzzy inference systems for disaster response 83

fuzzy-rule-based systems, fuzzy expert systems, fuzzy modeling, fuzzy associative mem-
ory, fuzzy logic controllers, and simply fuzzy systems.
Fuzzy inference process comprises of five parts: fuzzification of the input variables,
application of the fuzzy operator (AND or OR) in the antecedent, implication from the
antecedent to the consequent, aggregation of the consequents across the rules, and defuzzi-
fication.
A typical FIS can be described in four steps which are; fuzzification, fuzzy rules, fuzzy
inference and defuzzification (Figure 4.2).
Step 1: (Fuzzification) Fuzzification process involves the definition of the membership
functions of input/output variables by linguistic variables. In this study, triangular mem-
bership functions are used for the representation of linguistic expressions of input/output
variables because of simplicity.
Step 2: (Fuzzy rules) A FIS with i-input variables has r = pi rules, where p is the
number of linguistic terms per input variable. As the dimension and complexity of a system
increase, the size of the rule base increases exponentially.
The rules for the present study are structured as:

IF I1 is A1j AND I2 is A2j AND . . . In is Anj THEN y is B j for j = 1, 2, . . . , r (4.1)

where Ii (i = 1, 2, . . . , n) are input variables and y is the output variable, A1j , A2j , . . . , Anj and
B j are the linguistic terms used for the membership function of the corresponding input
and output variables for the jth rule, respectively.
Step 3: (Fuzzy inference) Fuzzy inference is an inference procedure to derive a con-
clusion based on a set of if–then rules. In this study, Mamdani method is applied for fuzzy
inference. The Mamdani inference method is manually constructed on the basis of expert
knowledge and the final model is neither trained nor optimized. The method considers
fuzzy inputs and returns fuzzy outputs (Mamdani, 1974). Since Mamdani approach is not
exclusively dependent on a data set, with sufficient expertise on the system involved, a gen-
eralized model for effective future predictions can be obtained (Keshwani et al., 2008). The
mechanism of Mamdani inference method is as follows: (1) If there is more than one input
in the rule, fuzzy set operations should be applied to achieve a single membership value;
(2) then implication method (min) is applied to reach each rule’s conclusion; (3) the out-
puts obtained for each rule are combined into a single fuzzy set, using a fuzzy aggregation
operator (max).
For the case where input variables Ii (i = 1, 2, . . . , n) are crisp variables and the fuzzy
rules are described by Equation (4.1), so for a set of disjunctive rules, where j = 1, 2, . . . , r,

[email protected]
84 Decision Aid Models for Disaster Management and Emergencies

the output using Mamdani inference method is formulated as follows (Ross, 2004);

μ j (y) = max min μ j (I1 ), μ j (I2 ), . . . , μ j (In ) (4.2)

B j A1 A2 An

Step 4: (Defuzzification) Through fuzzy inference, the result can be converted into a
definite value. In this study the center of area defuzzification method is used in order to
get the advantage of its continuity and disambiguity. Center of area (also called centroid
method or center of gravity method) is the most prevalent and physically appealing of all
the available defuzzification methods (Ross, 2004). It is given by the algebraic expression
as follows:

μC · c · dc
c∗ = , c∈C (4.3)
μC dc

where C is a fuzzy set having the membership function μC .

The graphical illustration of the used FIS is represented in Figure 4.2.

Fig. 4.2 Graphical Mamdani (max-min) inference method

[email protected]
Fuzzy inference systems for disaster response 85

4.4 Literature Review

Disasters are unpredictable and have highly destructive effects that cause a great un-
certainty in disaster management. Fuzzy set theory is used to deal with the uncertainty
in disaster management areas such as forecasting earthquake damages, flood damages and
landslide susceptibility analysis, safety monitoring of hillsides, developing fire protection
policies. . . etc.
Cret et al. (1993) use the fuzzy set theory and fuzzy decision analysis to estimate earth-
quake damage. They proposed a system to formalize knowledge acquired from experience
and assess earthquake damage from ground conditions and ground motion characteristics.
Fuzzy damage indices are generated from the results of the proposed system. Fuzzy de-
cision analysis is used to transform this imprecise information into distinct decision on
whether to cut or maintain the gas supply in a given area of the network.
Huang and Leung (1999) propose a hybrid fuzzy neural network to determine the rela-
tionship between isoseismic area and earthquake magnitude.
Şen (2011) proposes a method to classify the buildings in terms of their weaknesses for
a possible earthquake. The most important building factors that would affect the impacts
of earthquake on the buildings and their relationships with the five hazard categories are
determined by fuzzy numbers. The relationships are presented through a supervised hazard
center classification method.
Chang and Chang (2006) use the adaptive network-based FIS to build a prediction
model for reservoir management. A reservoir in Taiwan which came across a large number
of typhoon and heavy rainfall events in past years was used as a case study to illustrate the
applicability and capability of the adaptive network-based FIS.
Shiri and Kisi (2010) investigate the application of hybrid wavelet-neuro-fuzzy model
to model daily, monthly and yearly stream flows which are important as it determines the
reservoir inflow as well as the flooding events.
Akter and Simonovic (2005) propose a methodology to capture the views of multiple
stakeholders which are associated with flood management decision making problems by
using fuzzy set theory and fuzzy logic. By using fuzzy expected value three different
possible forms of individual stakeholders’ input are analyzed to obtain the aggregated input.
Simonovic and Akter (2006) use fuzzy expected value input with the multi-criteria
decision-making tool named Fuzzy Compromise Programming. The methodology has been
applied to floodplain management in the Red River Basin, Canada that faces periodical
flooding.

[email protected]
86 Decision Aid Models for Disaster Management and Emergencies

Chang et al. (2008) propose the counter propagation fuzzy-neural network for extract-
ing flood control knowledge in the form of fuzzy if–then rules to simulate a human-like
operating strategy in a city flood control system through storm events. The Yu-Cheng
pumping station, Taipei City, is used as a case study.
Topuz et al. (2011) propose an approach that integrates environmental and human
health risk assessment for industries using hazardous materials in order to support environ-
mental decision makers with quantitative and directive results. Analytic hierarchy process
and fuzzy logic are used as tools to handle problems caused by complexity of environment
and uncertain data.
Cheng and Ko (2002) describe the development of a decision support system for safety
monitoring of hillsides by using fuzzy set theory.
Saboya Jr. et al. (2006) use fuzzy logic to evaluate the susceptibility of occurrence
of landslides in the areas where landslides are susceptible to occur. The methodology is
applied to identify the susceptibility of landslides in a chaotic occupied urban area of Rio
de Janeiro, Brazil, where some occurrences have been reported.
Dokas et al. (2009) integrate fuzzy expert systems, fault tree analysis and worldwide
web technologies and applied in the development of the landfill operation management
advisor which is a novel early warning and emergency response system for solid waste
landfill operations.
Sezer et al. (2011) establish the results of an adaptive neuro-FIS model using remote
sensing data and geographical information system for landslide susceptibility analysis in a
part of the Klang Valley areas in Malaysia.
Oh and Pradhan (2011) present landslide susceptibility mapping using an adeptive
neuro FIS using a geographic information system environment.
Chang and Chien (2006) use a fuzzy–neural hybrid system to simulate typhoon waves.
A membership function based on the fuzzy theory is expressed by a union Gaussian func-
tion to illustrate the rapid wave decaying.
Iliadis and Spartalis (2005) propose a two level method to develop a rational and sensi-
ble protection policy from natural disasters. On the first level the annual forest fire risk for
each area of Greece is estimated by using fuzzy trapezoidal membership function. On the
second level a narrow expected closed interval for the burned area are forecasted by using
a fuzzy expected interval model.

[email protected]
Fuzzy inference systems for disaster response 87

Iliadis (2005) proposes a decision support system which applies an inference mech-
anism based on various aspects of fuzzy sets and fuzzy machine learning techniques to
develop a rational and sensible forest fire prevention and protection policy.
Liu et al. (2010) present a geographical information system and information diffusion
based methodology for risk analysis of grassland fire disaster to livestock production in the
grassland area of the northern China.
Huang and Inoue (2007) discuss three soft risk maps of a flood, a drought and an earth-
quake which are aimed at the visualization of risk levels of natural disasters defined by
fuzzy probabilities.
In the above literature review, it is found that the papers using fuzzy set theory in dis-
aster management focus on estimating the magnitude of the disasters such as fires, earth-
quakes and floods. Therefore we can conclude that, to the best of our knowledge, there is
limited number of fuzzy set theory based studies on disaster management, particularly on
disaster response. Hence we aim at modeling a FIS for spontaneous volunteer management,
which is one of the important activities in disaster response.

4.5 A Fuzzy Inference System for Spontaneous Volunteer Management

Spontaneous volunteer management is one of the important parts of disaster response

process. Potential spontaneous volunteers are individuals or groups of people who seek or
are invited to contribute their assistance during and/or after an event and who are unaffil-
iated with any part of the existing official emergency management response and recovery
system and may or may not have relevant training, skills or experience (URL-2). The spon-
taneous, unaffiliated volunteers, who are often our neighbors, friends, and ordinary citizens,
frequently arrive at a disaster scene to give help. But because they are not connected with
any part of the existing emergency management response system, group, or organization,
their offers of help are regularly underutilized and can even be challenging to most profes-
sional responders (Gallant, 2008). Uncoordinated or uncontrolled spontaneous volunteers
can interfere with disaster relief operations and cause a secondary disaster. To avoid such a
situation volunteer management centers should be established away from the disaster scene
and classification of volunteers should be managed in these centers in order to efficiently
utilize volunteers.
At this stage, the classification decision is highly important for disaster response be-
cause of the limitation on the amount of water, food and waste disposal in the disaster area.

[email protected]
88 Decision Aid Models for Disaster Management and Emergencies

Table 4.1 Input variables of FIS

Input Fuzzy variables Fuzzy numbers
young (18,18,45)
Age (A) middle age (35,45,55)
old (45,65,65)
Compatibility of the physical characteristics (PC)
Leading and management skills (LM) low (0,0,4)
Human relations skills (HR) medium (2,5,8)
high (6,10,10)
Computer knowledge (CK)
First aid knowledge (FA)
Driving experience (DE)

Moreover, the time limitation is also an important factor which will force decision makers
to make decisions in a limited time. Applicants for volunteer may have various motivation
and skills. The aim of the volunteer management centers will be effectively assigning the
volunteers to the pre-disaster or post-disaster activities by checking their knowledge and
skills.
In this section, we offer a FIS developed by using MATLAB Fuzzy Logic ToolboxTM
for spontaneous volunteer management. The first step of the FIS is the fuzzification process
which involves the definition of the membership functions of input/output variables by
linguistic variables. The input variables of the inference system will be the skills and
qualifications of volunteers for the classification process. Since the aim of the volunteer
management centers are effectively assigning the volunteers to the pre-disaster or post-
disaster activities by checking their knowledge and skills, the output variables of the system
will be these activities. For the input/output variables, triangular membership functions are
used to represent the linguistic values.
For the volunteer classification process, the inputs of the FIS are determined as age,
compatibility of the physical characteristics, leading and management skills, computer
knowledge, first aid knowledge and driving experience. The input variables and the re-
lated linguistic variables and fuzzy numbers are represented in Table 4.1. Furthermore,
we collected pre/post disaster activities into five groups as search and rescue, health and
sanitation, safety and security, logistics and supply and communication to be the outputs
of the proposed FIS. The compatibility of the volunteers to working in these activities is
determined by linguistic variables represented by triangular fuzzy numbers (Table 4.2).
After the definition of the membership functions of input/output variables by linguistic
variables, the fuzzy rules are structured. Fuzzy rules consist of IF-THEN rules and produce

[email protected]
Fuzzy inference systems for disaster response 89

Table 4.2 Output variables of FIS

Output Fuzzy variables Fuzzy numbers
Search and rescue (S&R)
Health and sanitation (H&S) very compatible (0,0,0.4)
compatible (0.1,0.5,0.9)
Safety and security (S&S) incompatible (0.6,1,1)
Logistics and supply (L&S)
Communication (COM)

output values for the given input values to construct a classification framework for sponta-
neous volunteer management. In this study we totally structured 15 rules and these rules
can be increased for a more detailed analysis. Three examples of structured rules are given
below:

IF “A” is young and “PC” is high and “LM” is high and “HR” is high and “FA” is high,
THEN S&R is very compatible,

IF “PC” is medium and “LM” is medium and “DE” is medium, THEN S&S is compatible,

IF “CK” is low, THEN COM is incompatible,

Subsequent to the construction of fuzzy rules, the fuzzy inference process is performed
to derive a conclusion based on a set of IF-THEN rules. In this study, Mamdani method is
used for fuzzy inference (Eq. (4.2), Figure 4.2). The input parameters for each volunteer
are evaluated and the compatibilities of the volunteer to pre-determined activities are deter-
mined by FIS developed by using MATLAB Fuzzy Logic ToolboxTM. The proposed FIS
is represented in Figure 4.3.
Let we have a volunteer with the following properties:

Age: 25
Compatibility of the physical characteristics: 8
Leading and management skills: 5
Human relations skills: 3
Computer knowledge: 9
First aid knowledge: 2
Driving experience: 6
When we operate FIS for the above input parameters we observe that this volunteer is
very compatible to communication activity with 0.863 membership degree and he is also
compatible to search and rescue, safety and security or logistics and supply activities with

[email protected]
90 Decision Aid Models for Disaster Management and Emergencies

Fig. 4.3 FIS for spontaneous volunteer management

0.5 membership and this volunteer is incompatible for health and sanitation with 0.153
membership degree (Fig. 4.4).
The output of the FIS shows that a volunteer with the above properties is a very suitable
candidate for communication activity. However, if there is no more person required for
communication activity than this volunteer, he/she can be assigned to search and rescue,
safety and security or logistics and supply activities.

4.6 Conclusion

In this chapter we study disaster management as it is one of the most important phe-
nomena in today’s world. We particularly focused on the disaster response, which is one of
the critical stages of disaster management. We used a FIS to deal with uncertainties inher-
ent in a disaster response activity. As a result, this chapter describes potential applications
of FISs in a disaster response system.
In the literature review it is found that the papers using fuzzy set theory in disaster
management focus on estimating the magnitude of the disasters such as fires, earthquakes
and floods. However, to the best of our knowledge there is limited number of works on
disaster response using fuzzy set theory. Therefore our study can be considered as an initial

[email protected]
Bibliography 91

Fig. 4.4 Matlab output of fuzzy rules for spontaneous volunteer management

attempt to use fuzzy set theory in disaster response that provides a FIS for spontaneous
volunteer management, which is one of the most critical activities of disaster response.
The proposed FIS is composed of fuzzy rules where the input and output parameters are
both represented by fuzzy numbers. In the proposed FIS, we used Mamdani type inference
in which the knowledge is obtained by extracting information thoroughly from experts. The
system gives quick and flexible solutions to the problem defined in spontaneous volunteer
management. For further research, we suggest using other type of inference systems based
on data driven approaches including adaptive learning mechanisms such as Sugeno type
inference system.
FIS can also be used for other activities of disaster response systems such as donation
management, evacuation systems, emergency operations management, and also in other
fields of disaster management.

Bibliography

[1] Akter, T. and Simonovic, S.P. (2005) Aggregation of fuzzy views of a large number of stake-
holders for multi-objective flood management decision-making, Journal of Environmental
Management 77, 133–143.
[2] Alexander, D. (2002) Principles of Emergency Planning and Management, Oxford University
Press, USA.

[email protected]
92 Decision Aid Models for Disaster Management and Emergencies

[3] Altay, N., Green, III and Walter, G. (2006) OR/MS research in disaster operations management,
European Journal of Operational Research 175, 475–493.
[4] Aoyama, T. (1988) Study on offshore wave control by a submerged plate, M. Eng. Thesis,
Department of Civil Eng., Univ. of Tokyo, 121. (in Japanese).
[5] Arora, H., Raghu, T.S., and Vinze, A. (2010) Resource allocation for demand surge mitigation
during disasternext term response, Decision Support Systems 50, 304–315.
[6] Burke, J. E. (1964) Scattering of surface waves on an infinitely deep fluid, J. Math. Phys. 6,
805–819.
[7] Chang, F.J. and Chang, Y.T. (2006) Adaptive neuro-fuzzy inference system for prediction of
water level in reservoir, Advances in Water Resources 29, 1–10.
[8] Chang, F.J., Chang, K.Y. and Chang, L.J. (2008) Counterpropagation fuzzy-neural network for
city flood control system, Journal of Hydrology 358, 24–34.
[9] Chang, H.K. and Chien, W.A. (2006) A fuzzy-neural hybrid system of simulating typhoon
waves, Coastal Engineering 53, 737–748.
[10] Cheng, M.Y. and Ko, C.H. (2002) Computer-aided decision support system for hillside safety
monitoring, Automation in Construction 11, 453–466.
[11] Confort, L.K., Ko, K., and Zagoracki, A. (2004) Coordination in Rapidly Evolving Disaster
Response Systems : The Role of Information, American Behavioral Scientist 48, 295–313.
[12] Coppola, D.P. (2006) Introduction to International Disaster Management, Elsevier, Amsterdam.
[13] Cret, L., Yamazaki, F., Nagata, S. and Katayama, T. (1993) Earthquake damage estimation and
decision analysis for emergency shut-off of city gas networks using fuzzy set theory, Structural
Safety 12, 1–19.
[14] Dokas, I.M., Karras, D.A. and Panagiotakopoulos, D.C. (2009) Fault tree analysis and fuzzy
expert systems: Early warning and emergency response of landfill operations, Environmental
Modelling & Software 24, 8–25.
[15] Fritz, C.E. (1961) Disasters, pp. 651–694 in R.K. Merton and R.A. Nisbet (eds.) Contemporary
Social Problems. New York: Harcourt.
[16] Gallant, B.J. (2007) Managing the spontaneous volunteer, Disaster Management Handbook,
Edited by Jack Pinkowski, CRC Press.
[17] Hall, L.O. and Kandel, A. (1991) The evolution from expert systems to fuzzy expert systems,
in Fuzzy Expert Systems, edited by A. Kandel, CRC press.
[18] Huang, C. and Inoue, H. (2007) Soft risk maps of natural disasters and their applications to
decision-making, Information Science 177, 1583–1592.
[19] Huang, C. and Leung, Y. (1999) Estimating the relationship between isoseismal area and earth-
quake magnitude by a hybrid fuzzy-neural-network method, Fuzzy Sets and Systems 107, 131–
146.
[20] Iliadis, L.S. and Spartalis, S.I. (2005) Fundamental fuzzy relation concepts of a D.S.S. for the
estimation of natural disasters’ risk (The case of a trapezoidal membership function), Mathe-
matical and Computer Modelling 42, 747–758.
[21] Iliadis, L.S. (2005) A decision support system applying an integrated fuzzy model for long-term
forest fire risk estimation, Environmental Modelling & Software 20, 613–621.
[22] Jianshe, D., Shuning, W. and Xiaoyin, Y. (1994) Computerized support systems for emergency
decision making, Annals of Operations Research 51, 315–325.
[23] Kahraman, C., Gülbay, M. and Kabak, Ö. (2006) Application of Fuzzy Sets in Industrial En-
gineering: A Topical Classification. in: Kahraman, C. (Ed.), Fuzzy Applications in Industrial
Engineering, Springer-Verlag, Nederland, 1–55.
[24] Keshwani, D.R., Jones, D.D., Meyer, G.E. and Brand R.M. (2008) Rule-based Mamdani-type
fuzzy modelling of skin permeability, Applied Soft Computing 8, 285–294.
[25] Mendel, J. (2001) Uncertain rule-based fuzzy inference systems: Introduction and new direc-
tions, Prentice-Hall.

[email protected]
Bibliography 93

[26] Li, L.F., Wang, J.F. and Wang, C.Y. (2005) Typhoon insurance pricing with spatial decision
support tools, International Journal of Geographical Information Science 19, 363–382.
[27] Liu, X., Zhang, J., Cai, W. and Tong, Z. (2010) Information diffusion-based spatio-temporal
risk analysis of grassland fire disaster in northern China, Knowledge-Based Systems 23, 53–
60.
[28] Lu Y. and Yang D. (2011) Information exchange in virtual communities under extreme disaster
conditions, Decision Support Systems 50, 529–538
[29] Mamdani, E.H. (1974) Application of fuzzy algorithms for control of simple dynamic plant,
IEEE Proceeding 121 (12), 1585–1588.
[30] Mamdami, E.H. and Assilin, S. (1975) An experiment in linguistic synthesis with a fuzzy logic
controller, International Journal of Man-Machine Studies, 7 (1), 1–13.
[31] Oh, H.J. and Pradhan, B. (2011) Application of a neuro-fuzzy model to landslide-susceptibility
mapping for shallow landslides in a tropical hilly area, Computers & Geosciences 37, 1264–
1276.
[32] Rao, R.R., Eisanberg, J. and Schmitt, T. (2007) Improving Disaster Management, The National
Academies Press, Washington, D.C.
[33] Rodríguez, J.T., Vitorianoa, B. and Monteroa, J. (2010), A natural-disaster management term
DSS for Humanitarian Non-Governmental Organisations Knowledge-Based Systems 23, 17–
22.
[34] Ross, T.J. (2004) Fuzzy Logic with Engineering Applications, England: John Wiley & Sons
Ltd.
[35] Saboya, Jr.F., Alves, M.G. and Pinto, W.D. (2006) Assessment of failure susceptibility of soil
slopes using fuzzy logic, Engineering Geology 86, 211–224.
[36] Şen, Z. (2010) Fuzzy Logic and Hydrological Modeling, CRC Press, 195–228.
[37] Şen, Z. (2011) Supervised fuzzy logic modeling for building earthquake hazard assessment,
Expert Systems with Applications, doi:10.1016/j.eswa.2011.05.026.
[38] Sezer, E.A. , Pradhan, B. and Gokceoglu, C. (2011) Manifestation of an adaptive neuro-fuzzy
model on landslide susceptibility mapping: Klang valley, Malaysia, Expert Systems with Ap-
plications 38, 8208–8219.
[39] Sheu, J.B. (2007) An emergency logistics distribution approach for quick response to urgent
relief demand in disasters, Transportation Research Part E-Logistics And Transportation Re-
view 43, 687–709.
[40] Shiri, J. and Kisi, O. (2010) Short-term and long-term streamflow forecasting using a wavelet
and neuro-fuzzy conjunction model, Journal of Hydrology 394, 486–493.
[41] Simonovic, S. P. and Akter, T. (2006) Participatory floodplain management in the Red River
Basin, Canada, Annual Reviews in Control 30, 183–192.
[42] Simonovic, S.P., Nirupama, (2005) A spatial multi-objective decision-making under uncer-
tainty for water resources management, Journal of Hydroinformatics 7, 117–133.
[43] Sugeno, M. and Kang, G. T. (1988) Structure identification of fuzzy model, Fuzzy Set. Syst.,
28, 15–33.
[44] Syed, M.H. (2009) Encylopedia of Disaster Management, Himalaya Publishing House, India.
[45] Takagi, T. and Sugeno, M. (1985) Fuzzy identification of systems and its application to model-
ing and control. IEEE Transactions on System, Man and Cybernetics 15: 116–132.
[46] Topuz, E., Talinli, I. and Aydin, E. (2011) Integration of environmental and human health risk
assessment for industries using hazardous materials: A quantitative multi criteria approach for
environmental decision makers, Environment International 37, 393–403.
[47] Tsai, C.H., Chen, C.W., Chiang, W.L. and Lin, M.L. (2008) Application of geographic informa-
tion system to the allocation of disaster shelters via fuzzy models, Engineering Computations
25, 86–100

[email protected]
94 Decision Aid Models for Disaster Management and Emergencies

[48] Url-1: http://www.ifrc.org/en/what-we-do/disaster-management/

about-disaster-management/ accessed 02-08-2011.
[49] Url-2: http://boroondara.vic.gov.au/accessed06-08-2011
[50] Wallace, W.A. and Balogh, F.D. (1985) Decision Support Systems for Disaster Management,
Public Administration Review 45, 134–140.
[51] Yoon, S.W., Velasquez, J.D., Partridge, B.K., and Nof, S.Y. (2008) Transportation security
decision support system for emergency response: A training prototype, 46 139–148.
[52] Zadeh, L.A. (1965) Fuzzy Sets, Information and Control 8, 338-353.
[53] Zadeh, L.A. (1997) Toward a theory of fuzzy information granulation and its centrality in hu-
man reasoning and fuzzy logic, Fuzzy Sets and Systems, 90, 117–117.
[54] Zografos, K.G., Androutsopoulos, K.N., and Vasilakis, G.M. (2002) A real-time decision sup-
port system for roadway network incident response logistics, Transportation Research Part C:
Emerging Technologies 10, 1–18.

[email protected]
Chapter 5

Security Based Operation in Container Line

Supply Chain: A Literature Review

Dawei Tang 1,∗ , Dong-Ling Xu 2 , Jian-Bo Yang 2 , and Yu-wang Chen 2

1 Safety and Environment Assurance Center, Unilever Colworth, Colworth Science Park,
Sharnbrook, Bedfordshire, UK, MK44 1LQ
2 Manchester Business School, the University of Manchester, Manchester, UK, M15 6PB

∗ [email protected]

Container Line Supply Chain (CLSC) plays a dominant role in world cargo transportation,
but also subjects to various threats during its operation due to its inherent features. The
threats can not only cause disruptions for CLSC operation, they may also lead to various
disasters globally. To help to manage the potential disasters, in this chapter, current re-
search on security for CLSC operation is reviewed. Specifically, the research is divided
into three categories: (1)research from a general level, including standards, regulations,
codes, etc. issued by international, national and industrial organizations, aiming at improv-
ing CLSC security, (2) research on specific security issues in CLSC, e.g., threats faced by
CLSC, features of CLSC, criteria for security assessment of CLSC, etc., and (3) research
on different risk analysis tools and their applications in CLSC security related areas. Fur-
ther, the limitations of current research are also analyzed and potential directions for future
research in this field are suggested.

5.1 Introduction

One of the most prominent features of modern business is that more and more compa-
nies, instead of operating on their own, are operating cooperatively within a supply chain.
Supply chain, since its introduction into business operation, has played and will continue
to play an important role in modern business. However, the level of risks involved in sup-
ply chain is also increasing due to some features of contemporary business, for example,
trend of globalization and outsourcing (Chopra and Meindl, 2004; OECD, 2004), increas-
ing product and service complexity (GAO, 2005), more rapid consumer demand changes
(Sørby, 2003), shorter product lives (Sørby, 2003), and so on.

B. Vitoriano et al. (eds.), Decision Aid Models for Disaster Management 95

and Emergencies, Atlantis Computational Intelligence Systems 7,
DOI: 10.2991/978-94-91216-74-9_5, Ó Atlantis Press 2013

[email protected]
96 Decision Aid Models for Disaster Management and Emergencies

As one of the major categories of supply chain, Container Line Supply Chain (CLSC),
which transports cargo in containers, shares many common characteristics and risks with
general supply chains. At the same time, it also has its unique features.
Since their introduction in the 1950s, containers have become increasingly important
in world cargo transportation as it enables smooth and seamless transfer of cargo among
various modes of transportation, and thus makes cargo movement much more efficient
(Levinson, 2006; Wydajewski and White, 2002). It is estimated that approximately 95
percent of the world’s trade moves by containers (OECD, 2003) and approximately 250
million containers are shipped annually around the world (DHS, 2007). These two figures
clearly indicate that CLSC is a dominant means to ship cargo around the globe (Fransoo
and Lee, 2011; OECD, 2005).
Despite the dominant role of CLSC in world cargo transportation, CLSC is also subject
to many threats due to the following reasons:

• CLSC is complex. A typical container transaction involves as many as 30 different

physical documents and at least 25 different organizations (Cooperman, 2004), includ-
ing raw material vendors, semi-finished and finished product manufactures, exporters,
shippers, freight forwarders, importers, consignees, and so on (Yang, 2011). Further,
documents and organizations involved in CLSCs may spread all over the world. In ad-
dition, among many organizations involved, there is no single organization governing
the international movement of containers (Bakir, 2007) and there is no single organi-
zation that has full responsibility for the CLSC security (OECD, 2003).
• CLSC is vulnerable. During the transportation process of a container, many different
kinds of threats, including cargo theft, smuggling, stowaway, terrorist activity, piracy
and even labor protest, can have a serious impact on CLSC. In addition, any breach in
security in one part of CLSC may compromise the security of the entire chain (Bakir,
2007; Ø. Berleetal et al., 2011; Khan and Burnes, 2007; Sarathy, 2006).
• CLSC operates with insufficient preventative measures. Despite the complexity and
vulnerability of CLSC mentioned above, corresponding preventative measures against
various threats are not sufficient. For example, nowadays, only about 2 percent of the
imported containers are physically inspected in most countries (Closs and McGarrell,
2004), and the bill of lading, which states the contents of containers, is rarely veri-
fied through inspections of containers after packing or during transportation (OECD,
2003).

[email protected]
Security Based Operation in Container Line Supply Chain 97

It can be easily concluded from the above discussion that there is a relatively high prob-
ability for the occurrence of disruptions and even failures of CLSC. On the other hand, the
consequences of the disruptions or failures, which may include immediate consequences,
cascading consequences and long-term consequences, may be severe. They may cause
great human causalities, considerable financial loss, serious environmental pollutions, and
even global disasters. For example, if a port is seriously damaged by the explosion of an
atomic weapon, it may cause 100 billion dollars in port lock-out losses and 5.80 billion
dollars in port recovery losses (Yang, 2011). It can be seen from the above that CLSC is
operating in a highly risky environment.
Facing the fact that CLSC is a dominant but high risky means to transport world cargo,
scholars and researchers have started paying increasing attention on studying security is-
sues of CLSC in recent years, especially after the terrorist attack on September 11th, 2001.
In this chapter, a brief review of current research on security issues in CLSC is provided to
investigate what has been achieved in this field, to discuss the limitations of current research
and to suggest potential directions for future research. Through the discussion in the chap-
ter, the ultimate aim is to provide assistance for decision making relevant to management
of potential disasters caused by CLSC security incidents.
The chapter is organized as follows. In Section 5.2, the concepts of risk, security and
other related terms which will be used throughout this chapter are clarified. Section 5.3 is
dedicated to reviewing current research on security issues in CLSC, including research on
security issues in CLSC from a general level, research on different specific issues about
security in CLSC and research on the applications of different risk analysis tools in CLSC
security related areas. In Section 5.4, the limitations of current research are discussed on
the basis of the above review, followed by the discussion of potential directions for future
research. The chapter is concluded in Section 5.5.

5.2 Basic Definitions

Prior to reviewing current research on security issues in CLSC, some concepts need to
be defined to clarify the boundary and provide a basis of the discussion in this chapter.
Specifically, as the chapter mainly focuses on CLSC security, the concepts of security
should be defined. In addition, for some other terms which are closely related to security,
such as risk, threat, hazard and especially safety, their concepts should also be defined for
the clarification of the scope of security.

[email protected]
98 Decision Aid Models for Disaster Management and Emergencies

Currently, for different purposes, there are different definitions of risk, safety, security,
hazard, threat and other related terms from different points of view (Firesmith, 2003; Jon-
sson, 1998; Lau, 1998; Sørby, 2003; Willis and Ortiz, 2004). According to the content
of the research in this chapter and the opinions of different Port Facility Security Officers
(PFSO) from interviews, the definitions which are used in this chapter are based on those
proposed in (Firesmith, 2003):

• Safety: the degree to which accidental harm is prevented, detected, and reacted to;
• Security: the degree to which malicious harm is prevented, detected, and reacted to;
• Hazard: a situation that increases the likelihood of formation of one or more related
accidental harms;
• Threat: a situation that increases the likelihood of formation of one or more related
malicious harms;
• Risk: a term which is used to describe the likelihood of occurrence and the conse-
quences of a hazard or a threat. Accordingly, risk can be categorized as hazard based
risk and threat based risk. The ‘risk’ discussed in this chapter mainly refers to threat
based risk.

From the above definitions, we can see that threat, threat based risk and security are the
terms regarding malicious harm, while hazard, hazard based risk and safety are the terms
regarding accidental harm. In addition, the relation among threat, threat based risk and
security can be analyzed as follows: threat represents a certain state of a situation; threat
based risk considers both likelihood of the threat and potential consequence caused by the
threat; in addition to the likelihood and the potential consequence, security also considers
the features of the party which is under the threat. Similar conclusion can be drawn for the
relation among hazard, hazard based risk and safety.

5.3 Current research on security issues in CLSC

5.3.1 Research on security issues in CLSC from a general level

One of the most typical documents in this category is the ISPS Code (IMO, 2002a),
which was issued by IMO in 2002. This code is released in response to the “perceived
threats to ships and port facilities in the wake of the 9/11 attacks in the United States”
(PECC, 2004). It is a “comprehensive set of measures to enhance the security of ships and
port facilities” (IMO, 2002a), which covers the specifications of general responsibilities

[email protected]
Security Based Operation in Container Line Supply Chain 99

of contracting governments and ship companies; the general responsibilities of security

officers in ship companies, individual ships and ports; the descriptions of different security
levels of both ships and port facilities; the general requirements on development, training
and drilling of ship and port facility security plans; the verification and certification for
ships, and so on.
As nearly all CLSCs are operating internationally, customs, with their unique author-
ities and expertise, play a central role in ensuring CLSC’s security (WCO, 2007). Corre-
spondently, in 2007, WCO issued a SAFE Framework of Standards (WCO, 2007) to secure
and also facilitate the movement of global trade. This framework is mainly based on two as-
pects: Customs-to-Customs network arrangements and Customs-to-Business partnerships.
The former has 11 standards while the latter has 6 standards. In the standards, the respon-
sibilities of different organizations along a whole chain of cargo custody, from stuffing site
to unloading site, which were always ambiguous in the past, are clearly stated.
Another set of important documents relevant to CLSC security is the ISO 28000 series
(ISO, 2007a; ISO, 2007b; ISO, 2007c; ISO, 2007d), which are the standards on security
management systems for supply chains (LRQA, 2009; Piersall, 2007). Among the series,
ISO 28000 (ISO, 2007a) is a general specification which introduces the elements for secu-
rity management systems, including security management policy, security risk assessment
and planning, implementation and operation for security management, checking and cor-
rective actions, management review and continual improvement. ISO 28004 (ISO, 2007d)
is a detailed explanation on ISO 28000, which explains each part of ISO 28000 in 4 dimen-
sions, i.e., intent, typical inputs, process and typical output of each part.
Besides the documents issued by international organizations, some regional initiatives
are also developed. For example, in Europe, the ISPS Code is incorporated into the EC
Regulation 725/2004 (EC, 2004; TRANSEC, 2011); EC Regulation 884/2005 sets the pro-
cedures for conducting EC inspections in the field of maritime security (EC, 2005a); and
EC Directive 65/2005 aims at enhancing security throughout ports (EC, 2005b; TRANSEC,
2011). In addition, Authorized Economic Operator (AEO) is introduced by EC to CLSC
operators in Europe in 2005 (EC, 2005b) to encourage organizations involved in CLSCs to
enhance security in their operation.
All the documents mentioned above focus on sea transportation of cargo. However, in
CLSC, a container’s voyage contains not only sea transportation but also inland transporta-
tion, the security issues of which need to be considered as well. As such, the International
Shippers and Freight Forwarders Security Code (ISFFS Code) was proposed in 2003 by In-

[email protected]
100 Decision Aid Models for Disaster Management and Emergencies

ternational Trade Procedures Working Group (ITPWG) of United Nations Centre for Trade
Facilitation and Electronic Business (UN/CEFACT) (ITPWG, 2003). This code mainly
develops a set of requirements to ensure the security of cargo transported by road, rail or
inland waterways, including requirements on stuffers and packers; requirements on ware-
houses, storage areas and terminals; requirements on forwarders and transporters; require-
ments on information processors, and so on. For each category, the requirements are further
categorized according to pre-defined security levels.
Apart from the efforts of international/regional organizations, U.S. government also
issued initiatives concerning CLSC security under the threats of terrorists. Among the
initiatives, the Container Security Initiative (CSI) (CBP, 2002a) and the Customs-Trade
Partnership against Terrorism (C-TPAT) (CBP, 2002b) are two of the most important ones.
Both the initiatives were issued around 2002 by Customs and Border Protection (CBP),
a component of Department of Homeland Security (DHS). Both of them are developed
in response to “security vulnerabilities created by ocean container trade and the concern
that terrorists could exploit these vulnerabilities to transport or detonate Weapons of Mass
Destruction (WMD) in the United States” (GAO, 2003). The emphasis of CSI is the re-
quirement to examine highly risky cargo at foreign ports before they are loaded on a vessel
heading to the United States (Robert and Kelly, 2007). It is a government to government
initiative. On the other hand, the emphasis of C-TPAT is the requirement to improve global
supply chain security by private sectors along the whole supply chain (GAO, 2003). To
be more specific, it is a voluntary program between private sectors and customs, which
contains 22 key elements. It is a government to business initiative. In addition to CSI and
C-TPAT, another major program to improve U.S. marine security is the 24-hour Advance
Cargo Manifest Declaration Rule, which requires that containers must be manifested at
least 24 hours before they are loaded to any US-bound vessel. The information submitted
facilitates the targeting and pre-screening of suspected containers. Similar to the 24-hour
rule, a 96-hour rule, which relates to ships rather than cargo, is also proposed by DHS. The
rule requires that all ships calling at U.S. ports should provide a notice of arrival 96 hours
in advance to the U.S. government, which makes it possible for the U.S. government to
target particular ships for which it has security concerns (Pinto, et al., 2008).
The effective area of CSI, C-TPAT and ISPS Code along CLSC can be shown in the
Figure 5.1 as follows (OECD, 2003):
More recently, DHS issued a “Strategy to Enhance International Supply Chain Secu-
rity” (DHS, 2007) in response to the Security and Accountability For Every Port Act (SAFE

[email protected]
Security Based Operation in Container Line Supply Chain 101

Fig. 5.1 Effective area of CSI, C-TPAT and ISPS Code

Port Act) (US Congress, 2006), which is a public law aiming to improve maritime and cargo
security through enhanced layered defenses. The strategy issued by DHS intends to estab-
lish an overarching framework for the secure flow of cargo through supply chains. The
strategy identifies critical nodes along an international supply chain, delineates the roles
and responsibilities of different organizations involved, and most importantly, explains
necessary responsive activities and factors that need to be considered during the recov-
ery process after a disruption. These response and recovery issues are seldom mentioned
in other similar documents.
In addition to the documents issued by governmental and international/regional organi-
zations, some industrial organizations also developed certain initiatives for CLSC security.
For example, Transported Asset Protection Association (TAPA) developed a set of require-
ments and standards to assess the security of organizations involved in CLSCs, such as
Freight Security Requirement (FSR) which specifies the minimum acceptable standards
for security throughout the supply chain and the methods to be used in maintaining those
standards (TAPA, 2011), and Trucking Security Requirement (TSR) which specifies the
minimum acceptable standards for security throughout the supply chain utilizing truck-
ing and associated operations and the methods to be used in maintaining those standards
(TAPA, 2008). TSR may be used in conjunction with FSR.
Further, some academic papers also discuss security related issues in CLSC from a gen-
eral level. For example, current legislations on port safety and security are reviewed and
current security situations faced by ports and EU inter-model transportation are discussed
(Psaraftis, 2005). Security measures taken by the U.S. government and international or-

[email protected]
102 Decision Aid Models for Disaster Management and Emergencies

ganizations are reviewed and the development of a global agreement to ensure security of
CLSC is also suggested to link security and other maritime trade-related issues together
(Stasinopoulos, 2003). Key shore-based and near shore activities associated with maritime
operations, which are currently not covered by ISPS Code, are identified, relationships
among the activities are investigated, and key criteria for a good marine security manage-
ment system are studied (Paulsson, 2003). The impacts of CSI on maritime supply chains,
especially financial impacts are analyzed in general by Banomyong (2005). In addition,
Helmick discusses what had been done and what should be done in the field of port and
marine security (Helmick, 2008), indicating that further refinement and standardization of
risk based decision methodologies and applications are clearly needed, including compre-
hensive threat assessment, consideration of vulnerability variables through the whole global
supply chain, quantification of relative risks and uniform risk assessment methodologies,
etc.
In a word, this stream of research focuses on a general level, aiming at developing
and discussing strategies, policies, principles, specifications, requirements, etc. to enhance
CLSC security. It is the basis and general guidelines of all the research on security issues
in CLSC. However, this stream of research is too general for the development of analytical
CLSC security analysis models and the generation of detailed and specific measurements
to ensure CLSC security.

5.3.2 Research on specific issues of security in CLSC

5.3.2.1 Research on features of CLSC and threats faced by CLSC

Considering the features of CLSC and threats faced by CLSC, OECD issued several
reports. In the report issued in 2005 (OECD, 2005), which concentrates on container trans-
port security across inland and marine transport mode under the potential threat of con-
tainers being used by terrorists as a delivery vehicle for chemical, biological, radiological
or nuclear (CBRN) weapons, the features of a container transport chain are analyzed in
detail. Based on the analysis, the nature of CBRN threat is also revealed. In the report is-
sued in 2003 (OECD, 2003), threats faced by maritime transport are analyzed based on the
following categories: cargo, vessels, people, finance/logistics support and trade disruption.
Research in the above areas can provide general knowledge on how CLSC is operating and
how vulnerable CLSC is to different threats. The knowledge provides a background for
CLSC security analysis.

[email protected]
Security Based Operation in Container Line Supply Chain 103

5.3.2.2 Research on security assessment criteria of CLSC and components of

security plans in CLSC

In some literature, general criteria for CLSC security assessment are analyzed and the
essential components of security plans are discussed.
In ISPS Code (IMO, 2002a), the topics of security assessment and security plans for
both ships and port facilities are two of the most important contents. In the ISPS Code,
data required by security assessment and components required to develop a security plan
are specified in detail.
In the SAFE Framework of Standards (WCO, 2007), elements considered by AEOs
and Customs can be broadly divided into several categories, including cargo security, con-
veyance security, premises security, personnel security, trading partner security and crisis
management & incident recovery. These categories indicate high-level criteria when CLSC
security needs to be assessed.
In ISO 28001 (ISO, 2007b), the best practices for implementing supply chain security
assessments and security plans are discussed, including process and criteria for security
assessment and essential components for a security plan in a general level.
Another literature about assessment criteria is a report issued by RAND Cooperation
in 2004, which is one of the first of a series of studies on the topic of supply chain security
(Willis and Ortiz, 2004). In the report, five capabilities, regarding the efficiency and secu-
rity of global container supply chain, are proposed, and the capabilities can be considered
as general criteria for CLSC security assessment.
The criteria discussed in the above literature are proposed according to different em-
phases on security issues in CLSC from different points of view. A comprehensive under-
standing of the criteria can help to construct a set of high-level attributes for CLSC security
assessment. Research on essential components of a security plan reveals aspects to be
considered to respond to different security incidents, which is also important for security
assessment since responding capability is one of the elements which needs to be considered
when CLSC security is analyzed.

5.3.2.3 Research on countermeasures of CLSC facing different threats

The countermeasures of CLSC against different threats can be roughly divided into 3
categories: managerial measures, operative measures and technical measures.
Managerial countermeasures refer to policies, regulations, requirements or general
methodologies used to respond to threats faced by CLSC. For example: in ISO 28001 (ISO,

[email protected]
104 Decision Aid Models for Disaster Management and Emergencies

2007b), a general methodology for developing countermeasures is proposed; in ISO 28003

(ISO, 2007c), regulations for audit or certification agencies of supply chain security man-
agement systems are discussed; in the WCO Safe Framework of Standards (WCO, 2007),
requirements on the information of imported and exported cargo are provided, which needs
to be submitted to customs; regulations about how to provide critical data of maritime
security incidents to first responders are also developed (Wydajewski and White, 2002).
Operative countermeasures refer to actions taken by different operators in CLSC to
make it more secure. In C-TPAT (CBP, 2002b), 22 key elements are proposed. The op-
erative countermeasures mentioned in the elements include employee background checks,
inspection of empty containers, and so on. The operative countermeasures proposed by
Bakir (2007) include access control, security awareness training, standardization of pa-
perwork security and maintaining the security of warehouse perimeters. Other operative
countermeasures include continuously reviewing and updating security procedures (Closs
and McGarrell, 2004), developing contingency plans (Tang, 2006), securing container in-
tegrity (OECD, 2005), and so on.
Technical countermeasures refer to technologies which can be used to enhance CLSC
security. The countermeasures include the application of newly developed information
technologies (Noda, 2004) and data mining technologies (Lee and Wolfe, 2003), the imple-
mentation of Non-Intrusive Inspection (NII) technologies like X-ray or Gamma-ray scan-
ning (Hessami, 2004), the introduction of so-called ‘smart containers’ (Kim, et al., 2008;
Robert and Kelly, 2007), Radio Frequency Identification (RFID) technique (Yoon. et al,
2007), tracking technique (David, 2005 ; Tsamboulas, 2010), high capable seals (McCor-
mack, et al., 2010; Tirschwell, 2005 ; Tsamboulas, 2010) and so on.
Note that the above 3 categories of countermeasures are not independent of each other,
e.g., managerial countermeasures are implemented through operative countermeasures,
while technical countermeasures provide support to both managerial countermeasures and
operative countermeasures.
All the 3 categories of countermeasures mentioned above can provide ideas on how to
improve CLSC security and which factors should be considered when CLSC security is
analyzed.

[email protected]
Security Based Operation in Container Line Supply Chain 105

5.3.2.4 Research on cost and performance estimation for implementation of se-

curity related measures

The report issued by OECD in 2003 (OECD, 2003) proposes a method to estimate
costs for implementation of different initiatives to enhance maritime security. Specifically,
it mainly estimates the implementation costs of ISPS Code through the estimation of costs
to implement each part of the Code.
Performance estimation can be found from a series of reports issued by United States
Government Accountability Office (GAO). One of the roles of GAO is to assess the perfor-
mance of CSI and C-TPAT during their implementation. Based on the assessment, recom-
mendations can be generated to help CBP improve the performance of CSI and C-TPAT.
In 2003, shortly after CSI and C-TPAT were implemented, GAO issued the first report to
assess their performance (GAO, 2003). One of the problems revealed by GAO in the report
is that there lacks a set of criteria to measure the performance and achievements of the
two initiatives. In 2005, another report (GAO, 2005) was issued by GAO to follow up the
recommendations proposed in the previous report. In this latter report, it was stated that
progress had been made in developing performance criteria for assessing the initiatives’
performance, but the criteria mainly focused on the performance of information sharing
and collaboration among CSI and host country personnel, and they could not be used to
measure the effectiveness of CSI targeting and inspection activities. Following this assess-
ment result, CBP refined overall CSI performance criteria, but the criteria for core CSI
functions are still absent, as indicated in a report issued by GAO in 2008 (GAO, 2008).
In CLSC security analysis, one of the most important tasks is to allocate resources, e.g.,
budgets, human resources, hardware facilities, etc. to improve the security of organizations
involved in CLSC. As resources are always limited, it is necessary to utilize resources in an
efficient and effective way. Accordingly, the consumption of resources for different alterna-
tives to improve security should be estimated. As budgets are the most common resources
for security improvement, the estimation of costs incurred by implementing different secu-
rity improvement alternatives is very important. In addition, different measures for security
improvement have different impact on security, so performance estimation is also essential
for security analysis as the impact of different measures on the performance of relevant
factors related to CLSC security should be estimated. The literature reviewed in this part
can provide a rough and initial idea on how cost and performance can be estimated.

[email protected]
106 Decision Aid Models for Disaster Management and Emergencies

5.3.2.5 Summary

The research reviewed in this section aims at exploring different aspects of CLSC and
is more specific than what is discussed in Section 3.1. Several preliminary ideas, which
can be applied and further developed in the research on security analysis in CLSC, are
discussed in this section. However, nearly all these ideas are proposed in a subjective and
descriptive way and there lacks an analytical and structured model for CLSC security anal-
ysis, which can help to generate practical and specific suggestions on how CLSC security
can be ensured and maintained.

5.3.3 Research on risk analysis methods with their application in the areas rele-
vant to CLSC security analysis

When risk and security analysis methods are discussed, the most fundamental question
is how to model the concept of risk and security. In other words, what are the basic compo-
nents of risk and security? Usually, risk is described by two components, i.e., the likelihood
of occurrence of an undesirable event and the severity of its consequences (Aagedal et al.,
2002; Bahr, 1997; Butler, 2002; IMO, 2002b; Li and Cullinane, 2003). Although security
shares some common characteristics with risk, there are still subtle differences between
these two concepts, as discussed in Section 2, and thus components used to analyze and
model risk and security may not be exactly the same. However, components appropriate
for security modeling are not widely discussed in previous research.
Based on the components of risk, different methods for risk analysis are proposed, but
few of them are specifically applied to risk analysis under the context of CLSC. Thus, the
methods reviewed in the following are mainly about risk analysis in general supply chains
or risk analysis in individual marine operations.
The research on risk analysis in general supply chains began only recently (Khan and
Burnes, 2007; Rao and Goldsby, 2009), and most research is conducted in a descriptive and
qualitative way. For instance, the relation between product design and supply chain risk was
discussed (Khan et al., 2008); a conceptual framework for supply chain risk management
was developed (Manuj and Mentzer, 2008); a general framework for natural disaster re-
sponse of a supply chain was proposed based on interviews with logistics managers (Perry,
2007); while Christopher and Lee (2004) discussed the impact of visibility on supply chain
risk; and Giaglis et al. (2004) proposed an architecture for minimization of logistics risk
by routing vehicles in real time using mobile technologies. On the other hand, for limited
quantitative research related to risk issues in general supply chains, some discussions lie in

[email protected]
Security Based Operation in Container Line Supply Chain 107

the analysis of inventory risk, demand risk, supply risk and transportation risk for individ-
ual organizations in supply chains (Tomlin, 2006; Towill, 2005; Wilson, 2007), while other
discussions focus on modeling relationships among supply chain risk and supply chain
efficiency and profitability (Agarwal and Seshadri, 2000; Wang and Webster, 2007). In
summary, for risk analysis in general supply chains, there is not enough analytical research
conducted, and among the limited research conducted quantitatively, information needed
for risk analysis models is measured numerically. In addition, in existing quantitative re-
search, very limited attention has been paid to the analytical risk assessment of the whole
supply chain.
Among the methods for risk analysis related to marine operations, Formal Safety As-
sessment (FSA) is widely applied, which is introduced by IMO as “a rational and system-
atic process for assessing the risks associated with shipping activity and for evaluating the
costs and benefits of IMO’s options for reducing these risks” (IMO, 2002b). According to
FSA, safety assessment is conducted through the following 5 steps: hazard identification,
risk analysis, Risk Control Options (RCO) development, Cost Benefit Assessment (CBA)
and recommendations for decision making. In addition to the introduction by IMO, there
are also some academic papers discussing the topic of FSA. For example, FSA is applied
to analyze risk in individual containerships (Wang and Foinikis, 2001), cruise ships (Lois,
et al., 2004) and general ships (Wang, 2001); it is also introduced with several practical
applications in the UK, Germany and some Scandinavian countries (Soares and Teixeira,
2001); in addition, a review process, i.e., FSA qualification, is introduced to support the
consolidation of confidence in FSA results (Rosqvist and Tuominen, 2004) and a critical
review of FSA with detailed introduction and analysis for each step of FSA is also proposed
(Kontovas and Psaraftis, 2009). Although FSA has been adopted by IMO since 2002, and
it has been applied in various situations by researchers, it also has its limitations. For ex-
ample, FSA only provides a general framework and process for safety analysis, and there
is limited practical guidance on how to conduct different steps in the process; when FSA
is applied in different situations, risk is usually represented by an index number (Kontovas
and Psaraftis, 2009; Lois, et al., 2004; Rosqvist and Tuominen, 2004; Wang and Foinikis,
2001), which may lead to information loss (Kontovas and Psaraftis, 2009); in addition,
the uncertainty, which is prevalent in risk and safety analysis in maritime operation, are
seldom discussed in the applications of FSA; further, all the applications of FSA focus on
individual maritime operators instead of a whole supply chain.

[email protected]
108 Decision Aid Models for Disaster Management and Emergencies

Another category of methods for risk analysis related to marine operation is based on
probabilities. For example, Event Tree Analysis (ETA) are applied for vulnerability as-
sessment of a maritime transportation system (Ø. Berleetal et al., 2011); both Fault Tree
Analysis (FTA) and ETA are introduced to risk assessment in shipping and ports (Bichou,
2008); in addition, Fault Trees and Event Trees are used to model a general risk manage-
ment framework for a maritime supply chain (Yang, 2011), and ETA and FTA are also
introduced under the framework of FSA (Kontovas and Psaraftis, 2009). In addition to
FTA and ETA, Bayesian Network (BN) is another tool used for risk analysis in the areas
relevant to CLSC. Specifically, BN is used under the framework of FSA by Kontovas and
Psaraftis (2009), it is also used to assess safeguards to secure supply chains (Pai, et al.,
2003) and to assess risk of container supply chain (Yang, 2006). Although ETA, FTA and
BN have the capability to handle uncertainty involved in security analysis in CLSC, they
can only handle the uncertainty caused by randomness. However, due to the complexity
of CLSC operation, not all uncertainty involved in CLSC security analysis are caused by
randomness, and uncertainty can also be caused by fuzzy information or ignorance in sub-
jective judgments. In addition, the precise probabilities required by ETA, FTA and BN
are usually very difficult to generate under the context of CLSC security analysis, as there
is usually insufficient historic data available to generate probabilities in an objective way
(Bichou, 2008). Even available information is usually not sufficient for experts to specify
probabilities according to their subjective knowledge.
To avoid specification of precise probabilities, Fuzzy Logic is applied for risk analysis
in port operations (Ung, 2007), offshore engineering (Ren, et al., 2009) and container sup-
ply chains (Yang, 2006). However, the rationality of fuzzy arithmetic is always arguable,
and the way to aggregate information based on fuzzy logic leads to information loss.
Moreover, some methods in Artificial Intelligence are applied for risk analysis for ma-
rine operations and one of the examples is the application of Artificial Neural Network
(ANN) for risk assessment in port operations (Ung, 2007). Although ANN is a well devel-
oped method, it is a ‘black box’ method which cannot explicitly show its inference process.
Apart from the above methods, the Evidential Reasoning (ER) approach, which is based
on Dempster-Shafer theory (Shafer, 1976), was developed in early 1990’s (Yang and Singh,
1994) and improved in 2000’s (Yang and Xu, 2002). The ER approach has been applied
to analyze risks in offshore engineering systems (Liu, et al. 2005; Ren, et al., 2005; Sii,
et al., 2005) and to assess risk of container supply chains (Yang, 2006). Compared with the
methods reviewed above, the ER approach has the following two major advantages: 1) it

[email protected]
Security Based Operation in Container Line Supply Chain 109

has a solid mathematical basis (Shafer, 1976); and 2) with the introduction of the concept of
belief distribution, information with different features and different kinds of uncertainty can
be accommodated and handled by the ER approach under a unified framework, and there
is no information loss during the reasoning process. Based on the ER approach, RIMER
was proposed (Yang, et al., 2006). Under the framework of RIMER, belief distributions
are used to model individual factors threatening CLSC security, and Belief Rule Bases
(BRB), which incorporates belief distributions into conventional rule bases, are applied to
model the relations among the factors. Apart from the advantages of the ER approach as
mentioned above, RIMER is capable and flexible in representing knowledge contained in
inference models, and unlike ANN which is a ‘black box’ method, the inference process of
RIMER is transparent.
From the above discussion, it can be seen that compared with other methods as reviewed
above, RIMER can be considered as a potential basic tool for CLSC security analysis.

5.4 Summary and limitations of current research and potential directions for
future research for CLSC security

From the above discussions, we can find several features of current research relevant to
CLSC security analysis:

• There is preliminary research on CLSC security. However, the research is either in a

very general level, e.g., regulations, codes, initiatives issued by different organizations,
or only subjective and descriptive in discussing specific security issues of CLSC, and
the analytical discussions on CLSC security are not enough (Rao and Goldsby, 2009;
Tsamboulas, 2010; Yang, 2011), which makes practical and specific guidance on how
to improve CLSC security absent;
• There are a number of methods available for analytical risk analysis, and some of them
are applied in the areas close to CLSC security analysis. However, most of the methods
have limitations when they are directly applied for security analysis in CLSC;
• The research on analytical CLSC security analysis is seldom conducted

In addition, the characteristics of CLSC relevant to security analysis can be summarized

as follows according to the literature reviewed in this chapter:

• CLSC is dominant in world cargo transportation, the operation of CLSC is very com-
plex, and CLSC is vulnerable to various threats during its operation;

[email protected]
110 Decision Aid Models for Disaster Management and Emergencies

• Organizations involved in a CLSC are not operating independently, there are interac-
tions among organizations;
• Due to the complexity of CLSC, the factors which can influence CLSC security may
spread all over the world, and it is unlikely that all the factors can share the same
nature;
• Due to the complexity of CLSC, uncertainty is inevitable and prevalent in CLSC oper-
ation (Bichou, 2008; Rao and Goldsby, 2009). In addition, the sources of uncertainty
are various;
• Although CLSC security has started attracting attention of different organizations and
various researchers recently, historical data regarding CLSC security incidents are very
limited (Bichou, 2008; Kontovas and Psaraftis, 2009);

Based on the above features of current research and the above characteristics related to
CLSC security analysis, the potential directions for future research for CLSC security and
the corresponding requirements are suggested as follows:

• Developing structured frameworks to identify, model and measure different factors as

well as analytical relationships among the factors for CLSC security analysis. The
framework should be developed under the context of the whole supply chain instead
of individual organizations within supply chains. In other words, the relations and
interactions among different organizations in a CLSC should be considered in the
framework. In addition, the framework should be able to accommodate factors and
knowledge involved in the security analysis process with different features and differ-
ent kinds of uncertainty;
• Developing new methods or enhancing the capability of existing methods to rationally
assess the security level of a certain CLSC with different kinds of uncertainty involved
in assessment processes taken into consideration. The generation of parameters needed
for the methods should not be heavily dependent on historical data; experts’ judgments
should play a key role in the specification of the parameters; the bias of judgments
should be minimized and the consistency of the judgments should be maintained;
• Developing new methods to improve CLSC security in a cost-effective way based on
CLSC security assessment result. The outcome generated by the methods should be
able to provide practical and specific suggestions on how to maintain and improve
CLSC security;
• Applying the frameworks and methods to analyze the security of specific CLSCs.

[email protected]
Bibliography 111

5.5 Conclusion

In this chapter, the concept of risk, safety, security and other related terms are dis-
cussed first and current research on security issues in CLSC is reviewed. In general, the
current research can be divided into 3 categories, i.e., research on security issues in CLSC
from a general level, research on specific issues about security in CLSC and research on
the application of risk analysis tools in CLSC and related areas. From this review, it can
be concluded that the research on CLSC security is still at its early stage, and is mainly
descriptive and subjective. There is a clear need to develop analytical and/or quantitative
methods to analyze CLSC security, which should be developed in the context of a whole
supply chain with capability to accommodate different forms of information with different
kinds of uncertainty. The specification of parameters in such methods should be based on
experts’ knowledge and outcomes generated should provide practical and specific sugges-
tions on how CLSC security can be maintained and improved and how potential disasters
which may be caused by CLSC security incidents can be managed.

Acknowledgments

This work was supported by EPSRC under the Grant No. EP/F024606/1. This work
was completed when the first author was conducting his PhD research at Manchester Busi-
ness School (MBS) of The University of Manchester which was also supported by the UK
Overseas Research Scholarship.

Bibliography

[1] J.O. Aagedal, F. Braber, T. Dimitrakos, B.A. Gran, D. Raptis and K. Stolen, “Model-based risk
assessment to improve enterprise security," Proceedings of the Sixth International Enterprise
Distributed Object Computing Conference, Lausanne, Switzerland, pp. 51–62, 2002.
[2] V. Agarwal and S. Seshadri, “Risk intermediation in supply chains”, IIE Transactions, Vol. 32,
No. 9, pp. 819–831, 2000.
[3] N.J. Bahr, “System safety engineering and risk assessment”, Taylor & Francis: London, 1997.
[4] N.O. Bakir, “A brief analysis of threats and vulnerabilities in the maritime domain”, in I. Linkov
et al. (eds.), Managing Critical Infrastructure Risks, Springer: Netherlands, pp. 17–49, 2007.
[5] R. Banomyong, “The impact of port and trade security initiatives on maritime supply chain
management”, Maritime Policy & Management, Vol. 32, No. 1, pp. 3–13, 2005.
[6] Ø. Berleetal, B.E. Asbjørnslett and J.B. Rice, “Formal Vulnerability Assessment of a maritime
transportation system”, Reliability Engineering and System Safety, Vol. 96, pp. 696–705, 2011.
[7] K. Bichou, “Security and risk-based models in shipping and ports: review and critical analysis”,
OECD International Transport Forum, Discussion Paper No. 2008-20, 2008.

[email protected]
112 Decision Aid Models for Disaster Management and Emergencies

[8] S.A. Butler, “Security attribute evaluation method: a cost-benefit approach”, Proceedings of the
24th International Conference on Software Engineering, Florida, U.S., pp. 232–240, 2002.
[9] CBP, “Container Security Initiative”, Custom Border Protection, 2002a.
[10] CBP, “Customs-Trade Partnership Against Terrorism”, Custom Border Protection, 2002b.
[11] S. Chopra and P. Meindl, “Supply chain management: strategy, planning and operations”, 2nd
edition, Pearson Prentice Hall: New Jersey, 2004.
[12] M. Christopher and H. Lee, “Mitigating supply chain risk through improved confidence”, Inter-
national Journal of Physical Distribution & Logistics Management, Vol. 34, No. 5, pp. 388–96,
2004.
[13] D. Closs and E. McGarrell, “Enhancing Security Throughout the supply chain”, IBM Center
for the Business of Government Special Report Series, 2004.
[14] S. Cooperman, “Tracking Cargo”, Security, Vol. 41, No. 8, pp. 20–22, 2004.
[15] M. David, “Editorial; High-tech tracking tools secure incoming cargo”, Electronic Design, Feb.
17, 2005.
[16] DHS, “Strategy to Enhance International Supply Chain Security”, U.S. Department of Home-
land Security, 2007.
[17] EC, “Commission Regulation (EC) No. 884/2005: laying down procedures for conducting
Commission inspections in the field of maritime security”, European Commission, 2005a.
[18] EC, “Directive 2005/65/EC of the European Parliament and of the Council on Enhancing Port
Security”, European Commission, 2005b.
[19] EC, “Regulation (EC) No. 725/2004 of the European Parliament and of the Council on Enhanc-
ing Ship and Port Facility Security”, European Commission, 2004.
[20] D.G. Firesmith, “Common Concepts Underlying Safety, Security, and Survivability Engineer-
ing”, Technical Report CMU/SEI-2003-TN-033, Carnegie Mellon Software Engineering Insti-
tute, 2003.
[21] J.C. Fransoo and C.Y. Lee, “Ocean container transport: an underestimated and critical link in
global supply chain performance”, Production and Operations Management, in press, 2011.
[22] GAO, “Container Security: Expansion of Key Customs Programs Will Require Greater At-
tention to Critical Success Factors”, Government Accountability Office Report, GAO-03-770,
2003.
[23] GAO, “Container Security: A Flexible Staffing Model and Minimum Equipment Requirements
Would Improve Overseas Targeting and Inspection Efforts”, Government Accountability Office
Report, GAO-05-557, 2005.
[24] GAO, “Supply Chain Security: Examinations of High Risk Cargo at Foreign Seaports Have
Increased, but Improved Data Collection and Performance Measures Are Needed”, Government
Accountability Office Report, GAO-08-187, 2008.
[25] G. Giaglis, I. Minis, A. Tatarakis and V. Ziempekis, “Minimizing logistics risk through real
time vehicle routing and mobile technologies”, International Journal of Physical Distribution &
Logistics Management, Vol. 34, No. 9, pp. 749–64, 2004.
[26] J.S. Helmick, “Port and maritime security: A research perspective”, Journal of Transportation
Security, Vol. 1, No. 1, pp. 15–28, 2008.
[27] A.G. Hessami, ”A Systems Framework for Safety and Security: The Holistic Paradigm”. Sys-
tems Engineering, 7(2), pp. 99–112, 2004.
[28] IMO, “International Ship and Port Facility Security Code”, International Maritime Organiza-
tion, 2002a.
[29] IMO, “Guidelines for Formal Safety Assessment for Use in the IMO Rule-making process”,
IMO, 2002b.
[30] ISO, “ISO 28000:2007: Specification for security management systems for the supply chain”,
International Organization for Standardization, 2007a.

[email protected]
Bibliography 113

[31] ISO, “ISO 28001:2007: Security management systems for the supply chain – Best practices
for implementing supply chain security, assessments and plans – Requirements and guidance”,
International Organization for Standardization, 2007b.
[32] ISO, “ISO 28003:2007: Security management systems for the supply chain – Requirements
for bodies providing audit and certification of supply chain security management systems”,
International Organization for Standardization, 2007c.
[33] ISO, “ISO 28004:2007: Security management systems for the supply chain – Guidelines for
the implementation of ISO 28000”, International Organization for Standardization, 2007d.
[34] ITPWG, “The International Shippers and Freight Forwarders Security Code”, International
Trade Procedures Working Group of United Nations Centre for Trade Facilitation and Elec-
tronic Business, 2003.
[35] E. Jonsson, “An integrated framework for security and dependability”, In Proceedings of the
New Security Paradigms Workshop, pages 22–25, Charlottesville, VA, USA, September 1998.
[36] O. Khan and B. Burnes, “Risk and supply chain management: creating a research agenda”, The
International Journal of Logistics Management, Vol. 18 No. 2, pp. 197–216, 2007.
[37] O. Khan, M. Christopher and B. Burnes, “The impact of product design on supply chain risk:
a case study”, International Journal of Physical Distribution & Logistics Management, Vol. 38
No. 5, pp. 412–32, 2008.
[38] S.J. Kim, G.F. Deng, S.K.S. Gupta, M. Murphy-Hoye, “Intelligent networked containers for
enhancing global supply chain security and enabling new commercial value”, the 3rd Interna-
tional Conference on Communication System Software and Middleware, Bangalore, Jan.6-10,
2008, pp. 662–669.
[39] C.A. Kontovas and H.N. Psaraftis, “Formal Safety Assessment: A Critical Review”, Marine
Technology, Vol. 46, No. 1, pp. 45–59, 2009.
[40] O. Lau, “The ten commandments of security”, Computers & Security, Vol. 17, Issue. 2, pp.
119–123, 1998.
[41] K.X. Li and K. Cullinane, “An Economic Approach to Maritime Risk Management and Safety
Regulation”, Maritime Economics and Logistics, Vol. 5, No. 3, pp. 268–284, 2003.
[42] H.L. Lee and M. Wolfe, “Supply chain security without tears”, Supply Chain Management
Review, Vol. 7, No. 1, pp. 12–20, 2003.
[43] M. Levinson, “The box: How the shipping container made the world smaller and the world
economy bigger”, Princeton University Press: Princeton, 2006.
[44] J. Liu, J.B. Yang, J. Wang and H.S. Sii, “Engineering system safety analysis and synthesis
using fuzzy rule-based evidential reasoning approach”, Quality and Reliability Engineering
International, Vol. 21, pp. 387–411, 2005.
[45] P. Lois, J. Wang, A. Wall and T. Ruxton, “Formal safety assessment for cruise ships”, Tourism
Management, Vol. 25, pp. 93–109, 2004.
[46] LRQA, “Supply Chain Security Management Systems and the role of the Assurance Provider”,
Lloyd’s Register Quality Assurance, 2009.
[47] I. Manuj and J.T. Mentzer, “Global supply chain risk management”, Journal of Business Logis-
tics, Vol. 29, No. 1, pp. 133–56, 2008.
[48] E. McCormack, M. Jensen, and A. Hovde, “Evaluating the Use of Electronic Door Seals
(E-Seals) on Shipping Containers”, International Journal of Applied Logistics, 1 (4), 13–29,
October-December 2010.
[49] S. Noda, “Container Shipping and Security Issues: the Carriers’ Responsibility in the Fight
against Terrorism”, Maritime Economics & Logistics, Vol. 6, No. 2, pp. 157–186, 2004.
[50] OECD, “Security in Maritime Transport: Risk Factors and Economic Impact”, Organisation
for Economic Co-operation and Development Report, 2003.
[51] OECD, “Report on Container Transport Security across modes: Executive Summary and Con-
clusions”, Organisation for Economic Co-operation and Development Report, 2004.

[email protected]
114 Decision Aid Models for Disaster Management and Emergencies

[52] OECD, “Container Transport Security across Modes”, Organisation for Economic Co-
operation and Development Report, 2005.
[53] R.R. Pai, V.R. Kallepalli, R.J. Caudill and M.C. Zhou, “Methods toward supply chain risk anal-
ysis”, IEEE International Conference on Systems, Man and Cybernetics, Vol. 5, Washington,
U.S., pp. 4560–4565, 2003.
[54] U. Paulsson, “Managing risks in supply chains: an article review”, Presented at NOFOMA,
Oulu, Finland, 2003.
[55] PECC, “Study on the Mutually Supportive Advancement of APEC’S Trade Facilitation and
Secure Trade Goals post September 11”, Analysis and Case Studies prepared for APEC, Sin-
gapore Pacific Economic Cooperation Council, 2004.
[56] M. Perry, “Natural disaster management planning: a study of logistics managers responding to
the tsunami”, International Journal of Physical Distribution & Logistics Management, Vol. 37,
No. 5, pp. 409–33, 2007.
[57] C.H. Piersall, “Securing the global supply chain”, ISO Focus, October, pp. 33–35, 2007.
[58] C.A. Pinto, G. Rabadi and W.K. Talley, “U.S. Port Security”, in W.K. Talley. (eds.), “Maritime
Safety Security and Piracy”, Informa: London, 2008.
[59] H.N. Psaraftis, “EU Ports Policy: Where do we Go from Here?”, Maritime Economics & Lo-
gistics, Vol. 7, No. 1, pp. 73–82, 2005.
[60] S. Rao and T. Goldsby, “Supply chain risks: a review and typology”, The International Journal
of Logistics Management, Vol. 20, No. 1, pp. 97–123, 2009.
[61] J. Ren, I. Jenkinson, H.S. Sii, J. Wang, D.L. Xu and J.B. Yang, “An offshore safety assessment
framework using fuzzy reasoning and evidential synthesis approaches”, Journal of Marine En-
gineering & Technology, No. A6, pp. 3–16, 2005.
[62] J. Ren, J. Wang, I. Jenkinson, D. L. Xu, J. B. Yang, “An offshore risk analysis method using
fuzzy Bayesian network”, Journal of Offshore Mechanics and Arctic Engineering, Vol. 131,
No. 4, 2009.
[63] W. Robert and J.D. Kelly, “Containing the Threat: Protecting the Global Supply Chain through
Enhanced Cargo Container Security”, The Reform Institute, reform brief, 2007.
[64] T. Rosqvist and R. Tuominen, “Qualification of Formal Safety Assessment: an exploratory
study”, Safety Science, Vol. 42, pp. 99–120, 2004.
[65] G. Shafer, “A mathematical theory of evidence”, Princeton, N.J.: Princeton University Press,
1976.
[66] R. Sarathy, “Security and the Global Supply Chain”, Transportation journal, Vol. 45, No. 4, pp.
28–51, 2006.
[67] H.S. Sii, J. Wang, A.G. Eleye-Datubo, J.B. Yang and J. Liu, “Safety assessment of FPSO turret-
mooring system using approximate reasoning and evidential reasoning”, Journal of Marine
Technology, Vol. 42, No. 2, pp. 88–102, 2005.
[68] C.G. Soares and A.P. Teixeira, “Risk assessment in maritime transportation”, Reliability Engi-
neering and System Safety, Vol. 74, pp. 299–309, 2001.
[69] K. Sørby, “Relationship between security and safety in a security-safety critical system: Safety
consequences of security threats”, Master thesis, Norges Teknisk-Naturvitenskapelige Univer-
sitet, Trondheim, Norway, 2003.
[70] D. Stasinopoulos, “Maritime Security – The Need for a Global Agreement”, Maritime Eco-
nomics & Logistics, Vol. 5, No. 3, pp. 311–320, 2003.
[71] C. Tang, “Robust strategies for mitigating supply chain disruptions”, International Journal of
Logistics Research and Applications, Vol. 9, No. 1, pp. 33–45, 2006.
[72] TAPA, “Freight Suppliers Minimum Security Requirements”, Transported Asset Protection As-
sociation, 2011.
[73] TAPA, “Freight Suppliers Minimum Trucking Security Requirements”, Transported Asset Pro-
tection Association, 2008.

[email protected]
Bibliography 115

[74] P. Tirschwell, "An opportunity for container seals," Journal of Commerce, Feb. 7, 2005.
[75] B. Tomlin, “On the value of mitigation and contingency strategies for managing supply chain
disruption risks”, Management Science, Vol. 52, No. 5, pp. 639–57, 2006.
[76] D.R. Towill, “The impact of business policy on bullwhip induced risk in supply chain manage-
ment”, International Journal of Physical Distribution & Logistics Management, Vol. 35, No. 8,
pp. 555–75, 2005.
[77] TRANSEC, “A brief overview of the United Kingdom National Maritime Security Pro-
gramme”, Department of Transport, the UK, 2011.
[78] D. Tsamboulas, “Terrorism and the Threat to Multimodal Transport – An Overview”, in
M. Bell, et al. (eds.), Security and Environmental Sustainability of Multimodal Transport,
Springer: Netherlands, pp. 3–22, 2010.
[79] S.T. Ung, “The Development of Safety and Security Assessment Techniques and their Appli-
cation to Port Operations”, PhD Thesis, Liverpool John Moores University, 2007.
[80] U.S. Congress, “Security and Accountability For Every Port Act”, Public Law, 109–347, 2006.
[81] J. Wang, “The current status and future aspects in formal ship safety assessment”, Safety Sci-
ence, Vol. 38, pp. 19–31, 2001.
[82] J. Wang and P. Foinikis, “Formal safety assessment of containerships”, Marine Policy, Vol. 25,
No. 2, pp. 143–157, 2001.
[83] C. Wang and S. Webster, “Channel coordination for a supply chain with a risk neutral manu-
factures and a loss averse retailer”, Decision Sciences, Vol. 38, No. 3, pp. 361–89, 2007.
[84] WCO, “SAFE Framework of Standards”, World Customs Organization, 2007.
[85] H.H. Willis and D.S. Ortiz, “Evaluating The Security Of The Global Containerized Supply
Chain”, RAND Technical Report Series, TR-214-RC, 2004.
[86] M. Wilson, “The impact of transportation disruptions on supply chain performance”, Trans-
portation Research, Part E, Vol. 43, No. 4, pp. 295–320, 2007.
[87] K.J. Wydajewski and B.L. White, “Processes and techniques for providing critical data to first
responders to maritime security incidents”, Oceans 2002 IEEE/MTS Conference Proceedings,
Vol. 2, Mississippi, U.S., pp. 1180–1190, 2002.
[88] Y.C. Yang, “Risk management of Taiwan’s maritime supply chain security”, Safety Science,
Vol. 49, pp. 382–393, 2011.
[89] Z.L. Yang, “Risk Assessment and Decision Making of Container Supply Chains”, PhD Thesis,
Liverpool John Moores University, 2006.
[90] J. B. Yang, J. Liu, J. Wang, H. S. Sii and H. W. Wang, “A belief rule-base inference methodology
using the evidential reasoning approach – RIMER”, IEEE Transactions on Systems, Man, and
Cybernetics – Part A, Vol. 36, No. 2, pp. 266–285, 2006.
[91] J.B. Yang and M.G. Singh, "An evidential reasoning approach for multiple attribute decision
making with uncertainty", IEEE Transactions on Systems, Man, and Cybernetics, Vol. 24, No.
1, pp. 1–18, 1994.
[92] J.B. Yang and D.L. Xu, “On the evidential reasoning algorithm for multi-attribute decision
analysis under uncertainty”, IEEE Transactions on Systems, Man, and Cybernetics - Part A:
Systems and Humans, Vol. 32, No. 3, pp. 289–304, 2002.
[93] W.J. Yoon, S.H. Chung, S.J. Lee and Y.S. Moon, “Design and Implementation of an Active
RFID System for Fast Tag Collection," 7th IEEE International Conference on Computer and
Information Technology, Fukushima Prefecture, Japan, pp. 961–966, 2007.

[email protected]
Chapter 6

A Linguistic-Valued Information Processing

Method for Fuzzy Risk Analysis

Li Zou 1 , Zheng Pei 2 , Da Ruan 3 , and Yang Xu 4

1 School of Computer & Information Technology, Liaoning Normal University, Dalian

116029, China
2 School of Mathematics & Computer Engineering, Xihua University, Chengdu 610039,

China
3 SCK • CEN & Ghent University, Boeretang 200, 2400 Mol, Belgium
4
Intelligent Control Development Center, Southwest Jiaotong University, Chengdu
610031, China

E-mail: [email protected], [email protected], [email protected]

Risk analysis is a crucial issue to be handled in disaster management. Fuzzy risk analysis,
i.e., risk analysis model using fuzzy set theory aims to assess the risk of hazard event under
incomplete or imprecise environment. In practice, the evaluations of risk is always ex-
pressed by linguistic values in natural language. In this chapter, we present two approaches
for fuzzy risk analysis with linguistic evaluation values. The first approach is based on the
unbalanced linguistic weighted geometric operator, which can be used to deal with aggre-
gation of unbalanced linguistic risk evaluation values with numerical weights or linguistic
weights. The advantage of the approach is that the evaluation result is linguistic value
which is no need of approximation processing and easier to communicate to decision and
policy-makers. The other approach is based on linguistic truth-values lattice implication
algebra from the logical algebraic point of view. We discuss the operations and special
properties of the lattice implication algebra with evaluation-10 linguistic evaluation values.
The reasoning and aggregation process directly act on the linguistic evaluation values in the
risk analysis process. This approach can better express and handle both comparable and
incomparable linguistic information in risk analysis domains. The proposed approaches
aim at provide a support for risk analysis in different application context including disaster
management under uncertain environment.

B. Vitoriano et al. (eds.), Decision Aid Models for Disaster Management 117
and Emergencies, Atlantis Computational Intelligence Systems 7,
DOI: 10.2991/978-94-91216-74-9_6,  Atlantis Press 2013

[email protected]
118 Decision Aid Models for Disaster Management and Emergencies

6.1 Introduction

From the system point of view, risk analysis means to combine the individual responses
to one statement on the system’s performance for the purpose of decision-making. It deals
with the occurrence of individual failure events (e.g., changes in components or in relations
among components as distinct points in space and time) and their possible consequences on
the system level [1]. In the procedure of risk analysis, the estimation of the likelihood (e.g.,
frequencies) and the consequences of hazard occurrence are included. The estimation of
the likelihood of hazard occurrence depends greatly on the reliability of the system’s com-
ponents, the interaction of the components taking the system as a whole and human-system
interactions. Risk evaluation needs a systematic research of accidental scenarios, includ-
ing failure rates for the component (e.g., safety barriers) as well as for operator behavior
(human factor) within an evolving environment [3].
In practice, information of risk analysis stems from historical data of complex systems
or knowledge of decision makers, experts and regulators. Accordingly, there exist qual-
itative and quantitative methods in risk analysis to identify the risk drivers, assess their
likelihood of occurrence and their potential consequences, and find ways to monitor and
then mitigate the risks [2]. The differences between qualitative and quantitative methods
for risk analysis are focused on the evaluation of the likelihood of accident sequences. In
the quantitative methods, probability and Bayesian networks are the main tools for risk
analysis due to their ability to model probabilistic data with dependencies between events,
compute the distribution probabilities in a set of variables according to the observation of
some variables and the prior knowledge of the others, and quantify low probability events
[3, 23, 24].
Qualitative methods to risk analysis are largely based on expert judgement under the
assumed boundary conditions, in which statements and implications are considered in per-
forming a risk analysis, and the result is expressed by a linguistic value, e.g., safe or un-
safe. Due to uncertainty and complexity, information about the probabilities of various risk
items is vaguely known. This makes researchers to consider risk analysis based on fuzzy
logic, i.e., fuzzy risk analysis. In recent years, many methods to fuzzy risk analysis based
on fuzzy numbers have been discussed, e.g., based on fuzzy arithmetic operations [35], the
similarity measure of fuzzy numbers [5, 7], interval fuzzy numbers, the alpha level sets, and
the ranking of fuzzy numbers [4, 8–13] and fuzzy partition tree [37]. In all of these fuzzy
risk analysis, many shapes (membership functions) associated to fuzzy numbers are used
to represent fuzziness of the evaluating value of the risk of each subcomponent [34], e.g.,

[email protected]
A Linguistic-Valued Information Processing Method for Fuzzy Risk Analysis 119

triangular fuzzy numbers, trapezoidal fuzzy numbers and interval fuzzy numbers. Then
fuzzy arithmetic operations, the similarity measure of fuzzy numbers and the ranking of
fuzzy numbers could help us to evaluate the result of the risk analysis system.
To the best of our knowledge, there are the following three drawbacks when we use
membership functions associated to fuzzy numbers to represent fuzziness in risk analysis:

(1) Fuzzy arithmetic operations, the similarity measure of fuzzy numbers and the ranking
of fuzzy numbers depend on membership functions associated to fuzzy numbers, e.g.,
triangular fuzzy numbers, trapezoidal fuzzy numbers and interval fuzzy numbers. Dif-
ferent shapes of fuzzy numbers lead to different results. Thus, in practice, which one
is the best for a given risk analysis is a problem;
(2) In risk analysis, linguistic values or fuzzy numbers rather than membership functions
are used to represent fuzziness of the evaluation value. The result of risk analysis does
not depend on their membership functions;
(3) In many cases, processing membership functions associated to linguistic values in-
crease the computational complexity, and membership functions no longer keep the
same form after fuzzy arithmetic operations. As the results do not exactly match any
of the initial linguistic values, an approximation process must be developed to express
the results in the initial expression domain which induces the consequent loss of infor-
mation and hence the lack of precision.

From the practical point of view, a main characteristic of fuzzy risk analysis is that
they always use natural languages and the result of fuzzy risk analysis is always expressed
by natural language. Risk analysis based on Computing with Words (CWW) provides
an alternative method to process these natural languages. CWW, which was proposed by
Zadeh, is a methodology for reasoning, computing and decision-making with information
described in a natural language [14, 21, 43, 47]. Basically, CWW is a system of computa-
tion two important capabilities [30]: a) the capability to preciseate the meaning of words
and propositions drawn from natural language; b) the capability to reason and compute
with precisiated words and propositions. In CWW, linguistic values rather than member-
ship functions or numerical values play an important or key role, i.e., linguistic values are
computational variables. This avoids the loss of information in the results of risk analysis,
which in this way are also easier to communicate with decision- and policy-makers.
Informally, risk analysis based on CWW belongs to qualitative methods. In [15], fuzzy
number indexes of linguistic evaluation values was used to deal with fuzzy risk analysis
problems, in which the linguistic evaluation values were represented by their fuzzy number

[email protected]
120 Decision Aid Models for Disaster Management and Emergencies

indexes, and the final evaluation value was obtained by linguistic information fusion based
on fuzzy number indexes of linguistic values [17, 20]. In [16], we considered unbalanced
linguistic information in fuzzy risk analysis, and proposed a new method for fuzzy risk
analysis with unbalanced linguistic evaluation values.
There exists incomparable linguistic evaluation values in fuzzy risk analysis. Xu, Pei
et al. characterized the set of linguistic values by a lattice-valued algebraic structure and
investigated the corresponding logic systems with linguistic truth-value based on lattice
implication algebra (LIA) [38, 39]. From the lattice-valued logic system point of view [40,
41], linguistic truth-values by the lattice implication algebra can express both comparable
and incomparable elements [44, 45]. Zou et al. [46] proposed a framework of linguistic
truth-valued propositional logic and developed the reasoning method of linguistic truth-
valued logic system.

6.2 Preliminaries

In this section, we first outline the notion of fuzzy risk analysis, the 2-tuple fuzzy lin-
guistic representation model, unbalanced linguistic term sets and linguistic truth-valued
lattice implication algebra.

6.2.1 Fuzzy risk analysis

The aim of fuzzy risk analysis is to evaluate the probability of failure of every compo-
nent consisted of many sub-components, in which estimations of severity of loss and prob-
ability of failure of sub-components are included. Formally, risk analysis can be described
as follows: assume that there are r components P1 , P2 , . . . and Pr made by r manufactories
M1 , M2 , . . . and Mr , respectively. Each component Pi consists of s sub-components pi1 ,
pi2 , . . . and pis . Each sub-component pi j (1  i  r, 1  j  s) is evaluated by means of its
severity of loss Li j and its probability of failure Fi j . Each probability of failure Fi of compo-
nent Pi is calculated through both the correspondent severity of loss Li j and the probability
of failure Fi j (1  j  s). Then the larger the value of Fi is, the higher the probability of
failure of component Pi is made by manufactory Mi (1  i  r). The structure of fuzzy risk
analysis can be shown in Fig. 6.1 [8].
According to the evaluation of the likelihood of accident sequences, qualitative and
quantitative approaches to risk analysis have been discussed. In the quantitative approaches
to risk analysis, every severity of loss Li j and probability of failure Fi j of sub-components

[email protected]
A Linguistic-Valued Information Processing Method for Fuzzy Risk Analysis 121

component Pi evaluated by
probability of failure Fi

sub-component pi1 sub-component pi2 sub-component pis

severity of loss Li1 severity of loss Li2 ··· severity of loss Lis
probability of failure Fi1 probability of failure Fi2 probability of failure Fis

Fig. 6.1 The structure of fuzzy risk analysis

pi j (1  i  r, 1  j  s) are described by probabilities. Then Bayesian networks are used

for obtaining final evaluation results. As a general modeling approach, Bayesian networks
offers a compact presentation of the interactions in a stochastic system by visualizing sys-
tem variables and their dependencies. Formally, a Bayesian network consists of two main
parts: a qualitative part and a quantitative part. The qualitative part is a directed acyclic
graph mirrored the nodes of the system variables and the conditional dependence between
variables are represented by the edges of the graph. Conditional probability functions are
included in the quantitative part according to the relations between the nodes of the graph.
According to Figure 6.1, the probability of failure Fi of component Pi can be calculated
by the conditional probability distribution P(Fi | Pa(Li1 , Fi1 ), Pa(Li2 , Fi2 ), . . . , Pa(Lis , Fis )),
where Pa(Li j , Fi j ) (1  j  s) is joint distribution function of sub-component pi j .
Due to the complexity of systems and lack of data, risk assessments often rely on ex-
perts’ opinion. The experts’ opinion has its inherent vagueness, which is expressed by
linguistic evaluation values, such as very low, fairly low, pretty low, medium, almost high,
more or less high, very high and absolutely high, etc. According to the concept of linguistic
variable proposed by Zadeh [42], membership functions can be used to represent linguistic
evaluation values and handle risk analysis. Generally, risk analysis based on fuzzy numbers
is presented as follows:

(1) Aggregate all evaluation values {(Li j , Fi j ) | 1  j  s} of component Pi by fuzzy num-

ber aggregation operators, e.g., fuzzy weighted mean method and the generalized fuzzy
number arithmetic operator, where Li j and Fi j are represented by corresponding fuzzy
s
F ⊗Li j
j=1 i j
numbers (membership functions), i.e., Fi = s , in which ⊗ is a generalized
L
j=1 i j
fuzzy numbers multiplication and Fi is a fuzzy number;
(2) Rank fuzzy numbers {Fi | 1  i  r} by the ranking values of the fuzzy numbers. The
larger the value of Fi is, the higher the risk of the manufactory Mi is;

[email protected]
122 Decision Aid Models for Disaster Management and Emergencies

(3) Approximate fuzzy number max{Fi | 1  i  r} to linguistic values, the final evaluation
result is a linguistic evaluation value corresponding to such a fuzzy number.

In [15, 16], alternative methods have been presented, i.e., CWW is used to handle risk
analysis and compared with methods for risk analysis based on fuzzy numbers. Evalu-
ation values are linguistic evaluation values instead of fuzzy numbers (their membership
functions) and the final evaluations are directly represented by linguistic evaluation values
without approximation from a fuzzy number to a linguistic evaluation value. Generally,
risk analysis based on CWW is presented as follows:

(1) Aggregate all linguistic evaluation values {(Li j , Fi j ) | 1  j  s} of component Pi by

linguistic aggregation operators, i.e., Fi = f ({(Li j , Fi j ) | 1  j  s}), in which, f is a
linguistic aggregation operator;
(2) Rank linguistic evaluation values {Fi | 1  i  r} by the ranking method of linguistic
values. The larger the value of Pi is, the higher the risk of the manufactory Mi is.

6.2.2 The 2-tuple fuzzy linguistic representation model

The 2-tuple linguistic representation model was introduced by Herrera [25]. Let S =
{s0 , . . . , sg } be the initial finite linguistic value set. Formally, the 2-tuple linguistic repre-
sentation model is formed by (si , α ), in which si ∈ S (i ∈ {0, 1, . . . , g}) and α ∈ [−0.5, 0.5),
i.e., linguistic information is encoded in the space S × [−0.5, 0.5). Based on the represen-
tation (si , α ), we can easily obtain the following symbolic translation of linguistic values
from β ∈ [0, g] to S × [−0.5, 0.5), i.e., Δ : [0, g] → S × [−0.5, 0.5), β �→ (si , α ), in which
i = round(β ) (round(·) is the usual round operation) and α = β − i ∈ [−0.5, 0.5). Intu-
itively, Δ(β ) = (si , α ) expresses that si is the closest linguistic value to β , and α is the
value of the symbolic translation. Additionally, there is a Δ−1 function such that from a 2-
tuple it returns its equivalent numerical value β ∈ [0, g], i.e., Δ−1 : S × [−0.5, 0.5) → [0, g],
Δ−1 (si , α ) = i + α = β .
In fact, it defines a set of transformation functions between linguistic values and 2-
tuples linguistic representations as well as numeric values and 2-tuples linguistic represen-
tations. Evidently, an order relation on S × [−0.5, 0.5) can be deduced by Δ−1 , i.e., for any
(si , αi ), (s j , α j ) ∈ S × [−0.5, 0.5), (si , αi )  (s j , α j ) if and only if Δ−1 (si , αi )  Δ−1 (s j , α j ).

[email protected]
A Linguistic-Valued Information Processing Method for Fuzzy Risk Analysis 123

6.2.3 Representation of unbalanced linguistic terms

Unbalanced linguistic terms proposed in [26] are used to deal with scales for assessing
preferences where the experts need to assess a number of terms in a side of reference
domain higher than in the other one. Generally, an unbalanced linguistic term set S has a
minimum label, a maximum label and a central label. And the remaining labels are non-
uniformly and non-symmetrically distributed around the central one on both left and right
lateral sets, i.e., we can represent S on the form S = Sl ∪ Sc ∪ Sr , in which Sl contains all
left lateral labels but the central label, Sc just contains the central label and Sr contains all
right lateral labels higher than the central label.

Example 6.1 ([27]). S = {none (N), low (L), medium (M), almost high (AH), high (H),
quite high (QH), very high (V H), almost total (AT ), total (T )} is an unbalanced linguistic
term set, in which Sl = {N, L}, Sc = {M} and Sr = {AH, H, QH,V H, AT, T }.
N L M AH H QH VH AT T

Fig. 6.2 Scale with more values on the right of the midterm.

To obtain 2-tuple fuzzy linguistic representations of unbalanced linguistic terms, we

need the concept of linguistic hierarchies LH = t l(t, n(t)) [29], which takes into account
a set of levels where each level is a linguistic term set with different granularity from the
remaining levels of the hierarchy, where l(t, n(t)) is a linguistic hierarchy with t being
a number that indicates the level of the hierarchy and n(t) the granularity of the linguistic
term set of t. The linguistic term set Sn(t+1) of the level t +1 is obtained from its predecessor
Sn(t) as l(t, n(t)) → l(t + 1, 2 × n(t) − 1). Transformation function of LH is defined as
follows [28]: for any t and t � , T Ftt� : l(t, n(t)) −→ l(t � , n(t � )) such that

�t−1 (si , α n(t) ) × (n(t �) − 1)

n(t)
, α n(t) ) = �t �
n(t)
T Ftt� (si . (6.1)
n(t) − 1

Example 6.2. Let

LH = l(1, 3) ∪ l(2, 5) ∪ l(3, 9) ∪ l(4, 17)

= s30 , s31 , s32 ∪ s50 , s51 , . . . , s54 ∪ s90 , s91 , . . . , s98 ∪ s17
0 , s1 , . . . , s16
17 17

[email protected]
124 Decision Aid Models for Disaster Management and Emergencies

be a linguistic hierarchy. (s95 , 0.3) is a 2-tuple fuzzy linguistic representation of level 3, it’s
2-tuple fuzzy linguistic representation in level 2 is
�−13 (s5 , 0.3) × (5 − 1)
9
T F23 (s95 , 0.3) = �2
9−1
5.3 × 4
= �2 = �2 (2.65)
8
= (s53 , −0.35).

By using linguistic hierarchies LH = t l(t, n(t)), we can obtain the 2-tuple fuzzy lin-
guistic representation of each term of unbalanced linguistic term set in LH by using the
algorithm presented in [26].

Example 6.3. Continuing Example 6.1. For unbalanced linguistic term set S =
n(2)−1
{N, L, M, AH, H, QH,V H, AT, T }, 1) Due to n(2) = 5 and 2 = |Sl | = |{N, L}| = 2,
the representation of Sl is obtained from level 2 of LH as follows:
{L ← s51 , N ← s50 };
n(3)−1
2) Due to n(3) = 9, n(4) = 17 and 2 = 4 < |Sr | = |{AH, H, QH,V H, AT, T }| = 6 <
n(4)−1
2 = 8, we use level 3 and level 4 to represent Sr = {AH, H, QH,V H, AT, T }, according
to lab3 and lab4 of [26], {AH, H} and {QH,V H, AT, T } are represented in level 3 and level
4, respectively,
{AH ← s95 , H ← s96 }, {QH ← s17
13 ,V H ← s14 , AT ← s15 , T ← s16 };
17 17 17

3) According to density and bridging representation gaps [26], the upside and the downside
of the central label M are represented in level 2 and 3 of LH by means of s52 and s94 , respec-
tively. The upside and the downside of the label H are represented in level 3 and 4 of LH
by means of s96 and s17
12 , respectively;
4) the final 2-tuple fuzzy linguistic representations of S in LH are Sl : {N ← s50 , L ← s51 },
Sc : {M ← s52 ∪s94 }, Sr : {AH ← s95 , H ← s96 ∪s17
12 , QH ← s13 ,V H ← s14 , AT ← s15 , T ← s16 }.
17 17 17 17

Let S be an unbalanced linguistic term set. Formally, for any 2-tuple fuzzy linguistic
representation (si , αi ) (si ∈ S and αi ∈ [−0.5, 0.5)), (si , α ) can be converted by the following
unbalanced linguistic transformation functions in LH and vice versa, i.e.,
L H : S × [−0.5, 0.5) −→ LH × [−0.5, 0.5),
G(i) G(i)
(si , αi ) �−→ (sI(i) , αi ) such that sI(i) ∈ LH;
L H −1 : LH × [−0.5, 0.5) −→ S × [−0.5, 0.5),
n(t)
(sk , αk ) �−→ (si , λ ),
n(t)
in which sk ∈ Sn(t) ⊂ LH. (si , λ ) is decided by cases as follows:

[email protected]
A Linguistic-Valued Information Processing Method for Fuzzy Risk Analysis 125

n(t)
(1) If there exists si ∈ S such that si ← sk , then we consider two possible situations
depending on how si is represented in LH:
(a) if si is represented with only one label in LH, e.g., L ← s51 in Example 6.3, then
(si , λ ) = (si , αk );
(b) if si is represented with two labels in LH from levels, e.g., H ← s96 ∪ s17
12 in
Example 6.3, then (si , λ ) depends on the localization of si ∈ S, and λ = αk or
�t−1 (sk �t−1 (sk ,αk )×(n(t+1)−1)
n(t) n(t)
,αk )×(n(t+1)−1)
n(t)−1 − round( n(t)−1 ).

(2) If there exists no si ∈ S such that si ← sk , then L H −1 ((sk , αk )) =

n(t) n(t)

L H −1 (T Ftt� (sk , αk )), in which t � is a level of LH such that T Ftt� (sk , αk ) =

n(t) n(t)

n(t � ) n(t � )
(sk� , αk� ) and ∃s j ∈ S, s j ← sk� .

Example 6.4. Continuing Example 6.3. We have

L H (H, 0.3) = (s96 , 0.3),

L H −1 (s17
13 , −0.2) = (QH, −0.2),

L H −1 (s17
12 , −0.2) = L H
−1
(T F34 (s17
12 , −0.2)) = L H
−1 9
(s6 , −0.1) = (H, −0.1),
L H −1 (s17
12 , 0.4) = L H
−1
(T F34 (s17
12 , 0.4)) = L H
−1 9
(s6 , 0.2) = (H, 0.4).

6.2.4 Linguistic truth-valued lattice implication algebra

Inspired by hedge algebra [31, 32], we construct lattice implication algebra of linguistic
truth values in which both comparable and incomparable linguistic truth values can be
expressed.

Definition 6.1 ([38]). Let (L, ∨, ∧, �, O, I) be a bounded lattice with universal boundaries O
(the least element) and I (the greatest element) respectively, and “�” be an order-reversing
involution. For any x, y, z ∈ L, if a mapping →: L × L → L satisfies:
(I1 ) x → (y → z) = y → (x → z),
(I2 ) x → x = I,
(I3 ) x → y = y� → x� ,
(I4 ) if x → y = y → x = I, then x = y,
(I5 ) (x → y) → y = (y → x) → x,
(I6 ) (x ∨ y) → z = (x → z) ∧ (y → z),
(I7 ) (x ∧ y) → z = (x → z) ∨ (y → z),
then (L, ∨, ∧, �, →, O, I) is a lattice implication algebra.

[email protected]
126 Decision Aid Models for Disaster Management and Emergencies

Table 6.1 Negation of H

h absolutely very exactly somewhat slightly
h� slightly somewhat exactly very absolutely

Table 6.2 Disjunction of H

∨ absolutely very exactly somewhat slightly
absolutely absolutely absolutely absolutely absolutely absolutely
very absolutely very very very very
exactly absolutely very exactly exactly exactly
somewhat absolutely very exactly somewhat somewhat
slightly absolutely very exactly somewhat slightly

Table 6.3 Conjunction of H

∧ absolutely very exactly somewhat slightly
absolutely absolutely very exactly somewhat slightly
very very very exactly somewhat slightly
exactly exactly exactly exactly somewhat slightly
somewhat somewhat somewhat somewhat somewhat slightly
slightly slightly slightly slightly slightly slightly

Example 6.5. We choose the five common linguistic hedges to construct the
linguistic values, i.e., absolutely, very, exactly, somewhat, slightly (denoted by H =
{absolutely, very, exactly, somewhat, slightly}), and the partially ordered relation on H as
slightly < somewhat < exactly < very < absolutely. According to the properties of LIA,
we can define the negation, disjunction, conjunction and implication of the linguistic
hedges in Tables 6.1–6.12.

Definition 6.2. Let H = { absolutely, very, exactly, somewhat, slightly } and C = {ci | c1 =
c�2 , c2 = c�1 , i = 1, 2}. We define L10 = (V, ∨, ∧,� , O, I) as follows, its operations “∨” and
“∧” are shown in the Hasse diagram of L10 in Fig. 6.3, (hi , c1 )� = (hi , c2 ), (hi , c2 )� = (hi , c1 )

[email protected]
A Linguistic-Valued Information Processing Method for Fuzzy Risk Analysis 127

Table 6.4 Implication of H

→ absolutely very exactly somewhat slightly
absolutely absolutely very exactly somewhat slightly
very absolutely absolutely very exactly somewhat
exactly absolutely absolutely absolutely exactly exactly
somewhat absolutely absolutely absolutely absolutely very
slightly absolutely absolutely absolutely absolutely absolutely

and its operation “→” is defined as follows:

(hi , c1 ) → (h j , c2 ) = hmax{0,i+ j−4}, c2

(hi , c2 ) → (h j , c1 ) = hmin{4,i+ j} , c1
(hi , c1 ) → (h j , c1 ) = hmin{4,4−i+ j} , c1
(hi , c2 ) → (h j , c2 ) = hmin{4,4− j+i} , c1
Then L10 = (V, ∨, ∧,� , O, I) is an LIA.

(absolutely, c1) 
 
 
(very, c1 )
)  
  (slightly, c2 )
(exactly, c1
)  
  (somewhat, c2)
(somewhat, c1  
(slightly, c1 )  
 (exactly, c2 )


 (very, c2 )
(absolutely, c2)
Fig. 6.3 Hasse Diagram of L10 .

In Definition 6.2, C is called the basic evaluation word set if it satisfies C = {ci | c1 = c�2 ,
c2 = c�1 , i = 1, 2}. Let V be a linguistic truth-value set. Every linguistic evaluation value
v ∈ V is composed of a linguistic hedge operator h and a basic word c, i.e., V = H × C
where the linguistic hedge operator set H is a totally ordered and finite set.
Note that there exist incomparable linguistic truth values (denoted by “�”) in L10 ,
e.g.,(exactly, c1 )�(somewhat, c2).

Proposition 6.1. For any (hi , c j ) ∈ L10 ,

(1) (absolutely, c2 ) → (hi , c j ) = (absolutely, c1 ),

(2) (absolutely, c1 ) → (hi , c j ) = (hi , c j ),
(3) (hi , c j ) → (absolutely, c2 ) = (hi , c j )� ,

[email protected]
128 Decision Aid Models for Disaster Management and Emergencies

(4) (hi , c j ) → (absolutely, c1 ) = (hi , c j ).

From the logical point of view, we obtain if the antecedent is false, the conclusion is
always true regardless what the consequent is.
If two linguistic hedges of linguistic values are equal or complement, we have the
following results.

Proposition 6.2. For any hi , h j ∈ H,

(1) If hi = h�j , then (hi , c1 ) → (h j , c2 ) = (h0 , c2 );

(2) If hi = h�j , then (hi , c2 ) → (h j , c1 ) = (hn , c1 );
(3) If hi = h j , then (hi , c1 ) → (h j , c1 ) = (hn , c1 );
(4) If hi = h j , then (hi , c2 ) → (h j , c2 ) = (hn , c1 ).

As a special case, we have:

(1) (absolutely, c1 ) → (absolutely, c1 ) = (absolutely, c1 );

(2) (absolutely, c1 ) → (absolutely, c2 ) = (absolutely, c2 );
(3) (absolutely, c2 ) → (absolutely, c1 ) = (absolutely, c1 );
(4) (absolutely, c2 ) → (absolutely, c2 ) = (absolutely, c1 ).

If (absolutely, c1 ) is “T” and (absolutely, c2 ) is “F”, we can obtain the classical logic
implication. Hence linguistic truth-valued logic is an extension of classical logic and fuzzy
logic.

6.3 Risk Analysis Based on Linguistic Aggregation

6.3.1 The unbalanced linguistic weighted geometric operator

To aggregate unbalanced linguistic values, we propose the unbalanced linguistic

weighted geometric operator in this section.

Definition 6.3 ([22]). An weighted geometric operator of dimension n is a mapping g :

(R+ )n → R+ that has associated with it a weighting vector W = (w1 , w2 , . . . , wn ), with
n n wi
wi ∈ [0, 1] and i=1 wi = 1, such that g(a1 , a2 , . . . , an ) = j=1 ai .

Definition 6.4 ([22]). Assume that the set of unbalanced linguistic values V = {si | i =
1, 2, . . . , n} be aggregated, in which si ∈ S = {s0 , s1 , . . . , sm } (m  n) is an unbalanced lin-
n
guistic value. A weighting vector is W = (w1 , w2 , . . . , wn ) with wi ∈ [0, 1] and i=1 wi = 1.

[email protected]
A Linguistic-Valued Information Processing Method for Fuzzy Risk Analysis 129

Then the unbalanced linguistic weighted geometric operator (the ULW G operator) is de-
fined as fULW G ({(wi , si ) | i = 1, . . . , n}) = fULW G ({(wi , T Ftt0i (L H (si ))) | i = 1, 2, . . . , n}) =
−1 n(t0 )
L H −1 (�t0 ( n ti
i=1 (�t0 (T Ft0 (L H (si ))))wi )) = L H −1 (sk , αk ), where ti is the level
n(t )
of L H (si ) in LH, t0 is a level of LH fixed by users, sk 0 ∈ Sn(t0 ) ⊂ LH and αk ∈
[−0.5, 0.5) such that
n
k + αk = (�t−1
0
(T Ftt0i (L H (si ))))wi . (6.2)
i=1

Remark 6.1. In Definition 6.4, T Ftt0i (L H (si )) means that unbalanced linguistic values
are represented at level t0 of LH and t0 is decided by users. In fact, by using T Ftt0i (·)
n(t0 )
and L H (·), S is converted in Sn(t0 ) of LH. By using L H −1 (sk , αk ), fULW G (V ) is
converted to the 2-tuple fuzzy linguistic representation of unbalanced linguistic value.

Example 6.6. Let S = {N, L, M, AH, H, QH,V H, AT, T } be a set of unbalanced lin-
guistic values. Suppose that {(0.25, AH), (0.35, QH), (0.4, H)} will be aggregated.
Here we select t0 = 3 and densitySr = extreme [26], hence, T F3ti (L H (AH)) =
T F33 (s95 ) = s95 , T F3ti (L H (QH)) = T F34 (s17 ti
13 ) = (s6 , 0.5), T F3 (L H (H)) = T F3 (s6 ) =
9 3 9
.
s96 , according to Eq. (6.2), we have k + αk = 50.25 × 6.50.35 × 60.4 = 5.93,
.
fULW G ({(0.25, AH), (0.35, QH), (0.4, H)}) = L H −1 (s96 , −0.07) = (H, −0.07).

In many cases, weight wi is a linguistic weight rather that number in [0, 1], e.g., in
fuzzy risk analysis, every severity of loss Li j of sub-component pi j (1  i  r, 1 
j  s) acts as the weight in aggregation, i.e., wi is a linguistic value in Eq. (6.2). We
use the following method to obtain numbers in [0, 1] from unbalanced linguistic val-
ues: 1) For any unbalanced linguistic values {l1 , . . . , lk } and LH = t l(t, n(t)), we have
t
{T Ftt01 (L H (l1 )), . . . , T Ft0k (L H (lk ))}, where ti is the level of L H (li ) (i ∈ {1, . . ., k}) in
LH and t0 is a level of LH fixed by users; 2) The function f : {l1 , . . . , lk } → [0, 1] is defined
by
�t−1 (T Ftt0i (L H (li )))
f (li ) = k
0
−1 ti
. (6.3)
j=1 �t0 (T Ft0 (L H (l j )))
�t−1 (T Ft i (L H (li )))
t
k
Evidently, for any li ∈ {l1 , . . . , lk }, f (li ) ∈ [0, 1] and i=1
0 0
= 1.
�−1 (T Ft i (L H (l j )))
k t
j=1 t0 0

Definition 6.5 ([22]). Assume that the set of unbalanced linguistic values V = {(li , si ) |
i = 1, . . . , n} is aggregated, and every li is a linguistic weight corresponding to si .
Then the ULW G operator with linguistic weights (the ULW GLW operator) is defined as

[email protected]
130 Decision Aid Models for Disaster Management and Emergencies

fULW GLW ({(li , si ) | i = 1, . . . , n}) = fULW G ({( f (li ), T Ftt0i (L H (si ))) | i = 1, 2, . . . , n}) =
n(t0 )
L H −1 (�t0 ( n −1 ti
i=1 (�t0 (T Ft0 (L H (si )))) f (li ) )) = L H −1 (sk , αk ), where ti is the
n(t )
level of L H (si ) in LH, t0 is a level of LH fixed by users, sk 0 ∈ Sn(t0 ) ⊂ LH and
αk ∈ [−0.5, 0.5) such that
n
k + αk = (�t−1
0
(T Ftt0i (L H (si )))) f (li ) . (6.4)
i=1

in which every f (li ) is decided by Eq. (6.3).

Proposition 6.3 ([22]). Let unbalanced linguistic values V = {(wi , si ) | i = 1, . . . , n} be

n
aggregated, where wi ∈ [0, 1] and i=1 wi = 1. If for any i ∈ {1, . . . , n}, wi = 1, then
fULW G ({(w1 , s1 ), . . . , (wn , sn )}) = si .

As special cases, we have

1) Denote s j = max{s1 , . . . , sn }, if w j = 1 then fULW G ((w1 , s1 ), . . . , (wn , sn )) = s j ;

2) Denote sk = min{s1 , . . . , sn }, if wk = 1 then fULW G ((w1 , s1 ), . . . , (wn , sn )) = sk ;
3) If w j = wk = 0 then fULW G reduces to the linguistic Olympic operator, i.e., the smallest
and largest linguistic values are deleted from linguistic evaluation values.

Proposition 6.4 ([22]). Let unbalanced linguistic values V = {(wi , si ) | i = 1, . . . , n} be

n
aggregated, where wi ∈ [0, 1] and i=1 wi = 1. The ULW G operator satisfies:

1) min{s1 , . . . , sn }  fULW G ({(w1 , s1 ), . . . , (wn , sn )})  max{s1 , . . . , sn };

2) fULW G is idempotent, i.e., fULW G ({(w1 , s1 ), . . . , (wn , sn )}) = s1 when s1 = · · · = sn ;
3) fULW G is monotone in relation to the input values si , i.e., for any s�i  si ,
fULW G ((w1 , s1 ), . . . , (wi , s�i ), . . . , (wn , sn ))  fULW G ((w1 , s1 ), . . . , (wi , si ), . . . , (wn , sn ));
4) fULW G is commutative;
5) fULW G reduces to the linguistic geometric mean if wi = n for all i = 1, . . . , n, i.e.,
1

−1
fULW G ({(w1 , s1 ), . . . , (wn , sn )}) = (sk , αk ) and k + αk = n n ti
i=1 �t0 (T Ft0 (L H (si ))).

6.3.2 Illustrative example based on the ULW G operator

In this subsection, we apply the ULW G operator to deal with fuzzy risk analysis prob-
lems. We firstly use an example to illustrate the fuzzy risk analysis process of our method,
where linguistic evaluation values are unbalanced linguistic evaluation values. In the ex-
ample, assume that the following unbalanced linguistic evaluation values are considered:

[email protected]
A Linguistic-Valued Information Processing Method for Fuzzy Risk Analysis 131

(1) Unbalanced linguistic evaluation values for severity of loss: L = {none (N), low (L),
medium (M), almost high (AH), high (H), quite high (QH), very high (V H), almost
total (AT ), total (T )};
(2) Unbalanced linguistic evaluation values for probability of failure: F = {none (N),
small (S), medium (M), almost big (AB), big(B), quite big (QB), very big (V B), almost
total (AT ), total (T )}. The proposed fuzzy risk analysis algorithm is now presented
as follows: 1) Unbalanced linguistic evaluation values for severity of loss and proba-
bility of failure are represented in the level t0 of LH = t l(t, n(t)), where t0 is fixed
by decision makers; 2) According to transformation function (6.1), the transformed
unbalanced linguistic evaluation values are 2-tuple fuzzy linguistic representations;
(3) Aggregate all linguistic evaluation values {(Li j , Fi j ) | 1  j  s} of component Pi by
the ULW G operator with linguistic weights, i.e.,

Fi = fULW GLW ({(Li j , Fi j ) | 1  j  s})

= fULW G ({( f (Li j ), T Ftt0i (L H (Fi j ))) | 1  j  s})
n
= L H −1 �t0 (�t−1
0
(T Ftt0i (L H (Fi j )))) f (Li j )
i=1
−1 n(t0 )
=LH (sk , αk ),
n
k + αk = (�t−1
0
(T Ftt0i (L H (Fi j )))) f (Li j ) ,
i=1
�t−1 (T Ftt0i (L H (Li j )))
f (Li j ) = k
0
−1 ti
;
j=1 �t0 (T Ft0 (L H (Li j )))
(4) Rank linguistic evaluation values {Fi | 1  i  r} by the ranking method of linguistic
n(t ) n(t )
values, i.e., Fi1  Fi2 if and only if �t−1
0
(sk1 0 , αk1 )  �t−1
0
(sk2 0 , αk2 ). The larger the
linguistic evaluation value Fi of Pi is, the higher the risk of the manufactory Mi is.

Example 6.7. Assume that there are three manufactories M1 , M2 and M3 producing the
components P1 , P2 and P3 , respectively, where P1 , P2 and P3 are the same product made
by different manufactories. Each component Pi consists of three sub-components Pi1 , Pi2
and Pi3 , where 1  i  3. Assume that there are two evaluation items Li j and Fi j to derive
the probability of failure Fi of component Pi made by manufactory Mi , where Li j denotes
the severity of loss of the sub-component Pi j , Fi j denotes the probability of failure of the
sub-component Pi j , 1  i  3 and 1  j  3. The linguistic evaluation values of every
sub-component Pi j are shown in Table 6.5.

[email protected]
132 Decision Aid Models for Disaster Management and Emergencies

Table 6.5 Linguistic evaluation values of the sub-components made by manufactories

manufactory sub-components the severity of loss the probability of failure
P11 L AT
M1 P12 QH S
P13 AH M
P21 M QB
M2 P22 H M
P23 N T
P31 L VB
M3 P32 M B
P33 AT S

In the example, the linguistic hierarchies be LH = l(1, 3) ∪ l(2, 5) ∪ l(3, 9) ∪ l(4, 17) for
the severity of loss and the probability of failure of sub-components, the levels t0 of the
severity of loss and the probability of failure are fixed by 2 and 3, respectively. Transfor-
mations of linguistic evaluation values are shown in Table 6.6. We can obtain every Fi of
component Pi according to Eq. (6.3), Eq. (6.4) and Table 6.6, respectively.

Table 6.6 Transformations of linguistic evaluation values

manufactory sub-components Li j T F2ti (L H (Li j )) Fi j T F3ti (L H (Fi j ))
P11 L (s51 , 0) AT (s98 , −0.5)
M1 P12 QH (s53 , 0.25) S (s92 , 0)
P13 AH (s53 , −0.5) M (s94 , 0)
P21 M (s52 , 0) QB (s97 , −0.5)
M2 P22 H (s53 , 0) M (s94 , 0)
P23 N (s50 , 0) T (s98 , 0)
P31 L (s51 , 0) VB (s97 , 0)
M3 P32 M (s52 , 0) B (s96 , 0)
P33 AT (s54 , −0.25) S (s92 , 0)

[email protected]
A Linguistic-Valued Information Processing Method for Fuzzy Risk Analysis 133

For component P1 with sub-components P11 , P12 and P13 , we have

�−1
2 (s1 , 0)
5
.
f (L) = = 0.15,
�2 (s1 , 0) + �2 (s3 , 0.25) + �−1
−1 5 −1 5
2 (s3 , −0.5)
5

�−1
2 (s3 , 0.25)
5
.
f (QH) = = 0.48,
�−1 −1 5 −1 5
2 (s1 , 0) + �2 (s3 , 0.25) + �2 (s3 , −0.5)
5

−1 5
�2 (s3 , −0.5) .
f (AH) = = 0.37,
�2 (s1 , 0) + �−1
−1 5 −1 5
2 (s3 , 0.25) + �2 (s3 , −0.5)
5

F1 = fULW GLW ({(L, AT ), (QH, S), (AH, M)})

= fULW G ({( f (L), (s98 , −0.5)), ( f (QH), (s92 , 0)), ( f (AH), (s94 , 0))})
= L H −1 (�3 ((�−1
3 (s8 , −0.5))
9 f (L)
× (�−1
3 (s2 , 0))
9 f (QH)
× (�−1
3 (s4 , 0))
9 f (AH)
))
= L H −1 (�3 (7.50.15 × 20.48 × 40.37))
.
= L H −1 (�3 (3.13)) = (M, −0.43).

Table 6.7 The probability of failure of the component made by manufactory

manufactory the component the probability of failure
M1 P1 F1 = (M, −0.43)
M2 P2 F2 = (AH, −0.15)
M3 P3 F3 = (M, −0.325)

According to Table 6.7, the probability of failure F2 of the component P2 made by

manufactory M2 is (AH, −0.15), i.e., almost big with the value of the symbolic translation
−0.15, it is the largest linguistic evaluation value among the linguistic evaluation values of
F1 , F2 and F3 . Hence the risk of the manufactory M2 is the highest.

Example 6.8 ([6]). Assume that the balanced linguistic values are { Absolutely -low, Very-
low, Low, Fairly-low, Medium, Fairly-high, High, Very-high, Absolutely-high }. Their cor-
responding interval valued fuzzy numbers are shown in Table 6.8. The linguistic evaluation
values of every sub-component Pi j are shown in Table 6.9. In Table 6.9, wi j denotes the
degree of confidence of the decision-maker’s opinion with respect to sub-components Pi j .
The method proposed in [6] is presented as follows:

(1) Aggregate the linguistic evaluation values of sub-components Pi j of each component Pi

made by manufactory Mi based on the fuzzy weighted mean method and the interval-
valued fuzzy numbers arithmetic operators proposed in [6] to get the probability of

[email protected]
134 Decision Aid Models for Disaster Management and Emergencies

failure Fi , e.g., for F1 , according to Table 6.8 and Table 6.9, we have
(low ⊗ f airly low) ⊕ ( f airly high ⊗ medium) ⊕ (very low ⊗ f airly high)
F1 =
low ⊕ f airly high ⊕ very low
(A3 ⊗ A4 ) ⊕ (A6 ⊗ A5) ⊕ (A2 ⊗ A6 ) (A∗3 ⊗ A∗4 ) ⊕ (A∗6 ⊗ A∗5) ⊕ (A∗2 ⊗ A∗6 )
= =
A3 ⊕ A6 ⊕ A2 A∗3 ⊕ A∗6 ⊕ A∗2
= [(0.240, 0.328, 0.656, 0.905; 0.288), (0.195, 0.306, 0.678, 0.949; 0.763)],

where A∗2 , A∗3 , A∗4 , A∗5 and A∗6 are type-1 fuzzy numbers of A2 , A3 , A4 , A5 and A6 ,
respectively, e.g., A∗2 = , , ,
0.0075+0 0.0075+0 0.015+0.02 0.0525+0.07 0.5+1
2 2 2 2 ; 2 = (0.00375,
∗ ∗
0.00375, 0.0175, 0.06125; 0.75), A2 ⊗ A6 = (0.00375 × 0.615, 0.00375 × 0.65125,
0.0175 × 0.77875, 0.06125 × 0.825; min(0.75, 0.75)), A∗2 ⊕ A∗6 = (0.00375 ⊕ 0.615,
0.00375 ⊕ 0.65125, 0.0175 ⊕ 0.77875, 0.06125 ⊕ 0.825; min(0.75, 0.75));
(2) Calculate the degree of similarity between the upper fuzzy numbers of the interval-
valued fuzzy numbers Fi and every linguistic value shown in Table 6.8, respectively,
e.g., for F1 and A1 , we have

X (F1 , A1 ) = SX ((0.195, 0.306, 0.678, 0.949; 0.763), (0, 0, 0, 0; 1))

SU U U U

|0.195 − 0| + |0.306 − 0| + |0.678 − 0| + |0.949 − 0|

= 1− = 0.468;
4
(3) Calculate the spread between the upper fuzzy numbers of the interval valued fuzzy
numbers Fi and every linguistic value shown in Table 6.8, respectively, e.g., for F1 and
A1 , we have

ST DU (F1U , AU1 ) = |ST DFU − ST DAU | = 0.347,

1 1
0.195 + 0.306 + 0.678 + 0.949
x1 = = 0.532,
4
(0.195 − x1)2 + (0.306 − x1)2 + (0.678 − x1)2 + (0.949 − x1)2
ST DF U = ;
1 4−1
(4) Calculate the degree of similarity on the X-axis between the interval-valued fuzzy num-
bers Fi and every linguistic value shown in Table 6.8, respectively, e.g., for F1 and A1 ,
we have SX (F1 , A1 ) = 1 − (|(0.195 − 0.24) − (0 − 0)| + |(0.306 − 0.328) − (0 − 0)| +
|(0.678 − 0.656) − (0 − 0)| + |(0.949 − 0.905) − (0 − 0)|)/4 = 0.96675;
(5) Calculate the degree of similarity on the Y -axis between the interval-valued fuzzy num-
bers Fi and every linguistic value shown in Table 6.8, respectively, e.g., for F1 and A1 ,
we have SY (F1 , A1 ) = 1 − |yF1 − yA1 | = 0.915, where yF1 and yA1 are associated to areas
of F1U , F1L , AU1 and AL1 , respectively;

[email protected]
A Linguistic-Valued Information Processing Method for Fuzzy Risk Analysis 135

(6) Calculate the degree of similarity between the interval-valued fuzzy numbers Fi and
every linguistic value shown in Table 6.8, respectively, which is decided by the above
mentioned degrees of similarity and the spread, e.g., for F1 and A1 , we have
X (F1 , A1 ) × (1 − |0.763 − 1|)
SU U U
S(F1 , A1 ) = × SX (F1 , A1 ) × SY (F1 , A1 ) = 0.234.
1 + STDU (F1U , AU1 )
By using the method, the probability of failure of P1 made by manufactory M1 is
“Medium”, the probability of failure of P2 is “fairly-high” and the probability of failure
of P3 is “fairly-high”. Hence, the risk of manufactories M2 and M3 is the highest, see
[6] for more detail.
Table 6.8 Linguistic values and their corresponding interval-valued fuzzy numbers

Linguistic values Interval-valued fuzzy numbers

Absolutely-low (s90 ) A1 = [(0, 0, 0, 0; 1), (0, 0, 0, 0; 1)]
Very-low (s91 ) A2 = [(0.0075, 0.0075, 0.015, 0.0525; 0.5), (0, 0, 0.02, 0.07; 1)]
Low (s92 ) A3 = [(0.0875, 0.12, 0.16, 0.1825; 0.5), (0.04, 0.1, 0.18, 0.23; 1)]
Fairly-low (s93 ) A4 = [(0.2325, 0.255, 0.325, 0.3575; 0.5), (0.17, 0.22, 0.36, 0.42; 1)]
Medium (s94 ) A5 = [(0.4025, 0.4525, 0.5375, 0.5675; 0.5), (0.32, 0.41, 0.58, 0.65; 1)]
Fairly-high (s95 ) A6 = [(0.65, 0.6725, 0.7575, 0.79; 0.5), (0.58, 0.63, 0.8, 0.86; 1)]
High (s96 ) A7 = [(0.7825, 0.815, 0.885, 0.9075; 0.5), (0.72, 0.78, 0.92, 0.97; 1)]
Very-high (s97 ) A8 = [(0.9475, 0.985, 0.9925, 0.9925; 0.5), (0.93, 0.98, 1, 1; 1)]
Absolutely-high (s98 ) A1 = [(1, 1, 1, 1; 1), (1, 1, 1, 1; 1)]

Table 6.9 Linguistic evaluation values of the sub-components made by manufactories

manufactory sub-components the severity of loss the probability of failure
P11 low fairly-low (w11 = 0.9)
M1 P12 fairly-high medium (w12 = 0.7)
P13 very-low fairly-high (w13 = 0.8)
P21 low very-high (w21 = 0.85)
M2 P22 fairly-high fairly-high (w22 = 0.9)
P23 very-low medium (w23 = 0.9)
P31 low fairly-low (w31 = 0.95)
M3 P32 fairly-high high (w32 = 0.8)
P33 very-low fairly-high (w33 = 1.0)

[email protected]
136 Decision Aid Models for Disaster Management and Emergencies

In the following, we use our method to get the probability of failure Fi made by manu-
factory Mi (i = 1, 2, 3). In this example, because linguistic evaluation values are balanced,
transformation function T Ftt� : l(t, n(t)) −→ l(t � , n(t � )) are unnecessary, or transformation
function T Ftt� is such that t = t � = 3.
According to (6.3), (6.4) and Table 6.9, we have

F1 = fULW G ({(low, f airly low), ( f airly high, medium), (very low, f airly high)})
= fULW G ({(s92 , s93 ), (s95 , s94 ), (s91 , s95 )})
f (s5 )
= L H −1 (�3 ((�−1 f (s2 )
× (�−1 × (�−1 f (s1 )
9 9 9
3 (s3 , 0)) 3 (s4 , 0)) 3 (s5 , 0)) )),
9 9 9

�−1 (s92 )
in which f (s92 ) = 3
= 2
= 0.25, f (s95 ) = 5
= 0.625
�−1 (s92 )+�−1 (s95 )+�−1 (s91 ) 2+5+1 2+5+1
3 3 3
−1 .
and f (s91 ) = 1
2+5+1 = 0.125. Hence F1 = L H (�t0 (30.25 × 40.625 × 50.125 ))
=
−1 −1
LH (�t0 (3.83) = L H (s4 , −0.17) = (Medium, −0.17).
9

Similarly, F2 and F3 can be obtained as

F2 = fULW G ({(s92 , s97 ), (s95 , s95 ), (s91 , s94 )})

f (s5 )
= L H −1 ((�−1 f (s2 )
× (�−1 × (�−1 f (s1 )
9 9 9
3 (s7 , 0)) 3 (s5 , 0)) 3 (s4 , 0)) )
9 9 9

. −1 9
= L H (s5 , 0.3) = (Fairly − high, 0.3),

F3 = fULW G ({(s92 , s93 ), (s95 , s96 ), (s91 , s95 )})

f (s5 )
= L H −1 ((�−1 f (s2 )
× (�−1 × (�−1 f (s1 )
9 9 9
3 (s3 , 0)) 3 (s6 , 0)) 3 (s5 , 0)) )
9 9 9

.
= L H −1 (s95 , −0.07) = (Fairly − high, −0.07).

Accordingly, the probability of failure of P1 made by manufactory M1 is “(Medium,-0.17)”,

the probability of failure of P2 is “(Fairly-high,0.3)" and the probability of failure of P3 is
“(Fairly-high,-0.07)”. Hence the risk of the manufactory M2 is the highest.
Compared our method with the method proposed in [6], because interval valued fuzzy
numbers are used to represent evaluation values, computation of the method is complex and
the result of the method is the lack of precision, i.e., P2 and P3 have the same probability of
failure “fairly-high”. This is loss of information due to the degree of similarity between the
interval-valued fuzzy numbers. However, the result of our method is that the probability of
failure of P2 is “(Fairly-high, 0.3)” more than “(Fairly-high, -0.07)” of P3 . There is no loss
of information due to 2-tuple fuzzy linguistic representation and no complex computation
due to linguistic aggregation operator.

[email protected]
A Linguistic-Valued Information Processing Method for Fuzzy Risk Analysis 137

Table 6.10 The evaluations of expert 1

Risk types Expert 1
SR ((exactly, small), (exactly, low))
CR ((exactly, small), (exactly, low))
QR ((absolutely, small), (absolutely, low))
PR ((very, small), (very, low))
RR ((exactly, big), (exactly, high))

Table 6.11 The evaluations of expert 2

Risk types Expert 2
SR ((very, small), (very, low))
CR ((very, small), (very, low))
QR ((slightly, small), (slightly, low))
PR ((exactly, small), (exactly, low))
RR ((very, small), (very, low))

Table 6.12 The evaluations of expert 3

Risk types Expert 3
SR ((slightly, small), (slightly, low))
CR ((slightly, small), (slightly, low))
QR ((somewhat, small), (somewhat, low))
PR ((slightly, big), (slightly, low))
RR ((exactly, small), (exactly, low))

6.4 Risk Analysis Based on L10

In a project management, risk management is important. For the best benefit from the
project invest, the manager must do the risk analysis. In the project risk analysis, severity
of loss is expressed by the linguistic evaluation values, e.g., very low, fairly low, pretty low,
medium, almost high, more or less high, very high, and absolutely high. Furthermore, the
probability of failure is also used in fuzzy risk analysis.
Assume that there are r components P1 , P2 , . . . and Pr made by r manufactories
M1 , M2 , . . . and Mr , respectively. Each component Pi consists of s sub-components

[email protected]
138 Decision Aid Models for Disaster Management and Emergencies

pi1 , pi2 , . . . and pis . Each sub-components pi j (1  i  r, 1  j  s) is evaluated by

the severity of loss Li j and a probability of failure Fi j , in which Li j and Fi j come from
the linguistic evaluation values, respectively. Now we will give a software project risk
analysis model based on L10 , in which the basic evaluation word sets are selected by
{c1 = high, c2 = low} and {p1 = big, p2 = small}, respectively. The linguistic hedge set is
H = { absolutely, very, exactly, somewhat, slightly } which is accepted by all experts.

(1) The linguistic evaluation values for severity of loss: L = { absolutely low, very low,
exactly low, somewhat low, slightly low, slightly high, somewhat high, exactly high,
very high, absolutely high }, which can be constructed by L10 of severity of loss;
(2) The linguistic evaluation values for probability of failure: F = { absolutely small, very
small, exactly small, somewhat small, slightly small, slightly big, somewhat big, ex-
actly big, very big, absolutely big }, which can be constructed by L10 of probability of
failure.

component Pi

sub-component pi1 sub-component pi2 ··· sub-component pis

Fig. 6.4 The structure of fuzzy risk analysis

The structure of fuzzy risk analysis is shown in Fig. 6.4, in which Pi is evaluated
by Li (severity of loss) and Fi (probability of failure), and every sub-component pi j
( j = 1, 2, . . . , s) is evaluated by Li j and Fi j . Formally, we can rewrite the result of fuzzy
risk analysis as linguistic evaluation of Li and linguistic evaluation of Fi . From the alge-
braic point of view, the result of fuzzy risk analysis (linguistic evaluation of Li , linguistic
evaluation of Fi ) is embedded in L × F = {((hi , c j ), (hk , pl )) | (hi , c j ) ∈ L, (hk , pl ) ∈ F},
because L and F are all lattice implication algebras L10 , and L × F is a L10 × L10 lat-
tice implication algebra. On the other hand, there exist associable relationships among
sub-components of fuzzy risk analysis, e.g., sub-component pi1 affects sub-component
pi2 , i.e., pi1 → pi2 . Naturally, the corresponding results of pi1 and pi2 are translated by

[email protected]
A Linguistic-Valued Information Processing Method for Fuzzy Risk Analysis 139

Table 6.13 The aggregation results

Risk types Experts
SR ((very, small), (very, low))
CR ((very, small), (very, low))
QR ((absolutely, small), (absolutely, low))
PR ((very, small), (very, low))
RR ((very, small), (very, low))

((hipi1 , c pj i1 ), (hkpi1 , plpi1 )) → ((hipi2 , c pj i2 ), (hkpi2 , plpi2 )) in L × F. Hence, we obtain the result
of fuzzy risk analysis by the following three steps:

(1) Integrate experts’ results of each sub-component Pi . From the algebraic point of view,
every binary operator can be used as an aggregation to integrate experts’ results, e.g.,
∨ or ∧;
(2) Consider associable relationships among sub-components of fuzzy risk analysis. In the
lattice implication algebra, every associable relationship is translated by implication
formula;
(3) Integrate all results of steps 1 and 2. In the lattice implication algebra, we select ∨ or
∧ to aggregate all results.

Example 6.9. In a repair information analysis system, there exist some risk because cus-
tomers cannot explain their requirement exactly or some new technology is applied. This
project risk may be schedule risk (SR), cost risk (CR), quality risk (QR), personnel risk
(PR) and requirement risk (RR), in which personnel risk affects schedule risk, require-
ment risk affects quality risk. To analysis the project risk, the manager asked three experts
(requirement analysis expert, quality analysis expert and project management expert) to
assess these five kinds of risk, respectively. The evaluation results of experts are shown in
Table 6.10. For example, requirement analysis expert gives the severity of loss which is
“exactly small” and the probability of failure which is “exactly low” for schedule risk.
According to the steps of fuzzy risk analysis, we aggregate the evaluation results of
experts, e.g., for SR, we use operator ∨ and obtain ((exactly, small), (exactly, low)) ∨
((very, small), (very, low)) ∨ ((slightly, small), (slightly, low)) = ((exactly, small) ∨
(very, small) ∨ (slightly, small), (exactly, low) ∨ (very, low) ∨ (slightly, low)) = ((very,
small), (very, low)), others are shown in Table 6.13.

[email protected]
140 Decision Aid Models for Disaster Management and Emergencies

Table 6.14 The transformed results

Risk types Experts
PR → SR ((absolutely, big), (absolutely, high))
CR ((very, small), (very, low))
RR → QR ((very, big), (very, high))

We consider associable relationships “personnel risk affects schedule risk” and “re-
quirement risk affects quality risk”, i.e., ((very, small), (very, low)) → ((very, small),
(very, low)) = ((very, small) → (very, small), (very, low) → (very, low)) = ((absolutely,
big), (absolutely, high)) and ((very, small), (very, low)) → ((absolutely, small),
(absolutely, low)) = ((very, small) → (absolutely, small), (very, low) → (absolutely,
low)) = ((very, big), (very, high)). Formally, the fuzzy risk analysis is transformed by
Table 6.14.
Based on Table 6.14, we finally obtain the result of the fuzzy risk analysis as
((absolutely, big), (absolutely, high)) ∧ ((very, small), (very, low)) ∧ ((very, big),
(very, high)) = ((absolutely, big) ∧ (very, small) ∧ (very, big), (absolutely, high) ∧ (very,
low) ∧ (very, high)) = ((very, small), (very, low)).

6.5 Conclusion

In practice, a main characteristic of fuzzy risk analysis is the use of linguistic values
to represent the risk value instead of numerical values, the result of fuzzy risk analysis is
also expressed by linguistic values in natural language. Inspired by the weighted geometric
operator as a typical aggregation operator and lattice implication algebra as a logical alge-
bra for the reasoning purpose, we have presented two methods for fuzzy risk analysis with
linguistic evaluation values. In the first method, the unbalanced linguistic weighted geo-
metric operator has been used to deal with aggregation of unbalanced linguistic evaluation
values with numerical weights as well as linguistic weights. By the linguistic aggregation
operator, we can directly represent the final linguistic evaluation values and without loss of
information. In the second method, we have chosen five common linguistic hedges in risk
analysis and the basic words including two antonyms. The evaluation linguistic values have
been composed of linguistic hedges and basic words. Using the lattice implication algebra
approach to the nature structure of domains of linguistic values is feasible to handle both
comparable and incomparable evaluation linguistic values of fuzzy risk analysis. Some

[email protected]
Bibliography 141

examples were also provided to illustrate the proposed approaches. These approaches will
provide some methodology support for fuzzy risk analysis.

Acknowledgments

This work is partly supported by National Nature Science Foundation of China

(Grant No.61105059,61175055,61173100), China Postdoctoral Science Foundation
(2012M510815), Liaoning Excellent Talents in University (LJQ2011116), Sichuan Key
Laboratory of Intelligent Network Information Processing (SGXZD1002-10) and Key Lab-
oratory of Radio Signals Intelligent Processing (Xihua University) (XZD0818-09).

Bibliography

[1] Kirchsteiger, C. (1999). On the use of probabilistic and deterministic methods in risk analysis,
Journal of Loss Prevention in the Process Industries 12, pp. 399–419.
[2] Rosqvist, T. (2010). On the validation of risk analysis, Reliability Engineering and System
Safety 95, pp. 1261–1265.
[3] Weber, P., Medina-Oliva, G., Simon, C. and Iung, B. (2010). Overview on Bayesian networks
applications for dependability, risk analysis and maintenance areas, Engineering Applications
of Artificial Intelligence, (in press, doi:10.1016/j.engappai.2010.06.002).
[4] Chen, S. M. and Wang, C. H. (2009). Fuzzy risk analysis based on ranking fuzzy numbers
using α -cuts, belief features and signal/noise ratios, Expert Systems with Applications 36, pp.
5576–5581.
[5] Chen, S. M. (1996). New methods for subjective mental workload assessment and fuzzy risk
analysis, Cybernetics and Systems 27, 5, pp. 449–472.
[6] Chen, S. M. and Chen, J. H. (2009). Fuzzy risk analysis based on similarity measures between
interval-valued fuzzy numbers and interval-valued fuzzy number arithmetic operators, Expert
Systems with Applications 36, pp. 6309–6317.
[7] Chen, S. J. and Chen, S. M. (2003). Fuzzy risk analysis based on similarity of generalized fuzzy
numbers, IEEE Transactions on Fuzzy Systems 11, 1, pp. 45–56.
[8] Chen, S. J. and Chen, S. M. (2007). Fuzzy risk analysis based on the ranking of generalized
trapezoidal fuzzy numbers, Applied Intelligence 26, 1, pp. 1–11.
[9] Chen, S. J. and Chen, S. M. (2008). Fuzzy risk analysis based on measures of similarity between
interval-valued fuzzy numbers, Computers and Mathematics with Applications 55, 1, pp. 1670–
1685.
[10] Tang, T. C. and Chi, L. C. (2005). Predicting multilateral trade credit risks: Comparisons of
Logit and fuzzy logic models using ROC curve analysis, Expert Systems with Applications 28,
3, pp. 547–556.
[11] Wang, Y. M. and Elhag, T. M. (2006). Fuzzy TOPSIS method based on alpha level sets with an
application to bridge risk assessment, Expert Systems with Applications 31, 2, pp. 309–319.
[12] Yong, D. and Qi, L. (2005). A TOPSIS-based centroid-index ranking method of fuzzy numbers
and its application in decision-making, Cybernetics and Systems 36, 6, pp. 581–595.
[13] Lee, L. and Chen, S. M., (2008). Fuzzy risk analysis based on fuzzy numbers with different
shapes, Expert Systems with Applications 34, pp. 2763–2771.

[email protected]
142 Decision Aid Models for Disaster Management and Emergencies

[14] Pei, Z., Liu, X. and Zou, L. (2010). Extracting Association Rules Based on Intuitionistic Fuzzy
Sets, International Journal of Innovative Computing, Information and Control 6, 6, pp. 2567–
2580.
[15] Pei, Z. (2009). Fuzzy risk analysis based on linguistic information fusion, ICIC Express Letters
3, 3A, pp. 325–330.
[16] Sun, F., Li, B., Zou, L. and Pei, Z. (2011). Unbalanced linguistic information in fuzzy risk
analysis, ICIC Express Letters 5, 5, pp. 1661–1666.
[17] Pei, Z., Ruan, D., Liu, J. and Xu, Y. (2009). Linguistic Values based Intelligent Information
Processing: Theory, Methods, and Applications, Atlantis Computational Intelligence Systems,
Vol. 1 (Atlantis Press/ World Scientific).
[18] Pei, Z. and Yi, L. (2006). A new aggregation operator of linguistic information and its proper-
ties, 2006 IEEE International Conference on Granular Computing, pp. 486–489.
[19] Pei, Z., Ruan, D., Xu, Y. and Liu,J. (2007). Gathering linguistic information in distributed
intelligent agent on the internet, Internat.J. Intell. Systems 22, pp. 435–453.
[20] Meng, D. and Pei, Z. (2010). The linguistic computational models based on index of linguistic
label, The Journal of Fuzzy Mathematics 18, 1, pp.9–20.
[21] Pei, Z., Ruan, D., Liu, J., Xu, Y. (2012). A linguistic aggregation operator with
three kinds of weights for nuclear safeguards evaluation, Knowledge-Based Systems
(doi:10.1016/j.knosys.2011.10.016).
[22] Pei, Z. and Shi, P. (2011). Fuzzy risk analysis based on linguistic aggregation operators, Inter-
national Journal of Innovative Computing, Information and Control 7, 12, pp. 7105–7118.
[23] Okuhara, K., Yeh, K. Y., Shibata, J., Ishii,H. and Hsia, H. C. (2008). Evaluation and assignment
of traffic volume for urban planning based on planner and user stances, International Journal
of Innovative Computing Information and Control 4, 5, pp. 1151–1160.
[24] Mazurov, A. and Pakshin, P. (2009). Stochastic dissipativity with risk-sensitive storage function
and related control problems, ICIC Express Letters 3, 1, pp. 53–60.
[25] Herrera, F. and Martínez, L. (2000). A 2-tuple fuzzy linguistic representation model for com-
puting with words, IEEE Transactions On Fuzzy Systems 8, 6, pp. 746–752.
[26] Herrera, F., Herrera-Viedma, E. and Martinez, L. (2008). A fuzzy linguistic methodology to
deal with unbalanced linguistic term sets, IEEE Transactions On Fuzzy Systems 16, 2, pp. 354–
370.
[27] Herrera, F. and Alonso, S. et al, (2009). Computing With Words in Decision Making: Founda-
tions, Trends and Prospects, Fuzzy Optimization and Decision Making 8, 4, pp. 337–364.
[28] Herrera, F. and Martinez, L. (2001). A model based on linguistic 2-tuples for dealing with
multi-granularity hierarchical linguistic contexts in multi-expert decision-making, IEEE Trans.
Syst., Man Cybern. B, Cybern. 31, 2, pp. 227–234.
[29] Cordón, O., Herrera, F. and Zwir, I. (2001). Linguistic modeling by hierarchical systems of
linguistic rules, IEEE Transactions On Fuzzy Systems 10, 1, pp. 2–20.
[30] Mendel, J.M. and Zadeh, L.A. et al., (2010). What Computing with Words means to me, IEEE
Computational Intelligence Magazine 5, 1, pp. 20–26.
[31] Ho, N. and Wechler, W. (1990). Hedge algebras: an algebraic approach to structure of sets of
linguistic turth values, Fuzzy Sets and Systems 35, pp. 281–293.
[32] Ho, N. and Wechler, W. (1992). Extended hedge algebras and their application to fuzzy logic,
Fuzzy Sets and Systems, 52, pp. 259–281.
[33] Ho, N. C. and Long, N. V. (2007). Fuzziness measure on complete hedge algebras and quanti-
fying semantics of terms in linear hedge algebras, Fuzzy Sets and Systems 158, pp. 452–471.
[34] Huang, C. and Ruan, D. (2008). Fuzzy risks and an updating algorithm with new observations,
Risk Analysis 28, 3, pp. 681–694.
[35] Kangari, R. and Riggs, L.S. (1989). Construction risk assessment by linguistics, IEEE Trans-
actions on Engineering Management 36, 2, pp. 126–131.

[email protected]
Bibliography 143

[36] Lawry, J. (2004). A framework for linguistic modeling, Artificial Intelligence 155, pp. 1–39.
[37] Shinkai, K. (2008). Decision analysis of fuzzy partition tree applying fuzzy theory, Interna-
tional Journal of Innovative Computing Information and Control 4, 10, pp. 2581–2594.
[38] Xu, Y., Ruan, D., Qin, K. Y. and Liu, J. (2004). Lattice-Valued Logic, Springer.
[39] Xu, Y., Ruan, D., Kerre, E.E., and Liu, J. (2000). α -resolution principle based on lattice-valued
propositional logic LP(X), Information Sciences 130, pp. 195–223.
[40] Xu, Y., Liu, J., Ruan, D. and Lee, T.T. (2006). On the consistency of rule bases based on lattice-
valued first-order logic LF(X), International Journal of Intelligent Systems, 21, pp. 399–424.
[41] Xu, Y., Chen, S.W. and Ma, J. (2006). Linguistic truth-valued lattice implication algebra and
its Properties, Proc. IMACS Multiconference on “Computational Engineering in Systems Ap-
plications” (CESA2006), pp. 1413–1418.
[42] Zadeh, L.A. (1975). The concept of linguistic variable and its application to approximate rea-
soning, Parts 1, 2, 3, Information Sciences 8, 9, pp. 199–249, pp. 301–357, pp. 43–80.
[43] Zadeh, L.A. (1996). Fuzzy logic = computing with words, IEEE Transactions on Fuzzy Systems
4, 2, pp. 103–111.
[44] Zou, L., Ruan, D., Pei, Z., and Xu, Y. (2011). A Linguistic-Valued Lattice Implication Algebra
Approach for Risk Analysis. Journal of Multi-Valued Logic and Soft Computing 17, 293-303.
[45] Zou, L., Ruan, D., Pei, Z. and Xu, Y. (2008). A linguistic truth-valued reasoning approach
in decision making with incomparable information, Journal of Intelligent and Fuzzy Systems
19,4-5, pp. 335-343.
[46] Zou, L., Liu, X., Wu, Z. and Xu, Y. (2008). A uniform approach of linguistic truth values in
sensor evaluation, Fuzzy Optimization and Decision Making 7, 4, pp. 387–397.
[47] Zou, L., Pei, Z., Liu, X. and Xu, Y. (2009). Semantic of Linguistic Truth-valued Intuitionistic
Fuzzy Proposition Calculus, International Journal of Innovative Computing, Information and
Control 5, 12(A), pp. 4745–4752.

[email protected]
Chapter 7

A Belief Rule-Based Generic Risk Assessment

Framework

Jun Liu 1 , Shuwei Chen 1 , Luis Martinez 2 , and Hui Wang 1

1 School of Computing and Mathematics, University of Ulster, Northern Ireland, UK
2 Department of Computer Science, University of Jaen, Spain

E-mail: [email protected]

The present work aims to propose a generic risk assessment framework for modeling, an-
alyzing and synthesizing risk-related information with various uncertainties. Such infor-
mation may be of very different nature: it can be quantitative or qualitative, and uncertain,
incomplete, imprecise, conflicting, for which the traditional quantitative approach (e.g., sta-
tistical approach) does not give an adequate answer. IF-THEN rules place the key role in
this framework which is then combined with Dempster-Shafer theory of evidence and de-
cision theory in order to form a generic IF-THEN belief rule based risk assessment frame-
work with the aim to address different types of information. The general risk assessment
specification for the risk assessment of a hazard event is reviewed firstly. The belief-rule-
based risk assessment framework is then illustrated and outlined including rule-base rep-
resentation, inference procedure, rule-base generation and multi-source synthesis aspects.
This is then followed by the illustrative analysis of hierarchical structure nature of the real
world risk assessment model. How can the above discussed risk assessment framework
be incorporated to form a multi-layer risk assessment model is finally specified. In the
proposed generic framework, various types of information from different sources can be
transformed and used in the inference process. It provides a flexible and effective way to
represent and a rigorous procedure to deal with hybrid uncertain assessment information to
arrive at rational conclusions.

7.1 Introduction

Risk management (e.g. disaster management), which encompasses monitoring, predict-

ing, preventing, preparing for, responding to, mitigating and recovering from unexpected
hazards (e.g., disasters), aims to deal with any potential and actual hazard (e.g., disaster) by
effective and efficient organization, communication, interaction and utilization of counter

B. Vitoriano et al. (eds.), Decision Aid Models for Disaster Management 145
and Emergencies, Atlantis Computational Intelligence Systems 7,
DOI: 10.2991/978-94-91216-74-9_7,  Atlantis Press 2013

[email protected]
146 Decision Aid Models for Disaster Management and Emergencies

hazard (e.g., disaster) resources [1]. It is the process of assessing hazard risks and taking
steps to either eliminate or to reduce them (as far as is reasonably practicable) by introduc-
ing control measures.
Risk assessment is one of the key elements of risk management. Expressions such
as “risk assessment”, “risk evaluation” and “risk analysis” are used in a somewhat inter-
changeable way to describe a variety of techniques and processes involved in the overall
management of risk [2]. Despite this lack of clarity, Frosdick [2] and other researchers
consistently use the term “risk assessment” as a catch-all to include all those activities that
are needed before appropriate risk reduction methods can be decided upon.

Define scope
and objectives

Identify hazards / define

potential accident
Risk
Analysis
Evaluate the event Estimate the potential Risk
consequences/effect accident frequencies Assessment

Risk
Estimate the risk Management

Risk Evaluation
-tolerability decisions
-analysis of options

Risk reduction / control

-decision making
-implementation/monitoring

Fig. 7.1 A simplified relationship between risk analysis, risk assessment and risk management

Fig. 7.1 shows the different processes in risk management procedure and presents a
simplified relationship between risk analysis, risk assessment and risk management. This
definition of the risk management process has been adopted from the International Electro-
technical Commission (IEC). There are other relationships between the definitions of risk

[email protected]
A Belief Rule-Based Generic Risk Assessment Framework 147

analysis, risk assessment and risk management, where risk assessment is part of the risk
analysis. In this paper, we focus only on risk assessment issue.
Risk assessment (especially in disaster management) is a very complicated subject
where risk is determined by numerous factors, including human error. The general area
of risk assessment is vast, with many methods and tools available to use for assessing risk
of various environments [3]. Many risk assessment techniques currently used in the disas-
ter management are comparatively mature tools. However, there are still some limitations
for these techniques to be widely and effectively applied to provide useful solutions to
risk-based decision making. This is due to the following problems among others:

• Under prevailing circumstances, only limited or no previous experience exists on a

hazard, for which the statistical accuracy is poor, in particular for a risk scenario where
undesired events are extremely rare.
• It is extremely difficult to quantify the effects and consequences of hazards as they
involve too many factors with a high level of uncertainty, even in those cases where
the physical processes are clearly understood.
• It is extremely difficult to generate a mathematical model to represent and describe
the risk behavior/discipline of disaster hazards as risk is a multiple-level and multiple-
variable optimization problem.
• A large number of assumptions, judgments and opinions are involved subjectively in
a risk quantification process. Therefore, it may require considerable skill for a risk
analyst to interpret the results produced.

Very importantly, handling uncertainty is one of the crucial issues in the risk assessment
in complex systems with diverse environments. Uncertainty in risk assessment can be
considered essentially of two different types: randomness due to inherent variability in
the system (i.e., in the population of outcomes of its stochastic process of behavior) and
imprecision due to lack of knowledge and information on the system. The former type of
uncertainty is often referred to as objective, aleatory, stochastic whereas the latter is often
referred to as subjective, epistemic, state-of-knowledge [4, 5]. In most risk assessment, we
have to rely on such imperfect information through appropriate risk management.
In certain circumstances, probability theory can be a powerful tool. Traditional study of
risk analysis is conducted using probabilistic tools and techniques. However, it is not diffi-
cult to see that aspects related to imprecision or vagueness clearly have a non-probabilistic
character since they are related to imprecision of meanings. Very often the type of uncer-
tainty encountered in hazard does not fit the axiomatic basis of probability theory, simply

[email protected]
148 Decision Aid Models for Disaster Management and Emergencies

because uncertainty in these hazards is usually caused by the inherent fuzziness of the pa-
rameter estimate rather than randomness. Traditional approaches, e.g., quantitative risk
assessment (QRA) including probabilistic safety/risk analysis (such as Fault Tree Analy-
sis (FTA) and Failure Mode, Effects and Criticality Analysis (FMECA)) have been widely
used, but often fail short in their ability to permit the incorporation of subjective and/or
vague terms as they rely heavily on supporting statistical information that may not be avail-
able [7].
Extensive research has been devoted to the analysis and management of the risks under
uncertainty. Some detailed overviews of uncertainty aspects in risk and safety management
can be found in [8–15]. Chowdhury et al. (2009) provided a detailed review of uncertainty
analysis in risk management studies associated with disinfection by-products (DBPs) in
drinking water and human health risk [9]. Markowski (2010) discussed and presented the
sources and types of uncertainties encountered in process safety analysis and also methods
to deal with them [13]. From the literature, many different formal techniques have been
developed over the past two decades for dealing with uncertain information for risk assess-
ment in decision making, where Bayesian probability theory [16, 17], Dempster-Shafer
theory of evidence [18, 19], and fuzzy logic [20–22] are three of the most common meth-
ods of representing and reasoning with uncertain knowledge. Liu et al. (2011) provided a
detailed review of risk assessment based on computing with words [23].
Novel risk analysis methods are therefore required to identify major hazards and assess
the associated risks in an acceptable way in various environments where mature tools can-
not be effectively or efficiently applied. The subject of our work would be to establish a
general framework that provides the basis and tool for risk analysis in different application
area, especially to deal with information that may be unquantifiable due to its nature, and
imprecise, too complex, ill-defined, incomplete etc., for which the traditional quantitative
approach (e.g., statistical approach) does not give an adequate answer.
The rest of this paper is organized as follows: Risk specification parameters for a haz-
ard event assessment are briefly overviewed in Section 7.2. The generic belief rule based
hazard risk assessment framework is then justified and outlined in Section 7.3. In order to
handle more complex situation, the hierarchical structure of risk nature in complex hazard
situation are illustrated using some realistic examples and the multi-layer belief rule-based
risk assessment framework is introduced in Section 7.4. Conclusions are drawn in Sec-
tion 7.5.

[email protected]
A Belief Rule-Based Generic Risk Assessment Framework 149

7.2 Hazard Associated Risk Specification Parameters

As can be seen from Fig. 7.1, risk assessment is a key element of risk management;
it can be further separated depending on how detailed are the analysis and the labour re-
sources available in at least three levels:

• qualitative methods
• semi-quantitative methods
• quantitative methods

During risk analysis, all three levels can be used in sequence. Qualitative methods are
used to determine which scenarios are relevant to continue with the quantitative risk analy-
sis. Initially, risk assessments were qualitative because of their subjectivity. The search for
greater objectivity, led to the development of quantified risk analysis techniques. Quanti-
tative risk analysis methods fall under the broad category of probabilistic risk assessment
(PRA). Risk is generally characterized by the severity (or magnitude) of an adverse conse-
quence that can result from an action and the likelihood of occurrence of the given adverse
consequence. In PRA, consequences are expressed numerically and their likelihoods of
occurrence are expressed as probabilities or frequencies. Accordingly, risk is defined as the
product of likelihood and severity. Determination of needed basic event probabilities is the
most difficult task in applying this technique. Many references explain all aspects of PRA
in great detail [24–26].
Risk is sometimes characterized not only by likelihood and severity, but also with some
additional parameters depending on the different applications. The definitions of those
parameters also vary according to different application contexts. For example, FMEA ap-
proach is a widely used engineering technique for defining, identifying and eliminating
known and/or potential failures, problems, errors and so on from system, design, process,
and/or service before they reach the customer [27]. The traditional FMEA determines the
risk priorities of failure modes through the risk priority number (RPN), which is the product
of the occurrence (O), severity (S) and detection (D) of a failure. That is:

RPN = O × S × D,

where O and S are the frequency and seriousness (effects) of the failure respectively, and
D is the ability to detect the failure before it reaches the customer. The failure modes
with higher RPNs are assumed to be more important and will be given higher priorities for
correction.

[email protected]
150 Decision Aid Models for Disaster Management and Emergencies

Table 7.1 Qualitative and quantitative parameters

Risk parameter Qualitative descriptions Quantitative descriptions
Minor injury No deaths per event
Consequence (C) Marginal: one death or permanent [10−2 , 10−1 ] probable deaths per
injury event
Critical: several deaths [10−1 , 1] probable deaths per event
Catastrophic: many deaths > 1 probable deaths per event
Rare < 10% of time
Exposure (F)
Frequent  10% of time
Possible 90% probability of avoiding hazard
Avoidance (P)
 90% probability of avoiding
Not likely
hazard
Very low < 1 in 30 years ≈< 0.03 per year
Demand rate (W ) 1 in [3, 30] years ≈ [0.03, 0.3] per
Low
year
Relatively high 1 in [0.3, 3] years ≈ [0.3, 3] per year

Wang et al. [27] used three fundamental parameters to assess the safety level of an
engineering system on a subjective basis: failure rate (FR), consequence severity (CS) and
failure consequence probability (FCP). FR describes failure frequencies in a certain period,
which directly represents the number of failures anticipated during the design life span
of a particular system or an item. CS describes the magnitude of possible consequences,
which is ranked according to the severity of failure effects. FCP defines the probability that
consequences happen given the occurrence of the event.
In addition, four risk parameters, considered to be sufficiently generic to deal with a
wide range of applications, have been combined to risk assessment as well. These parame-
ters are: consequence (C), frequency and exposure time (F), possibility of avoiding hazard
(P), and probability of the unwanted occurrence (W ). All parameter aspects imply a quanti-
tative or qualitative valuation of undesired events or harmful events effects. Table 7.1 shows
an example of a risk graph as used in the UKOOA guidelines and quantitative definitions
of risk parameters [28, 29]. Baybutt (2007) has developed an improved risk assessment
with the following four parameters: initiating cause frequency, enabling events/conditions,
safeguards failure probability and consequences of the hazardous event [30].
In summary, how to represent and determine those risk relevant factors, and then model
the relationships between risk factors and risk level are the key issues to provide the risk
estimation of a hazard. For example, to assess the safety associated with an event, it is re-
quired to synthesize the associated occurrence likelihood, consequence severity and failure

[email protected]
A Belief Rule-Based Generic Risk Assessment Framework 151

probability. The way of synthesis of risk factors into risk estimation can be represented in
different ways based on cause and effect relationship. Different methodologies have been
proposed accordingly in different applications. The following section provides a general
risk assessment framework.

7.3 A Belief-Rule Based Generic Hazard Risk Assessment Framework

7.3.1 Motivation and Rationale

As discussed in Section 7.2, the nature of risk is usually affected by numerous factors
including human errors, in many circumstances; it may be extremely difficult to assess the
associated risks due to the great uncertainty involved. The quantitative risk assessment
approaches fall short in their ability to permit the incorporation of subjective and/or vague
terms as they rely heavily on supporting statistical information that may not be available.
Quantification of risk in scalar values is subject to uncertainties for many reasons including
difficulties in defining the likelihood and consequence severity and the mathematics of
combining them. Collecting sufficient data to base a statistical probability of risk is costly,
and in many situations, such data are limited or unavailable due to a lack of research or
the complexity of the system/process considered. In addition, it may be extremely difficult
to construct an accurate and complete mathematical model to correlate the risk factors and
risk levels.
Those quantitative assessments of risk are particularly challenging in domains where
undesired events are extremely rare, and the causal factors are difficult to quantify and
non-linearly related, e.g., include the difficulty of determining both the probabilities of rare
events (such as a nuclear accident, or only incomplete information is available during the
very early phases of the system life cycle), and their severity, further, the probabilities may
be dynamic, and vary with a variety of factors which are not known in advance.
To overcome the above drawbacks, many approaches have been proposed, among many
alternative approaches for knowledge representation and reasoning, it is widely recognized
that the rule-based system seem to be one of the most common framework for expressing
various types of knowledge in an intelligent system (systems that have elements of human
reasoning and certain level of intelligence) [6, 31]. As discussed in the above section, in the
design and implementation of rule-based systems for supporting human decision making,
it is necessary and inevitable to use a scheme for representing and processing imprecise,
incomplete and uncertain information in conjunction with precise data. During the last

[email protected]
152 Decision Aid Models for Disaster Management and Emergencies

quarter of a century many different types of rule-based systems emerged, certainly includ-
ing the fuzzy rule-based system, which, as one of the dominant and main framework in
rule-based system, has been widely accepted, investigated and applied in many application
areas. Fuzzy logic has been also widely applied in risk assessment to different areas and
has been regarded as one of the promising methods for the reduction of uncertainties in risk
assessment. An important contribution of fuzzy system theory is that it provides a system-
atic procedure, i.e., fuzzy rule base approach, for capture the uncertainty and the non-linear
relationships among the system input and output parameters. Fuzzy rule-based system is
constructed using human knowledge in the form of IF-THEN rules. In risk assessment risk
factors are inputs, and risk estimation is the output. The relationship between risk factors
and risk is described by IF-Then Rules. For example, the following is a classical IF-THEN
rule for safety analysis [10]:
IF FR of a hazard is frequent AND CS is catastrophic AND FCP is
(7.1)
likely, THEN risk estimate is high
The detailed review of fuzzy logic based risk assessment methodology could be found
in Refs [32–34] and mostly recently in Refs. [23] and [35–38], this showed the successful
application of fuzzy logic into risk assessment.
Moreover, in recognition of the need to handle hybrid information with uncertainty
in human decision making, a new methodology has been proposed recently for modeling
a hybrid rule-base using a belief structure and for inference in the belief rule-based sys-
tem using the evidential reasoning (ER) approach [39]. The methodology is referred to
as a belief Rule-base Inference Methodology using the Evidential Reasoning approach –
RIMER [40], where a rule-base is designed with belief degrees embedded in the consequent
term of a rule, called belief rule-base (BRB), is used to capture nonlinear causal relation-
ships as well as uncertainty. The inference of a rule-based system is implemented using
the ER approach. RIMER approach has been further investigated and applied to different
areas, such as, among others, work in [10, 11, 41–43].
The main methodologies and technical framework involved in RIMER is summarized
in Fig. 7.2 below.
The RIMER approach is summarized here, and see ref. 40 for more details. Suppose a
belief rule-base is given by R = {R1 , . . . , RL } with the kth rule represented:
Rk : IF U is Ak THEN D with belief degrees β k , with a rule weight θk and
(7.2)
attribute weights δk1 , . . . , δkTk
where U represents the antecedent attribute vector (U1 , . . . ,UTk ), Ak the packet antecedents
{Ak1 , . . . , AkTk }, Aki (i = 1, . . . , Tk ) the referential value of the ith antecedent attribute in the kth

[email protected]
A Belief Rule-Based Generic Risk Assessment Framework 153

Fig. 7.2 Main methodologies and technical framework in RIMER

rule; Tk is the number of antecedent attributes used in the kth rule. Suppose T is the total
number of antecedent attributes used in the rule base, D the consequent vector (D1 , . . . , DN ),
and β k the vector of the belief degrees (β1k , . . . , βNk ) for k ∈ {1, . . . , L}, and βik the belief
degree to which Di is believed to be the consequent if in the kth packet rule the input
satisfies the packet antecedents Ak . θk is the relative weight of the kth rule and δkTk the
relative weights of the Tk antecedent attributes in the kth rule. L is the number of all the
i=1 βik
N
packet rules in the rule-base. If = 1, the kth packet rule is said to be complete;
otherwise, it is incomplete. Rule (7.1) is referred to as a belief rule.
Accordingly fuzzy rule in (7.1) is extended into fuzzy rules with belief degrees for
multiple possible consequent terms in the safety assessment [10]:
Rk : IF the FR is frequent AND the CS is critical AND the FCP is unlikely
(7.3)
THEN risk estimate is {(high, 0.7), (Average, 0.3), (Fair, 0), (Low, 0)}
where risk estimate is a belief distribution representing that we are 70% sure that risk level
is Fair, and 30% sure that it is Poor, which reflects another kind of uncertainty caused
because evidence available is not sufficient or experts are not 100% certain to believe in a
hypothesis but only to degrees of belief. This kind of belief rule is able to capture uncer-
tainty and nonlinear relationships between these three parameters and the risk level.

[email protected]
154 Decision Aid Models for Disaster Management and Emergencies

Once given an input, the activation weight wk for Ak , which measures the degree to
which the kth rule is weighted and activated, is calculated by:
Tk L Tk
δ̄i δ̄l
wk = θk ∗ αik θi ∗ αli with δ̄i = δi max {δi } (7.4)
i=1,...,Tk
i=1 i=1 l=1
where αik ∈ [0, 1] called the individual matching degree, is the degree of belief to which the
input for Ui belongs to Aki of the ith individual antecedent in the kth rule, and αik could be
generated using various ways depending on the nature of an antecedent attribute which will
be seen in details in Ref. [40].
Having determined the activation weight of each rule in the rule base, the ER ap-
proach [39] can be applied to combine the belief distributions of the consequence of all
the rules and generate final conclusions. The conclusion generated by aggregating all ac-
tivated rules by the actual input vector I = (I1 , . . . , ITk ) for the antecedent vector U can be
represented as follows:
S(I) = {(D j , β j ), j = 1, . . . , N} (7.5)
The result in Eq. (7.5) reads that if the input is given by I, then the consequent is D1 to
a degree β1 , . . ., and DN to a degree βN . Using the analytical ER algorithm [41], the overall
combined degree of belief β j for D j in Eq. (7.5) is generated as follows:
L N L N
μ× wk β j,k + 1 − wk βi,k − 1 − wk βi,k
βj = k=1 i=1
L
k=1 i=1
, j = 1, . . . , N (7.6)
1−μ× (1 − wk )
k=1
where
−1
N L N L N
μ= wk β j,k + 1 − wk βi,k − (N − 1) 1 − wk βi,k .
j=1 k=1 i=1 k=1 i=1
Note that the final result in Eq. (7.5) is still a belief distribution on an estimate, which
gives a panoramic view about estimate for a given input. It clearly shows its nonlinear
features in aggregating a large number of consequent attributes. The belief theory together
with fuzzy logic provides a flexible approach to represent uncertain evidence and uncertain
knowledge. The rules represent functional mappings between imprecise information with
uncertainty. A fuzzy rule-base designed on the basis of a belief structure is used to capture
uncertainty and non-linear relationships between these parameters and the safety level;
and the inference of the rule-based system is implemented using the evidential reasoning
algorithm. RIMER are an extension of traditional rule based systems and are capable of
representing more complicated causal relationships using different types of information
with uncertainties.

[email protected]
A Belief Rule-Based Generic Risk Assessment Framework 155

7.3.2 Hazard risk assessment framework based on RIMER

The proposed framework for risk analysis consists of five major components, which are
outlined as follows:
Component #1: Identification of causes/factors of a hazard event
In this step, as discussed in Section 7.2, all the anticipated causes/factors to a hazard
event are identified. This can be done by involvement of a panel of experts during a brain-
storming session.
Component #2: Specify and represent input (e.g., failure rate, consequence severity
and failure consequence probability for risk assessment) and output/solution variables (i.e.,
risk estimates) and create the fuzzy membership functions or utility function for all the
related input variables.
Some fundamental parameters used to assess the risk level of system on a subjective
basis (using linguistic variables instead of ultimate numbers in probabilistic terms) need to
be defined, such as failure rate, consequence severity and failure consequence probability,
or some other parameters as discussed in Section 7.2, are more appropriate for analysis
using these three parameters as they are always associated with uncertainty. These lin-
guistic assessments can become the criteria for measuring risk levels. The definitions and
identifications of those parameters differ from different applications according to different
requirements in codes and standards (e.g., safety/risk guidelines, regulations, laws etc.) and
different aspects of engineering systems such as fire, explosions, structure, safety system,
etc., for example, different definitions can be found in Refs. [27, 44, 45].
Component #3: Construct the rule base with belief structure

7.3.2.1 Optimization models based on training sample data

Optimization models procedures are investigated in the RIMER framework based on the
optimization models for training general belief rules based systems [41], in order to help
generate desirable belief rules and other system parameters simultaneously. Figure 7.1
shows the process of training a belief rule base with additional elements to deal with fuzzi-
ness, where x̂m is a given input, ŷm the corresponding observed output, either measured
using instruments or assessed by experts, ym the simulated output generated by the belief
rule based system, and ξ (P) the difference between ŷm and ym , as defined later.
In the learning process, a set of observations on the system input and output is required.
In the following, we assume that a set of observation pairs (x, y) is available, where x is an
input vector and y the corresponding output vector. Both x and y can be either numerical,

[email protected]
156 Decision Aid Models for Disaster Management and Emergencies

Observed output ( ŷ m )
Real System
Input ( x̂m ) [ (P)

Simulated output ( y m )
Belief rule base

Fig. 7.3 Illustration of Optimal Learning Process

judgmental or both. The format of the objective functions is important for the parameter
optimization. Depending on the type of input and output, the optimal learning model can
be constructed in different ways, as discussed in details in Refs. [41, 42]: i) optimization
algorithms based on the numerical output form; ii) optimization algorithm based on the
output in the form of the subjective judgment; iii) optimization algorithm based on the
input-output in form of interval.
It is worth note that in this case (ii), a training set composed of M input-output pairs
(xm , ym ) (m = 1, . . . , M), where ym can be a subjective judgment, i.e., a distributed assess-
ment on the linguistic value with belief. Notice that the single judgment as one linguis-
tic value can be regarded as a special case of the distribution assessment. The output of
RIMER-based risk assessment is actually a distribution safety assessment instead of a sin-
gle numerical score, which provides a panoramic view about the output status, from which
one can see the variation between the original output and the revised output on each linguis-
tic term. A distribution is easy to understand and flexible to represent output information
than a single average value. Especially it is very useful in the case that the output is dif-
ficult to quantify. For example, the linguistic terms in risk (e.g., high, fair, or low, etc.)
have no clearly defined bases and are difficult to define quantitatively. So, it is better to
draw a conclusion using the same linguistic terms as the ones in the consequent but with
different degrees of belief. Such subjective judgments with belief are useful in such areas
as safety/risk classification.
In summary, due to the use of belief rules, the RIMER method provides an analytical
description of relationships between system’s inputs and outputs. This enables a belief
rule-based system to act as a generic functional mapping from system inputs to outputs and
allows powerful learning techniques to be used for parameter training and system updating
on the basis of both numerical data and human judgments, this is one of the main feature
in the training models.

[email protected]
A Belief Rule-Based Generic Risk Assessment Framework 157

7.3.2.2 Construct the rule base with belief structure based on experts

Due to the qualitative nature of linguistic rule base, the causal relationship between
IF part and THEN part is not possible to be obtained due to lack of complete and accu-
rate information regarding the training set, e.g., a new system or rarely happened events.
Such a situation typically arises when learning examples are generated by one or several
experts, whose subjective evaluation of consequence may be tainted with imprecision and
uncertainty. For each expert may not always be able to classify clearly the consequence
class with full certainty. He or she may, however, be able to assess the “likelihood” that a
certain phenomenon is present in the data. Hence, an approach to obtain belief rule base
is to collect the opinions of several experts, and consider for each example the empirical
distribution of expert opinions about its class belief. Belief distribution in Eq. (7.2) can
be obtained in several different ways [46], called belief elicitation method, which in this
framework means any aid that is used to acquire a probability from an expert, including:
direct elicitation from an expert, who is asked to quantify by a real number between
0 and 1 the degree of belief that IF part Ak belongs to each of the N evaluation grade
D1 , . . . , DN . However, people find words more comfortable than numbers probably be-
cause the vagueness of words captures the uncertainty they feel about their probability
assessments. Since directly assessed numbers tend to be biased, various indirect elicitation
methods have been developed. So an alternative is suggested, i.e.,
from an empirical distribution of expert opinions, using the relative frequency:
uik
βik = , (7.7)
M
where uik denotes the number of experts (out of M) who assigned IF part Ak (k = 1, . . . , L)
into the evaluation Di (i = 1, . . . , N) in THEN part. Alternatively, we could use possibilistic
histograms [47] to generate a relative frequency as follows:
uik 1
βik = , where uik = min(uik , u jk ) (7.8)
utk M
j=1,...,N
t=1,...,N

The above feature and approaches associated with RIMER shows that it can be regarded
as data-driven and domain knowledge based risk assessment model. This belief rule base
is a great extension of traditional fuzzy rule based systems. In a belief rule based system,
while human expert knowledge is used to construct a roughly correct belief rule base, the
optimal learning mechanism can help to fine tune system performance if the system input–
output data are available. As such, we believe that reasoning with fine-tuned logical rules
is more acceptable to human users than the recommendations given by a black box system,

[email protected]
158 Decision Aid Models for Disaster Management and Emergencies

because such reasoning is comprehensible, provides explanations, and can be validated by

human inspection. It also increases confidence in the system, and may help to discover
important relationships and combinations of features [41].
Component #4: Reasoning mechanism based on belief rule-base and evidential rea-
soning algorithm.
Once rule-base is built up, the knowledge contained is used to perform the inference
procedure while the inputs are given. For each global antecedent attribute, the matching
degree between the input and each individual antecedent need to be determined and finally
get the global matching degree by combining the individual matching degree. These global
matching degrees are used as the strength weights of the global attribute. The inference
procedure is basically composed of the following steps, and can be summarized as follows:

Step 1: Discretization of the input variables into the distributed representation of linguistic
terms in the antecedent;
Step 2: Selection of “AND” connectives to reflect the dependencies of attribute in the an-
tecedent of the rule;
Step 3: Determination of the actual degree of belief of the rule;
Step 4: Fuzzy rule expression matrix of knowledge-base;
Step 5: Rule combination by using evidential reasoning (ER) algorithm.

Component #5: Multi-source risk synthesis

In this component, some aggregation approaches are used in the later stage of the frame-
work to deal with risk synthesis with complexity involving multi-experts or multi-sources,
this is to integrate all the possible risk estimates made by a panel of experts or based on dif-
ferent sources. The ranking and interpretation of the final safety/risk synthesis of a system
is given.
Assume that there are several sources (including experts) Si (i = 1, . . . , K). Without
loss of generality, suppose input comes from different sources evaluated by different ex-
perts. Note that it is likely for selected information sources to be of different importance
(or reliability), so the weights of information sources need to be taken into account. The
assessment of weight for each information source is an important decision for the analyst to
make in view of the safeguards system under scrutiny. Each information source is assigned
a weight to indicate the relative importance/reliability in contributing towards the overall
evaluation. The analyst must decide which information source has higher reliability and
then assign weights accordingly.

[email protected]
A Belief Rule-Based Generic Risk Assessment Framework 159

We may assume that different sources have different reliability weights, wsi (i =
1, . . . , K). Suppose Aei ( j) is an input vector derived from ei for an event I j . For each
input, we may get a corresponding process estimate Dei using the above RIMER approach,
which can be formulated as follows:

IF I j is Ae1 ( j) THEN Pe1 ( j) is {(D1 , η11 ), (D2 , η21 ), . . . , (DN , ηN1 )}

IF I j is Ae2 ( j) THEN Pe2 ( j) is {(D1 , η12 ), (D2 , η22 ), . . . , (DN , ηN2 )}
...
IF I j is Aei ( j) THEN Pei ( j) is {(D1 , η1i ), (D2 , η2i ), . . . , (DN , ηNi )}
...
IF I j is AeK ( j) THEN PeK ( j) is {(D1 , η1K ), (D2 , η2K ), . . . , (DN , ηNK )}

where {(D1 , η1i ), (D2 , η2i ), . . . , (DN , ηNi )} results from Eq. (7.5) obtained using the
RIMER approach. Then the actual risk estimate of a specific hazard event P( j) can be gen-
erated by synthesizing multi-expert assessments, i.e., by aggregating {Pe1 , Pe2 , . . . , PeK }
using the ER algorithm, which is represented as

S(P( j)) = (Di , ηi ); i = 1, . . . , N ,

j
j = 1, . . . , d (7.9)

The risk assessment and synthesis framework is summarized and illustrated in Fig. 7.4.

7.4 A Belief-Rule Based Generic Hierarchical Risk Assessment Framework

Section 7.3 outlines the general individual hazard risk assessment framework using
RIMER approaches. However, due to the complexity and uncertainty of real world systems
and risks, it is necessary to assess the risks by considering the organization of risk informa-
tion and risk mechanism in different level. Hierarchical framework method is considered
to be straightforward and effective to deal with complexity. Then hierarchical assessments
are conducted to evaluate the risks of components first and then the whole system. In risk
assessment, one of the most valuable and critical contributions of hierarchical framework
is its ability to facilitate the evaluation of subsystems risks and their corresponding con-
tributions to the risks of the overall system. Particularly, its ability to model the intricate
relationships among the various subsystems and to account for all relevant and important
elements of risk and uncertainty renders the risk assessment process more tractable, repre-
sentative, and encompassing [48].
In general, a hierarchical risk assessment model is composed of risk attributes (some-
times also referred to as risk variables or risk parameters) that represent risk subproblems.

[email protected]
160 Decision Aid Models for Disaster Management and Emergencies

Inputs MF/Utility
Function Fact base, i.e.
data base

Rule base with

belief structure
Input distribution
transformation

RIMER
Activation weight Engine

Risk Estimate For individual input

Multi-expert/Multi-source Aggregation using ER algorithm

For an overall
assessment of a
hazard event
Risk Synthesis

Ranking and interpretation of the final risk synthesis of a hazard

Fig. 7.4 Generic individual hazard risk assessment framework using RIMER Approaches

They are organized hierarchically so that the risk attributes that occur on higher levels of
the hierarchy depend on lower-level risk attributes.

Example 7.1. A hierarchical model for the assessment of patient’s risk status is shown in
Fig. 7.5 [49]. It consists of three main groups of attributes: history (data about previous
ulcers and amputations), present status (data on symptoms, deformities and other changes),
and the results of tests (loss of protective sensation, absence of pulse).

[email protected]
A Belief Rule-Based Generic Risk Assessment Framework 161

RISK

History Present status Tests

Other Loss of port. Absence of

Ulcers Amputation Symptoms Deformities
changes sensation pulse

Fig. 7.5 The hierarchical structure of model for risk assessment in diabetic foot care

Associated with each node, there are transition links that explicitly describe the risk
relationships between each other. Belief rule base can be carefully designed so as to deter-
mine the higher level risk level based on lower level factors. It is clear that the attributes in
a lower level are different from each other in types and representations, also different from
the high level attributes. The leads to the difficult to establish the mathematical model to
correlate their relationship, IF-THEN rule is more appropriate here to model the nonlinear
relationship, and belief IF-THEN rule could be further utilized to reflect the uncertainty.

Example 7.2. A hierarchical model for the assessment of cancer risk is shown in
Fig. 7.6 [49], which is assessed using a four-valued scale, and is derived from features
such as age, regularity of menstruation and fertility duration.

Similarly, the risk of cancer can be evaluated by hieratical belief-rule base, which can
be defined by the experts. As an example, consider Table 7.2 that shows the rule that
determines the risk with respect to menstrual cycle and derives it from two input attributes:
fertility duration and regularity: stability of menstruation. Here, the fertility duration can
be either short (up to 30 years of menstrual period), average (30–40 years), or long (longer
than 40 years). Menstruation cycle can be either regular with a period of less than 28 days
(R-28), regular with a period longer than 29 days (R29_) or irregular (N).
For each combination of fertility duration and menstruation regularity: stability, the ex-
perts assessed the risk related to menstrual cycle and expressed it using a three-level scale:
low, medium, and high. Each row in the table can thus be interpreted as an elementary
if-then rule that assesses the risk for the corresponding values of the two input attributes.

[email protected]
162 Decision Aid Models for Disaster Management and Emergencies

RISK

Hormonal Personal
circumstances characteristics Others

Menstrual Oral Quetel’s Cancerogenic

cycle Fertility contaceptiveses index exposure Demographic
circumstances

Fertility Age Family Physical

duration history factors

First delivery
Reg. and stab. Chemical
Menopause factors
of menstr.

# deliveries

Fig. 7.6 The structure of breast cancer risk assessment model

Also, each row represents a point in the space determined by all the three attributes. It
is clear from Table 7.1 that those classical IF-THEN rules shown some limitation which
cannot distinguish some rules even they have the different antecedent terms, e.g., when
Fertility duration is Long, whatever Regularity is R-28, R29+ or N, Menstrual cycle is al-
ways High risk. Using belief IF-THEN rule could simply solve this problem without loss
of any information:

IF Fertility duration is Long and Regularity is R28+ THEN Menstrual cycle is {(High
risk, 0.9), (Moderate risk, 0.1), (Low risk, 0)}
IF Fertility duration is Long and Regularity is R-28 THEN Menstrual cycle is {(High
risk, 0.8), (Moderate risk, 0.2), (Low risk, 0)}

[email protected]
A Belief Rule-Based Generic Risk Assessment Framework 163

Table 7.2 Decision rule to assess the risk specific to menstrual cycle [49]
Fertility duration Regularity/stability of menstruation Menstrual cycle
1 Average R-28 High risk
2 Long R-28 High risk
3 Long R29+ High risk
4 Long N High risk
5 Short R-28 Moderate risk
6 Average R29+ Moderate risk
7 Short R29+ Low risk
8 Short N Low risk
9 Average N Low risk

IF Fertility duration is Long and Regularity is N THEN Menstrual cycle is {(High risk,
0.7), (Moderate risk, 0.2), (Low risk, 0.1)}

Belief rule base can be established to model the relationship between lower level and
higher level following the hierarchical structure in Fig. 7.6. The overall risk can be achieved
by integrating all the sub-belief rule based inference processes.

Example 7.3. A scheme of the three-level hierarchical social risk assessment system is
presented in Fig. 7.7 [50].

A social risk model can be designed as a hierarchical structure with several inputs
and one output. The social risk of the studied region is assessed based on the hyrdo-
meteorological and seismic information taking into account the social vulnerability. The
social risk assessment from the natural hazards is done under the subjective and uncertain
conditions. The number of inputs corresponds to the linguistic variables (indicators), which
described the environmental risk and social vulnerability. The output represents a social
risk assessment from natural disasters. Five indicators for the social risk assessment are
defined using the expert knowledge, statistical data and published thematic maps for the
seismic, and flood hazards [50]. The inputs are defined as follow:

• Input 1 “Extreme temperature”;

• Input 2 “Floods”;
• Input 3 “Seismic hazard”;
• Input 4 “Population density”;
• Input 5 “Socioeconomical status”.

[email protected]
164 Decision Aid Models for Disaster Management and Emergencies

Social risk

Environmental risk Social vulnerability

Climatic risk Seismic hazard

Floods Extreme temperatures Population density Socio-economical status

Fig. 7.7 Three-level hierarchical system for social risk assessment

The risk model is designed as a three-level belief rule based hierarchical model with
previously defined five inputs. The first level includes one RIMER subsystem. The second
level includes two RIMER subsystems. The third level includes only one subsystem. The
each RIMER subsystem has two inputs. The hierarchical system output gives the social risk
assessment of the natural hazards. The value of the complex assessment is a criterion for
final decision making about the degree of social risk for the considered areas. The higher
value corresponds to the higher risk degree.
Inherently qualitative features of the indicators are rather than quantitative values,
which are usually represented by linguistic variables. Some of the belief rules are defined
as follow:

IF “Extreme temperature” is “Middle” and “Floods” is “Low” THEN “Climatic risk”

is {(Very high, 0), (High, 0), (Middle, 1), (Low, 0), (Very low, 0)};
IF “Extreme temperature” is “High” and “Floods” is “Middle” THEN “Climatic risk”
is {(Very high, 0), (High, 0.8), (Middle, 0.2), (Low, 0), (Very low, 0)};
IF “Climatic risk” is “Middle” and “Seismic hazard” is “Low” THEN “Environmental
risk” is {(Very high, 0), (High, 0), (Middle, 0.1), (Low, 0.9), (Very low, 0)};
IF “Climatic risk” is “High” and “Seismic hazard” is “Middle” THEN “Environmental
risk” is {(Very high, 0), (High, 0.2), (Middle, 0.8), (Low, 0), (Very low, 0)};

[email protected]
A Belief Rule-Based Generic Risk Assessment Framework 165

IF “Population density” is “Low” and “Socio-economic status” is “Middle” THEN

“Social vulnerability” is {(Very high, 0), (High, 0), (Middle, 0.1), (Low, 0.9), (Very
low, 0)};
IF “Population density” is “Middle” and “Socio-economic status” is “High” THEN
“Social vulnerability” is {(Very high, 0), (High, 0.2), (Middle, 0.8), (Low, 0), (Very
low, 0)};
IF “Environmental risk” is “Low” and “Social vulnerability” is “Low” THEN “social
risk from natural disasters” is {(Very high, 0), (High, 0), (Middle, 0), (Low, 0.2), (Very
low, 0.8)};
IF “Environmental risk” is “High” and “Social vulnerability” is “Low” THEN “social
risk from natural disasters” is {(Very high, 0), (High, 0.7), (Middle, 0.2), (Low, 0.1),
(Very low, 0)};
IF “Environmental risk” is “High” and “Social vulnerability” is “High” THEN “social
risk from natural disasters” is {(Very high, 0.7), (High, 0.3), (Middle, 0), (Low, 0),
(Very low, 0)}.

Similarly, after setting up the belief rule base, once given the input in the bottom level,
then the resulting social risk can be estimated using the hierarchical risk assessment scheme
based on the framework proposed in Section 3 from bottom level to top level.

7.5 Conclusions and Outlook

Based on the brief review of risk assessment specification and hierarchical nature in the
complex situation, the paper proposes a generic hazard risk assessment framework based
on the belief rule based inference methodology – called RIMER for modeling, analyzing
and synthesizing risk-related information with various uncertainties. In this framework,
risk factors are described using linguistic variables with belief, and an IF-THEN rule base
with a belief structure, i.e., IF-THEN rules with belief degrees for all possible output terms
in the consequent, is to capture uncertain causal relationships between the risk factors and
the special risk estimate. Moreover, the antecedent of each IF-THEN rule is considered as
an overall attribute, which is assessed to an output term in the consequent of a rule with
a belief degree. Actual input can be transformed into a distributed representation for a
linguistic term of an individual antecedent indicator. Finally, the inference of the rule-base
is implemented using the evidential reasoning algorithm, where an activation degree is used
as the weight of an overall attribute. The input information for each risk can be provided

[email protected]
166 Decision Aid Models for Disaster Management and Emergencies

from different sources which are normally managed by different experts. Once given an
input, RIMER can be used to inference and generate an output. The modeling framework
of multi-experts/multi-source synthesis is then provided based on the evidential reasoning
approach, i.e., a combination of evidences from different sources.
This general individual hazard risk assessment framework is then followed by the il-
lustrative analysis of hierarchical structure nature of the real world risk assessment model.
How can the above discussed risk assessment framework be incorporated to form a multi-
layer belief rule-base risk assessment model is finally specified.
The proposed risk assessment framework is able to analyzing risks and safety of com-
plex systems, products and processes involves multiple factor analysis and handling uncer-
tain information of both a quantitative and qualitative nature. The belief rule based system
is flexible, can be adapted to represent complicated systems, and is a valid novel approach
for risk assessment. In order to make all the processes of rule generation and rule-based in-
ference more efficient, a software package called an intelligent decision support tool based
on belief rule-based inference methodology has been developed [51], which provided nec-
essary support for the work proposed and enhanced the applicability of the methodology in
different application.

Dedications

This paper is dedicated to the memory of our close friend, an excellent, passionate, and
dedicated scientist, Prof. Dr. Da Ruan, who recently passed away in the age of only 50.
We will miss you forever, Da!

Acknowledgments

This work is partially supported by the research project TIN2009-08286.

Bibliography

[1] J.C. Augusto, H. Wang, and J. Liu. Situation assessment during disaster management. Interna-
tional Journal of Computational Intelligence Systems, 1 (3) (2008), pp. 237–247.
[2] S. Frosdick. The techniques of risk analysis are insufficient in themselves. Disaster Prevention
and Management, 6 (3) (1997), pp. 165–177.
[3] P.A. Ralston, J.H. Graham, and S.C. Patel. Literature Review of Security and Risk Assess-
ment of SCADA and DCS Systems, T.R. TR-ISRL-06-01, Editor. Intelligent Systems Research

[email protected]
Bibliography 167

Laboratory, Dept. of Computer Engineering and Computer Science, University of Louisville:

Louisville, KY 40292. 2006, pp. 20.
[4] G.E. Apostolakis. The concept of probability in safety assessments of technological systems.
Science, 250 (4986) (1990), pp. 1359–1364.
[5] J.C. Helton. Alternative representations of epistemic uncertainty. Special Issue of Reliability
Engineering and System Safety, 85 (1-3) (2004), pp. 1–10.
[6] P. Angelov and R. Yager. A simple fuzzy rule-based system through vector membership and
kernel-based granulation. In: Proc. of 2010 IEEE Int. Conference on Intelligent Systems, pp.
349–354, 7-9 July, 2010. London, UK.
[7] J. Wang and T. Rxton, Design for safety, Journal of American Society of Safety Engineer, 1997,
pp. 387–411.
[8] K.Y. Cai. System failure engineering and fuzzy methodology: an introductory overview. Fuzzy
Sets and Systems, 83 (1996), pp. 113–133.
[9] S. Chowdhury, P. Champagne, and P.J. McLellan. Uncertainty characterization approaches for
risk assessment of DBPs in drinking water: A review. Journal of Environmental Management,
90 (5) (2009), pp. 1680–1691.
[10] J. Liu, J.B. Yang, J. Wang, H.S. Sii, and Y.M. Wang. Fuzzy rule-based evidential reasoning
approach for safety analysis. International Journal of General Systems, 33 (2004), pp. 183–
204.
[11] J. Liu, J.B. Yang, J. Wang, H.S. Sii. Engineering system safety analysis and synthesis using
fuzzy rule-based evidential reasoning approach. Quality and Reliability Engineering Interna-
tional, 21 (2005), pp. 387–411.
[12] Y.J. Liu, R. Wang, C.L. Chen. Robust adaptive fuzzy controller design for a class of uncertain
nonlinear time-delay systems. International Journal of Uncertainty, Fuzziness and Knowledge-
Based Systems, 19 (2) (2011), pp. 329–360.
[13] A.S. Markowski, M.S. Mannan, A. Kotynia, and D. Siuta. Uncertainty aspects in process safety
analysis. Journal of Loss Prevention in the Process Industries, 23 (3) (2010), pp. 446–454.
[14] S. Samson, J.A. Reneke, and M.M. Wiecek. A review of different perspectives on uncertainty
and risk and an alternative modeling paradigm. Reliability Engineering and System Safety, 94
(2) (2009), pp. 558–567.
[15] D. Ruan, J. Kacprzyk, and M. Fedrizzi. Soft Computing for Risk Evaluation and Management.
Physica-Verlag, 2001.
[16] J. Pearl. Probabalistic Reasoning in Intelligent Systems: Networks of Plausible Inference, Mor-
gan Kaufmann Publishers, 1988.
[17] C. Wu. Fuzzy reliability estimation using Bayesian approach. Computer and Industrial Engi-
neering, 46 (3) (2004), pp. 467–493.
[18] A.P. Dempster. Upper and lower probabilities induced by a multi-valued mapping. Annals
Mathematical and Statistics, 38 (2) (1967), pp. 325–339.
[19] G. Shafer. Mathematical Theory of Evidence, Princeton Press, NJ., 1976.
[20] L.A. Zadeh. Fuzzy sets. Information Control, 8 (1965), pp. 338–353.
[21] L.A. Zadeh. Fuzzy logic = computing with words. IEEE Transactions on Fuzzy Systems, 4 (2)
(1996), pp. 103–111.
[22] L.Z. Zadeh. The concept of a linguistic variable and its application to approximate reasoning.
Part 1, Information Sciences, 8 (1975), pp. 199-249, Part 2, Information Sciences, 8, pp. 301–
357, Part 3, Information Sciences, 9, pp. 43–80.
[23] J. Liu, L. Martinez, H. Wang, RM. Rodriquez, and V. Novozhilov. Computing with words in
risk assessment. International Journal of Computational Intelligence Systems, 3 (2010), pp.
396–419.
[24] A. Brandsæter. Risk assessment in the offshore industry. Safety Science, 40 (2002), pp. 231–
269.

[email protected]
168 Decision Aid Models for Disaster Management and Emergencies

[25] E. Henley and H. Kumamoto. Probabilistic Risk Assessment and Management for Engineers
and Scientists, IEEE Press. New York, 1996.
[26] M. Stamatelalos. Probabilistic Risk Assessment Procedure Guide for NASA Managers and
Practitioners. NASA Office of Safety and Mission Assurance: Washington, DC., 2002, pp.
323.
[27] Y.M. Wang, K.S. Chin, G.K.K. Poon, and J.B. Yang. Risk evaluation in failure mode and ef-
fects analysis using fuzzy weighted geometric mean. Expert Systems with Applications, 36 (2)
(2009), p. 1195–1207.
[28] S. Dean. IEC 61508-Assessing the Hazard and Risk, in S.C. Ltd, Editor.
[29] D.J. Smith and K.J.L. Simpson. Functional Safety: A Straightforward Guide to Applying IEC
61508 and related Standards. Elsevier Butterworth-Heinemann, 2004.
[30] P. Baybutt. An improved risk graph approach for determination of safety integrity levels (SILs).
Process Safety Progress, 26 (1) (2007), pp. 66–76.
[31] R. Sun. Robust reasoning: Integrating rule-based and similarity-based reasoning, Artifical In-
telligence, 75 (2) (1995), pp. 241–295.
[32] L. Pokorádi. Fuzzy Logic-Based Risk Assessment. Academic and Applied Research in Military
Science, 1 (1) (2002), pp. 63–73.
[33] I. Elishakoff and B. Ferracuti. Fuzzy sets based interpretation of the safety factor. Fuzzy Sets
and Systems, 157 (18) (2006), pp. 2495–2512.
[34] A.B. Zolotukhin and O.T. Gudmestad. Application of fuzzy sets theory in qualitative and quan-
titative risk assessment. International Journal of Offshore and Polar Engineering, 12 (4) (2002),
pp. 288–296.
[35] F. Herrera, S. Alonso, F. Chiclana, and E. Herrera-Viedma. Computing with words in decision
making: foundations, trends and prospects. Fuzzy Optimization and Decision Making, 8 (4)
(2009), pp. 337–364.
[36] M. Jablonowski. Implications of fuzziness for the practical management of high-stakes risks.
International Journal of Computational Intelligence Systems, 3 (1) (2010), pp. 1–7.
[37] K.M. Tay and C.P. Lim. On the use of fuzzy inference techniques in assessment models: part
I—theoretical properties. Fuzzy Optimization and Decision Making, 7 (2008), pp. 269–281.
[38] K.M. Tay and C.P. Lim. On the use of fuzzy inference techniques in assessment models: part
II: industrial applications. Fuzzy Optimization and Decision Making, 7 (2008), pp. 283–302.
[39] J.B. Yang and D.L. Xu. On the evidential reasoning algorithm for multiple attribute decision
analysis under uncertainty. IEEE Transactions on Systems, Man, and Cybernetics, Part A, 32
(2002), pp. 289–304.
[40] J.B. Yang, J. Liu, J. Wang, H.S. Sii, and H.W. Wang. Belief rule-base inference methodology
using the evidential reasoning approach – RIMER, IEEE Transactions on Systems, Man, and
Cybernetics, Part A, 36 (2006), pp. 266–285.
[41] J.B. Yang, J. Liu, D.L. Xu, J. Wang, and H.W. Wang. Optimization models for training belief-
rule-based systems, IEEE Transactions on Systems, Man, and Cybernetics-Part A, 37 (2007),
pp. 569–585.
[42] J. Liu, J.B. Yang, D. Ruan, L. Martinez, and J. Wang. Self-tuning of fuzzy belief rule bases for
engineering system safety analysis. Annals of Operations Research, 163 (2008), pp. 143–168.
[43] D.L. Xu, J. Liu, J.B. Yang, G.P. Liu, J. Wang, I. Jenkinson, and J. Ren, Inference and learning
methodology of belief-rule-based expert system for pipeline leak detection, Expert Systems
with Applications, 32 (1) (2007), pp. 103–113.
[44] H.S. Sii and J. Wang. A design-decision support framework for evaluation of design op-
tions/proposals using a fuzzy-logic-based composite structure methodology. Journal of Process
Mechanical Engineering, 217 (1) (2003), pp. 59–76.
[45] R. Nait-Said, F. Zidani, and N. Ouzraoui. Modified risk graph method using fuzzy rule-based
approach. Journal of Hazardous Materials, 164 (2-3) (2009), pp. 651–658.

[email protected]
Bibliography 169

[46] S. Renooi. Probability elicitation for belief networks: issues to consider, The Knowledge Engi-
neering Review, 16 (2001), pp. 255–269.
[47] D. Dubois and H. Prade. Unfair coins and necessity measures: towards a possibilistic interpre-
tation of histograms. Fuzzy Sets and Systems, 10 (1983), pp. 15–20.
[48] Y.Y. Haimes. Risk Modeling, Assessment, and Management, John Wiley & Sons Inc., 2004.
[49] M. Bohanec, B. Zupan, and V. Rajkovic. Applications of qualitative multi-attribute decision
models in health care. International Journal of Medical Informatics, 58-59 (2000), pp. 191–
205.
[50] P. Zlateva, L. Pashova, K. Stoyanov, and D. Velev. Social risk assessment from natural hazards
using fuzzy logic. International Journal of Social Science and Humanity, 1 (3) (2011), pp.
193–198.
[51] A. Calzada, J. Liu, H. Wang, L. Martinez, and A. Kashyap. An intelligent decision support
tool based on belief rule-based inference methodology. in Proc. 2011 IEEE Int. Conf. on Fuzzy
Systems (FUZZ-IEEE 2011, Taipei, Taiwan; June 27-30, 2011), pp. 2638–2643.

[email protected]
Chapter 8

Fuzzy semantics in closed domain question

answering

Mohammed-Amine Abchir, Isis Truck, and Anna Pappa

University Paris-8, 2, rue de la Liberté, 93526 Saint-Denis, France

In this chapter we describe fuzzy semantics in a disaster question answering domain.

Emergency management implies to go from reports or descriptions of the disaster into
rescue work. These descriptions are usually expressed linguistically by witnesses and must
be transformed into comprehensible data.
The aim of our approach is to achieve a multiple interest: modeling data through
(un)balanced fuzzy 2-tuples should permit to decrease the CPU time while keeping an
adequate approximation. Another interest lies in the recourse to the natural language pro-
cessing (NLP) that should offer a multi-lingual interface to non-expert users. Thanks to our
method, these users will be helped to express their diagnosis (of the disaster consequences)
which will be translated into a dataset representing the catastrophic event.

8.1 Introduction

To be able to help populations just after a disaster takes place, it is necessary to con-
ceive tools such as decision support systems for logistics of intervention in disasters and
emergencies that can forecast unknown variables as soon as possible to support the decision
making process. The aim is to be able to give clues (fatalities, injuries, homeless people,
etc.) to NGOs to estimate the magnitude of a disaster, according to quantifiable and hardly
quantifiable data. Regarding the hardly quantifiable data (e.g. a vulnerability measure of
the area), we know how imprecise and uncertain they may be, that is why fuzzy logic typ-
ically is a good framework to manage them. In this chapter two main fields are referred
to. We consider both the fuzzy 2-tuple framework [12] and the natural language processing
(NLP). Besides, ontologies coupled with NLP will provide for a general framework able to
adapt to different kinds of disaster. Since ontology defines concepts and relations between

B. Vitoriano et al. (eds.), Decision Aid Models for Disaster Management 171
and Emergencies, Atlantis Computational Intelligence Systems 7,
DOI: 10.2991/978-94-91216-74-9_8, Ó Atlantis Press 2013

[email protected]
172 Decision Aid Models for Disaster Management and Emergencies

concepts, research focuses on ontology technology involved risk management systems [18]
for acquisition, extraction and creation of specified knowledge process. Long term, the aim
is to be able to obtain ad hoc data, whatever the disaster type.
The fuzzy 2-tuple framework should be used to create semantic dependencies in both
clearly explicitly stated expressions and vague ones according to user’s diagnosis of con-
sequences of the disaster. The query processing expressed in natural language is reflect-
ing different types of real world reasoning (classical reasoning or plausible reasoning). A
grammar of affirmative, interrogative and negative sentence forms is needed and parts of
text recognition should be determine thanks to syntax. Syntax is related to semantics, con-
sequently used to disambiguate closed assigned semantic roles. However, disambiguation
meaning and plausible inference from hidden structure needs semantics contribution to
create reasonable evidence. Fuzzy semantics in this disaster management question answer
domain is how to combine incomplete or hidden information to a similar meaning words
set so we can produce a clear diagnosis of the catastrophe.
This chapter is structured as follows: Section 8.2 introduces disaster or emergency
management revising in short the different phases that have to be addressed. Section 8.3
reviews various NLP techniques that should be useful in our context while Section 8.4
presents a way to deal with imprecision in the particular context of emergency management.
First, the 2-tuple fuzzy linguistic representation model is reviewed and second, an ad hoc
method is proposed including both another 2-tuple representation model inspired by the
first one and some NLP techniques. Section 8.5 concludes this chapter.

8.2 Disaster or Emergency Management

Disaster management is an important domain for researchers since several years. When
a catastrophe happens, it is crucial to be able to obtain as clear an assessment of the situ-
ation as is possible in order to alleviate short-term suffering and to learn enough form this
assessment to prevent future hazards to become disasters. For example, the ALADDIN
(Autonomous Learning Agents for Decentralised Data and Information Networks) project
aims to develop techniques, methods and architectures for building decentralised systems
that can bring together information from a variety of heterogeneous sources in order to take
informed actions and to recover the situation (see http://www.aladdinproject.org/).
Another example lies in SEDD (a Spanish acronym for Expert System for Disaster
Diagnosis) [21] that is a decision support system to enable intervention in disasters. The

[email protected]
Fuzzy semantics in closed domain question answering 173

system aims at forecasting unknown or ill-known variables as soon as possible as the catas-
trophe has occurred. It helps non-governmental organizations (NGOs) to know or at least
to estimate the magnitude of the catastrophe through information such as reports on dam-
age, fatalities, injured, homeless population, etc.). This kind of information is very often
tainted with imprecision or incompleteness, that makes the estimation more difficult.
Experts such as Cuny defines emergency management as being composed of four
phases in a loop [6]: mitigation, prepareness, response, and recovery (see Figure 8.1).

Mitigation Preparedness

Recovery Response

Fig. 8.1 General scheme for disaster management.

Mitigation deals about prevention, i.e. the efforts that must be done in order to avoid
disasters, or in order hazards or risks don’t turn into catastrophes. Authors such as Zou
et al. have proposed linguistic approaches for risk analysis. The presented method bet-
ter expresses information, comparable and incomparable, by use of linguistic hedges and
antonyms [26]. They construct a linguistic-valued lattice implication algebra whis is suit-
able for expressing semantics of ten evaluating linguistic values. The reasoning process is
directly performed and linguistic values are given to describe the risk analysis. Due to the
imprecision of knowledge and of linguistic statements, the proposed approach deals with
fuzzy environments.
Disaster management is also often linked to climate change and sustainable develop-
ment and energy. Prevention can lead to evaluate scenarios to avoid risks. For example, a
model has been proposed for evaluating long-term sustainable energy scenarios [7]. The
selection of the energy policy is based on the decision analysis that can manage different
types of information (numerical, interval-valued and linguistic) with the aim to verbalize
the results to facilitate their understanding.
The second phase, prepareness, follows mitigation and is just after the catastrophic
event has occurred. It corresponds to the way behaviors are changed to limit the impact of
the disaster on people. This phase may have something to do with psychology and applied
behavior analysis. This is quite far away from our major issue in this chapter.

[email protected]
174 Decision Aid Models for Disaster Management and Emergencies

Response is the phase where the necessary resource is mobilized. It corresponds to the
action phase. The mobilization of the emergency services requires two steps. First select
the services that have the needed functional properties (e.g. firemen, electrician, doctors,
policemen, etc.); second, choose among those services the best one(s) (i.e. the quickest (the
most available), the most qualified, the one(s) with the best equipment, etc.). We have pro-
posed in [4] a model (the LCP-nets for Linguistic Conditional Preference Networks) that
represents preferences on a graph (actually a network) to deal with this kind of problems.
Indeed the choice between alternatives is usually complex because it depends on several
other criteria. For instance, if the service is very far away from the injured (let us say that
the injured have musculoskeletal trauma), it is preferable to have an emergency medical
transport very quickly. Otherwise it is preferable to ask for a doctor that is an orthopedic
surgeon. If no orthopedic surgeon is available, then another surgeon or a general practi-
tioner will do. This means that the preferences about a service depend on the features of
the available services.
Figure 8.2 shows an LCP-net with three nodes: D stands for the distance between the
service and the injured; A for the ambulance or medical transportation and S for surgeon
or doctor that can look after the injured. Three preference tables are attached to each node
(the glossary is given below) and a line must be understood as follows: “If D is (more or
less) DVC then the preference is (more or less) H”. When the table has two dimensions
it means that the rule has two inputs. A simple arc (e.g., between D and A) exhibits a
conditional preference (see the table attached to A), an arc with a black triangle  exhibits
an unconditional preference (see the table attached to S that does not depend on D), and an
arc with a black square  (a ci-arc) exhibits a conditional relative importance preference
(in the example, the preference on A or S depends on the value taken by node D: if D is
(more or less) DVC then optimizing S is prefered to optimizing A). Indeed the conditional
importance table attached to the ci-arc between A and S contains three lines. Each line
exhibits a relation such as X z Y where  represents a preference relation, i.e. optimizing
X is prefered to optimizing Y if Z = z.

The glossary that has been used in this example is the following:

[email protected]
Fuzzy semantics in closed domain question answering 175

DVC DC DF
89:;
?>=<
D
H M L  ???
 ??
 ??
 ??
 ??
??
 ??
 D 
89:;
?>=<
A  89:;
?>=<
S
AL AM AH
S DVC A
DVC L M H SL SH
S DC A
DC VL L VH L H
A DF S
DF VV L VL VV H
Fig. 8.2 An LCP-net for a vital supply (medicine).

DVC service is very close to the injured

DC service is close to the injured
DF service is far away from the injured
AL ambulance has low equipment (similar to a taxi)
AM ambulance has medium equipment
AH ambulance has high equipment (a real emergency medical transport)
SL “low-level” surgeon (a general practitioner)
SH “high-level” surgeon (a specialist of the disease)
VV L very very low preference
VL very low preference
L low preference
M medium preference
H high preference
VH very high preference
VV H very very high preference

This model permits to pre-compute the functional services upstream. Downstream, at

the very last moment, we can ask for the best service(s) available among the functional
ones. This system is conceived to adapt to services that are not stable (services appear and
disappear any time).
Recovery is the last phase, when it is time to restore the area to its previous state. It deals
with several kinds of tasks: restore traffic, restore electricity, supply water, but also restore

[email protected]
176 Decision Aid Models for Disaster Management and Emergencies

livelihood (fishery, agriculture, buildings with rehabilitation and reconstruction program,

etc.). This last phase is also quite far away from our major issue in this chapter.
From a computer scientist point of view, to manage such events is necessary to have a
total view on information and knowledge fusion. It is also required to have feedback from
the decisions taken (what were the consequences of such a decision? Were they what was
expected?) in order to link sensors and effectors in the system. Moreover, the objectives
that must be achieved are in particular environments: imprecision, incompleteness, ambi-
guity, uncertainty, vagueness, changing data, etc. And all the related works we have cited
share a common point: to deal with language and imprecision. This is probably the most
important thing since NGOs are composed of human beings and only human beings can or-
ganize themselves to rescue injured people. That is why we now propose natural language
processing techniques to cope with disaster management.

8.3 Using Natural Language Processing (NLP) Techniques

Relating NLP techniques to Risk management we find two kinds of relationship: the
first one is the probability approach relationship which deals with risk definitions in doc-
uments and considers the difficulty of assigning to linguistic terms the relation with risk
definition and management [14]. So we consider that the dimension of risk could be seen
as a fuzzy value between the uncertainty (something that could happen) and the effect di-
mension (the risk materialized), instead of attempting to determine meanings associated to
some particular common probability-related terms. The second relationship considers that
the probability is interpreted as a constant value in a fuzzy logic approach to risk inference
from a natural language description. Actual applications of NLP techniques like automatic
text classification, clustering and opinion extraction are related to the risk management
domain by the analysis of large collection of textual feedback data [22].
The first NLP techniques between the sixties and the eighties to deal with natural lan-
guage were the use of complex sets of hand-written rules [23, 24], but machine learning
algorithms have revolutionized NLP in introducing decision trees and statistical models
(cache language models) to try to understand and disambiguate sentences [5, 17]. Of
course, the context that is highlighted in the speech corpus is also very important in NLP.
Transcription of oral corpora is a large field of research where style (meaning the character
of the speech) is coming after spelling, grammar or syntax like the PFC (French acronym
for Phonology of Contemporary French) project which is based on orthography indexa-

[email protected]
Fuzzy semantics in closed domain question answering 177

tion [15]. Among the various tasks that are very common in NLP there are the identifi-
cation of the discourse structure, the morphological segmentation (using e.g. word stems)
and the syntactic analysis (parsing). We note that grammar component is used through-
out most automatic linguistic tools mainly of its role to specifying the acceptable phrases
in a human language. Grammar dictates syntactic rules that allow words combination for
well formed sentences. There are many research works in grammar, which inspired most
syntactic parsers. We mention some of them that influenced our work.
First, we follow Harris school for its theoretical framework, which extends grammar
so that incorrect expressions appear in a regular way [10]. The parser is used like a pre-
processor. The technique is about filling a questionnaire which assigns parts of a phrase in
table columns. Although this kind of tree analysis is restrictive in knowledge representa-
tion, it is easier to convert to an intermediate treatment like transfer to query language of
database.
Second, we adapt unification grammars in a syntactic model with semantic features.
It’s interesting to mention that all formalisms proposed by functional unification grammars
are computational adaptive. Most linguistic description paradigms are directly encoded
in a feature/value structure. These structures were proposed by computer scientists like
generalized mechanisms of data type knowledge representation, and were largely used in
language comprehension and speech understanding. Semantic grammar is a free context
grammar in which the choice of the non terminal and the rule production are governed by
an altogether syntactic and semantic function. But in these functions is quite difficult to
describe linguistic patterns like deletions (unspecified nouns and verbs, lack or referential
index), distortions (complex equivalence, presuppositions) or generalizations like universal
quantifiers and modal operator of necessity or possibility.
Semantics comes just after, not because the meaning is not the goal of any phrase anal-
ysis but because we need all the elements of grammar and syntax altogether so we could
determine the best “machine” comprehension or interpretation if translation is operated.
Therefore, there is an order of different operations leading to “understanding”; in the first
place, language recognition; is it about English or French? Does this word belong to En-
glish dictionary or another one? Second, the words’ order or syntax, which confirms the
language choice and sets the rules for further semantic analysis. For instance, in some
languages like French we have both syntactical and semantic opposition: the place of the
adjective in the phrase illustrates this opposition, adjective could precede or follow the

[email protected]
178 Decision Aid Models for Disaster Management and Emergencies

noun. In English this redundancy is removed, but other problems still remain like words’
polysemy as written below.
To make discourse analysis, part-of-speech (PoS) tagging — which is a grammatical
tagging [8], i.e. algorithms which associate terms in accordance with a set of descriptive
tags — can be used to help disambiguating words (e.g. “close” can be an adjective, a
noun, an adverb, or a verb) [10]. This kind of ambiguity is common in languages with
little inflectional morphology because there is a lot of such words. When analyzing a text,
sequences of tokens are defined to determine its structure according to a grammar, for
instance noun inflexion for singular or plural number. Inflected forms of words express
different grammatical categories, e.g. tense, conjugation, person, gender, etc. However
these techniques permit to “understand” sentences without ambiguity in a closed domain
context but they don’t consider any imprecision or vagueness in the meaning.

8.4 How to deal with imprecision in our context?

The first approaches to deal with this uncertainty come from Zadeh when he introduced
in 1965 the fuzzy set theory, the fuzzy logic and the concept of linguistic variables [25]. The
fuzzy sets could be employed to integrate vagueness throughout the relational structure of
meaning including both the concept of structure and reference that a term denotes. Experi-
mental psychology demonstrated the difference that exists between what people ’would do’
under some supposed conditions and what people do when the conditions are real. Some
NLP techniques can help to identify these language regularities which real speakers follow
and/or establish in discourse or short texts empirically [20].
In practice, how to deal with imprecision in disaster management? First of all an ad hoc
lexicon is necessary, dedicated to emergency management. There exist several lexicons in
English that gather typical terms used for disaster description and management, see e.g. [3,
16]. As an example in this chapter, we have chosen the vocabulary about a hurricane, i.e.
the Beaufort scale, and the wave height. Richer than the lexicons are the ontologies that
formally represent not only knowledge as a set of concepts within a domain but also the
relationships between those concepts. The consequence of this richer representation is that
it can be used to reason about the entities within that domain. In theory, an ontology is a
“formal, explicit specification of a shared conceptualisation” [9].
Second, a convenient model is necessary to represent the entities.

[email protected]
Fuzzy semantics in closed domain question answering 179

8.4.1 The 2-tuple fuzzy linguistic representation model

We quickly review the fuzzy linguistic 2-tuple model that we were inspired by. Since
1965, many models have been proposed but one seems the most appropriate in our case: the
2-tuple fuzzy linguistic model [12] because it deals with words and uses a simple internal
representation of them.
The idea is to deal only with words or linguistic expressions in translating them into a
linguistic pair (si ,α ) where si is a triangular-shaped fuzzy set and α a symbolic translation,
both representing a linguistic term. It is required that in the linguistic term set there exist:
(1) a negation operator: Neg(si = s j ) such that j = g − i (g + 1 is the cardinality), (2) an
order: si  s j ≡ i  j. Therefore, there exists a min and a max operator. The semantics of
the terms are given by fuzzy numbers defined in the [0, 1] interval.
If α is positive then si is reinforced otherwise si is weakened. This model is also in-
teresting because it permits to reduce CPU time: we free ourselves from the fuzzy sets
— at least in the computations — while keeping a linguistic semantics attached to the
data. Indeed, the 2-tuple is composed of two floating-point numbers instead of a mem-
bership function (represented by equations of the first degree or higher-degree equations)
that consumes more CPU resources. Besides, it is necessary to keep the linguistic terms
in the reasoning process because the estimation of the disaster magnitude is performed for
humans (from NGOs).
If the information is perfectly balanced (i.e. the distance between words is exactly
the same, then all the si values are equally distributed on the axis). But if not — that
may happen when talking about distance, for instance, “at the same place” and “close to”
are closer to each other than “near” and “far away” — the si values may not be equally
distributed on the axis. That is why another model has been proposed by the same team
to deal with such information that they call multigranular linguistic information [13]. To
do this, they use linguistic hierarchies composed of an odd number of triangular fuzzy sets
of the same shape, equally distributed on the axis, as a fuzzy partition. A word may have
one or two linguistic herarchy(ies) and the representation of a notion (such as “distance”)
can be made up of several hierarchies. Thus the fuzzy sets obtained are still triangular-
j
shaped but not always isosceles triangular-shaped and the notation si permits to keep both
the hierarchy and the linguistic term. See [19] for a deeper review of these models and [11]
for a methodology to deal with unbalanced linguistic term sets.

[email protected]
180 Decision Aid Models for Disaster Management and Emergencies

8.4.2 An ad hoc method to deal with disaster management

However, in recent papers, we have shown that despite its advantage, the 2-tuple model
or unbalanced linguistic term sets doesn’t fit our needs perfectly especially when one (or
more) linguistic expression is very far away from its next neighbor [1, 2]. The new model
we have proposed fully takes advantage of the symbolic translations α that become a very
important element to generate the data set. Our 2-tuples are twofold. Indeed, except the
first one and the last one of the partition, they all are composed of two half 2-tuples: an
upside and a downside 2-tuple. The choice of our 2-tuple model is relevant since the lin-
guistic terms used in the emergency management, such as geolocation context are usually
unbalanced.
The methodology we have used to deal with imprecision inside the natural language
is inspired by the PoS recognition and the PoS tagging. We simplify the analysis using
a semantic tagging because the context (disaster management) is known. As an exam-
ple, let us take the following preference (see also Figure 8.2 where an LCP-net about
the same example is given): “If the injured are far away, I prefer to optimize the qual-
ity of the ambulance than the specialized medical field. Otherwise, it is the contrary”.
Tokens are processed (classified and recognized) using NLP techniques detailed above
(see section 8.3) and ontologies. A first tentative of ontology has been proposed for
disaster management, where each abstracted object is identified through a URI (Uni-
fied Resource Identifier) or more precisely a URN (Unified Resource Name), such as
urn:univ-paris8:disaster:medicalEmergency:medicalEquipment:ambulance
or urn:univ-paris8:disaster:affectedPopulation:injured.

Then tokens are tagged in an XML file, such as the following one:

<tokens>
<token gram="CONJ" sem="COND_PREF">If</token>
<token gram="DET">the</token>
<token gram="NOUN" sem="ANTECEDENT">injured</token>
<token gram="VERB">are</token>
<token gram="ADV" sem="FUZZY_DIST">far away</token>
<token gram="ADV">I</token>
<token gram="VERB">prefer</token>
<token gram="VERB">to optimize</token>
<token gram="DET">the</token>

[email protected]
Fuzzy semantics in closed domain question answering 181

<token gram="NOUN">quality</token>
<token gram="PREP">of</token>
<token gram="DET">the</token>
<token gram="NOUN" sem="THEN_CONSEQUENT">ambulance</token>
...
<token gram="NOUN" sem="ELSE_CONSEQUENT">contrary</token>
</tokens>

A tree using a simplified tree-adjoining grammar (TAG)-based is then created, where

each leaf node represents the semantic tag of a token from the lexicon, e.g. in Relax NG
compact syntax:

PREF=TYPE,ANTECEDENT,THEN_CONSEQUENT,ELSE_CONSEQUENT?
TYPE=COND_PREF|UNCOND_PREF
ANTECEDENT=FUZZY_MODIFIER?,FUZZY_VAR
FUZZY_VAR=FUZZY_DIST|FUZZY_TIME|...
...
THEN_CONSEQUENT=FUZZY_MODIFIER?,FUZZY_VAR
...

Upstream a fuzzy partition with our 2-tuple model has been performed by experts
for each imprecise piece of data. For example, the expert gives 5 terms: A, B, C, D,
E that are by default uniformly distributed on their axis. But, looking for synonyms (e.g.
http://www.crisco.unicaen.fr/cgi-bin/cherches.cgi that exhibits a French dic-
tionary) and the distance between the terms, our algorithm permits to obtain 5 synonym
bags. Resemblance rates between bags are then computed in order to obtain new positions
of A, B, C, D, E on the axis. Downstream the semantic tokens (e.g. “close”) are expressed
through 2-tuples and compared to the fuzzy partition. It is to notice that adverbs such as
“very” modify their associated 2-tuple through the α value. In the following we propose a
modelling of a use case with two different fuzzy models.

The use case we have chosen is about an alert when wave height becomes too big
(Beaufort scale). This scale admits 12 levels that are defined through a Beaufort number, a
linguistic description, the wind speed, the wave height, the sea conditions and/or the land
conditions (see e.g http://en.wikipedia.org/wiki/Beaufort_scale):

[email protected]
182 Decision Aid Models for Disaster Management and Emergencies

• 0: calm
• 1: light air
• 2: light breeze
• 3: gentle breeze
• 4: moderate breeze
• 5: fresh breeze
• 6: strong breeze
• 7: strong wind
• 8: gale
• 9: strong gale
• 10: storm
• 11: violent storm
• 12: hurricane

The software has been written in Java, using jFuzzyLogic with the extension we have
proposed [2] that implements the FCL (Fuzzy Control Language) specification (IEC 61131-
7). We remind that FCL is a simple language, actually a “control language” to define a
fuzzy inference system. Its main concept is a “control block” which has some input and
output variables.
The FCL script below uses three different partitions for the input Distance.

FUNCTION_BLOCK Hurricane
VAR_INPUT Wave_Height_V1:LING; Wave_Height_V2:LING;
END_VAR
VAR_OUTPUT Alert_Level:REAL;
END_VAR
FUZZIFY Wave_Height_V1
TERM S := ling Calm Light_Air Light_Breeze Gentle_Breeze Moderate_Breeze
| Fresh_Breeze Strong_Breeze Strong_Wing
| Gale Strong_Gale Storm Violent_Storm Hurricane, extreme extreme;
END_FUZZIFY
FUZZIFY Wave_Height_V2
TERM S := pairs (Calm, 0.0) (Light_Air, 0.2)
(Light_Breeze, 0.5) (Gentle_Breeze, 1.0)
[...]

[email protected]
Fuzzy semantics in closed domain question answering 183

(Hurricane, 30.0);
END_FUZZIFY
DEFUZZIFY Alert_Level // 5 linguistic terms
TERM Min := trian 0.0 0.0 5.0;
[...]
TERM Max := trian 85.0 100.0 100.0;
METHOD : LM; // ’Center Of Gravity’
END_DEFUZZIFY
RULEBLOCK Rules
RULE 1: IF Wave_Height_V1 IS Calm THEN Alert_Level IS Min;
...
END_RULEBLOCK
END_FUNCTION_BLOCK

Figure 8.3 and Figure 8.4 show both partitions for the input: the partition generated
by the linguistic 2-tuple model from Herrera et al. (Wave_Height_V1) and our 2-tuple
partition (Wave_Height_V2). The comparison between both models permits to show that
our partition fits better reality while keeping a bigger unbalancement than the first one.
What is interesting to keep in mind is that with these models, any linguistic assessment
can be modelled (when there is much unbalancement, our model fits better). In emergency
management such models are thus very useful.

[email protected]
184 Decision Aid Models for Disaster Management and Emergencies

Fig. 8.3 Partition with Herrera et al. linguistic 2-tuple model.

[email protected]
Fuzzy semantics in closed domain question answering 185

Fig. 8.4 Partition with our linguistic 2-tuple model.

[email protected]
186 Decision Aid Models for Disaster Management and Emergencies

A diagram to illustrate the overall technical framework of our proposal is given in

Figure 8.5.

dialogue, sentences,
interface, preferences,
questionnaire... linguistic descriptions...

PoS recognition ad hoc

and tagging ontologies

tokens with semantics

tree−adjoining
grammar

TAG−based tree
with rules for
the reasoning

ad hoc synonyms dictionary,

2−tuples ontologies

Variables represented
by linguistic 2−tuples

use of
FCL

‘‘Business’’ rules

reasoning process

Decision

Fig. 8.5 General scheme of our proposal.

8.5 Conclusions

In this chapter we have presented principles of disaster management that is composed

of four phases where knowledge is expressed through natural language. As usual in such

[email protected]
Bibliography 187

cases, imprecision and imperfection taint the knowledge that let the decision harder to
make. The discussions and methods we propose aim at reducing imprecision in taking into
account it through a 2-tuple representation model coupled with natural language process-
ing techniques. A linguistic conditional preference network paradigm described in previous
works is also shown to be efficient in such situations. Indeed two in four phases are ad-
dressed in the paper and the proposed methods have shown to be useful. Future works will
have to cope with a best understanding of the natural language and with the definition of
more dedicated ontologies.

Acknowledgement

This work is partially funded by the French National Research Agency (ANR) under
grant number ANR-09-SEGI-012.

Bibliography

[1] Abchir, M. (2011). A jFuzzyLogic Extension to Deal With Unbalanced Linguistic Term Sets,
Book of Abstracts, pp. 53–54.
[2] Abchir, M. and Truck, I. (2011). Towards a New Fuzzy Linguistic Preference Modeling Ap-
proach for Geolocation Applications, in Proc. of the EUROFUSE Workshop on Fuzzy Methods
for Knowledge-Based Systems, pp. 413–424.
[3] Blanchard, W. (2008). Guide to emergency management and related terms, definitions, con-
cepts, acronyms, organizations,
programsn guidance, executive orders & legislation, http://training.fema.gov/EMIWeb/
edu/docs/terms\%20and\%20definitions/Terms\%20and\%20\\Definitions.pdf.
[4] Châtel, P., Truck, I. and Malenfant, J. (2010). LCP-Nets: A Linguistic Approach for Non-
functional Preferences in a Semantic SOA Environment, J. UCS 16, 1, pp. 198–217.
[5] Chomsky, N. (1986). Knowledge of language, its nature, origin, and use, (Praeger, New York).
[6] Cuny, F. (1983). Disasters and development (Oxford University Press, New York).
[7] Espinilla, M., Ruan, D., Liu, J. and Martínez, L. (2010). A heterogeneous evaluation model for
assessing sustainable energy: A Belgian case study, in FUZZ-IEEE, pp. 1–8.
[8] Green, B. and Rubin, G. (1971). Automated Grammatical Tagging of English, Tech. rep., De-
partment of Linguistics, Brown University.
[9] Gruber, T. R. (1993). A Translation Approach to Portable Ontology Specifications, Knowledge
Acquisition 5, 2, pp. 199–221.
[10] Harris, Z. S. (1991). A theory of language and information : a mathematical approach / Zellig
Harris (Clarendon Press ; Oxford University Press, Oxford [England]: New York).
[11] Herrera, F., Herrera-viedma, E. and Martínez, L. (2008). A fuzzy linguistic methodology to
deal with unbalanced linguistic term sets, IEEE Transactions on Fuzzy Systems , pp. 354–370.
[12] Herrera, F. and Martínez, L. (2000). A 2-tuple fuzzy linguistic representation model for com-
puting with words, IEEE Transactions on Fuzzy Systems 8, 6, pp. 746–752.

[email protected]
188 Decision Aid Models for Disaster Management and Emergencies

[13] Herrera, F. and Martínez, L. (2001). A model based on linguistic 2-tuples for dealing with
multigranular hierarchical linguistic contexts in multi-expert decision-making, IEEE Transac-
tions on Systems, Man, and Cybernetics, Part B 31, 2, pp. 227–234.
[14] Hillson, D. A. (2005). Describing probability : The limitations of natural language dimensions
of risk, Risk Management, pp. 1–7.
[15] J. Durand, J.-M. T. (2006). Pfc, corpus et systèmes de transcription, Cahiers de Grammaire 30,
pp. 139–158.
[16] Kelman, I. (2008). Disaster lexicon. version 7, 19 may 2008 (version 1 was 10 january 2007),
http://www.ilankelman.org/miscellany/DisasterLexicon.rtf.
[17] Kuhn, R. and Mori, R. D. (1990). A cache-based natural language model for speech recognition,
IEEE Transactions on Pattern Analysis and Machine Intelligence 12, 6, pp. 570–583.
[18] Makki, J., Alquier, A. and Prince, V. (2009). Ontology Population via NLP techniques in Risk
Management, International Journal of Humanities and Social Sciences 3, 3, pp. 210–217.
[19] Martínez, L., Ruan, D. and Herrera, F. (2010). Computing with words in decision support
systems: An overview on models and applications, International Journal of Computational
Intelligence Systems 3, 4, pp. 382–395.
[20] Rieger, B. B. (1979). Fuzzy structural semantics. on a generative model of vague natural lan-
guage meaning, in Trappl/ Hanika/Pichler (Eds.): Progress in Cybernetics and Systems Re-
search, Vol. V, New York/London/Sydney, pp. 495–503.
[21] Rodríguez, J. T., Vitoriano, B. and Montero, J. (2010). A natural-disaster management DSS for
Humanitarian Non-Governmental Organisations, Knowledge-Based Systems 23, 1, pp. 17–22,
special Issue on Intelligent Decision Support and Warning Systems.
[22] Tulechki, N. (2011). Des outils de TAL en support aux experts de sûreté industrielle pour
l’exploitation de bases de données de retour d’expérience, in TALN, 12 p.
[23] Weizenbaum, J. (1966). Eliza — a computer program for the study of natural language com-
munication between man and machine, Communications of the ACM 9, 1, pp. 36–45.
[24] Winograd, T. (1972). Procedures as a representation for data in a computer program for under-
standing natural language, Cognitive Psychology 3, 1, pp. 1–191.
[25] Zadeh, L. A. (1965). Fuzzy sets, Information Control 8, pp. 338–353.
[26] Zou, L., Ruan, D., Pei, Z. and Xu, Y. (2011). A Linguistic-Valued Lattice Implication Algebra
Approach for Risk Analysis, Multiple-Valued Logic and Soft Computing 17, 4, pp. 293–303.

[email protected]
Chapter 9

Decision Making with Extensions of Fuzzy Sets:

An Application to Disaster Management

H. Bustince, M. Pagola, E. Barrenechea, C. Lopez-Molina, J. Fernandez

Universidad Publica de Navarra

Email: {bustince,miguel.pagola,edurne.barrenechea,carlos.lopez,
fcojavier.fernandez}@unavarra.es

In this chapter we propose an extension of the weighted voting method to the interval-
valued setting. In this way, we allow the use of interval-valued information in those sit-
uations in which the lack of knowledge prevents the expert from providing more precise
information. We also provide some simple examples of our method.

9.1 Introduction

A disaster can be defined as as a situation in which the capacities and means of a given
group to recover the previous equilibrium are overcome, in such a way that survival of the
considered group is seriously compromised.
The literature on expert systems or decision making in disaster management is wide,
starting from the 80s last century Wallace and De Balogh [26]. Nevertheless, the different
approaches that have been considered display common problems. One of the most relevant
ones is that they are not suitable to be applied to any disaster at any place Aleskerov et al.
[1], Mendonca et al. [22]. In order to reach such a solution, we shall first analyze the
common characteristics of disaster situations. In this work we focus on how to model the
potential inaccuracy of the information the systems are fed with. In this sense, we do not
pretend to provide a general solution, but rather an improvement of an already existing
solution that can be applied to more cases.
Automatic disaster management is a complicated goal that is likely to require a long
research before being solved, especially if we intend to generate scenario-in pendent solu-

B. Vitoriano et al. (eds.), Decision Aid Models for Disaster Management 189
and Emergencies, Atlantis Computational Intelligence Systems 7,
DOI: 10.2991/978-94-91216-74-9_9,  Atlantis Press 2013

[email protected]
190 Decision Aid Models for Disaster Management and Emergencies

tions. However, automatic systems become of paramount importance to solve the problems
associated with human decision-based strategies, whose performance is severely affected
by factors such as stress and death risk (e.g. crowd disasters Helbing et al. [16], Krausz
and Bauckhage [21]). A initial step can be the development of algorithms that do not de-
pend closely on the different situations Mendonca et al. [22] but can be applied to other
simple situations. In this, sense one of the most widely used algorithms in decision making
is the weighted voting, mainly due to its ability to adapt to rather different situations that
require to choose an alternative between several of them Huellermeier and Brinker [17].
In particular, this algorithm only requires a fuzzy preference relation to be constructed, in-
dependently of the way in which it is created Huellermeier and Vanderlooy [18]. In some
situations, however, this algorithm may not be able to provide a single alternative as an
answer, due, for instance, to the lack of precise knowledge by the experts Sanz et al. [23].
The use of extensions of fuzzy sets to overcome this situation has a long history Burillo
and Bustince [5, 6], Chen and Tan [10], Tan and Chen [25]. In this way, in the present
chapter we propose the use of interval-valued fuzzy sets as a mean to solve this difficulty
in some particular situations.
Our goal is to extend the weighted voting method to an interval-valued setting in such a
way that our proposal is able to decide between several alternatives in some cases in which
the usual weighted voting method is not able to do so. Of course, we are not providing
a universally working algorithm, but rather a hint on how the use of interval-valued fuzzy
sets can help to improve the behaviour of classical decision making algorithms at least in
some particular cases. On the other hand, our approach is quite simple, so we expect that
our generalization can be further extended or adapted. Notice that along the text we will
freely use the mathematical equivalence between Atanassov’s intuitionistic fuzzy sets and
interval-valued fuzzy sets Atanassov [2], Deschrijver and Kerre [13]
The structure of this is work is as follows. In the next Section we recall some prelimi-
nary definitions. Next, in Section 9.3 we present a construction method of interval-valued
fuzzy preference relations from fuzzy preference relations. In Section 9.4 we propose two
algorithms for solving decision-making problems, and in Section 9.5 we illustrate them
with some examples. Finally, conclusions are presented in Section 9.7.

[email protected]
Decision Making with Extensions of Fuzzy Sets 191

9.2 Preliminary concepts

To start with, we recover some well-known concepts that are necessary for our subse-
quent developments. The first step is to recall the definition of a fuzzy set.

Definition 9.1 (Zadeh [29]). A fuzzy set A defined on a non-empty universe U is given
by
A = {(ui , μA (ui )) | ui ∈ U}.

F S (U) will denote the set of all fuzzy sets over the referential U. We know that a
binary fuzzy relation R is a fuzzy subset of the Cartesian product U ×U, that is R : U ×U →
[0, 1]. We denote by F R(U × U) the set of all fuzzy relations over U × U.
In fuzzy theory, a function N : [0, 1] → [0, 1] is a strict negation if N is strictly decreas-
ing, continuous, N(0) = 1 and N(1) = 0. If N is also involutive (i.e., N(N(x)) = x for all
x ∈ [0, 1]), then N is called a strong negation.

Definition 9.2. An automorphism of the unit interval is a continuous, strictly increasing

function ϕ : [0, 1] → [0, 1] with ϕ (0) = 0 and ϕ (1) = 1.

A triangular norm (t-norm for short) T : [0, 1]2 → [0, 1] is an associative, commutative,
non-decreasing function such that T (1, x) = x for all x ∈ [0, 1]. Examples of t-norms are
the minimum t-norm (TM (x, y) = min(x, y)) and the product t-norm (TP (x, y) = x · y). A
triangular conorm (t-conorm for short) S : [0, 1]2 → [0, 1] is an associative, commutative,
non-decreasing function such that S(0, x) = x for all x ∈ [0, 1]. Examples of t-conorms
are the maximum t-conorm (SM (x, y) = max(x, y)) and the probabilistic sum t-conorm
(SP (x, y) = x + y − x · y).
The associativity allows to extend each t-norm T and t-conorm S in a unique way to an
n-ary operation in the usual way by induction, defining for each n-tuple (x1 , x2 , . . . , xn ) ∈
[0, 1]n (see page 9 of Klement et al. [20]).
n n−1
T xi = T T xi , xn = T (x1 , x2 , . . . , xn )
i=1 i=1
n n−1
S xi = S S xi , xn = S(x1 , x2 , . . . , xn )
i=1 i=1
Therefore, a t-norm T or a t-conorm S can take a different number of arguments.
In this chapter we are going to deal with interval-valued fuzzy sets.Let’s consider the
set

L([0, 1]) = x = [x, x] | 0  x  x  1

[email protected]
192 Decision Aid Models for Disaster Management and Emergencies

Given an interval x ∈ L([0, 1]), one of its characteristic features is its length, given by
W (x) = x − x.

Definition 9.3. An interval-valued fuzzy set (IVFS) over the finite referential set U is an
expression A given by

A = { (ui , [A(ui ), A(ui )]) | ui ∈ U

{
with [A(ui ), A(ui )] ∈ L([0, 1]) for all ui ∈ U.

When we consider an interval-valued membership, its length can be understood as a

measure of the lack of knowledge of the expert in assigning the membership of the con-
sidered element Bustince et al. [9]. We denote by I V F S (U) the set of all IVFS over
the referential set U. Notice that any fuzzy set A = {(ui , μA (ui )) | ui ∈ U} can be seen as
the interval-valued fuzzy set A = {(ui , [μA (ui ), μA (ui )]) | ui ∈ U}; that is, a fuzzy set can be
understood as an interval-valued fuzzy set such that the memberships of all the elements in
the referential set are given by intervals of length equal to zero.
In L([0, 1]) we consider the following partial order: For any x, y ∈ L([0, 1]), x L y if
x  y and x  y. With this order L([0, 1]) becomes a complete lattice with infimum given
by 0L = [0, 0] and supremum given by 1L = [1, 1].
This order can be extended to a total one as follows Xu and Yager [27]. Let x, y ∈
L([0, 1]). We say that x < y if and only if

• score(x) < score(y)

• score(x) = score(y) and accuracy(x) < accuracy(y)

where score(x) = x + x − 1 and accuracy(x) = x − x.

Let’s consider now the extension of the usual fuzzy theory concepts to the setting of
IVFSs.

Definition 9.4 (Bustince et al. [7], Deschrijver et al. [12]). An interval-valued fuzzy
negation is a function N : L([0, 1]) → L([0, 1]) that is decreasing (with respect to L ) such
that N(0L ) = 1L and N(1L ) = 0L . If for all (x ∈ L([0, 1]), N(N((x)) = x it is said that N is
involutive.

An interval-valued fuzzy binary relation R� Burillo and Bustince [5, 6] is defined as

an interval-valued fuzzy subset of U × U, that is, R� : U × U → L([0, 1]). We denote by
I V F R(U × U) the set of all interval-valued fuzzy binary relations (IVFR)on U.

[email protected]
Decision Making with Extensions of Fuzzy Sets 193

Following the ideas of Gerstenkorn and Manko [15], for A, B ∈ I V F S (U), its cor-
relation can be defined as

C(A, B) = A(ui )B(ui ) + (1 − A(ui ))(1 − B(ui )) (9.1)

ui ∈U

Notice that the value provided by this correlation is a real number.

9.2.1 Implication operators

In this work we consider a fuzzy implication operator as an implication in the sense

of Fodor and Roubens [14].

Definition 9.5. An implication is a function I : [0, 1]2 → [0, 1] that satisfies the following
properties:

(I1) x  z implies I(x, y)  I(z, y) for all y ∈ [0, 1].

(I2) y  t implies I(x, y)  I(x,t) for all x ∈ [0, 1].
(I3) I(0, x) = 1 for all x ∈ [0, 1] (dominance of falsity).
(I4) I(x, 1) = 1 for all x ∈ [0, 1].
(I5) I(1, 0) = 0.

Remark 9.1. Properties (I3), (I4) and (I5) imply that I is an extension of the standard
Boolean implication. Indeed, it holds that I(0, 0) = I(0, 1) = I(1, 1) = 1 and I(1, 0) = 0.

Depending on the application, the following properties can also be demanded:

(6) I(1, x) = x (neutrality of truth).

(7) I(x, I(y, z)) = I(y, I(x, z)) (exchange property).
(8) I(x, 0) = N(x) where N is a strong negation.
(9) I(x, y)  y.
(10) I is a continuous function (continuity).

A wide study of these properties can be found in Baczynski and Jayaram [4], Bustince
et al. [8].

[email protected]
194 Decision Aid Models for Disaster Management and Emergencies

9.3 Construction of Interval-valued Fuzzy Preference Relations from Fuzzy

Preference Relations

Let us consider the problem of constructing interval-valued fuzzy preference rela-

tions from fuzzy preference relations. We start introducing a way of building elements
in L([0, 1]) from two elements in [0, 1] such that

• the first one represents the degree of preference given by the expert of one alternative
over another;
• the second one represents the ignorance of the expert in the assignation of the first
value.

The construction method of elements in L([0, 1]) is described in Theorem 9.1.

Theorem 9.1. Let Fμ : [0, 1]2 → [0, 1] such that

(F1) Fμ (x, 0) = x for all x ∈ [0, 1];

(F2) Fμ (x, y) is increasing in the first argument;
(F3) Fμ (x, y) is decreasing in the second argument;
(F4) Fμ (1, x) = 1 − x for all x ∈ [0, 1].

In these conditions the following items hold:

(1) Fμ (x, y)  x for every x, y ∈ [0, 1];

(2) Fμ (x, 1) = 0 for every x ∈ [0, 1];
(3) Fμ (0, x) = 0 for every x ∈ [0, 1];
(4) The mapping F : [0, 1]2 → L([0, 1]) given by

F(x, y) = [Fμ (x, y), Fμ (x, y) + y]

is such that W (F(x, y)) = y for all x, y ∈ [0, 1], where W denotes the length of the
interval.

Proof. (1) By (F1), (F3) x = Fμ (x, 0)  Fμ (x, y). (2) By (F2), (F4) Fμ (x, 1)  Fμ (1, 1) =
0. (3) By (F1), (F3) Fμ (0, x)  Fμ (0, 0) = 0. (4) By (F2), (F4) it is clear that Fμ (x, y)+y  1.
By construction, we have that W (F(x, y)) = y. 

Example 9.1.
y y
max 0, x − if x + 1
Fμ (x, y) = 2 2
1−y otherwise

[email protected]
Decision Making with Extensions of Fuzzy Sets 195

Next we propose a method for constructing Fμ functions from implication operators in

such a way that we can use the different construction methods of implication operators that
appear in the literature.

Lemma 9.3.1. Let Fμ : [0, 1]2 → [0, 1] satisfying (F1)–(F4). Then, the function

I : [0, 1]2 → [0, 1] given by

I(x, y) = N(Fμ (x, y))

is an implication operator for all negation N.

Proof. (I1) and (I2) follow from (F2) and (F3) respectively. (I3) I(0, x) = N(Fμ (0, x)) =
N(0) = 1. (I4) I(x, 1) = N(Fμ (x, 1)) = N(0) = 1. (I5) I(1, 0) = N(Fμ (1, 0)) = N(1) = 0
from (F4). 

Theorem 9.2. A function Fμ : [0, 1]2 → [0, 1] satisfies properties (F1)-(F4) if and only if
there exists a function I : [0, 1]2 → [0, 1] satisfying (I1), (I2), (I6) and (I9) with respect to
the standard negation such that

Fμ (x, y) = 1 − I(x, y). (9.2)

Proof. (Sufficiency) (I1), (I2) by (F2), (F3). (I6) I(1, x) = 1 − Fμ (1, x) = 1 − (1 − x) = x

by (F4). (I9) I(x, 0) = 1 − Fμ (x, 0) = 1 − x by (F1).
(Necessity) Fμ (x, 0) = 1 − I(x, 0) = x by (I9) and with N the standard negation. (F2), (F3)
by (I1), (I2). Fμ (1, x) = 1 − I(1, x) = 1 − x by (I6). 

Remark 9.2. From Theorems 9.1 and 9.2, F : [0, 1]2 → L([0, 1]) is given by

F(x, y) = [1 − I(x, y), 1 − I(x, y) + y].

For F to be well-constructed, I must satisfy I(x, y)  y (I10). Bustince et al. [8] proved
that if I satisfies (I1), (I6) then I also satisfies (I10).

Example 9.2.

• If taking Reichenbach implication I(x, y) = 1 − x + xy then

F(x, y) = [x(1 − y), 1 − (1 − x)(1 − y)].

• If taking Lukasiewicz implication I(x, y) = min(1, 1 − x + y) then

F(x, y) = [max(0, x − y), max(x, y)].

[email protected]
196 Decision Aid Models for Disaster Management and Emergencies

• If taking Kleene-Dienes implication I(x, y) = max(1 − x, y) then

F(x, y) = [min(x, 1 − y), max(0, 1 − x − y)].

Corollary 9.1. In the setting of Theorem 9.1 and 9.2, if I is strictly monotone and also
satisfies (I7) and (I13), then there exists an automorphism in the unit interval such that

Fμ (x, y) = ϕ −1 (ϕ (x) − ϕ (x)ϕ (y)). (9.3)

Proof. Direct taking into account that I(x, y) = N(ϕ −1 (ϕ (x)ϕ (N(y)))), as shown
in Bustince et al. [8]. 

Remark 9.3. Notice that for Expression (9.3) to satisfy (F4) it is necessary to take N(x) =
1 − x. So Fμ (x, y) = ϕ −1 (ϕ (x)ϕ (1 − y)).

Example 9.3. If taking ϕ (x) = x then

F(x, y) = [x(1 − y), x(1 − y) + y].

9.4 Two generalizations of the weighted voting strategy

In this section we apply our previous developments for presenting two new decision-
making algorithms that generalize the weighted voting strategy (see Subsection 9.4.1) and
that make use of interval-valued fuzzy preference relations. Our starting point will be a
finite set of alternatives, X = {x1 , x2 , . . . , xn } and we assume that we have homogenized the
information about the preferences into a FPR.
Recall that a fuzzy preference relation R on X is a fuzzy binary relation R : X × X →
[0, 1]. The value R(xi , x j ) = Ri j denotes the degree to which alternative xi is preferred to
alternative x j Chiclana et al. [11].
In a similar way, an interval-valued fuzzy preference relation R� on X is defined as R� :
X × X → L∗ . The interval-valued membership value R� (xi , x j ) = R�i j = [R� i j , R� i j ] provides
an upper and a lower bound for the “actual” value of the preference of alternative xi over
alternative x j .
In order to deal also with fuzzy data, in both of our proposed algorithms we start from a
FPR and construct an IVFPR taking into account the ignorance of the expert when provid-
ing that FPR. The concept of weak ignorance is defined by Sanz et al. [23] as a measure
of the lack of knowledge that the expert suffers in the assignation of a numerical value to
the membership of an element to a set (see also Bustince et al. [9]).

[email protected]
Decision Making with Extensions of Fuzzy Sets 197

Definition 9.6. A function g : [0, 1] → [0, 1] is called a weak ignorance function if it satisfies
the following conditions:

(g1) g(x) = g(1 − x) for all x ∈ [0, 1];

(g2) If x = 0.5 then g(x) = 1;
(g3) g(x) = 0 if and only if x = 0 or x = 1.

Next we justify the three properties demanded to weak ignorance functions.

• Based on the additive reciprocal property of FPR, that is, Ri j + R ji = 1, the ignorance
associated to the preference Ri j must be the same as the ignorance associated to R ji , so
g(Ri j ) = g(1 − Ri j ) = g(R ji ).
• If an expert assigns the value 0.5 to his/her preference of alternative xi over x j , we
understand that he/she is not able of choosing between both alternatives and hence
his/her ignorance is maximal.
• We consider that the expert’s ignorance is minimal when he/she is absolutely sure of
his/her preference of one alternative over another. That is, if Ri j = 1 or Ri j = 0.

In Theorem 9.3 we present a construction method of weak ignorance functions from

t-norms.

Theorem 9.3. Let T be a t-norm such that T (x, y) = 0 if and only if x · y = 0 and without
zero divisors. Then the function
T (x, 1 − x)
if T (x, 1 − x)  T (0.5, 0.5)
T (0.5, 0.5)
g(x) =
T (0.5, 0.5)
otherwise
T (x, 1 − x)
is a weak ignorance function.

Proof. Direct. 

Example 9.4. If taking T = TM then

g(x) = 2 · min(x, 1 − x) for all x ∈ [0, 1].

In Theorem 9.4 we present the construction method of IVFPRs from F functions stud-
ied in Section 9.3 and from weak ignorance functions.

Theorem 9.4. Let R ∈ F R(X × X ) and let g be a weak ignorance function. The relation
R� given by
R�i j = F(Ri j , g(Ri j )) for all Ri j ∈ R

[email protected]
198 Decision Aid Models for Disaster Management and Emergencies

is an interval-valued fuzzy relation such that W (R�i j ) = g(Ri j ).

Proof. Direct. 

Remark 9.4. The amplitude of each element of the IVFPR is the same as the ignorance of
that element of the FPR by construction method of Theorem 9.1.

One of the most commonly required properties in decision-making algorithms is the

reciprocity of the FPR Chiclana et al. [11], Kacprzyk [19], Switalski [24]. In this work
we say that R ∈ F R(X × X ) is reciprocal if there exists a negation N such that Ri j = N(R ji )
for every i, j ∈ {1, . . . , n} with i �= j. Analogously, we say that R� ∈ I V F R(X × X ) is
reciprocal if there exists an interval-valued fuzzy negation such that R�i j = N(R�ji ) for every
i, j ∈ {1, . . . , n} with i �= j. If using N([x, y]) = [1 − y, 1 − x] we have that R� i j = 1 − R� ji and
1 − R�i j = 1 − R� ji Deschrijver et al. [12].
In Theorem 9.5 we analyze under which conditions the IVFPR built as in Theorem 9.4
is a reciprocal IVFPR.

Theorem 9.5. In the setting of Theorem 9.4, let R ∈ F R(X × X ) be reciprocal with respect
to N(x) = 1 − x. If taking Fμ (x, y) = x(1 − y) then R� ∈ I V F R(X × X ) is also reciprocal
with respect to N([x, y]) = [1 − y, 1 − x].

Proof. Direct taking into account that g(Ri j ) = g(1 − Ri j ) = g(R ji ). 

9.4.1 Weighted voting strategy

The weighted voting strategy Huellermeier and Brinker [17], Huellermeier and Vander-
looy [18] is one of the simplest and most widely used methods for solving decision-making
problems. In this algorithm, the preference value Ri j is considered as a weighted vote for
the alternative xi . In this way, the final evaluation of each alternative xi is calculated as a
sum of votes:

Ri j . (9.4)
1i�= jn

The alternative that obtains the highest amount of votes is the one chosen as solution.

9.4.2 First generalization of the weighted voting strategy

In this subsection we present the first adaptation of the voting method to solve decision-
making problems with IVFPRs. Algorithm 1 is based on the arithmetic operations on

[email protected]
Decision Making with Extensions of Fuzzy Sets 199

interval-valued fuzzy sets obtained by extending those given given by Atanassov [3] for
Atanassov’s intuitionistic fuzzy sets. Take [x1 , y1 ], [x2 , y2 ] ∈ L∗ , the sum operator is defined
as

[x1 , y1 ] + [x2, y2 ] = [x1 + y1 − x1 y1 , x2 + y2 − x2 y2 ] ∈ L([0, 1]). (9.5)

Remark 9.5. This expression can be rewritten in terms of the probabilistic sum SP as fol-
lows: [x1 , y1 ] + [x2 , y2 ] = (SP (x1 , x2 ), SP (y1 , y2 )).

To calculate the evaluation of each alternative we replace the sum of preferences given
in Expression (9.4) by the sum of elements of L([0, 1]) given in Expression (9.5). We will
use the total order relation based on score and accuracy functions to determine the final
ranking of alternatives.
The schema of Algorithm 1 is the following:
1. Select a function Fμ and a weak ignorance function g.
2. Build the IVFPR R� from the FPR R as in Theorem 9.4.
3. FOR each alternative xi ∈ X calculate the interval

SP Ri j, SP R �i j
1 j�=in 1 j�=in

END FOR
4. Order in a decreasing way the alternatives, following the score and accuracy of the
values calculated in step 3.

Algorithm 1

Although Algorithm 1 is a natural adaptation of the voting strategy, we can see by

means of the following example that, in some cases, it is not a good approach to solve
decision-making problems. Given the FPR:

− 0.9 0.9 0.9

0.1 − 0.4 1
R=
0.1 0.6 − 0.3
0.1 0 0.7 −
1. We take Fμ (x, y) = x(1 − y) and g(x) = 2 · min(x, 1 − x).
2. We build the IVFPR R� (See Table 9.1 )
3. We calculate the intervals associated to each of the alternatives

• x1 : [0.9780, 0.9995]

[email protected]
200 Decision Aid Models for Disaster Management and Emergencies

Table 9.1 Interval-valued fuzzy binary relation R� constructed by Algorithm 1.

x1 x2 x3 x4
x1 - [0.72,0.92] [0.72,0.92] [0.72,0.92]
x2 [0.08,0.28] - [0.08,0.88] [1,1]
x3 [0.08,0.28] [0.12,0.92] - [0.12,0.72]
x4 [0.08,0.28] [0,0] [0.28,0.88] -

• x2 : [1, 1]
• x3 : [0.2876, 0.9839]
• x4 : [0.3376, 0.9136]

4. We order the alternatives

• x1 : score([0.9780, 0.9995]) = 0.9775

• x2 : score([1, 1]) = 1
• x3 : score([0.2876, 0.9839]) = 0.2715
• x4 : score([0.3376, 0.9136]) = 0.2512

The order of the alternatives is x2 > x1 > x3 > x4 . Observe that, attending to the pref-
erence of the expert given in the FPR, x1 is preferred over the rest of the alternatives.
However, x1 is not the solution that we obtain. This is due to the fact that if the expert
expresses the preference of an alternative with a high value, as in the case R24 = 1, the final
evaluation of this alternative is maximum, without taking into account the rest of prefer-
ences of this alternative. This is due to the use of t-conorms to calculate the sum of votes
for each alternative.
This behaviour leads us to propose a new generalization to solve the problems detected
in Algorithm 1.

9.4.3 Second generalization of the weighted voting strategy

In this subsection we present a new adaptation of the weighted voting strategy. The
scheme of Algorithm 2 is the following:

[email protected]
Decision Making with Extensions of Fuzzy Sets 201

1. Select a function Fμ and a weak ignorance function g.

2. Build the IVFPR R� from the FPR R as in Theorem 9.4.
3. FOR each alternative xi ∈ X evaluate

(R� i j + R� i j − 1)
1 j�=in

END FOR
4. Take as solution the alternative xi of greatest evaluation.

Algorithm 2

Notice that we are adding the scores of the interval memberships of each of the rows.
On the other hand, we can we justify Algorithm 2 from the point of view of correlation
between interval-valued fuzzy sets

9.4.4 Algorithm 2 and correlation

Correlation between interval-valued fuzzy sets is used in the literature about decision-
making problems to rank the alternatives [28]. In particular, Ye [28] considers the “excel-
lent” alternative given by

A∗ = {((xi , x j ), [1, 1]) | j ∈ {1, . . ., n}, i �= j and (xi , x j ) ∈ X × X}. (9.6)

In the present work we also consider the worst valued alternative, namely,

A∗ = {((xi , x j ), [0, 0]) | j ∈ {1, . . ., n}, i �= j and (xi , x j ) ∈ X × X}. (9.7)

We can represent each alternative xi as an IVFS Ai given by

Ai = {((xi , x j ), [R� i j , R� i j) | j ∈ {1, . . . , n}, i �= j and (xi , x j ) ∈ X × X} (9.8)

Using correlation between interval-valued fuzzy sets given in Expression (9.1), we

present the following results.

Proposition 9.1. Let A∗ , A∗ and Ai given by Expressions (9.6), (9.7) and (9.8). The follow-
ing items hold:

i) C(Ai , A∗ ) = R� i j .
1 j�=in
ii) C(Ai , A∗ ) = 1 − R� i j .
1 j�=in

Proof. Direct. 

[email protected]
202 Decision Aid Models for Disaster Management and Emergencies

Given a decision-making problem such that each alternative is represented by an A-IFS

as in Expression (9.8), we consider that the best alternative xi is that one whose correlation
with A∗ (excellent alternative) is maximal and at the same time its correlation with A∗
(worst alternative) is minimal. In this way, the evaluation of each alternative is exactly that
we have considered in Algorithm 2.

9.4.5 Semiautoduality

Our definition of sum of intervals was based on the use of the probabilistic sum. Nev-
ertheless, other definitions can be considered. In particular, we want to mention here those
based in the use on semiautodual functions.

Definition 9.7. Let A : [0, 1]2 → [0, 1] be an aggregation function and N a strong negation.
A is said to be semiautodual (with respect to N) if:

A(x, y) + N(A(N(x), N(y))  1

The minimum provides an example of a semiautodual aggregation function with respect

to the standard negation. Now observe that, if A is semiautodual, it follows that for any
x, y ∈ [0, 1] it holds that [A(x, y), 1 − N(A(N(x), N(y)))] is an interval on [0, 1]. In particular,
if taking N(x) = 1 − x as strong negation, we have that, if A is semiautodual with respect
to N, then [A(x, y), A(1 − x, 1 − y)] is also an interval, whatever x, y ∈ [0, 1] are. In this way,
we can define other additions. Nevertheless, these would require a deep study that it is out
of the scope of this chapter.

9.5 Two examples

In this section we consider two decision-making problems that make clear the effec-
tiveness of Algorithm 2. To check how they work, we compare the results of Algorithm 2
with those of the voting strategy. We see that Algorithm 2 is efficient in some problems in
which the weighted voting is not able to provide a unique solution.
In the following example we assume that we are dealing with the following situation.
There are four possible routes to provide help after an earthquake, none of them completely
safe. An expert is asked to provide the preference of route i over route j. Of course, this is
a toy problem, but we expect them to clarify the advantages of our approach.

[email protected]
Decision Making with Extensions of Fuzzy Sets 203

In this section, we take for step 1.

Fμ (x, y) = x(1 − y)
g(x) = 2 · min(x, 1 − x).

9.5.1 First example: Obtaining the same result as with the voting strategy

Given the reciprocal FPR:

− 0.33 0.48 0.65

0.67 − 0.61 0.58
R=
0.52 0.39 − 0.6
0.35 0.42 0.4 −
2. We build the IVFPR R� (see Table 9.2)

Table 9.2 Interval-valued fuzzy binary relation R� from Example 9.5.1.

x1 x2 x3 x4
x1 - [0.1122,0.7722] [0.0192,0.9792] [0.1950,0.8950]
x2 [0.2278,0.8878] - [0.1342,0.9142] [0.0928,0.9328]
x3 [0.0208,0,9808] [0.0858,0.8658] - [0.12,0.92]
x4 [0.1050,0.805] [0.0672,0.9072] [0.08,0.88] -

3. We evaluate each alternative

• x1 : −0.0272.
• x2 : 0.1896.
• x3 : −0.0068.
• x4 : −0.1556.

4. We choose the alternative with greatest evaluation.

x2 > x 3 > x 1 > x 4

Notice that results are in agreement with the ones obtained with the weighted voting
strategy.

9.5.2 Second example: Improving the solution obtained with the voting strategy

There exist problems for which the weighted voting strategy is not able to choose be-
tween two alternatives (two alternatives obtain the same sum of votes). We present an

[email protected]
204 Decision Aid Models for Disaster Management and Emergencies

example in which Algorithm 2 solves this problem. Take the reciprocal FPR:
− 0.78 0.60 0.28
0.22 − 0.75 0.69
R=
0.40 0.25 − 0.44
0.72 0.31 0.56 −
2. We build the IVFPR R� (see Table 9.3)

Table 9.3 Interval-valued fuzzy binary relation R� from Example 9.5.2.

x1 x2 x3 x4
x1 - [0.4368,0.8768] [0.12,0.92] [0.1232,0.6832]
x2 [0.1232,0.5632] - [0.375,0.875] [0.2622,0.8822]
x3 [0.08,0.88] [0.125,0.625] - [0.0528,0.9328]
x4 [0.3168,0.8768] [0.1178,0.7378] [0.0672,0.9472] -

3. We evaluate each alternative

• x1 : 0.16.
• x2 : 0.0808.
• x3 : −0.3044.
• x4 : 0.0636.

4. We choose the alternative with largest evaluation

x 1 > x2 > x4 > x3

Notice that if we apply the weighted voting strategy to the FPR R, the ranking that we
obtain is

x 1 = x2 > x4 > x3

In this case, this strategy is not able to decide between x1 and x2 for the best alternative.
However, using the ignorance in the construction of the interval-valued fuzzy preference
relation, Algorithm 2 allows to order the alternatives and select the best one.

9.6 Advantages of the method

Natural disaster management requires sudden decisions to be taken, forcing the envi-
ronmental data to be quickly collected. This data can come either in the shape of physical
measurements or as human evaluations derived from experience. In any case, time limita-
tions do not allow careful evaluations or repeated measurements, what might affect severely

[email protected]
Decision Making with Extensions of Fuzzy Sets 205

the quality of the data. We have considered those environments in which we can estimate
the inaccuracy of the measurement units or the expert, and manage to express it in a nu-
merical way. Intervals (in our case, interval-valued information) provide a mathematical
framework to handle situations where the system is provided with both a (possibly inaccu-
rate) evaluation and a prediction of its uncertainty.
Although practical and intuitive, the interval information incorporates several problems,
among which we must highlight the lack of proper ranking methods, which constitute the
grounds of any decision making algorithm. Indeed, very few proposals have appeared in
the literature able to handle interval valued information, what forces the information to
be considered precise, and lead to unwanted results. Moreover, note that by not using
the information about the uncertainty of the data, we are missing half of the available
information. Hence, we believe that the ability to deal with imprecise information will be
a key in further developments of disaster management systems, whether this is done by
expressing the data as intervals or using any other mathematical tool.

9.7 Conclusions

In this work we have presented a decision-making algorithm to obtain a solution in

some cases in which classical methods are not able to decide between two alternatives. To
do so, we have proposed a construction method of IVFPRs from the FPRs given by experts.
This method quantifies by means of weak ignorance functions the lack of knowledge of the
expert in the assignation of the numerical values of the fuzzy preferences.
Our construction method, our algorithm departs from the FPR given by the expert, so
the results of our algorithm can be compared to those obtained with any other classical
decision-making method. In the illustrative examples, we have checked that Algorithm 2
provides a solution in some cases in which the voting strategy does not distinguish between
two alternatives. This algorithm can hence be a good choice for those cases in which the
weighted voting method does not arrive to a unique solution.

Acknowledgment

This work has been partially supported by the National Science Foundation of Spain,
reference TIN2010-15055.

[email protected]
206 Decision Aid Models for Disaster Management and Emergencies

Bibliography

[1] Aleskerov, F., Iseri Say, A., Toker, A. and G., A. (2005). A cluster based decision support
system for estimating earthquake damage and casualties, Disasters 29, pp. 255–276.
[2] Atanassov, K. T. (1986). Intuitionistic Fuzzy-Sets, Fuzzy Sets and Systems 20, 1, pp. 87–96.
[3] Atanassov, K. T. (1994). New operations defined over the intuitionistic fuzzy-sets, Fuzzy Sets
and Systems 61, 2, pp. 137–142.
[4] Baczynski, M. and Jayaram, B. (2008). Fuzzy Implications, Studies in Fuzziness and Soft Com-
puting, Vol. 231 (Springer, Berlin).
[5] Burillo, P. and Bustince, H. (1994a). Intuitionistic fuzzy relations. (Part I), Mathware and Soft
Computing 2, 1, pp. 5–38.
[6] Burillo, P. and Bustince, H. (1994b). Intuitionistic fuzzy relations. (Part II) Effect of
Atanassov’s operators on the properties of the intuitionistic fuzzy relations, Mathware and Soft
Computing.
[7] Bustince, H., Barrenechea, E. and Pagola, M. (2008). Generation of interval-valued fuzzy and
Atanassov’s intuitionistic fuzzy connectives from fuzzy connectives and from K-alpha opera-
tors: Laws for conjunctions and disjunctions, amplitude, International Journal of Intelligent
Systems 23, 6, pp. 680–714.
[8] Bustince, H., Burillo, P. and Soria, F. (200). Automorphisms, negations and implication opera-
tors, Fuzzy Sets and Systems 134, 2, pp. 209–229.
[9] Bustince, H., Pagola, M., Barrenechea, E., Fernandez, J., Melo-Pinto, P., Couto, P., Tizhoosh,
H. R. and Montero, J. (2010). Ignorance functions. An application to the calculation of the
threshold in prostate ultrasound images, Fuzzy Sets and Systems 161, 1, pp. 20–36.
[10] Chen, S. M. and Tan, J. M. (1994). Handling multicriteria fuzzy decision-making problems
based on vague set-theory, Fuzzy Sets and Systems 67, 2, pp. 163–172.
[11] Chiclana, F., Herrera, F. and Herrera-Viedma, E. (1998). Integrating three representation mod-
els in fuzzy multipurpose decision making based on fuzzy preference relations, Fuzzy Sets and
Systems 97, 1, pp. 33–48.
[12] Deschrijver, G., Cornelis, C. and Kerre, E. E. (2004). On the representation of intuitionistic
fuzzy t-norms and t-conorms, IEEE Transactions on Fuzzy Systems 12, 1, pp. 45–61.
[13] Deschrijver, G. and Kerre, E. E. (2003). On the relationship between some extensions of fuzzy
set theory, Fuzzy Sets and Systems 133, 2, pp. 227–235.
[14] Fodor, J. and Roubens, M. (1994). Fuzzy preference modelling and multicriteria decision sup-
port (Kluwer, Dordrecht).
[15] Gerstenkorn, T. and Manko, J. (1991). Correlation of Intuitionistic Fuzzy-Sets, Fuzzy Sets and
Systems 44, 1, pp. 39–43.
[16] Helbing, D., Johansson, A. and Al-Abideen, H. Z. (2007). Dynamics of crowd disasters: An
empirical study, Physical Reviews E 75, p. 046109.
[17] Huellermeier, E. and Brinker, K. (2008). Learning valued preference structures for solving
classification problems, Fuzzy Sets and Systems 159, 18, pp. 2337–2352.
[18] Huellermeier, E. and Vanderlooy, S. (2010). Combining predictions in pairwise classification:
An optimal adaptive voting strategy and its relation to weighted voting, Pattern Recognition 43,
1, pp. 128–142.
[19] Kacprzyk, J. (1986). Group decision-making with a fuzzy linguistic majority, Fuzzy Sets and
Systems 18, 2, pp. 105–118.
[20] Klement, E., Mesiar, R. and Pap, E. (2000). Triangular Norms (Kluwer Academic Publishers,
Dordrecht).
[21] Krausz, B. and Bauckhage, C. (2012). Loveparade 2010: Automatic video analysis of a crowd
disaster, Computer Vision and Image Understanding 116, 3, pp. 307–319.

[email protected]
Bibliography 207

[22] Mendonca, D., Beroggi, E. and Wallace, W. (2001). Decision support for improvisation during
emergency responsa operations, International Journal of Emergency Management 1, pp. 30–38.
[23] Sanz, J., Fernandez, A., Bustince, H. and Herrera, F. (2011). A genetic tuning to improve the
performance of fuzzy rule-based classification systems with interval-valued fuzzy sets: Degree
of ignorance and lateral position, International Journal of Approximate Reasoning In Press,
Corrected Proof, 10.1016/j.ijar.2011.01.011.
[24] Switalski, Z. (2003). General transitivity conditions for fuzzy reciprocal preference matrices,
Fuzzy Sets and Systems 137, 1, pp. 85–100.
[25] Tan, C. and Chen, X. (2010). Intuitionistic fuzzy Choquet integral operator for multi-criteria
decision making, Expert Systems with Applications 37, 1, pp. 149–157.
[26] Wallace, W. and De Balogh, F. (1985). Decision support systems for disaster management,
Public Administration Review 45, pp. 134–146.
[27] Xu, Z. and Yager, R. R. (2006). Some geometric aggregation operators based on intuitionistic
fuzzy sets, International Journal of General Systems 35, 4, pp. 417–433.
[28] Ye, J. (2010). Fuzzy decision-making method based on the weighted correlation coefficient
under intuitionistic fuzzy environment, European Journal of Operational Research 205, 1, pp.
202–204.
[29] Zadeh, L. A. (1965). Fuzzy Sets, Information and Control 8, 3, pp. 338–353.

[email protected]
Chapter 10

Classification of Disasters and Emergencies

under Bipolar Knowledge Representation

J. Tinguaro Rodríguez∗, Begoña Vitoriano, Daniel Gómez, Javier Montero

Faculty of Mathematics, Complutense University of Madrid 28040, Madrid, Spain

E-mail: {jtrodrig,bvitoriano,javier_montero}@mat.ucm.es

A fully precise numerical evaluation of disasters’ effects is unrealistic in the time-pressured,

highly uncertain decision context taking place just after a disaster strike. This is mainly due
to some features of the available information in such a context, but also because of the im-
precise nature of some of the relevant categories (think about the number of affected people,
for instance). Instead of a numerical evaluation, in this work is considered that it is rather
more plausible and realistic to classify the severity of the consequences of a disaster in terms
of the relevant scenarios for the NGO’s decision makers. Therefore, the abovementioned
practical problem of evaluation of disaster consequences leads to a classification problem
in which the classes are identified with the linguistic terms that describe those relevant sce-
narios. In order to carry out this classification and ensure the linguistic adaptation and the
understandability of the proposed solution, the methodology of the descriptive fuzzy rule-
based classification systems has been adopted in this work. Nevertheless, some features of
that context, as the ordering and gradation of the consequences or the need of avoiding the
risk of underestimation of the effects of disasters, entail the necessity of considering and
assuming an structure over the set of classes or linguistic labels, somehow modeling those
features inside of the classification model. Such an structure is introduced here by means
of the notion of dissimilarity between classes, leading to a bipolar knowledge representa-
tion framework which allows to adequate the classification models to the constraints and
requirements of the NGO context.

10.1 Introduction

Natural and anthropogenic disaster management constitutes a main research stream not
only because of the tremendous impact of disasters in directly affected people, but also be-
cause of the deep consequences they can have in infrastructures, economy and the political
∗ Corresponding author. E-mail address: [email protected]

B. Vitoriano et al. (eds.), Decision Aid Models for Disaster Management 209
and Emergencies, Atlantis Computational Intelligence Systems 7,
DOI: 10.2991/978-94-91216-74-9_10,  Atlantis Press 2013

[email protected]
210 Decision Aid Models for Disaster Management and Emergencies

system of the affected regions, which could lead to political and social instability, espe-
cially in developing countries, and affect international relations and security. Therefore,
the development of technologies and procedures to mitigate the potential consequences of
disasters is a key issue in today’s global world, beside humanitarian arguments.
At this respect, developed countries invest a huge amount of funds, usually by means
of development assistance and other more disaster-mitigation concerned policies in regions
they consider geo-strategically sensible, in order to avoid the risks of political and social
destabilization produced by disasters (see [17]). However, most of this funding is channeled
through United Nations agencies and international Non-Governmental Organizations, on
the basis of their supposed political neutrality and links with the target populations. This
allows these actors, for example, to access politically unstable countries which otherwise
would not accept any external interference. For these reasons, the importance of these
agencies and organizations is, in practice, highly significant.
Moreover, in the context of these organizations, it is fundamental to manage the avail-
able resources with transparency and efficiency. As a consequence, humanitarian logistics
[29] is an emerging field of research, with a growing importance, which is at the same
time drawing attention to some interesting problems in management sciences and opera-
tions research. In this sense, humanitarian response to disasters implies a complex decision
making, in which a number of key strategic and operational decisions have to be faced in a
time-pressured context and with low quality information.
The long term objective of the project containing this work is to develop and implement
a general, standard decision support system (DSS) for disaster management, specifically
designed to address a part of this complexity and help NGO’s decision makers involved in
the design of humanitarian relief operations. Particularly, this work focuses on SEDD (the
Spanish acronym for Disaster Diagnostic and Evaluation System), which is the part of such
a global DSS concerned with the assessment of the consequences of disasters with the very
first information available after the strike (see [30] for a description of other parts of such
a global DSS being developed in this project). As discussed in [21, 24], SEDD constitutes
one of the first proposals in the direction of providing NGOs with disasters’ consequences
evaluation procedures specifically designed for them.
Let us recall with [21, 23] that SEDD’s inference capability is based on fuzzy ma-
chine learning procedures (see [9]), particularly on the methodology of descriptive fuzzy
rule based classification systems (FRBCS, see [6, 10–12]). This is, SEDD’s methodology
is a mixture between fuzzy inference systems [6], that enable inference to be carried out

[email protected]
Classification of Disasters and Emergencies under Bipolar Knowledge Representation 211

in terms of a linguistically expressed (see [32, 33]), interpretable set of rules and infor-
mation, and data mining (see for instance [1, 9]), that allows these rules to be obtained
from adequate databases. In this sense, SEDD extracts its rules from EM-DAT (Emergency
Database, see www.em-dat.be), the most exhaustive and complete public database about
disasters and emergencies. Moreover, as shown in [25], fuzzy rule based systems outper-
forms ordinary statistical techniques and most state-of-the-art machine learning techniques
in the task of providing a simultaneously accurate and interpretable (in the sense of [16])
assessment of disaster consequences.
This paper continues the work in [23] by studying the characteristic structure of the
classification problem that underlies the disaster severity assessment provided by SEDD.
Particularly, the relationships between the structure assumed on the set of classes and the
requirements of the decision context are analyzed, showing that a significant improvement
in the behavior of the classification methodology of SEDD in terms of its adaptation to
such requirements is obtained when certain structures are assumed and introduced inside
the classification models.
In this sense, as shown for instance in [22], in a supervised classification context it
is possible to introduce relations between the classes by means of the notion of semantic
antagonism or dissimilarity proposed in [20]. Thus, in this work the effect of different
dissimilarity structures is discussed in terms of their ability to replicate and adapt to some
of the requirements and features of the disaster management context.
This work is structured as follows: in Section 10.2, the problem SEDD tries to address
is described with some detail, focusing on its strategic features, that will allow to identify
it with a structured classification problem in Section 10.3. SEDD’s extension to these
problems is presented in Section 10.4, discussing the adaptation of diverse dissimilarity
structures to the task of modeling different context requirements. Finally, some conclusions
are shed in Section 10.5.

10.2 Problem description

Just after a disaster strikes somewhere in the world, international NGOs start a decision
process intended to reach a conclusion about the pertinence of a relief operation and about
whether or not suitable conditions exist to initiate it. It is possible that a specific disaster
scenario does not fit the NGO’s requirements or constraints regarding the nature of an
intervention, size of disaster, or logistical capabilities. Therefore, the decisions to be made

[email protected]
212 Decision Aid Models for Disaster Management and Emergencies

in a first stage have an strategic nature, being more concerned with assessing the degree
of involvement of the organization in a possible response operation than with the specific
content of such an operation.
However, precisely because they determine the shape and guidelines of the actions to be
done, strategic decisions have a major influence on the subsequent logistical and on terrain
decision processes (see Fig. 10.1 below), which evaluate the amount of aid to send and how
it will be delivered to the affected country and the suffering population. As a consequence,
strategic decision-making takes place in a highly time-pressured context, since any delay at
this stage could affect the position of the NGO in the international coalition delivering aid,
thus affecting the NGO’s prestige and reputation, and slow down the subsequent decision
processes, thus delaying the reception of aid by the affected population.
Notice that this strategic decision process has to be flexible, in the sense that it has to
be able to be carried out for every combination of disaster type and place, since NGOs are
possibly specialized in covering some part of the relief tasks (as water sanitation, shelter
and site management, health care, etc.) but are not specialized in response to, for instance,
earthquakes in Haiti, floods in Pakistan or any other specific disaster scenario (despite
the geo-strategic priorities a particular NGO could have as a result of the interests of its
donors). This is, rather than the specific geographic features of the affected location or
the nature of the implied natural or anthropogenic phenomenon, what matters in terms of
NGO’s strategic decisions is the extent and the severity of the humanitarian disaster that is
taking place as a consequence of whatever factors.

DECISIONS

DISASTER INITIAL
STRATEGIC
ASSESSMENT

LOGISTIC

ON TERRAIN

Fig. 10.1 Hierarchy of decisions for disaster response. An initial assessment of disaster conse-
quences is crucial in order to start the decision process.

[email protected]
Classification of Disasters and Emergencies under Bipolar Knowledge Representation 213

Therefore, as shown in Fig. 10.1, strategic decision-making (and thus all the subsequent
decision processes) is strongly dependent on a correct assessment of the consequences of a
disaster and the resulting needs of the affected population. In fact, recall that humanitarian
relief operations are assumed to follow the quality standards established by the Humani-
tarian Charter of the Sphere Project (see [28] and www.sphereproject.com), which, among
other things, emphasize the correct estimation of needs as a critical step for success in
these operations. This is, the fulfillment of the quality standards in the delivery of aid
presupposes an adequate assessment of the kind of relief actually needed by the affected
population.
However, the available information just after the impact of a disaster uses to be affected
by different kinds of uncertainty. The first reports are usually pretty incomplete, if not
directly confusing and contradictory. Moreover, when it exists, relevant information is
usually expressed linguistically, and thus it could be vague and imprecise. In fact, due to
the effects an adverse phenomenon has on the informational system of a region, a more
or less exhaustive and precise picture of the situation could not be obtained until some
days (or even weeks) after the moment of the strike. This poses a strong difficulty in the
development of the mentioned strategic decision-making process, since in this context the
urgency of the decisions to be made in order to relief the people affected by a disaster
clashes with the need of a correct estimation of the consequences of such a disaster, i.e. of
the needs of the affected populations.
In the spirit of [31], all this complexity suggests as a promising alternative the de-
velopment and application of inference techniques enabling a fast, flexible and correct
assessment of disasters’ consequences in the presence of uncertainty. Nevertheless, some
constraints must be imposed on the nature of these techniques. For example, the procedures
leading to such an assessment have to be understandable and interpretable by the decision
makers, in order to guarantee the usability of such an inference tool.
Even more important are the infrastructure and data requirements of a decision support
tool operating in the context of NGOs. It is necessary to make realistic assumptions in this
area if the DSS have to be used in organizations or countries where the operational infras-
tructure can not supply highly sophisticated data, as, for example, real-time remote sensing
information [18], exhaustive building census [26] or precise meteorological data [4]. It is
important to remark that this question is a critical issue when considering disaster manage-
ment from the point of view of NGOs or developing countries, as complex data could be

[email protected]
214 Decision Aid Models for Disaster Management and Emergencies

unavailable or the trained personnel requirements for its usage could be too high (see [2, 13]
for a discussion about the rigidity of most current DSS for disaster management).
With these ideas in mind, SEDD was designed to be a low-cost, tailor-made solution
for NGO-oriented assessment of disasters consequences. It is not the aim of this paper
to describe SEDD’s methodology in deep. However, it is important to remark that, as
shown in [23], SEDD’s inference capability relies on a set of data easily accessible from
the first description of a potential disaster. Moreover, the reasoning method that sustains
such an inference capability is interpretable, since it is based on the usage of (a rather
small number of) rules expressed in terms of a natural language (i.e. fuzzy rules). Also,
since its knowledge (i.e. the rules) is learnt from the EM-DAT database by means of
appropriate machine learning procedures, SEDD is able to produce an assessment of any
disaster scenario for which enough similar historical information is available. Therefore,
SEDD’s flexibility depends on the existence of appropriate data, but its methodology is in
principle able to face any possible combination of disaster type and location. Some further
details about SEDD are given in Section 4 of this chapter. A complete description of its
methodology and its performance can be found in [25].

10.3 Strategic disaster severity assessment as an structured classification

problem

As explained above, the strategic decision-making about the involvement of an NGO

on a disaster response operation is strongly dependent on the initial assessment of the con-
sequences of such a disaster. However, a fully precise numerical evaluation of disasters’
effects, as casualties, homeless people or the extension of the material damage, is unrealis-
tic in such a decision context. In fact, even a more or less complete and precise description
of these consequences is usually not available by the time in which such (urgent) strategic
decisions have to be already taken.
This is mainly due to the uncertainty and the referred features of the available informa-
tion just after a disaster strike, but it is also because of the imprecise nature of some of the
relevant categories. For example, the notion of affected people shows such an imprecision,
since it could be not always clear whether a person has been affected or not. As a con-
sequence, the number of affected people is usually stated through an implicitly imprecise
quantity, as happens when it is said that a disaster produced, for instance, 40.000 affected
people.

[email protected]
Classification of Disasters and Emergencies under Bipolar Knowledge Representation 215

Nevertheless, it is important to notice that a totally precise and exhaustive evaluation

of consequences is not actually needed in order to perform the above described strategic
decision-making. As pointed out above, strategic decisions determine the shape of an oper-
ation but not its specific contents. As the decision process in Fig. 10.1 develops, decisions
need information to be more and more precise, since decisions become more and more
concrete. In this sense, NGOs usually deliver experts on the affected location in order to be
able to acquire such a more precise evaluation for its logistical and on terrain decisions. But
the decision of acquiring such a further evaluation is a strategic decision that has to be taken
in the first moments after the strike, when little information is available, i.e. on the basis
of the initial reports of the disaster. However, such an initial assessment of consequences
needs to be correct or accurate, but not necessarily fully precise.
For example, consider the estimation of the variable number of homeless people. This
variable measures the number of people that become homeless as a consequence of a dis-
aster. For NGOs, such a quantity constitutes a key indicator of the size of the efforts a
potential relief operation should place in matter of temporary shelter and site management.
This also provides an idea of the efforts to be placed in the water sanitation area, for in-
stance. In practice, at a first stage, for an NGO decision maker it is not important at all
to distinguish whether 50.000 or 70.000 people become homeless as a consequence of a
disaster, since anyway such a number is going to be considered large and the strategic de-
cisions in terms of both the subsequent actions and the size and nature of the required relief
operation are going to be similar.
Therefore, in the first moments after a disaster strike, NGO decision makers assess
disaster severity (and thus also the needs of a potential intervention) in a qualitative (rather
than quantitative) way. As said, their problem has a big-scale, strategic nature, rather
than a small-scale, tactic or operative nature. In other words, their problem consists on
evaluating, in a context of highly uncertain and imprecise information, the magnitude of
the consequences of a disaster in relation with the relevant scenarios and decisions that
can arise regarding the implementation of a relief operation. In such a context, it is even
possible that an assessment stated in crisp and precise terms could result little trustable to
those decision makers.
However, a linguistic description of the magnitude order of the consequences coming
from a reliable source, for example stating that there are a lot of casualties or that buildings
took a several damage, will be much more trustable to decision makers, despite its implicit
imprecision. This kind of linguistic information is enough relevant to elaborate a first

[email protected]
216 Decision Aid Models for Disaster Management and Emergencies

perception of the disaster scenario, providing a base for the subsequent strategic decision
making.
Thus, in order to obtain such an initial assessment giving rise to an adequate strategic
decision making, we consider that instead of a numerical evaluation, it is rather more plau-
sible and realistic to classify the severity of the consequences of a disaster in terms of the
relevant scenarios for the NGO’s decision makers. Therefore, the abovementioned practical
problem of evaluation of disaster consequences leads to a classification problem in which
the classes are identified with the linguistic terms that describe those relevant scenarios, as
no casualties or a lot of injured people. These linguistic labels or classes, assessing the
magnitude of the different relevant consequences of a disaster, have to be assigned on the
basis of the description or attributes of such a disaster given by the first available informa-
tion, as the type of disaster, its intensity and the features of the affected location (e.g. its
vulnerability, see [15]).
Notice that, as shown in [23], SEDD is able to produce assessments with different
levels of precision, accordingly to the characteristics of the information needed in each
decision stage. Particularly, SEDD provides three types of output or assessment of each of
the disaster consequences it addresses (those present in the EM-DAT database, as we shall
see in Section 10.4): a) numerical; b) intervals; and c) linguistic labels or classes. In this
work we focus on the classification methodology of SEDD, which is therefore associated
to the support of the strategic decision-making of NGOs.

10.3.1 Structure of the set of classes

Consider now one of the variables that have to be linguistically evaluated in order to
obtain such a first initial assessment. For instance, let us focus on the variable number
of casualties (CAS). This variable estimates the number of people that were killed as a
result of the strike of an adverse phenomenon. As just explained, in a first stage a fully
precise estimation of such a number is not strictly necessary, but just a qualitative, linguistic
assessment. In this way, for instance we can measure the magnitude of a disaster scenario,
in terms of the casualties it produced, by means of the labels no casualties, very few, few,
quite a lot and a lot of casualties. These labels represent the classes in which such a
disaster scenario has to be classified in order to provide an initial assessment of the relevant
consequence CAS.
The particular meaning of these labels has to be specified by means of intervals, or
more generally, through fuzzy subsets of the range of the underlying numerical variable, in

[email protected]
Classification of Disasters and Emergencies under Bipolar Knowledge Representation 217

Table 10.1 Intervals associated to the linguistic labels defined for the variable
number of casualties (CAS).
Class Label Interval
CAS1 No casualties [0, 10)
CAS2 Very few [10, 100)
CAS3 Few [100, 1000)
CAS4 Quite a lot [1000, 10000)
CAS5 A lot [10.000, +∞)

this case the positive integers. Moreover, such meanings have to be related to the different
scenarios that are relevant in terms of the decisions to be made. For example, each label
can be associated with a different order of magnitude of the number of casualties, in an
increasing way, as shown in Table 10.1 above. Notice that these intervals are just a sim-
plification of the more flexible and robust semantics specified by the fuzzy sets depicted in
Figure 10.2.

Fig. 10.2 Fuzzy partition for the variable number of casualties (CAS).

Notice that, as they are associated to different orders of magnitude of the consequences
of a disaster, these labels can be considered as ordered from the lowest (no casualties) to the
greatest (a lot of casualties) level of magnitude of such consequences. This is, the classes
associated to the variable CAS (in the first column of Table 10.1) are linearly ordered, i.e.
CAS i < CAS j whenever i < j.
This assumption of linearity on the effects of a disaster is quite general, since what
decision makers try to assess in a first stage is precisely the order of magnitude or severity
of these consequences, or equivalently the order of magnitude of the efforts that a relief op-
erations should consider in order to adequately alleviate such consequences. As explained
above, these orders of magnitude are not necessarily powers of ten of the underlying nu-
merical variable, but rather they are associated to the different relevant scenarios that can

[email protected]
218 Decision Aid Models for Disaster Management and Emergencies

arise, which can be also considered as ordered attending to the gravity of the humanitarian
crisis taking place in each of such scenarios.
Thus, in this setting classes are not independent, unrelated items, but they conform a
valuation structure, in which some relationships hold between the valuation states given
by the classes. This is, following [14] in this context the set of classes presents a relevant
structure, which should be taken into account in the process of assessing the severity of
a disaster scenario under study from the very first information available. For example,
we do not commit the same error when a disaster scenario having no casualties (CAS1)
is assessed as one with very few casualties (CAS2), than when it is evaluated as having
produced a lot of casualties (CAS5). Therefore, the abovementioned problem of disaster
severity assessment can be understood as a classification problem with a structured set of
classes, i.e., as a structured classification problem.

10.3.2 Context requirements

As stated above, the notion of structure of the set of classes is introduced in order
to capture some relevant relationships that hold between the concepts represented by the
classes. These relationships are given by the features of each particular application context,
which provides the specific meaning or semantics attributed to the classes. However, it is
important to notice that these relationships not only depend on the semantics of the classes,
but they also have to reflect the criteria and requirements of the decision context.
For instance, the set of classes introduced before for the variable CAS in principle fits
into a linear structure, since classes are semantically associated to orders of magnitude of
the consequences. As a consequence of this assumption, the classifier could be required to
show a gradable, smooth behavior, in the sense that small variations on the attributes that
describe a disaster scenario should not produce a large variation on the predicted conse-
quences. Some classes are closer than others to a given class, and thus different error levels
can be distinguished according to such a distance.
However, notice that the error committed by assessing a CAS1 scenario as a CAS5 one
is also different from that committed when a CAS5 scenario is evaluated as verifying CAS1.
In the first case, overestimation error is committed, while in the second case scenarios are
underestimated. Though the two types of errors are relevant, notice that underestimation of
disaster consequences could lead to much more dangerous situations than overestimation
in terms of the prestige of an NGO and the relief of the affected population.

[email protected]
Classification of Disasters and Emergencies under Bipolar Knowledge Representation 219

In this sense, overestimation of consequences may lead to an initial overreaction, but as

soon as observers are deployed on terrain and further information is available the scenario
can be reassessed and the decisions reconsidered without too many difficulties. However,
when a disaster scenario is underestimated, it uses to attract less attention and to be con-
sidered as less important, which could lead to not properly ask for further information or
even to ignore it in a first moment, thus potentially affecting the timing of the strategic
decision-making stage (with the resultant delays on the subsequent logistical and opera-
tional decision phases) as well as the NGO’s reputation.
Consequently, NGO decision makers usually tend to avoid the risk of underestimation
of disaster effects, for instance by carrying out a worst-case analysis of the scenarios un-
der study. In this way, initial assessments of a disaster scenario could be required to be
developed under the assumption of avoiding underestimation risk. In this sense, such a
decision-related requirement entails introducing a somehow asymmetric configuration in
the linear structure of the classes, since different error levels are then attained depending
on whether a disaster scenario under study is underestimated or overestimated. Therefore,
as pointed out above, the structure of the set of classes has to capture both the relevant
aspects of the semantics of the classes as well as the objectives and requirements related
with the decision context in which the classifier is used.

10.3.3 Dissimilarity structures

Notice that the assumption of asymmetry on the linear structure of the set of classes
forces to look for more general structures than orders. In this work, we adopt the notion of
dissimilarity structure proposed in [20] in order to provide a formal definition of the struc-
ture of the set of classes in an structured classification problem. Recall that dissimilarity
structures are based on the notion of semantic antagonism (also proposed in [20]), that pro-
vide a formal framework to model the opposition relationships between a set of concepts in
which such an opposition is allowed to be asymmetric. Therefore, by adopting the notion
of dissimilarity or antagonism instead of that of linear order, we somehow translate the se-
mantic distance between two classes, coming from the linear ordering of the consequences,
into the degree of opposition among them, which is however allowed to be asymmetric in
order to reflect the requirement of underestimation risk avoidance.
Moreover, as we shall see in next section, dissimilarity structures provide an easy and
effective method of introducing the relationships between the classes into the classification
models, i.e. in the learning and reasoning processes of the classifiers. In this context, the

[email protected]
220 Decision Aid Models for Disaster Management and Emergencies

opposition between classes represented in the dissimilarity structure enable to distinguish

significant exceptions to a classification rule from simple, logical counterexamples, which
leads to introduce a negative confidence degree of the rules. Such a negative degree together
with the usual, positive confidence degree, constitute then a bipolar evidence pair for the
evaluation of classification rules, which leads to a bipolar fuzzy rule-based classification
framework ([18, 22] for further details).
Therefore, let us denote by ζ = {C1 , . . . ,CN } the set of concepts or classes into con-
sideration. Recall that a dissimilarity structure can be built upon this set by means of a
dissimilarity matrix Δ = (di j )N×N , such that the value di j ∈ [0, 1] expresses the degree up
to which the class C j is opposite, antagonistic or dissimilar to the class Ci , i, j = 1, . . . , NC .
Notice that Δ is allowed to be non-symmetric, thus enabling the underlying dissimilarity
notion to be asymmetric.
This way, for example, a matrix ΔI = 0 represents a situation in which no class is
opposite to any other. On the other side, a matrix ΔIII = 1 − Id describes a situation in
which every class is totally opposite to each other. As we shall see in the next section, the
consideration of a dissimilarity matrix allows to introduce and take into account some of the
requirements and constraints of the application context inside of the classification model.
Therefore, there exist a wide range of possibilities, lying between these extreme cases, in
order to model specific dissimilarity conditions. The choice of a particular dissimilarity
matrix will of course depend on the specific semantic requirements to be fulfilled.

10.4 SEDD’s bipolar classification methodology

Recall that, as it was described in [23] and [25], SEDD’s fuzzy rule-based classifica-
tion methodology does not explicitly consider any structure on the set of classes (though a
multi-classification procedure was given in order to enable predictions formed by several
adjacent classes, introducing a pessimistic reasoning method in order to avoid underesti-
mation, see [23] for the details). Therefore, in this section we illustrate the ideas above by
adapting the methodology of SEDD to an structured classification framework. To this end,
we adopt the bipolar fuzzy rule-based classification framework proposed in [18, 22]. Thus,
here we study the capability of different dissimilarity structures (i.e. of different dissimilar-
ity matrices Δ) to capture the semantics and the requirements of the disaster response NGO
strategic decision context as well as to produce a correct classification result, i.e. a correct
assessment of disaster consequences, through the bipolar classifiers proposed in [22].

[email protected]
Classification of Disasters and Emergencies under Bipolar Knowledge Representation 221

10.4.1 Some basics about SEDD

For the sake of an adequate understanding of the example we will propose in next
section in order to illustrate the feasibility of the proposed approach, let us first recall some
basics about SEDD.
Firstly, recall that SEDD can be understood as a fuzzy rule based classification systems
(see [6, 10, 11]). As such, the knowledge or rules that guides the classification process
have to be learned from training examples, and the reasoning process that assigns a class
to a query needs such a query to be described in the same terms as the learning exam-
ples. Training examples for SEDD are provided by EM-DAT (Emergency Database, see
www.em-dat.be), the most exhaustive and complete public database about disasters and
emergencies. However, EM-DAT provides a rather incomplete description of the locations
affected by disasters. For this reason, EM-DAT has been merged with both UNDP data
about the Human Development Index (HDI), that provides an estimation of the vulner-
ability of the affected locations in the moment of the strike, as well as with US Census
historical data on population densities (POP). Also, EM-DAT informs on the type of ad-
verse phenomenon that produced each registered disaster, its magnitude (MAG) or intensity
as well as on the effects it produced in terms of a set of consequences, that range from the
number of casualties (CAS) to the number of homeless people or the extent of the material
damages (see Table 10.2 below).
The descriptive variables MAG, HDI and POP are taken as explanatory or independent,
while those related to effects or consequences, as CAS, are taken as dependent variables to
be assessed. As a consequence, SEDD need a disaster scenario to be described in terms of
these three independent variables in order to produce an assessment of its consequences.
Also, SEDD is only able to assess the consequences registered in EM-DAT, though these
constitute a enough relevant description for NGO decision makers in the first moment after
the strike. It is important to notice that EM-DAT poses a difficult classification problem,
since the variability of the consequences or classes is huge for similar values of the indepen-
dent variables, and also the sample is highly unbalanced (e.g. there are much more training
examples from class CAS1 than from class CAS5) as can be observed in Figure 10.3 below.
Secondly, SEDD needs to translate the raw, numerical data of EM-DAT in terms of
the linguistic labels that are used in both the premises and the consequents of the rules.
To this end, a fuzzy partition is defined on the range of each variable, according to the
desired semantics of the labels, similarly to what was done when defining the semantics
of the labels for the variable CAS, see Fig. 10.2 again. The labels and partitions defined

[email protected]
222 Decision Aid Models for Disaster Management and Emergencies

Table 10.2 Relevant variables contained in EM-DAT (after being merged) for each disaster
type. In SEDD, the first three variables are taken as independent and the remaining ones as
dependent.
Variable Description

Magnitude Intensity of the adverse phenomena:degrees on the Richter scale

for earth quakes, inundated area in km2 for floods, etc.
Human Development Index: an estimation of the affected
HDI
country’s vulnerability at the moment of the strike.
Population density of the affected country at the moment of the
Population Density
strike: an estimation of the affected place’s population-at-risk.
Casualties Number of casualties produced by the disaster.
Injured Number of injured people.
Homeless Number of homeless people.
Affected Number of affected people.
An estimation of the amount of infrastructural damage in
Damage thousands of US dollars.

Fig. 10.3 EM-DAT sample for disaster type = earthquakes. Classes for the variable CAS are ob-
tained through the intervals defined in Table 10.1.

for the variable MAG and HDI are depicted in Fig. 10.4. The labels and partitions for the
remaining variables are omitted, since they are not relevant for the example exposed in next
section.
Thirdly, a learning procedure has to be applied in order to extract the rules from the
data. Therefore, let us denote by X1 , . . . , Xn the n attributes or independent variables that are
used to describe a disaster scenario, and assume that a set of mhistorical disaster scenarios
(x1p , . . . , xnp ;C p ) is available as learning sample, where for each p = 1, . . . , m, C p is one
of the classes in ζ = {C1 , . . . ,CN } defined to (linguistically) assess the consequence or

[email protected]
Classification of Disasters and Emergencies under Bipolar Knowledge Representation 223

Very low Medium - low Medium - high High Very high

Human Development Index

Very weak Weak Medium Strong Very strong

Magnitude of the eartquake

Fig. 10.4 Fuzzy partition and linguistic labels for the variables Magnitude of an earthquake (MAG)
and Human Development Index (HDI).

dependent classification variable Y . Assume also that a dissimilarity matrix Δ = (di j )N×N
has been defined modeling the opposition relationships between the classes in ζ . For all
i = 1, . . . , n let us also denote by Ai1 , . . . , Aici the ci linguistic labels defined for each attribute
Xi , in such a way that μAi j (xi ) ∈ [0, 1] represents the degree up to which the value xi fulfills
the label Ai j . A premise A of a rule is given by the combination of a different label Ai ji
of each attribute, i.e. A = A1 j1 × · · · × An jn . The degree of fulfillment of a premise A
by a description x = (x1 , . . . , xn ) is usually obtained through a t-norm (see [27]) T , i.e.
μA (x) = T (μA1 j1 (x1 ), . . . , μAn jn (xn )).
In these conditions, following [22], a rule with premise A and having as a consequent a
class C j ∈ ζ is evaluated by means of a pair of confidence degrees
m
μA (x p ) μA (x p ) · μΔC j (C p )
p=1,...,m/C p =C j p=1
r+j (A) = m , r−j (A) = m ,
μA (x p ) μA (x p )
p=1 p=1

where μΔC j (Ci ) = d ji . Notice that r+ , r− ∈ [0, 1] and that it holds that r+ + r−  1 (and
see [3]). As explained in [22], r+j (A) and r−j (A) respectively estimate the proportion of
positive examples and significant exceptions (or negative examples) of the rule A ⇒ C j out
of the total number of training patterns fulfilling the premise A. Therefore, the dissimilarity
structure is introduced in the evaluation of the rules by means of the negative confidence
degree. Different procedures (see for instance [5]) can be used to obtain the set of premises
A for which rules A ⇒ C j ( j = 1, . . . , N) have to be built, basically ensuring that each
training pattern is covered by at least one rule. Once this learning stage is finished, a set

[email protected]
224 Decision Aid Models for Disaster Management and Emergencies

of rules or rule base is available, which represents the knowledge of the classifier. It is
important to remark that a main advantage of descriptive fuzzy classifiers (like SEDD)
is that they provide rules expressed in terms of a natural language, so their knowledge is
explicit and interpretable.
Lastly, it is possible to apply a fuzzy reasoning method (see [6]) in order to produce
an assessment of a disaster scenario x = (x1 , . . . , xn ) under study. Following [19, 22], here
we apply the VA1 (one-dimensional additive veracity) reasoning method, that proceeds by
computing the degree of veracity t j (A) = max{r+j (A)− r−j (A), 0} for each of the R available
rules Rq : Aq ⇒ C j , and then obtaining the evidence degree t j (x) for the classification of the
scenario x in each class C j by means of the expression
μAq (x) · t j (Aq )
q=1,...,R
t j (x) = .
μAq (x)
q=1,...,R

Therefore, the vector t(x) = (t1 (x), . . . ,tN (x)) constitutes the final output of SEDD, as-
sessing the degree of evidence for each class or level of consequences. If a crisp prediction
is needed, then it is usual to assign the scenario x to the class with maximum evidence, i.e.
to the class Ch such that th (x) = max j t j (x).

10.4.2 Dissimilarity structures for disaster assessment

Here we illustrate the effect of different dissimilarity structures on the assessment pro-
vided by SEDD. Particularly, in order to be able to produce a picture of the assessments
obtained by using each dissimilarity matrix Δ, we drop the population density POP from
the set of independent variables, which leaves the variables MAG and HDI as the only
explanatory variables to be used. Similarly, in this example we will focus on just one
consequence variable, the number of casualties CAS, and on one type of disaster, earth-
quakes. Therefore, it is n = 2 and N = 5 (the same classes as in Table 10.1 are used for
the variable CAS). The training sample for these explanatory and dependent variables is
shown in Fig. 10.3, and it is m = 386. As the maximum number of premises is quite small
(5 · 5 = 25), we adopt a grid-based learning procedure, i.e. rules are built for all possible
premises. However, a support threshold (δ = 0.01) is defined in order to avoid those rules
built from a too small sample. Therefore, those premises A such that
m
μA (x p ) < δ
p=1

are discarded.

[email protected]
Classification of Disasters and Emergencies under Bipolar Knowledge Representation 225

In order to illustrate and compare the effects of each matrix Δ, the behavior of the
resulting VA1 classifiers is simulated in a dense mesh of points of the input space of the at-
tributes MAG and HDI, in such a way that a picture of the predictions and class boundaries
produced by each dissimilarity structure is obtained. Also, two error measures are used in
order to measure the performance of the different classifiers applied below: 1) %CC rep-
resents the rate of correct classifications obtained over the training sample, thus evaluating
the predictive accuracy of each classifier; 2) to measure both the deviation of the predic-
tions from the real classes and the risk of underestimation, the average cost AVCOST of the
predictions is computed over the training sample, where the cost of classifying a instance
from the class i in the class j is given by the element COS Ti j of the matrix

0 1 2 3 4
2 0 1 2 3
COS T = 4 2 0 1 2 .
6 4 2 0 1
8 6 4 2 0

10.4.2.1 No dissimilarity

Let us start by assuming that no opposition relationships hold between the classes, i.e.
by taking Δ ≡ 0. In this case, it is r−j (A) = 0 for every premise A and consequent C j , so it
is t j (A) = r+j (A) for all rules A ⇒ C j . Therefore, this case corresponds to a non-structured
classification framework, in which a usual non-bipolar fuzzy classifier, identical to that
used by SEDD in [23], is obtained. As no structure is assumed on the set of classes, the
resulting classifier treats all the classes as independent items. Consequently, the classifier
will be biased towards the classes with more training examples. This is clearly shown
in Fig. 10.5, where the results of the simulation are depicted. Note that a huge part of
the input space of the attributes is assigned to the lowest class CAS1 = no casualties, i.e.
that with the highest proportion (54.4%) of training patterns. This entails a great risk of
underestimation of consequences. Furthermore, the behavior of the classifier is not smooth
at all, since predictions present sharp variations. Table 3 presents the performance measures
for this non-bipolar classifier. Though we will use these results for comparison with the
rest of classifiers, it is important to remark that almost all the correct classification rate
(%CC= 54.15) is due to examples of class CAS1. In fact, notice that %CC is in this case
almost equal to the proportion of examples from the class CAS1.

[email protected]
226 Decision Aid Models for Disaster Management and Emergencies

Fig. 10.5 Simulation result of the non bipolar classifier with Δ ≡ 0.

Table 10.3 Performance measures of the non-bipolar classifier (Δ ≡ 0).

%CC AVCOST
54,15 1,453

10.4.2.2 Total opposition

As explained above, dissimilarity matrices move between the extreme cases given by
Δ ≡ 0 and Δ = 1 − Id. Let us analyze now this last case, that corresponds to a situation in
which each class is totally opposite to the others. Notice that now it is r−j (A) = 1 − r+j (A),
and thus t j (A) = max{2r+j (A) − 1, 0}. Therefore, since all classes are equally related,
the same bias as before towards the more abundant classes is obtained. However, a rule
A ⇒ C j will obtain t j (A) = 0 unless r+j (A) > 0.5, i.e. unless more than a half of the training
examples compatible with the premise A belong to class C j . If no class fulfill this condition,
then it is t j (A) = 0 for all j. As a result, it is possible for a query x to keep unclassified if
no activated rule have a positive veracity degree. The simulation of this classifier clearly
illustrates this point, as shown in Fig. 10.6. A half of the input space is left unclassified,
while the other half is assigned to the majority class CAS1, the only one that reaches the 0.5
threshold for some premises. In this sense, if the previous classifier predicts the class with
the greatest (positive) evidence, then it is possible to say that the present classifier only give
a prediction if the evidence for a class is much larger than for the others. To some extent,
this classifier can be associated with a requirement of not giving an assessment unless
robust, strongly supported predictions are feasible. In this sense, it can be used for a first
assessment of whether a scenario have no consequences at all or not. Table 10.4 shows the

[email protected]
Classification of Disasters and Emergencies under Bipolar Knowledge Representation 227

Table 10.4 Performance measures under the assumption of a total opposition

between classes (Δ ≡ 1 − Id).
%NC %CC AVCOST
43.78 38.86 0.811

performance measures of this classifiers. Notice that 43.78% and 38.86% of the m = 386
training examples are respectively left unclassified and correctly classified. This gives an
error rate of only 17.36% (the previous was 45.85%), though at the price of not-classifying
almost a half of the sample. Similarly, the average cost is significantly lower, since a great
part of the instances from the highest classes are left unclassified.

Fig. 10.6 Simulation result of the non bipolar classifier with Δ ≡ 1 − Id.

10.4.2.3 Restricted asymmetric linear order

In order to reproduce the semantics of linear order associated with the classes, let us
now introduce the dissimilarity matrix
0 0.2 0.2 0.5 1
0 0 0.2 0.5 1
Δ1 = 0 0 0 0 0 .
0 0 0 0 0
0 0 0 0 0
This matrix produces a structure in which the higher classes are gradually more and
more dissimilar to the two lowest classes, CAS1 and CAS2, but not symmetrically, i.e.,
the lower classes are not dissimilar to the higher ones. Moreover, classes CAS3-CAS5

[email protected]
228 Decision Aid Models for Disaster Management and Emergencies

Table 10.5 Performance measures for the restricted asymmetric lin-

ear structure (Δ ≡ Δ1 ).
%CC AVCOST
55.18 1.069

are completely unrelated between them. Therefore, in this situation the classes CAS1 and
CAS2 receive negative information from the higher ones in a progressive way, but not con-
versely, representing the linear structure of the classes together with the requirement of
avoid underestimation risk. As a consequence, classes CAS1 and CAS2 will obtain a lower
veracity degree in the presence of the higher classes, thus requiring more evidence for the
former classes in order to be predicted. However, if it is estimated that such a risk can be
disregarded for the higher classes (e.g. if CAS3 and higher scenarios are always further
assessed), then it is possible to restrict the linear order structure to the first two classes
and allow the classes to compete freely between them, similarly to what happened for all
the classes when Δ ≡ 0. These are the assumptions behind the matrix Δ1 above. Com-
pared to those of the non-bipolar classifier in Fig. 10.5, the simulation results now show
a smoother behavior, in which the class CAS3 (in light blue in Fig. 10.7) appears in the
transition zone between the lower and the upper classes. Notice that, in general, the upper
classes obtain a greater portion of the input space than before. In fact, as shown in Ta-
ble 10.5, this classifier obtains an average cost of 1.069, thus reducing the underestimation
risk of the non-structured case. Furthermore, this classifier obtains a better classification
rate (%CC = 55.18) than the non-bipolar one, i.e. the consideration of a dissimilarity struc-
ture leads in this case to a more accurate classifier than without it.

10.4.2.4 Worst-scenario analysis

Finally, let us consider the requirement of a total avoidance of the underestimation risk.
To this end, consider now the dissimilarity matrix
0 1 1 1 1
0 0 1 1 1
Δ2 = 0 0 0 1 1 .
0 0 0 0 1
0 0 0 0 0
This leads to an structure in which each class is totally opposite to all the classes lower
than it, but not conversely. Therefore, in this setting there is not a explicit linear structure,
but just the assumption of a total asymmetry between the lower and the higher classes.

[email protected]
Classification of Disasters and Emergencies under Bipolar Knowledge Representation 229

Fig. 10.7 Simulation result of the non bipolar classifier with Δ ≡ Δ1 .

Table 10.6 Performance measures for the worst case scenario analysis (Δ ≡ Δ2 ).
%CC AVCOST
50.52 0.968

Each class receives the positive confidence of all the higher classes as negative information.
Thus, the lower the class, the harder it is for such class to be predicted. In this sense,
matrix Δ2 fits to the requirement of performing a worst case analysis of the disaster scenario
under study. Notice that in this setting, no class is guaranteed to be predicted unless it
attains a positive confidence r+ > 2/3, except the highest, which is predicted whenever
its confidence is bigger than 1/3. Consequently, the class CAS5 appears for the first time
in the simulation results of this classifier, as shown in Fig. 10.8. Note also the improved
smooth behavior of the classifier, producing a soft transition between classes. In fact, a
straight line could be drawn in the input space passing through all the classes in order.
More importantly, this classifier enable to distinguish a clear trend in the consequences,
in such a way that worst consequences are associated with lower HDI values (and thus
with a greater vulnerability) and greater intensities of earthquakes. This trend is logically
expected, but notice that no one of the previous classifiers could express it so clearly. The
performance measures of this classifier, shown in Table 10.6, presents a further reduction
of the average cost (AVCOST = 0.968) and thus of the underestimation risk, that could be
even more important since the rate of correct classification (%CC = 50.52) is lower than
before (with the subsequent increment of non-zero costs).

[email protected]
230 Decision Aid Models for Disaster Management and Emergencies

Fig. 10.8 Simulation result of the non bipolar classifier with Δ ≡ Δ2 .

10.5 Conclusions

In [23], the problem of assessing the severity of a potential disaster from the first in-
formation available, a key problem in humanitarian logistics, was reformulated as a clas-
sification problem, in which the classes are associated to the different orders of magnitude
of the consequences of such a disaster. However, the semantics of these classes makes it
necessary to consider them as related, in such a way that the set of classes can be regarded
as having a certain structure. This leads to the notion of structured classification prob-
lem, which has been studied in this work in relation with the abovementioned problem of
disaster severity assessment.
Particularly, it has been shown that different semantic (e.g. the linear order of the
classes) and decision-related (e.g. the avoidance of the underestimation risk) requirements
of the NGO context can be associated to different structures of such set of classes, which
have been constructed and introduced in the classification models by means of the notions
of semantic antagonism and dissimilarity structure. As a result, a bipolar methodology for
SEDD has been obtained, which allows a more accurate and adapted assessment of disaster
consequences.

Acknowledgments

This work has been partially supported by grant TIN2009-07190.

[email protected]
Bibliography 231

Bibliography

[1] Agrawal, R., Imielinski, T., Swami, A. (1993) Mining Association Rules Between Sets of Items
in Large Databases, SIGMOD Conference, 207–216.
[2] Aleskerov, F., Iseri Say, A., Toker, A., Akin, H.L., Altay, G. (2005) A cluster-based decision
support system for estimating earthquake damage and casualties, Disasters, 29 (3) 255–276.
[3] Atanassov, K.T. (1986) Intuitionistic Fuzzy-Sets, Fuzzy Sets and Systems, 20 (1) 87–96.
[4] Billa, L., Shattri, M., Mahmud, A.R., and Ghazali, A.H. (2006). Comprehensive planning and
the role of SDSS in flood disaster management in Malaysia. Disaster Prevention and Manage-
ment, 15 (2), 233–240.
[5] Casillas, J., Cordón, O., Herrera, F. (2002) COR: A Methodology to Improve Ad Hoc Data-
Driven Linguistic Rule Learning Methods by Inducing Cooperation Among Rules. IEEE Trans.
on Systems, Man and Cybernetics-Part B: Cybernetics, 32 (4) 526–537.
[6] Cordón, O., del Jesús, M.J., Herrera, F. (1999) A proposal on reasoning methods in fuzzy rule-
based classification systems, International Journal of Approximate Reasoning, 20 (1) 21–45.
[7] Drabek, T.E., Hoetmer, G.J. (1991) Emergency management: principles and practice for local
government, International City Management Association, Washington DC.
[8] Dubois, D., Prade, H. (2008) An introduction to bipolar representations of information and
preference, International Journal of Intelligent Systems, 23 (8) 866–877.
[9] Hullermeier, E. (2005) Fuzzy methods in machine learning and data mining: Status and
prospects, Fuzzy Sets and Systems, 156 (3) 387–406.
[10] Ishibuchi, H., Nakashima, T. (2001) Effect of rule weights in fuzzy rule-based classification
systems, Ieee Transactions on Fuzzy Systems, 9 (4) 506–515.
[11] Ishibuchi, H., Yamamoto, T. (2005) Rule weight specification in fuzzy rule-based classification
systems, Ieee Transactions on Fuzzy Systems, 13 (4) 428–435.
[12] Kuncheva, L.I. (2000) Fuzzy classifier design, Physica-Verlag, Heidelberg.
[13] Mendonça, D., Beroggi, E.G., Wallace, W.A. (2001) Decision support for improvisation during
emergency response operations, International Journal of Emergency Management, 1 30–38.
[14] Montero, J., Gomez, D., Bustince, H. (2007) On the relevance of some families of fuzzy Sets,
Fuzzy Sets and Systems, 158 (22) 2429–2442.
[15] Morrow, B.H. (1999) Identifying and mapping community vulnerability, Disasters, 23 (1) 1–18.
[16] Nauck, D., Kruse, R. (1999) Obtaining interpretable fuzzy classification rules from medical
data, Artificial Intelligence in Medicine, 16 (2) 149–169.
[17] Olsen, G.R., Carstensen, N., Hoyen, K. (2003) Humanitarian crises: What determines the level
of emergency assistance? Media coverage, donor interests and the aid business, Disasters, 27
(2) 109–126.
[18] Prins, E.M. and Menzel, W.P. (2004) “Geostationary satellite detection of biomass burning in
South America,” Intl. J. Remote Sensing, 13: 49–63.
[19] Rodríguez, J.T. (2010) Clasificación de desastres y emergencias con representación bipolar del
conocimiento, Ph.D. Thesis, Complutense University of Madrid.
[20] Rodriguez J.T., Franco C.A., Vitoriano B., Montero J. (2011) An axiomatic approach to the
notion of semantic antagonism, Procs IFSA-AFSS’11 FT104-1/6.
[21] Rodríguez, J.T., Vitoriano, B., Montero, J. (2010) A natural-disaster management DSS for
Humanitarian Non-Governmental Organisations, Knowledge-Based Systems, 23 (1) 17–22.
[22] Rodríguez, J.T., Vitoriano, B., Montero, J. (2011) Rule-based classification by means of bipolar
criteria. 2011 IEEE Symposium on Multicriteria Decision Making (SSCI-MCDM) 197–204.
[23] Rodríguez, J.T., Vitoriano, B., Montero, J. (2012) A general methodology for data-based rule
building and its application to natural disaster management, Computers & Operations Research,
39 (4) 863–873.

[email protected]
232 Decision Aid Models for Disaster Management and Emergencies

[24] Rodríguez, J.T., Vitoriano, B., Montero, J., Omaña, A. (2008) A decision support tool for
humanitarian operations in natural disaster relief, Computational Intelligence in Decision and
Control, 1 805–810.
[25] Rodríguez, J.T., Vitoriano, B., Montero, J., Kecman V. (2011) A disaster-severity assessment
DSS comparative analysis. OR Spectrum, 33 (3) 451–479.
[26] Schneider P.J., Schauer B.A. (2006) Hazus - its development and its future. Natural Hazards
Review, 7 (2) 40–44.
[27] Schweizer B., Sklar A. (1983) Probabilistic Metric Spaces. North–Holland/Elsevier, New York.
[28] Stoddard, A. (2003) Humanitarian NGOs: Challenges and Trends. In: J. Macrae and A. Harmer
(eds.), Humanitarian Action and the ‘Global War on Terror’: A Review of Trends and Issues
HPG Report 14, ODI, London.
[29] Van Wassenhove, L.N. (2006) Humanitarian aid logistics: supply chain management in high
gear, J. Oper. Res. Soc, 57 (5) 475–489.
[30] Vitoriano, B., Ortuño, M.T., Tirado, G., Montero, J. (2010) A multi-criteria optimization model
for humanitarian aid distribution, Journal of Global Optimization (JOGO), 51 189–208.
[31] Wallace, W. A., De Balogh, F. (1985) Decision Support Systems for Disaster Management,
Public Administration Review, 45 134–146.
[32] Zadeh, L.A. (1965) Fuzzy Sets, Information and Control, 8 (3) 338–353.
[33] Zadeh, L.A. (1975) Concept of a Linguistic Variable and Its Application to Approximate Rea-
soning 1., Information Sciences, 8 (3) 199–249.

[email protected]
Chapter 11

A Network Transshipment Model for Planning

Humanitarian Relief Operations after a Natural
Disaster

Alistair Clark∗ and Bernadette Culkin

Department of Engineering Design and Mathematics, University of the West of England,

Bristol, BS16 1QY, United Kingdom

Every year, natural disasters and humanitarian crises affect approximately 200 million peo-
ple, requiring the quick movement of goods and people to ease the human suffering and to
return the population to some sense of normality. With contributions to humanitarian relief
programmes falling short of what is required, programme managers need to become more
cost-efficient and do more with less. The field of operational research (OR) has developed
many models to help the commercial sector examine current practices and find ways of
becoming more cost efficient. However, much of this good practice has not transferred to
the humanitarian field.
This paper develops a mathematical transshipment multi-commodity supply-chain flow
model for use within humanitarian relief operations. A small data set, based on real life
data from the South Asian Earthquake of October 2005, is used to validate the model so-
lutions compared to the real life situation. Several variants of the model are developed to
add realism and flexibility over a number of possible scenarios. From the variant solutions
several recommendations are made to provide guidance on planning for humanitarian relief
operations.

11.1 Introduction

Every year, natural disasters and humanitarian crises affect approximately 200 million
people, often temporarily displacing up to five million people [36]. Responding rapidly
to these situations relies on the quick movement of goods and people to ease the human
suffering and to return the population to some sense of normality.
A disaster, which can take a variety of forms, is described in Van Wassenhove [42]
as a disruption that physically affects a system as a whole and threatens its priorities and
∗ Corresponding author: [email protected]

B. Vitoriano et al. (eds.), Decision Aid Models for Disaster Management 233
and Emergencies, Atlantis Computational Intelligence Systems 7,
DOI: 10.2991/978-94-91216-74-9_11,  Atlantis Press 2013

[email protected]
234 Decision Aid Models for Disaster Management and Emergencies

goals. However, the occurrence of a disaster does not necessarily mean that humanitarian
assistance is required. Many disasters occur throughout the year without becoming a crisis.
It is neither the size nor magnitude of the disaster that constitutes a crisis, but rather the need
for intervention. If the disaster can be dealt with using existing routines then the situation is
an emergency rather than a crisis [38]. A crisis evolves into a disaster when it challenges the
traditional values shared by the organization, the incumbent response mechanisms become
saturated and overwhelmed, and external help is required [38]. Such foreign intervention of
food, shelter and services by other governments and non-government organisations (NGOs)
is known as relief or humanitarian assistance [23, 26, 36].
When a disaster occurs, the right goods and people must be sent to the right place, at
the right time and in the right quantity [21]. This movement of goods and people accounts
for 80% of any disaster relief operation and is often the most expensive part of any relief
operation, making it a crucial factor in determining whether the humanitarian effort has
been a success or a failure [36, 42].
However, each year contributions of resources have often fallen short of what is re-
quired, leading to humanitarian organisations managing programmes more efficiently, do-
ing more with less [26]. Cost efficiency has been a particular focus within the commercial
sector for a number of years, achieved by making improvements to supply chains through
the examination of cross-functional solutions to address some of the barriers that inhibit
improvements [15]. However, much of this good practice has not crossed over to the hu-
manitarian sector where organizations are 15 to 20 years behind their private sector coun-
terparts in terms of supply chain development [36, 42]. Indeed, as one researcher puts it,
it is paradoxical that a sector which has such extreme requirements in terms of timeliness,
affordability and oversight is so underdeveloped. It is precisely this paradox that creates
what we see as a great opportunity for advancement of the field and of the humanitarian
mission [36].
Operational Research (OR), known as Operations Research in the USA, has contributed
widely to the commercial sector by developing models to help plan logistics and transporta-
tion operations. However the humanitarian sector has been relatively neglected [5] until
recently. A post-1980 literature survey [1] of OR research and disaster management opera-
tions found just 109 articles relating specifically to the use of OR in disaster management,
of which 31 were related to natural disasters and just 1 to humanitarian situations. How-
ever, since 2006, there has been an increased interest in humanitarian logistics as evidenced
by recent overviews [8, 43].

[email protected]
A Network Transshipment Model for Planning Humanitarian Relief Operations after a Natural Disaster 235

Yet the aim of humanitarian relief operations is to mitigate the urgent needs of a pop-
ulation with sustainable reduction of their vulnerability in the shortest amount of time
and with the least amount of resources [38], a challenge that is ideal for OR approaches
such as mathematical optimisation and discrete-event simulation. The Association of Euro-
pean Operational Research Societies has also recognised the growing need to develop new
methodologies or new variants of old ones, such as emergency logistics, devoting its Man-
agement Science Strategic Innovation Prize in 2006 to this area. The British OR Society’s
2006 annual Blackett Lecture was also devoted to humanitarian aid logistics [42].
As a contribution in that direction, this paper develops a transshipment network flow
model to assist policy makers and planners to design an effective supply chain so that goods
and service reach those in need as quickly as possible. The paper is organised as follows:
Section 11.2 defines the humanitarian principles and discusses the planning required for
effective supply chain management. Section 11.3 then considers the models that have been
developed for use in humanitarian relief operations. Section 11.4 goes on to develop an
initial network flow model. Several variants of this model are proposed and tested with
South Asian data in Section 11.5. The modelling and optimisation outcomes are discussed
from a strategic planning perspective. Section 11.6 concludes by discussing some general
implications for planners of supply chains in humanitarian relief operations.

11.2 Planning For Humanitarian Relief Operations

Planning for relief operations needs to abide by the humanitarian principles which are
fundamental to the operations of humanitarian organisations [42]. Three principles define
humanitarianism: humanity, impartiality and neutrality. These state, in short, that suffering
will be alleviated wherever it is found, giving priority to the most urgent needs and without
discrimination [38, 42].
These principles are described in the Code of Conduct for the International Federation
of Red Cross and Red Crescent Societies [19], developed in 1994 to define a standard
of behaviour expected of NGOs working in disaster-affected countries. Following this, a
group of organisations drew up a humanitarian charter in 1997, together with minimum
standards to be attained during any relief operation which came to be known as the Sphere
Project [35]. These principles guide the work of NGOs and their planning activities.
Planning, or preparedness, outlines a set of actions to be taken in the event of a disaster
occurring [1]. It is essential to ensure proper co-ordination and anticipate problems that

[email protected]
236 Decision Aid Models for Disaster Management and Emergencies

may occur in the supply chain at an early stage [2, 13, 24, 27, 28, 37, 42]. Put simply, pre-
paredness is essential for a timely, competent and cost-effective emergency response [27].
Planning for efficient humanitarian supply chains can be seen as a part of disaster recov-
ery planning, defined by the United Nations International Strategy for Disaster Reduction
as a set of actions, arrangements and procedures taken in anticipation of an emergency to
ensure a rapid, effective and appropriate response that may save lives and livelihoods [40].
The World Conference on Disaster Reduction in January 2005 called for governments to
prepare for effective response and recovery by identifying and allocating existing resources
from the establishment, development and emergency budgets for disaster and risk manage-
ment to greater effect in the realization of sustained risk reduction [41]. This is endorsed
by NGOs who state in their Code of Conduct that governments should seek to provide a
coordinated disaster information and planning service [20].
Many authors, e.g. Oloruntoba and Gray [28], report that there is frequently a lack of
planning in humanitarian supply chains, resulting in inefficiencies. Some authors believe
that planning is difficult during the initial stages of the humanitarian response because each
disaster is unique [23, 27]. However, the Pan-American Health Organisation [32] disputes
this point, stating that most disasters and their arising needs are usually predictable and so,
by studying past humanitarian assistance programmes, accurate forecasts can be made [27],
allowing regions at risk to prepare themselves and for relief agencies to prepare their efforts
[23, 46]. In this way, an international humanitarian organisation was able to devise a top
ten list of commodities required in most emergencies and arrange pre-purchase agreements
with suppliers [25].
Rather than planning for cost-effective programmes, McGuire [27] describes the ob-
jective of humanitarian relief operation in its initial stages as being able to improvise and
set up a supply chain which can deliver at all, rather than contemplating optimal and cost-
efficient solutions. However, Van Wassenhove [42] affirms that a successful response to
a disaster is not improvised. This view is also endorsed by the PAHO who state that the
erroneous idea that logistics may be improvised at the moment of a disaster depending on
needs indicated by the situation must be eliminated [32].
Two types of planning are identified in [22], namely, strategic and tactical.

(1) Strategic Planning can be thought of as long term planning. It identifies available re-
sources and allows policy-makers and planners to assess the strengths and weaknesses
of a system based on a number of likely disaster scenarios.

[email protected]
A Network Transshipment Model for Planning Humanitarian Relief Operations after a Natural Disaster 237

(2) Tactical/Operational Planning is short term planning and plans current daily or weekly
operations. In the humanitarian sector this would occur when a disaster strikes and
humanitarian aid is needed.

Strategic planning is common within the commercial sector. For example, the Kellogg
Company inputs estimated costs and demand forecasts into their strategic model in order to
optimally source products, allowing them to establish financial budgets, space for inventory
and transportation requirements [6]. The models are also used to test alternative scenarios
of a particular problem. Proctor and Gamble have also used such models to look at the
impact of closing particular plants [7]. The use of strategic models, as in the commercial
sector, can aid in identifying necessary resources and budgets.
Strategic planning practice can be used within the humanitarian sector through the use
of data from past operations which can provide important post-event learning [36]. The use
of feedback models can help operational decision-makers in humanitarian organizations
understand the complexity of humanitarian relief efforts and learn how to design, plan and
manage such operations [39]. One reason why these types of strategies have not yet filtered
into the humanitarian sector could be due to lack of financing as NGOs can find funds for
relief but not for planning for relief [42].

11.3 Models Developed by other Researchers

Many studies into disaster management make use of mathematical programming

and incorporate well established models such as goal programming [29, 45], inventory-
allocation models [17, 34], supply chain models [33], vehicle routing models [31], location-
distribution models [47], network models [3, 9, 30], multi-criteria optimization models
[16, 44] and inventory-control models [4]. A variety of objective functions have been used,
the most common of which is minimising the cost of the operation [3]. However, planning
operations based on costs alone may conflict with the IFRC Code of Conduct [19], which
states that aid should be distributed based on need.
The IFRC Code of Conduct is adhered to by Ozdamar et al. [31] who sets the ob-
jective function to minimise the amount of unsatisfied demand in their model. Unsatisfied
demand is also included in the objective function by Yi and Ozdamar [47] together with a
priority weighting for each commodity. The model developed by Hwang [17] minimises
the amount of pains and starving , but there is no information about how this is calculated.
However, cost cannot be totally ignored because humanitarian organisations are becoming

[email protected]
238 Decision Aid Models for Disaster Management and Emergencies

increasingly accountable to their donors on how funds are spent [42] and are required to
plan more cost-effective programmes due to a shortfall in funding contributions [27]. This
is acknowledged in [3] who not only minimise the costs of the operation but also include a
penalty cost for unmet demand.
All the authors above acknowledge that their models can be used to plan for humanitar-
ian relief operations. However, few authors makes recommendations for use in preparing
for a disaster. Instead, the focus tends to be on inputting data and analysing the results from
that particular data set.
One of the difficulties in planning for humanitarian relief operations is that information
changes quickly, particularly at the beginning of operations, a reality that must be reflected
in any model that is developed. The model in [31] incorporates this by allowing new plans
to be developed at given time intervals once updated information has been received. The
model in [3] is an attempt at planning where supply and arc capacities are determined in
the pre-event stage for each earthquake scenario. This information is then extracted for the
actual event.
It is not clear whether these models are flexible enough to use in other kinds of disasters.
The model in [31] focuses specifically on logistics planning after a natural disaster, but it
is only tested on a relief operation after an earthquake, and gives no indication of whether
it is suitable in other natural disasters. The model in [3] also gives little indication of its
generality. Although its development was based on a number of earthquake scenarios, it is
not clear whether it is applicable to other types of disaster.

11.4 A Network Transshipment model for Emergency Relief

We now develop a basic transshipment model for exploratory strategic planning. A

humanitarian supply chain can be thought of as starting at multiple suppliers, then passing
through various transshipment points and finally arriving at the multiple recipients.
Figure 11.1 shows a humanitarian supply chain as a network transshipment model in
which items are pulled through the supply chain to meet the demands of the recipient. The
quantity of items sent along the chain is dependent upon the demand at the final destination.

The model is strategic, rather than operational, and aims to identify whether there are
sufficient resources, suppliers, modes of transport, inventory and storage capacity along the
supply chain. It includes all the necessary nodes within the supply chain as defined in the

[email protected]
A Network Transshipment Model for Planning Humanitarian Relief Operations after a Natural Disaster 239

Supplier Supplier

Airport, Seaport or
Train Terminal

Warehouse

Recipient Recipient Recipient

Fig. 11.1 A humanitarian supply chain

conceptual model. Suppliers, warehouses, airports, seaports and rail terminals and recip-
ients are all nodes along the supply chain with possible intermediate demand for certain
items.
Users of the model can flexibly define supply and demand parameters at each node
as well as capacity levels for flows along each arc. The model can represent a variety of
resources and constraints such as the type and number of transport vehicles available or
limits on the number of items sent to each warehouse. The use of a flow model also allows
restrictions on arcs to model the impacts of natural disasters on the supply chain such as
impassable roads.

Comment. To specify and formulate the model more precisely, consider the following
index sets:

S suppliers
W warehouses
R recipients (people)
A airports

[email protected]
240 Decision Aid Models for Disaster Management and Emergencies

F seaports
B rail terminals
I items
E types of vehicles available
K of nodes k ∈ {S ∪W ∪ R ∪ A ∪ F ∪ B} of the supply chain
G of intermediate nodes k ∈ {S ∪W ∪ A ∪ F ∪ B}
L links between successive nodes

The index i represent items transported, and v denotes vehicles. The indices j and
k represents the network nodes in the supply chain such as suppliers, airports, seaports,
terminals, warehouses, and recipients .
The index t represents a day within the planning horizon. Thus t = 1, . . . , 10 means a
ten-day horizon with decisions at the start of each day, including the current one (for which
t = 1). Non-positive values of t represent the past, so that, for example, t = 0 would be
the day before the current one, t = −1, the day before that, and so on. The inclusion of
on-positive values of t is allowed to appropriately model the effect of transportation lead
times, as in, for example, the definition of data parameter Tk0jvit below.
The following parameters contain all the information that planners should (ideally)
know:

Capt available capacity time in each period t.

Iki0 current initial inventory at node k of item i (i.e., at the end of day 0)
Dkit demand at node k for item i on day t
Uki0 current unsatisfied demand at node k for item i
VCv maximum capacity that each vehicle v can carry in metric tonnes
VAkvt number of vehicles available at node k of a vehicle type v on day t (integer)
Lk jv lead time from node k to node j using a vehicle of type v
Tk0jvit amount of item i already sent from node k to node j, using a vehicle of type v, i
at past times t = −Lk jv , . . . , 0
Fk jv fixed transport cost to include driver, fuel, maintenance etc. of using each vehi-
cle type v between node k and node j
Wi weight of each item i
Ck capacity limits of each node k

The decisions to be taken correspond to the following variables:

[email protected]
A Network Transshipment Model for Planning Humanitarian Relief Operations after a Natural Disaster 241

Tk jvit amount of item i sent from node k to node j, using vehicle type v, on day t
Vk jvt number of vehicles of type v sent from node k to node j on day t (integer)

The outcomes of these decisions are the following variables:

Ikit inventory at node k of item i at the end of day t

Ukit unsatisfied demand (backlog) at node k for item i at the end of day t

The model, being strategic, is designed to plan and analyse flows through the supply chain
in advance of a natural disaster occurring, exploring possible scenarios. The decision vari-
ables give information about the amount Tk jvit of items transported from one node to an-
other and the number of vehicles Vk jvt needed to do so. The inventory outcome variable
Ikit allows planners to investigate the amount of storage required for the supply chain to
operate effectively and efficiently. However, the unsatisfied demand outcome variable Ukit
is probably the most important as it provides information on how many recipients will still
be waiting for each item at the end of each day in the planning horizon.
Any humanitarian operation has at its heart the requirement to ease the suffering of
others, and so the objective function of this model is to minimise unmet need, defined as
unsatisfied demand accumulated over time. However, cost cannot be completely ignored
and so needs to be minimized and included in the objective function. This leads to the
use of the weighted-sum multi-objective function (11.1) below that not only minimises
need but also transportation and inventory costs. In using such a multi-objective model,
the organisation should carefully consider its priorities as these can affect the cost of the
humanitarian relief operation. Is minimising unsatisfied demand in the shortest amount of
time more important than minimising transportation costs? What action should be taken
if some recipients do not receive items? How much capacity can be unused in each mode
of transport or should vehicles be filled to capacity? These discussions are then useful in
determining the weightings that should be given to each of the components in the objective
function.
The humanitarian organisations’ Code of Conduct requires that aid must be delivered
based on need and not cost [20, 35]. Accordingly, the objective function (11.1) below
pre-emptively places much greater priority on minimising need before transportation and
inventory costs:
Minimise Ukit + 0.001 Fk jvVk jvt + 0.01 Ikit + Vk jvt (11.1)
kit k jvt kit k jvt
The weight values of 1, 0.001, 0.01 and 1 in expression are somewhat arbitrary, and here
simply reflect the relative magnitude each particular component in the policy of the orga-

[email protected]
242 Decision Aid Models for Disaster Management and Emergencies

nization. For this paper, we wished the items to be delivered to recipients as quickly as
possible, hence the high relative weighting given to kit Ukit . The units of measurement of
each component must also be taken into account. At first glance it may appear that there is
a greater weighting given to inventory than transport. However, the magnitude of the trans-
portation costs means that its weighting has to be scaled down in order for transportation
costs to approximately equal inventory costs. The danger with using weightings is that if
there was a small transportation cost and a large inventory then these weightings may be
inappropriate leading to the inventory component having a greater influence than the trans-
portation costs weighting. The weightings have been verified so that this problem should
not occur. Technical normalisation methods can also be used to help determine appropriate
relative weightings [11].
The final component k jvt Vk jvt of the objective function also minimises the number of
vehicles used. It has been added because transportation costs from suppliers are assumed
to be paid for by the supplier, and so the number of vehicles used are not minimised by the
fixed costs component k jvt Fk jvVk jvt in the objective function. If policy makers knew the
cost for these journeys then this component could be removed from the objective function.
Varying the relative weights in the objective function (11.1) allows different polices to
be explored, and the consequent impact on the time taken to satisfy demand. However, a
more direct and comprehensive way of investigating the impact of different policies, would
be to directly introduce constraints on, for example, the permitted amount of inventory, or
other components of the objective function. This type of multi-objective decision making
would need to be undertaken interactively with planners, making use of methods that ex-
plore the efficient frontier of Pareto-optimal decisions [14], for example, the eta-constraint
technique [10].
To establish initial conditions, constraint (11.2) and (11.3) specifies the current unsat-
isfied demand and inventory of each item at each node, i.e., at the end of day 0:

Uki0 = Uki0 for all nodes k, and items i (11.2)

Iki0 = Iki0 for all nodes k and items i (11.3)

Each time the model is re-run during the course of the relief operation, the values of pa-
rameters Uki0 and Iki0 would be updated to reflect the situation at the time of planning.
Initially, unsatisfied demand Uki0 at day 0 is set at zero because it is assumed that the
recipients had no need for any extra items before the natural disaster occurred, i.e., demand
was being dealt with by existing systems. However, when using the models during relief

[email protected]
A Network Transshipment Model for Planning Humanitarian Relief Operations after a Natural Disaster 243

operations, unsatisfied demand would be the number of recipients still requiring items at
the current point in time.
Constraint (11.4) defines the amount of each item sent between pairs of nodes before
the disaster has occurred but which are still in transit and so have not yet arrived at the
destination node:

Tk jvit = Tk0jvit

for all nodes k & j, vehicles v, items i, and past days t = −Lk jv , . . . , 0 (11.4)

The parameter Tk0jvit has a default value of 0 as it is assumed that it is not known that a
disaster is imminent and therefore items were not pre-ordered. If the natural disaster is
expected then operations planners may have pre-ordered certain items which can then be
entered into the model using Tk0jvit for past days t = −Lk jv , . . . , 0. The same reasoning
applies when the model is re-run during a humanitarian relief operation.
In a network flow model, supply must balance demand, inventories, and backlogs over
time. Constraint (11.5) ensures that no more items are sent than are received and/or taken
from the local inventory:

T jkvi,t−L jkv + Iki,t−1 − Uki,t−1 = Tk jvit + Ikit − Ukit + Dkit

jv jv

for all nodes k, items i and days t (11.5)

Constraint (11.5) states that, on any given day and at any given node, the items arriving from
previous nodes, together with the items inherited from the previous day’s ending inventory,
less the unmet demand from the previous day, should equal in quantity the items sent to
the next nodes, plus the demand and the amount put into inventory, less any unsatisfied
demand.
In the model, unsatisfied demand Ukit can be thought of as backlog. The model does
not allow backlogs to occur at intermediate nodes because it is only at the final destina-
tion (where recipients are based) that planners would need to know whether there is still
unsatisfied demand. Thus unsatisfied demand at intermediate nodes is fixed to be zero.
Observe how the lead time value L jkv in constraint (11.5) is used to correctly calculate
the amount of an item i coming into a given node k via vehicle v during day t. It is not
needed to calculate the amount of items going out of a node. It thus partly determines the
values of Ukit and IUkit which feature directly in the objective function. As a result, lead
time features only indirectly in the objective function (11.1).

[email protected]
244 Decision Aid Models for Disaster Management and Emergencies

The use of lead time in constraint (11.5) highlights one of the main differences between
this and other models developed for humanitarian operations after a natural disaster. Lead
time is an important consideration as it affects the decisions made about sourcing supplies
and consequently how long recipients are deprived of help. It will influence decisions
about the minimum inventory of items to keep in local warehouses or pre-agreed purchase
agreements to deal with initial demand.
Constraint (11.6) ensures that there are enough vehicles to transport items between
nodes:
0.001 i Tk jvit Wi
Vk jvt  for all nodes k and j, vehicles v, and days t (11.6)
VCv
The coefficient 0.001 converts the item weights in kilograms to metric tonnes, the units of
vehicle capacity.
Constraint (11.7) ensures that the weight of all inventory items is within the node ca-
pacity limits, again multiplying by 0.001 to convert from kilograms to metric tonnes:

0.001 Ikit Wi  Ck for all nodes k, items i, and days t (11.7)

i

Thus our basic model is formulated as objective function (11.1) and constraints (11.2) to
(11.7).

11.5 Computational Analysis and Model Refinement

The basic model was tested and further developed over a 10-day planning horizon with
data representing 5 items, 4 vehicle types and 54 nodes as follows: 25 suppliers, 6 ware-
houses, 8 recipients, 8 airports or heliports, 1 seaport, and 6 rail terminals. A larger instance
was then implemented with 5 items and 4 vehicle types as before, but now with 83 nodes
as follows: 49 suppliers, 6 warehouses, 13 recipients, 8 airports or heliports, 1 seaport,
6 rail terminals. The data was adapted from that on the UNJLC website for the Pakistan
Earthquake in 2005.
The model was implemented using AMPL [12] and solved using Cplex 9.1 [18] on a
Sun V208 Dual Opteron 252 processor running under Linux with a processing speed of 2.6
GHz and 4GB of RAM. Almost all the model refinements below took less than 10 seconds
to optimally solve, as discussed in more detail later in the paper.

[email protected]
A Network Transshipment Model for Planning Humanitarian Relief Operations after a Natural Disaster 245

11.5.1 Analysis with the Uncapacitated Model

If there are no capacity limits, then all supplies are dispatched to recipients in the dis-
aster zone on day 1. All supplies are sent to the airport that has the shortest lead times
and cheapest transportation costs as the objective function requires the model to reduce
unsatisfied demand in the shortest amount of time, at minimum cost and with the the least
inventory.
When deciding where to send items, the model considers which recipients are closest
to the point of entry, i.e., which have the shortest lead times. If there is no difference in the
lead time of two or more recipients then the model considers which of these has cheaper
transportation costs.
Helicopters are used to transport items to heliports. Even though they are more expen-
sive, helicopters are chosen because they can get items to where they are needed quickly
thus satisfying the objective that unsatisfied demand should be minimised in the shortest
amount of time. The recipients with the cheapest helicopter rates are chosen to receive
items. Trucks are then used to transport items from heliports to warehouses and onward to
recipients.
Those recipients without access to heliports are in danger of not receiving items be-
cause these need to be delivered by truck, a journey which has a longer lead time than the
same journey by helicopter. Consequently the time taken to minimise unsatisfied demand
increases. In addition, these recipients are also further away from the initial points of en-
try and so have larger transportation costs, increasing the value of the objective function.
Special attention might need to be given to this group of recipients to ensure that they are
given consideration when distributing aid.

11.5.2 Airport Arrival Capacity

The model does not take into account a problem that is identified in many situation
reports, namely airport congestion. Is there enough space for all aeroplanes to land? Are
there sufficient resources to deal with the volume of items being delivered? If not, then
this will delay the arrival of aeroplanes on subsequent days and so prolong the time taken
to deliver items, extending time needed to meet (or minimise) unsatisfied demand. To be
realistic the model should be adapted to include a limit on the number of aeroplanes that
can arrive at each airport on any given day.
To allow planners to investigate the impact of imposing limits on the number of aero-
planes arriving at each airport, an airport capacity parameter AC jvt is added to the model in

[email protected]
246 Decision Aid Models for Disaster Management and Emergencies

an additional constraint (11.8) limiting the number of aeroplanes that can arrive each day.

Vk jv,t−Lk jv  AC jvt for all nodes j, vehicles v, and days t (11.8)

k

Constraint (11.8) states that the total number of vehicles Vk jv,t−Lkvt , sent t −Lk jv days ago
from nodes k to j to arrive on day t, must not exceed the airport capacity AC jvt at node j for
each vehicle type v on a given day t. Indexing the airport capacity parameter over t allows
planners to change the limits on a given day to take account of the number of staff available
to deal with arriving items, space for aeroplanes, weather, etc., at each airport.
When airport capacity is limited, the number of days over which airports are used in-
creases, because there are not enough spaces at each airport for the planes to land and so
the model staggers the delivery of items. This leads to an increase in the amount of time in
which recipients are waiting for items.
Items initially arrive at the airport closest to the disaster zone until it has reached its
capacity. At this point, arrivals are switched to the airport that has the next shortest lead
times and cheapest transportation costs, and so on. The total number of planes increases,
due to the transfer of items by air from airports further away than those closest to the dis-
aster zone. Items are transferred by plane rather than by road due to the objective function,
which requires that unsatisfied demand should be minimised in the shortest amount of time.
The model chooses the route and vehicle with the shortest lead times which results in items
being transferred by aeroplane to the airport closest to the disaster zone where helicopters
and trucks are then used to transport items to their final destination.

11.5.3 Limiting Vehicle Availability at Nodes

So far, there is an unlimited supply of vehicles. Although this provides useful infor-
mation about where vehicles are needed and the ideal number required, it is not realistic.
Adding constraint (11.9) below allows planners to examine the impact on the objective
function of a limited availability VAkvt of vehicles:

Vk jvt  VAkvt for all nodes k, vehicles v, and days t (11.9)

j

Indexing by day t allows planners to flexibly specify the number of vehicles available over
time so as to model, for example, NGOs donating extra vehicles. Note that this is still a
simplistic constraint - it does not model the movement of vehicles among nodes as they
transport items, such as in [9]. Such information has to be explicitly provided on a day-by-
day basis via the values of the parameter VAkvt .

[email protected]
A Network Transshipment Model for Planning Humanitarian Relief Operations after a Natural Disaster 247

11.5.3.1 Limited Helicopters

Initially, VAkvt = 10 for v = Helicopters at each node k and day t. This has no ef-
fect on the arrival of items into the disaster zone. All items are still sent by suppliers on
day 1, arriving at the airport closest to the disaster zone. The model initially uses heli-
copters to transfer items to those recipients with cheaper transportation costs. When all
helicopters have been used, the model then has to consider alternative forms of transport
which still minimise unsatisfied demand in the shortest amount of time. This leads the
model to choose trucks and to send these to the place with the cheapest transportation costs
for trucks. Transferring remaining items by truck ensures that there are no items left in the
inventory, fulfilling the third component of the objective function which states that inven-
tory should be minimised. However, the time taken to deliver items increases, compared
to the uncapacitated model, as items sent by road have longer lead times than if sent by
helicopter.

11.5.3.2 Limited Trucks

A limited availability of trucks at each node causes the use of a greater number of
airports to receive items, even though capacity is not reached at each airport. The reason
for this is because of a problem further down the supply chain.
The objective function requires that items should be delivered as quickly as possible
in order to minimise unsatisfied demand in the shortest amount of time. This results in
items being sent to the airport closest to the disaster zone. Items are initially transferred
by helicopter to those recipients with the cheapest transportation costs. However, delivery
stalls when there are no more trucks available to transport items from heliport to warehouse,
causing an increase in the time taken to deliver items.
One may think that items could be delivered to the heliport and stored there until they
are ready to be transferred by truck but the next component in the objective function states
that inventory should be minimised. Thus, rather than have items stored in inventory, items
are sent to recipients by road, to be delivered the following day. So when after as many
items as possible are delivered by helicopter, the model then uses all its trucks to transfer
deliver items to recipients. When all these trucks have been used, suppliers send items to
the next airport so that the trucks there can be used to deliver items to recipients.
Items are delivered to a greater number of places than when transport is not limited.
Some of these recipients are further away from the initial point of entry, and consequently
have higher transportation costs which contribute to increased transportation costs.

[email protected]
248 Decision Aid Models for Disaster Management and Emergencies

11.5.4 Helicopters and Road Access

Delivery by road when helicopters are unavailable works well when there is no disrup-
tion to the road network. However, should this happen, as occurs frequently after earth-
quakes, then delivery of items may be affected.
Items continue to be sent by helicopter to the recipient with the cheapest transportation
costs, in line with the objective function of minimising unsatisfied demand in the shortest
amount of time and at minimum cost. When the VAkvt = 10 helicopters available on day
t at node k have been used, the model has to look for the next quickest mode of transport
available that will still minimise unsatisfied demand as quickly as possible, namely trucks.
The model then considers which recipient has the shortest lead time and cheapest trans-
portation cost. If this recipient cannot be reached, then the model looks at the next shortest
lead times and cheapest routes, and so on.
Planners may wish to restrict the use of helicopters to serve those recipients where
there is no road access or introduce a new mode of transport, such as an airlift, to drop
items directly to recipients.
When the availability of helicopters is limited, then all helicopters at the initial point of
entry are used. The remaining helicopters at other heliports are not used which is a waste
of a valuable resource which would not occur in the real world. The model needs to be
adapted so that better use is made of this valuable resource.

11.5.5 Pooling Vehicles

Instead of allocating vehicles equally amongst all nodes, they can be pooled to allow a
more effective use of resources. An additional constraint (11.10) is needed.
Vk jvt  V Pvt for all vehicles v, and days t (11.10)
kj
where the new parameter VP vt specifies the number of vehicles of type v in the pool on
day t. Again, indexing the vehicle pool parameter over t allows planners to add additional
vehicles to the pool should they become available. Note that this still does not model the
movement of vehicles among nodes as in [9].

11.5.5.1 Pooling Helicopters

Pooling resources actually increases transportation costs, due to the greater use of he-
licopters than when not pooled. This is due to the fact that helicopters journeys cost twice
as much as trucks, an increase that has a great effect on transportation costs.

[email protected]
A Network Transshipment Model for Planning Humanitarian Relief Operations after a Natural Disaster 249

Pooling helicopters allows them to be allocated where they are best used, i.e., at the
initial point of entry to deal with the incoming supplies. In a similar way to the previous
model, the items are sent to those recipients with short lead times and cheaper transporta-
tion routes and therefore satisfies the objective function which states that unsatisfied de-
mand should be minimised in the shortest amount of time with respect to transportation
costs.
When all helicopters in the pool are used, items are sent by road to the recipient with
shortest lead times and cheapest trucking rates. Items could be stored in inventory for de-
livery next day but there is a component in the objective function that states that inventory
should be minimised. In addition, storing items in inventory would not reduce the time
taken to minimise unsatisfied demand nor minimise transportation costs, therefore increas-
ing the objective function.
Pooling helicopters allows a greater number of items to be delivered and therefore un-
satisfied demand is minimised more quickly, compared to when vehicles are not pooled.

11.5.5.2 Pooling Trucks

The pooling of trucks allows the model to allocate them to where they are most needed.
Once again, items are sent to recipients with the cheapest transportation costs and shortest
lead times. The model solution shows that items are initially transported by helicopter to
recipients with cheaper transportation costs. However, transportation then switches to road
which, initially, appears quite surprising as there are an unlimited amount of helicopters
available to transport these items with a lead time of zero. However, if items were trans-
ferred by helicopter, then there are not enough vehicles available to transfer items from the
heliport to the warehouse and then onto recipient. At this point, items would need to be
stored in the inventory. But this does not happen because the minimisation of inventory in
the objective function causes items to be then sent to the recipient with the next cheapest
trucking route, therefore minimising the transportation costs component of the objective
function.
There is a difference in the time taken to deliver items due to the fact that initial delivery
to the disaster zone is delayed so as to minimise the inventory component of the objective
function.

[email protected]
250 Decision Aid Models for Disaster Management and Emergencies

11.5.6 Emergency Contingency Stocks

The models examined so far have assumed that there is no inventory in any of the ware-
houses. We now assumed that there are emergency contingency stocks held in warehouses
within the disaster zone to examine the effect that this has.
Even with the use of emergency contingency stocks, the number of aeroplanes arriving
each day does not differ from the uncapacitated model. This is not surprising as the ob-
jective function states that unsatisfied demand should be minimised in the shortest amount
of time and so all items are sent as quickly as possible to the airport closest to the disaster
zone. There are an unlimited number of helicopters and trucks available to transport items
to warehouses, so inventory will be minimised.
The addition of emergency contingency stocks into the supply chain causes an increase
in the transportation costs (due to the large increase in the number of trucks used to trans-
port the emergency stocks from warehouses to recipients). In addition, unsatisfied demand
initially decreases sharply, but then follows a similar pattern to the uncapacitated model as
more items enter the disaster zone from external suppliers.
The initial unsatisfied demand decreases because all emergency contingency stocks are
sent to the closest recipient on day 1. This is because the objective function states that
unsatisfied demand should be minimised in the shortest amount of time with respect to
transportation costs. This journey has zero lead time and, as items are delivered by truck,
transportation costs are relatively cheap.
Items from external suppliers are sent to those recipients that have short lead times and
cheap transportation costs, as in the uncapacitated model. This is to satisfy the objective
function which states that unsatisfied demand should be minimised in the shortest amount
of time with respect to transportation costs.
Once again, there are recipients who do not receive any items. These recipients are
at a considerable disadvantage to other recipients for a number of reasons. Firstly, they
are furthest way from the initial point of entry and so have larger transportation costs.
Secondly, they do not have a heliport nearby so they have larger lead times as items have to
be delivered by truck. Finally, they do not have a warehouse nearby and so initial demand is
not minimised. The combination of all these factors leads to these recipients not receiving
items.
In all the models examined so far, the decision of where items are delivered is based on
shortest lead times and cheapest transportation costs. However, delivering items based on
speed and cost may violate the NGO’s Code of Conduct [19] which states that humanitarian

[email protected]
A Network Transshipment Model for Planning Humanitarian Relief Operations after a Natural Disaster 251

aid should be delivered based on need. Thus the model should be adapted to include a
parameter identifying relative need and then prioritise aid to where it is most needed.

11.5.7 Prioritising Greatest Need

Prioritising need ensures that those in greatest need receive items first. The objective
function (11.11) includes a weighted average of unsatisfied demand so those recipients with
a greater priority weighting Pk will have a less unsatisfied demand than those deemed to be
of lower priority.

Minimise PkUkit + 0.001 Fk jvVk jvt + 0.01 Ikit + Vk jvt (11.11)

kit k jvt kit k jvt

The weightings for Pk range from 10 (for the location of the greatest need) down to 1
(indicating the lowest need).
The number of aeroplanes landing is the same as in the unprioritised uncapacitated
model. This is to be expected as there are no constraints on how many aeroplanes allowed
to land at each airport nor on the number of vehicles available to transport items from
airports.
However, transportation costs can increase, depending on how the priority weighting
Pk has been allocated. If recipients with a higher priority status are further away from the
initial point of entry, then they will tend to have higher transportation costs. If they do
not have a heliport nearby, then items need to be transported by road. The model solution
in this situation shows an increase in the usage of all forms of transport in order to reach
recipients, contributing to increased transportation costs.

11.5.8 Minimum Delivery to each Recipient

Another way of ensuring a fairer distribution of food is to enforce a distribution policy

whereby a minimum amount Mi j of each item i is delivered to each node j is incorporated
into a new constraint:

Tk jvi  M ji for all nodes j and items i (11.12)

kv

Indexing Mi j by end-node j allows planners to set minimum delivery levels for each recip-
ient based on a number of factors such as number of people affected, particular needs of
each recipient, and so on.

[email protected]
252 Decision Aid Models for Disaster Management and Emergencies

Again, the number of aeroplanes arriving is the same as in the original uncapacitated
model, the reason being that items will arrive as quickly as possible so that unsatisfied
demand can be minimised in the shortest amount of time.
When a minimum delivery policy is introduced, transportation costs again increase
slightly compared to the original uncapacitated model, as some formerly neglected recipi-
ents have larger transportation costs. The number of helicopters is reduced because some of
these new recipients do not have a heliport nearby and so their items are delivered by road,
for example, Abbottabad. The original helicopter-aided recipients of these items receive
less than previously, thus reducing the need for helicopters.
All recipients at node j receive the minimum number Mi j of items i, as declared in the
data set. The remaining items are sent by helicopter to those recipients with the shortest
lead times and cheaper transportation costs, as the objective function requires that unsatis-
fied demand should be minimised in the shortest amount of time with respect to transporta-
tion costs. If demand has been satisfied then any remaining items are sent by helicopter to
the recipient with the next cheapest transportation costs. Items are sent to other recipients
via heliports, as this is the quickest way of getting items closest to recipients and allows
unsatisfied demand to be minimised more quickly.
It is possible to combine the use of a priority weighting Pk and minimum delivery
amount Mi j so that all recipients receive a minimum number of items but then priority is
given to those recipients with a higher priority weighting.

11.5.9 Reducing Unused Vehicle Capacity

In all the models examined so far, there have been situations when the number of items
transported has been very small, leading to a great deal of unused space in vehicles. This
is not a cost-effective way of delivering items which, in addition, may cause upset to those
recipients who are in desperate need of goods if only 10 items arrive for over 100,000
people. Even though unsatisfied demand is minimised in the shortest amount of time, it is
not a policy that would be endorsed by most organisations who would want resources to be
used to maximum effect. One way of overcoming this is to reduce the amount of slack that
is left in each vehicle. Thus more items will be put onto each vehicle enabling much more
cost effective operations.

[email protected]
A Network Transshipment Model for Planning Humanitarian Relief Operations after a Natural Disaster 253

Constraint (11.13) calculates the amount V Sk jvt of vehicle slack:

V Sk jvt = Vk jvt VCv − 0.001 Tk jvit Wi  0

i

for all nodes k & j, vehicles v, and days t (11.13)

Slack is then minimised in the revised objective function (11.14):

Minimise Ukit Pk + 0.001 Fk jvVk jvt + 0.01 Ikit + Vk jvt + V Sk jvt

kit k jvt kit k jvt k jvt
(11.14)
When vehicle slack is minimised, there is no effect on the higher priority components of
the objective function. The total number of items delivered to each recipient in the amount
of time taken to deliver these items remains the same. However, as desired, the number of
deliveries of less than 100 items is reduced, as are the number of vehicles used and trans-
portation costs. This further reduces the value of the objective function as transportation
costs are reduced, therefore providing a much more cost effective operation.

11.5.10 Computational times

With the smaller data set, the models had between 10,500 and 11,600 integer variables.
All but two of the model refinements above took less than 10 seconds to optimally solve,
with half taking under 1 second. However, the truck pooling model took 57.7 seconds
to solve, probably due to the small number of trucks available, resulting in a tightly con-
strained model. The capacitated landing model, which took 13.1 seconds, was also tightly
constrained with few aeroplanes being allowed to land at each airport.
The fast solution times make the model highly suitable for multiple runs during an
interactive session with decision-making end-users to explore different policies.
However, under the larger data set, no optimal solution was provably found after 4
hours for capacitated-trucks model. The long solution time is due to the very small number
of trucks available to transport all items to recipients, making it tightly constrained. The
addition of extra suppliers and recipients has increased the complexity of the model to
find an optimal solution. Under these circumstances, computation needs to be stopped
after some predetermined amount of time, and the incumbent solution used or some other
heuristic method used.

[email protected]
254 Decision Aid Models for Disaster Management and Emergencies

11.6 Recommendations and Conclusions

This paper develops a mathematical transshipment multi-commodity supply-chain net-

work flow model to aid policy makers plan for effective humanitarian relief operations after
an earthquake. A small data set, based on real life data from the South Asian Earthquake
of October 2005, is used to validate the model solutions compared to the real life situation.
The model aids the planning process by allowing policy makers to test a number of scenar-
ios so that a better understanding of the strengths and weaknesses of the supply chain can
be developed.
The model can be initially used as an uncapacitated model to provide useful informa-
tion about budget requirements, numbers and types of vehicles required, capacity limits at
initial points of entry and at warehouses, time taken to satisfy demand, whether the num-
ber of existing suppliers is satisfactory to meet the needs of a given scenario, etc. Several
variants of the model are developed to add realism and flexibility over a number of pos-
sible scenarios thus allowing policy makers to examine the consequences of their real-life
situation such as limited availability of vehicles, restricted capacity at initial points of en-
try and warehouses, road closures etc. The multi-criteria nature of the objective function
also obliges planners to consider and clarify their priorities - indeed this is one of the most
useful purposes of the model.
Based on the analysis above, general guidance on the use of the model can be synthe-
sised. Table 11.1 lists a number of problem scenarios that planners may be faced with,
together with recommendations on how to use the model to overcome each problem. The
recommendations concern the allocation of resources and the kind of parameters that need
to change and in which direction. They are extrapolated from the computational results of
the particular case study in this paper, and are intended as helpful guidance for policy mak-
ers and strategic planners to consider and act upon before a disaster occurs. However, the
quantitative and optimising characteristics of the model also make it very useful for tactical
operations, enabling field planners to flexibly redirect transportation and material resources
on a daily basis, if reliable and timely data is available. If some of the data is of doubtful
quality or needs to be very roughly estimated, then the models’ outputs will need to be
treated accordingly. Nevertheless, such outputs will help to compare the consequences of
different courses of action.

[email protected]
Bibliography 255

Table 11.1 Recommendations for effective humanitarian relief operations

Scenario Recommendation
Increase weighting on transportation cost component in objective function.
Low Budget. Minimise vehicle slack.
Reduce the number of helicopters used.
Stagger arrival of supplies from external suppliers.
Limited Storage.
Send items to warehouses by truck to make use of storage in transit.
Introduce a priority status.
Recipients
Introduce a minimum delivery policy.
neglected.
Consider placing emergency contingency stocks close to these areas.
Small number of
helicopters and Pool resources to make maximum use of resource.
trucks available.
Need to reduce
the amount of Increase the number of helicopters available at the initial point of entry.
time taken to Increase the number of trucks available at heliports and warehouses.
minimise Increase the number of aeroplanes allowed to land at airports.
unsatisfied Consider the use of suppliers close to the disaster zone.
demand.
Limited Road Restrict the use of helicopters to those recipients that have no road access.
Access. Introduce airlifts as a new mode of transport to these recipients.

Acknowledgements

We wish to thank the referees for their careful reading of the article. Their constructive
comments, queries and suggestions enabled us to improve its quality, clarity and readability.

Bibliography

[1] Altay, N. and Green, W. G. (2006). OR/MS research in disaster operations management, Euro-
pean Journal of Operational Research 175, 1, pp. 475–493.
[2] Balcik, B., Beamon, B. M., Krejci, C. C., Muramatsu, K. M. and Ramirez, M. (2010). Coor-
dination in humanitarian relief chains: Practices, challenges and opportunities, International
Journal of Production Economics 126, 1, pp. 22–34, Improving Disaster Supply Chain Man-
agement - Key supply chain factors for humanitarian relief.
[3] Barbarosoğlu, G. and Arda, Y. (2004). A two-stage stochastic programming framework for
transportation planning in disaster response, Journal of the Operational Research Society 55,
pp. 45–53.
[4] Beamon, B. and Kotleba, S. (2006). Inventory modelling for complex emergencies in humani-
tarian relief operations, International Journal of Logistics: Research and Applications 9, 1, pp.
1–18.
[5] Beamon, B. M. (2004). Humanitarian relief chains: issues and challenges, in Proceedings of
the 34th International Conference on Computers and Industrial Engineering (San Francisco,

[email protected]
256 Decision Aid Models for Disaster Management and Emergencies

CA), pp. 77–82.

[6] Brown, G., Keegan, J., Vigus, B. and Wood, K. (2001). The Kellogg company optimizes pro-
duction, inventory, and distribution, Interfaces 31, 6, pp. 1–15.
[7] Camm, J. D., Chorman, T. E., Dill, F. A., Sweeney, D. J. and WWegbyn, G. (1997). Blending
OR/MS, judgment, and GIS: Restructuring P&G’s supply chain operations, Interfaces 27, 1,
pp. 128–142.
[8] Caunhyea, A. M., Niea, X. and Pokharel, S. (2012). Optimization models in emergency logis-
tics: A literature review, Socio-Economics Planning Sciences 46, 1, pp. 4–13.
[9] Clark, A. R. (2011). Humanitarian supply chain network models for disaster relief, LAND-
TRANSLOG II Workshop, Puerto Varas, Chile, 12-15 December 2011, http://www.cems.
uwe.ac.uk/~arclark, accessed 11 April 2012.
[10] Ergott, M. and Gandibleux, X. (2002). Multiobjective combinatorial optimization - theory,
methodology and applications, International Series in Operational Research and Management
Science 52, pp. 369–444.
[11] Erhgott, M. and Gandibleux, X. (2002). emphMultiple Criteria Optimization: State of the Art
Annotated Bibliographic Surveys, Vol. 52 (Kluwer).
[12] Fourer, R., Gay, D. and Kernighan, B. (2003). emphA Mathematical Programming Language,
2nd edn. (Thomson, Calif, USA).
[13] Fritz Institute (2005). Logistics and the effective delivery of humanitarian relief, www.
fritzinstitute.org, accessed 11 April 2012.
[14] Goodwin, P. and Wright, G. (2004). emphDecision Analysis for Management Judgement, 3rd
edn. (Wiley, Chichester).
[15] Graves, S. C. and Willems, S. P. (2000). Optimizing strategic safety stock placement in supply
chains, Manufacturing & Service Operations Management 2, 1, pp. 68–83.
[16] Haghani, A. and Oh, S. (1996). Formulation and solution of a multi-commodity, multi-modal
network flow model for disaster relief operations, Transport Research - Part A 30, 3, pp. 231–
250.
[17] Hwang, H. (1999). A food distribution model for famine relief, Computers and Industrial En-
gineering 37, pp. 335–338.
[18] Ilog (2004). emphCPLEX 9.1 User’s Manual, ILOG S.A, www.cplex.com.
[19] IRFC (1994a). International federation of the red cross and red crescent, code of conduct, www.
ifrc.org, accessed 11 April 2012.
[20] IRFC (1994b). International federation of the red cross and red crescent, code of conduct:
Annex 1, www.ifrc.org, accessed 11 April 2012.
[21] IRFC (2006). International federation of the red cross and red crescent, disasters, www.ifrc.
org, accessed 11 April 2012.
[22] Long, D. (2003). emphInternational Logistics: Global Supply Chain Management (Kluwer
Academic Publishers, Massachusetts).
[23] Long, D. and Wood, D. (1995). The logistics of famine relief, Journal of Business Logistics 16,
1, pp. 213–229.
[24] M T Melo, S. N. and da Gama, F. S. (2005). Dynamic multi-commodity capacitated facility
location: a mathematical modelling framework for strategic supply chain planning, Computers
and Operations Research 33, pp. 181–208.
[25] Martin, R. (2005). The balancing act: Speed, agility versus cost in effective disruption manage-
ment, http://www.itconversations.com, podcast accessed 11 April 2012.
[26] McGuire, G. (2000). Supply chain management in the context of international humanitarian
assistance in complex emergencies – part 1, Supply Chain Practice 2, 4, pp. 30–43.
[27] McGuire, G. (2001). Supply chain management in the context of international humanitarian
assistance in complex emergencies – part 2, Supply Chain Practice 3, 1, pp. 4–18.

[email protected]
Bibliography 257

[28] Oloruntoba, R. and Gray, R. (2005). Humanitarian aid: an agile supply chain? Supply Chain
Management 11, 2, pp. 115–120.
[29] Ortuño, M., Tirado, G. and Vitoriano, B. (2011). A lexicographical goal programming based
decision support system for logistics of humanitarian aid, TOP: An Official Journal of the
Spanish Society of Statistics and Operations Research 19, 2, pp. 464–479.
[30] Ozdamar, L. (2011). Planning helicopter logistics in disaster relief, OR Spektrum 33, 3, p. 655–
672.
[31] Ozdamar, L., Ekinci, E. and Kucukyazici, B. (2004). Emergency logistics planning in natural
disasters, Annals of Operations Research 129, pp. 217–245.
[32] PAHO (2000). Pan-american health organisation. manual logistical management of humanitar-
ian supply, www.disaster-info.net, podcast accessed 11 April 2012.
[33] Poojari, C. A., Lucas, C. and Mitra, G. (2008). Robust solutions and risk measures for a supply
chain planning problem under uncertainty, Journal of the Operational Research Society 59, 1,
pp. 2–12.
[34] Rottkemper, B., Fischer, K., Blecken, A. and Danne, C. (2011). Inventory relocation for over-
lapping disaster settings in humanitarian operations, OR Spectrum 33, pp. 721–749.
[35] Sphere Project (2004). Humanitarian charter and minimum standards in disaster response, www.
sphereproject.org, accessed 11 April 2012.
[36] Thomas, A. (2003). Humanitarian logistics: Enabling disaster response
www.fritzinstitute.org, accessed 11 April 2012.
[37] Thomas, A. (2005). Matching recognition with responsibility, www.fritzinstitute.org,
accessed 11 April 2012.
[38] Tomasini, R. and Van Wassenhove, L. N. (2004). A framework to unravel, prioritize and co-
ordinate vulnerability and complexity factors affecting a humanitarian operation, knowledge.
insead.edu, accessed 11 April 2012.
[39] Tsui, E. (2002). Initial response to complex emergencies and natural disasters, in K. M. Cahill
(ed.), Emergency Relief Operations, chap. 2 (Fordham University Press, New York), pp. 32–54.
[40] UNISDR (2005a). United Nations international strategy for disaster reduction, Emergency pre-
paredness for effective response: Strengthening institutional capacities, www.unisdr.org, ac-
cessed 11 April 2012.
[41] UNISDR (2005b). United Nations international strategy for disaster reduction, Proceedings of
the conference building the resilience of nations and communities to disasters, www.unisdr.
org, accessed 11 April 2012.
[42] Van Wassenhove, L. N. (2006). Humanitarian aid logistics: supply chain management in high
gear journal of the operational research society, Journal of the Operational Research Society
57, 5, pp. 475–489.
[43] Van Wassenhove, L. N. and Pedraza Martinez, A. J. (2010). Using or to adapt supply chain
management best practices to humanitarian logistics, International Transactions in Operational
Research.
[44] Vitoriano, B., Ortuño, M., Tirado, G. and Montero, J. (2011). A multi-criteria optimization
model for humanitarian aid distribution, Journal of Global Optimization 51, pp. 189–208.
[45] Vitoriano, B., Ortuño, T. and Tirado, G. (2009). HADS, a goal programming-based humanitar-
ian aid distribution system, Journal of MultiCriteria Decision Analysis 16, 2009, pp. 55–64.
[46] Woods, D. F., Barone, A., Murphy, P. and Wardlow, P. (1995). International Logistics (Chap-
man & Hall, London).
[47] Yi, W. and Ozdamar, L. (2007). A dynamic logistic coordination model for evacuation and
support in disaster response activities, European Journal of Operational Research 179, 3, pp.
1177–1193.

[email protected]
Chapter 12

A Travel Behaviour Study through Learning and

Clustering of Fuzzy Cognitive Maps

Maikel León 1,2,∗ , Lusine Mkrtchyan 2,3 , Benoît Depaire 2 , Da Ruan 2,3 †, Koen Vanhoof 2
1 Center of Studies on Informatics, Central University of Las Villas, Santa Clara, Cuba
2 Transportation Research Institute, Hasselt University, Diepenbeek, Belgium
3 Belgian Nuclear Research Centre (SCK • CEN), Mol, Belgium
∗ [email protected]

Transport management and behaviour modelling appears in modern societies because of

the importance for all social and economic processes. Using in this field advanced com-
puter techniques like the Artificial Intelligence ones is really relevant from the scientific,
economic and social point of view. In this paper we deal with Fuzzy Cognitive Maps as an
approach in representing the behaviour and operation of such complex systems. We also
show how travellers base their decisions on knowledge of different transport mode proper-
ties at different levels of abstraction depending on their perception of available information.
These levels correspond to the abstraction hierarchy including different scenarios of travel-
ing, different set of benefits while choosing a specific travel mode, and different situations
and attributes related with those benefits. We use learning and clustering of Fuzzy Cogni-
tive Maps to describe travellers’ behaviour and change trends in different abstraction levels.

12.1 Introduction

Travel behaviour studies are very important for decreasing travel-related energy con-
sumption, depressing high weight on urban infrastructure, and predicting individuals’ be-
haviours, etc. Many attempts have been made to influence individuals’ unsustainable travel
behaviour towards more sustainable forms, improving the transportation use by reducing
travel demand or distributing it in time and space. However, these behaviour studies can
be effectively and efficiently implemented if they are developed founded on a profound
† Prof. Da Ruan deceased on 31 of July 2011 while he was actively involved in this study.

B. Vitoriano et al. (eds.), Decision Aid Models for Disaster Management 259
and Emergencies, Atlantis Computational Intelligence Systems 7,
DOI: 10.2991/978-94-91216-74-9_12,  Atlantis Press 2013

[email protected]
260 Decision Aid Models for Disaster Management and Emergencies

understanding of the basic causes of travel, such as people’s reasons and inclinations, and
comprehensive information of individuals’ behaviours [1].

12.1.1 Problem presentation

In the process of transportation planning, travel demand forecast is one of the most
important analysis instruments to evaluate various policy measures to influence travel sup-
ply and demand. In past decades, increasing environmental awareness and the generally
accepted policy paradigm of sustainable development made transportation policy measures
to shift from facilitation to reduction and control. Some objectives of those measures are:
to alter travel behaviour without necessarily embarking on large-scale infrastructure ex-
pansion projects and to encourage better use of available transport resources avoiding the
negative consequences of continued unrestrained growth in private mobility.
Individuals’ activity travel selections can be considered as real decision problems, pro-
ducing the generation of a mental representation or Cognitive Map (CM) of the decision
situation and alternative action courses in the expert’s mind. This CM concept is often
referred to, in theoretical frameworks of travel demand models, especially related to the
representation of spatial dimensions, but many features can be taken into account (see Fig-
ure 12.1).

Fig. 12.1 Abstraction levels of mind related to Travel Behaviour.

However, actual model applications are scarce, mainly due to problems in modelling
and measuring the construct. To this end, the development of the mental map concept can
benefit from the knowledge provided by individual tracking technologies. Many studies
are focusing on that direction, to improve developed models and to produce a better quality

[email protected]
A Travel Behaviour Study through Learning and Clustering of Fuzzy Cognitive Maps 261

of systems. At an individual level it is important to realize that the relationship between

travel decisions and the spatial characteristics of the environment is established through the
individuals’ perception and cognition of environment [2].
Records regarding individuals’ decision making processes can be used as input to gen-
erate mental models. Such models treat each individual as an agent with mental qualities,
such as viewpoints, objectives, predilections, inclinations, etc. For modelling, several Arti-
ficial Intelligence techniques can be used, in this case Fuzzy Cognitive Maps (FCMs) will
be applied; through them we will try to simulate individuals’ decision making processes.
Consequently, FCMs can be used not only to understand people’s travel behaviours, but
also to pretend the changes in their actions due to some factors in their decision atmo-
sphere. This technique is very well known by its self-explicability.
From a computational point of view, FCMs are a mixture of several aspects from Fuzzy
Logic and Neural Networks; combining the heuristic and common sense rules of Fuzzy
Logic with the learning heuristics of the Neural Networks. They were introduced by B.
Kosko, who enhanced CMs with fuzzy reasoning, that had been previously used in the
field of socio-economic and political sciences to analyse problems such as social decision-
making problems, etc. The use of FCMs for many applications in different scientific fields
had been proposed.

12.1.2 Contribution in Disaster Management studies

Disaster event monitoring as one of the steps in risk and crisis management is a very
complex system with uncertain input parameters. Fuzzified inputs, the fuzzy rule base,
which is constructed using objective and subjective definitional, causal, statistical, and
heuristic knowledge, is able to present the problem in a user-friendly form. The com-
plexity of the system can be managed by hierarchically-structured reasoning models, and
if necessary, gained risk factor structure. Crisis or disaster event monitoring provides basic
information for many decisions in today’s social life.
The disaster recovery strategies of countries, the financial investments plans of in-
vestors, or the level of the tourism activities all depend on disaster different groups or
crisis factors [3]. Transport problems constitute medullar in disaster management, espe-
cially in humanitarian logistic in which most of the problems involved could be addressed
as transportation or travel problems. Disaster behaviour or travel behaviour after a disas-
ter is an interesting topic less considered and that certainly contributes with society and
environment.

[email protected]
262 Decision Aid Models for Disaster Management and Emergencies

On the other hand, cognitive mapping is an extraordinarily revealing window on a per-

son’s ability to cope with these problems. It also probes the subject’s expertise and potential
as a front-line emergency manager. As an exercise, it instructs trainee emergency managers
and students of crisis management in the process of assessing the absolutely fundamental
spatial dimension of disasters. An overall reasoning in this study presents that an extraordi-
nary variety of spatial perceptions comes to the fore when many emergency managers are
asked to assess the same situation.
This has implications for management strategies, as these will be influenced strongly
by distance, space and place, by the time taken to assemble and deploy resources, and by
the relative nearness or remoteness of critical sites, all of which are geographical prob-
lems. The results also have implications for emergency management training, in particular,
that spatial cognition is a vital skill that should be developed along with others such as
communication and decision-making.

12.1.3 Case study

In this work, FCMs constitute a good alternative to study individuals during their de-
cision making process. A decision maker activates a temporary mental representation in
his working memory based on his previous experiences or existing knowledge. There-
fore, constructing a mental representation requires a decision maker to recall, reorder and
summarize relevant information in his long-term memory. It may involve translating and
representing this information into other forms, such as a scheme or diagram, supporting
coherent reasoning in a connected structure.
When only taking into account real situations and the benefits people want to gain in
certain decision, analysis will be only done in a non-depth mind level, results constitute
useful, but more exploitation could certainly be done (see Figure 12.2). It is known that
people are able to have deeper levels of reasoning, correlating their actions and goals, and
to explore this as real as possible will guide us to a better representation of decision making
process, specifically in Travel Behaviour research.
A study related to Travel Behaviour has been made in the city of Hasselt, capital of the
Flemish province of Limburg, Belgium. The city has around 73 000 habitants, with a traffic
junction of important circulation arteries from all directions. Hasselt made public transport
by bus zero-fare from 1 July 1997 and bus use was said to be as much as "13 times higher"
by 2010, being the first city in the world that had entirely zero-fare bus services on the
whole of its territory.

[email protected]
A Travel Behaviour Study through Learning and Clustering of Fuzzy Cognitive Maps 263

Fig. 12.2 Human mind level abstraction.

The aim of this study is three fold: in one hand we used learning of FCMs as a predic-
tive tool to explore the possible preferences of users giving certain circumstances, on the
other hand, we used clustering algorithms as a descriptive tool to analyse different groups
of users and understand the main features of choosing a transport mode. The results of
learning will serve as a guide for transportation policy decision makers for future plans
if, for example, one of the circumstances is changing, what are the most expected actions
to be taken. Consequently, the results of FCMs clustering will help decision makers to
program activities considering the specific needs of different groups of users. In addition,
we explore how travellers’ preferences are changing through different levels of information
processing abstraction.

12.2 Modelling with Fuzzy Cognitive Maps

Firstly, in this section we introduce the main concepts and definitions of CMs/FCMs
that we need as a theoretical base for the next chapters, then we detail the current ap-
proaches of using FCMs/CMs in transportation related topics. We summarize the appli-

[email protected]
264 Decision Aid Models for Disaster Management and Emergencies

cation domains that use cognitive mapping in decision making process. Next, we briefly
explain the main advantages of using CMs/FCMs as a modelling tool and we compare it
with some other wide used techniques.

12.2.1 Knowledge Representation with Fuzzy Cognitive Maps

Cognitive Maps were first introduced by Axelrod [4] who focused on the policy do-
main. Since then many researchers have used CMs applying them in various fields where
the problems are ill structured or not well-defined. The application areas of CMs or FCMs
are very broad and diverse, and, in general, they can be used in different decision making
problems such as failure system modelling, network security, electrical circuits, environ-
mental management, information management, data mining, web mining, risk analysis and
management, social sciences, health care and medical decisions, human reliability anal-
ysis, industry applications, etc. The reader can refer to [5] for the short description and
references of mentioned fields in which FCMs have been applied.
Cognitive Maps have two types of elements: concepts and causal beliefs. The former
are variables while the latter are relationships between variables. Causal relationships can
be either positive or negative, as specified by a ‘+’, respectively a ‘−’, sign on the arrow
connecting two variables. The variable that causes a change is called cause variable and the
one that undergoes the effect of the change is called an effect variable. If the relationship is
positive, an increase or decrease of the cause variable causes the effect variable to change
in the same direction (e.g., an increase in the cause variable causes increase of the effect
variable). In the case of a negative relationship, the change of the effect variable is in
the opposite direction (e.g., an increase in the cause variable causes decrease of the effect
variable).
Figure 12.3(a) shows a simple CM for the travel behaviour problem described previ-
ously. This example shows some of the concepts to decide which transportation mode is
more comfortable: taking a bike or a car. Here only the change directions are shown;
for example the increase/decrease of number of bags causes the decrease/increase of bike
comfort: therefore the sign is negative. On the other hand, if we consider the causal rela-
tionship between bike infrastructure and bike comfort, the link is positive as the change in
cause node changes the effect node in the same direction.
In his work Axelrod introduces also Weighted CMs and Functional CMs. In weighted
CMs the sign in the map is replaced with a positive or a negative number, which shows the
direction of the effect as well as its magnitude. In functional CMs a function is associated

[email protected]
A Travel Behaviour Study through Learning and Clustering of Fuzzy Cognitive Maps 265

with each causal relationship showing more precisely the direction and the magnitude of
the effect. These two types of CMs give more flexibility as they can handle and provide
more detailed information.
However, CMs, whatever their type, are not easy to define and the magnitude of the
effect is difficult to express in numbers. Usually CMs are constructed by gathering in-
formation from experts/users and generally, they are more likely to express themselves in
qualitative rather than quantitative terms.
To this end, it may be more appropriate to use FCMs, suggested by Kosko [6]. Actually,
FCMs are weighted CMs with fuzzy weights. The degree of relationship between concepts
in an FCM is either a number in [−1; 1], or a linguistic term, such as often, extremely, some,
etc. Figure 12.3(b) shows a simple example of FCM for the same problem where the causal
relationships are expressed by using fuzzy linguistic terms.

Fig. 12.3 (a) An example of a CM. Fig. 12.3 (b) An example of an FCM.

For example, if we consider again the relationship between bike infrastructure and bike
comfort, the increase/decrease of cause variable will cause high increase/decrease in effect
variable. It is useful to mention that the nature of our data allowed us to model maps so
that all links have only positive sign.
Note that each concept represents a characteristic of the system; in general it stands for
events, actions, goals, values, trends of the modelled system, etc. In addition, each concept
is characterized by a number that represents its value and it results from the renovation of
the real value of the system variable [7].
Beyond the graphical representation of the FCM there is its mathematical model. It
consists of a 1 × n state vector A which includes the values of the n concepts and a n × n
weight matrix (adjacency matrix) W which gathers the weights Wi j of the interconnections

[email protected]
266 Decision Aid Models for Disaster Management and Emergencies

between the n concepts. The value of each concept is influenced by the values of the
connected concepts with the appropriate weights and by its previous value. So the value
Ai for each concept Ci can be calculated, among other possibilities, by the following rule
(12.1):

Ai = f A j × W ji (12.1)
j=1
j�=i

where Ai is the activation level of concept Ci , A j is the activation level of concept C j and W ji
is the weight of the interconnection between C j and Ci , it is to say, the value of Ai depends
of the weighted sum of its input concepts, and f is a threshold or normalization function.
The nonlinear function, f , can be a simple thresholding operation with a threshold value T
resulting for example binary concept values.
To produce continuous concept values, a continuous-output transformation function
may be used. The most widely used function is the sigmoid function. As for our application
the links of FCMs have only positive signs, we choose the normalization function given
below that fits our task the best (see (12.2) and Figure 12.4).
1
f (x) = (12.2)
1 + e−9(x−0.5)
Binary FCMs are suitable for highly qualitative problems where only representation of
increase or stability of a concept is required, Trivalent FCMs are suitable for qualitative
problems where representation of increase, decrease or stability of a concept is required,
but Sigmoid FCMs are suitable for qualitative and quantitative problems where represen-
tation of a degree of increase, a degree of decrease or stability of a concept is required and
strategic planning scenarios are going to be introduced.

Fig. 12.4 The normalization function and its graphical representation.

[email protected]
A Travel Behaviour Study through Learning and Clustering of Fuzzy Cognitive Maps 267

So, the new state vector Anew is computed by multiplying the previous state vector Aold
by the weight matrix W . The new vector shows (12.3) the effect of the change in the value
of one concept in the whole FCM [8].

Anew = f (Aold × W ) (12.3)

Some other notions that we use in the next chapters are introduced below.
The conceptual centrality of a node Ci is denoted by CEN(Ci) and is

CEN(Ci ) = IN(Ci ) + OUT(Ci ),

n n
(12.4)
where IN(Ci ) = wki and OUT(Ci ) = wik
k=1 k=1

The conceptual centrality represents the importance of the node for the causal flow on
the CM. A node Ci is called transmitter if IN(Ci ) = 0 and OUT(Ci ) > 0, and is called
receiver if IN(Ci ) > 0 and OUT(Ci ) = 0. The total number of receiver nodes in a map is
considered a complexity index. A large number of receiver variables indicate a complex
map while a large number of transmitters show a formal hierarchical system where causal
nodes do not collaborate with each other [9].
Among several ways of developing CMs and FCMs, the most common methods are:

• extracting knowledge from questionnaires,

• extracting knowledge from written texts,
• conducting interviews, and
• drawing maps from data.

To obtain a CM from questionnaires requires first to identify the most important vari-
ables for the given problem, and then to give experts ordered pairs of variables in a ques-
tionnaire format. Afterward, experts decide the strength of causal links relying on their
knowledge and experience. The second method is a type of content analysis in which the
causal relationships are identified by analysing texts. The main problem related to this
method is that usually the relationships are not explicitly stated, and the language structure
in which texts are written can vary from one language to another.
The detailed description of how to construct a map through interviews is provided in
[9]. The methodology is composed by the following steps: to decide the most important
variables, to provide experts with a sample map, unrelated with the problem at hand, and
finally, to ask them to draw their own maps of the issue under investigation. The automatic
construction of CMs (particularly FCMs) based on user provided data is discussed in [10].

[email protected]
268 Decision Aid Models for Disaster Management and Emergencies

This method first finds the degree of similarity between any of the two variables, then
decides whether the relation between the two variables is direct or inverse, and with the
use of a fuzzy expert system tool determines the causality among variables. Note that these
methods can be used also in combinations, such as questionnaires with interviews. In the
next section we discuss the data gathering approach that we followed in this study.
To have more realistic and reliable results, more than one expert participates in FCMs
drawing or in knowledge acquiring process to construct FCMs. The group map can be
generated based on individual maps.
This is one advantage of FCMs over other approaches like Petri Nets (PN) or Bayesian
Networks (BN). PN are another graphical and mathematical modelling tool consisting of
places, transitions, and arcs that connect them, that can be used as a visual-communication
aid similar to flow charts, block diagrams, and networks. As a mathematical instrument,
it is possible to set up state equations, algebraic equations, and other mathematical models
governing the performance of systems.
It is well known that the use of PN has as a disadvantage the drawing process by a non-
expert in this technique, that is the reason of having only a limited numbers of tools usable
for this purpose, and it is not well established how to combine different PN that describe
the same system.
BN is a powerful tool for graphically representing the relationships among a set of vari-
ables and for dealing with uncertainties in expert systems, but demanding effort caused by
specification of the net (structure and parameters) and an expensive algorithm of propa-
gation of probabilities. Also is not evident for a non-expert in this field how to construct
a BN, and even more difficult how to compare or combine different BN that describe the
same system.
Another observation that we consider important to mention is the fact that there are only
few software tools developed with the intention of drawing FCMs by non-expert users with
different backgrounds and technical knowledge, as FCM Modeler [11] and FCM Designer
[12]. The first one is a simple incursion, while the second one is a better implementation,
but still difficult to interact with and it does not have experimental facilities. In [13] a
software tool is described with more functionalities; although it was conceived for general
purposes, we have included special options and developed specific method to deal with data
requirement used in the travel behaviour study.

[email protected]
A Travel Behaviour Study through Learning and Clustering of Fuzzy Cognitive Maps 269

12.2.2 Cognitive Maps and Transportation

In the previous sections we briefly mentioned the application fields of FCMs. Obvi-
ously, the application areas are diverse covering very broad range of domains. However,
as far as we found, the links between CMs and travel behaviour is less well-developed.
Specifically, research on CMs and travel focuses primarily, in fact almost exclusively on
route choice. In contrast, the other steps such as trip generation (how many trips?), trip
distribution (where to go?) have been given far less attention by cognitive mapping re-
searchers [14].
So it is to the potential role of CMs in trip generation, trip distribution, and mode split
we now turn. In [15] it is explained CM modelling to extract the mental representation of
individuals in the decision making and planning of trips, related to daily travels.
Current opinion appears to indicate that, because factors such as CM ability, knowledge
of feasible alternatives, navigation and way finding strategies, and preferences for path
selection criteria all are presumed to have a substantial impact on travel choices, there is a
growing need to include spatial cognition explicitly in models.
Cognitive mapping and travel behaviour research have centred on what is known about
the location, possible destinations, and feasible alternatives for any choice affecting the
network over which the traveling must take place [16]. The literature on household ac-
tivity modelling, as a conceptually sound and robust way to predict travel behaviour than
traditional travel demand modelling is large and increasing. Activity modelling could be
enhanced significantly with better information on how modal experience shapes individu-
als’ CMs.
In other words, the CMs of people who mostly walk and use public transit may vary
systematically from those who are mostly chauffeured in private vehicles, and from those
who usually drive [17]. This line of reasoning is consistent with research on job search
behaviour among low wage workers. Those with regular access to private vehicles tend not
only to search larger geographic areas work for work, but tend to perceive job opportunities
in less spatially constrained ways.
In order to get rid of such cognitive barriers to job opportunities experienced by those
without regular access to cars, compensatory solutions such as trip planning services, guar-
anteed ride home services, and overall improvements to transit service could be imple-
mented.
Another means to compensate for limitations in individuals’ CMs could be the dis-
semination of Intelligent Transportation Systems (ITS). Such systems reduce individuals’

[email protected]
270 Decision Aid Models for Disaster Management and Emergencies

overall reliance on their own CMs potentially increasing access to known destinations. ITS
would not influence necessarily how prior spatial knowledge informs the initial portions of
travel behaviour sequence, trip generation and trip distribution [14].
Individuals would still rely on their CMs when choosing to make a trip and selecting
a particular destination for that trip. Public transit planning could potentially benefit from
CMs researches [2].
While researchers have recognized the connection between travel and spatial learning,
little is known yet about how the existing transportation infrastructure itself shapes CMs
and, in turn, affects route selection as well as other aspects of travel including trip fre-
quency, trip purpose and destinations, and mode choice. However, the limited available
research suggests that transportation infrastructure and, in particular, way finding on over-
lapping, yet distinct, modal networks, sidewalks, bike lanes, transit routes, local streets and
roads, and freeway networks, affects the development of CMs and, in turn, travel behaviour.

12.3 Capturing individuals’ mental representation: Automatic knowledge

engineering and FCM modelling

Formalizing knowledge for Artificial Intelligence processing is of vital importance.

In the knowledge engineering process after the variables selection that better describe the
problem, and the interaction among them, it is necessary to construct abstract structures that
facilitates the organization of the stored information, such a way that they can be interpreted
by a computational technique [18], allowing the extraction of useful information, and even
more: knowledge.
A cognitive model is an abstract structure containing all information related with visu-
alization ideas that takes place in the decision making process [19], the following design is
used in our approach:

• Personal information about individual: Useful for demographic processing and analy-
ses, etc.
• Cognitive subsets: The variable interaction that takes place in the decision making
process conduces the construction of related variables, such as “situation-benefit-
attribute”.
• Expert criteria: After defining all cognitive subsets the criteria of the individual is cap-
tured using artificial scenario. Situational variables are assigned with random states,

[email protected]
A Travel Behaviour Study through Learning and Clustering of Fuzzy Cognitive Maps 271

and the respondent specifies how much utility reports in that conditions, the use of bus,
car or bike.
• Causal influences among variables: It is considered a causal relation among the vari-
ables that participate in the modelling of a cognitive subset, and the experts must ex-
press their opinions.
• Benefits importance: To all the benefit variables participating in the modelling it is
assigned an importance level, experts give their criteria.

Definitely the informatics software products success depends on its architecture de-
sign and user interface flexibility, especially when the software product is an Automatic
Knowledge Engineer [19], in this case generating individual mental representation for after
forward processing. In order to achieve effective knowledge acquisition, a dynamically ap-
pearance according to the behaviour of the experts is required. This enables, among other
things, to capture and formalize in the knowledge bases the relationships among the vari-
ables involved in their decisions, and the degree of causal influence. Thus, there is a miti-
gation of “bottlenecks” problem, typically in the knowledge based systems creation [20].
We have taken into consideration some features that are practically indispensable in the
successful performance of the proposed application:

• wizard with non-fixed pages number generated taking into account previous expert
elections (ensuring the dynamism of the application),
• randomly ordered variables lists (in case the users use to select the same variables
according to their position in the lists),
• explanatory pages with detailed instructions that will guide the users with some flexi-
bility (these messages and instructions are considered “sophisticated” as it is generated
according to the individual respondent behaviour, using the natural language obvious
advantages),
• detailed description of each variable when receiving the mouse focus.

In an imaginary case where an expert considers modelling his considerations, accord-

ing to his own experience, the situational variable Precipitation will be the cornerstone of
his decisions, the benefit variable Physical comfort describes the benefit that reports to the
expert the selection of Precipitation variable and Stay dry characterizes the relationship be-
tween the two variables. Later, a list of imaginary scenarios (involving the chosen variables
by an expert) is constructed to know the expert criteria under specific circumstances.

[email protected]
272 Decision Aid Models for Disaster Management and Emergencies

For a No rain state of the Precipitation variable, the respondent preferred bike for trans-
portation, with less preference considers taking bus and finally private transport (car), be-
cause the benefit that brings his selection is the physical comfort and aims to stay dry.
Generally, a cognitive model contains instructions that define individual’s mental repre-
sentations in situations relating to travel behaviour, which the software generates to knowl-
edge bases and can be summarized in four aspects:

• construction of the cognitive subsets,

• expert assessment by building scenarios,
• causal influence among the various categories of variables, and
• importance of the benefits involved in the model.

In the referred knowledge acquiring process, the developed Automated Knowledge En-
gineer codes in knowledge bases all information concerning the mental maps that take
place in the experts who are in situations relevant to travel behaviour. However, the stored
knowledge is worthless if it cannot be represented in terms of computational structures
that enable further analysis, automated or not, but basically for simulations or predictions
purposes that are useful to science and application area. The construction of FCMs from
the knowledge bases generated by the Automated Knowledge Engineer helps to solve this
problem.
To express the stored knowledge in terms of components that allow automatically build-
ing FCMs we identify three primary structures as part of cognitive models contained in the
knowledge bases related to FCMs:

• Cognitive subsets: represent triplet concepts namely: situational-benefit-attribute,

defining the structure of the map. The concepts for the decisions (Car, Bus, Bike)
and for the final utility (Utility) are implicit concepts and mandatory for every cog-
nitive subset. For example, for the cognitive subset Precipitation-Stay dry-Physical
comfort, the topology of the resulting map is shown in Figure 12.5.
• Concept causal influences: values in the interval [0, 1] representing the causal influ-
ences among the variables of the cognitive subsets.
• Weights of the benefit variables involved in the problem (Ratings): represent a list of
all weighted benefits with a value in the interval [0, 1]. This value represents the impact
of a benefit concept in the map.

One of the most important features of this topology is the fact that the concepts C6
and C7 are considered vectors variables and not linear ones, thus, allowing when running

[email protected]
A Travel Behaviour Study through Learning and Clustering of Fuzzy Cognitive Maps 273

Fig. 12.5 General topology of a minimum FCM.

the inference process that only one of three decisions was active (property exclusivity in
the decision node), so the final result of these concepts will be a three-dimensional vector
(making an inference for each decision). This philosophy allows the utility calculation of
each decision under specific circumstances.

12.4 A learning method for FCMs inspired on Particle Swarm Optimization

Problems associated with development of FCMs encourage researchers to work on au-

tomated or semi-automated computational methods for learning of FCM structures using
mainly historical data. Semi-automated methods still require a relatively limited human
intervention, whereas fully automated approaches are able to compute an FCMs model
solely based on historical data. Researches on learning of FCMs models from data have
resulted in a number of alternative approaches. One group of methods is aimed at provid-
ing a supplement tool that would help experts to develop an accurate model based on their
knowledge about a modelled system. Algorithms from the other group are oriented toward
eliminating human factor from the entire development process, only historical data are nec-
essary to establish FCMs models. A number of algorithms for learning of FCM structure
from data have been proposed. They can be categorized into two groups based on the

[email protected]
274 Decision Aid Models for Disaster Management and Emergencies

learning paradigm used, i.e., Hebbian-based learners and methods based on evolutionary
algorithms [21].

12.4.1 Conception of the learning method using Particle Swarm Optimization

Particle Swarm Optimization (PSO) method, which belongs to the class of Swarm Intel-
ligence algorithms, can be used to learn FCMs structure based on historical data, consisting
of a sequence of state vectors that leads to a desired fixed-point attractor state. PSO is a
population based algorithm, which performs a search by maintaining and transforming a
population of individuals [22]. This method improves the quality of resulting FCMs mod-
els by minimizing an objective or heuristic function. The function incorporates human
knowledge by adequate constraints, which guarantee that relationships within the model
will retain the physical meaning defined by experts. Figure 12.6 shows the main idea of the
proposal.

Fig. 12.6 Using PSO for readjusting an FCM.

The flow chart illustrates the application of PSO in the readjusting of the weight matrix
that characterized an FCM, trying to find a better configuration that guaranty a convergence
or expected results. PSO is applied straight forwardly using an objective function defined
by the user. Each particle of the swarm is a weight matrix, encoded as a vector.
First the concepts and relation are defined, and the construction of an FCM is made,
and then is possible to make simulations and obtain outputs due to the inference process. If
the new values are not adequate, known by the execution of the heuristic function, then it

[email protected]
A Travel Behaviour Study through Learning and Clustering of Fuzzy Cognitive Maps 275

is necessary a learning process (in this case through the use of PSO metaheuristic) having
as results new values for the weight matrix.
The heuristic function h(x) will guide the search (12.5); the optimal case is when the
new found causal matrix offers the expected output:
n
h(x) = ri (x) (12.5)
i=1

where
A, ExpDecisioni(x) = InfDecisioni (x)
ri (x) =
c > 0, in other case
m
A= ExpDecisioni (x) − InfDecisioni (x)
k=1

where n is the number scenarios, m is the number of transport mode decisions, h(x) means
that the kth decision of ith scenario from map x according to the expert criteria (ExpDeci-
sion) is compared with the homolog kth decision that is inferred by the map x (InfDecision)
and its correspondent causal matrix w. We used c = 30 as this value was best fitting our
data.
The pseudocode in Figure 12.7 displays the general philosophy of the proposed method.
In this case genetic algorithm operators are used at initialization steps. Mixed approaches,
like using Genetic Algorithm and PSO have been performed so far. Some results in the
related literature are really promising and encouraging further researches and applications
in this area. Using this approach, new zones of the search space are explored in a particular
way, through the crossover of good initial particles and the mutation of some others, just to
mention two possible alternatives.
As it is known, the optimization capability of PSO algorithm is quite sensitive to pa-
rameters, through experimentation some of them were deducted to best perform the task,
but in general the tactic itself offered a good solution in finding a satisfactory solution, this
according to the very complex explored search space.

12.4.2 Implementing the learning method based on PSO for the FCMs
readjustment

A tool based on FCMs for the modelling of complex systems was developed, supporting
facilities for the creation, definition of parameters and options to make the inference process
more comprehensible, understanding and used for simulations experiments [23]. In the

[email protected]
276 Decision Aid Models for Disaster Management and Emergencies

Generate initial population using Wi j as initial approximation

Calculate initial evaluation
Cross over good particles
Mutation of selected particles
Initialize Xpbest with best solutions found by each particle
Initialize Xgbest as the best global found
Initialize Wmax = 1.4, Wmin = 0.4, c1 = 2.05, c2 = 2.05, k = 0.381966011
For t = 0 to Ngenerations
wk = (Wmax − Wmin ) ∗ ((Nc max − t)/(Nc max + Wmin ))
For each Xi
Calculate Vi (t + 1) and limit to [−Vmax, +Vmax ] using wk
Calculate Xi (t + 1) = Xi (t) + k ∗ Vi(t + 1) and normalize
Analyse the Swarm with Xi (t + 1) and Speed with Vi (t + 1)
Evaluate the particle Xi (t + 1)
Analyse the vector Xpbest with the best solutions
Update Xgbest with the best global particle
endFor
endFor

Fig. 12.7 Pseudocode of the proposed method.

learning package, the necessary initialization of parameters is done through the window
shown in Figure 12.8.
In simulation and experimenting in general, the visualization process is a fundamental
aspect (that is why a panel was conceived where the learning progression can be observed,
Figure 12.9 shows an example). It is possible to see how the FCM is updated with a new
weight matrix that better satisfies the expected results.

12.4.3 Validating learned structures against other approaches

In Table 12.1 we detail the data organization for the statistic experiment, through a
population comparison, to validate the performance of an FCM against other classical ap-
proaches such as Multilayer Perceptron (MLP), ID3 Decision Tree, or Naive Bayes (NB)
classifier. The same knowledge had been modelled with these techniques. The idea is

[email protected]
A Travel Behaviour Study through Learning and Clustering of Fuzzy Cognitive Maps 277

Fig. 12.8 Window for the PSO parameter specificatio12.

Fig. 12.9 Learning visualization panel.

to analyse the difference among their classification percentage (CP) obtained in a cross-
validation process with 10-folds.
After applying Kolmogorov-Smirnov test and having a non-normal distribution in our
data, we apply non parametric Friedman test, where a signification less than 0.05 suggests
to reject main hypothesis, therefor we can conclude that it exists a significant difference
among groups. Looking to the mean ranks, the best value is given to FCM, however, it is
not possible yet to affirm that our technique performs better than the others.

[email protected]
278 Decision Aid Models for Disaster Management and Emergencies

Table 12.1 Data organization for processing.

FCM MLP ID3 NB
Expert 1 CPFCM1 CPMLP 1 CPID3 1 CPNB 1
Expert 2 CPFCM 2 CPMLP 2 CPID3 2 CPNB 2
... ... ... ... ...
Expert 221 CPFCM n CPMLP n CPID3 n CPNB n

It is important to mention that as the data reflects classification percentage, a higher

value in the mean rank corresponds to a better performance. Using a Wilcoxon test for
related samples it is possible to analyse per pairs, and in all cases the main hypothesis of
the test is rejected and it is confirmed that there exists a significant difference between pairs.
Summing up, FCMs offer better results than the other approaches.
Table 12.2 contains the average percentages. First the learning scenarios serve for train-
ing, then for calculating optimistic estimation (resubstitution technique, empirical error) of
the convergence. The resubstitution test is absolutely necessary because it reflects the self-
consistency of the method, a prediction algorithm certainly cannot be deemed as a good
one if its self-consistency is poor.
Later, the testing scenarios were used to obtain a pessimist estimation (cross-validation,
real error) of the convergence through a cross-validation process with 10-folds. A cross-
validation test for an independent testing data set is needed because it can reflect the effec-
tiveness of the method in future practical application. The prediction capability had been
measured in the forecast of the first possible decision and in three decisions given by the
experts.
A recent study in [24] reports that a pressing issue is that currently the most accurate
automated methods for learning cannot scale problems exceeding “several dozens of con-
cepts”. In our case we have FCMs composed by more than 120 concepts; forthcoming a
study about temporal and spatial complexity in our proposal will be issued.
An analysis about the advantages that our method has over the others is productive,
thus we clarify that FCMs had a higher accuracy than other classical well-known tech-
niques when predicting the best decision, in this case the transport mode selection or the
whole preferences of a transport mode order. In one hand, the existing knowledge in the
elaborated topology definitely provides benefit to the classification structure, together with
the effectiveness in the learning method based on PSO, which also contributes in this task,
consequently the learned FCMs gained in the performed comparison.

[email protected]
A Travel Behaviour Study through Learning and Clustering of Fuzzy Cognitive Maps 279

Table 12.2 Classification percentage per technique, experiment and model.

FCM MLP ID3 NB
FIRST DECISION
Optimistic Model
99.47 97.38 94.26 95.63
Pessimistic Model
93.74 92.06 89.39 91.37
THREE DECISIONS
Optimistic Model
96.27 94.38 87.29 93.12
Pessimistic Model
88.72 82.40 77.59 80.25

On the other hand, FCMs not only performed better, but also the most important is its
capacity of presenting visual understanding information, combined with the classification
skills, makes them seems a good approach for these kinds of tasks.

12.5 Clustering of Fuzzy Cognitive Maps

The aim of this section is threefold: first we refer to the existing approaches of similarity
measurement algorithms of FCMs, second we detail the clustering method details, particu-
larly we explain our approach to find the optimum number of clusters by using FCMs, and
finally we analyse our results in the last subsection.

12.5.1 The distance matrix of FCMs

While there are many cases to use CMs or FCMs in different application fields, there
are only few attempts to develop algorithms to compare CMs. Similar studies of finding
the similarity between CMs are done in [25] and in [26]. First study is restricted only on
similarity algorithm without further use of it for cluster analysis. Second study extends the
similarity measurement of the first approach and, in addition, cluster analysis is provided.
However there are some drawbacks of this study.
For cluster analysis Ward linkage method is applied which commonly uses squared
Euclidean distance for similarity matrix calculation. In addition, the number of optimum
clusters is not discussed; neither any validation method is applied for cluster analysis. An
study of FCMs clustering is done in [27].

[email protected]
280 Decision Aid Models for Disaster Management and Emergencies

However, while the clustering based on map structure is comprehensive enough to be

used for different application domains, the clustering based on map content is rather case
specific and is restricted by principal component analysis.
Moreover, neither the weights of the links nor the links signs are not considered: thus
the functionalities of CMs are not discussed fully in cluster analysis. Another study of
clustering is done in [28]. This approach differs from previous ones as it clusters not FCMs
into different groups but the nodes of an FCM. Similar approach of clustering the nodes
into hierarchical structure is proposed in [29]. These last two studies are less relevant for
us because of two reasons: first we already have structured representation of the nodes in
our dataset, and, secondly, we are interested in FCMs clustering (not nodes clustering).
In general CMs/FCMs can be compared in two dimensions: comparing the content
and the structure of each map. The content difference is associated with the differences in
elements in both maps and the differences of the relationships between those elements. The
structural difference, on the other hand, is associated with the varying complexity degrees
of the maps structure. In our application we have hierarchical maps for all users therefore
we will focus only on the content difference analysis.
In [25] three types of differences between two individuals (maps) are proposed:

• Existence or non-existence of elements: thus one expert considers certain element as

important for the given domain; the other has the opposite opinion. In this case the
adjacency matrix for the CM of the first expert contains the element/elements while
the other matrix does not contain.
• Existence or non-existence of beliefs: thus one expert considers that there is a casual
relationship between two concepts, while the other has the opposite opinion. In this
case two experts should agree upon the fact that the nodes are important for the given
domain, but have opposite opinions towards the causal link.
• Different values for identical beliefs: thus two experts agree that there is a relationship
between two nodes, but one expert holds the belief more strongly than the other. In
adjacency matrices this difference is expressed by non-identical nonzero values for the
cell showing the causal link between two nodes.

In [26] the authors suggest an improvement of FCMs similarity measurement algorithm

described in [25] mentioning that the algorithm does not consider the missing values prop-
erly as well as it lacks of generalizability. We will take into account only the comment
about generalizability, and will adjust the algorithm to be applicable in our study.

[email protected]
A Travel Behaviour Study through Learning and Clustering of Fuzzy Cognitive Maps 281

Notice that by generalizability, we mean that in [25] the number of linguistic terms
is fixed and the comparison formula cannot handle different number of linguistic terms.
More specifically, the study fixed the number of linguistic terms to 7, assigning maximum
strength to 3, minimum strength to −3, and consequently the similarity measurement al-
gorithm cannot be applied for the cases with more/less linguistic terms. Therefore, for our
task, we use the following (12.6) distance ratio (DR):
p p
(a∗i j − b∗i j )
i=1 j=1
DR(uA , uB ) = (12.6)
2p2c + 2pc(puA + puB ) + p2uA + p2uB − (2pc + puA + puB )
where

1, if mi j �= 0, and i or j �∈ Pc
m∗i j = (12.7)
mi j , otherwise
In (12.6), ai j and bi j are the adjacency matrices of the first and second map respectively,
p is the total number of possible nodes, pc is the set of common nodes for both maps, pc
is the number of such nodes, is the number of nodes unique to user ua /ub respectively. In
(12.7) mi j is the value of the ith row and jth column in the zero augmented adjacency matrix
of any of two maps.

12.5.2 Hierarchical clustering of FCMs

The aim of this section is to use cluster analysis as a descriptive tool to group trav-
ellers and analyse each group finding similarity in people thinking in making a decision
about some transport mode. The understanding of travellers’ behaviour tendencies will
help policy makers in more realistic assessment when some concepts are changing over the
time.
As we have already mentioned in the previous chapters, we have 221 users from the city
of hasselt who were asked to choose the most important variables of some transport mode
choice. In addition different scenarios have been developed and asked users to provide their
choice according to the circumstances in a scenario. The nodes in FCMs are divided into
three groups: benefit nodes, situation nodes and attribute nodes. A traveller chooses the
main benefits of choosing a particular transport mode depending on the situation and the
attributes related with that situation.
For example, consider the set {situation, attribute, travel mode, benefit} with the fol-
lowing nodes in each set respectively {car availability}, {flexibility, independency, travel
time, speed}, {car, bike, bus}, {convenient, freedom}. Here the car availability is a node

[email protected]
282 Decision Aid Models for Disaster Management and Emergencies

showing the situation which is related with attributes such as flexibility, independency,
travel time, speed. The main benefits of choosing one out of three transport modes are
convenient, freedom. The expected results of cluster analysis will help the transportation
policy decision makers to understand which group of people has the inclination of which
benefits, and what the situations leading to the chosen benefit are.
In general cluster analysis is an unsupervised learning method to examine the dataset
by dividing it into the groups so that the similarity within the clusters and dissimilarity
between different clusters are maximized. Cluster analysis methods have been applied in
different fields such as engineering, social science, medical sciences, economics etc. The
detailed description of all clustering techniques is out of scope of this paper. The interested
reader can refer to [30] and [31].
In the previous subsection we presented the distance ratio algorithm to find the distance
between two FCMs. There are two wide used clustering techniques based on similarity
or distance measurements: the hierarchical approach and the partitional approach (e.g.
K-means). Hierarchical clustering algorithms produce a nested series of partitions based
on a criterion for merging or splitting clusters based on similarity. Partitional clustering
algorithms identify the partition that optimizes a clustering criterion.
The preference of one approach over another depends on the task on hand and on the
main goal of the study. As in our case we do not know the number of clusters in advance
and we do not have time complexity issue for the maps in our dataset we prefer using
hierarchical clustering.
The most wide used linkage methods are single, complete and ward linkage methods
[30]. Ward method is not efficient for our study as we do not use euclidian distance for
similarity measurement. To cluster FCMs the single-linkage method takes the distance
between two clusters as the minimum of the distances between all pairs of maps from the
two clusters. On the other hand, the complete-linkage algorithm calculates the distance
between two clusters as the maximum of all pairwise distances between maps in the two
clusters.
Single linkage suffers from a chaining effect producing elongated clusters. The study
in [30] shows that complete linkage produces more compact and more useful hierarchies
in many applications than the single-linkage algorithm. Besides to decide which algorithm
best suits our data, we calculated the Cophenetic Coefficient (CC) for single, complete,
weighted, average, ward and centroid methods.

[email protected]
A Travel Behaviour Study through Learning and Clustering of Fuzzy Cognitive Maps 283

Note that the CC shows how strong the linking of maps in the cluster tree has a corre-
lation with the distances between maps in the distance vector. This coefficient usually is
used to compare different linkage methods. The closer the CC to one the more accurately
the clustering solution reflects the data. For our dataset, the results of CC calculation show
that complete and centroid methods are the best for our data as they gave maximum values
for CC. However, the centroid method also suffers from chaining effect, and the best option
that we used for our cluster analysis is complete linkage.

12.5.3 Finding the optimum number of FCMs clusters

After clustering of FCMs there are several important questions to be considered: How
good is the clustering? What is the optimum number of clusters? Or what are the main
patterns to be explored in the clusters? In this section we discuss the optimum number of
the clusters and we propose the adjustment of well-known Davies-Bouldin Index (DBI).
The results are twofold: first we find the optimum number of clusters; second, we use the
concept of central map to analyse the clusters separately.
There are several cluster validity indexes that lead to the decision of the optimum num-
ber of clusters. We will explore two of them only, namely the Silhouette index (SI) and DBI
[32], [33]. We compare the results of SI and the proposed extension of DBI index.
For each sample in a cluster the SI assigns a confidence value showing how good the
sample has been classified. The SI for ith sample in cluster X j ( j = 1, . . . , c) where c is the
number of clusters is defined as follow in (12.8).
b(i) − a(i)
s(i) = (12.8)
max{a(i), b(i)}
where a(i) is the average distance between ith sample and all other samples included in X j ,
b(i) is the minimum average distance between and all samples in cluster Xk (k = 1, . . . , c;
k �= j). The closer s(i) to 1 the better the ith sample assigned to its cluster. The cluster Sil-
houette index which shows the heterogeneity and isolation properties is defined as follows
in (12.9).
m
1
S( j) = s(i) (12.9)
m
i=1

To find the optimum number of clusters the average value of S( j) is calculated for
different number of clusters, and the one with maximum value is taken as the optimum
number. As SI, DBI also aims to find the optimum number of clusters. DBI is defined as

[email protected]
284 Decision Aid Models for Disaster Management and Emergencies

follows in (12.10).
c
1 Δ(Xi ) + Δ(X j )
DBI = max (12.10)
c i= j δ (Xi , X j )
i=1

where ΔXi /ΔX j is the average distance of the samples in ith / jth cluster to centre of the
cluster, δ (Xi , X j ) is the distance between the centres of the ith and jth clusters. The clus-
ter configuration minimizing DBI is takes as an optimum number of clusters. As we are
clustering FCMs, we adjust DBI to be applied for FCMs. We first derive a central map for
each cluster, then calculate the distance of each map in a cluster from its central map, the
distance between the central maps of different clusters and derive BDI index defined above.
To find the central map first we do some observations about the structure of the maps
that we have for our task. As all the maps in our dataset have the same hierarchical struc-
ture, some nodes are only transmitters (situation nodes), some are only receivers, and some
others are both transmitters and receivers. For each node we compute the conceptual cen-
trality as in explained before.
A node is included in the central map if it exists in more than half on maps of a cluster.
The weights of the links are calculated as the average value of the weights from all maps
that contain both nodes that comprise the link. Once we have all central maps of all clusters
we calculate BDI for different number of clusters to find the optimum number of clusters.
Note that once we identify the number of clusters, we use central maps for further analysis
of each cluster.
In the next section we show the results explained here and above sections about clus-
tering of FCMs.

12.5.4 Cluster estimation and validation

In this section we explain the results of cluster analysis implemented in Matlab environ-
ment. The dendrogram in Figure 12.10 illustrates the arrangement of the clusters produced
by Equations (12.6) and (12.7) for distance matrix and complete linkage for inter-cluster
distance. Note that we also tested other linkage methods on our data set. However, the re-
sults were not satisfactory as the resulted clusters were not compact as some of the clusters
had only few maps while the others were with very big size.
To find the optimum number of clusters we calculated the SI for ten clusters as shown
in Table 12.3. Moreover, we derived also DBI for ten clusters first finding the central maps
for each cluster configuration. The optimum number of clusters with both SI and DBI is
found equal to six (see Figure 12.11) with 15, 53, 33, 23, 59 and 38 maps in each cluster

[email protected]
A Travel Behaviour Study through Learning and Clustering of Fuzzy Cognitive Maps 285

Fig. 12.10 The clustering results’ interpretation with a dendrogram.

respectively. After finding the optimum number of clusters, we first analyse “situation-
benefit-attribute” sets for each cluster, afterwards, we analyse the demographic features of
the clusters.
Due to that we use central maps that we derived for DBI: thus, we have six maps
representing each cluster. For demographic analysis we take into consideration the users’
age, gender, income, household size, education level, occupation, parking type (paid or
free), bike ownership, number of owned cars and having a bus card. The results of chi-
square test indicate that some of these variables namely the income, occupation, bus card,
parking availability and number of cars are dependent variables.
In the next section we explain each of six clusters in more details in terms of maps
structure, nodes, links as well as demographic features of the users in each cluster.

12.6 Clustering and learning: digging deeper into individuals mental maps

As we have already mentioned in the previous chapters, we used learning and cluster-
ing of FCMs as predictive and descriptive tools to analyse the preferences of travellers,
different travellers groups under certain situations corresponding to different scenarios. In
addition we can use clustering and learning results to analyse the users decision making
preferences in the different levels of abstraction. Particularly if we analyse clustering with

[email protected]
286 Decision Aid Models for Disaster Management and Emergencies

Table 12.3 SI for different clustering configurations.

Cluster
S S1 S2 S3 S4 S5 S6 S7 S8 S9 S10
Number
2 0.354 0.342 0.367
3 0.389 0.449 0.376 0.342
4 0.555 0.535 0.392 0.449 0.472
5 0.612 0.768 0.659 0.435 0.648 0.552
6 0.829 0.848 0.734 0.868 0.735 0.948 0.842
7 0.707 0.789 0.633 0.748 0.634 0.868 0.535 0.742
8 0.524 0.652 0.518 0.489 0.473 0.571 0.456 0.678 0.357
9 0.529 0.491 0.632 0.352 0.418 0.691 0.548 0.468 0.435 0.342
10 0.419 0.461 0.515 0.591 0.422 0.252 0.318 0.389 0.448 0.368 0.435

Fig. 12.11 SI and DBI values for different number of clusters.

maps before learning, we will obtain the groups of people that are similar to each other in
their initial mental representation of transport mode choice. Figure 12.12 shows the gen-
eral overview of our approach of using clustering and learning of FCMs to analyse users’
preferences in different stages of reasoning.
If we go further and provide more information to the users about different scenarios
and ask them to evaluate their behaviour in different circumstances, we will acquire deeper
knowledge representation. In this section we analyse the results of clustering of user pro-
vided FCMs before learning and after learning: thus we investigate their preferences and
similarities when they choose only the nodes and the weighted links (first case) and when
in addition they evaluate also different scenarios (second case).

[email protected]
A Travel Behaviour Study through Learning and Clustering of Fuzzy Cognitive Maps 287

Fig. 12.12 Decision making analysis procedure.

As we have already stated in the previous section, we obtained six different clusters
when we take the user provided maps, while after learning the maps and clustering them, we
get only four clusters. Figure 12.13 (the width of the relations corresponds to the frequency
of links evaluation by group members) shows two central maps of two clusters out of six
clusters while Figure 12.14 shows two central maps of clusters out of four learnt maps
clusters. For simplicity of the map visualization we omit the three decision nodes and the
final utility node.
Before learning of FCMs the travellers are grouped as follows (for simplicity we do not
show all six central maps):

• In the first group are young travellers (less than 30 years old), with medium income
(from 2 000 to 4 000 Euros) having more than two cars and a small household size
(living alone or in two). This group is the most homogenous in terms of their preferred
nodes and evaluated links: thus there are not links or nodes that are chosen only by few
members of the group. The main benefits combining the travellers in this group is nor-
mally their desire to be free while traveling, to have convenience, paying attention on
the required time and effort of a specific travel mode. The attributes that they consider

[email protected]
288 Decision Aid Models for Disaster Management and Emergencies

Fig. 12.13 Central maps of two different clusters before maps learning.

as very important for chosen benefits are the travel time as well as the flexibility and
independence. In addition they take into account also the travel time, the treatment of
bags; mental effort needed and direct traveling possibility.
• In the second group are older-retired people (more than 60 years old), having only
one car, with household more than two persons and who mainly use paid parking.
This group has most of the preferences as the previous one, in addition users in this
group pay more attention on physical comfort, easiness of parking, physical effort,
and the situation precipitation taking into account the shelter availability as well as the
reliability of a transport mode under the chosen situation variables.
• In the third cluster are grouped the travellers having more than 40 years old who are
either employed or retired, with high income (more than 4000 Euros), using mainly the
paid parking. The preferences of this group are very similar to the second group with
the difference of choosing a benefit being healthy, convenience and paying no attention
on reliability. Actually, this group and the previous one are in the same cluster after
learning of maps (shown in left of Figure 12.13).
• The fourth group is different from previous ones (shown in right of Figure 12.13). In
this group mainly are students and young employers with medium or high income,
and they mainly use free parking. There is one benefit variable unique to this group:
having fun. Another difference with respect to the other groups is the set of situation
variables that the group chooses in making a decision of a transport mode. Namely the
group considers the situations related with precipitation, number and size of the goods
purchased, as well as the available time.

[email protected]
A Travel Behaviour Study through Learning and Clustering of Fuzzy Cognitive Maps 289

• The fifth group clusters the travellers with low education, high income preferring not
to provide information about their income and declaring themselves as unemployed.
This group has a unique benefit variable that is the assurance and certainty related with
their decision. The unique situation that the group considers and other groups do not,
is the availability of parking and some attributes that the group considers as important
are the reliability, accessibility, travel time, transport mode preference, etc.
• The last group is the only group that gives importance to safety and security, saving
money, and the traveling cost Here travellers have low education, low income (less
than 2 000 Euros), mainly students and unemployed or retired with small household.

These were the preferences of users when they choose the variables and the links be-
tween variables that they consider important for travel decisions. However, these prefer-
ences are changing when users are provided more information in terms of different possible
scenarios. Indeed, in different levels of information, knowledge and experience people tend
to act differently, in the beginning it is more reactive response, than with deeper understand-
ing and analysing gives more situational awareness leading to more rational decisions.

Fig. 12.14 Central maps of two different clusters after learning of maps.

For example, in our case study we found that in the beginning a group of travellers con-
sider the cost and saving money as important factors for traveling decision. However, with
awareness of different circumstances and scenarios, their preferences towards the men-

[email protected]
290 Decision Aid Models for Disaster Management and Emergencies

tioned factors are decreasing. Therefore, the travellers with these preferences (the sixth
cluster described above) are distributed among other clusters: thus only four cluster groups
have been identified with adjusted maps after applying PSO learning algorithm.
As a concluding remark, one should notice that for policymakers it is important to
distinguish in which case they have to refer to the initial clusters and when to the clustering
of the adjusted maps. Clustering of learnt maps takes into account different scenarios;
therefore it gives more comprehensive view of traveller’s subconscious decision making
preferences. Consequently, policymakers can refer to these clustering results when the task
is related with a complex policy testing.
On the other hand, the clustering of non-learnt maps can be used when testing new
policies related with simpler measures.

12.7 Conclusions

In this study we proposed FCMs as a modelling tool to analyse the behaviour of com-
plex systems, where it is very difficult to describe the entire system by an accurate math-
ematical model. Consequently, it is easier and practical to represent it in a graphical way
showing the causal relationships between concepts. Since this symbolic method of mod-
elling and control of a system is easily adaptable and relies on human expert experience
and knowledge, it can be considered intelligent.
As a case study we discussed the travel behaviour and we analysed the main concepts
that affect on the travellers choice of a specific transport mode such as car, bike or bus. We
explored the learning and clustering algorithms for FCMs to offer policymakers a frame-
work and real data to play with, in order to study and simulate individual behaviour and
produce important knowledge to use in the development of city infrastructure and demo-
graphic planning. We used clustering and learning of FCMs to give deeper inside of trav-
ellers mind while taking a decision of a transport mode.
We showed the benefits of the application of a learning method inspired in the PSO
metaheuristic, obtaining an improvement on the knowledge structures originally modelled.
It is an unsupervised weight adaptation methodology that have been introduced to fine-tune
FCM causal links and accompanied with the good knowledge of a given system or process
can contribute towards establishment of FCMs as a robust technique. Experimental results
based on simulations verify the effectiveness, validity and advantageous behaviour of the
proposed algorithm. The area related to the learning of FCMs is very promising because

[email protected]
Bibliography 291

the FCMs obtained are directly interpretable by humans, representing useful tool to extract
information from data about the relations among the concepts or variables inside a domain.
The clustering results can be used by policy decision makers to analyse the preferences
of different groups of travellers in terms of their preferred benefits, situation and attribute
variables. We use the concept of a central map as a representative for each group and
obtained central maps with both initial maps and maps after applying the PSO learning
algorithm. The comparison of two clustering results shows that travellers change their
preferences while providing more information of different scenarios.In keeping with a more
rigorous approach, future work should attempt to probe the backgrounds, experience and
knowledge of participants rather than merely to analyse the resulting maps. This would
also require the investigator to monitor participants’ general performance in emergency
management courses more closely and to relate this quantitatively to the attributes of their
emergency maps. Moreover, there are unresolved issues about how much prior experience
and knowledge the emergency manager would normally have of the geographical area in
which the disaster takes place. In this respect the present exercise is merely a pilot study:
the paper therefore serves as a preliminary attack of an interesting problem about which
there is much more to learn.
As a future extension of this study, it could be interesting to explore an approach where
first clustering is executed and one map represents each cluster (the central one or an aggre-
gation of all maps), so the learning process can be done for this structure but considering
all scenarios of the maps included in the cluster. A current work dealing with this approach
is in progress, and results of different approaches will be reported to community.

Bibliography

[1] J. Gutiŕrez, Análisis de los efectos de las infraestructuras de transporte sobre la accesibilidad y
la cohesión regional. Estudios de Construcción y Transportes. Ministerio de Fomento. España
(2006).
[2] D. Janssens, Tracking Down the Effects of Travel Demand Policies. Urbanism on Track. Re-
search in Urbanism Series. IOS Press (2008).
[3] J.T. Rodríguez, B. Vitoriano, J. Montero & V. Kecman, A disaster-severity assessment DSS
comparative analysis. OR Spectrum, 33, pp. 451–479 (2011).
[4] R. Axelrod, Structure of decision: the cognitive maps of political elites. Princeton University
Press (1976).
[5] L. Mkrtchyan, D. Ruan, Using Belief Degree Distributed Fuzzy Cognitive Maps for Energy
Policy Evaluation, Risk Management in Decision Making. Intelligent Methodologies and Ap-
plications. Springer (in press) (2011).
[6] B. Kosko, Fuzzy Cognitive Maps. International Journal of Man-Machine Studies. Vol. 24, pp.
65–75 (1986).

[email protected]
292 Decision Aid Models for Disaster Management and Emergencies

[7] J. Aguilar, A Dynamic Fuzzy-Cognitive-Map Approach Based on Random Neural Networks.

Journal of Computational Cognition. Vol. 1, pp. 91–107 (2003).
[8] C. Carlsson, Adaptive Fuzzy Cognitive Maps for Hyperknowledge Representation in Strategy
Formation Process. IAMSR, Abo Akademi University (2005).
[9] U. Ozesmi, S. L. Ozesmi, Ecological model based on people’s knowledge: a multi–step cogni-
tive mapping approach. Ecological Modeling. Vol. 176, pp. 43–64 (2004).
[10] M. Schneider, E. Shnaider, A. Kandel, G. Chew, Automatic construction of FCMs. Fuzzy Sets
and Systems. Vol. 93, pp. 161–172 (1998).
[11] S. Mohr, Software Design for a Fuzzy Cognitive Map Modeling Tool. Tensselaer Polytechnic
Institute (1997).
[12] J. Contreras, Aplicación de Mapas Cognitivos Difusos Dinámicos a tareas de supervisión y
control. Trabajo Final de Grado. Universidad de los Andes. Mŕida, Venezuela (2005).
[13] M. León, Fuzzy Cognitive Maps for modeling Complex Systems. Advances in Artificial Intel-
ligence. Part I, LNAI 6437, pp. 166–174. Springer-Verlag Berlin Heidelberg (2010).
[14] P. Taco, Trip Generation Model: A New Conception Using Remote Sensing and Geographic
Information Systems. Photogrammetrie Fernerkundung Geoinformation (2000).
[15] M. León, Cognitive Maps in Transport Behavior. MICAI Mexican International Conference on
Artificial Intelligence. IEEE Computer Society Press (2009).
[16] M. Dijst, Spatial policy and passenger transportation. Journal of Housing and the Built Envi-
ronment. Vol. 12, pp. 91–111 (1997).
[17] C. Khisty, Transportation Engineering: an introduction. Prentice Hall. pp. 388–397 (1990).
[18] P. Cassin, Ontology Extraction for Educational Knowledge Bases. Spring Symposium on Agent
Mediated Knowledge Management. Stanford University, American Association of Artificial
Intelligence (2003).
[19] M. León, A Revision and Experience using Cognitive Mapping and Knowledge Engineering
in Travel Behavior Sciences. “POLIBITS” Computer Science and Computer Engineering with
Applications. ISSN 1870-9044, pp. 43–50 (2010).
[20] C. Ros,́ Overcoming the knowledge engineering bottleneck for understanding student language
input. International Conference of Artificial Intelligence and Education (2003).
[21] D. Koulouritios, Efficiently Modeling and Controlling Complex Dynamic Systems using Evo-
lutionary Fuzzy Cognitive Maps. International Journal of Computational Cognition. Vol. 1, pp.
41–65 (2003).
[22] J. Mcmichael, Optimizing Fuzzy Cognitive Maps with a Genetic Algorithm AIAA 1st Intelli-
gent Systems Technical Conference. Chicago, Illinois (2004).
[23] M. León, A Fuzzy Cognitive Maps Modeling, Learning and Simulation Framework for studying
Complex System. IWINAC International Workconference on the interplay between Natural and
Artificial Computation. Part II, LNCS 6687, pp. 243–256. Springer-Verlag Berlin Heidelberg
(2011).
[24] M. Glykas, Fuzzy Cognitive Maps: Advances in Theory, Methodologies, Tools and Applica-
tions. Studies in Fuzziness and Soft Computing. Vol. 247. Springer. ISBN: 978-3-642-03219-6
(2010).
[25] K. Langfield-Smith, A. Wirth, Measuring differences between cognitive maps. The Journal of
the Operational Research Society. Vol. 43, No. 12, pp. 1135–1150 (1992).
[26] L. Markoczy, J. Goldberg, A method for eliciting and comparing causal maps. Journal of Man-
agement. Vol. 21, No. 2, pp. 305–333 (1995).
[27] L. Ortolani, N. McRoberts, N. Dendoncker, M. Rounsevell, Analysis of Farmers’ Concepts of
Environmental Management Measures: An Application of Cognitive Maps and Cluster Anal-
ysis in Pursuit of Modeling Agents’ Behavior. Journal of Fuzzy Cognitive Maps. Springer, pp.
363–381 (2010).

[email protected]
Bibliography 293

[28] S. Alizadeh, M. Ghazanfari, M. Fathian, Using Data Mining for Learning and Clustering FCM.
International Journal of Computational Intelligence. Vol. 4, No. 2 (2008).
[29] C. Eden, Analyzing cognitive maps to help structure issues or problems. European Journal of
Operational Research. Vol. 159, No. 3. Elsevier, pp. 673–686 (2004).
[30] A. Jain, M. Murty, P.Flynn, Data clustering: a review. Journal of ACM computing surveys
(CSUR), Vol. 31, No. 3, pp. 264–323 (1999).
[31] R. Wunsch, Survey of clustering algorithms. IEEE Transactions on Neural Networks. Vol. 16,
No. 3, pp. 645–678 (2005).
[32] D. Davies, D. Bouldin, A cluster separation measure. IEEE Transactions on Pattern Analysis
and Machine Intelligence, No.2, pp. 224–227 (1979).
[33] N. Bolshakova, F. Azuaje, Cluster validation techniques for genome expression data, Journal
of Signal Processing, Vol. 83, No. 4, pp. 825–833 (2003).

[email protected]
Chapter 13

Fire and explosion safety assessment in container

line supply chain

Yu-wang Chen 1 , Dong-Ling Xu 1 , Jian-Bo Yang 1 , and Dawei Tang 2,∗

1Manchester Business School, the University of Manchester, Manchester, UK, M15 6PB
2Safety and Environment Assurance Center, Unilever Colworth, Colworth Science Park,
Sharnbrook, Bedfordshire, UK, MK44 1LQ
∗ [email protected]

Safety assessment is an important part of maritime safety analysis and management as it

can provide decision support to safety-based operations. In this chapter, we first review
fire and explosion safety related regulations and assessment approaches, and then present a
hierarchical assessment framework for evaluating safety against fire and explosion hazards
in container line supply chain (CLSC). Main criteria and factors influencing fire and explo-
sion safety on containerships are investigated, and a safety modeling and synthesis method
using the Evidential Reasoning (ER) approach is briefly discussed. The outcomes gener-
ated by the proposed safety assessment method are highly consistent with “goal-setting”
safety regulations.

13.1 Introduction

Container line supply chain (CLSC), which transports cargo efficiently across seas and
into ports throughout the world, has contributed significantly in facilitating global eco-
nomic development and prosperity. Approximately 90 percent of world trade moves in
shipping containers, and over 250 million containers are transported annually (RAND,
2003). However, due to its characteristics of having complex physical and information
flows, CLSC is also highly vulnerable to many possible undesirable accidents such as ma-
chinery failure, contact, collision, fire and explosion, grounding, etc. (EMSA, 2008; MAIB,
2008). Among these “major accidents hazards”, fire and explosion may be one of the most
dangerous hazards with the potential to cause disastrous consequences (Wang, 2002). For
example, on March 21, 2006, M.V. Hyundai Fortune, a 64,054 gt. Panama flag contain-

B. Vitoriano et al. (eds.), Decision Aid Models for Disaster Management 295
and Emergencies, Atlantis Computational Intelligence Systems 7,
DOI: 10.2991/978-94-91216-74-9_13,  Atlantis Press 2013

[email protected]
296 Decision Aid Models for Disaster Management and Emergencies

ership, was severely damaged in an accidental fire and explosion in aft on-deck container
stacks and eventually abandoned after efforts to contain the fire failed. Besides the po-
tential disastrous consequences, MAIB (Marine Accident Investigation Branch) statistics
indicate that about 12% accidents of UK merchant vessels ( 100 gt) from 1997 to 2009
were caused by fire and explosion hazards as shown in the following Fig. 13.1 (MAIB,
2008).

Fig. 13.1 Statistics on accidents of UK merchant vessels

In response to the fact that CLSC is a dominant way to transport cargo but highly
vulnerable to fire and explosion hazards, in past decades a set of regulations were issued
by ISO (International Standard Organisation), IMO (International Maritime Organisation),
UK HSE (Health & Safety Executive), etc. Meanwhile, a number of safety assessment
approaches, such as Quantitative Risk Analysis (QRA) and Formal Safety Analysis (FSA),
have been widely used to assess the safety levels corresponding to various hazards in mar-
itime transportation (Wang, 2001). However, limited work has been carried out specifically
on the fire and explosion safety assessment in CLSC. In this chapter, a well-structured
assessment framework is proposed to assess safety against fire and explosion hazards in
CLSC. The rest of this chapter is organised as follows. A brief review of safety-related reg-
ulations and assessment approaches is provided in Section 13.2. A hierarchical assessment
framework for evaluating safety against fire and explosion hazards in CLSC is presented

[email protected]
Fire and explosion safety assessment in container line supply chain 297

and a safety modeling and synthesis method using the Evidential Reasoning (ER) approach
is briefly discussed in Section 13.3. Conclusions are drawn in Section 13.4.

13.2 Review of safety-related regulations and assessment approaches

Safety is very important in CLSC. Any disruption or damage on CLSC caused by ac-
cidents may result in non-availability of goods at the expected time and place, and in the
expected conditions. It can further incur economic losses and damage credibility with in-
vestors and other stakeholders. Traditionally, the safety strategies of maritime transpiration
and offshore installations over hazards are completely relied on ISO, IMO and classification
society regulations (HSE, 2001). These regulations are usually guided by expert judgments
responding to serious marine accident experience.
Internationally, IMO SOLAS (International Convention for the Safety of Life at Sea)
specifies detailed fire safety provisions for all ships and specific measures for cargo ships
(IMO, 2000).
In the UK, PFEER (Offshore Installations - Prevention of Fire and Explosion and Emer-
gency Response Regulations) provides general requirements for preventative and protective
measures to manage fire and explosion hazards, to secure effective emergency response,
and to ensure compliance with regulations (HSE, 1995); UKOOA (Offshore Operators As-
sociation) fire and explosion guidance provides an integrated approach to the management
of fires and explosions (UKOOA, 2003).
On the basis of safety related regulations, many typical safety assessment approaches
have also been widely applied to the evaluation of likelihood and consequences of haz-
ards, such as Fault Tree Analysis (FTA), Failure Modes, Effects and Criticality Analysis
(FMECA), QRA and FSA (HSE, 2001). QRA is a complicated technique for risk assess-
ment, and its frequency analysis and consequence modeling methods have been success-
fully applied to the risk management of fires and explosions (Franks, 2000; Vinnem, 2007).
FSA is probably the most formal and comprehensive framework for general risk manage-
ment in the field of maritime safety (Wang, 2001; HSE, 2002). It was first developed by the
UK Maritime and Coastguard Agency (MCA) and later incorporated into the IMO interim
guidelines for safety assessment. The approach mainly consists of five steps: (1) hazard
identification, (2) risk assessment, (3) risk control options, (4) cost-benefit analysis, and
(5) decision making.

[email protected]
298 Decision Aid Models for Disaster Management and Emergencies

Despite the variety of safety assessment approaches, FSA and other conventional as-
sessment techniques only provide an assessment process instead of a well-structured assess-
ment framework. Furthermore, these approaches are not designed to specifically address
fire and explosion safety issues in CLSC.

13.3 Fire and explosion safety assessment

To improve safety against fire and explosion hazards in CLSC, safety assessment frame-
work needs to be developed on the whole CLSC level. Generally, CLSC can be character-
ized by the following physical stages: shipment consolidation, inland transportation, port
of loading, maritime transportation, port of unloading, inland transportation, shipment de-
consolidation (as shown in Fig. 13.2).

Supplier/ Shipment Inland Port of Maritime Port of Inland Shipment

Importer
Factory Consolidation Transportation Loading Transportation Destination Transportation Deconsolidation

Fig. 13.2 Typical physical stages in CSLC

Internal vulnerability exists at each element of each stage. Thus many factors may influ-
ence the safety of CLSC. Basically, we can classify those factors into different categories.
Thus, at each stage, a hierarchical structure with all contributing factors can be constructed
to assess safety against fire and explosion hazards.

13.3.1 Assessment Criteria Hierarchy

In the framework, the safety level against hazard-based risks is modeled by three basic
parameters, namely occurrence likelihood, probability of occurrence of potential conse-
quence, and potential consequence. The evaluation of these three parameters can be further
decomposed into the assessment of contributing factors as shown in Fig. 13.3. To evaluate
occurrence likelihood of fire and explosion events, we need assess the inherent features
of cargo and a containership and prevention measures. To evaluate probability of occur-
rence of potential consequence, we need assess detection, control, mitigation measures and
recovery difficulty. Potential consequence can be evaluated by potential losses to people,
property, environment and business systems.

[email protected]
Fire and explosion safety assessment in container line supply chain 299

Safety against fire and

explosion hazards

Inherent Recovery Potential

Prevention Detection Control Mitigation
features difficulty consequence
Consequence
Occurrence likelihood Probability to potential consequence
severity
Codes, guidelines, standards and regulations Knowledge

Fig. 13.3 Assessment criteria hierarchy of fire and explosion safety

It is worth noting that the performance standards of inherent features, prevention, de-
tection, control and mitigation measures can be learnt from fire and explosion safety related
codes, guidelines, standards and regulations, and recovery difficulty and potential conse-
quence can be assessed by using experts’ knowledge and historical accident data. As a case
study, in the following we briefly discuss the main factors influencing the safety against fire
and explosion hazards on containerships.
(1) Occurrence likelihood
The occurrence likelihood defines the probability that a fire and explosion event occurs
onboard a containership.

• Inherent features include flammable and explosive characteristics of cargo (partic-

ularly for dangerous goods and hazardous substances, e.g., oils, chemicals, acids,
flammable solids, substances liable to spontaneous combustion, substances which can
emit flammable gases on contact with water, etc.), and inherent fire and explosion
safety of the containership (e.g., basic structure and layout, ventilation system of ma-
chinery spaces, etc.).
• Preventative measures, including managerial, operative and technical measure, are
the most effective means to minimise the occurrence likelihood of fire and explosion
events and its associated risks. The performance of preventative measures can be eval-
uated by design measures, procedural controls, early warning systems, and prevention
and protection systems.

(2) Probability of occurrence of potential consequence

[email protected]
300 Decision Aid Models for Disaster Management and Emergencies

This basic parameter represents the probability that potential consequence happens,
given the occurrence of a fire and explosion event. It is affected by detection, control,
mitigation measures and recovery difficulty.

• Detection measures can reduce undesired consequences of fire and explosion events
by detecting and locating fires and alerting the navigation bridge and fire teams. Gen-
erally, the detection system is installed to transmit information to control points. Its
performance is determined by automatic sprinklers, fire detection and alarm systems,
and fire patrols.
• Control measures can limit the scale, intensity and duration of potential consequence.
If a fire or explosion event happens, the containership has the capability of limiting the
fire growth within every space by fire-extinguishing systems and fire brigade.
• Mitigation measures are deployed to mitigate the effects of fire and explosion events,
including emergency training and drills, life-saving appliances, emergency instruc-
tions, and communication systems. For examples, it is required that at least one escape
route must remain available during and after a fire and explosion event.
• Recovery difficulty is characterized by those activities that are needed to provide initial
recovery and to provide the basis to facilitate long-term recovery activities, e.g., cargo
replacement and ship repair, and support the resumption of trade within accident areas
(those areas directly impacted by the effects of the accident) and non-accident areas
(those areas indirectly affected by the consequences of the accident) (DHS, 2007).

(3) Potential consequence

The potential consequence describes the magnitude of possible consequence. It can be
evaluated on the following dimensions:

• Human cost (e.g., physical and psychological harm to people, human death)
• Property damage (e.g., damage on cargo and containership, property losses)
• Environmental damage (e.g., environmental pollution, pollution on ecosystem)
• Losses to business systems (e.g., corporate image cost, economic losses to the com-
munity)

To assess the potential consequence of fire and explosion events, a fire and explosion
scenario with a specific set of conditions usually needs to be defined.

[email protected]
Fire and explosion safety assessment in container line supply chain 301

13.3.2 Subjective safety modeling and synthesis

In safety analysis, the assessment of each contributing factor can be described by sub-
jective linguistic variables. For example, one may often use such evaluation grades as
“catastrophic”, “critical”, “marginal”, and “negligible” to evaluate the parameter of po-
tential consequence (PC). In addition, uncertainties are always associated with maritime
safety assessment due to lack of reliable assessment data and lack of confidence. As such,
a belief structure (Yang and Singh, 1994) can be applied to model the subjective safety
assessment with uncertainty. For example, the subjective safety description S(PC) can be
expressed in the following form:
S(PC) = (β1,PC , catastrophic), (β2,PC , critical), (β3,PC , marginal), (β4,PC , negligible)
where βi,PC (i = 1, 2, 3 or 4) represents the extent to which PC is assessed to the ith
subjective expression.
In the assessment criteria hierarchy, the safety level on a general or higher level criterion
y is usually evaluated on the basis of evidence information collected from the bottom level
safety factors. It is therefore important to synthesize the safety assessment information
on the bottom level safety factors in a rational way so as to obtain the safety assessment
of the whole CLSC. The following Evidential Reasoning (ER) approach can be employed
to synthesize the subjective assessment information represented by the belief structures
(Wang, 2000; Liu et al., 2008).
Suppose there are L basic safety factors associated with assessing a general safety cri-
terion y in the criteria hierarchy. We can define a set of basic factors as follows:
E = {e1 , e2 , . . . , ei , . . . , eL }
Suppose the weights of the basic factors are given by w = {w1 , w2 , . . . , wi , . . . , wL },
where wi is the relative weight of the ith factor ei with 0  wi  1. The weights of the
basic factors may be estimated using some existing methods such as simple rating and
pairwise comparison methods. Further, suppose N mutually exclusive evaluation grades
Hn (n = 1, . . . , N) can be defined as a complete set of standards for assessing a basic safety
factor. In the assessment of potential consequence as discussed above, Hn represents a set
of four linguistic variables. Therefore the assessment on the safety factor ei (i = 1, . . . , L)
may be represented as the following belief distribution:
S(ei ) = (Hn , βn,i ), n = 1, . . . , N , i = 1, . . . , L
where βn,i  0, n=1 βn,i 
N
1, and βn,i denotes a degree of belief. The above distributed
assessment represents that the factor ei is assessed to the grade Hn with the degree of belief
of βn,i , n = 1, . . . , N.

[email protected]
302 Decision Aid Models for Disaster Management and Emergencies

Let βn be a degree of belief to which the general safety criterion y is assessed to the
grade Hn . Thus βn (n = 1, . . . , N) can be generated by aggregating the assessments for all
the basic safety factors ei (i = 1, . . . , L). Let mn,i be a basic probability mass representing
the degree to which the ith basic factor supports the hypothesis that the general criterion y
is assessed to Hn . Let mH,i be a remaining probability mass unassigned to any individual
grade after all N the grades have been considered for assessing the general criterion y as far
as ei is concerned. mn,i and mH,i can be calculated from βn,i as follows:

mn,i = wi βn,i ; n = 1, . . . , N; i = 1, . . . , L
N N
mH,i = 1 − mn,i = 1 − wi βn,i , i = 1, . . . , L
n=1 n=1

mH,i can be decomposed into mH,i and mH,i as follows:

N
mH,i = 1 − wi, mH,i = wi 1 − βn,i , i = 1, . . . , L
n=1

with mH,i = mH,i + mH,i .

Then, the final assessment distribution can be generated using the following analytical
ER algorithm (Wang et al., 2006),
L L
{Hn } : mn = μ (mn,i + mH,i + mH,i ) − (mH,i + mH,i ) , n = 1, . . . , N
i=1 i=1
L L
{H} : mH = μ (mH,i + mH,i ) − (mH,i )
i=1 i=1
L
{H} : mH = μ (mH,i )
i=1
N L L −1

μ= (mn,i + mH,i + mH,i ) − (N − 1) (mH,i + mH,i )

n=1 i=1 i=1
mn
{Hn } : βn = , n = 1, . . . , N
1 − mH
mH
{H} : βH = , n = 1, . . . , N
1 − mH
Then the aggregated safety assessment for the general criterion y can be described by
the following belief distribution:

S(y) = {(Hn , βn ), n = 1, . . . , N}

[email protected]
Fire and explosion safety assessment in container line supply chain 303

13.3.3 Implementation of IDS software tool

The synthesis process of using the ER approach above can be implemented with the
support of the IDS (Intelligent Decision System) tool (Xu and Yang, 2003). There are 3
major steps in using the IDS tool for safety assessment: model implementation, information
collection and aggregation, and assessment result generation.
(1) Model implementation
Using the IDS tool, the construction of the assessment criteria hierarchy is straightfor-
ward. The IDS main window is shown in Fig. 13.4, where there are a tree view window for
displaying the criteria hierarchy and a list view window for displaying the alternatives for
assessment.

Fig. 13.4 IDS main window and the assessment criteria hierarchy

After the criteria hierarchy is structured, each criterion needs to be further defined. The
definition includes an explanation of the criterion, whether it is a quantitative or qualitative
type and the number of evaluation grades as described in Section 13.3.2.
(2) Information collection and aggregation
The safety assessment model implemented using the IDS tool as described above can be
distributed to safety managers for collecting assessment information. The belief structure
and the ER aggregation process can make maximum use of different types of raw informa-
tion, including probabilistic data, incomplete and missing data (Yang and Xu, 2002). The

[email protected]
304 Decision Aid Models for Disaster Management and Emergencies

synthesis of the assessment information from bottom level factors to higher level criteria
is through the ER approach (Yang and Singh, 1994; Yang, 2001; Yang and Xu, 2002),
which is built into the IDS tool as the kernel aggregation engine. The aggregation process
is automatic.
(3) Assessment result generation
The IDS tool can generate different types of assessment results in graphical formats,
such as safety distribution and safety ranking. The graphs enable the comparisons among
selected alternatives to be carried out on any selected areas in different levels of the criteria
hierarchy. Fig. 13.5 presents the overall safety assessment distribution of a containership
against fire and explosion hazards.

Fig. 13.5 Safety assessment distribution with belief structure

In addition, with the support of the IDS tool, sensitivity analysis can be conducted to
identify the “cause and effect” relationship between the safety factors, the criterion weights
and the assessment outcomes.

13.4 Conclusion

In this chapter, a hierarchical assessment framework is presented for evaluating safety

against fire and explosion hazards in CLSC. The main criteria and factors influencing fire
and explosion safety on containerships are investigated, and a safety modeling and synthe-

[email protected]
Bibliography 305

sis method using the ER approach is briefly discussed. The proposed fire and explosion
safety assessment model has been tested using data collected from the Port of Liverpool,
and the generated outcomes are highly consistent with the “goal-setting” safety regulations.
Sensitivity and tradeoff analysis for supporting fire and explosion safety-related decision
making could be considered in the future research.

Acknowledgements

This work forms part of the project supported by the UK Engineering and Physical
Science Research Council under Grant No.: EP/F024606/1.

Bibliography

[1] DHS (2007). Strategy to enhance international supply chain security. Department of Homeland
Security.
[2] EMSA (2008). Maritime Accident Review 2008. European Maritime Safety Agency.
[3] Franks A., G. Hughes, S. Hanif (2000). A comparison of accident experience with quantitative
risk assessment methodology, Health and Safety Executive, HSE Books.
[4] HSE (1995). Offshore Installations – Prevention of Fire and Explosion, and Emergency Re-
sponse (PFEER) on Regulations 1995 (2nd ed.). Health and Safety Executive, HSE Books.
[5] HSE (2001). Marine risk assessment. HSE offshore technology report 2001/063.
[6] HSE (2002). A safety based decision support system for the design of large offshore engineering
products. HSE offshore technology report 2002/08.
[7] IMO (2000). International Convention for the Safety of Life at Sea (SOLAS). International
Maritime Organization.
[8] Liu J., Yang J.B., Ruan D., Martinez L. and Wang J., (2008). Self-tuning of fuzzy belief rule
bases for engineering system safety analysis, Annals of Operations Research, 163 (1): 143–168.
[9] MAIB (2008). Annual Report 2008. Marine Accident Investigation Branch.
[10] RAND (2003). Maarten van de Voort, Kevin A. O’Brien, Seacurity: Improving the Security of
the Global Sea-Container Shipping System. Rand Europe.
[11] UKOOA (2003). Fire and Explosion Guidance Part 0: Fire and Explosion Hazard Management.
UK Offshore Operators Association.
[12] Vinnem J.E. (2007). Offshore risk assessment: principles, modelling and applications of QRA
studies (2nd edition). Springer series in reliability engineering, Kluwer Academic Publishers,
The Netherlands.
[13] Wang J., (2000). A subjective modelling tool applied to formal ship safety assessment. Ocean
Engineering 27: 1019–1035.
[14] Wang J. (2001). The current status and future aspects in formal ship safety assessment. Safety
Science 38: 19–30.
[15] Wang J. (2002). Offshore safety case approach and formal safety assessment of ships. Journal
of Safety Research, 33: 81–115.
[16] Wang Y.M., Yang J.B., Xu D.L. (2006). Environmental Impact Assessment Using the Evidential
Reasoning Approach. European Journal of Operational Research, 174 (3), 1885–1913.
[17] Xu D.L., Yang J.B. (2003). Intelligent Decision Systemfor Self-Assessment. Journal of Multi-
Criteria Decision Analysis 12, 43–60.

[email protected]
306 Decision Aid Models for Disaster Management and Emergencies

[18] Yang J.B., Singh M.G. (1994). An evidential reasoning approach for multiple attribute decision
making with uncertainty. IEEE Transactions on Systems, Man, and Cybernetics 24 (1), 1–18.
[19] Yang J.B. (2001). Rule and utility based evidential reasoning approach for multiple attribute
decision analysis under uncertainty. European Journal of Operational Research 131 (1), 31–61.
[20] Yang J.B., Xu D.L. (2002). On the evidential reasoning algorithm for multiattribute decision
analysis under uncertainty. IEEE Transactions on Systems, Man, and Cybernetics - Part A:
Systems and Humans 32 (3), 289–304.

[email protected]
Chapter 14

A Human Situation Awareness Support System

to Avoid Technological Disasters

Mohsen Naderpour and Jie Lu

Decision Systems & e-Service Intelligence Lab, Centre for Quantum Computation &
Intelligent Systems, School of Software, Faculty of Engineering and Information
Technology, University of Technology Sydney, PO Box 123, Broadway, NSW 2007,
Australia

E-mail: [email protected], [email protected]

In many complex technological systems, accidents have primarily been attributed to human
error. In the majority of these accidents the human operators were striving against signifi-
cant challenges. They have to face data overload, the challenge of working with a complex
system and the stressful task of understanding what is going on in the situation. There-
fore, to design and implement complex technological systems where the information flow
is quite high, and poor decisions may lead to serious consequences, Situation Awareness
(SA) should be appropriately considered. A level 1 SA is highly supported in these systems
through the various heterogeneous sensors and signal-processing methods but, for levels 2
and 3 there is still a need for concepts and methods. This work develops a system called the
Human Situation Awareness Support System (HSASS) that supports the safety operators
in an ever increasing amount of available risky status and alert information. The proposed
system includes a new dynamic situation assessment method based on risk, which has the
ability to support the operators’ understanding of the current state of the system, predict
the near future, and suggest appropriate actions. The proposed system does not control the
course of action and allows the human to act at his/her discretion in specific contexts.

14.1 Introduction

A technological disaster is an event caused by the failure of a technological system

and/or human error in controlling or handling the technology. Since the beginning of the
industrial revolution many serious large-scale technological systems’ accidents that had
grave consequences, such as those of Three Mile Island, Bhopal and Chernobyl, have pri-
marily been attributed to “operator error”. For instance, the release of methyl isocyanate

B. Vitoriano et al. (eds.), Decision Aid Models for Disaster Management 307
and Emergencies, Atlantis Computational Intelligence Systems 7,
DOI: 10.2991/978-94-91216-74-9_14, Ó Atlantis Press 2013

[email protected]
308 Decision Aid Models for Disaster Management and Emergencies

from the Union Carbide chemical plant in Bhopal, India, in 1984 caused 2000 human ca-
sualties, 10,000 permanent disabilities, and over 200,000 injuries, arguably making it the
worst industrial disaster in history where the accident was officially blamed on human er-
ror [1]. Human error is the biggest challenge within most industries and on the surface,
would seem to imply that people are merely careless, poorly trained, or somehow not very
reliable. In fact, in the vast majority of these accidents the human operator was striving
against significant challenges. Operators have to face both data overload and the challenge
of working with a complex system. They are drilled with long lists of procedures and
checklists designed to cope with most of these difficulties, but from time to time they are
apt to fail. In fact, the person is not the cause of these errors but so much as the final dump-
ing ground for the inherent problems and difficulties in the technologies that engineers have
created [2]. Operators generally have no difficulty in physically performing their tasks, and
no difficulty in knowing what is the correct thing to do, but they are stressed by the task of
understanding what is going on in the situation. Over the last two decades, great deal of
research has been undertaken in the area of Situation Awareness (SA).
Today, in technological systems, operators rely on the principles and design of human
computer interaction to observe and comprehend the overwhelming amount of process data
that varies rapidly. They have often been moved to a control room far away from the phys-
ical process, so that their role becomes more of a monitor or supervisor of the automation
system, which is able to pass more and more information to the operator. It is widely ac-
cepted that more data does not equate to more information. In many cases automation has
only worsened the problem [3], and operators are required to handle more data and more
responsibility. For instance, in the 1970s, a typical operator manually controlled approxi-
mately 45 control valves in one process unit. Today, an operator controls, on average 175
control valves through an automation system interface. More specifically, the number of
observable process variables in the power distribution sector grew from 200,000 to 700,000
between the years 1990 and 2000 [4]. Although experienced users tend to filter through the
overabundance of data to generate information and acquire good SA, even the most expert
operator can become swamped by the excessive amount of data provided by new technolo-
gies. In the presence of all this data, operators are finding that they are even less aware than
ever before about the situations they are controlling. This has led to a huge gap between the
massive amount of data produced and disseminated and the operator’s ability to effectively
assimilate the required data and to make a timely, accurate decision [5].

[email protected]
A Human Situation Awareness Support System to Avoid Technological Disasters 309

SA can be described as knowing and understanding what is going on around you and
predicting how things will change [6]. The problem of poor operator SA continues to
worsen as technology advances whether the operator is a pilot, a manufacturing operator,
or a manager, and it can be seen through automation-facilitated accidents throughout the
world. For example, on March 23, 2005, at Texas City, TX BP Amoco Refinery explosion,
15 workers were killed and 170 injured when a column was overfilled, overheated, and
over-pressurized on startup. A key problem identified in this catastrophic event was the
difficulty experienced by the operator in maintaining an accurate awareness of the situation
while monitoring a complex, fast moving environment [7]. Several other studies of accident
throughout many industries have found that loss of, or poor operator SA, was related to
accidents classified as human error. For instance, loss of SA has been associated with 88%
of major air carrier accidents that involved pilot errors and 58.6% of operational error in air
traffic control operations [8]. Due to the severity of the accidents that have occurred over
the last ten years, SA has become the focus of research that aims to understand operator
performance in critical, dynamic environments [9].
This research considers the applicability of SA concepts to safety in the control of
complex systems. Safety supervision continues to increase in degree of automation and
complexity as operators are decreasing. As a result, each safety operator must be able to
comprehend and respond to an ever increasing amount of available situations with risky
status and alert information. This study introduces a new system for SA enhancement
called the Human Situation Awareness Support System (HSASS).
This chapter presents the basic concepts of SA, the proposed HSASS system and how
it will be implemented, and looks at related areas of research for the future.

14.2 Basic Concepts and Related Works

14.2.1 Situation Awareness

One of the widely applicable SA definitions introduced by Endsley in 1995, describes

SA as “the perception of the elements in the environment within a volume of time and
space, the comprehension of their meaning and the projection of their status in the near
future” [10]. Endsley’s model is arranged into three hierarchical levels of SA, each stage
being a necessary precursor to the next higher level (Fig. 14.1). This model follows a chain
of information processing, from perception, through interpretation, to prediction. From the
lowest to the highest, the levels of SA are [11, 12]:

[email protected]
310 Decision Aid Models for Disaster Management and Emergencies

• Perception: Perception involves the sensory detection of significant environmental

cues. For example, operators need to be able to see relevant displays or hear an alarm.
• Comprehension: Comprehension is understanding the meaning or significance of that
information in relation to goals. This process includes developing a comprehensive
picture of the world.
• Prediction: Projection consists of extrapolating information forward in time to deter-
mine how it will affect future states of the operating environment. The higher levels
of SA allow operators to function in a timely and effective manner, even with very
complex and challenging tasks.

x System Capability
x Interface Design
x Stress & Workload
x Complexity
x Automation

Task/System Factors
Feedback

Situation Awareness

Perception of Comprehension Projection

State of the Elements in of Current of Future Performance
Decision
Environment Current Situation Status of Actions
Situation
Level 1 Level 2 Level 3

Individual Factors

Information Processing
x Goals & Objectives Mechanisms
x Preconceptions
(Expectations) Long Term Automaticity
Memory Stores

x Abilities
x Experience
x Training

Fig. 14.1 Situations Awareness in Dynamic Decision Making [10].

Endsley’s model has been used in a number of studies as a justification for structuring
the computer-supported SA process in a variety of complex positions, such as air traffic
controllers, nuclear power plant operations, anesthesiologists, military commanders, elec-
tronic warfare tacticians, automobile drivers, power plant operators, and so on [2, 5, 11, 13].

[email protected]
A Human Situation Awareness Support System to Avoid Technological Disasters 311

14.2.2 Situation Assessment

SA is a state of knowledge that has to be distinguished from the processes underlying

the achievement of SA, which should be addressed as situation assessment [10]. The situ-
ation assessment models describe basic principles and general features about how people
process information or interact with the environment to attain their SA. In fact, awareness
information for a situation is derived as the results of situation assessment. Since SA is re-
garded as a dynamic and collaborative process, assessing a situation is often required data
integration or called data fusion with support of computer based intelligent techniques.
The enhancement of operators’ SA in complex systems is a major design goal in develop-
ing operator interfaces, automation concepts and training plans in a wide variety of fields
[14–16].
As SA aims to predict the status of a situation in the near future, which is the third level
of the SA model, we need proper and effective situation assessment approaches and tools
to conduct the prediction. For example many studies have reported that machine learning
techniques could be an effective method for intelligent prediction by extracting rules from
previous data to generate new assessment results [14], but their use has been limited, possi-
bly because of the lack of rich training data for this problem [17]. In some research, authors
developed a quantitative model based on Bayesian inference and information theory, and
described the process of knowledge-driven monitoring and the revision of operators’ under-
standing of the environments [15, 18]. Other studies considered computational methods,
but these approaches often do not satisfactorily handle all forms of uncertainty, especially
when information conflicts. Therefore, human behavior models have been developed from
cognitive architectures. The limitations of these systems are that they do not easily in-
corporate cognitive factors. Consequently, some approaches used the fuzzy logic system to
address the limitations of traditional models in producing the full range of human behaviors
[19–21].

14.2.3 Representing Situation Awareness

Endsley developed a methodology to determine the aspects of a situation that are im-
portant for a particular user’s SA requirements. This methodology is known as the Goal-
Directed Task Analysis (GDTA) and it is a specific form of cognitive task analysis that
focuses on identifying the goals and critical information needs for a task context. The
GDTA process has been used in many domains to detail SA requirements. As such, it
forms an exemplary template for incorporating human cognition into an actionable model

[email protected]
312 Decision Aid Models for Disaster Management and Emergencies

by describing in detail not only a user’s information data needs (Level 1) , but also how that
information needs to be combined to form the comprehension (Level 2), and projection of
future events (Level 3) that are critical to SA, thereby providing a critical link between data
input and the decisions to be made in a goal-directed environment [22].
In this analysis, the major goals of a particular job class are identified, along with the
major sub-goals necessary for meeting each goal. Associated with each sub-goal, the major
decisions that need to be made are then identified. The SA needed for making these deci-
sions and carrying out each sub-goal are identified (Fig. 14.2). These requirements focus
not only on what data the operator needs, but also on how that information is integrated, or
combined, to address each decision.

1.0 Major Goal

1.1 1.2 1.3

Subgoal Subgoal Subgoal

Decisions Decisions Decisions

SA Requirements: SA Requirements: SA Requirements:

Level 3-Projection Level 3-Projection Level 3-Projection
Level 2-Comprehension Level 2-Comprehension Level 2-Comprehension
Level 1-Perception Level 1-Perception Level 1-Perception

Fig. 14.2 Goal-Directed Task Analysis for Determining SA Requirements [10].

This type of analysis is based on goals or objectives, not tasks. This is because goals
form the basis for decision making in many complex environments. Conducting such an
analysis is usually carried out using a combination of cognitive engineering procedures
such as expert elicitation, observation of operator performance and analysis of documenta-
tion [2].

[email protected]
A Human Situation Awareness Support System to Avoid Technological Disasters 313

Table 14.1 Safety Goals and Decisions.

1.0 Eliminate or reduce the risks to a level that is as low as reasonably practicable
1.1 Determine the risks
1.1.1 Hazards identification
• Past hazards
1.1.2 Likelihood determination
• Prior likelihood
• Posterior likelihood
1.1.3 Severity determination
• Past consequences
• Degree of losses
1.1.4 Level of Risk
• Current level
1.2 Reduce the risks
1.2.1 Establish the practical options
• Available reduction and containment options
1.2.2 Impact of the options
• New level of risk

14.3 A Human Situation Awareness Support System

14.3.1 A HSASS General Model

As discussed earlier, SA involves perceiving critical factors in the environment (SA

level 1), understanding what those factors mean, particularly when integrated together in
relation to the operator’s goals (SA level 2), and at the highest level, an understanding
of what will happen with the system in the near future (SA level 3) [2]. To determine the
features that are important for an operator’s SA, we use GDTA. The SA requirements focus
not only on what data the operator needs, but also on how that information is integrated or
combined to address each decision. In this analysis process, SA requirements are defined
as those dynamic information needs associated with the major goals, or subgoals of, the
operator to perform his/her job. The results are showed in Table 14.1 [23].
The information provided for situational awareness must be more than just information
gathering. This implies collecting the right multi-domain information across a net-centric
environment for shared awareness and presenting the results for the human to understand

[email protected]
314 Decision Aid Models for Disaster Management and Emergencies

and make quick decisions. Any new approach must efficiently bring together the human
operator, sensor equipment data, and real world events to provide a subset of actionable
information [17]. Fig. 14.3 shows the general model of HSASS.

Info. Fusion & Situation SA/ GUI Human

Hazard Iden. Assessment Operator

Alarm
Objects Current Risk Level
Data Causes
Relations Actions
Consequences
Hazards Future Risk Level
Options
Conditions

Fig. 14.3 The HSASS General Model.

14.3.2 Information Fusion and Hazard Identification Component

Critical factors in the environment (SA level 1) should be collected from various sen-
sors, which are distributed in an intended area. This aims to locate and identify objects, and
also provide a global picture of the situation, which is reported by fusing the attributes of an
object from multiple sources. Any variety of relations - physical, organizational, informa-
tional, and perceptual - can be considered, as appropriate to the given information system’s
mission. Knowledge and experience of the past can be applied to the determination of the
potential hazardous circumstances. The outcome of this step is identification of hazardous
entities or relations that have a potential to lead to loss. As can be seen from the Fig. 14.3,
this forms what we defined as the information fusion and hazard identification component.

14.3.3 Situation Assessment Component

SA level 2 relates to the operator’s understanding of the system as a whole, and emer-
gent events (such as alarms or casualties), that arise from hazardous situations. Therefore it
is necessary to understand the causality of the current situation. SA level 3 requires that the
system and operator should understand the future state so consequences of the current sit-
uation should be determined; depending on the consequent losses, the current hazards may
subsequently require risk elimination, mitigation, transfer, control or combination thereof,
so it is necessary to assist the operators to choose the appropriate actions.

[email protected]
A Human Situation Awareness Support System to Avoid Technological Disasters 315

Based on above mentioned points, the situation assessment component involves causal,
consequence and option analysis. The causal analysis techniques are predominately ap-
plied within reliability engineering and are generally supported by mathematical founda-
tions and a suite of computer based tools. The quantification of causal models entails an
objective assessment of the potential frequency or likelihood for the causal factors. These
are combined according to the rules of probability calculus and Boolean logic to generate a
normalized or absolute measure for the realization of existing hazards. Consequence anal-
ysis is concerned with what may potentially follow the occurrence of a hazardous situation.
Adverse outcomes associated with the current hazardous situation should be considered,
including various degrees of harm to people, commercial detriment to an enterprise, dam-
age to the ecology of the environment, or a combination of these factors. It is useful for all
three components to be converted and expressed in a common currency, such as money, for
potential comparison and aggregation in order to provide a coherent view of the totality of
loss associated with a hazardous situation. Options analysis provides the future necessary
actions that should be implemented to eliminate or reduce the risks. On identification and
recording, it is essential to estimate the likely effects and potential benefits of each option
on the consequent safety, commercial and environmental losses, in order to establish the
objective and systematic criteria for selection and implementation. This is a requirement
of the statutory legal framework in some countries (e.g. the UK) to ensure that the safety
risks are reduced to As Low As Reasonably Practicable (ALARP) levels.

14.4 HSASS Implementation

14.4.1 Environment Description

To illustrate how to implement the HSASS into a real world environment, we use the
example of a petrochemical plant with expert systems as artificial intelligence tools. An
ethylbenzene process plant, involving two reactors and two distillation columns, as shown
in Fig. 14.4, is chosen.
An exothermic reaction occurs in Reactor 1 (R1) at 160 ◦ C and 9 bar, in which benzene
(B) and ethylene (E) react to produce ethylbenzene (EB). The undesirable reaction of ethy-
lene and ethylbenzene to produce higher-order species, for example, diethylbenzene (DEB)
is suppressed by the large excess of benzene in R1. Any DEB produced is separated from
ethylbenzene and recycled to R2, which operates adiabatically as DEB reacts with benzene
to produce ethylbenzene. Benzene, in the D1 distillate, is recycled to R1. A mixture of EB

[email protected]
316 Decision Aid Models for Disaster Management and Emergencies

Fig. 14.4 Process Flow Sheet of Ethylbenzene Process.

and DEB in the D1 bottoms product is fed to D2, with EB recovered in the distillate, and
DEB recycled to R2 [12].
Chemical processes are characterized by many variables that allow representation of
the behaviour, using a set of rules. These factors allow us to use them appropriately for SA.
We use three types of rules: fact, intermediate and decision rules. The format of the rule is
as follows:

IF Antecedent; THEN Consequent

Inference direction can be generally divided into three types: (1) backward reasoning:
starts with the target, intending to prove the target to be true or false; (2) forward reason-
ing: starts from the fact, reasoning towards target; (3) mixed reasoning: reasoning in both
directions. In our work, the forward reasoning strategy is used.

14.4.2 Hazard Identification Implementation

Hazards are often obtained through the design and implementation phase, and various
models have been developed to identify them. For example, HAZOP is one of the most
powerful hazard identification methods available and has been clearly described in the re-

[email protected]
A Human Situation Awareness Support System to Avoid Technological Disasters 317

Table 14.2 Temperature Limits (◦ C).

Unit Operating value Six-sigma quality High alarm Automatic shutdown
R1 160 165 170 180
R2 166 170 175 185
D1 186 190 195 200
D2 200 205 210 220

search literature [30]. Fault tree, event tree, bow-tie and experts’ knowledge are adopted as
the knowledge acquisition techniques.
To show and store the hazardous situations we use fact rules. In a fact rule, ‘antecedent’
refers to conditions that have potential to harm, while ‘consequent’ is a name for the current
situation. For example:

IF TCR1 > 170 ◦ C; THEN the temperature of R1 is high

For the ethylbenzene process, hazardous situations include those due to controller fail-
ure, loss of cooling, disturbances in the feed temperatures and flow rates, the reboiler heat
duty, and flooding in the distillation columns. The safety systems are assigned temperature
limits, as shown in Table 14.2, including limits for the six-sigma quality (by definition,
when the controller fails to maintain the temperature within the six-sigma quality limit, the
controller “Fails”), high alarm and automatic shut-down. For each abnormal event, when
these limits are exceeded, time logs for the safety systems are recorded [12]. For example
Fig. 14.5 shows a bow-tie diagram for a high-temperature abnormal event associated with
reactor R1. The consequences include continued-operation (CO), shut-down (SD), release
(REL), and explosion (EXP), based on the performance of six safety systems shown in
rectangles across the left. The six safety systems are: S1 (high alarm), S2 (operator obser-
vation), S3 (operator correction), S4 (automatic shut-down), S5 (manual shut-down), and
S6 (emergency relief system).

[email protected]
318 Decision Aid Models for Disaster Management and Emergencies

EXP REL SD EXP REL SD SD CO

ERS

Manual

Automatic
shut down

Operator
correction

Operator
observation

Alarm

High temperature
in R1

Poor/No Tempera Low level

cooling ture in input
sensor
fail

Valve V
fails to
run

Pump Pump
1fails 2fails to
to run run

Fig. 14.5 Bow-tie Diagram for High Temperature in R1.

[email protected]
A Human Situation Awareness Support System to Avoid Technological Disasters 319

14.4.3 Situation Assessment Implementation

The situation assessment component will provide the comprehension and projection
levels of the SA. To support these levels, we use intermediate rules. In an intermediate rule,
‘antecedent’ is a situation, while ‘consequent’ is a description of causes and consequences.

14.4.3.1 Likelihood Determination

To determine the likelihood of a target event, the concept of physical reliability models
can be used. Reliability models aim to explain the reliability (or failure) of a component
as a multivariate function of operational physical parameters. Among different types of
physical reliability models, covariate models and static models can be used to estimate the
primary events and consequently the probability of the target event [24].
Operational physical parameters, also called covariates, may be temperature, velocity,
pressure, or vibration amplitude. Covariate models explain the failure rate of a component
as a function e.g.:
β

t
F(t) = 1 − Exp − n
Exp a i xi
i=0

where any change in covariate xi will change the failure probability F(t). Static models do
not consider time as an influential parameter and only consider the component’s strength
and stresses. Both stress and strength can be constant or considered as random variables
having probability distribution functions. For example, the failure probability of compo-
nent Q having a constant strength, k, and being under a random stress, Y , can be defined as
the probability of Y being greater than k, e.g.:
∞
Pr(Q) = Pr(Y > k) = fy (y) dy
kk
where fY (y) is the probability density function (PDF) of stress, Y . Assuming an exponential
distribution for fY (y), Pr(Q) can written as:
∞
Pr(Q) = λ e−λ y dy = e−λ k
kk
Therefore the failure probability of component Q can be reassessed when a new value
for k is observed.

[email protected]
320 Decision Aid Models for Disaster Management and Emergencies

Table 14.3 Consequence Severity Matrix.

Severity Monetary
Human loss Ass Loss Environment loss
class Value
Minorrepairs that can
Very Around the area, easy
<10k One minor injury be done immediately
little recovery
by own crew
Repairsthat take
One or two minor Within plant, short
Little 10-100k several days to carry
injury term remediation effort
out
Damagethat takes
Minor offsite impact,
100k- Multiple major months to repair and
Medium remediation cost will
1million injuries cause serious
be less than 1 million
consequences
Communityadvisory
One fatality or
1-10 Very large material issued, remediation
High multiple injuries
million damage cost remain below 5
with disabilities
million
Communityevacuation
Very Significant parts of the for longer period,
>10million Multiple fatalities
high system destroyed remediation cost in
excess of 5 million

14.4.3.2 Consequence Determination

Generally, consequences of an abnormal situation may be categorized into four groups;

asset loss, human fatality, environmental loss, and confidence or reputation loss. The sever-
ity matrix used in this study is outlined in Table 14.3 including equivalent dollar value of
damage associated with each consequence category based on the severity of damage [25].
The failure probability of each safety system can be determined by the probabilistic
method and end states are determined by multiplying the related probabilities together.
Previous probabilities are called “Priors”, representing our belief about the system be-
fore observing the new data. After the initiation of the process, accident precursor data
which are the near misses and incidents occurring in the process (defined as events that are
not characterized as accidents but indicate the increasing likelihood of an accident occur-
rence) can be collected from the system. The ASP data can be used to form the likelihood
function, which in turn updates the prior knowledge about the occurrence probability of
every end state resulting in the formation of the posterior function. Bayesian theory is a
probabilistic approach that applies the conditional probability principals to reason with un-
certainties. The results obtained by application of this theory in this study will yield the
“Posterior” which is the updated knowledge about the end states of the system. Considering

[email protected]
A Human Situation Awareness Support System to Avoid Technological Disasters 321

Table 14.4 Risk Matrix.

S.
Very little Little Medium High Very high
P.
Very likely Significant Significant High High High
Likely Medium Significant Significant High High
Even Low Medium Significant High High
Unlikely Low Low Medium Significant High
Very Unlikely Low Low Medium Significant Significant

x as the failure probability of the system and f (x)as the probability distribution function
(prior distribution), f (xData) will present the posterior distribution that is derived using
the following equation:

f (x | Data) ∝ g(Data | x) f (x)

where f (x | Data)is the posterior function, g(Data | x) is the likelihood function and f (x)
is the prior [26].

14.4.3.3 Current Risk Level

To obtain the risk level of the target hazard we use a fuzzy risk analysis model. We
present probability of hazards with five linguistic values, e.g. very likely, likely, even, un-
likely, and very unlikely, and will explain the severity by five linguistic values e.g. very
little, little, medium, high and very high. The risks are represented by low, medium, sig-
nificant and high. Triangular and trapezium membership functions can be used together to
increase the sensitivity in some bound points. Fig. 14.6 shows membership functions for
probability, severity and risk variables.
To construct the fuzzy risk analysis model, we considered the risk matrix as shown in
Table 14.4, which has 25 rules, e.g.:

IF the probability is likely AND the severity is medium

THEN the risk is high

14.4.3.4 Risk Reduction

If the estimated risk is unacceptable, it is necessary to identify risk-reducing measures

that reduce either the frequency or the consequences of the occurrence of the unwanted
event. For each individually considered risk element, a decision must be made whether or

[email protected]
322 Decision Aid Models for Disaster Management and Emergencies

not the investment in the risk-reducing measures has the effect of reducing the risk to an ac-
ceptable level [33]. A list of available reduction and containment options can be presented
as decision rules where ‘antecedent’ is a situation, while ‘consequent’ is suggested actions
to remove or eliminate the risk. Based on the operator’s decision, a new level of risk can
be calculated.

Fig. 14.6 Membership Functions.

[email protected]
A Human Situation Awareness Support System to Avoid Technological Disasters 323

Table 14.5 Examples of Knowledge Rules.

Rule 1 IF TR1 > 170 ◦ C Fact
THEN Hazard (H1R1) : high temperature in R1
Rule 2 IF TR2 > 175 ◦ C Fact
THEN Hazard (H1R2) : high temperature in R2
Rule 3 IF TD1 > 195 ◦ C Fact
THEN Hazard (H1D1): high temperature in D1
Rule 4 IF TD2 > 210 ◦ C Fact
THEN Hazard (H1D2): high temperature in D2
...
Rule 21 IF (H1R1) Intermediate
THEN Poor cooling or TCR1 fail or input low level
Rule 22 IF (H1R2)
THEN ...
...
Rule 40 IF (H1R1) Decision
THEN switch to redundancy pump in cooling system and administrative checks
...

Assume the temperature in Reactor1 is increased to 170 ◦ C. The system is initialized

an abnormal situation occurred; the results are sent to the inference machine and stored in
the integrated database. Rule 1 is selected and returned according to the knowledge rules.
The system reports that the hazard (H1R1) occurred and an alarm will be shown on the
operator’s interface. At the same time, H1R1 characteristics are recalled from the database.
The posterior probability is calculated using Bayesian inference. For this example the
abnormal event occurred at interval 20 and the posterior probability is 0.0132. Accord-
ing to the fuzzy risk analysis model the current risk level is 0.65 and the system presents
“significant” risk level on the GUI. The causes of hazardous situation are searched by the
inference machine and Rule 21 is written to the cause’s area of GUI, and Rule 40 is se-
lected. The operating suggestions are displayed in the monitoring windows of the system.
Usually, during a short time period when multiple alarms occur, it is not possible to remove
all of them. One has to attribute priorities, and our approach has this ability.

[email protected]
324 Decision Aid Models for Disaster Management and Emergencies

14.5 Conclusion and Future Study

During the operation of complex systems that include human decision making, ac-
quiring and interpreting information from the environment forms the basis for the state of
knowledge of a decision maker. This state is often referred to as situation awareness (SA).
Lacking or inadequate SA has been identified as one of the primary factors in accidents
attributed to human error and it is especially important in work domains where the infor-
mation flow can be quite high, and poor decisions may lead to serious consequences. As
technological systems continue to increase in degree of automation and complexity, the task
of providing actionable information for SA becomes more difficult and costly to achieve.
In this study we proposed a new system to support SA for the safety operators. Initially, our
system conducts the complicated task of understanding what is going on in the situation,
and it then assesses the current situation by the risk analysis concept through a case study.
The enhancement of SA is a major design goal for developers of operator interfaces,
automation concepts, and training programs in a verity of fields. To evaluate the degree to
which new technologies or design concepts actually improve operator SA, it is necessary to
systematically evaluate them based on a measure of SA, which can determine those ideas
that have merit and those that have unforeseen negative consequences [8]. Therefore, de-
veloping a SA measuring method and a system prototype evaluation based on the proposed
SA measurement will form the basis for future study in this work.

Bibliography

[1] Shrivastava, P., Bhopal: anatomy of a crisis. 1992: P. Chapman Pub.

[2] Endsley, M.R., Situation Awareness, in Handbook of Human Factors and Ergonomics, G. Sal-
vendy, Editor. 2006, John Wiley and Sons Press p. 528–542.
[3] Endsley, M.R. and E.O. Kiris, The out-of-the-loop performance problem and level of control in
automation. Human Factors: The Journal of the Human Factors and Ergonomics Society, 1995.
37 (2): p. 381–394.
[4] Burmester, M., T. Komischke, and L. Wust, Innovative User Interfaces in Automation En-
gineering by Application of Usability Engineering Methods Shown by the Example of a
Three-Dimensional Plant Representation. International Journal of Human-Computer Interac-
tion, 2000. 12 (3-4): p. 359–373.
[5] Endsley, M.R. and D.J. Garland, Situation awareness: analysis and measurement. 2000:
Lawrence Erlbaum Associates.
[6] Vincenzi, D.A., M. Mouloua, and P.A. Hancock, Human performance, situation awareness and
automation: current research and trends : HPSAA II. 2004: L. Erlbaum Associates.
[7] Pridmore, J.L., Designing for the improvement of operator situation awareness in automation
systems. 2007, Auburn University: United States – Alabama.

[email protected]
Bibliography 325

[8] Endsley, M., Measurement of Situation Awareness in Dynamic Systems. Human Factors: The
Journal of the Human Factors and Ergonomics Society, 1995: p. 65–84.
[9] Garland, D.J., J.A. Wise, and V.D. Hopkin, Handbook of aviation human factors. 1999: Erl-
baum.
[10] Endsley, M.R., Toward a theory of situation awareness in dynamic systems. Human Factors,
1995. 37 (1): p. 32–64.
[11] Endsley, M.R. and E.S. Connors. Situation awareness: State of the art. in IEEE Power and
Energy Society 2008 General Meeting: Conversion and Delivery of Electrical Energy in the
21st Century, PES. 2008.
[12] Stanton, N.A., P.R.G. Chambers, and J. Piggott, Situational awareness and safety. Safety Sci-
ence, 2001. 39 (3): p. 189–204.
[13] Banbury, S. and S. Tremblay, A Cognitive approach to situation awareness: theory and appli-
cation. 2004: Ashgate Pub.
[14] Lu, J., Liu, B., Zhang, G., Hao, Z., Xiao, Y., A situation assessment approach using support
vector machines as a learning tool. International Journal of Nuclear Knowledge Management,
2008. 3 (1): p. 82–97.
[15] Kim, M.C. and P.H. Seong, An analytic model for situation assessment of nuclear power plant
operators based on Bayesian inference. Reliability Engineering & System Safety, 2006. 91 (3):
p. 270–282.
[16] Lu, J., G. Zhang, and D. Ruan, Intelligent multi-criteria fuzzy group decision-making for situ-
ation assessments. Soft Computing, 2008. 12 (3): p. 289–299.
[17] Brannon, N.G., et al., Coordinated machine learning and decision support for situation aware-
ness. Neural Networks, 2009. 22 (3): p. 316–325.
[18] Kim, M.C. and P.H. Seong, A computational method for probabilistic safety assessment of I&
C systems and human operators in nuclear power plants. Reliability Engineering & System
Safety, 2006. 91 (5): p. 580–593.
[19] Jones, R.E.T., E.S. Connors, and M.R. Endsley. Incorporating the human analyst into the data
fusion process by modeling situation awareness using fuzzy cognitive maps. in 2009 12th In-
ternational Conference on Information Fusion, 2009.
[20] Chandana, S. and H. Leung. Distributed situation assessment for intelligent sensor networks
based on hierarchical fuzzy cognitive maps. in IEEE International Conference on Systems,
Man and Cybernetics. 2008.
[21] Lu, J., G. Zhang, and F. Wu, Team situation awareness using web-based fuzzy group decision
Support systems. International Journal of Computational Intelligence Systems, 2008. 1 (1): p.
50–59.
[22] Jones, R., et al., Using fuzzy cognitive mapping techniques to model situation awareness for
army infantry platoon leaders. Computational & Mathematical Organization Theory, 2011. 17
(3): p. 272–295.
[23] Naderpour, M., J. Lu, and E. Kerre. A Conceptual Model for Risk-Based Situation Awareness.
in International Conference on Intelligent Systems and Knowledge Engineering (ISKE). 2011.
Shanghai, China: Springer Berlin / Heidelberg.
[24] Khakzad, N., F. Khan, and P. Amyotte, Dynamic risk analysis using bow-tie approach. Relia-
bility Engineering & System Safety, 2012. 104 (0): p. 36–44.
[25] Naderpour, M. and J. Lu, A Fuzzy Dual Expert System for Managing Situation Awareness
in a Safety Supervisory System, in IEEE International Conference on Fuzzy Systems. 2012:
Brisbane-Australia. p. 715–721.
[26] Kalantarnia, M., F. Khan, and K. Hawboldt, Dynamic risk assessment using failure assessment
and Bayesian theory. Journal of Loss Prevention in the Process Industries, 2009. 22 (5): p.
600–606.

[email protected]