Lecture Notes in Economics

and Mathematical Systems 394

Founding Editors:
M. Beckmann
H. P. Klinzi

Editorial Board:
H. Albach, M. Beckmann, P. Dhrymes, O. Feichtinger, W. Hildenbrand
W. Krelle, H. P. Klinzi, K. Ritter, U. Schittko, R. Selten

Managing Editors:
Prof. Dr. O. Fandel
Fachbereich Wirtschaftswissenschaften
Femuniversitat Hagen
Feithstr. 140/AVZ II, W-5800 Hagen I, FRO
Prof. Dr. W. Trockel
Institut flir Mathematische Wirtschaftsforschung (IMW)
Universitat Bielefeld
Universitatsstr. 25, W-4800 Bielefeld I, FRO
Young-Jou Lai Ching-Lai Hwang

Fuzzy Mathematical
Programming
Methods and Applications

Springer-Verlag
Berlin Heidelberg New York
London Paris Tokyo
Hong Kong Barcelona
Budapest
Authors
Prof. Young-Jou Lai
Prof. Ching-Lai Hwang
Department of Industrial Engineering
Kansas State University
Manhattan, KS 66506, USA

ISBN-13: 978-3-540-56098-2 e-ISBN-13: 978-3-642-48753-8

DOl: 10.1007/978-3-642-48753-8

This work is subject to copyright. All rights are reserved, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, re-use of
illustrations, recitation, broadcasting, reproduction on microfilms or in any other way,
and storage in data banks. Duplication of this publication or parts thereof is permitted
only under the provisions of the German Copyright Law of September 9, 1965, in its
current version, and permission for use must always be obtained from Springer-Verlag.
Violations are liable for prosecution under the German Copyright Law.
© Springer-Verlag Berlin Heidelberg 1992

Typesetting: Camera ready by author/editor

42/3140-543210 - Printed on acid-free paper
 PREFACE

This monograph is intended for an advanced undergraduate or graduate course as well as

for researchers who want a compilation of developments in the rapidly growing fields of
operations research, management science, economics, banking/finance, engineering and computer
science. This is a sequel of our previous works entitled Multiple Objective Decision Making -
Methods and Applications, Multiple Attribute Decision Making - Metlwds and Applications,
Group Decision Making Under Multiple Criteria - Metlwds and Applications, and Fuzzy Multiple
Attribute Decision Making - Methods and Applications (respectively, Nos. 164, 186, 281 and 375
of the Lecture Notes).
The literature on methods and applications of fuzzy mathematical programming (FMP) has
been systematically and thoroughly reviewed and classified. This state-of-the-art survey provides
readers with a capsule look into the existing methods, and their characteristics and applicability
to analysis of FMP problems.
In the last 25 years, the fuzzy set theory has been applied in many disciplines such as
operations research, management science, control theory, artificial intelligence/expert system, etc.
The growth of applications of the fuzzy set theory has been accumulating. This study presents
systematically the state-of-the-art of fuzzy mathematical programming in both techniques and
applications. In addition to fuzzy mathematical programming, we also present all existing
possibilistic mathematical programming problems, methods and applications, which are among
the latest important topics.
We will present our material in two volumes: the current volume covers single-objective
fuzzy and possibilistic mathematical programming; and the second includes fuzzy multiple
objective decision-making and possibilistic multiple objective decision-making problems.
To realize practical fuzzy modelling, we present important real-world problems including
production/manufacturing, transportation, assignment, game, environmental management, water
resources management, nutrition management, personnel management, accounting, banking/
finance, marketing management, trade balance, and agricultural economics. It should be noted
that all concepts and methods are developed for solving practical problems.
For this volume, we first systematically and thoroughly define and classify all existing
fuzzy and possibilistic (single-objective) mathematical programming problems and solution
techniques in Chapter 1.
In Chapter 2, basic fuzzy set theories, membership functions, fuzzy decisions, operators
and fuzzy arithmetic are introduced. We present them with simple numerical examples in an
easy-ta-read and easy-ta-follow manner.
Basic concepts, computational procedures and characteristics of all existing fuzzy
mathematical programming are concisely presented in Chapter 3. The computational procedure
of each approach is illustrated by solving a practical problem or a simple numerical example.
Six problems and eleven approaches of fuzzy linear programming are discussed. Fuzzy non-
linear programming and fuzzy integer programming are also discussed in this chapter. We also
present our expert decision-making support system IFLP which improves flexibility and
robustness of linear programming techniques by considering and solving all possibilities of a
specific domain of linear programming problems.
The practical applications discussed in this chapter include production/manufacturing,
transportation, environmental management and agriculture economics problems under fuzzy
environments.
Since the 1980s, the possibility theory has become more and more important in the fields
of operations research, management science, expert systems, etc. In Chapter 4, we concisely
present primary concepts, computational procedures and characteristics of all existing possibilistic
mathematical programming. The computational procedure of each approach is illustrated by
solving a practical problem or a simple numerical example. Seven problems and fourteen
approaches of possibilistic linear programming are discussed.
The practical applications discussed in this chapter include project investment, game,
banking and finance problems under imprecise environments.
In Chapter 5, brief discussions of probability theory versus fuzzy set theory, and stochastic
versus fuzzy mathematical programming are given. Concluding remarks and future studies are
also presented.
An updated bibliographical listing of 62 books, monographs or conference proceedings,
and about 300 selected papers, reports or theses is presented in the end of this study.
We would like to thank all the researchers in this field, especially all scholars listed in
Bibliography. The scope of this study could not have been completed without many excellent
researchers carrying out and publishing their research. We would like to especially thank
Professor L. A. Zadeh for his creation of fuzzy set theory; Professors H.-J. Zimmermann, C.
 VII

V. Negoita, H. Tanaka, K. Asai, S. Lee and P.L. Yll for their momentous contributions in
operations research under fuzziness; Professors D. Dubois and H. Prade for their material
contributions in fuzzy set theory and possibility theory; and Professors A. Kaufmann and M. M.
Gupta for their important contributions in fuzzy arithmetic.
Finally, we thank our wives for their patience while we devoted the many hours it takes
to see a project of this nature to completion. We also thank Ms. lochun Wu and Ms. M. Rankin
for editing this monograph.
This study was partly supported by HTX International, Inc.

Y. 1. Lai C. L. Hwang
Kansas State University Kansas State University
Manhattan, Kansas Manhattan, Kansas
TABLE OF CONTENTS

PREFACE v

1 INTRODUCTION 1

1.1 Objectives of This study 1

1.2 Fuzzy Mathematical Programming Problems 2

1.3 Classification of Fuzzy Mathematical Programming 5

1.4 Applications of Fuzzy Mathematical Programming 8

1.5 Literature Survey 10

2 FUZZY SET THEORY 14

2.1 Fuzzy Sets 14

2.2 Fuzzy Set Theory 19

2.2.1 Basic Terminology and Definition 20

2.2.1.1 Definition of Fuzzy Sets 20

2 . 2 . 1. 2 Support 21

2.2.1.3 a-level Set 21

2.2.1.4 Normality 22

2.2.1.5 Convexity and Concavity 22

2.2.1.6 Extension Principle 23

2.2.1.7 Compatibility of Extension Principle

with a-cuts 23
2.2.1.8 Relation 24

2.2.1.9 Decomposability 24

2.2.1.10 Decomposition Theorem 25

2.2.1.11 Probability of Fuzzy Events 25

2.2.1.12 Conditional Fuzzy Sets 27

2.2.2 Basic Operations 27

2.2.2.1 Inclusion 27

2.2.2.2 Equality 27

2.2.2.3 Complementation 28
 x

2.2.2.4 Intersection .28

2.2.2.5 Union 29

2.2.2.6 Algebraic Product 29

2.2.2.7 Algebraic Sum 29

2.2.2.8 Difference 29

2.3 Membership Functions 30

2.3.1 A Survey of Functional Forms 31

2.3.2 Examples to Generate Membership Functions 34
2.3.2.1 Distance Approach 35
2.3.2.2 True-Valued Approach 40

2.3.2.3 Payoff Function 45

2.3.2.4 Other Examples 48

2.4 Fuzzy Decision and Operators 48

2.4.1 Fuzzy Decision 49

2.4.2 Max-Min Operator 51

2.4.3 Compensatory Operators 53
2.4.3.1 Numerical Example for Operators 56
2.5 Fuzzy Arithmetic 58
2.5.1 Addition of Fuzzy Numbers 58
2.5.2 Subtraction of Fuzzy Numbers 61

2.5.3 Multiplication of Fuzzy Numbers 63

2.5.4 Division of Fuzzy Numbers 65
2.5.5 Triangular and Trapezoid Fuzzy Numbers 66
2.6 Fuzzy Ranking 72

3 FUZZY MATHEMATICAL PROGRAMMING 74

3.1 Fuzzy Linear Programming Models 75
3.1.1 Linear Programming Problem with Fuzzy Resources 75

3.1.1.1 Verdegay's Approach 79

3.1.1.1a Example 1: The Knox Production-Mix

Selection Problem 80

3.1.1.1b Example 2: A Transportation Problem 85

 3.1.1.2 Werners's Approach 87
3.1.1.2a Example 1: The Knox Production-Mix
Selection Problem 89
3.1.1.2b Example 2: An Air Pollution Regulation
Problem 90
3.1.2 Linear Programming Problem with Fuzzy Resources and
Objective 95
3.1.2.1 Zimmermann's Approach 95
3.1.2.1a Example 1: The Knox Production-Mix
Selection Problem 97
3.1.2.1b Example 2: A Regional Resource Allocation
Problem 98
3.1.2.1c Example 3: A Fuzzy Resource Allocation
Problem 101
3.1.2.2 Chanas's Approach 104
3.1.2.2a Example 1: An optimal System Design
Problem 107
3.1.2.2b Example 2: An Aggregate Production
Planning Problem 110
3.1.3 Linear Programming Problem with Fuzzy Parameters
in the Objective Function 119
3.1.4 Linear Programming with All Fuzzy Coefficients 121
3.1.4.1 Example: A Production Scheduling Problem 123
3.2 Interactive Fuzzy Linear Programming 127
3.2.1 Introduction 127
3.2.2 Discussion of zimmermann's, Werners's
Chanas's and Verdegay's Approaches 128
3.2.3 Interactive Fuzzy Linear Programming - I 130
3.2.3.1 Problem Setting 130
3.2.3.2 The Algorithm of IFLP-I 136
3.2.3.3 Example: The Knox Production-Mix
 XII

Selection Problem 139

3.2.4 Interactive Fuzzy Linear Programming - II 146
3.2.4.1 The Algorithm of IFLP-II 151
3.3 Some Extensions of Fuzzy Linear Programming Problems 155
3.3.1 Membership Functions 155
3.3.1.1 Example: A Truck Fleet Problem 161
3.3.2 Operators 168
3.3.3 sensitivity Analysis and Dual Theory 170
3.3.4 Fuzzy Non-Linear Programming 170
3.3.4.1 Example: A Fuzzy Machining Economics
Problem 171
3.3.5 Fuzzy Integer programming 174
3.3.5.1 Fuzzy 0-1 Linear Programming 176
3.3.5.1a Example: A Fuzzy Location Problem 180

4 POSSIBILISTIC PROGRAMMING 187

4.1 Possibilistic Linear Programming Models 189
4.1.1 Linear Programming with Imprecise Resources and
Technological Coefficients 189
4.1.1.1 Ramik and Rimanek's Approach 189
4.1.1.1a Example: A Profit Apportionment Problem 191
4.1.1.2 Tanaka, Ichihashi and Asai's Approach 196
4.1.1.3 Dubois's Approach 198
4.1.2 Linear Programming with Imprecise Objective
Coefficients 203
4.1.2.1 Lai and Hwang's Approach 203
4.1.2.1a Example: A Winston-Salem Development
Management Problem 208
4.1.2.2 Rommelfanger, Hanuscheck and Wolf's
Approach 213
4.1.2.3 Delgado, Verdegay and Vila's Approach 217
4.1.3 Linear Programming with Imprecise Objective and
 XIII

Technological Coefficients 223

4.1.4 Linear Programming with Imprecise Coefficients 225

4.1.4.1 Lai and Hwang's Approach 225
4.1.4.2 Buckley's Approach 227
4.1.4.2a Example: A Feed Mix (Diet) Problem 230
4.1.4.3 Negi's Approach 232
4.1.4.4 Fuller's Approach 236
4.1.5 Other Problems 237
4.2 Some Extensions of Possibilistic Linear Programming 238
4.2.1 Linear Programming with Imprecise Coefficients and
Fuzzy Inequalities 238
4.2.1a Example: A Fuzzy Matrix Game Problem 241
4.2.2 Linear Programming with Imprecise Objective
Coefficients and Fuzzy Resources 243
4.2.2a Example: A Bank Hedging Decision Problem 247
4.2.3 Stochastic Possibilistic Linear Programming 256
4.2.3a Example: A Bank Hedging Decision Problem 260

5 CONCLUDING REMARKS 268

5.1 Probability Theory versus Fuzzy Set Theory 268
5.2 Stochastic versus Possibilistic Programming 270
5.3 Future Research 275
5.4 Introduction of the Following Volume 276
5.5 Fuzzy Multiple Attribute Decision Making 277

BIBLIOGRAPHY 279
Books, Monographs and Conference Proceedings 279
Journal Articles, Technical Reports and Theses 283
1 INTRODUCTION

At the tum of the century, reducing complex real-world systems into precise mathematical
models was the main trend in science and engineering. In the middle of this century, Operations
Research (OR) began to be applied to real-world decision-making problems and thus became one
of the most important fields in science and engineering. Unfortunately, real-world situations are
often not so deterministic. Thus precise mathematical models are not enough to tackle all
practical problems.
To deal with imprecision/uncertainty, the concepts and techniques of probability theory
are usually employed. In the 1960s, meanings of the probability theory has been reconsidered
and criticized when modelling practical problems, especially in artificial intelligence. Around
the same time as the development of chaos theory to handle non-linear dynamic systems in
physics and mathematics, fuzzy set theory was developed in 1965 by L.A. Zadeh. Since then,
fuzzy set theory has been applied to the fields of operations research, management science,
artificial intelligence/expert system, control theory, statistics and many other fields.
Fuzzy set theory is a theory of graded concepts (a matter of degree), but not a theory of
chance. Therefore, figures and numerical tables are considered paramount in the study of fuzzy
set theory. They, unlike confusing mathematical slang, difficult functions, etc., provide the best
ways to communicate with outsiders. This is an important concept in the new generation of
operations research.
In operations research, fuzzy set theory has been applied to techniques of linear and non-
linear programming, dynamic programming, queueing theory, multiple criteria decision-making,
group decision-making and so on. In this monograph, we provide readers with a capsule look
into all existing fuzzy linear and nonlinear mathematical programming problems, methods, and
applications.

1.1 Objectives of This Study

This study is an introduction to the applications of fuzzy set theory toward Fuzzy
Mathematical Programming. It gives a state-of-the-art survey of all existing problems, methods
and practical applications concerning fuzzy mathematical programming. Many diversified
methods and applications are thoroughly and critically reviewed and systematically classified.
Some basic concepts and terminologies are defined so that we can explain the literature in a
 2

consistent manner. This study thus offers its readers a capsule look into the existing methods,
and their characteristics and applicability to analysis of fuzzy mathematical programming
problems. It is the first time these problems, methods and applications have been presented
together with both deep and broad discussions.
Practical applications are never over-emphasized. All concepts and techniques are
developed for solving real-world problems. Thus the most important issue is to involve real
applications with real decision makers. Unfortunately, these real applications have very rarely
been reported in the literature. To fill the gap between researchers and real decision makers for
solving real applications, discussions of practical problems are considered as important as
concepts and techniques in this study. Practical problems discussed here include
production/manufacturing, transportation, assignment, game, environmental management,
banking/finance and agricultural economics. We hope that through the discussions of these
practical problems, real decision makers will obtain more confidence in the usefulness of fuzzy
modelling.

1.2 Fuzzy Mathematical Programming Problems

While L. V. Kantorovich, a Soviet mathematician and economist, formulated and solved linear
programming dealing with the organization and planning of production in 1939, G. B. Dantzig,
M. Wood and their associates of the U. S. Department of the Air Force developed the general
problem of linear programming named by A. W. Tucker. In order to apply mathematical and
related techniques to military programming and planning problems, Dantzig proposed "that
interrelations between activities of a large organization be viewed as a linear programming type
model and the optimizing program determined by minimizing a linear objective function."
Conceptually, some concepts originated from Leotief's input-output model. Since that time,
linear programming has become an important tool of modem theoretical and applied mathematics.
Linear programming is concerned with the efficient allocation of limited resources to
activities with the objective of meeting a desired goal such as maximizing profit or minimizing
cost. The distinct characteristic of linear programming models is that the interrelations between
activities are linear relationships which are the satisfactions of the proportionality and additivity
requirements. Symbolically, the general linear programming problem may be stated as:

maximize z = f(c,x) = cx
subject to g(A,x) = Ax ~ b and x ~ 0 (1.1)
 3

where c is the vector of profit coefficients of the objective function and b is the vector of total
resources available. x is the vector of decision variables (or alternatives), and A is the matrix
of technical coefficients. These input data (of c, b and A) are usually fuzzy/imprecise because
of incomplete or non-obtainable information. For instance, available labor hours and available
material (b) may be "mmnd, 1220" hours and "ibwt 1550" units, respectively. Similarly, the
unit profit (c) of new products may be expected to be "ibwt $2.3" per unit and the estimates of
technological coefficients (A) may be "m:mmd 3" units per labor hour.
To formulate these fuzzy/imprecise numbers, we can use membership functions or
possibility distribution (depending on specific problems). The functional forms of membership
functions and possibility distributions are depicted in Figures 1.1 and 1.2, respectively. With
these fuzzy/imprecise input data, Equation (1.1) is then called fuzzy/possibilistic (linear)
programming. In this study, the grade of a membership function indicates a subjective degree
of satisfaction within given tolerances. On the other hand, the grade of possibility indicates the
subjective or objective degree of the occurrence of an event. It is important to realize this
distinction while modelling fuzziness/imprecision in mathematical programming problems.
In Figure 1.1, we illustrate four cases of the preference-based membership functions of
fuzzy resources. When the constraints are Ax ::s; ii, the rational preference-based membership
functions can be assumed to be non-increasing. Similarly, non-decreasing functions can be
assumed for Ax ~ b. For equality constraints, triangular or trapezoid functions might be
appropriate. For the maximization (or minimization) problem, the preference-based membership
c
functions of can be assumed to be non-decreasing as b in Ax ~ b or (or non-increasing as b
in Ax ::s; b). As to the preference-based membership functions of A, they may be either non-
increasing for Ax ::s; b, or non-decreasing for Ax ~ b. Sometimes, triangular or trapezoid
functions might be adopted. For possibilistic Oinear) programming, the possibility distributions
of A, band/or care often assumed to be triangular or trapezoid functions as shown in Figure
1.2.
It is noted that non-linear functions, such as piece-wise linear, exponential, power,
hyperbolic,etc., may be adopted as indicated in Figures 1.1 and 1.2.
When any off(c,x) and/or g(A,x) is a non-linear function, Equation (1.1) becomes a non-
linear programming problem. If x is restricted to be an integer, then Equation (1.1) will become
art integer programming problem. Both cases with fuzzy/imprecise input data as shown in
Figures 1.1 and 1.2, Equation (1.1) becomes fuzzy (possibilistic) non-linear and integer
programming, respectively.
 4

IJ
1 +-----".. IJ
1

o b/A
o b/A
Ax S b
Ax ~ b

IJ IJ
1 1

o z/c o z/c
min z max z

Jl Jl
1 1

o b/A o b/A
Ax b Ax b

Fiqure 1.1 Preference-based membership functions of fuzzy b, fuzzy A,

fuzzy c or fuzzy z

71' 71'
1 1

o AlbIc 0 AlbIc

Fiqure 1.2 Possibility distributions of imprecise A, imprecise b or

imprecise c
 5

1.3 Classification of Fuzzy Mathematical Programming

The most recent literature surveys of fuzzy mathematical programming are Zimmermann
[304,306, 1985; BM64, 1987], Slowinski [227, 1986], Wierzchon [264, 1987], Leung [BM38,
1988] and Luhandjula [161, 1989]. Zimmermann first classified fuzzy mathematical
programming into symmetrical and nonsymmetrical models. Leung classified fuzzy mathematical
programming problems into the following four categories: (1) a precise objective and fuzzy
constraints; (2) a fuzzy objective and precise constraints; (3) a fuzzy objective and fuzzy
constraints; (4) robust programming (which is considered as one of the possibilistic mathematical
programming problems in this study). Luhandjula categorized fuzzy mathematical programming
into flexible programming, mathematical programming with fuzzy parameters and fuzzy
stochastic programming. Fuzzy parameters or so-called fuzzy numbers are essentially
characterized by possibility distributions. For methodologies, he further classified flexible
programming methods into symmetrical and asymmetrical approaches as Zimmermann did. For
mathematical programming problems with fuzzy parameters, Luhandjula grouped them into two
major classes: problems with a deterministic objective function and problems with a fuzzy
objective function. Fuzzy stochastic programming concerns the parameters involving both fuzzy
and stochastic natures.
In this study, we systematically classify all possible problems and existing approaches after
an extensive and thorough literature survey. We distinguish fuzzy linear programming problems
from possibilistic linear programming problems. Fuzzy linear programming problems will
associate fuzzy input data which should be modelled by subjective preference-based membership
functions. On the other hand, possibilistic linear programming problems will associate with
imprecise data which should be modelled by possibility distributions. The possibility distributions
are an analogue of the probability distributions and can be either subjective or objective.
Furthermore, based on the possible combinations of crisp and/or imprecise coefficients (input
data), we also sub-classify fuzzy linear programming and possibilistic linear programming
problems as shown in Figures 1.3 and 1.4. Obviously, these classifications are problem-oriented.

Based on possible combinations of the fuzziness of A, b, c and/or z, we classify fuzzy

(linear) programming into the following five major problems: (I)~; (2) fuzzy b and fuzzy
z;; (3) ~; (4) fuzz;y A, fuzzy b and fuzzy c, fuzz;y A and fuzz;y b, fuzzy A and fuzz;y c, or
fuzzy A. fuzzy b and fuzzy c; (5) fuzz;y A and fuzz;y z, or fuzzy A. fuzz;y band fuzz;y z.
Moreover, we include the sixth problem of "an expert decision-making support system" which
 6

max/min z = cx Fuzzy z Fuzzy c

s.t. Ax {:s,=,~} b Fuzzy b or {:s,=,~} Fuzzy A

Problem Approach

1
r1FuZZY b ~ 3.1.1.1
3.1.1.1
Parametric programming
Verdegay's [250]
3.1.1.2 Werners's [257]
2
HFuZZY b & z
~ 3.1.2.1
3.1.2.2
Zimmermann's [296]
Chanas's [34]
3.2.3 Lai & Hwang's IFLP-I [137]
3
-1FUZZY c 3.1.3 Parametric programming
3.1.3 Verdegay's [251]
Fuzzy
linear I-
programming 4 Fuzzy A; b & c; H3.1.4 Carlsson & Korhonen's [31]
- A & b; A & c;
or A, b & c
5
lFuZZY Z & A; 3.2.4 Lai & Hwang's [141]
or z, A & b
6
An expert DM 3.2.4.1 Lai & Hwang's IFLP-II[141]
support system

Piqure 1.3 A taxonomy of fuzzy linear programming (Lai and Hwang [141])
 7

max/min z = cx Imprecise c
s.t. Ax {S,=,ee} b Imprecise b Imprecise A

Problem Approach

1
r1Imprecise A & b j - 4.1.1.1 Ramik & Rimanek's [210]
4.1.1.2 Tanaka et al. [240]
4.1.1.3 Dubois's [72]
2
~Imprecise b or cj- 4.1.2.1 Lai and Hwanq's [138]
4.1.2.2 Rommelfanqer et al.[216]
4.1.2.3 Delqado et al. [56]
3
Possibil-
istic
linear
proqramminq
r--
~Imprecise A,
b &c
I 4.1.4.1
4.1.4.2
4.1.4.3
4.1.4.4
Lai & Hwanq's [138]
Buckley's [24]
Neqi's [175]
Fuller's [97]
4
HImprecise A; 4.1.3 Lai & Hwanq's [138]
A & c; or b & c
5
~Imprecise A & b, 4.2.1 Delqado et al. [55]
and fuzzy S (ee)
6
nImprecise c and 4.2.2 Lai & Hwanq's [140]
fuzzy b

7 Stochastic & H4.2.3 Lai & Hwanq's [142]

- imprecise c;
(and fuzzy b)

Piqure 1.4 A taxonomy of possibilistic linear proqramminq (Lai [139])

 8

systematically integrates all possible problems listed above. Approaches to solve each fuzzy
(linear) programming problem are listed.
Similarly, we group possibilistic (linear) programming problems into the following four
major problems: (1) imprecise A and inmrecise b; (2) imprecise b, or imprecise c; (3) inmrecise
A, imprecise b and imprecise c; (4) imprecise A, imprecise A and imprecise c, or imprecise b
and imprecise c. Occasionally, some coefficients have preference-based membership functions
and others have possibilistic distributions. Here, we classify this problem into the category of
possibilistic linear programming as shown in Figure 1.4. The fifth and sixth possibilistic (linear)
programming problems are (5) imprecise A. imprecise b and fuzzy inequalities ($ or C:l, and
(6) imprecise c and fuzzy b, respectively. When there are coefficients involving both fuzziness
and randomness simultaneously, we have a stochastic possibilistic (linear) programming problem.
In this case, we also classify it into the seventh possibilistic programming problem as shown in
Figure 1.4. That is: (7) stochastic and imprecise c. and/or fuz~ b. The reason for this
arrangement is for ease of our presentation in this study. Of course, this classification is rather
SUbjective. Approaches to solve each problem are listed.
Finally, it is noted that for a non-linear programming problem or an integer programming
problem, the concepts and approaches as shown Figures 1.3 and 1.4 may also be applied.

1.4 Applications of Fuzzy Mathematical Programming

As indicated by White [262], mathematical programming has been applied to many
disciplines such as advertising, assignment, bacteria identification, blending, blood control,
budgeting, computer selection, credit, diamond dressing, diet, disease control, drilling,
employment, engineering design, finance, income tax, information services, investment, location,
maintenance, managerial compensation, manpower, marketing, materials handling, media,
medicare, mergers, metal cutting, networks, organizational design, perishable products, personnel
assignment, pollution, production, promotions, projection selection, portfolio analysis,
purchasing, quality control, racial balance, reliability, replacement, research and development,
safety, salaries, sales, security scheduling, stock control, tendering, timetabling, transportation
and visitor quotas. However, in these real-world problems, the input data or parameters are
often fuzzy/imprecise because of incomplete or non-obtainable information. Traditional
mathematical programming techniques, obviously, cannot solve all fuzzy programming problems.
The drive to have practical models for management systems under fuzziness has been the
impetus towards the development of many excellent methods for fuzzy mathematical
 9

~abl. 1.1 List of practical problems

Problem Fuzziness/ Approach Section

Imprecision
I. PRODUCTION/MAHUFACTURINGI

Production-Mix Fuzzy Resources Verdegay's 3.1.1.1a

Selection Werners's 3.1.1.2a
Fuzzy Resources
& Fuzzy Objective Zimmermann's 3.1. 2 .1a
Lai & Hwang's 3.2.3.3
Optimal System Design Fuzzy Resources
& Fuzzy Objective Chanas's 3.1.2.2a
Aggregate Production Fuzzy Resources
Planning & Fuzzy Objective Chanas's 3.l.2.2b
Production Scheduling Fuzzy Coefficients Carlsson &
Korhonen's 3.1.4.1
Machining Economics Fuzzy Resources
& Fuzzy Objective Zimmermann's 3.3.4.1
II. ~SPOR~A~ION:

Transportation Fuzzy Resources Verdegay's 3.1.1.1b

Truck Fleet Fuzzy Resources Nakamura's 3.3.1.1
III. ASSIGNMENT:

Network Location Fuzzy Resources Zimmermann & 3.3.5.1a

Pollatschek's
IV. GAMB:
Matrix Game Impr. Coeffs. &
Fuzzy Inequalities Campos's 4~2.1a

V. ERVIRONMENTAL IlANAGBMEIIT:

Air Pollution Fuzzy Resources Werners's 3.1.1.2b

Regulation
VI. BANKING/FINANCE:
Profit Apportionment Impr. Resources & Ramik &
Impr. Tech. Coeffs. Rimanek's 4.1.1.la
Project Investment Impr. Coeffs. Lai & Hwang's 4.1.2.1
Bank Hedging Decision Impr. Objective
Coeffs. & Fuzzy
Resources Lai & Hwang's 4.2.2a
Impr. & Stochastic
Objective Coeffs.,
& Fuzzy Resources Lai & Hwang's 4.2.3a
VII. AGRICUL~ BCONOMICS:

Regional Resource Fuzzy Resources

Allocation & Fuzzy Objective Zimmermann's 3.1.2.1b
Feed Mix Impr. Coeffs. Buckley'S 4.1.4.2a
Note: Impr. --> Imprecise; Coeffs. --> Coefficients.
 10

programming problems. However, few methods have been yet tested with real-world problems
by real decision makers. To fill the gap between researchers and real decision makers for
solving real applications, discussions of practical problems are considered as important as
concepts and techniques in this study.
The practical fuzzy/possibilistic problems discussed in this monograph include the
following seven fields: (I) production/manufacturing, (ll) transportation, (Ill) assignment, (IV)
game, (V) environmental management, (VI) banking/finance and (Vll) agricultural economics (as
shown in Table 1.1).
In Table 1.1, the first column indicates the types of problems studied in this volume. The
assumed fuzziness and/or imprecision are listed in the second column and approaches to solve
each problem are listed in the third column. Finally, the sections where these applications are
discussed are indicated in the last column.
It is our belief that this concise presentation of practical problems is necessary for teaching
and research in order to realize the meaning of fuzziness while modelling real-world problems.
After all, all concepts and methods are developed for solving real-world problems. We expect
that this study will persuade more students, researcher and real decision makers/practitioners to
discover, investigate and solve all possible real problems.

1.S Literature Survey

A thorough and systematical literature survey has been carried out. Books, monographs
and proceedings, related to the concerned problems, are listed in Table 1.2. Among 300 journal
articles, technical reports and theses, the most related are listed in Table 1.3. Both tables provide
pertinent readings for each category of problems.
Because of the broad interdisciplinary character of this field, the literature is diversified
in many journals as shown in Table 1.4. Journals in which related articles on fuzzy approaches
appear frequently are indicated in the lists.
This monograph will provide readers and researchers with a capsule look into the existing
methods, their characteristics and applicability to analyzing fuzzy mathematical programming.
Although we have tried to give a reasonably complete survey, those papers not included were
either inadvertently overlooked or considered not to bear directly on the topics in this survey.
We apologize to both the readers and the researchers if we have omitted any relevant papers.
Before going into the actual review of fuzzy and possibilistic mathematical programming
in Chapters 3 and 4, we first review necessary backgrounds of fuzzy set theory in Chapter 2.
 11

Table 1.2 Classification of books, monoqraphs and proceedinqs

Class Reference

Fuzzy Set Theory and BM7, BM8, BM29, BM31, BM32, BM35, BM36,
Its operations BM46, BM47, BM55, BM56, BM57, BM63
Fuzzy Arithmetic BM31, BM32
Fuzzy statistics BM27, BM36, BM41
Fuzzy Mathematical BM2, BM4, BM27, BM30, BM32, BM40, BM47,
Proqramminq BM49, BM60, BM62, BM63, BM64
Fuzzy Multiple Objective BM1, BM26, BM27, BM38, BM45, BM64
Decision Makinq
Fuzzy Multiple Attribute BM3, BM64
Decision Makinq
Fuzzy Modellinq in BM10, BM22, BM23, BM27, BM30, BM32, BM55,
Industrial Enqineerinq BM56, BM59, BM62
Fuzzy Modellinq in Other BM1, BM2, BM6, BM14, BM15, BM16, BM22,
Decision-makinq Problems BM23, BM25, BM27, BM38, BM34, BM37, BM39,
BM43, BM44, BM49, BM50, BM51, BM52, BM53,
BM55, BM56, BM60, BM62, BM63, BM64
 12

Tel. 1.3 Classification of journal articles, technical reports and

theses

Class Reference
FUZZY LINEAR PROGRAMMING (FLP)
Fuzzy Resources 12, 43, 77, 122, 139, 176-181,
185, 222, 231, 236, 250-252,
257-259, 266, 306

Fuzzy Resources & Objectives 12, 34, 35, 37, 43, 79, 87-89,
91, 105, 120, 122, 128, 139,
143, 145, 148, 149, 151-155,
174, 176-181, 192-197, 202,
220, 255, 259, 263, 296-306
Fuzzy Objective Coefficients 12, 139, 251, 252
Fuzzy Coefficients 12, 31, 139
Sensitivity Analysis 112, 199, 215, 233, 242
Interactive FLP - An Expert 12, 137, 139, 141, 309
Decision-Making Support System
FUZZY NON-LINEAR PROGRAMMING 12, 84, 139, 248
FUZZY INTEGER PROGRAMMING 12, 83, 84, 139, 303
POSSIBILISTIC LINEAR PROGRAMMING (PLP)
Imprecise Resources & 20, 64, 66, 69, 72, 86, 121,
Technological Coefficients 139, 144, 175, 200, 201, 205,
209-212, 225-228, 232, 240,
243
Imprecise Objective Coefficients 52, 53, 56, 138, 139, 159
Imprecise Objective & 138, 139
Technological Coefficients
Imprecise Coefficients 20, 24-27, 43, 97-99, 121,
138, 139, 151, 157-161, 175,
192-197, 216, 217, 221, 225-
228, 238-244, 264
Imprecise Coefficients & Fuzzy 29, 30, 55, 56, 139
Inequalities
Imprecise Objective Coefficients 140, 217
& Fuzzy Resources
STOCHASTIC POSSIBILISTIC LINEAR 131, 142, 156, 264a
PROGRAMMING
 13

Table 1.4 List of journals

1. Automata
2. Colloquia Mathematica Societatis Janos Bolyai
3. Computers and Mathematics with Applications
4. computers and Operations Research
5. * Control and Cybernetics
6. Decision sciences
7. Decision Support Systems
8. Economic Computer and Economic Cybernetic Studies
and Researches
9. Environment and Planning A
10. * European Journal of Operational Research
11. * Fuzzy Sets and Systems
12. Geographical Analysis
13. Human Systems Management
14. * IEEE Transactions on Systems, Man, and Cybernetics
15. Information and Control
16. Information Science
17. International Journal of General System
18. International Journal of Intelligent System
19. International Journal of Man-Machine Studies
20. * International Journal of Production Research
21. International Journal of System Science
22. Journal of Cybernetics
23. Journal of Experimental Psychology: General
24. Journal of Experimental Psychology: Human Perception
and Performance
25. Journal of Mathematical Analysis and Its Applications
26. Journal of Operational Research Society
27. Kybernetika
28. Kybernetes
29. Large Scale System
30. Literary and Linguistic Computing
31. Machine Intelligence
32. Management Sciences
33. Mathematical Analysis and Applications
34. Mathematical Modelling
35. Naval Research Logistics
36. ORSA Journal on Computing
37. Tekhnicheskaya Kibernetika
* indicates that articles of fuzzy mathematical programming and fuzzy
multiple objective decision-making appear frequently in this journal.
2 FUZZY SET THEORY

The term "fuzzy" was proposed by Zadeh in 1962 [273]. In 1965, Zadeh formally
published the famous paper "Fuzzy Sets"[274]. The fuzzy set theory is developed to improve
the oversimplified model, thereby developing a more robust and flexible model in order to solve
real-world complex systems involving human aspects. Furthermore, it helps the decision maker
not only to consider the existing alternatives under given constraints (optimize a given system),
but also to develop new alternatives (design a system).
The fuzzy set theory has been applied in many fields, such as operations research,
management science, control theory, artificial intelligence/expert system, human behavior, etc.
In this study we will concentrate on fuzzy mathematical programming and fuzzy multiple
objective decision-making. To do so, we will first introduce the required knowledge of the fuzzy
set theory and fuzzy mathematics in this chapter.

2.1 Fuzzy Sets

Because of incomplete knowledge and information, precise mathematics are not sufficient
to model a complex system. Traditionally, the probability theory is the prevailing approach to
handle this incompleteness/uncertainty. One of the fundamentals in the probability theory is the
law of the excluded middle (P(AUAC) = 1) and contradiction (p(AnAC) = 0). For instance, a
fruit is either an apple or not an apple, an animal (normally speaking) is eith~r male or female.
For this case, the probability is certainly a good approach to represent some knowledge or
information whose boundaries can be clearly defined. Throwing coins into the air, you can guess
that either heads or tails will be up. Rotating a dice, the result will be 6, 5, 4, 3, 2, or 1, but
never be 4.5 or 2.1. Of course, there are a lot of problems satisfying the laws of the excluded
middle and contradiction. But, intuitively and commonsensically, this is not true in other
problems. An evidence favoring a particular hypothesis to some degree does not simultaneously
disconfirm it to any degree, because it may not give any support to the contrary. For example,
a man may be nnm:t, not smart, or a little smart. A color can be ~, lWl....!l:d, or ~.
Therefore, it is difficult to define the sets of smart men and red colors with sharp/crisp
boundaries. Similarly, how can we define and classify "&QQd credit," "beautiful women," "&QQd
taste," "reasonable price," "1QQd personality" and so on? Obviously, the probability theory
cannot model all the possible problems of incompleteness. The fuzzy set theory is developed to
 15

define and solve these problems without sharp boundaries. That is: fuzzy set theory considers
the partial relationship/membership.
To clearly distinguish the fuzzy sets from classical (crisp) sets, let us first consider the
following examples.

Example 1. In the classical set theory, an element may either "belong to" set A or "not belong
to" set A in the given universe. For example, suppose there is a rather large target and shooters
always hit inside the target (see Figure 2.1). A circle is located in the center of the target. If
a shooter hits inside the circle, region A, he is given the title "good shooter." Otherwise, he is
called a "poor shooter." Shooter a 1 shoots inside region A, so he is a good shooter. On the
other hand, shooter a2 , who shoots far away from region A, is a poor shooter. The classical set
theory with binary relationship shows some problems. For example, if there are three shooters,
a\ a4 and as, who hit the target within close range of one another (see Figure 2.1), yet only a3
is within the target, then shooter a3 is a good shooter, and shooters a4 and as are poor shooters.
This is obviously unreasonable.
As a matter of fact, these three people should obtain some similar designation, at least up
to a certain degree. Therefore, a measure of the degree to which the shooter belongs to the set
"good shooters" should be developed in order to discern how good the shooter is. The set
composed of good shooters is actually fuzzy because there is no crisp boundary. It is rational
to consider the distance from the boundary of the region A as a measure for indicating the degree
to which shooter ai belongs to the set of "good shooters." In Figure 2.1, a 1 and a3 are absolutely
good shooters. On the other hand, a4 and as are not absolutely good shooters or absolutely poor
shooters. They are, to some degree, good shooters. By giving a numerical measure which is
assumed linearly proportional to the distance, d, of each shooter from region A, one can say that
shooter a4 has 0.8 degree of membership in the set of good shooters versus 0.2 for shooter as.
p.(a4 ) = 0.8 and p.(aS) = 0.2. Of course, the numerical measure, p., can be any number. A
normalized measure which is in [0, 1] is always adapted. In the above example, the preference
concept is implied while assigning the numerical measures for each shooter to represent the
degree of "A SHOOTER BELONGING TO THE SET OF GOOD SHOOTERS." The shorter
the hit spot is from the center of the target, the larger the grade assigned (small d is preferred
to large d). The grade values are actually preference values. The function p. (called membership
function in fuzzy set theory), constituted by these grades as shown in Figure 2.1, is then a
preference function. Similarly, in practice, the decision maker may feel that: "Around $20,000
 16

U Target

Assigned grades Assumed distances from

indicating how good center of the target
the shooters are
0 7 <-- a2
0
0 6
0
0 5
0.2 <-- as
0.5 4

0.8 3 <-- a4
1.0 <-- a3
1.0 2
<-- a1
1.0 1

1.0 0

P. d (distance)

p.(d)

1
I

------1----'1--
I I
0.8 I I I
I I I
I I I
0.5
I r I

: I:+ ______ _
I I 1

I ____ 1
______ '- I __ I
0.2
1 I I I
I : I I
o d
o 1 2 3 4 5 6

a1

Figure 2.1 Graphic explanation for the fuzzy set of GOOD SHOOTER
 17

profit is acceptable;" "The budget should be iW!.1lm! $9,000;" "Dividends should not be him
than 6%;" "Overtime should be ~ than 5% of the regular man-hours;" and so on. Obviously,
these linguistic statements cannot be described by probability. The fuzzy set theory, on the other
hand, gives us a way to handle such linguistic situations. Meanwhile, the preference concept is
often assumed in building the membership functions of the above linguistic statements.
In order to be able to more clearly understand the fuzzy sets concept, let us consider one
more example.

Example 2. Consider a universe composed of 4 male students with the same height of 5'9" ,
Hans, John, George and Young. That is, U = { Hans, John, George, Young}. The weights
for the 4 male students are given as follows:

Hans : 164 pounds John: 190 pounds

George : 180 pounds Young: 160 pounds.

Now, let consider the linguistic proposition "fill male students." The students who belong
to "fat male students" then constitute a fuzzy set, A. Is Hans E A, John E A, George E A or
Young E A?
One plots the weights on a real line (see Figure 2.2) in order to present the relative
differences. According to common sense, a male student weighing more than 185 pounds at 5'9"
in height is absolutely considered a fat male student (actually the term "fat" depends on culture,
race, etc., which are beyond our research). On the other hand, a male student weighing less than
160 pounds is not fat at all. Therefore, it is obvious that John is absolutely fat and Young is
absolutely not fat. How about Hans and George? Actually Hans approaches Young with respect
to weight and George is close to John. It is true that the heavier the male student is, the greater
the degree to which he belongs to fuzzy set A. Thus, a degree-scaled line (see Figure 2.2) can
be drawn corresponding to the previous weight-scaled line in order to represent the degree of
membership indicating that a male student belongs to A. The scale on the degree-scaled line is
linearly proportional to the weight-scaled line when the weight belongs to the interval [160, 185].
As a result, the following grades of membership are available:

degree(Hans E A) = ll(x=H) = 0.2

degree(John E A) = ll(x=J) = 1
degree(George E A) = ll(x=G) = 0.8
 18

degree(Young E A) = ,,(x=y) = 0

where ,,(.) (detailed definition is given in the next section) is the membership function of the
fuzzy subset A of the set X. Like the previous example, the preference concept is used to elicit
the membership function ,,(.).

y H G J

I I x
0 160 165 170 180 185 190 pound

0
I
0
I
0.2
I
0.8
I
1
I
1
y
degree

y y = ~(x)

1 --------------------- ------ ---- ---:;r-----

0.8 -----------------------------

0.2

o x
160 165 170 180 185 190 pound

Figure 2.2 Derived degree of membership for the fuzzy set A

Example 3. Consider the statement of "the interest rate will be ~ 8% to 8.2%." The most
possible interest rate is between 8% and 8.2%. The least possible interest rate is 6.5%
(optimistic value) and 10% (pessimistic value). It is not likely that the interest rate will be less
than 6.5% or larger than 10% in the opinion of the decision maker. Thus, we give 8% - 8.2%
interest rate a 1 possibility, and 6.5% or less and 10% or more a 0 possibility. Between them,
let us assume the grades are as shown in Figure 2.3. Here, we did not use the preference
concept to grade various interest rates, but used the possibility concept: most possible, least
possible, or in between. Therefore, we call this membership function a possibility function or
possibility distribution, denoted by T(·).
 19

'/1'(1)

o ------~~----------~--r_------------;_- I
6.5% 8% 8.2% 10%

interest rate

Figure 2.3 The membership (possibility) function of the interest rate

2.2 Fuzzy Set Theory

Consider a crisp set A = {X I x = 2y, Y is nature number(N)}. Obviously, A is a set of
all even nature numbers. Thus, any nature number YEN is either belonging to A if it is even,
or not belonging to A if it is odd. The crisp set A can then be represented as:

A = {(I, 0), (2, 1), (3, 0), (4, 1), ... }

where the second number of the ordered pairs, 0 or 1, is the measure of membership. 0
indicates that the first number of the ordered pairs is not even, and 1 indicates it is even. By
using this notation, one can apply it to Example 2 and get the representation of the fuzzy set A
as:

A = {(Hans, 0.2), (John, 1), (George, 0.8), (Young, O)}

where the second number is no longer just 0 and 1, but is [0, 1]. This is the main characteristic
that makes the fuzzy sets different from the classical (crisp) sets. Now, it is time to consider the
following terminologies and definitions.
 20

2.2.1 Basic Terminology and nef"mition

2.2.1.1 nef"mition of Fuzzy Sets

Let X be a classical set of objects, called the universe, whose generic elements are denoted
by x. The membership in a crisp subset of X is often viewed as characteristic function IlA from
X to {O, I} such that:

IlA(X) = 1 if and only if x e A

=0 otherwise (2.1)

where {O, I} is called a valuation set.

If the valuation set is allowed to be the real interval [0, 1], A is called a fuzzy set
proposed by Zadeh [274]. IlA(X) is the degree of membership of x in A. The closer the value
of IlA(X) is to 1, the more x belongs to A. Therefore, A is completely characterized by the set
of ordered pairs:

A = {(x, IlA(X» I x eX}. (2.2)

It is worth noting that the characteristic function can be either a membership function as shown
in Examples 2 and 3, or a possibility distribution as shown in Example 1. In this study, if the
membership function is preferred, then the characteristic function will be denoted as Il(x). On
the other hand, if the possibility distribution is prefered, the characteristic function will be
specified as '/I"(x).
Along with the expression of Equation (2.2), Zadeh [280] also proposed the following
notations. When X is a finite set {XI' x2 , ••• , x,,}, a fuzzy set A is then expressed as:

(2.3)

For Example 2, A can be expressed as:

A = IlA(Hans)lHans + IlA(John)/John + IlA(George)/George + IlA(Young)/Young

= 0.2/Hans + l/John + 0.8/George + O/Young.

When X is not a finite set, A then can be written as:

 21

(2.4)

Sometimes, we might only need objects of a fuzzy set (but not its characteristic function),
that is to transfer a fuzzy set into a crisp set. To do so, we need two concepts - support and a-
level cut.

2.2.1.2 Support
The support of a fuzzy set A is the crisp subset of X and is presented as:

supp A = {x I I-'A(X) > 0 and x EX}. (2.5)

For Example 2, supp A = {Hans, John, George}, where Young does not belong to supp A
because of I-'(Young) 'j. O.

2.2.1.3 a-level Set (a-cut)

The a-level set (a-cut) of a fuzzy set A is a crisp subset of X and is denoted by (Figure
2.4 shows a continuous case):

A", = {x I I-'A(X) ~ a and x EX}. (2.6)

For Example 2, Ao.2 = {Hans, John, George} and Ao.7 = {John, George} .

1 - - - - - - - - - - - - y - - --.

I
I
I I
I
- - - -t - - - J.._ - - - -
I

o x

Fiqure 2.4 An a-level set

 22

Although characteristic function can be assigned by any number, a normalized value

between 0 and 1 is always preferred. Thus let us introduce the normality as follows:

2.2.1.4 Normality
A fuzzy set A is normal if and only if Supx IlA(X) = 1, that is, the supreme of IlA(X) over
X is unity (see Figure 2.4 also). A fuzzy set is subnormal if it is not normal. A non-empty
subnormal fuzzy set can be normalized by dividing each IlA(X) by the factor SUPx IlA(X), (A fuzzy
set is empty if and only if IlA(X) = 0 for Vx EX).
It is noted that a characteristic function is always a normalized function through this study.

2.2.1.5 Convexity and Concavity

A fuzzy set A in X is convex if and only if for every pair of point Xl and X2 in X, the
membership function of A satisfies the inequality:

(2.7)

where 0 E [0, 1] (see Figure 2.5). Alternatively, a fuzzy set is convex if all a-level sets are
convex.

P.A(X)
1 ------------;...----~_

I
I
I
I
-r- --- - - - - - - - - 1----
I I I
I I
I I
I I
I
I
o x

Figure 2.5 A convex fuzzy set

 23

Dually, A is concave if its complement AC is convex. It is easy to show that if A and B

are convex, so is A n B. Dually, if A and B are concave, so is A U B.

2.2.1.6 Extension Principle

Let X be a Cartesian product of universe, X = Xl x X2 x ... x X n, and AI' A2, ... , An
be n fuzzy subsets of Xl, X 2, ... , Xn, respectively. The Cartesian product of AI' A2, ... , An
is defined as:

Let f be a mapping from Xl x X2 x ... x Xn to a universe Y such that y = f(xI' x2' ... ,
Xn), where Xl f Xl' ... , Xn f X n· The extension principle (Zadeh [280)) allows us to induce a
fuzzy subset B on Y through f from n fuzzy sets Aj such that:

J.'B(y) = sup min (J.'AI(xI)' J.'A2(xV, ... , J.'An(Xn» (2.9)

y=f(xI,···,xn)

= 0 if r-l(y) = (J)

where r-l(y) is the universe image of y. J.'B(y) is the greatest among the membership values

J.'Alx ... xAn(XI' ... , Xn) of the realization of y using n-ary (xl' ... , Xn).

For example, let A = E J.'A(X)/X = 0.6/-1 + 0.7/0 + 111 + 0.3/2, and f(x) = x2. Then,

we have:

J.'B(Y) = sup J.'A(x) if r-l(y) =1= (J) (2.10)

Xfr-l(y)
= 0 otherwise.

By applying the extension pronciple, B = 0.7/0 + 111 + 0.3/4 as shown in Figure 2.6:

2.2.1.7 Compatibility of Extension Principle with a-cuts

Denoting the image of AI' ... An by B = f(Al> ... , An>, the following proposition holds:

(2.11)
 24

(4, 0.3)

(2, 0.3)

sup (0.6, 1) 1
(1, 1) (1, 1)

(0, 0.7) (0, 0.7)

(-1, 0.6)

(x, PoA(x» (y, PoB (y) )

Piqure 2.' The extension principle

iff for'Vy E Y, there exists (Xl"" ... , Xu....) and #LB(Y) = #LAlx ... xAn(xl"', ... , Xu.J. (The upper
bound in Equation (2.9) is obtained.)

2.2.1.8 Relation
A fuzzy relation, R, in the product space X X Y = {(x, y) I X E X and y E Y} is a fuzzy
set in X x Y characterized by a membership function #LR which associates with each ordered pair
(x, y) a grade of membership #LR(x, y) in R. More generally, an n-ary fuzzy relation in a
product space X = Xl x X 2 x ... x Xn is a fuzzy set in X characterized by an n-ary variate
membership function #LR(xI' x2' ... , xn>, ~ E Xi' i = 1, 2, ... , n.
For example, let X = Y = RI, where RI is the real line (-00, (0). Then x > > y is a
fuzzy relation in R2. A subjective expression for #LR' in this case, might be:

#LR(X, y) =0 for x ~ y
= (1 + (x-yr2r l for x > y.

2.2.1.9 Decomposability
Let X = {x}, Y = {y} and let C be a fuzzy subset in the product space Z =X x Y
defined by a membership function #LC<X, y). Then, C is decomposable along X and Y if and only
if C admits of representation C =A n B (A is a fuzzy subset in X and B is a fuzzy subset in
 25

Y). That is:

(2.12)

where P-A(X) and P-B(X) are the membership functions of A and B. The same holds for a fuzzy
subset in the product of any finite number spaces.

2.2.1.10 Decomposition Theorem

Any fuzzy subset A may be decomposed in the following form:

(2.13)

where ai € (0, 1), i = 1, 2, ... , n. For example, A = «XI> 0.2), ("-2, 0.7), (X3' 1), (X4' 0)).
Obviously, we should take a l = 0, a2 = 0.2, a3 = 0.7 and a 4 = 1. Then, we have the
following results:

X XI X2 X3 X4

J.LA(X) 0.2 0.7 1.0 0.0

Ao.o 1.0 1.0 1.0 1.0

Ao.2 1.0 1.0 1.0 0.0
Ao.7 0.0 1.0 1.0 0.0
AI. O 0.0 0.0 1.0 0.0

0.0 (Ao.o) 0.0 0.0 0.0 0.0

0.2 (Ao.2) 0.2 0.2 0.2 0.0
0.7 (Ao.7) 0.0 0.7 0.7 0.0
1.0(Al. o) 0.0 0.0 1.0 0.0

max ai QjAai 0.2 0.7 1.0 0.0

2.2.1.11 Probability of Fuzzy Events

Let P be a probability measure on O. A fuzzy event A in 0 is defined to be a fuzzy subset
A of 0 whose membership function /lA is measurable. Then the probability of A is defined by
the Lebesgue-Stieltjes integral:

P(A) = f
0 P-A(X) dP.
(2.14)

Equivalently, P(A) = E(/lJ where E0 denotes the expectation operator. In the case of a normal
 26

crisp set, the above equation will reduce to the conventional definition of the probability of a
crisp event.

Example 4[BM38]. Let per capita income be depicted by the fuzzy subset A = "around $a" and
its membership function is defined as:

JLA(X) = exp{-(x-a)l}, x E R.

Assume that the probability distribution of per capita income be normal distribution as depicted
by the following probability density function:

exp{ -1I2[(x-m)/u]l}
p(x) =
oV2i

where m and u are mean value and standard deviation, respectively, and 7r = 0.341416... Then
the probability of the fuzzy event A is:

I +'"'
-,",
exp{ -(x-a)l} exp{ -1I2[(x-m)/u]l} dx

P(A) =--------------

where P(A) can be defined as the expectation of its membership function.

For a discrete case, let X be a set of events (Xl' Xl' ... , x,,} with probabilities P(X;) where
P(X;) E [0, 1] and E j P(X;) = 1. A fuzzy event A is a fuzzy subset on X whose membership
function is JLA(X;). The probability of the fuzzy event A is then defined as:

(2.15)

Finally, it is worth noting that P(A) satisfies the following properties:

(a) 0 :s: P(A) :s: 1
(b) P(0) = 0 and P(O) = 1
(c) P(A) :s: P{B) if AcB
(d) P(AC) = 1 - P(A)
(e) P(AUB) = P(A) + P{B) - p(AnB)
(f) P(AUB) = P(A) + P(B) if A n B = 0.
 27

2.2.1.12 Conditional Fuzzy Sets

A fuzzy set B(x) in Y = {y} is conditioned on x if its membership function depends on
x as a parameter. This dependence is expressed by JJ.s(Y i x).
Suppose that the parameter x ranges over a space X, so that to each x in X corresponds
a fuzzy set B(x) in Y. Thus, we have a mapping, characterized by JJ.s(Y i x) - from X to the space
of fuzzy sets in Y. Through this mapping, any given fuzzy set A in X induces a fuzzy set B in
Y which is defined by:

JJ.s(y) = supx min (JJ.A(X), JJ.s(Y i x»

= VX (JJ.A(X) A JJ.s(yix» (2.16)

where JJ.A(X) and JJ.s(x) are the membership functions of A and B. Note that this equation is
analogous - but not equivalent - to the expression for the marginal probability distribution of the
joint distribution of two random variables, with JJ.s(y i x) playing a role analogous to that of a
conditional distribution.

2.2.2 Basic Operations

In this sub-section, we will summarize some basic set-theoretic operations which is useful
in fuzzy mathematical programming and fuzzy mUltiple objective decision-making. These
operations are based on the definitions from Bellman and Zadeh [12].

2.2.2.1 Inclusion
Let A and B be two fuzzy subsets of X. Then, A is included in B if and only if:

JJ.A(X) ::::;; JJ.s(x) for 'Ix € X. (2.17)

This will be denoted by ACB. For example, suppose X = {Xl' x2, X3, ~}, A = O.l/x l + 0.5/X2
+ 0/X3 + l/~ and B = O.l/Xl + 0.6/X2 + 0.3/X3 + l/~. According to Equation (2.17), A
is included in B (or ACB), since 0.1 ::::;; 0.1,0.5 ::::;; 0.6, 0 ::::;; 0.3 and 1 ::::;; 1.

2.2.2.2 Equality
A and B are called equal if and only if:

JJ.A(X) = JJ.s(x) for 'Ix € X. (2.18)

 28

This will be denoted by A = B. For example, suppose X = {Xl' X2' X3 , x4 }, A = O.l/xl +

0.5/x2 + O/x3 + l/~ and B = O.l/Xl + 0.5/x2 + 0/x3 + l/~. According to Equation (2.18),

A equals to B (or A = B), since 0.1 = 0.1,0.5 = 0.5, 0 = 0 and 1 = 1.

2.2.2.3 Complementation
A and B are complementary if and only if:

J.'A(X) = 1 - J.'B(X) for Vx E X. (2.19)

This will be denoted by B = N or A = Be, where N and Be are the complements of A and B,
respectively. For example, if B = O.l/Xl + 0.6/X2 + 0.3/x3 + l/~, then A = 0.9/XI + 0.4/x2
+ 0.7/x3 + 0/x4 = Be.

2.2.2.4 Intersection
The intersection of A and B may be denoted by A n B which is the largest fuzzy subset
contained in both fuzzy subsets A and B. When the min operator (aggregator) is used to express
the logic "and," its corresponding membership is then characterized by:

J.'Ar1B(X) = min (J.'A(X), J.'B(X» for Vx E X

= J.'A(X) A J.'B(X) (2.20)

where A is conjunction here.

Example 5. Consider X = {Hans, John, George, Young}. Suppose A is the fuzzy subset of
"good looking students" and B is the fuzzy subset of "intelligent students." The following data
are also assumed:

A = 0.3/Hans + O.4/John + 0.7/George + O.8/Young

B = 0.7lHans + O.4/John + O.l/George + 0.3/Young.

Then, the intersection of A and B (if the min-operator is adopted) A n B = O.3/Hans +

O.4/John + O.l/George + 0.3/Young which is the fuzzy subset of "good looking and intelligent
student" (see Table 2.1 also). Besides, A is considered as a hard "and."
 29

2.2.2.5 Union
The union (A U B) of A and B is dual to the notion of intersection. Thus, the union of
A and B is defined as the smallest fuzzy set containing both A and B. The membership function
of A U B is given by:

~AUB(X) = max (~A(X), ~B(X» for Vx e X

= ~A V ~B (2.21)

where V is disjunction. For Example 5, A U B = O.3lHans + O.4/John + O.l/George +

O.3/Young is the fuzzy subset of "good looking or intelligent students" (see Table 2.1 also).
Besides, V is considered as a hard "or."

2.2.2.6 Algebraic Product

The algebraic product AB of A and B is characterized by the following membership
function:

(2.22)

This algebraic product is considered as a soft "and." For Example 5, AB = O.2l/Hans +

O.16/John + O.07/George + O.24/Young is the fuzzy subset of "good looking and intelligent
students" (see Table 2.1 also).

2.2.2.7 Algebraic Sum

The algebraic sum A ffi B of A and B is characterized by the following membership
function:

(2.23)

This algebraic sum is considered as a soft "or." For Example 5, A ffi B = O.79lHans +
O.64/John + O.73/George + O.86/Young is the fuzzy subset of "good looking or intelligent
students" (see Table 2.1 also).

2.2.2.8 Difference
The difference A - B of A and B is characterized by:

(2.24)
 30

Table 2.1 Solutions of Example 5

X Hans John George Young

~A(X) 0.3 0.4 0.7 0.8

~B(X) 0.7 0.4 0.1 0.3

~AnB(x) 0.3 0.4 0.1 0.3

~AUB(x) 0.7 0.4 0.7 0.8

~AB(x) 0.21 0.16 0.07 0.24

~AeB(x) 0.79 0.64 0.73 0.86

where BC is the complement of B with ~Bc(x) =I - ~B(x). For Example 5, we have A - B as

shown in the following table:

X Hans John George Young

~A(X) 0.3 0.4 0.7 0.8

~B(x) 0.7 0.4 0.1 0.3

~Bc (x) 0.3 0.6 0.9 0.7

~AnBc (x) 0.3 0.4 0.7 0.7

2.3 Membership Functions

While classical (crisp) mathematics are dichotomous in character, fuzzy set theory
considers the situations involving the human factor with all its vagueness of perception,
subjectivity, attitudes, goals and conceptions. By introducing vagueness and linguistics into the
crisp set theory, fuzzy set theory becomes more robust and flexible than the original
dichotomous classical set theory.
There are two essential features of the fuzzy set theory (Zimmermann [301]):
(I) Membership function of fuzzy sets and operators play a crucial role in fuzzy set theory.
(II) Fuzzy set theory is essentially a very general, flexible, formal theory. If it is to be applied
to a real problem, it can and has to be adapted carefully. Neither the concept of membership nor
the operator has a unique semantic interpretation. The context-dependent semantic interpretation
will lead to different mathematical definitions and appropriate operators.
 31

Thus membership functions and operators are actually the cornerstones of the fuzzy set
theory. In this section, let us discuss some approaches to generate membership functions, while
leaving the discussion of operators (aggregator) for the Section 2.4.

2.3.1 A Survey of Functional Forms

In this section, we will present a state-of-the-art survey of functional forms of membership
(both preference-based membership functions and possibility distributions). This survey will
provide researchers a direction of future studies in fuzzy and possibilistic mathematical
programming problems.
Based on Dombi [61] (with modifications), we classify all existing membership functions
into the following four classes:

I. Membership functions based on heuristic determination.

(a) Zadeh's unimodal functions:

/l-you.,,(x) = J1I{1+[(x-25)/Sf} if x > 2S

11 ifx~25 (2.25)

P.Old(X) = {01l{1+[(X-SO)/SJ 2} if x ~ SO
ifx<SO. (2.26)

(b) Dimitru and Luban's power functions:

p.(x) = x2 /a2 + 1, x E [0, a] (2.27)

p.(x) = -r/a2 -2x/a + 1, x E [0, a]. (2.28)

(c) Svarowski's sin function:

p.(x) = 112 + (1I2)[sin{[1r/(b-a)][x-(a+b)/2]), x E [a, b]. (2.29)

n. Membership functions based on reliability concerns with respect to the particular problem.

(a) Zimmermann's linear function:

 32

p.(x) = 1 - x/a, x E [0, a]. (2.30)

(b) Tanaka, Uejima and Asai's symmetric triangular function:

p.(x) = {I - I b-x I la, ifb-a SS x b+a

o otherwise. (2.31)

(c) Hannan's piecewise linear function:

p.(x) = Ej OJ I x-ll.i I + {3x + r, j = 1,2, ... , N (2.32)

OJ = (~+I - 9 /2
{3 = (tN + 1 + t l )/2
r = (SN+I + sl)/2

where p.(x) = ~x + Sj for each segment i, 1\-1 S X S 1\, Vi. Here, ~ is the slope and Sj is the
y-intercept for the section of the curve initiated at 1\-1 and terminated at 1\.

(d) Leberling's hyperbolic function:

p.(x) = 112 + (1I2)tanh(a(x-b)), -00 S x S 00 (2.33)

where a is a parameter.

(e) Sakawa and Yumine's exponential and hyperbolic inverse functions, respectively:

p.(x) = c(1_e(b-X)/(b-a»), x E [a, b] (2.34)

p.(x) = 112 + ctanh-I(d(x-b)) (2.35)

where c and d are parameters.

(f) Dimitru and Luban's function:

p.(x) = 11(1 +x/a) (2.36)

where a is a parameter.
 33

(g) Dubois and Prade's L-R fuzzy number:

L«a-x)/a) if x < a
IL(X) = { R«x-b)/~) if x > b
1 ifa:S;x:S;b (2.37)

where L(') and R(') are reference functions.

m. Membership functions based on more theoretical demand.

I
(a) Civanlar and Trussel's function:

lL(x) = ap(x) if ap(x) :S; 1

o otherwise (2.38)

where a E [0, 1] is a parameter and p(x) is the probability density function.

(b) Svarovski's function:

0 if x < a
lL(x) = K(x-a)2 if a S x :S; b
K2X2 + K1x + Ka ifb<x:S;c
1 if x > c (2.39)

where K, Ka, Kl and K2 are parameters.

VI. Membership functions as a model for human concepts.

(a) Hersh and Caramazza's function (which was experimentally formed in order to determine the
nature of context effects upon the interpretation of a set of nature language terms, such as short,
very short, sort of short, etc.):

IL(X) = 112 + d(r/l0) (2.40)

where d(x) = 1 for "yes" reponses and d(x) = -1 for "no" responses and r is a confidence value.
 34

(b) Zimmennann and Zysno's (see Section 2.3.2.1) function:

p.(x) = 1/2 + (1/d)[1/(l +e-a(x-b» - c]. (2.41)

(c) Dombi's function:

(2.42)

where s (the characteristic value of the shape) is the intersection value of y = p.(x) and y = x.
As discussed in Chapter 1, membership functions can be distinguished into two classes of
preference-based membership functions and possibility distributions. The preference-based
membership function is constructed by eliciting the preference infonnation from the decision
makers. On the other hand, the possibility distribution, which is an analogous of probability
distribution, is constructed by considering the possible occurrence of the events. However, this
section is to show the existing functional fonns, yet not to discern the differences between
preference-based membership functions and possibility distributions.
The functions, which are most often used in modelling fuzzy and possibilistic mathematical
programming problems, have been shown in Figures 1.1 and 1.2. Detailed discussions will be
given in Chapters 3 and 4.
When modelling fuzziness and imprecision of the input data (A, b and c) in fuzzy and
possibilistic mathematical programming problems, preference-based membership functions and
possibility distributions are always assumed to be rationally given. Thus, how to generate
preference-based membership functions and possibility distributions becomes a natural question.
In the next section, we provide some examples of generating membership functions as references
for interested readers.

2.3.2 Examples to Generate Membership Functions

There are two approaches to generate membership functions: axiomatic and semantic
approaches (Giles [106]). The axiomatic approach, which is similar to the approaches used in
utility theory (Zimmennann [301]), is centered on the mathematical consideration. On the other
hand, the semantic approach is concentrated on the practical interpretation of the tenns but not
on the mathematical structure which is emphasized in the axiomatic concept. The semantic
approach actually follows the perceptions of pragmatism in its insistence that all conclusions
should be finnly based on the practical meaning of the concepts involved. There is no axiom or
 35

law - in principle not even the syntax - laid down in advance. The purpose is to start with an
analysis of the immediate practical meanings of the concepts. These properties are then
formulated in precise terms as axioms. Other concepts may then be defined in terms of the initial
ones, and so a formal theory is gradually built up. Unlike the axiomatic approach, the semantical
approach provides a derivation of a particular mathematical theory, and from the way the
concepts are introduced, it yields a specific interpretation of this theory at the time.
In this study, we concentrate on the applications and not pure mathematics. The practical
meaning and interpretation of membership are considered more important than mathematical
terms. The semantical approaches, thus, are emphasized and discussed in the following
subsections. However, this section is not required to understand fuzzy and possibilistic
mathematical programming.

2.3.2.1 Distance Approach

When eliciting membership functions, the determination of the lowest necessary scale level
of membership for a specific application is very important. Generally, the required scale level
should be as low as possible in order to facilitate data acquisition which usually affords the
participation of human beings. Besides, a suitable numerical handling is desirable in order to
insure mathematically appropriate operations. Among the five classical scale levels (nominal,
ordinal, interval, ratio, and absolute scale), Zimmermann and Zysno [305] proposed that the
interval scale level seems to be most adequate, since the intended mathematical operations require
at least interval scale quality.
The concept of using a distance d(x) from a reference as an evaluation criterion has been
popular in many fields, such as mUltiple attribute decision-making, multiple objective decision-
making, goal decision-making, et al. Zimmermann and Zysno used this concept to derive
membership functions. The rationality of this concept is that if the object has all the ideal
features, the distance should be zero. By contrast, if no similarity between the object and the
ideal exists, the distance then should be 00. If this evaluation concept is formally represented
by a fuzzy set A in X, then a certain degree of membership JLA(X) (or simply JL) will be assigned
to each element x. Membership can then be defined as a function of the distance between a given
object x and a standard (ideal). Thus Zimmermann and Zysno proposed the following relation
for JL and d(x):

JL = 1/(1 + d(x» (2.43)

 36

where d(x) = 0 - > p. = 1, and d(x) = 00 - > p. = O. Obviously, Equation (2.43) is just a
transformation rule (or a normalization process) which maps real number R into the interval [0,
1].
However, experience shows that ideals are very rarely ever fully realized and distance
function is quite context dependent. The context-dependent parameters a and b are then created
to represent the evaluation unit and the reference/standard, respectively. For instance, a may be
the unit of length such as feet, meter, yards, etc., and b may be a fast and rough pre-evaluation
such as "rather positive," "rather negative," etc. Since the relationship between physical units
and perceptions is generally exponential, Zimmermann and Zysno proposed the following
distance function as:

d(x) = 1/ea(x-b) (2.44)

and then:

p. = 1/[1 + e-a(x-b)] (2.45)

where a, b can be considered as semantic parameters from a linguistic point of view.

Since concepts or categories, which are formally represented by sets, are normally
linguistically described, the membership function is the formed representation of meaning. The
vagueness of the concept is operated by the slope a and the identification threshold by b. For
managerial terms such as "appropriate dividend or "good utilization of capacities, " the parameter
H

a models the slope of the membership function in the tolerance interval and b represents the point
at which the tendency of the subject's attitude changes from rather positive into rather negative.
Equation (2.45), however, is still too general to fit subjective models of different persons.
Frequently, only a certain part of the logistic function is needed· to represent a perceived
situation. In order to allow for such a calibration, it is assumed that only a certain interval of
the physical scale is mapped into the open interval (0, 1). And, the membership grades of the
lower and upper bounds are assigned to be 0 and 1, respectively.
Since an interval scale is requested, the interval of the degrees of membership may be
transformed linearly. On this scale level the ratios of two distances are in variant. Let P.u and
p.L be the upper and lower bounds of the normalized membership scale, respectively, and P.i is
a degree of membership between these bounds, p.L < P.i < P.u, and let p.L' and P.u, be the
corresponding values on the transformed scale. Then we have:
 37

(2.46)

For the normalized membership function, one may have p.L = 0 and P.u = 1. Hence,

(2.47)

Generally, it is preferable to define the range of validity by specifying the interval d with the
center c (see Figure 2.7) as follows:

(2.48)

Then Equation (2.47) will become:

P.i' = p.i(d) + p.L'

= dP.i - d/2 + d/2 + p.L'
= dP.i - d/2 + [(p.u, _p.L')/2 + p.L']
= d(P.i - 112) + c (2.49)

which can be specified by two parameters, a and b, if Jl.i is replaced by P.i'.

iJ.

o x

Figure 2.7 Calibration of the interval for measurement

 38

Solving Equation (2.49) for "'j will lead to the following complete membership model:

1
"'j = I (lId)(
1 + e-a(x-b)
-c) + 1/2 (2.50)

where "'j E [0, 1], "'j(x) = 1 for x > xu, and "'j(x) = 0 for x < xL.
The determination of the parameters from an empirical data base does not pose any
difficulties in the general model of Equation (2.45). That is:

'" = 1/[1 +e-a(x-b)] = > In [p./(l-",)] = a(x-b). (2.51)

Now, suppose Y = In [",/(1-",)]. A linear relationship between x and y is obvious. The straight
line of the model is then defined by the least squares of deviations.
The estimation of the parameter c and d in the extended model still poses some problems.
There is no direct way for a numerically optimal estimation. Thus Zimmermann and Zyson
proposed the following interactive procedure. First, they assumed that a set of stimuli which is
equally spread over the physical continuum was chosen such that the distance between any two
of the neighboring stimuli is constant:

Xj+1 - Xj = s (2.52)

where s is a constant distance. This condition serves as a criterion for precision. If c and d are
correctly estimated, then those scale values Xj' are reproducible and are invariant with respect
to Xj with the exception of the additive and multiplicative constant. It is obvious that Equation
(2.50) can be rewritten as:

or

1 - [d("'j-I12) + c] (2.53)

Let s' be the distance between the pair Xj+l' and Xj', and M' be their mean value. If the
estimated values de and ce are equal to their true values, then the estimated distance se, and the
mean Me, are equal to their respective true values and vice versa:
 39

(d" = d) and (c" = c) < = > (s"' = s) and (Me, = M). (2.54)

Our aim, therefore, is to reach the equivalence of se' and s as well as Me' = M.

Aieorithm.

Step O. Determine the initial appropriate values c, and d, by the following formula:

c, = ( Ej Jl.j)/n (2.55)
d, = min (I-11k, 2(1-c), 2c) (2.56)

where n is the member of stimuli and k is the number of different degrees of membership. If
only the values 0 and 1 occur, d, = 112. Only the linear interval in the middle of the logistic
function is used. With increasing k, d converges to 1; the entire range of the function is then
used. In any case d must not exceed the minor part of 2c or 2(I-c).

Step 1. Determine the Xj' which corresponds to the empirically determined Jl.j:

(2.57)

se' = [E j (Xj+,-X'j)]/(n-1)
= (x,,'-x,')/(n-1). (2.58)

Step 2. Calculate the absolute difference between two estimates. That is:

Me" _ Me' I ~ ~ (2.59)

sell _ se' I ~ Os (2.60)

where ~ and Os are predetermined tolerances. If Equations (2.57) and (2.58) are satisfied, the
last estimate is accepted as sufficiently exact and then STOP. Otherwise, go to the next step.

Step 3. Determine the interval of the base variable Xj' which corresponds to the (0, 1) interval
of the membership value. That is to compute:

xU, = Me' + (n/2)se, (2.61)

XL, = Me' - (n/2)s·' (2.62)

 40

which is the upper and lower bounds, respectively.

Now, by Equation (2.53), the corresponding P.u, and p.L' can be computed, and then by
Equation (2.48), new parameters c· and d· are estimated. Then go to Step 1.

Example. From the empirical evidence that 64 subjects (16 for each set) from 21 to 25 years
of age individually rated 52 different statements of age concerning one of the four fuzzy sets:
"very young man," "young man," "QlQ man" and "very old man," Zimmermann and Zyson
obtained monotonic membership functions as shown in Figure 2.8, and unimodal membership
functions as shown in Figure 2.9.
The monotonic memberShip functions of "old man" and "very old man" (Figure 2.8) are
rather similar. They differ only with respect to their inflection points, indicating a difference of
about five years between "old man" and "very old man." The same holds for the monotonic
membership functions of "very young man" and "young man"; their inflection points differ by
nearly 15 years. It is interesting to note that the modifier "very" has a greater effect on "young"
than "old," but in both cases it can be formally represented by a constant.
Finally, the meaning of "young" is less vague than that of "old" if the slope is an indicator
for vagueness as shown in Figure 2.8. On the other hand, the variability of membership
functions may be regarded as an indicator of ambiguity. Thus, though being less vague, "young"
seems to be more ambiguous.

2.3.2.2 True-Valued Approach

Smets and Magrez [229] provided a definition of what is meant by a proposition with a
truth value of 0.35 or any other intermediate value between True (or 1) and False (or 0). A
canonical scale for the truth values is defined, and sets of fuzzy propositions are constructed for
which the truth value has a unique numerical value on the canonical scale. Such a set of
propositions with well-defined intermediate truth values is necessary to give meaning to the
assertion "the truth P is 0.35."
Fuzzy logic has two characteristics: the truth domain is the whole [0, 1] interval, and the truth
value can be a fuzzy subset of [0, 1]. Smets and Magrez considered the first characteristic, that
is to reduce fuzzy logic to its multi-valued logic components.
According to Zadeh [274], fuzzy sets are those for which one can define, for each
element, a grade of membership in [0, 1]. Smets and Magrez then postulated the relation
between degree of truth and grade of membership so that the degree of membership P.A(X) of an
 41

.9

'" (x) .8

•7

.6

.5

.4
very old man
•3

•2
.1

0 x
10 20 30 40 50 60 70 80 90 100

Figura 2.8 Monotonic membership functions of "very young man," "young

man," "old man" and "very old man" [305]

.9

'" (x) .8
young man
•7
very young man
.6

.5

•4

•3

•2

.1

0 x
10 20 30 40 50 60 70 80 90 100

Figure 2.9 unimodal membership functions of "very young man" and "young
man" [305]
 42

element x to a fuzzy set A is numerically equal to the degree of truth v(x is A') that the fuzzy
predicate A' describing the fuzzy set A applies to the element x:

P-A(X) = a < => v(x is A) = a. (2.63)

For instance, if A is the fuzzy set of "tall men," the degree of membership p-A(John) of John to
the set A is numerically equal to the degree of truth of the proposition "John is tall". Therefore,
the presentation could have been based on both approaches when fuzzy logic is used. Whatever
is deduced for degree of truth can be applied to degree of membership.
The semantical interpretation of "A - > B is true" is that the consequent B is at least as
true as the antecedent A. The degree of truth of A - > B quantifies the degree by which B is at
least as true as A. It will be Truth whenever B is truer than A. Otherwise, it will be False
whenever A is False and B is Truth. When A is somehow truer than B, the degree by which B
is at least as true as A might be intermediate between True and False.
Now, let v(A) be the truth value of proposition A with A f 0, where 0 is a Boolean
algebra of propositions with T its tautology and F its contradiction. Then v: 0 - > (1 is a
mapping from 0 to a truth domain (1, where (1 is a bounded set with its greatest element Truth
= v(T), its least element False = v(F) and the order relation "at least as true as." Thus Smets
and Magrez defined a strictly increasing transformation w from to [0, 1] such that w·v: 0 - >
°
(1

[0, 1] with w·v(T) = 1, w·v(F) = and' A is as true as B' is equivalent to "w·v(A) ~ w·v(B)."
If v = w·v, then v will measure the degree of truth of propositions in 0 on the interval [0, 1].
For simplicity, v(A) is the truth value of A (in fact the truth value is defined on (1 and v(A) is
defined on [0, 1]), because they are uniquely related.
In order to construct a logic with multi-value truth domain for generating the degree of
truth, the following axioms for the implication connective - > should first be considered:

AI. Truth-Functionality: v(A- > B) depends only on v(A) and v(B).

A2. Contrapositive Symmetry: v(A- > B) = v(--, B- > --, A), where the Trillas's strong negation
--, is defined: there exists a strictly decreasing continuous function n: [0, 1] - > [0, 1] such that
v(--, A) = n(v(A», n(O) = 1, n(1) = °and n(n(a» = a.
AJ. Exchange Principle: v(A->(B->C» = v(B-> (A-> C».
A4. Monotony: v(A- > B) ~ v(C- > D) if v(A) ~ v(C) and v(B) ~ v(D).
AS. Boundary Condition: v(A- > B) = 1 iff v(A) ~ v(B).
A6. Neutrality Principle: v(T- > A) = v(A).
 43

A7. Continuity: v(A- > B) is a continuous function of v(A).

With these axioms, Smets and Magrez have proved that the implication and the negation
operators are necessary such that:

v(A->B) = f'(min{f(I)-f(a)+f(b),f(l)}) (2.64)

v(-, A) = f'(f(I)-f(a)) (2.65)

where a = v(A), b = v(B), and the generator f is any bounded, continuous, monotonically
increasing function from [0, 1] to [0, (0) with f(O) = 0 and f(1) < 00.

The f generator is defined up to any strictly monotone transformation. As the truth scale
v(A) is also defined up to any strictly monotone transformation, we can then define the canonical
scale for the truth value v(A) of a proposition by selecting the f generator such that f(v(T)) =
f(l) = I and f(v(A)) = f(a) = a. In that case, we obtain the Lukasiewicz operator for the
material implication connective:

v(A->B) = min {1-v(A)+v(B), I} (2.66)

v(-, A) = I - v(A). (2.67)

This conclusion is quite natural as f and v can always be adapted in order to obtain such a scale,
and it is obviously the simplest we can construct.
However, this canonical scale does not explain what is meant by a truth value of 0.35.
It must mean something more than just that it is between 0.34 and 0.36. Thus we need a set of
propositions for which the truth value is uniquely defined. These propositions could be used later
as a reference scale to measure the truth of other propositions.
According to Gaines, Smets and Magrez constructed a reference scale which defines the
meaning of any truth value in [0, 1] as follows:
1. Let A = "John is tall" and --, A = "John is not talL" The truth of A and of --, A
depends on the height h of John. If h = 110cm, A is false and --, A is true. If h = 190 cm, A
is true and --, A is false (see Table 2.2). Thus, when h increases from 110 to 190, a = v(A)
increases continuously from 0 = v(F) to 1 = v(T) (where T and F are the tautology and
contradiction) and na = v(-, A) decreases continuously from 1 to O.
 44

Table 2.2 Heights and true values of A = "John is tall"

and B "Paul is tall" [372]

height 110cm k' h' 190cm

yeA) 0 0.5 1
V(,A) 1 0.5 0
v(B) 0 0.25 0.5 1
v(, B) 1 0.75 0.5 0
v(, B -> B) 0 0.50

There exists a value h' such that a = na. In such a case, .., A - > A is true, therefore
v(.., A- > A) = 1 becomes fl(f(I)-f(na) +f(a» = 1. As f(na) = f.fl(f(1) - f(a» = f(l) - f(a),
we have fl(2f(a» = 1, thus 2f(a) = f(I).
As f(1) < 00, we can use an fscale such that f(l) = 1, in which case f(a) = f(na) = 0.5.
It is noted that the result would have been directly obtained by considering the strong negation
and the canonical scale as a = na is identical to a = I-a.
2. Let B = "Paul is tall" and.., B = "Paul is not tall." The truth of B and of.., B
depend on the height k of Paul. Let us further postulate that k ~ h', thus b = v(B) ~ 0.5.
Consider the proposition.., B - > B, i.e., 'B is at least as true as .., B'. Its truth value c
= v(-, B->B) depends on k. Ifk = h', c = 1 as v(B) = v(-, B). Ifk = 1l0cm, B is false and
.., B is true, therefore c = O. By increasing k from 1l0cm to h', c is increasing from 0 to 1.
Thus, there exists a k' such that v(-, B- > B) = v(A) = 0.5 implies f(l) - f(nb) + f(b) = 2f(b)
= f(a) = 0.5 = > f(b) = 0.25 and f(nb) = 0.75.
3. Let Po = A, PI = B, ho = h' and hi = k'. The procedure is iterated with the
propositions Pi == "Xi is tall" such that v(-, Pc> Pi) = V(Pi. I), i = 1, 2, ... , and v(P;) < V(Pi_I).
The values of the heights hi of Xi' all i, are derived as above. For such hi and with Pi = v(Pi),
we obtain f(P;) = 2-i.1 and f(npi) = 1 - 2-i-l.
4. Other values of f can be obtained from expression based on the truth of.., Pi - > Pj'
as it corresponds to f(P;) + f(P) = 2-i-1 + 2-j-l. Further values are obtained from expressions like
«.., Pi->Pj) -> PJ -> p. whose true value is f(Pi) + f(Pj) + f(pJ + f(pJ, etc.
Appropriate sequences of implications based on propositions Pi can be constructed in order
to obtain any f value. For instance, let us construct the sequence to obtain f = 0.65625. As
0.65625 = 0.50 + 0.125 + 0.03125, 0.65625 = f<Po) + f(pJ + f(P4), it corresponds to the true
value of the implication (.., Po- > PJ - > P4 with Pi = "Xi is tall" and the height of Xi is 11;; thus,
 45

thus, 0.65625 is the degree of truth of the proposition "P4 is at least as true as not .:£2 is at least
as true as""l Po,:" or "P4 is at least as true as 'P2 is less true than Po,:."
Finally, it is noted that f function is defined up to any strictly monotonic transformation
such that f(O) = 0 and f(1) = 1.
Thus we can state without loss of generality that f(x) = x, for
computational efficiency. By use of the previous canonical scale, v(Pj) = 2- j-l , and all other
truth values can be equated to some sequence of implications based on propositions Pj. We have
derived a reference scale for the truth value of any proposition Q by the use of the relation "Q
is as true as" where P is a sequence of implications based on propositions Pj . Thus P and Q
share the same numerical truth value.

2.3.2.3 Payoff Function

Giles [106] pointed out that the concepts of a set (fuzzy or not) are closely related to that
of a property: any set determines a property which belongs to the set, and any property
determines a set which contains all objects with the property. For example, to the property
"tall," there corresponds the set of all tall people (objects). Then any statement about sets can
be translated into a corresponding statement about properties, and vice versa. The concepts of
the set and property thus appear to be equivalent. Nevertheless, the concept of a property is
actually less abstract and closer to common usage than that of a set. Thus the analogue of degree
of membership in a fuzzy set can be equivalent to that of the • degree of possession" of the
corresponding property. For instance, the degree of membership of John in the set of all tall
people may be identified with the degree of which John has the property, "tall."
There is a third notion related to these two concepts. This is "degree of truth" - for the
previous example, the degree to which John is tall with the "degree of truth" of the statement,
"John is tall." Every assignment of a degree of membership in a set or of a degree of possession
of a property can be equally well-regarded as an assignment of a degree of truth to a
corresponding statement. Indeed, we can hardly contemplate intermediate (other than 0 and 1)
grades of membership in a set unless we are also prepared to consider intermediate (other than
true and false) degrees of truth for a statement. Conversely, should we succeed in the task of
attaching a well-defined meaning to intermediate degrees of truth, we would at the same time
have given a meaning to intermediate grades of membership and intermediate "degrees of
possession" of a property.
Once we have attached a clear meaning to statements with other than classical truth values,
we can deduce how one should reason with such statements: the appropriate logic is determined
 46

as soon as a clear understanding of the meaning of the statements with which it has to operate
has been reached. Thus the problem of the interpretation of grades of membership cannot be
considered independently of the problem of fuzzy reasoning itself: it is part and parcel of the
latter; a solution of one is a solution of both.
Giles further pointed out that the "assertions" concept, instead of the statements
themselves, is fundamental, and the meaning of an assertion can be identified by its payoff
function in terms of the decision theory. For example, consider the assertion" John is mil." The
meaning of the fuzzy statement "John is tall" is that information specific to this fuzzy statement
which in conjunction with one's state of belief regarding the relative possibility of different world
states (here the relevant is the height of John in these various possible states) will allow him to
decide whether (and how willingly) to assert "John is tall."
Now, let us use this example, "John is mil," to discuss how to construct the degree of
membership from the n assertions" via the payoff concept.
Usually, one who asserts "John is tall" will receive approval from his peers to an extent
dependent primarily on the height of John. In other words, the payoff of the assertion in any
world state is approximately a function f(h) of the height h of John in that state. Insofar as this
is the case, we can represent the payoff function of the assertion by the graph of the function f.
Clearly, the utility will increase with height from some negative initial value to positive values.
For example, the graph might be as shown in Figure 2.10. In this figure the payoff value f(h)

utility f (xl truth value

10

5 1.0

o
140 150 160 200
John 0.5

-5

-10 0.0

Fiqure 2.10 Payoff function of the assertion "John is tall" [106]

 47

may be described as the degree of willingness of an agent to assert "John is tall." In particular,
a positive (negative) value for the payofff(h) indicates that the agent will (won't) choose to assert
"John is mIl" when he knows John's height h. So one can tell where the graph crosses the axis
just by asking for the least height at which John would be described as tall.
We can then determine other points on the graph of f(h) by offering bribes. For instance,
the graph suggests that the agent would be willing to assert "John is mll" even if h = 170,
provided he was offered as a reward a prize worth 5 units of payoff. Actually, this argument
is not quite correct, since when two events occur together their payoffs do not necessarily
increase. Fortunately, one can flip a fair coin. If it shows a head, he will get the reward.
Otherwise, he will be obligated to assert "John is mil." Therefore, if he agrees to this proposal,
provided that he was sure John's height was at least 170cm, then f(170) > -5 is obtained. By
using a lottery instead of the coin, one can similarly determine all other negative ordinates on the
graph. On the other hand, the positive ordinates are also found by using a penalty instead of a
reward.
It is noted that the previous discussion of the payoff function for the assertion "John is
1all" has been simplified in the following two respects:
I. It is only approximately true that the payoff is a function of height, according to the
following two reasons:
a) In a normal society the term "tall" does not solely represent geometrical height. For instance,
a child is more likely to be considered tall than an adult of the same height.
b) Even the assertion that the payoff depends only on the nature of John must be qualified.
n. The meaning of an assertion, as represented by its payoff function, is determined by
the society in which the assertion is used, and does not depend on the particular person who
makes the assertion.
As a matter of fact, an assertion can be brought into standard form by a scaling and a
shift. Here let us call the payoff function of the resulting standardized assertion "the truth
function of the asserted fuzzy statement." The function is then the truth function of the sentence.
Truth functions of fuzzy statements thus coincide with payoff functions of standardized assertions
except that the extreme values 0 and 1 are necessarily attained in the latter case. Therefore, the
graph of the truth function of the asserted fuzzy statement may be obtained from the graph of the
payoff function of the assertion simply by changing the calibration on the y-axis. From Figure
2.10, the degree of the truth function for the fuzzy statement "John is mll" can be obtained, by
reading the right hand scale. The degree of membership of John in the fuzzy set of tall men is
 48

then identified with the truth value of the fuzzy statement "John is 1all."
For the payoff function, the following two significant points should be noted:
(1) It is clear that for any practical assertion, one value must be positive and the other negative.
Otherwise, the assertion would either always or never be made, so that it would be useless for
conveying the beliefs of a speaker.
(2) A scaling of the payoff function (multiplying it by a positive scale factor) corresponds to a
change in emphasis of the assertion. For example, the payoff function for "John is tlIJl" may be
taken to differ only by a scale factor (> 1) from that for the less emphatic "John is tlIJl.·
Turning this observation around, one should regard any assertions a and a', whose payoff
functions differ only by a positive scale factor, as assertions of the same fuzzy sentence except
more (or less) emphasized.

2.3.2.4 Other Examples

Besides approaches presented in Sections 2.3.2.1, 2.3.2.2 and 2.3.2.3, there are several
other considerations and approaches. Rapoport, Wallsten and Cox [213] dealt with the synthesis
of subjective data for deriving membership functions. They provided direct (magnitude
estimation) and indirect (graded pair-comparison) concepts to establish membership functions for
probability phrases such as probable, rather likely, very likely, very unlikely, and so forth.
Eliciting membership function is considered as eliciting utility function in the decision theory.
Bandemer [7] developed a fuzzy counterpart to regression analysis methods accounting for
fuzzy observation. The fuzzy (functional) relations were then derived either directly, or via
families of parametrized functions.
Pedrycz [203] solved an identification problem in terms of fuzzy relational equations from
input/output fuzzy sets. Based on the clustering technique, Pedrycz proposed a general
methodological scheme which includes the following three steps: (i) structure determination, (ii)
parameter determination, and (iii) model valuation.
Bezdek and Hathaway [15] proposed the relational hard c-means (HCM) algorithm for the
classification of objects when information about pairwise relationship between these objects is
available. Their approach can be used to derive membership functions.

2.4 Fuzzy Decision and Operators

In a decision-making process, the principal ingredients include a set of
alternativesldecisions, a set of constraints on the choice between different alternatives, and a
 49

performance/objective function which associates with each alternative the gain (or loss) resulting
from the choice of that alternative. When this decision-making process is used in a fuzzy
environment, Bellman and Zadeh [12] proposed that the symmetry between goals and constraints
is the most important feature. A symmetry erases the differences between them and makes it
possible to relate in a relatively simple way the concept of a decision to those of the goals and
constraints of a decision process.
With this property of symmetry, Bellman and Zadeh proposed that a fuzzy decision is
defined as the fuzzy set of alternatives resulting from the intersection of the goals/objectives and
constraints.

2.4.1 Fuzzy Decision

Assume that objective(s) and constraints in an imprecise situation can be represented by fuzzy
sets. For an illustration, suppose that we have a fuzzy goal G and a fuzzy constraint C in a
decision space X expressed as follows:

G: x should be substantially larger than 10 , with

p-o(x) = [1 + (x-l 0)·2]"1 if x ;;:: 10

=0 if x < 10; and

C: x should be in the vicinity of 15,

Then, with the assumption of the symmetry we may make decisions which satisfy both the
constraint "and" the goal. That is: G and C are connected to another by the operator "and"
which corresponds to the intersection of fuzzy sets. This implies that in the example the
combined effect of the fuzzy goal G and the fuzzy constraint C on the choice of alternatives may
be represented by the intersection GnC, with the membership function (see Figure 2.11):

(2.68)

Then Bellman and Zadeh [12] proposed that a fuzzy decision may be defined as the fuzzy set of
alternatives resulting from the intersection of the goals and the constraints. That is: the decision
D = Gn C is a fuzzy set resulting from the intersection of G and C, and has its membership
 50

function J.l.o = J.l.G n J.l.c (see Figure 2 .11).

I" I"c

1 ------ - \ -

o x

Figure 2 . 11 The relationship of G, C and 0

A maximizing decision then can be defined as follows:

=0 elsewhere. (2.69)

If J.l.o(x) has a unique maximum at XM, then the maximizing decision is a uniquely defined crisp
decision which can be interpreted as the action which belongs to all fuzzy sets representing either
constraints or objective(s) with the highest possible degree of membership .
Suppose that the goals are defined as fuzzy sets G t , G2, ... , Gk in Y = {y} while the
constraints C t , C2 , ... , C m are defined as fuzzy sets in X = {x} . Now, given a fuzzy set G j in
Y, we can then find a fuzzy set G' j in X which induces Gj in Y. Specifically, with y = f(x) the
membership function of G'; is given by the equality:

(2.70)

The decision D, then, can be expressed as the intersection of G't, G'2' .. . , G\, and C t , C2, .. . ,
Cm. By using Equation (2.70), we can then obtain the following :

(2.71)

where f: X - > Y. In this way, the case where the goals and the constraints are defined as fuzzy
sets in different spaces can be reduced to the case where they are defined in the same space. The
 51

maximizing decision xM as shown in Equation (2.69) can then be obtained.

From the above analysis, it is obvious that the decision DM is obtained by using the max-
min operator. However, except for the max-min operator, the decision DM can also be obtained
by using oth~r operators. In the next section, we will discuss some properties of the max-min
operator. Other operators will be discussed in Section 2.4.3.

2.4.2 Max-Min Operator

Zadeh [274] proposed the max and min operators to define intersection and union of fuzzy
sets. Mathematically, Goguen [108] first considered the order structure of fuzzy sets and the
properties of the max and min operator. The most complete arguments were provided by
Bellman and Giertz [13] from a logic point of view, interpreting the intersection as "logical and,"
the union as "logical or" and the fuzzy set A as the statement. For an axiomatic justification,
Bellman and Giertz considered: two statements, S and T, for which the truth values are I'(S) and
I'(T), respectively. I'(S) and I'(T) E [0, 1]. The truth values of the "and" and "or" combination
of these two statements are I'(S and T) and I'(S or T), which belong to [0, 1] and are interpreted
as the values of the membership functions of the intersection and the union, respectively, of S
and T statements.
Bellman and Giertz assumed that whenever truth values have been assigned to (arbitrary)
statements S and T, then two real-valued functions f and g provide us with truth values for "S
and T" and "S or T" such that:

I'(S and T) = f[JL(S), I'(T)] (2.72)

I'(S or T) = g[JL(S), I'(T)]. (2.73)

They feel that the following restrictions are reasonably imposed on f and g. The original
statements are shown as follows:
(i) f and g are nondecreasing and continuous in both variables. Our willingness to accept
"S and T" or to accept "S or T," should not decrease if our willingness to accept S(or T)
increases. With suitably small changes in the level of acceptance of S or T, we should be able
to change the truth value of "S and T," or of "S or T," by an arbitrarily small amount.
(ii) f and g are symmetric, i.e., f(x, y) = fey, x) and g(x, y) = g(y, x). There is no
reason to assign different truth values to "S and T" and "T and S."
(iii) f(x, x) and g(x, x) are strictly increasing in x. If I'(Sl) = I'(Sz} > I'(S3) = I'(S,v,
we should be more willing to accept "Sl and S2" than to accept "S3 and S4.·
 52

(iv) f(x, y) :s; min {x, y} and g(x, y) C!: max {x, y}. Accepting"S and T" requires
more, and accepting "S or T" requires less than accepting S or T alone. Thus p.(S and T) :s;
p.(S) and so on.
(v) f(l, 1) = 1 and g(O, 0) = O. If both Sand T are completely accepted, we must
accept "S and T" completely as well, and if S and T are both completely rejected, then we must
also reject "S or T."
(vi) Logically equivalent statements have equal truth values. To illustrate with an
example, the statement:

is logically equivalent to:

We have no reason to be more willing to accept one of these statements over the other.
Condition (ii) is another example of this principle.
Bellman and Giertz then proved mathematically that

p.(S and T) = min {p.(S), p.(T)} (2.74)

p.(S or T) = max {p.(S), p.(T)}. (2.75)

The min and max operators also have been used in other fields, such as utility theory, game
theory and so on, for a long time. Sometimes, the rationality of max and min operators may be
reasonable for the human decision-making process. The decision maker may like to make a
decision which maximizes the minimal possible return, or which minimizes the maximal possible
penalty (or regret). Another good feature of the max and min operators is their computational
tractability and simplicity. This is very important in practice. Freeling [93] pointed out that
there are very strong reasons for the max-min operators to apply:
1) The computation of the fuzzy expected utilities is very simple.
2) The use of them allows the full machinery of fuzzy sets to be applied to the problem.
3) There is a very appealing interpretation of the fuzzy decision analysis as a multiple-level
sensitivity .
 53

2.4.3 Compensatory Operators

Even though the max-min operator has so many good properties, it is, however, not
omnipotent. The following shows the other side of the argument.
Yager [267] showed that if the information with respect to membership grades is expected
in the form of an ordinal preference function over the membership grades of the fuzzy subsets,
then Zadeh's max and min operators are the only meaningful definitions of union and
intersection. Furthermore, if the preference information is expressed in the form intensity
information over each fuzzy subset individually, then Bellman and Zadeh's soft union and
intersection V&ouc<x) = I'o(x) + l'C<x) - l'o(x)l'C<x) and l'onC<x) = l'o(x)l'C<x), respectively)
becomes a meaningful way to combine fuzzy subsets without losing the information. However,
if all we have is ratio information over each of the fuzzy subsets individually, then min and max
are not meaningful at all.
Thole, Zimmermann and Zysno [245] conducted an experiment testing the suitability of
minimum and product operators (l'onC<x) = l'o(x)l'c(x» for conjunction (here, "and" operator).
The result, in contrast to the findings of Oden's studies [184] and the min operator, was a slight
superiority, but was not perfect for the intersection operation in human categorizing processes.
However, Thole et al. stated [245]:

"The fact that the empirical values have a tendency to exceed those predicated by means
of the min-operator may be considered as a matter of no importance. However, from a slightly
different point of view, the occurrence of rather systematic deviations is an experimental
phenomenon, and possibly a lawful one. For it is conceivable that human beings use some kind
of "compensation" when combining fuzzy sets in the sense of "and." This means that in rating
objects with respect to a composite attribute they do not process the relevant information as if
they were choosing the smaller of two grades of membership, but proceed internally as if they
were using the smaller one only as an orientation and then modifying it in the direction of the
higher value."

Therefore, searching compensatory operators for "and· and • or" operations has become one of
the most important research fields in the empire of fuzzy set theory. Based on Chen and Hwang
[BM3], we summarized the existing various compensatory min and max operators in Table 2.3.
Zimmermann and Zysno [299] made a further experiment to test the minimum, maximum,
arithmetic mean and geometric mean operators. The results showed:
1) People use averaging operations when making judgments or evaluations resulting in
membership values between minimum and maximum.
2) The geometric mean is an adequate model for human aggregation of fuzzy sets when a
compensatory effect exists.
 54

Tabla 2.3 A taxonomy of compensatory min and max operators

COMPBHSATORY HZ. OPBRATORS:

1. Alqebraic Product: ~D(x) = ~G(x)~c(x)

2. Bounded Product: ~D(X) = max (0, ~G(x)+~c(X)-l)

3. Hamacher's Kin Operator (r-operator): for r E [0, 1]

~D(X) =

4. Yaqar's Kin Operator: for q ~ °

~D(X) = 1 - min {1, [(l-~G(x»q + (l-~c(x»q]l/q}

5. Dubois and Prade's Kin Operator: for r E [0, 1]

~D (x) =

6 . •erners's "Puzzy ADd" Operator: for r E [0, 1]

~D(x) = r min«~G(x), ~c(x» + (l-r) (~G(x)+~c(x»/2

COKPENSATORY MAX OPERATORS:

1. Alqebraic Sum: ~D(X) = ~G(x)+~C(X)-~G(X)~C(X)

2. Bounded Sum: ~D(X) = min [1, ~G(x)+~c(x)]

3. Hamacher's KaX Operator: for r E [0, 1]

[(l-r)~G(x)~c(x)+r(~G(x)+~c(x»]

[r+~G(x)~c(x)]

4. Yaqer's Kax Operator: for q ~ 1

~D(X) = min {1, [(~G(x»q + (~c(x»q]l/q}

5. Dubois and Prade's Kax Operator: for r E [0, 1]

{~G(X)+~c(X)-~G(X)~c(x)-min[l-r'~G(x),~c(X)]}
~D (x) =

6 . •erners's "Puzzy Or" operator: for r E [0, 1]

~D(X) = r max«~G(X)' ~c(x» + (1-r)(~G(x)+~c(x»/2

 55

3) Men use operators other than "and" and "or.·

Thus, Zimmermann and Zysno provided a more general operator:

(2.76)

where r E [0, 1] is the degree of compensation between min and max operators. I'Gnc(x) and
J.LouC<x) are the results from using any (compensatory) min and max operators (as shown in
Table 2.3), respectively.
Based on the recent development, the union and intersection operators should come from
a class of triangular norms associated with certain modem algebras. These triangular norms (t-
norms) share the following properties: (1) T(I,J.Lo(x» = J.Lo(x) for any J.Lo(x) in [0, 1); (2)
monotonicity; (3) commutativity; and (4) associativity. T(J.Lo(x), J.LC<x» is affiliated with the
interaction operator. A co-norm, S(J.Lo(x), J.LC<x», is affiliated with the union operator, and its
properties are identical to those of T(J.Lo(x), J.Lc(x» with the exception of (1) which is replaced
with S(O, J.Lo(x» = J.Lo(x) for any J.Lo(x) in [0, 1). T<I'G(x), J.Lc(x» s min (J.Lo(x), J.Lc(x» and
S(J.Lo(x), J.Lc(x» ~ max (J.Lo(x), J.Lc(x». These are not only Zadeh's original operators, but also
are the operators that we would obtain in probability if G and C are overlapped (maximally
dependent). Furthermore, T(J.Lo(x), J.LC<x» ~ max (J.Lo(x), J.LC<x» and S{J.Lo(x), J.LC<x» :s;; min
(1, J.Lo(x)+J.Lc(x», which are operators called "bounded sum" pair in fuzzy set theory and also
corresponding probability to conjunction and disjunction when two events are maximally
exclusive [BM47). In Table 2.3, compensatory min and max operators, except Werners's,
belong to t-norm and t-conorm, respectively.
Finally, among these various operators, how can we select a suitable operator to obtain
decisions? Zimmermann [BM64) proposed the following eight rules to justify a suitable operator
for a particular fuzzy decision problem:
1. Axiomatic stren&th. An operator with less axiomatic restrictions is better.
2. Empirical fit. An operator must be an appropriate model of real system behavior which can
normally only be proven by empirical testing.
3. Adaptability. An operator should be dependent on the context and the semantic interpretation.
4. Numerical efficiency. An operator should be computationally efficient.
5. Compensation. An operator should have compensation which is defined as: J.LD(x) =
J.Lo(x)· J.LC<x) , • is compensatory if a change in J.Lo(x) can be counteracted by a change in J.LC<x).
6. Rance of resultinc membership. The larger the range of resulting membership the better the
 56

operator.
7. AK&reptinK bebayior. Since the degree of membership in the aggregated set depends very
frequently on the number of sets combined, it is important that the resulting degree of
membership be nonincreasing [108].
8. Required scale level of membership functions. An operator that requires the lowest scale
level (nominal, interval, ratio, absolute) is the most preferable.

2.4.3.1 Numerical Examples for Operators

To illustrate computational procedure of various operators, let:

A = {(4, 0), (5, 0.2), (6,0.4), (7, 0.6), (8, 0.8), (9, 1.0), (l0, O)}
and
B = {(3, 0), (4, 0.5), (5, 0.7), (6, 1.0), (7,0.7), (8, 0.5), (9, O)}.

With the assumption ofr = q = r = 0.5, the results of intersection of A and B by using various
min operators are summarized below:

x 3 4 5 6 7 8 9 10

""A (x) 0 0 0.2 0.4 0.6 0.8 1.0 0

,,"B(X) 0 0.5 0.7 1.0 0.7 0.5 0 0

Min Operator 0 0 0.2 0.4 0.6 0.5 0 0

COMPENSATORY MIN OPERATORS:

Algebraic Product 0 0 0.14 0.4 0.42 0.4 0 0

Bounded Product 0 0 0 0.4 0.3 0.3 0 0

r-operator 0 0 0.16 0.4 0.45 0.42 0 0

Yager's Min Operator 0 0 0 0.4 0 0 0 0

Dubois and Prade's

Min Operator 0 0 0.2 0.4 0.6 0.5 0 0

Werners's "Fuzzy And" 0 0.13 0.33 0.55 0.63 0.58 0.25 0

For example, at x = 7 we bave:

Min Operator: P.D(x=7) = min [p.A(x=7), P.B(x=7)] = min [0.6, 0.7] = 0.6
Compensatory min operator:
 57

1. Algebraic Product: I1-D(I) = I1-A(7)I1-B(I) = (0.6)(0.7) = 0.42

2. Bounded Product: I1-D(7) = max [0, I1-A(7) + I1-B(7) - 1] = max (0, 0.3) = 0.3
3. r -operator:
I1-D(I) = I1-A(7)I1-B(7)/{0.5+0.5[p.A(7)+I1-B(7)-I1-A(I)I1-B(7)]} = 0.42/0.94 = 0.45
4. Yager's Min Operator:
I1-D(7) = 1 - min{l, [(p.A(I)O.S + (p.B(7»O.s:J lIo. s } = 1 - min (1, 2.60) = 0
5. Dubois and Prade's Min Operator:
I1-D(7) = I1-A(7)I1-B(7)/{max[p.A(7), I1-B(7), 0.5]} = 0.42/{max[0.6, 0.7, 0.5]} = 0.6
6. Werners's "Fuzzy And": I1-D(7) = 0.5 min(0.6, 0.7) + 0.5(0.6+0.7)/2 = 0.625.

With the assumption of r = q = r = 0.5, the results of union of A and B by using various max
operators are summarized below:

x 3 4 5 6 7 8 9 10

"'A(x)
"'B(x)
°0 °0.5 0.2

0.7
0.4

1.0
0.6

0.7
0.8
0.5
1.0
0
0

max operator 0 0.5 0.7 1.0 0.7 0.8 1.0 0

COMPENSATORY MAX OPERATORS:

alqebraic sum 0 0.5 0.76 1.0 0.88 0.90 1.0 0

bounded sum 0 0.5 0.9 1.0 1.0 1.0 1.0 0

Hamacher's max operator 0 0.5 0.96 1.0 0.93 0.94 1.0 0

Yaqer's max operator 0 0.5 1.0 1.0 1.0 1.0 1.0 0

Dubois and Prade's

max operator 0 0.5 0.7 1.0 0.76 0.8 1.0 0

Werners's "fuzzy or" 0 0.38 0.58 0.85 0.68 0.73 0.75

Zimmermann and Zysno's

r-operator 0 0 0.33 0.63 0.61 0.6 0 0

For example, at x = 7 we have:

Max Operator: I1-D(x=7) = max [p.A(x=7), I1-B(x=7)] = min [0.6, 0.7] = 0.7
Compensatory Max Operator:
1. Algebraic Sum: I1-D(7) = I1-A(x) + I1-B(x) - I1-A(I)I1-B(7) = 0.6 + 0.7 - (0.6)(0.7) = 0.88
2. Bounded Sum: I1-D(7) = min [1, I1-A(7) + I1-B(7)] = min (1, 1.3) = 1.0
 58

3. Hamacher's Max Operator:

P.D(7) = ((0.5)P.A(7)P.B(7)+0.5[p.A(7)+P.B(7)]}/[0.5+p.A(7)P.B(7)]
= 0.86/0.92 = 0.93
4. Yager's Max Operator:
P.D(7) = min{l, [(p.A(7»O.S + (P.B(7»O.s:J lIo.s } = min (1, 2.60) = 1.0
5. Dubois and Prade's Max Operator:

{p.A(7)+p.B(7)-p.A(7)p.B(7)-min[l-O.5, 0.6, 0.7]}

P.D(7) = = 0.38/0.5 = 0.76
{max [0.5, 1-P.A(7), 1-P.B(7)]}

6. Werners's "Fuzzy Or": P.D(7) = 0.5 max(0.6, 0.7) + 0.5 (0.6+0.7)/2 = 0.675.

As to Zimmermann and Zysno's operator of P.D(x) = p.GnC<x)l-rp.GuC<x)f, we assume that r =

0.5, p.GnC<x) = P.A(x)P.B(x) and p.GuC<x) = {I - [l-P.A(x)][l-P.B(x)]}. Then, at x = 7 we have
the following result:

2.5 Fuzzy Arithmetic

In the previous section, we have discussed some major operations for handling fuzzy sets
and obtaining decisions. However, those operators are not enough to calculate all the situations.
For example, they cannot solve an equation involving fuzzy numbers which is always composed
of addition, subtraction, multiplication and division. Equations involving fuzzy numbers are the
most important ingredients in mathematical programming problems and many other fields. Thus,
in this section we will discuss the operations of addition, subtraction, multiplication and division
of fuzzy numbers. More detailed discussion can be found in Kaufmann and Gupta
[BM31][BM32], and Chen and Hwang [BM3].

2.5.1 Addition of Fuzzy Numbers

Assume that A and B are two fuzzy numbers. The addition of X and Y can be
accomplished by using a-level cut and max-min convolution.
I. a-level cut. By use of the concept of confidence intervals, the a-level sets of X and Y are
 59

Xa = [XaL , xaU] and Ya = [YaL , YaU] where xaL , xau, YaL and Yau E R. Then the result Z of the
addition of the fuzzy numbers X and Y can be defined by the a-level sets. That is:

(2.77)

for every a E [0, 1]. Equation (2.77) obviously shows that the lower bound of the resulting fuzzy
set Z is the sum of the lower bounds of the fuzzy numbers X and Y, and the upper bound of the
resulting fuzzy set Z is the sum of the upper bounds of the fuzzy numbers X and Y.

Example 6. Assume the membership functions of fuzzy numbers X and Y are (see Figure 2.12):

0 if x ::;; 0 or x > 4
Jlx(x) = 1 x/2
(4-x)/2
if 0 < x ::;; 2
if2 < x ::;; 4

ify::;;30ry> 11
= J~Y-3)/5 if 3 < Y ::;; 8
1
Jly(Y)
(11-y)/3 if8<y::;;11.

Then, by using a-level cut, we have Jlx(x) ;;:: a and Jly(y) ;;:: a for a E (0, 1]. That is:

xl2 ;;:: a and (4-x)/2 ;;:: a => 2a::;; x ::;; 4-2a.

Therefore, we have Xa = [xaL , xaU] = [2a, 4-2a]. Similarly, we have Y a = [YaL, YaU]
[3+5a, 11-3a]. Then:

Za = Xa (+) Ya = [(2a)+(3+5a), (4-2a)+(11-3a)] = [7a+3, IS-Sa].

Now, let Za = [zaL , zaU] = [7a+3, IS-Sa] and we have zaL = 7a + 3 and zau = 15 - Sa, or
(zaL - 3)/7 = a and (15 - zaU)/5 = a, where a E (0, 1]. Finally, we obtain the following
membership function for the fuzzy number Z:

>
1~Z-3)/7
if z ::;; 3 or z 15

Jlz(z) = if 3 < z ::;; 10

(15-z)/5 if 10 < z ::;; 15.
 60

ll. Max-min convolution. By using the extension principle, the result Z of the addition of the
fuzzy numbers X and Y can be defined as:

Z(z) = max {min £Jtx(x), JLy(y)]} (2.78)

z=x+y

where x, y and Z E R.

IL()
x Y X(+)Y

XjY
-16 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10 12 14 16

Figure 2.12 The fuzzy numbers of X (+) Y and X (-) Y

Example 7. Assume that the fuzzy numbers X and Yare:

X x 1 2 3 4 5 6

0.1 0.5 1.0 0.7 0.1 0.0

Y Y 1 2 3 4 5 6 7

lLy (y) 0.0 0.1 0.7 1.0 0.8 0.5 0.0

By using Equation (2.78), we have:

JLz(z=6) = max [min(O.I, 0.8), min (0.5,1.0), min (1.0,0.7), min (0.7,0.1),

min (0.1,0.0))
= max [0.1, 0.5, 0.7, 0.1, 0.0] = 0.7
 61

where 6 =I +5 =2 + 4 = 3 + 3 =4 + 2 = 5 + I. Other possible values of z are

summarized in Table 2.4. From Table 2.4, we can easily and systematically obtain the resulting
fuzzy number Z. For example, the value of Pz(z=6) is the maximum of the diagonally blocked
numbers which are the minimum values of the corresponding row and column values. Thus we
have the following solution:

Z z 2 3 4 5 6 7 8 9 10 11 12 13

p.z(z) 0.0 0.1 0.1 0.5 0.7 1.0 0.8 0.7 0.5 0.1 0.1 0.0

Table 2.4 Results of X (+) Y for Example 7

X Y Y 1 2 3 4 5 6 7
x P. (x) p.(y) 0.0 0.1 0.7 1.0 0.8 0.5 0.0

1 0.1 (z=2) 0.0 0.1 0.1 0.1 0.1 0.1 0.1

2 0.5 (3) 0.0 0.1 0.5 0.5 0.5 0.5 0.0

-
3 1.0 (4) 0.0 0.1 0.7 1.0 0.8 0.5 0.0

4 0.7 (5) 0.0 0.1 0.7 0.7 0.7 0.5 0.0

5 0.1 (6) 0.0 0.1 0.1 0.1 0.1 0.1 0.1

6 0.0 (7) 0.0 0.0 0.0 0.0 0.0 0.0 0.0

(z=8) (9) (10) (11) (12) (13)

2.5.2 Subtraction of Fuzzy Numbers

Similar to the addition operation, we have the following two definitions of substraction of
fuzzy numbers:

I. a-level cut. By using a-level cut, we can define the resulting fuzzy number Z of the
subtraction of the fuzzy numbers X and Y as:

(2.79)

for every a E [0, I]. The calculation is quite similar to the addition operation.
 62

Example. Consider Example 6. Obviously, we have Z.. = [5a-1I, 1-7a). The membership
function of the fuzzy number Z is then (see Figure 2.12):

>
~z
if z S -11 or z 1
p.z(z) = { + 11 )/5 if -11 <z S -6
(l-z)17 if-6<zsl.

n. Max-min convolution. By using the extension principle, we can also define the subtraction
operation as:

p.z(z) = max {min Utx(x), ~y(y)]} (2.80)

z=x-y

= max {min [~x(x), ~y(-y)]}

z=x+y

= max {min [~x(x), ~.y(y)]}.

z=x+y

As we can see, the subtraction is equivalent to the addition of image of Y to X.

Example. Consider Example 7. Now, let Z = X (-) Y. The calculation will be similar to the
addition operation. Thus we can directly compute Table 2.5.

Tabla 2.5 Results of X (-) Y for Example 7

(z=O) (-1) (-2) (-3) (-4) (-5) (-6)

X Y Y 1 2 3 4 5 6 7
x ,,(x) ,,(y) 0.0 0.1 0.7 1.0 0.8 0.5 0.0

1 0.1 0.0 0.1 0.1 0.1 0.1 0.1 0.1

2 0.5 (z=l) 0.0 0.1 0.5 0.5 0.5 0.5 0.0

3 1.0 (2) 0.0 0.1 0.7 1.0 0.8 0.5 0.0

'---
4 0.7 (3) 0.0 0.1 0.7 0.7 0.7 0.5 0.0

5 0.1 (4) 0.0 0.1 0.1 0.1 0.1 0.1 0.1

6 0.0 (5) 0.0 0.0 0.0 0.0 0.0 0.0 0.0

 63

From the above table, we can easily and systematically obtain the resulting fuzzy number Z. For
example, the value of ILz(Z=O) = 0.7 is the maximum of the diagonally blocked numbers which
are the minimum values of the corresponding row and column values. Thus we have the
following solution:

Z z -6 -5 -4 -3 -2 -1 0 1 2 3 4 5

p.z(z) 0.1 0.1 0.5 0.8 0.7 1.0 0.7 0.7 0.1 0.1 0.0 0.0

2.S.3 Multiplication of Fuzzy Numbers

For the potential sign effect, we first assume that p.x(x) = 0 for x < 0 and ILy(Y) = 0 for
y < O. Then, as previously discussed, we have the following two definitions of the
multiplication of fuzzy numbers:

I. a-level cut. By using a-level cut, we can define the resulting fuzzy number Z of the
mUltiplication of the fuzzy numbers X and Y as:

(2.81)

for every a E [0, 1].

Example. Consider Example 6. Xa = [2a, 4-2a] and Ya = [3+Sa, 11-3a]. By using

Equation (2.81), we obtain:

Za = [(2a)(3+Sa), (4-2a)(1l-3a] = [10aZ+6a, 6aZ-34a+44]

which is not a simple triangular fuzzy number. The membership function of the fuzzy number

I:
Z is then (see Figure 2.13 also):

ifz~Oorz>44

p.z(Z) = + (36+40z) /20

If2 if 0 < Z ~ 16
34 - (100+24z) If2 /12 if 16 < Z ~ 44.
 64

o
o 5 10 15 20 25 30 35 40 45

Figure 2.13 FUZZy number Z x (.) y

II. Max-min convolution. When using the extension principle, the multiplication operation
becomes more complex. Thus Kaufmann and Gupta [BM31] proposed the following procedure:

I. First, find Zl ( the peak of the fuzzy number Z) such that P,Z(ZI) = 1. Then we calculate the
left and right legs.

2. The left leg of p'z(z) is defined as:

p'z(z) = max {min [P,x(x), p,y(y)]}. (2.82)

xy~z

3. The right leg of p'z(z) is defined as:

p'z(z) = max {min [P,x(x), p,y(y)]}. (2.83)

xy~z

Example. Consider Example 7. We can directly compute Table 2.6. From Table 2.6, we can
obtain the peak at x = 3 and y = 4. That is p'z(12) = 1. Next, we should find the left and
right legs of the fuzzy number Z by using Equations (2.82) and (2.83). The results are:
 65

z z 1 2 ••• 5 6 ••• 8 9 .•• 11 12 13 14

J£zCz) 0.0 0.1 0.1 0.5 0.5 0.7 0.7 1.0 0.8 0.8

z 15 16 ••• 20 21 24 25 35 36 ••• 42

0.8 0.7 0.7 0.5 0.5 0.1 0.1 0.0 0.0

Tabla 2.6 Results of X C·) Y for Example 7

X Y Y 1 2 3 4 5 6 7
x J£Cx) J£Cy) 0.0 0.1 0.7 1.0 0.8 0.5 0.0

1 0.1 0.0 0.1 0.1 0.1 0.1 0.1 0.1

2 0.5 0.0 0.1 0.5 0.5 0.5 0.5 0.0

3 1.0 0.0 0.1 0.7 1.0 0.8 0.5 0.0

4 0.7 0.0 0.1 0.7 0.7 0.7 0.5 0.0

5 0.1 0.0 0.1 0.1 0.1 0.1 0.1 0.1

6 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

Note: when z = 6, we choose the maximum among the bold

numbers

2.5.4 Division of Fuzzy Numbers

Like the relation between addition and subtraction,the division operation is essentially an
extension of the multiplication operation. Thus we can define the division of fuzzy numbers as
follows:

I. a-level cut. First, by using a-level cut, we have the following result:

(2.84)

for every a E [0, 1].

Example. Consider Example 6. X", = [2a, 4-2a] and Y", = [3+5a, 11-3a]. Z", = [2a/(ll-
3a), (4-2a)/(3+5a)] or
 66

> 4/3
{~lZ/(3Z+2)
ifz S; Oorz
p.z(z) = if 0 < z S; 114
(4-3z)/(2 + 5z) if 114 < z S; 4/3.

n. Max-min convolution. As discussed in the previous section, we must first find the peak of
the fuzzy number Z =X (:) Y when the max-min convolution is used. Then the left leg is
defined as:

IJ.z(z) = max {min [IJ.x(x), IJ.y(y)]} (2.85)

x/ySz

= max {min [p.x(x), IJ.y(1Iy)]}

xySz

= max {min [p.x(x), IJ.11Y(y)]}'

xySz

The right leg of p.z(z) is defined as:

p.z(z) = max {min [p.x(x), IJ.y(y)]} (2.86)

x/y";?z

= max {min [IJ.x(x), IJ.y(1Iy)]}

xy";?z

= max {min [IJ.x(x), IJ.11Y(Y)]}.

xy";?z

From the above equations, we can see that the division operation can be completed by using the
multiplication operation.

2.S.S Triangular and Trapezoid Fuzzy Numbers

Among the various types of fuzzy numbers, triangular and trapezoid fuzzy numbers are
of the most importance. They are especially useful in solving possibilistic mathematical
programming problems. Thus, in this section we will provide some tables to show the results
of applying fuzzy arithmetic on the triangular and trapezoid fuzzy numbers.
First, the triangular fuzzy numbers can be denoted as X = (x, a, (J) and Y = (y, r, a),
where x and y are the central values (p.x(x) = 1 and IJ.y(y) = I), a and r are the left spreads, and
(3 and /) are the right spreads (see Figure 2.14). Then the results of applying fuzzy arithmetic
 67

on the fuzzy numbers X and Y are shown in Table 2.7.

Similarly, we can denote the trapezoid fuzzy numbers as X = (x)' ~, ex, (:J) and Y = (y),
Y2' r, 0) as shown in Figure 2.14. The results of applying fuzzy arithmetic on the fuzzy numbers
X and Y are shown in Table 2.8.
On the other hand, the triangular fuzzy numbers can also be denoted as X = (xm, xP, XO)
and Y = (ym, yP, yO) where xm = x (ym = y), xI' = x - ex (yP = Y - r) and XO = x + {:J (yO =
Y + 0). Then the results of applying fuzzy arithmetic on the fuzzy numbers X and Y are shown
in Table 2.9.
Similarly, the trapezoid fuzzy numbers can also be denoted as X = (x)m, x2mxP, XO) and
Y = (y)m, Y2m, yP, yO) where x)m = x) (y)m = y), x2m = X2 (Y2m = Y2)'XP = x) - ex (yP = y) -
r) and XO = X2 + {:J (yO = Y2 + 0). Then the results of applying fuzzy arithmetic on the fuzzy
numbers X and Y are shown in Table 2.10.

J..£()

o x

Figure 2.14 The triangular and trapezoid fuzzy numbers

 68

Table 2.7 Fuzzy arithmetic on triangular fuzzy numbers

X = (x, a, fj) & Y (y, r, 0)

Image of Y: -Y (-y, 0, r)

Inverse of Y: y.l = (y'l, oy·2, ry·2)

Addition: X (+) Y (x+y, a+r, fj+cS)

Subtraction: X (-) Y X (+) -y (x-y, a+o, fj+r)

Multiplication:

X > 0, y > 0: X ( . ) y (xy, xr+ya, xcS+yfj)

X < 0, y > 0: X .
( ) y (xy, ya-xo, yfj-xr)
X < 0, y < 0: X ( .) y (xy, -xo-yfj, -xr-ya)

Scalar Multiplication:

a > 0, a E: R: a ( . ) X (ax, aa, afj)

a < 0, a E: R: a ( . ) X (ax, -afj, -aa)

Division:

X > 0, y > 0: X ( : ) Y (x/y, (xo+ya) /y2, (xr+yfj) /y2)

X < 0, y > 0: X (:) Y (x/y, (ya-xr) /y2, (yfj-xo) /y2)
X < 0, y < 0: X (:) Y (x/y, (-xr-yfj) /y2, (-xo-ya) /y2)
 69

Ta~le 2.8 Fuzzy arithmetic on trapezoid fuzzy num~ers

X = (x" X2, a, (3) & Y = (YI' Y2' r, 6)

Image of Y: -y

Inverse of Y:

Addition: X (+) y

S~traction: X (-) y

Multiplication:

X > 0, y > 0: X ( . ) y (XIYI' X2Y2' xlr+Yla - ar , X20+Y2/3-/3 0)

X < 0, y > 0: X ( . ) y (XIY2' x2Y" Y2a-xlo+ao, YI/3-x2r-/3r)
X < 0, y < 0: X ( . ) y (X2Y2' XIYI' -X2o-Y2/3-/3o, -xlr-Yla +ar )

Scalar Multiplication:

a > 0, a E: R: a ( . ) X (axI' ax2, aa, a(3)

a < 0, a E: R: a ( . ) X (axI' ax2, -a/3, -aa)

Division:

X > 0, Y > 0:

X < 0, Y > 0:

X < 0, Y < 0:
 70

Table 2.9 Fuzzy arithmetic on triangular fuzzy numbers

X = (xm, xP, XO) & Y = (ym, yP, yo)

Image of Y: -y (-ym, -yo, -yP)

Inverse of Y:

AdcHtion: X (+) Y

Subtraction: X (-) y X (+) -y

Multiplication:

X > 0, y > 0: X ( . ) y (xmym, xPyP, xV)

X < 0, y > 0: X ( . ) y (xmym, xPyo , x"yP)

X < 0, y < 0: X ( . ) y (xmym, XOYo, xPyP)

Scalar Multiplication:

a > 0, a Ii: R: a ( . ) X (axm, axP, axO)

a < 0, a Ii: R: a ( . ) X (axm, axo, axP)

Division:

X > 0, y > 0: X ( : ) Y (xm/ym, xP/yO , XO /yP)

X < 0, y > 0: X ( : ) Y (xm/ym, XO /yO, x P/yP)

X < 0, y < 0: X ( : ) Y (xm/ym, xO/yP, xP/yo)

 71

Table 2.10 Fuzzy arithmetic on trapezoid fuzzy numbers

X = (X1m, ~m XP, XO) & Y = (Ylm, Y2m, y P, yo)
Image of Y: -y

Inverse of Y: y-I = (1/Y2m, l/Ylm, l/YO, l/yP)

Addition: X (+) Y (X1m+Ylm, X2m+Y2m, xP+yP, ~+yo)

subtraction: X (-) Y (x1m-Y2m , X2m -Ylm, xp-yo, ~-yP)

Multiplication:

X > 0, y > 0: X ( . ) y (x1my1m, x 2my2m, xPyP, xoyO)

X < 0, y > 0: X ( ) y. (~mylm, X1my2m, xPy°, xoyP)

X < 0, y < 0: X ( .) y (X2my2m, x1my1m, xoyo, xPyP)

Scalar Multiplication:

a > 0, a E R: a .
( ) X (ax1m, ax2m, axP, a~)

a < 0, a E R: a .
( ) X ( ax2m, ax1m, a~, axP)

Division:

X > 0, y > 0: X ( : ) Y (X1m/Y2m, ~m/Ylm, xP/Yo, xO/yP)

X < 0, y > 0: X ( : ) Y (X2m/Y2m, X1m/Ylm, ~/yo, xP/yP)

X < 0, y < 0: X ( : ) Y (~m/Ylm, X1m/Y2m, ~/yP, xP/yo)

 72

2.6 Fuzzy Ranking

Fuzzy ranking is an important topic in fuzzy mathematical programming. Consider the
constraints of the fuzzy mathematical programming: Ax s; b. If A, b and/or x are fuzzy, the
inequalities Ax S; b will become a ranking problem. Chapter 4 will discuss how to use fuzzy
ranking approaches in solving fuzzy/imprecise constraints.
Fuzzy ranking approaches have been systematically and thoroughly classified and well
presented in Chen and Hwang [BM3] as shown in Table 2.11. For detailed discussions, readers
are invited to refer to Chen and Hwang's Fuzzy Multiple Attribute Decision Making: Methods
and Applications [BM3]. Most of these approaches can be used to handle the constraints (of
mathematical programming) involving fuzzy numbers. However, we should pay attention to the
meaning and usefulness of these approaches for various possible mathematical programming
problems.
 73

Table 2.11 A taxonomy of fuzzy ranking methods [BM3]

Comparison Technique Approach

Medium Involved

Fuzzy Preference Degree of Baas & Kwakernaak

ranking relation optimality Watson, Weiss & Donnell
Baldwin & Guild

Hamming Yager
distance Kerre
Nakamura
Kolodzijezyk

a-cut Adamo
Buckley & Chanas
Mabuchi

Comparison Dubois and Prade

function Tsukamoto et al.
Delgado et al.

Fuzzy mean Probability Lee & Li

and spread distribution

Fuzzy Proportion McCahon

scoring to optimal

Left/right Jain
scores Chen
Chen and Hwang

centroid Yager
index Murakami et al.

Area Yager
measurement

Linguistic Intuition Efstathiou & Tong

expression
Linguistic Tong & Bonissone
approximation
3 FUZZY MATHEMATICAL PROGRAMMING

Fuzzy set theory was first developed for solving the imprecise/vague problems in the field
of artificial intelligence, especially for imprecise reasoning and modelling linguistic terms. In
solving decision making problems, the pioneer work came from Bellman and Zadeh [12],
Tanaka, Okuda and Asai [236][237], Negoita et al. [176][178][179], Zimmermann [296][298],
Orlovsky [192], Yager [266] and Freeling [92].
Since Tanaka et al. [236], there have been a number of fuzzy linear programming models.
In this monograph, we classify linear programming models with imprecise information into two
main classes: fuzzy linear programming and possibilistic linear programming. In this chapter we
first discuss fuzzy linear programming problems and models. Possibilistic linear programming
problems and models will be discussed in the next chapter. Dyson [79] argued that the
membership function is equivalent to the utility function when the membership functions of fuzzy
linear programming are based on a preference concept like the utility theory. But, this is not true
for the possibility distributions in possibilistic linear programming.
Some important fuzzy linear programming problems and methods are discussed in Section
3.1. In this section we introduce two major models: symmetric and nonsymmetric [304]. The
symmetric models are based on the definition of fuzzy decision proposed by Bellman and Zadeh
[12]. They assumed that objective(s) and constraints in an imprecise situation can be represented
by fuzzy sets. A decision, then, may be stated as the confluence of the fuzzy objective(s) and
constraints, and may be defined by a max-min operator. That is: assume that we are given a
fuzzy objective set G and a fuzzy constraint set C in a space of alternatives X. Then, G and C
combine to form a decision D which is a fuzzy set resulting from the intersection of G and C,
and corresponding P-D = P-G n P-c. This relation between G, C, D is depicted in Figure 3.1.
The nonsymmetric models are based on the following two approaches [304]:
(I) The determination of the fuzzy set decision.
(2) The determination of a crisp maximizing decision by aggregating the objective function, after
appropriate transformations, with the constraints.
In Section 3.2 we will present our integrated linear programming decision support system.
Finally, we will discuss some extension of the fuzzy linear programming in Section 3.3.
 75

ll-
ll-o(X)
l

o x

Figure 3.1 The relation of the fuzzy sets C, G and D

3.1 Fuzzy Linear Programming Models

In this section, five main fuzzy linear programming problems will be discussed. The
methodology for each problem is presented under the pre-assumption of linear membership
function forms, and max-min operators. Other membership function forms and operators are also
assumed in some specified cases. For simplicity, we assume that these membership functions
and operators are consistent to the decision maker's judgement and rationality in the decision
processes.
It is worthy to note that the pre-assumed membership functions of fuzzy linear
programming models are based on the preference concept, but not the possibility concept which
will be discussed in Chapter 4.

3.1.1 Linear Programming Problem with Fuzzy Resources

In order to understand as clearly as possible how and where the fuzziness is applied in linear
programming problems with fuzzy resources available, a simplified linear programming case is
demonstrated .
The Hardee Toy Company makes two kinds of toy dolls. Doll A is a high quality toy with
a $0.40 per unit profit and doll B is of lower quality with a $0.30 per unit profit. Each doll A
requires twice as many labor hours as each doll B, and the total available labor hours are 500
hours per day. The supply of material is sufficient for only 400 dolls per day (both A and B
combined). It is assumed that all the dolls the factory has made can be sold. The manager, then,
 76

wishes to maximize the total profit in production scheduling.

The manager formulates the production scheduling problem as follows:

maximize z = O.4x l + 0.3X2 (profit)

subject to gl(X) = Xl + ~ ~ 400 (materials)
g2(X) = 2Xl + X2 ~ 500 (labor hours)
Xl' ~ <:?: 0

where Xl is the amount of doll A to be produced and X2 is the amount of doll B to be produced.
After solving the above LP problem by using the graphic method (see Figure 3.2), the manager
gets the optimal solution: he produces 100 units of doll A and 300 units of doll B, and earns
$130 profit. Subsequently, the manager feels that the available material may change in some
range because he can get additional materials from suppliers, and that available time may also
change because he can ask the workers to work overtime. Therefore, he begins studying the
fuzzy set theory and introduces it into his production scheduling problem. Now the manager
decides that the profit should be substantially larger than $130 because of the increased resources.
The membership functions of the material and time shown in Table 3.1 and Figure 3.3 are
provided by the manager. He explains that the values of the membership functions mean the
degree of his satisfaction (preference) on the constraints gl (material) and g2 (labor hours). For
example, when the total amount of material used is less than or equal to 400, the constraint "gl(X)
~ 400" is absolutely satisfied. When the total amount of material used is greater than or equal
to 500, the constraint "gl(X) ~ 400" is absolutely violated, so the degree of satisfaction is equal
to O. Between 400 and 500, the degree of membership function is monotonic linear decreasing.
The membership function of the time availability can be explained in the same way. Thus the
manager considers that the only feasible solutions should be those satisfying the membership
functions of the material, labor hours, and profit at the same time. The feasible alternative with
the maximum degree of satisfaction is then chosen as the optimal solution of the production
scheduling problem.
From Figure 3.4, when the available material and labor hours are 400 and 500,
= 130 with the satisfaction level = 1, at Xl = 100 and X2
respectively, the optimal solution is z
= 300 as indicated by point b. For gl(X) = 450 and g2(X) = 550 the optimal solution is z' =
140 with the satisfaction level = 0.5, at Xl = 100 and X2 = 350 as indicated by point b'.
Similarly, for gl(X) = 500 and g2(X) = 600, the optimal solution is z" = 160 with the
satisfaction level = 0, at Xl = 100 and X2 = 400 as indicated by point b". Obviously, the
 77

optimal solutions are fuzzy and their membership values are determined by the membership
values (for example, 0, .5, 1) of constraints.

500,

optimal Solution (100,300)

z (x)
300

200

100

0 XI
100 200 30b 400 500

Figure 3.2 The graphic approach for a linear programming problem

Table 3.1 The membership functions of the material and labor hours

material gl (x) 150 200 250 300 350 400 450 500

available iJ.,i (x) 1 1 1 1 1 1 0.5 0

time g2(X) 200 250 300 350 400 450 500 550 600

available iJ.g2(x) 1 1 1 1 1 1 1 0.5 0

 78

I
I
I
0.5 ---------------~----
l ,
I
I
,
I
I

o 250 300 350 400 450 500

,
I
I

0.5 ---------------T--- I

I I
I
I
I

o 350 400 450 500 550 600

Figure 3.3 The membership functions of the materials and labor hours

600

c
500
c'

300

200

100

a a'a"

o 100 200 300 400 500

Figure 3.4 Graphic solutions of the fuzzy linear programming problem

 79

In the above example, we can see that in fuzzy linear programming, the fuzziness of
available resources is characterized by the membership function over the tolerance range
(ambiguous range). The membership function of the optimal solutions can then be constructed
by the convolution of the membership functions of constraints.
The general model of this linear programming with fuzzy resources available is formulated
as:

max z=cx
s.t. (Ax)i S bi' i = I, 2, ... , m
x~O (3.l.a)

where bi, Vi, are in [hi, bi+Pi] with given Pi. We may also consider the following fuzzy
inequality constraints:

max z=cx
s.t. (Ax)i ~ bi, i = 1,2, ... , m
x ~ 0 (3.l.b)

where ~ is called fuzzy less than or equal to. Assume that tolerance Pi for each fuzzy constraint
is known. Then (Ax)i S bi will be equivalent to (Ax)i S (bi + Bpi)' Vi, where 8 is in [0, 1].
If the membership functions of both cases are the same, then Equations (3.l.a) and (3.l.b) will
be the same problem [137]. Thus we will consider both problems as equivalent in this study.
Verdegay [250] first proved that the problem of Equation (3.1) is equivalent to the crisp
parametric programming problem. Thus we can use the parametric programming methods to
solve our first fuzzy linear programming problem. Werners [257], on the other hand, considered
that the objective should be fuzzy because of the fuzzy constraints. Both approaches are
discussed as follows.

3.1.1.1 Verdegay's Approach - A Nonsymmetric Model

For Equation (3.1), Verdegay [250] considered that if the membership functions of the

!
fuzzy constraints:

1 if (Ax)i < bi
lLi(x) = 1 - [(Axkbi]/Pi if bi S (Ax)i S b i +Pi
o if (Ax)i > bi +Pi (3.2)
 80

are continuous and monotonic functions, and trade-off between those fuzzy constraints are
allowed, then Equation (3.1) will be equivalent to:

max cx
s.t. x E Xa (3.3)

where Xa = {x I Io'i(x) ~ a, Vi, x ~ O}, for each a E [0, 1]. The a-level cut concept is based
on the previous works of Tanaka et al. [236] and Orlovski [192].
The membership functions indicate that: (1) if (Ax)i S b i then the ith constraint is
absolutely satisfied, Io'i(x) = 1; (2) if (Ax)i ~ b i + Pi where Pi is the maximum tolerance from
bi and determined by the decision maker in any systematic or nonsystematic way, then the ith
constraint is absolutely violated, Io'i(x) = 0; (3) if (Ax)i E (hi' bi + Pi); then the membership
functions are monotonically decreasing. That is, the more resources consumed, the less
satisfaction the decision maker feels.
We can then substitute membership functions of Equation (3.2) into Equation (3.3) and
obtain the following problem:

max cx
s.t. (Ax)i S bi + (1 - a)Pi' Vi
x ~ 0 and a E [0, 1] (3.4)

which is equivalent to a parametric programming, while a = 1 - (). Thus the fuzzy linear
programming problem given by Equation (3.1) can be equivalent to a crisp parametric linear
programming problem when some proper forms of membership functions of the fuzzy constraints
are assumed. It is noted that for each a, we have an optimal solution, so the solution with a
grade of membership is actually fuzzy.
For illustrating this approach, let consider the following two examples.

3.1.1.1a Example 1: The Knox Production-Mix Selection Problem

Let us consider a product-mix selection problem [BM4I][137]. Suppose that the Knox
Mix company has the option of using one or more of four different types of production
processes. The first and second processes yield items of product A, and the third and fourth
yield items of product B. The inputs for each process are labor measured in man-weeks, pounds
of material Y, and boxes of material Z. Since each process varies in its input requirements, the
 81

profitabilities of the process differ, even for processes producing the same item. The
manufacturer's decision on a week's production schedule is limited in the range of possibilities
by the available amounts of manpower and both kinds of raw materials. The full technology and
input restrictions are given in Table 3.2. Suppose that production levels in processes 1, 2, 3, and
4 are Xl' x2' x3 and x4' respectively. The problem can then be formulated as the following linear
programming problem:

max 4xI + 5x2 + 9x3 + l1x4 (profit)

S.t. gl(x) = Xl + x2 + x3 + x4 ~ 15 (man-weeks)
g2(x) = 7xI + 5x2 + 3x3 + 2x4 ~ 120 (material Y)
g3(X) = 3xI + 5x2 + 10x3 + 15x4 ~ 100 (material Z)
Xl' x2' x3 and x4 ~ O.

We then solve this linear programming problem by use of the simplex method. The final tableau
is shown in Table 3.3. The optimal solution is: x· = (5017, 0, 5517, 0) = (7.14, 0, 7.86, 0)
and z· = $69517 = $99.29. The actual resources used are 15, 73.57 and 100 units for man-
weeks, material Y and material Z, respectively.
Let us assume that the available total man-weeks and material Z are imprecise and their
maximum tolerances are 3 and 20 units, respectively. Then, the membership functions (see
Figure 3.5) of the fuzzy constraints are:

Table 3.2 The input data for the Knox product-mix selection problem

Item Man-weeks Material-Y Material-Z unit

(pounds) (boxes) profit

One item of product A

process 1 1 7 3 4
process 2 1 5 5 5
One item of product B
process 3 1 3 10 9
process 4 1 2 15 11

Total
available :S 15 :S 120 :S 100 maximize
 82

Table 3.3 The final tableau of the simplex method for the Knox
product-mix selection problem

Basic 4 5 9 11 0 0 0
variable xl x2 x3 x4 sl s2 s3 RHS

4 xl 1 5/7 0 -5/7 10/7 0 -1/7 50/7

0 s2 0 -6/7 0 13/7 -61/7 1 4/7 325/7
9 x3 0 2/7 1 12/7 -3/7 0 1/7 55/7
zj - cj 0 3/7 0 11/7 13/7 0 5/7 695/7

I if gl(X) < 15
I'I(X) = {1 - [g1(x)-15]/3 if 15 ~ gl(X) ~ 18
o if gl(X) > 18

I if g3(X) < 100

{
l'3(x) = 1 - [g3(x)-I00]/20 if 100 ~ g3(X) ~ 120
o if g3(X) > 120.

1-'1 (x)

o 10 15 18

o
--.------+~-----...".,..--
80
~
100 120
93(X)

Fiqure 3.5 The membership functions of man-weeks and material Y

constraints
 83

Then we have the following problem:

max 4Xl + 5x2 + 9X3 + l1x4

S.t. P.1(X) ~ 0:
g2(X) = 7Xl + 5x2 + 3X3 + 2~ ::;; 120
P-3(x) ~ 0:
0: E [0, 1], and Xl, X2, X3 and ~ ~ 0
or
max 4Xl + 5x2 + 9X3 + lI~

S.t. gl(X) = Xl + X2 + X3 + ~ ::;; 15+3(1-0:)

g2(X) = 7Xl + 5x2 + 3X3 + 2X4 ::;; 120
g3(X) = 3x l + 5x2 + 10x3 + 15x4 ::;; 100+20(1-0:)
0: E [0, 1], and Xl, X2, X3 and ~ ~ O.

Set 0 = 1 - 0:. The following parametric programming problem is obtained:

max 4Xl + 5x2 + 9X3 + 11~

S.t. gl(X) = Xl + X2 + X3 + X4 ::;; 15 + 30

g2(X) = 7Xl + 5x2 + 3X3 + 2x4 ::;; 120
glx) = 3x l + 5x2 + IOx3 + 15~ ::;; 100 + 200
Xl' ~, X3 and ~ ~ O.

where 0 f [0, 1] is a parameter. By use of the parametric technique and the final table of the
simplex method shown in Table 3.3, we can obtain the following results:

(10/7,0, -1/7) ( ~o ) = 100/7

200

(-6117, 1, 4/7) (~O ) = -1030/7

200

(-3/7,0, 117) ( ~O) = lIO/7.

200
 84

The final simplex table is shown in Table 3.4. Since the RHS 5017 + 10817,32517 - 103817 and
5517 + 11817, for 8 E [0, 1], are always greater than zero, the optimal solution is then: x· =
(7.14 + 1.438, 0, 7.86 + 1.578, 0), and z· = $(99.29 + 19.868). To display these results
clearly, Table 3.5 is provided for the decision maker.

Ta~le 3.4 The final parametric tableau of the simplex method

for the Knox product-mix selection problem

Basic 4 5 9 11 0 0 0
variable xl x2 x3 x4 sl s2 s3 RHS

4 xl 1 5/7 0 -5/7 10/7 0 -1/7 50/7+10fJ/7

0 s2 0 -6/7 0 13/7 -61/7 1 4/7 325/7+l03fJ/7
9 x3 0 2/7 1 12/7 -3/7 0 1/7 55/7+11fJ/7

zj - cj 0 3/7 0 11/7 13/7 0 5/7 695/7+139fJ/7

Ta~le 3.5 The solutions of the parametric programming problem

fJ z
. resources actually used
man-weeks material Y material Z

0.0 99.29 15.00 73.57 100.00

0.1 101.28 15.30 75.04 102.00
0.2 103.27 15.60 76.51 104.00
0.3 105.26 15.90 77 .98 106.00

0.4 107.25 16.20 79.45 108.00

0.5 109.24 16.50 80.92 110.00
0.6 111.23 16.80 82.39 112.00
0.7 113.22 17.10 83.86 114.00

0.8 115.21 17.40 85.33 116.00

0.9 117.20 17.70 86.80 118.00
1.0 119.19 18.00 88.27 120.00
 85

3.1.1.1b Example 2: A Transportation Problem

Consider a transportation problem with fuzzy demand [252]:

min cx

s.t. E Xu ~ bj , j E J = {I, 2, ... , n-l}

iEl

E X;j ~ 8;, i E I = {I, 2, ... , m}

jEJ

X;j ~ 0, Vi,j.

The membership functions of the fuzzy constraints, IljO, j E J, and IljO, i E I, are assumed to be
continuous and strictly monotonic, respectively. Then it can be proved that:

E Ilj-l(a) ~ E Ilt(a), a E [0, 1].

iEl jeJ

Therefore, the fuzzy transportation problem can be solved by means of the crisp parametric
transportation problem. That is:

min cx

s.t. E Xu ~ Ilj-l(a), j E J
iEl

E X;j ~ Ilj-l (a) , i E I

jeJ

X;j ~ 0, Vi andj, a E [0,1].

In order to get Ej p.;-l(a) = Ej Ilt(a) , an n-th dummy destination is introduced. Thus, the
following parametric linear programming problem is obtained:

min cx

s.t. E X;j = Ilj-l(a), j EJ = {I, 2, ... , n}

iEl

E X;j = Ilj-l(a), i EI
jeJ
 86

X;j ~ 0, Y i andj, a E [0,1].

Now, consider the following numerical illustration of the fuzzy transportation problem:

4 5 2 1 (-co, 8, 13)

6 2 4 3 (-00, 6, 9)

I 3
I 1 1 1 (-00, 5, 6)

(3, 4, 00) (3, 5, 00) (3, 6, 00) (1, 4, 00)

where the triangular shaped notation of (S1o ~, ~) is introduced for simplicity (see the following
figure).

lJ.i or IJ.j
1

(-00,8,13)

(3,4,00)

----~--------~~~----------~--~------------------~ s
o 3 4 8 13

By introducing the 5-th dummy destination with its membership function (- 00, °, 18), we obtain:

13-5a

9-3a

6-a
3+a 3+2a 3+3a 1+3a 18-l8a

The optimal solution that provides a fuzzy solution to the fuzzy transportation is:
 87

For ex E [0, 3/5]:

3+a 6a 1+3a 9-15a

9-3a

3+2a 3-3a

For a € [3/5, 12/13]:

3+a 9-9a 1+3a

-9+15a 18-18a

12-13a -6+12a

For a € [12/13, 1]:

3+a -3+4a 13-10a

3+2a -12+13a 1a-18a

6-a

For transportation problems, OhEigeartaigh [185] considered the case where the
membership functions of the fuzzy demands are triangular forms. Sometimes the decision maker
may prefer to have some particular amount of demand. More or less than this amount is less
preferred. Thus, the triangular membership function is rational. Chanas and Kulej [35] also
provided an approach to solve a fuzzy linear programming problem with triangular membership
functions of fuzzy resources.

3.1.1.2 Werners's Approach - A Nonsymmetric Model

Werners [257, 258] proposed that the objective function of Equation (3.1) should be fuzzy
because of fuzzy total resources or fuzzy inequality constraints. Similar to the previous section,
let us assume that the tolerances pj for the fuzzy resources are available and given. Then for
solving Equation (3.1), Werners first defined ZO and Zl as follows:

ZO = max cx
s.t. (Ax)j :::; bj, Vi, and x ~ 0

Zl = max cx
s.t. (Ax)j :::; bj+Pi, Vi, and x ~ o. (3.5)
 88

Thus, we can construct a continuously nondecreasing linear membership function for the
objective by use of tJ and Zl. Since the optimal solution will be in between ZO and Zl, the
satisfaction of the optimal solution will increase when its value increases. The membership
function 1./.0 (see Figure 3.6) of the objective is:

I if cx > Zl

I./.o(x) = {1- (ZI_CX)/(ZI_ZO) if ZO s; cx S; Zl

o ifcx<tJ. (3.6)

With the above membership function, we can then use the max-min operator to obtain optimal
decision. Then Equation (3.1) can be solved by solving" maxx;,o a, where a = min [I./.o(x) ,
~I(X), ••• , ~m(x)]." That is:

max a
s.t. I./.o(x) ~ a
~j(x) ~ a, Vi
a E [0, 1] and x ~ O. (3.7)

Equation (3.7) is essentially a symmetrical model which is similar to the model proposed by
Zimmermann (see Section 3.1.2.1). The difference is the membership function of the objective
function.

1-'0

o --~-------------r---------- cx

Figure 3.6 The membership function of the objective function

 89

3.1.1.2a Example 1: The Knox Production-Mix Selection Problem

Let us consider the previous example. From the solution of the crisp linear programming
problem, we can see that 120 units for the constraint of material Y will have too many idle
resources. We now assume that the constraint of material Y will be 80 units, and that the
tolerances for man-weeks, and material Y and Z are 5,40,30, respectively. Our problem then
becomes:

max z = 4Xl + 5x2 + 9X3 + 11""

S.t. gl(X) = Xl + X2 + X3 + "" ::;; 15
g2(X) = 7Xl + 5X2 + 3X3 + 2X4 ::;; 80
g3(X) = 3xl + 5X2 + IOx3 + 15"" ::;; 100

where the membership function p-j(x) for the ith fuzzy constraints, Vi, are:

1 if gl(X) < 15
P-1(X) = 1 1 - [g1(x)-15]/5 if 15 ::;; gl(X) ::;; 20
o if gl(X) > 20

P-2(x) = j 1 if g2(X) < 80

1 - [gix)-80]/40 if 80 ::;; g2(X) ::;; 120
o if g2(X) > 120

< 100
~ ~
if g3(X)
",(x) { - [g,(x)-l 00] 130 if 100 ::;; gix) ::;; 130
if g3(X) > 130.

Then, solving Equation (3.5), we obtain:

ZO = z·(O = 0) = 99.29
Zl = z·(O = 1) = 130

where 0 is the parameter in the new parametric programming problem. The symmetric linear
programming problem can then be formulated as follows:
 90

max a
s.t. JLo ~ a
Ilj ~ a, i = 1, 2, 3
a E [0, 1]

where the membership function JLo of the fuzzy objective is defined as:

~~
if z > 130
",,(x) { - (130-z)/(130-99.29) if 99.29 ~ z ~ 130
if z < 99.29.

This problem is actually equivalent to:

min 0
s.t. z = 4xI + 5x2 + 9x3 + llx4 ~ 130 - 30.710
gl(x) = xl + x2 + x3 + x4 ~ 15 + 50
g2(x) = 7xI + 5x2 + 3x3 + 2x4 ~ 80 + 400
g3(x) = 3xI + 5x2 + IOx3 + 15x4 ~ 100 + 300
Xl' x2' x3 and x4 ~ 0 and 0 E [0, 1]

where 0 = 1 - a. The solution is: x· = (8.57,0,8.93,0) and z· = $114.65 at 0 = 0.5, while

the actually used resources are 17.5, 86.78 and 115.01 for man-weeks, material Y and material
Z, respectively.

3.1.1.2b Example 2: An Air Pollution Regulation Problem

Consider an air pollution regulation problem [231]. It is intended to improve the quality
of life indicated by the purity of air in a given area, where each point is expressed by the
coordinates (xl' xz) ofa cartesian coordinate system. The pollutant's concentration was modified
so that (a) given air quality standards are strictly complied to, (b) people's desire to reach an
even better concentration rate than the adapted standard is taken into consideration, and (c) the
reduction of the emission does not entail too large a decrease of the plant's production (d) by
minimizing the reduction cost.
Because of these requirements, the fuzzy optimization technique is needed to solve the air
pollution regulation problem.
First, let us make the following assumptions:
 91

(1) Consider a single kind of pollutant which is released from several resources in the control
area.
(2) The given air quality standards are to be complied with at scattered so-called receptor-points
in the control area.
(3) There is no change in weather and climate conditions. This implies that we do not need
parameters expressing quantities like wind-speed and direction, pressure change, moisture,
temperature, etc. This can be considered in a more elaborate model. A known "transfer
function" Sj exists for each source j, j = 1, 2, ... , J, which indicates the concentration of the
pollutant at the ground level with the coordinates (Xl> xJ. This function may be of generalized
Gauss type:

where the sources' special parameters <Xj and fJj are known functions of Xl' ~. The concentration
rate at a receptor-point i ( i = 1, 2, ... , I) caused by the source j is:

where (Xli, x,.i) are the coordinates of the receptor-point i. Sji is one of the given data and
expresses the actual concentration rates.
The decision variables Ej E [0, 1] serve as an indicator for the reduction of the pollutant
at the source j, so that:

Ej is the reduction rate, and

(I-Ej)sji is the concentration rate after reduction.

Finally, assuming that the total concentration at one receptor-point is the sum of the pollutant's
concentrations that are released from all sources, we determine:

J
E (I-Ej)sji
j=1

as the receptor-point's total concentration (after reduction, if there exist a j so that Ej > 0).
Now, the first type of fuzzy restriction is "Let, for each receptor-point i, Ej (I-Ej)sji not
exceed the standard d but try to go below ei which is the desirable quantity at the receptor-point
 92

i." This can be written as:

J
~ (I-Ej)sjj::;; ej; d, Vi .
j=1

with the membership functions:

The second type of fuzzy restriction is "Ensure, for each source j, that the reduction rate does
not exceed E/ and even try to go below wr This is:

with the membership functions

where Wj and E/ E [0, 1], Vj. The objective function is to minimize ~j cjEj where cj is the total
reduction cost at the source j, assuming (or having approximately determined) that the cost
function is linear on [0, Et]. Thus, we obtain:

min ~j cjEj
S.t. ~j (l-Ej)sjj ::;; ej + 8(d - eJ, Vi
Ej ::;; Wj + 8(E/ - wJ, Vj
Ej ~ 0, Vj

where Sjj, ejo d, wj' E/, Cj are given data.

Now, let us consider the following numerical example. Suppose that the transfer functions
are:

SI(XI, xJ = exp{-[l + (Jtz-l)2]/XI}/XI , Xl > °

~(Xh xJ = exp{-[2+Jtz2/4]/XI}/XI' Xl > °
S3(XI, xJ = exp{-[l +(Xz+ 1)2]/(XI-2)}/(XI-2), Xl > 2

for the three sources. Beyond this, assume that there are only six scattered receptor points as
depicted in Figure 3.7.
 93

*3
urban area

2
@ Xl
1 2 3

industrial area recreation area

3
-1
*4 @
*5 *6

j
@ : source * : receptor point
i

Figure 3.7 The control area of the air pollution regulation problem

The control area is divided into three different parts, so that the desirable concentration rate e j
depends on the location but the maximum permitted concentration d is a constant standard over
the entire control area. ej and dare:

area d

recreation 0.42 0.5

urban 0.44 0.5
industrial 0.49 0.5

The receptor points 2, 3, and 4 are not taken into account because they already have a smaller
sum of the plant's concentration rates than the desirable rates. The data of the sources' assumed
maximum permitted reduction rate Ej , the desirable rate Wj and the reduction cost cj are:

j Cj Ej Wj

1 2 0.2 0.04
2 4 0.35 0.15
3 1 0.55 0.3
 94

Other data are shown as follows:

receptor points actual concen. desirable

and its rates (before concen.
coordinates reduction) rate
i Xi
1 X2i Sli S2i S3i ~i Sji ei
1 1 1 .3679 .1054 .4733 .42
2 3 1 .2388 .1757 .0067 .4030 .44
3 5 1 .1638 .1275 .0630 .3543 .44

4 1 -1 .0067 .1054 .1121 .49

5 3 -1 .0630 .1575 .3679 .5884 .49
6 5 -1 .0736 .1275 .2388 .4399 .42

The problem becomes:

min z = Ej cjEj
s.t. .3679E,+.1054~ ~ .0533 - .080
.0630E,+.1575~+.3679~ ~ .0984 - .010
.0736E,+.1275~+.2388~ ~ .0199 - .080
EI ~ .04 + .160
~ ~ .15 + .20
~ ~ .3 + .250
Ej ~ 0, Vj.

We first obtain: zO = .2403 and Zl = .87643. We can then establish Wemers's membership
function for the objective function and solve the fuzzy linear programming problem. The optimal
decision is: E," = .04, ~" = .00 and~" = .26.
 95

3.1.2 Linear Programming Problem With Fuzzy Resources and Objective

For the same reason as mentioned in the previous section, the following two problems are
equivalent here:

max
~

z = cx
s.t. (Ax)i ~ b i, i = 1,2, ... , m
x~O (3.8.a)
and
max z = cx
s.t. (Ax)i ;§ b i, i = 1, 2, ... , m
x~O (3.8.b)

To solve Equation (3.8), the methods proposed by Zimmermann [296] and Chanas [34]
are presented in this section.

3.1.2.1 Zimmermann's Approach - A Symmetric Model

In this approach, the goal bo and its corresponding tolerance Po of the fuzzy objective are
given initially, and also are for the fuzzy resources: bi and its corresponding tolerances Pi' Vi.
The fuzzy objective and the fuzzy constraints are then considered without difference, and their
corresponding regions can be described in the intervals [hi, bi +Pi] , Vi. Thus, Equation (3.8) can
be considered as [296]:

find x
such that cx ~ bo
(Ax)i ~ bi, Vi,
x~O (3.9)

In the fuzzy set theory, the fuzzy objective function and the fuzzy constraints are defined
by their corresponding membership functions. For simplicity, let us assume that the membership
function 1"0 of the fuzzy objective is a non-decreasing continuous linear function, and the
membership functions I-'i' Vi, of the fuzzy constraints are non-increasing continuous linear
membership functions as follows (see Figure 3.8):
 96

~.
ifcx> bo

""x) = { (b,-cx)/p, if bo-Po ~ cx ~ bo

if cx < bo-Po (3.10)

I if (Ax); < b;
1l;(X) = { 1 - [(Ax);-bJ/p; if b; ~ (Ax); ~ b;+p;
o if (Ax); > b;+p;, Vi. (3.11)

llo(X)

1 - ----------------.r------

o CX
•• • I

Il; (x)

o (Ax);

Figure 3.8 The membership functions for the fuzzy objective constraint
c x ~ So and the ith fuzzy constraint (Ax); S ii;

Zimmermann then used Bellman and Zadeh's max-min operator to solve the Equation
(3.9). Thus the optimal solution can be obtained by:

max Ilo = max { min £JJ.o(x), IlI(X), ... , p.".(x)]} (3.12)

where Ilo is the membership function of the decision space D, and Ilo = min (Ilo, Ill' ..• , p.".).
 97

If 0: = P.o, the Equation (3.9), via the Equation (3.12), will be equivalent to:

max 0:

s.t. ILo(X) = 1 - (bo - cx)/Po ~ 0:

p.Jx) = 1 - [(Ax)i - bJ/Pi ~ 0:, i = 1, ... , m

P.i(X), Vi, and 0: € [0, 1] (3.13)
or
max 0:

s.t. cx ~ b o - (1 - O:)Po
(Ax)i ~ b i + (1 + O:)Pi, Vi
x ~ 0 and 0: € [0, 1] (3.14)

where c, A, b o, Po, b i and Pi' Vi, are given initially.

Obviously, Equation (3.14) is a crisp linear programming technique. A unique optimal
solution can be obtained. It should be noted that this approach is considered as the first practical
method to solve a linear programming problem with fuzzy resources and objective.

3.1.1.2a Example 1: The Knox Production-Mix Selection Problem

Consider the above product-mix selection problem and assume that bo = 111.57 and Po
= 10. Our problem becomes:

find x
such that &>(x) = z = 4XI + 5~ + 9X3 + 11~ ~ 111.57
gl(X) = Xl + X2 + X3 + ~ ~ 15
g2(X) = 7XI + 5X2 + 3X3 + 2~ ~ 80
g3(X) = 3x I + 5x2 + 10x3 + 15~ ~ 100
Xl' X2, X3 and ~ ~ O.

where the membership functions of the fuzzy constraints are described in previous subsections
and the membership function ILo of the fuzzy objective is defined as:

if &>(X) > 111.57

",(x) ~ ~ { - [1l1.57-g,,(x)]/IO if 101.57 ~&>(x) ~ 111.57
if &>(x) > 101.57
 98

which is then equivalent to:

max a
s.t. &>(X) = 4xl+5x2+9x3+11~ ~ 111.57 - 1O(1-a)
gl(X) = Xl+X2+X3+~ S 15 + 5(1-a)
g2(X) = 7Xl+5x2+3x3+2~ S 80 + 40(1-a)
g3(X) = 3x1 +5x2+ 10x3+ 15~ S 100 + 30(I-a)
Xl> X2, X3, ~ ~ 0 and ex E [0, 1].

The solution is: x· = (8.01,0,8.50,0) and z· = $108.54 at ex = 0.70, while the resources used
are 16.51, 81.57 and 109.03 for man-weeks, material Y and material Z, respectively.

3.1.2.1b Example 2: A Regional Resource Allocation Problem

Consider a simplified regional resource allocation problem which involves the communal
farming communities (the system of kibbutzim, in Israel) [148].
It is common for groups of kibbutzim to join together as a confederation to share common
technical services and to coordinate their production. Within the confederation, the over-all
planning is carried out by a coordinating technical office. The function of the office is to plan
agricultural production of the confederation for the coming year.
The allocation problem concerns a confederation of three kibbutzim. The agricultural
output of each kibbutz is limited by the amount of available irrigable land and the quantity of
water (see Table 3.6) (which can only be imprecisely specified) allocated for irrigation by the
Water Commissioner, a national government official.

Table 3.6 Irrigable land and water capacity

Kibbutz Fuzzy Interval

Irrigable Land (acres) Water Capacity (acre feet)

1 [400, 440)' [600, 660]

2 [600, 630] [800, 840]
3 [300, 320] [375, 450]

* The fuzzy specification for the fuzzy interval is corresponding

to "should be less than 400 or not much greater than 400" and the
membership function form is defined as equation (3.11).
 99

The crops under consideration are sugar beets, cotton and sorghum. These crops differ
in their expected net return and their consumption of water. Imprecise capacity of land that can
be devoted to each of the crops is also imposed by the Ministry of Agriculture as follows:

Table 3.7 Crop data for the confederation of kibbutzim

Crop Net Return (acre feet) Maximum Capacity (acres)

($/acre) / (acre) (fuzzy interval)
Sugar beets 400 3 [600, 650].
Cotton 300 2 [500, 540]
Sorghum 100 1 [325, 350]
* The fuzzy specification for the fuzzy interval is corresponding
to "should be less than 600 or not much greater than 600" and the
membership function form is defined as equation (3.11).

The three kibbutzim of the Confederation have agreed that every kibbutz will plant a pre-
specified proportion of its available irrigable land. Nevertheless, any combination of the crops
may be grown at any of the kibbutzim.
For the coming year, the coordinating technical office has to determine the amount ofland
to be allocated to each crop in the respective kibbutzim so that the above prescribed restrictions
are satisfied. In place of taking the maximization of total net return of the confederation as an
objective, it is decided that a target value, $ 350,000, should be employed and the total net return
is required to exceed the target value. In case the target value is too optimistic, the total net
return is allowed to fall below it. The bottom line is $ 250,000. Thus:
"Total net return should be greater than $350,000 or not much smaller than $ 350,000,"
and is defined by the membership function form of equation (3.10) with the associated fuzzy
interval [250,000, 350,000].
Based on the above information, the resource allocation problem of the Confederation can
be formulated as a fuzzy linear programming problem as follows:

Objective (fuzzy):
z = 400(xll +X 12 +x13) +300(x21 +X22 + x23 )+ lOO(x31 +X32+ X33) ~ 350,000; 250,000
Constraints:
Water consumption (fuzzy):
gll(X) = 3xll + 2X21 + X31 :s; 600; 660
gdx) = 3X l2 + 2X22 + X32 :s; 800; 840
 100

g13(X) = 3X13 + 2Xz3 + X33 ~ 375; 450

Crop-land (fuzzy):
g21(X) = Xu + XI2 + X13 ~ 600; 650 (Sugar beets)
g22(X) = X21 + X22 + X23 ~ 500; 540 (Cotton)
g23(X) = X31 + X32 + X33 ~ 325; 350 (Sorghum)
Irrigable land (fuzzy):
g31(X) = Xu + XzI + X31 ~ 400; 440
g32(X) = XI2 + X22 + X32 ~ 600; 630
g33(X) = X13 + Xz3 + X33 ~ 300; 320
Land appropriation (precise):
841(x) = (X +X21+ X31)/400 = (X I2 +X22+ X32)/600
U

84Z<x) = (X12 + Xz2 + x3J/600 = (X13 + Xz3 + X33)/300

843(X) = (XI3+X23+X33)/300 = (x u +XzI+x3I)1400
Non-negativity:
X;j ~ 0, i and j = 1, 2, 3

In the above table, X;j = the number of acres of land in kibbutz j to be allocated to plant crop
i, i = 1 (sugar beets), 2 (cotton) and 3 (sorghum).
Since the membership functions of the fuzzy objective and the fuzzy constraints are defined
as equations (3.10) and (3.11), respectively, we can then obtain the following model:

max ex
s.t. 1 - (35()()()()-z)/l()()()()() ~ ex
1 - [gu(x)-600]/60 ~ ex
1 - [gdx)-800]/40 ~ ex
1 - [g13(x)-375]175 ~ ex
1 - [g2I(X)-600]/50 ~ ex
1 - [g22(x)-500]/4O ~ ex
1 - [g23(x)-325]/25 ~ ex
1 - [g31(X)-400]/40 ~ ex
1 - [g32(X)-600]/30 ~ ex
1 - [g33(X)-300]/20 ~ ex
3(XU +XzI+X31)-2(X I2 +Xzz+Xn> = 0
 101

(X12+X22+X32)-2(X13+X23+X33) = 0
4(X13 +x23 +x33)-3(xll +X21 +X31) = 0
a E [0, I], and Xij ~ 0, Vi,j.

The optimal solution is:X\l = 152.78, X"'12 = 48.61, X"'13 = 128.47, x21 = 104.94, X"'22 =
337.96, X\3 = X\l = x32 = 0, and X"'33 = 64.82 with a'" = 0.213.

3.1.2.1c Example 3: A Fuzzy Resource Allocation Problem

Consider a fuzzy resource allocation problem [172]. The problem is that there are J
resources, indexed by j E {I, 2, ... , J}, which are allocated to I tasks, i E {I, 2, ... , I}. Each
task contributes to the completion of one or more activities, indexed by k E {I, 2, ... , K}. The
quantity of resources j allocated to task i is denoted by Xji for all i and j, where Xji is a non-
negative real number, and x = (Xji) is the Jx! allocation matrix. The total consumption Yj of the
resource j is then given by

and Yj should not be larger than the available quantity hj of resource j, where hj > 0 for all j.
Accordingly, for each j E {I, ... , J} introduce a fuzzy set Aj describing the intuitive requirement
that: "The consumption of resource j should be as small as possible and not greater than or equal
to hj ." Mathematically, Aj is defined by Aj = {yj' J'iYj)} where Yj E [0, (0) for all j. It is
assumed that J'j is continuous and that:

J'j(Y/) > J'j(Y/) if Y/ < Y/ and Y/' Y/ E [0, hj)

J'j(O) > 0
J'iYj) = 0 if Yj ~ hj .

For a given activity k, it is assumed that the total effectiveness on that activity of all the
allocation is given by:

where, for each G, i, k), r jik ~ 0 is a real number describing the effectiveness of an allocation
for each Xji = I.
A fuzzy set is introduced for each activity k in order to describe the requirement that:
 102

[0, 00) for all k, where r t is the membership function of B." assuming that r t is continuous and

rt{O) ~ 0
rt{xl/) < r t {xjt2) if Xtl < Xt2 and Xtl, Xt2 ~O
miOt rt{O) < mi~ I-'.;{O).

By using the max-min operator, an allocation matrix x = {Jt.jJ is selected on the basis of
the fuzzy set D (see Figure 3.1 also) defined by D = {x, z(x)} and x = (Jt.j;) ~ 0, where z{x)
= min {mi~ I-'.;(Yj), miOt rt{Xt)}. A decision x· with the highest value z· of the membership
function z{x) is then given by the solution of maximizing z{x) where x = (Jt.j;) ~ O. That is to
solve the following problem:

max a
s.t. J.Lj( E; Jt.j;) ~ a, Vj
r t { Ej E; rjitJt.jJ ~ a, Vk
x = (Jt.j;) ~ O.

To solve the above problem, let us extend the definition of r t as rt(1\} = rt{O)(o., + 1) for
0., E [-1, 0], showing that r t is a strictly increasing function of 0.,.
It is observed that the continuity and strict monotonicity assumptions on the functions J.Lj
and r t demand the existence of inverse functions defined by the requirements that:

J.Lj(Yj) = a < => J.L/{a) = yj' yjElO, hj), aE{O, J.Lj{O)]

rt(1\} = a < => rt-I(a) = 0." o.,E[-I, 0], aE[O, rt(oo)],

It is noted that I-'.;-I{O) is not uniquely defined, while I-'.;(Yj) = 0 and Yj ~ hj.
It follows that J.Lj-1 and rt -I are respectively strictly decreasing and strictly increasing
functions. For a E [0, mi~ 1-'.;(0)] (z· s mi~ J.Lj(O», our problem becomes:

max 5
s.t. 11:;5 - E; Jt.ji ~ lI:;a - J.L/(a), Vj
ht5 - Ej Ei rjitJt.ji S hta - rt-I(a), Vk
x = (Jt.jJ ~ 0

where II:; < 0 for V j and bt > 0 for Vk are given (selected) real numbers. Thus the optimal
 103

solution of our fuzzy resource allocation problem can be determined by the following procedure:
Step 1. Start with n = 1 and a selected value al of a to solve the above problem. If o·(al) =
aI' terminate the procedure; otherwise go to Step 2.
Step 2. Choose an+l > an if o·(an) > an and choose an+l < an if o·(an) < an' and then
solve the problem. Terminate the procedure if o*(an+l) = an+l or if z· E (an' an+l) where an
< an+l' or z· E (an+l' an> where an+l < an' and the difference I an+l - an I is sufficiently
small. Otherwise replace n by (n+ 1) and return to Step 2.
Now, suppose that the following data are given (J = I = K = 2):

2 1) (00)

r
= (
Wi!)
J 0 0
(1\2)=
J 01
if Yl E [0, 1]
-y,
JLl(Yl) = 07 /2 - (3/2)Yl ifYl E (I, 2)
ifYl E [2, 00)

JL2(y2) = (~- Y2 if Y2 E [0, 2)

if Y2 E [2, 00)

rl(xl) = { 3x, if Xl E [0, 1)

2 + Xl if xl E [1,00)

,,(x,) ~ {x' if x2 E [0, 1)

-1 + 2x2 if x2 E [1, 00)

and let al = a2 = -1 and b l = b2 = 1. This problem can then be formulated as the following
linear programming problem:

max 0
S.t. -0 - Xu - x12 - sl = -a - JLl-l(a)
-0 - x2l - x22 - s2 = -a - JL2- 1(a)
o - 2x11 - x12 - s3 =a - rl-1(a)
o - x22 - s4 =a - r£l(a)
Xji -> 0, V"J,1.

The choice al = 0 in Step 1 of the algorithm and the solution of this problem gives the basic
 104

variable 0, SI' X12' x22 with the values 1, 1, 1, 1 and 0·(0) (= 1) > al (= 0). So, let a2 = 2.
The solution is then o· = 112, xll = 112, x12 = 2/3 and x22 = 3/2. Now, 0·(2)= 112 > 2 and
demand a3 < 2. The choice of a3 = (al + avI2 gives the solution (0·, xll, x12' x2V = (1,
1, 1, 1). 0·(1) = 1 shows that an optimal solution is given by z· = 1 and (xll' x12' x21' x2V'"
= (1, 2,0, 1).

3.1.2.2 Chanas's Approach - A Nonsymmetric Model

For Equation (3.8), Chanas considered that the goal bo and its tolerance Po cannot be
specified initially because of lack of knowledge about the fuzzy feasible region. Therefore, in
order to help the decision maker to determine bo and Po, Chanas [34] first solves the following
problem, and then presents the results to the decision maker to determine bo and Po:

max cx

x;;:: O. (3.15)

where the tolerances Pi for each bi is given and the membership functions of the fuzzy constraints
are assumed as Equation (3.11). Then we obtain lLi;;:: 1 - [(Ax)i) - bi1/Pi ;;:: a, Vi, which are
equivalent to (Ax)i :S; bi + (1 - a)Pi' Vi. If () = 1 - a, then we will have (Ax)i :S; bi + ()Pi.
Obviously, Equation (3.15) is equivalent to the following parametric programming:

max cx
s.t. (Ax)i:S; b i + ()Pi, Vi
x ;;:: 0 and () f [0, 1] (3.16)

where c, A, bi and Pi' Vi, are given and () is a parameter. It is easy to see that for every
admissible solution x\O) with a fixed parameter 0, the condition:

lLi(Ax\0» ;;:: 1 - (), Vi

is valid. On the other hand, for every non-zero basic solution (if Pi > 0), there exists at least
one active constraint, i, such that lLi(Ax·«()) ;;:: 1 - () (or (Ax)i = b i + ()Pi) and consequently the
common degree of satisfaction of the constraints is:
 105

(3.17)

Hence, for every fJ we obtain a solution (if one exists) which satisfies jointly the constraints with
degree I - fJ.
The optimal solutions z*(fJ) and x*(fJ) of Equation (3.16) are then presented to the decision
maker. The decision maker can now choose bo and its corresponding Po. According to these
values, we can then construct the membership function Po of the objective function as (3.10).
Since the final optimal solution will exist at x*(fJ), Po becomes:

I if cx*(fJ) > bo
{
Po(x*(fJ» = 1- [bo-cx*(fJ))/po ifbo-Po s cx*(fJ) S bo
o if cx*(fJ) < bo-Po (3.18)

which is always piecewise linear function with respect to fJ.

In Equation (3.16), the overall membership function P-c of the fuzzy constraint set is I -
fJ, by use of the min-operator. Therefore, the final optimal solution x*(fJ*) and z*(fJ*) will exist
at:

max P-D(fJ) = max { min [Po(fJ), P-c(fJ)] } = max [Po(fJ) A P-c(fJ)] (3.19)

as depicted in Figure 3.9. To illustrate this approach, we proposed the algorithm as shown in
Figure 3.10.

1-'0 and P-c

1

o 8* 1 8
Figure 3.t The intersection of 1-'0 and I-'c
 106

step 1.
set up and solve
(3.16) to
obtain z.(8), x·(8)

step 2.
Determine the goal
of the objective Determine an
appropriate value
of a = 1-8 to get ZO

no

step 3.
Set up po and
determine #J.o· (c x)

step 4.
Find 8· satisfying
#J.o· (8·) = #J. c (8·) ,
and then obtain
JJD·' z·, x·

Figure 3.10 Solution procedure of Chanas's approach

 107

3.1.2.28 Example 1: An Optimal System Design Problem

Consider designing an optimal system problem proposed by Zeleny [293]. However, we
modify it as an example of a fuzzy linear programming problem.
Zeleny has proposed that modern production systems are increasingly characterized by
novel high-productivity attributes: minimal buffers, close-ta-zero inventories, full utilization of
resources, flexibility of layouts, resource portfolios and so on. Therefore, the crisp linear
programming, with its mathematically "built-in" under-utilization of resources, single-mindedness
of purpose, and fixation on the "fixed, " does not appear to be a sufficiently rich and flexible tool
to match these novel requirements.
Here, a simplified example of an optimal design system in a fuzzy environment is
considered and solved. The constraints and the objective are fuzzified.
A small company has been producing a highly profitable decorative material for the last
several Christmas seasons. This decorative material came in two versions, x and y. Five
different component ingredients were needed: golden thread, silk, velvet, silver thread and
nylon. The prices of these inputs and their technological contributions to both x and y are given
in Table 3.8. The profit margins are $400 per unit of x and $300 per unit of y.

Tabla 3.8 Inputs and technological coefficients

Resource Technological Coefficients Price/unit

x y ($)
golden thread 4 0 30
silk 2 6 40
velvet 12 4 9.S
silver thread 0 3 20
nylon 4 4 10

In order to maintain these margins, the company does not allow "substantially more than $2600"
to be spent on the purchase of the components. If the decision maker spends $2600 to buy 20,
24, 60, 10.5, 26 units of golden thread, silk, velvet, silver thread, nylon, respectively, we can
solve the following linear programming and get the optimal production scheduling in the given
situation:
 108

max 400x + 300y (profits)

s.t. 4x S 20 (golden thread)
2x + 6y S 24 (silk)
12x + 4y s60 (velvet)
3y s 10.5 (silver thread)
4x + 3y s 26 (nylon).

By graphic technique, the solution obtained is: x· = (3.125, 2.625 ) and the optimal profit =
2437.5.
Now let us formulate the above problem into the following fuzzy linear high-productivity-
system programming model (Lai [139]):

max 400x + 300y (profits)

s.t. 4x = bi (golden thread)
2x + 6y = b2 (silk)
12x + 4y = b3 (velvet)
3y = b4 (silver thread)
4x + 3y = bs (nylon)
30b I +40b2 +9.5b 3 +20b4 + 10bs S2600 (budget)
x, y, b I , b2 , b3 , b4 , bs 2! 0

where bi' i = 1, ... ,5, are the optimal resource levels to achieve high-productivity systems with
close-to-zero inventory. The solution procedure is shown as follows:
Step 1. Suppose the maximal tolerance for the budget constraint is $200, then the parametric
programming version becomes:

max 400x + 300y

s.t.30(4x)+40(2x+6y)+9.5(l2x+4y)+20(3y)+10(4x+3y)=354x+368yS2600+2000
x, y 2! 0, and 0 E [0, 1 ].

The optimal solution x· = «2600 + 2000)/354,0) and z·(O) = (400/354) (2600 + 2000)/350.
Step 2. For simplicity, assume that 7Q = 3100 is the goal of the profit objective.
Step 3. Assume that Po = 200 and the membership function of the profit objective is linear as
Equation (3.10). Thus, one can get the following membership of the objective:
 109

3100 - 400(2600 +2(08)/354

J.Lo·(8) = 1 - = 0.189 + 1.138
200

Step 4. Set P.c = J.Lo as shown in Figure 3.11. We obtain 0.189 + 1.138 = 1 - 8 and then 8'
= 0.38. Thus, P.D· = J.Lo = P.c = 1 - 0.38 = 0.62. The optimal solution is then: x' = [2600
+ 200(0.38)]/354 = 7.56, y" = 0, and z' = $3023.73. The necessary resources are: b l ' =
30.24, b2 ' = 15.12, b3 ' = 90.72, b4 * = 0 and bs' = 30.24 and the budget is $ 2676.24.
Table 3.9 represents the differences between the two solutions of the crisp linear
programming and fuzzy linear high-productivity programming.

.189

o
o o· .72 1

Figure 3.11 The optimal solution of the fuzzy

optimal system design problem
 110

Table 3.9 The comparison of the crisp solution and fuzzy

high-productivity solution

Crisp standard system Fuzzy high-productivity system

x. = 4.125
.
x = 7.56
Y" = .2.625 Y" = 0

Profit
z " = $2437.5 z" = 3023.73

Budget = $2600 $2676.24

Resources
b l = 20 b"
I = 30.24
b 2 = 24 ~" = 15.12
b 3 = 60 b 3" = 90.72
b 4 = 20.5 b"4 = 0
b s = 26 b s" = 30.24

3.1.2.2b Example 2: An Aggregate Production Planning Problem

Consider a single objective aggregate production planning (APP) problem [145]. APP
deals with matching supply and demand of forecasted and fluctuated customers' orders over the
medium time range, up to approximately twelve months into the future. The term "aggregate"
implies that the planning is done for a few aggregate product categories. The purposes of
aggregate production planning are (1) to set up overall production levels for each product
category to meet the fluctuating or uncertain demands in the near future; (2) to set up decisions
and policies on the issues of hiring, layoff, overtime, backorder, subcontracting, and inventory
level, which means that the APP will determine the appropriate resources to be used as well.
When using any of the aggregate production planning models, it is often assumed that the
goals and the model inputs (resources and demands) are deterministic/crisp. In practice,
demands, resources and costs are usually imprecise/fuzzy. The current APP model represents
the information in a fuzzy environment where the objective function and the parameters are not
completely defined and cannot be accurately measured. The best compromise APP will balance
the cost of building and holding inventory against the cost of adjusting activity levels to meet the
fluctuations in demands. The forecasted demand in a particular period could either be satisfied
or backordered. However, the backorder must be fulfilled within the next period. The single
objective linear programming model for the APP can then be formulated as:
 111

(ex. I)

....
s.t. Wt:S; W_ (ex.2)
Wt = Wt.l + Ht - ~ (ex.3)
kPrt :s; oWt (ex. 4)
kPot :s; PtoWt (ex.5)
Prt + Pot + ~-l - Bt_l ~ Ftmia (ex. 6)
~ - Bt = ~-l - Bt_l + P rt + Pot - Ft (ex.7)
P rt, Pot, W" It, B" Ht, ~ ~ 0, Vt (ex. 8)

where the parameters and decision variables are explained in the table:

Parameters and constants:

Cpt : production cost excluding labor cost in period t ($/unit)
Cot : overtime labor cost in period t ($/man-hour)
Crt : labor cost in period t ($/man-day)
Cit: inventory carrying cost in period t ($/unit-period)
Cbt : cost to hire one worker in period t ($/man-day)
Cit : cost to layoff one worker in period t ($/man-day)
Cbt : stockout cost in period t ($/unit-period)
k : conversion factor in hours of labor per unit of production
o : regular working hours per worker per day
Pt : fraction of working hours available for overtime production
T : planning horizon or number of periods
~ : initial inventory level (units)
Wo : initial work force level (man-day)
Bo : initial backorder level (units)
Wtmu. : maximum work force available in period t (man-day)
F t : forecasted demand in period t (units)
FtmiD : minimum demand in period t (units)
Decision variables:
P rt : regular time production in period t (units)
 112

Pot : overtime production in period t (units)

Wt : work force level in period t (man-day)
It : inventory level in period t (units)
S. : backorder level in period t (units)
~ : worker hired in period t (man-day)
~ : worker layoff in period t (man-day)

The fuzzy linear programming model for the single objective APP problem has a fuzzy
goal, fuzzy work force level allowed in each period, and fuzzy demands. Equation (ex.l) is
the objective of minimizing total cost. Equation (ex.2) is a policy that sets the limit of
maximum available labor force in any period. Equation (ex.3) represents the constraint that the
work force in period t is equal to the work force in period t-l plus new hires minus layoffs.
Equations (ex.4) and (ex.5) represent the constraints that limit the regular and overtime
production to the available hours. Equation (ex.6) stands for the constraint that a production
backorder in any period is not carried for more than one period. Equation (ex.7) requires a
solution that the inventory level or backorder level in period t be equal to what it was in period
t-l plus the regular time production, overtime production and minus the forecasted demand.
Equation (ex.8) is a non-negative constraint.
Now, let us consider the following numeric data:
a) There is a six-period planning horizon with fuzzy demands of intervals [250, 275], [285,
330], [430, 500], [260, 320], [300, 350], [270, 340] units. Here, it is assumed that the trend
of future demand is optimistic in periods 1 and 2. The decision maker has the highest
satisfaction level at demand levels 275 and 330, and the lowest satisfaction level at demand 250
and 285 for periods 1 and 2, respectively. The decision maker feels that the demands in periods
5 and 6 are pessimistic, and in-between in periods 3 and 4. The minimum demands are 250,
285,430,260,300 and 270 units, respectively.
b) The production cost other than labor cost is $20 per unit. Three hours of labor are needed
for each unit produced. The regular labor force works an eight-hour day.
c) The initial work force is 100 workers. The maximum numbers of workers allowed are [90,
100] and [100, 115], for periods 1 and 2, [120, 150] for periods 3, 4, 5, and 6. The costs
associated with the regular payroll, hiring and firing are $64, $30 and $40 per worker per day,
respectively.
 113

d) Overtime production is limited to no more than 30% of regular time production. The
overtime charge is based on $15 per worker per hour.
e) There is no beginning inventory. The inventory carrying cost is $2 per unit per period.
Backorder must not be carried over more than one period. The backorder cost is $3 per unit per
period.
The formulation of this APP model is:

6
min E [20(PrI + poJ + 64Wt + 15(3PoJ + 2~ + 3Bt + 30~ + 40LJ
-
t=1
s.t. WI ~ 90, Wz ~ 100, W3 ~ 120
W4 ~ 1io, Ws ~ 120, W6 ~ 120
WI = Wo + HI - Llo Wz = WI + Hz - Lz
W3 = Wz + H3 -~, W4 = W3 + ~ - L4
Ws = W4 + Hs - Ls, W6 = Ws + ~ - L6
3Prl ~ 8W I, 3Pa ~ 8Wz, 3Pt.! ~ 8W3
3Pr4 ~ 8W4, 3PrS ~ 8Ws, 3Pr6 ~ 8W6
3Pol ~ (0.3)8WIo 3Poz ~ (0.3)8Wz
3P03 ~ (0.3)8W3, 3P04 ~ (0.3)8W4
3Pos ~ (0.3)8Ws, 3P06 ~ (0.3)8W6
Prl+ Pol + 10 - Bo ~ 250
Pa + Paz + II - BI ~ 285
Pt.! + P03 + Iz - Bz ~ 430
Pr4 + P04 + 13 - B3 ~ 260
PrS + Pas + 14 - B4 ~ 300
Pr6 + P06 + Is - Bs ~ 270
II - Bl = 10 - Bo + Prl + Pol - 275 -
Iz - Bz = II - Bl + Pa + P02 - 330
13 - B3 = Iz - Bz + Pt.! + P03 - 450
14 - B4 = 13 - B3 + Pr4 + P04 - 300
Is - Bs = ~ - B4 + PrS + Pas - 300
16 - B6 = Is - Bs + Pr6 + P06 - 270
PrI , Pot, Wto ~, Bto ~, ~ ~ 0, t = 1, ... 6
 114

where the membership functions of the fuzzy demands and work force levels are given in Figure
3.12. The membership function of work force level in period 1, for example, can be represented
as follows :

~
if W'max < 90
/'w,_ = { - (W,_ - 90)/10 if 90 S W'max S 100
if 100 < W'max.

80 90 100 110 120 130 140 150 WIIIIU

250 300 350 400 450 500

Figure 3.12 Membership functions of the fuzzy work force levels and
fuzzy demands
 115

According to Chanas's concept [34, 35], we can obtain the following equivalent constraints for
the first fuzzy resource constraint WI S 90 + 106. The other fuzzy resources of the work force
level are W2 S 100 + 156, and W3, W4, Ws and W6 S 120 + 306. For the equality
constraints of fuzzy demands, we have:

10 + Bl - 11 - Bo + Prl + Pol + Xl = 275

Xl s 275 - 275 + (250+0)6 = 256
11 + B2 - 12 - Bl + P r2 + P02 + x2 = 330
x2 s 456
12 + B3 - 13 - B2 + P r3 + P03 + x3 = 450 + 506
x3 s 706
13 + B4 - 14 - B3 + P r4 + P04 + x4 = 300 + 206
X4 s 606
14 + Bs - Is - B4 + P rS + PoS + Xs = 300 + 506
Xs S 506
Is + B6 - 16 - Bs + P r6 + P06 + x6 = 270 + 706
~ S 706

where Kt is an auxiliary variable, t = 1, 2, ... , 6. Our problem becomes:

6
min z = E [20(Prt + PoJ + 64Wt + 15(3PoJ + 2It + 3Bt + 30Ht + 4OL.]
t=l
s.t. WI S 90 + 106, W2 S 100 + 156
Wi S 120 + 306, i = 3, 4, 5 and 6
WI = Wo + HI-L l , W2 = WI +H2-~
W3 = W2 + H3 -~, W4 = W3 + H4 - L4
WS = W4 + Hs - Ls , W6 = WS + H6 - L6
3Prl S 8W l , 3Pr2 S 8W2, 3Pr3 S 8W3
3Pr4 S 8W4, 3PrS S 8Ws, 3Pr6 S 8W6
3Pol S (0.3)8W l , 3P02 S (0.3)8W2
3P03 S (0.3)8W3, 3P04 S (0.3)8W4
3PoS S (0.3)8Ws' 3P06 S (0.3)8W6
Prl + Pol + Io-Bo ~ 250
P r2 + P02 + II-B l ~ 285
 116

P r3 + P03 + 12 - ~ ~ 430
P r4 + P 04 + 13 - B3 ~ 260
P rS + P oS + 14 - B4 ~ 300
P r6 + P06 + IS - BS ~ 270
Io + B1 - 11 - Bo + P rl + Pol + Xl = 275
11 + ~-12-B1 + P r2 + P 02 + X2 = 330
12 + B3 - 13 - B2 + P r3 + P 03 + x3 = 450 + 500
13 + B4 - 14 - B3 + P r4 + P04 + x4 = 300 + 200
14 + BS - IS - B4 + P rS + P oS + Xs = 300 + 500
Is + B6 - 16 - BS + P r6 + P 06 + x6 = 270 + 700
xl ~ 250, x2 ~ 450, x3 ~ 700
x4 ~ 600, Xs ~ 500, x6 ~ 700
P rt , Pot' Wt, ~, Bt, H t , ~ and ~ ~ 0, V t.

Now, we can use the parametric programming approach proposed by Chanas [35] to solve this
problem. However, we provided 11 possible solutions, as in Table 3.10, which are considered
most helpful for decision makers.
The total costs associated with the production plans are between $81,915 and $91,213.
For a = 1 - 0 = 1, the decision maker has the highest satisfaction level in the fuzzy resources
utilized. The results indicate that each fuzzy parameter has achieved the maximum satisfaction
level in the decision maker's mind. The production plan is to produce more by using overtime
production, hiring fewer people, and establishing a lower inventory level. On the other hand,
the decision maker will hire more people to produce more through regular time production and
build up inventory during the optimistic periods when a lower aspiration level (a) is presented.
Suppose that the total cost should be in [$81,000, $88,000] by referring to Table 3.10.
The goal is $81,000 with the tolerance $7,000. Then we can obtain the following membership
function of the objective function (z = total cost):

1 if z*(O) < 81000

p.z(z*(0» = 1-
~ [z*(0)-81000]17000 if 81000 ~ z*(O) ~ 88000
if z*(O) > 88000.

Therefore, the following relation is obtained:

 117

'l'able 3.10 Fuzzy solutions of the fuzzy APP problem

8 z·(8) 8 z·(8) 8 z·(8) 8 z·(8)
1.0 81915 0.7 84625 0.4 87419 0.1 90262
0.9 82809 0.6 85545 0.3 88360 0.0 91213
0.8 83717 0.5 86482 0.2 89311

Aqqreqate production plan:

8 1.0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0

Prl 267 264 261 259 256 253 251 248 245 243 240
Pr2 307 303 299 295 291 287 283 279 275 271 267
Pel 307 303 299 295 292 295 298 302 308 314 320
P", 280 283 286 289 292 295 298 302 308 314 320
PrS 280 283 286 389 292 295 298 300 300 300 300
Pr6 270 270 270 270 270 270 270 270 270 270 270

Pol 0 0 0 0 4 9 14 20 25 30 35
Pol 0 12 31 50 63 67 72 75 76 77 77
Pol 85 91 90 88 88 89 89 91 92 94 96
P04 0 0 0 0 0 0 0 0 0 0 0
PaS 0 0 0 0 0 0 0 0 0 0 0
P06 0 0 0 0 0 0 0 0 0 0 0

II 17 12 6 1 0 0 0 0 0 0 0
Iz 38 37 42 47 50 47 43 37 30 22 14
I3 0 0 0 0 0 0 0 0 0 0 0
I4 20 17 14 11 8 5 2 0 0 0 0
Is 0 0 0 0 0 0 0 0 0 0 0
I6 0 0 0 0 0 0 0 0 0 0 0

BI 0 0 0 0 0 0 0 0 0 0 0
Bz 0 0 0 0 0 0 0 0 0 0 0
B3 0 2 4 6 8 10 12 14 16 18 20
B4 0 0 0 0 0 0 0 0 0 0 0
Bs 0 0 0 0 0 0 0 0 0 0 0
B6 0 0 0 0 0 0 0 0 0 0 0

WI 100 99 98 97 96 95 94 93 92 91 90
W2 115 114 112 111 109 108 106 105 103 102 100
W3 115 114 112 111 110 111 112 113 116 118 120
W4 105 106 107 108 110 111 112 113 116 118 120
Ws 105 106 107 108 110 111 112 113 113 113 112
W6 101 101 101 101 101 101 101 101 101 101 101

HI 0 0 0 0 0 0 0 0 0 0 0
Hz 15 15 14 14 13 13 12 12 11 11 10
H3 0 0 0 0 1 3 6 8 13 16 20
H4 0 0 0 0 0 0 0 0 0 0 0
Hs 0 0 0 0 0 0 0 0 0 0 0
~ 0 0 0 0 0 0 0 0 0 0 0

LI 0 1 2 3 4 5 6 7 8 9 10
Lz 0 0 0 0 0 0 0 0 0 0 0
~ 0 0 0 0 0 0 0 0 0 0 0
L4 10 8 5 3 0 0 0 0 0 0 0
Ls 0 0 0 0 0 0 0 0 3 5 8
~ 4 5 6 7 9 10 11 12 12 12 11
 118

z'(8) s 81000 + 70008.

By use of Chanas's approach, the solution is presented in Table 3.11 and Figure 3.13.

/1-z and /1-<

0.8

0.6

0.2

o 0.2 0.4 0.6 0.8 1 0

O·

Figure 3.13 Intersection of /1-0 and /1-<

Table 3.11 Chanas's optimal solution of the APP problem

o= 0.62 (a = 0.38) z*(0.62) 85348

Production Level Inventory Level Work Force Level

P r1 257 11 0 W1 96
P r2 292 12 51 W2 109
P r3 292 13 0 W3 109
P r4 291 14 9 W4 109
P rS 291 Is 0 Ws 109
P r6 270 16 0 W6 101

Pol 3 B1 0 H1 0
P o2 62 B2 0 H2 13
P o3 87 B3 8 H3 0
P o4 0 B4 0 H4 0
Pos 0 Bs 0 Hs 0
P o6 0 B6 0 H6 0
 119

3.1.3 Linear Programming Problem with Fuzzy Parameters in the Objective Function

It often happens that the unit profit or cost cannot be precisely detennined. Therefore,
considering the linear programming with imprecise coefficients in the objective function is
valuable. This problem is formulated as:

max cx
s.t. (Ax)j S b j , Vi
x ~ O. (3.20)

For Equation (3.20), Verdegay [251, 252] proposed the following equivalent parametric
linear programming:

max cx
s.t. I'(c) ~ 1 - ex
(Ax)j S bj, Vi
ex f [0, 1] and x ~ 0 (3.21)

where C = (cl> ~, ... , cn), I'(c) = in~ l'iCj), and I'(c) is the membership function of (cx) and
l'iCj) , Vj, are the membership functions of Cj' The problem given by Equation (3.21) is
equivalent to:

max cx
s.t. l'iCj) ~ 1 - ex, Vj
(Ax)j S b j , Vi
ex e [0, 1] and x ~ 0

which is equivalent to:

max Ej CjXj
S.t. Cj ~ I'j-l(l - ex), Vj
(Ax)j S bj, Vi
ex e [0, 1] and x ~ 0
or
 120

max };j /Lj l(1 - a)x;

O

s.t. (Ax)j S; bj, Vi

a E [0, 1] and x ~ O. (3.22)

Obviously, Equation (3.22) is a parametric programming, so we can obtain the solution of

Equation (3.20) by solving Equation (3.22). The solution is fuzzy and for each (1 - a)-level cut
of the fuzzy objective set, the solution has an a grade of preference.
On the other hand, Verdegay proposed that for a given fuzzy linear programming Equation
(3.20) or (3.1), there always exists another one that is dual and has the same fuzzy solution.
Therefore, we can solve the dual problem of Equation (3.20) by using the approaches discussed
in Section 3.1.1.
The dual problem of Equation (3.4) is:

min [b+p(l-a)]u
S.t. uAT ~ c, U ~ 0 and a E [0, 1] (3.23)

which is equivalent to:

min au
s.t. a =b + p(l - a)
uAT ~ c, U ~ Oandae [0,1].

Let fJ = 1 - a. The problem becomes:

min au
s.t. a S; b + pfJ « => 1 - [bj-a;]/pj ~ I-fJ, Vi)
uAT ~ c, U ~ OandfJE [0,1]
or

min au
s.t. J.Li(8j) ~ 1 - fJ, Vi
uAT ~ c, U ~ OandfJE [0,1].

Finally, we obtain:
 121

min au
s.t. uAT ~ c, u ~ O. (3.24)

In conclusion, the linear programming with fuzzy parameters in the objective function can
be solved by the approaches discussed in Section 3.1.1.

3.1.4 Linear Programming with All Fuzzy Coefficients

Sometimes, all the coefficients of a linear programming are imprecise. This problem can
be formulated as:

max -
cx
s.t. (Ax)i S bi, Vi,
x~O (3.25)

For Equation (3.25), Carlsson and Korhonen [31] proposed a fully trade-off approach. They
argued that Chanas' s approach does not consider any continuous trade-off between grades of
constraint violation. For example, if ILl = 1 - 81 and lLi > 1 - 81, i = 2, ... , m, then ILc = 1 -
81. Thus, the better possible solution will come out through releasing the lLi' which is greater
than 1 - 81, to 1 - 8i by trade-off with lLi ~ 1 - 82, i = 2, .. ,m. It is obvious 82 will be less than
81. Therefore, the better solution with a higher degree of satisfaction may be obtained with ILn
= Ilo = ILl = IL2 = .... = Pm = 1 - 82 •
Before going into the solution procedure, the requirement of the solution z· = z·(c, -A,
b) of the non-fuzzy version of Equation (3.25) is an increasing function of the parameters c, -A
and b is needed. Also, it is assumed that the user can specify the intervals [cO, c 1), [AO, AI) and
[b0, b 1) for the possible values of the parameters. The lower bounds represent 'risk-free' values
in the sense that a solution most certainly should be implementable. The upper bounds, on the
other hand, represent parameter values which are most certainly unrealistic and impossible and
the solution obtained by using these values is not implementable. When we move from 'risk-
free' towards "impossible" parameter value, we move from solutions with a high grade to ones
with a low grade on implementing from secure to optimistic solutions. Then, our task is to find
an optimal compromise "in-between as a function of the grades of imprecision in the parameter.
II

Carlsson and Korhonen proposed a relationship between a solution in Equation (3.25) and
its parameters: the solution z· = z·(c, -A, b) of Equation (3.25) is an increasing function of the
parameter c, -A and b. Thus, we can reasonably assume that membership functions are
 122

monotonically decreasing functions of the parameters, c, -A, b. The monotonically decreasing

functions may be linear, piecewise linear, hyperbolic exponential, etc. For the purpose of
illustrating non-linear membership functions, we will discuss the exponential functions proposed
by Carlsson and Korhonen. That is (see Figure 3.14):

(3.26)

where be > 0 or be < 0, and ~ = lI[l-exp(-bJ] and c E [co, c l ]. ILe = 1 when c ~ co, ILe =
o when c > c l . be is specified by the decision maker. In the same manner, the membership
function of A and b can be obtained. It is worth noting that for a maximization problem, the
membership functions of the objective function should be non-decreasing in the sense of
preference. However, let us keep the originality of Carlsson and Korhonen [31] here.
After full trade-off between c, -A and b, the solution will always exist at:

IL = ILe = J.LA = J.Lb· (3.27)

Therefore, we can obtain the following equation:

(3.28)

where J.L E [0, 1] and ge' GA and gb are inverse functions of IJ.e' ILA and ILb. Then Equation
(3.25) becomes:

max [ge(J.L)]x
s.t. [GA(J.L)]x ~ gb(IL) and x ;::: O. (3.29)

1 +----r-- ....
"- ....
"
,,bc<O
,,
\
\

\
\
\

° ------T-----------~------ c

Figure 3.14 The membership function of fuzzy set c

 123

Obviously, Equation (3.29) is a nonlinear programming problem. However, it can be solved by

any linear programming technique if p. is given. Thus, we can obtain a set of solutions
corresponding to a set of p.s and then plot these solution pairs (z·, p.). After referring to this
relationship, the decision maker can choose his/her preferred solution for implementation.

3.1.4.1 Example: A Production Scheduling Problem

Consider a production scheduling problem. A leather goods factory manufactures 3 types
of handbags, which must pass through four work centers and the man-hours in each are: chinking
[18, 22), paring [10, 40), stitching [96, 110) and finishing [96, 110). The input data are
described as fuzzy parameters because they are estimates and cannot be stated precisely in nature.
The unit profits of the three products and the technical coefficients are given in the following
fuzzy linear programming model:

max [1, 1.5)x1 + [1, 3)X2 + [2, 2.2)X3

s.t. [3, 2)Xl + [2, 0)x2 + [3, 1.5)x3 ~ [18, 22)
+ [2, I)x2 + [1, 0)x3 ~ [10,40)
[1, 0.5)x 1
[9, 6)x1 + [20, I8)x2 + [7, 3)X3 ~ [96, 110)
[7, 6.5)x1 + [20, I5)x2 + [9, 8)X3 ~ [96, 110)

where the numerical data are based on Carlsson and Korhonen [31]. Suppose that the
membership function of c, A, b (Figure 3.15) are:

P.aij = (au - a./)/(a.jo - a./)

P.bj = {I - exp[-0.8{bj-b/)/{bjO-b/)]}/[I-exp(-0.8)]
P.ci = {I - exp[3(cj-c/)/(cjO-c/)]}/[1-exp(3)]

where the constants (-0.8) and (3) determined by the decision maker indicate that a positive
number implies convex exponential function and a negative number implies concave exponential
function.
Next, through the appropriate transformation with the assumption of absolute trade-off
between fuzzy numbers of a.j' bj, cj' Vi,j, we obtained:
 124

I-'aij 1
(1-'.11 )

0.8

0.6

0.4

0.2

0 a ij
2 2.2 2.4 2.6 2.8 3.0 (all)

I-'bi 1
(I-'bd
0.8

0.6

0.4

0.2

0 hi
18 18.8 19.6 20.4 21.2 21.6 22 (hi)

I-'cj 1
(I-'cd
0.8

0.6

0.4

0.2

0 cj
1 1.1 1.2 1.3 1.4 1.5 (cd

Fiqure 3.15 The membership functions of the fuzzy coefficients

 125

~j = ~/ + p.(~jO_~/)
b i = b i - (110.8)ln{1 - P.bi[1-eXP(-O.8)]}(bP - bh
Cj = Cj + (1I3)ln{1 - p.cjl1-exp(3.0)]}(cjo - c/).

Thus, the original fuzzy linear programming with absolute trade-off is then equivalent to the
following:

max [1.5-0. 1671n(1 + 19.1p.)]xI + [3-O.667In(1 + 19.1p.)]x2+ [2.2-0.067In(1 + 19.1p.)]x3

s.t. (2 + P.)xI + (2P.)x2 + (1.5 + 1.5P.)x3 :S: 22+5In(I-O.55p.)
(0.5+0.5p.)xI +(1 + P.)x2 + (p.)x3 S 4O+37.5In(I-0.55p.)
(6+3p.)xI +(18 +2P.)x2+(3 +4P.)x3 :S: 110+ 17.5In(1-O.55p.)
(6.5+0.5p.)xI +(15+5p.)x2+(8+p.)x3 S 110+ 17.51n(I-0.55p.)
p. E [0, 1] and Xj ~ 0, i = 1, 2, 3.

Obviously, the above nonlinear programming cannot be solved by any standard linear
programming technique. However, it can be solved if p. is pre-determined. That is: for each
specific value of p., one can get an optimal solution for the original solution. Therefore, one may
choose n number of experiments (n different p. value) in order to obtain n optimal solutions and
then present these optimal solutions to the decision maker.
The solution procedure is described as follows:
Step 1. Let us start at p. = 0 and take n = 10 experiment, and set a = 0.1.
Step 2. Solve the following linear programming:

max 1.5xI + 3x2 + 2.2x3

s.t. 2xI + 1.5x3 S 22
0.5xI + x2 S 40
6xI + 18x2 + 3x3 SilO
6.5xI + 15x2 + 8x3 S 110
Xl' x2, x3 ~ O.

The optimal solution is: x· = (0,0, 13.75) and z· = $30.25.

Step 3. Plot (p., Z) = (0,30.25) in Figure 3.16. Then, set p. = p. + 0.1. If p. > 1, and go
to Step 4. Otherwise, go to Step 2 and solve the linear programming with p. = p. + 0.1.
Step 4. Figure 3.16 and Table 3.12 show the relationship between the optimal profits and the
corresponding membership grades. According to this relationship, the decision maker can then
 126

get his optimal solution under a pre-determined imprecision allowable. For instance, when the
decision maker considers 0.3 degree imprecision is acceptable, the optimal solution is: z· = 15.8
and x· = (0, 2.53, 6.28).

Table 3.12 The optimal solution for the absolute trade-off

model of a fuzzy linear programming

J.I.
z
.
Profit
x
.
Decision Resources used
bl ~ b3 b4
0 30.25 (0, 0, 13.75) 19.9 0.0 39.8 110.0
0.1 28.35 (0, 0.17, 13 .14) 21.7 1.5 47.7 109.1
0.2 26.02 (0, 0.73, 11. 74) 21.4 3.1 58.0 108.0
0.3 23.76 (0, 1. 22, 10.44) 21.1 4.7 66.5 106.8

0.4 21.60 (0, 1.64, 9.26) 20.8 6.0 73.4 105.7

0.5 19.55 (0, 1.99, 8.18) 20.4 7.1 78.7 104.4
0.6 17.62 (0, 2.29, 7.19) 20.0 8.0 82.8 103.1
0.7 15.80 (0, 2.53, 6.28) 19.6 8.7 85.5 101. 4

0.8 14.24 (0, 0, 7.07) 19.1 5.7 43.8 62.2

0.9 13.08 (0, 0, 6.52) 18.8 5.9 43.0 58.0
1.0 12.00 (0, 0, 6) 18.0 6.0 42.0 54.0

Objective value
30

25

20

I
I
I
15 I
I
I
I
--------------------~-----------------
I
10 I
I
I
I
5

o
o 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 3.16 The relationship between optimal values and their

corresponding membership grades
 127

3.2 Interactive Fuzzy Linear Programming

In this section, an interactive fuzzy linear programming approach (Lai and Hwang [137]
is investigated to improve the flexibility and robustness of linear programming (LP) techniques.
This section provides a practical modelling method which is a symmetric integration of
Zimmermann's, Wemers's, Verdegay's and Chanas's fuzzy LP approaches and additionally
provides a decision support system for solving a specific domain of practical LP problems. The
interactive concept provides a learning process about the system, whereby the decision maker can
learn to recognize good solutions, relative importance of factors in the system and then, design
a high-productivity system, instead of optimizing a given system. This interactive fuzzy LP
system provides integration-oriented, adaptation and learning features by considering all
possibilities of a specific domain of LP problems which are integrated in logical order using an
IF-THEN rule.

3.2.1 Introduction
In the early twentieth century, reducing complex real-world systems into precise
mathematical models became the main trend in science and engineering. Since then, Operations
Research (OR) has been applied to real-world decision making problems and has become one of
the most important fields in science and engineering. However, practical situations are often not
well-defined and thus cannot be described precisely. This imprecise nature is actually fuzziness
rather than randomness. Therefore, the traditional OR approaches may not really be suitable for
solving practical decision making problems. Since the fuzzy set theory was proposed by Zadeh
[274] in 1965, we have been able to handle not only well-defined, precise data, but also vague
or fuzzy data such as ·profit is around 30 %,. "cost is essentially less than $55," "lifetime is
essentially larger than 68 hours," etc. The role of fuzzy sets in decision processes is best
described in the original statements of Bellman and Zadeh [12]:

Much of the decision-making in the real world takes place in an environment in which the
goals, the constraints and the consequences of possible actions are not known precisely. To deal
quantitatively with imprecision, we usually employ the concepts and techniques of probability
theory and, more particularly, the tools provided by decision theory, control theory and
information theory. In so doing, we are tacitly accepting the premise that imprecision --
whatever its nature -- can be equated with randomness. This, in our view, is a questionable
assumption ...

Thus, decision processes are better described and solved by using fuzzy set theory, rather
 128

than precise approaches. However, the decision maker (OM) himself is always playing the most
important role in using fuzzy set theory. Therefore, an interactive process between the DM and
the decision process is necessary to solve our problems. That is actually a user-dependent fuzzy
LP technique. Furthermore, a problem-oriented concept is also a vitally important concept in
solving practical problems, as noted by Simon [224]:

Perhaps what is most important in making rapid progress towards our goals is that we
adopt a problem-oriented point of view. We must let the problem that we are trying to solve
determine the methods we apply to it, instead of letting our techniques determine what problems
we must be willing and able to tackle. That means we must be willing to apply our entire
armory of techniques, both heuristic and algorithmic.

By use of fuzzy set theory, and user-dependent (interactive) and problem-oriented

concepts, the flexibility and robustness of linear programming (LP) techniques are improved.
An interactive fuzzy LP approach is then developed which is a symmetric integration of
Zimmermann's, Wemers's, Verdegay's and Chanas's approaches (as discussed in the previous
sections) and additionally provides a decision support system for solving a specific domain of
real-world LP systems.
Moreover, the interactive concept provides a learning process about the system and makes
allowance for psychological convergence of the DM, whereby, (s)he learns to recognize good
solutions, relative importance offactors in the system, and then design a high-productivity system
[293], instead of optimizing a given system. Thus, we can improve some deficits of operations
research methodology as indicated in Ackoff [1, 2].
This interactive fuzzy LP system provides integration-oriented, adaptation and learning
features by considering all possibilities of a specific domain of LP problems which are integrated
in logical order using an IF-THEN rule.
In the following sections, we will first discuss Zimmermann's, Wemers's, Chanas's and
Verdegay's approaches.

3.2.2 Discussion Of Zimmermann's, Werners's, Chanas's and Verdegay's Approaches

After a survey of various models of fuzzy linear programming techniques, we find that
Zimmermann's, Wemers's, Chanas's and Verdegay's approaches and concepts are the most
practical ones among various techniques in the literature. However, each of them solves one
kind of linear programming problems and decision making tasks.
In order to understand the differences and contributions of these approaches and concepts,
 129

it is necessary to discuss some major points among them in this section.

In the Zimmermann fuzzy linear programming model, goal bo and its maximum tolerance
Po should be given initially. In real-world problems, it is unrealistic to initially ask the decision
maker to give bo and Po without providing any information about them. Therefore, the
membership function of the fuzzy objective is questionable and so the solution is also
questionable. For example, if the bo given is too large, there will be no solution in
Zimmerman's FLP model. At the same time, if the Po given is too large, then there will be no
meaning for the membership function. Thus, the solution is questionable. Instead of asking the
decision maker for bo and Po to establish the membership function of the fuzzy objective,
Werners provided two possible extreme points zO and zl which are:

zO = inf (max cx)

xeX

Zl = sup (max cx)

xeX

where X = {xl (Ax)i S bi, Vi, and x ~ OJ. The difference between Werners's and
Zimmermann's membership function is depicted in Figure 3.17. In this figure, Zimmermann's
bo and Po are assumed to be rational, i.e., zO S bo S zl and bo - Po ~ zoo If so, the decision
maker may consider Zimmermann's membership function as a more acceptable one most of the
time.

#-,0
1 -------------~~~~=======
Werners
Zimmermann

o cx

piqura 3.17 The difference between Zimmermann's and Werners's membership

functions ""0
 130

While Tanaka et al. [236] and Zimmermann [296] connected fuzzy set theory and the max-
min operator with linear programming, Verdegay [250] and Chanas [34] obtained the equivalent
relation between parametric programming with parametric resources and fuzzy linear
programming with fuzzy resources defined by assumed linear membership functions. Thus, for
a symmetric fuzzy linear programming problem, we can first consider its sub-problem of
nonsymmetric fuzzy linear programming with a crisp objective and fuzzy constraints, and then
construct and solve its corresponding parametric programming problem. The solutions will be
presented in a tableau which not only provides an optimal solution with respect to a parameter,
but also provides the actual consumed resources. Therefore, the difficulty of providing bo and
Po initially is overcome by presenting the solution table of a parametric programming problem.
After referring to this solution table, the decision maker can precisely locate his subjective bo and
Po. The solution of Zimmermann's model is then reliable.
Chanas [34] suggested that the membership function of the fuzzy objective be constructed
directly by the parametric optimal solution. Thus, Ilo is a function of the parameter instead of
the general function of x. However, in any real-world problems, the number of constraints is
always rather large (say 50), and so are the decision variables. Therefore, Chanas's approach
for formulating the membership function of the fuzzy objective is not practical.

3.2.3 Interactive Fuzzy Linear Programming - I

3.2.3.1 Problem Setting

The interactive fuzzy linear programming proposed here is a problem-oriented and user-
dependent approach. It also considers a large variety of situations that the decision maker might
meet in modelling and solving a linear programming problem. Interactive fuzzy linear
programming provides an efficient and systematic approach to solving the indicated linear
programming problem, and then :;hows the results (solutions and resources used) to the decision
maker. The decision maker may be satisfied with this solution or he may want to modify some
situations and/or change the original model. If satisfied, the problem, of course, has been
solved. If not satisfied, an interactive process will then proceed. The problem-solving procedure
is continued until the decision maker obtains a satisfying solution. It is worth noting that the
approach proposed here is not only to solve a given LP problem, but also to design a high
productivity system. A high productivity system was stated by Zeleny [293] as:
Modem production systems are increasingly characterized by novel high-productivity
 131

attributes: minimal buffers, close-to-zero inventories, just-in-time operations, minimal waste, full
utilization of resources, flexibility of layouts and resource portfolios, and so on. . . . Finally, the
problem has shifted from "How to optimize a given system" to "How to design an optimal
system" .... It is therefore important to design high-productivity systems with no excess, waste
or buffers. Only then can the desirable "buffers" be added with some insight and competence.
If such optimal design is impossible, then at least imprOVed redistribution of the excess capacity
should be attempted.

Interactive fuzzy programming methods have been studied since 1980. Typical works are
Baptistella and Ollero [8], Fabian, Ciobanu and Stoica [84], Sakawa and Yano [220], Slowinski
[227], Werners [257] and Zimmermann [BM64]. Zimmermann [BM64] described some general
concepts and modelling methods of decision support systems and expert systems in a fuzzy
environment. Others developed interactive approaches to solve multiple criteria decision making
problems (crisp or fuzzy).
The approach proposed here is based on an interaction with the decision maker in order
to find a satisfying solution for a linear programming problem. In a decision process using an
LP model, resources available may be fuzzy, instead of a precisely given number as in the
traditional (crisp) LP model. For example, machine hours, labor force, material needed and so
on, are always imprecise, because of various potential suppliers and environments. Therefore,
they should be considered as fuzzy resources, and then the LP problem should be solved by use
of the fuzzy set theory. Thus under the fuzzy resource constraints, the objective function (e.g.,
profit function) to be maximized may be considered as either a crisp objective function or a fuzzy
objective function symmetrically. Therefore, we can define a linear programming problem with
crisp or fuzzy resource constraints, and a crisp or fuzzy objective as:

(-)
max z=cx
s.t. (Ax)j S'b;, i = 1,2, ... , m
x~O (3.30)

where fuzzy resources bj, Vi, have the same forms of membership functions as Figure 3.8. We
may also consider the following fuzzy inequality constraints:

,-.
max z=cx
s.t. (Ax)j,;e,bj, i = 1,2, ... , m
x~O (3.31)

where fuzzy inequality constraints have the membership functions of Figure 3.8. Even though
 132

Equations (3.30) and (3.31) are different in some points of view, we can use the same approach
to handle them under the pre-assumption of the membership functions of the fuzzy available
resources and fuzzy inequality constraints.
According to Equations (3.30) and (3.31), we can derive the following five problems:

Problem 1. When the resources can be determined precisely, a traditional linear programming
problem is considered as:

max z = cx
s.t. (Ax)j:::; bj, Vi
x~o (3.32)

where c, A, and bj, Vi, are precisely given. The solution of Equation (3.32) is a unique optimal
solution.

Problem 2. (Approach of Chanas [34] and Verdegay [250]) A decision maker might wish to
make a post-optimization analysis. Thus, a parametric programming problem is formulated as:

max z=cx
s.t. (Ax)j:::; bj + Opj' Vi
(J E [0, 1] and x ~ 0 (3.33)

where c, A, hj and Pj, Vi, are precisely given and (J is a parameter (a fraction of the maximum
tolerances). Pj' Vi, are maximum tolerances which are always positive. The solutions z*«(J) of
Equation (3.33) are functions of (J. That is, for each (J, we can obtain an optimal solution; with
the increasing use of resources, the optimal solution behavior can be figured out. The results are
displayed in Table 3.13 and then presented to the decision maker. The decision maker can then
choose his satisfying solution for implementation.
On the other hand, the available resources may be fuzzy. Then the linear programming
problem with fuzzy resources becomes:

max z = cx
s.t. (Ax)j:::; bj, Vi
x ~ o. (3.34)
 133

Table 3.13 Solutions of a parametric programming problem

Resources Actually Used

8 z* (8) b1 b2 bm
0.0
0.1
0.2

0.8
0.9
1.0

It is possible to determine the maximum tolerances Pj, of the fuzzy resources bj, Vi. Then we
can construct the membership functions J.'j assumed linear for each fuzzy constraints, as follows:

I if (Ax)j < bj
(
J.'j(x) = 1 - [(Axkbj]/pj if bj ~ (Ax)j ~ bj+Pj
o if (Ax)j > bj+Pj (3.35)

Verdegay and Chanas proposed that Equations (3.34) and (3.35), however, are equivalent to
Equation (3.33), a parametric linear programming where c, A, bj and Pj, Vi, are given,.by use
of the a-level cut concept, where a = 1 - (j.

Problem 3. (Werners's [257] approach) A decision maker may want to solve a fuzzy linear
programming problem with a fuzzy objective and fuzzy constraints, while the goal bo is not
given. That is:

--- z=cx
max
S.t. (Ax)j ~ -bj' Vi
x~O (3.36)

which is equivalent to:

max z = cx
s.t. (Ax)j ~ bj + ()Pj' Vi
(j E [0, 1] and x ~ 0 (3.37)
 134

where c, A, bi and Pi' Vi, are given, but the goal of the fuzzy objective is not given.
By using Wemers's approach, we first obtain zO and zl for Equation (3.37). Then
Wemers's membership function f.'o of the fuzzy objective is (see Figure 3.17):

I if cx > zl
{
f.'o(x) = 1 - (zl-cx)/(zl-z~ if zO s cx S zl
o if cx < zoo (3.38)

The membership functions J.Li' Vi, of the fuzzy constraints are defined as Equation (3.35). But
using the max-min operator, we have:

max ex
s.t. f.'o(x) and I-'i(x) ~ ex
ex, f.'o(x) and I-'i(x), Vi, ex E [0, 1] and x ~ 0

where c, A, bi and Pi' Vi, are given, and ex = min (f.'o, 1-'1' ... , I-'m). Let ex = 1 - e, then the
problem given becomes:

min e
s.t. cx ~ zl - e(zl - z~
(Ax)i S bi + Bpi, Vi
e E [0, 1] and x ~ 0 (3.39)

where c, A, bi and Pi' Vi, are given and e is a fraction of (zl - z~ for the first constraint and a
fraction of the maximum tolerances for others. The solution is a unique optimal solution.

Problem 4. (Zimmermann's [296] approach) A decision maker may want to solve a fuzzy
linear programming problem with a fuzzy objective and fuzzy constraints, when the goal bo of
the fuzzy objective and its maximum tolerance are given. That is:

max cx
s.t. (Ax)i S bi, Vi, and x ~ 0 (3.40)

where c, A, bo, Po, bi and Pi, Vi, are given. It is noted that the decision maker should refer to
Table 3.13 for deciding bo and Po. Zimmermann's membership function of the fuzzy objective
f.'o is defined as (see Figure 3.17):
 135

~
if cx > bO

",,(x) - { - (bo-cx)/Po if bo-Po S cx S bo

if cx < bo-Po' (3.41)

Let ex = 1 - (). Then Equation (3.40) becomes:

min ()
s.t. cx ~ bo - ()Po
(Ax)i S bi + ()Pi' Vi
() E [0, 1] and x ~ ° (3.42)

where c, A, bo, Po, bi and Pi' Vi, are given and () is a fraction of the maximum tolerances. The
optimal solution of the Equation (3.42) is unique.

Problem S. A decision maker may want to solve a fuzzy linear programming problem with a
fuzzy objective and fuzzy constraints, while only the goal bo of the fuzzy objective is given, but
its tolerance Po is not given. That is:

max cx
S.t. (Ax)i S bi' Vi
x~O (3.43)

where c, A, bo, b i and Pi' Vi, are given, but Po is not given. While Po is not given, we do know
that Po should be in between ° and bo - zO (zo as defined in the Problem 4 section). For each
Po E [0, bo - z~, we can obtain the membership function of the fuzzy objective as Equation
(3.41). Since in a high-productivity system the objective value should be larger then zO at () =
0, there is no meaning to giving a positive grade of membership for those which are less than
zO. Figure 3.18 depicts the possible range of Po.
The difference between Problem 5 and Problem 4 is that Po is not initially given in
Problem 5. Therefore, we may assume a set ofpos, where Po E [0, bo - z~. Then, a problem
with each given Po is a Problem 4. The solutions for this given set of PoS are presented in Table
3.14. After referring to this table, the decision maker may choose a refined Po. Then Problem
4 with the decision maker's refined Po is solved. This solution will be the final optimal solution
for Equation (3.43).
 136

110
1 ----1------ - - - - - - - - - - - - - - - - - - -
I
I
I
I
I
I
I

o cx

Fiqure 3.18 The reasonable range of Po

Table 3.14 optimal solutions of a symmetric fuzzy linear

programming for a given set of Pos

Resources Actually Used

Po o z** x** b1 b2 bm
o

3.2.3.2 The Algorithm of IFLP-I

Based on Lai and Hwang [137], the algorithm of interactive fuzzy linear programming -
I is (see Figure 3.19) then:

Step 1. Solve a traditional linear programming problem of Equation (3.32) by use of the simplex
method. The unique optimal solution with its corresponding consumed resources is presented to
the decision maker.

Step 2. Does this solution satisfy the decision maker? Consider the following cases:

IF THEN
1) solution is satisfied -> print out results and stop
2) resource i, for some i -> reduce resources available
are idle and go to step 1
3) available resources are -> make a parametric analysis
not precise and some and go to step 3
tolerances are possible
 "iI
...
IDecrease b i (and change Pi)' for some i ~ _________ ~11
1---1 CD
w
...
III

'"lj
I-'
yes 0
:E:
0
::r
11/
11
rt
yes 0
111
H
I '"lj
I
no I ~
6. Solve L ______ ~ __ _ I
Problem 3 H
~

..... t"' w
--J
P- 11/
--....
IV
01
s:: 11/
11. Solve ..... ::s
L _______________ _ P- .t::
'" 0-
Problem 5 (J
(Problem 4 IV
III 'tl
for n times) s:: ~
11/
IV
'..."
(J
'" ::s
IQ
s:: .....
......... .Q
"'-
IV IV IV
....W
III E III
0 --.J
IV
'..." III IV
'..."
(J ... (J
yes IV 0 IV
Q .... Q

I
--L_
I
I
I
I
I
L ______________________________________________________________________
-
..tI
 138

Step 3. Solve a parametric linear programming problem of the Equation (3.33). The results are
depicted in Table 3.13. At the same time, let us identify.p = z·(8 = 0) and zl = z·(8 = 1).
Step 4. Do any of these solutions shown in Table 3.13 satisfy the decision maker? Consider
the following cases:

IF THEN
1) solution is satisfied -> print out results and stop
2) resource i, for some i -> decrease b i (and change Pi)
are idle and then go to step 1
3) tolerance i, for some i -> change Pi as desired and go
are not acceptable to step 3
4) the objective should be -> go to step 5
considered as imprecise

Step S. After referring to Table 3.13, the decision maker is then asked for his subjective goal
bo and its tolerance Po for solving a symmetric fuzzy linear programming problem. If the
decision maker does not like to give his goal for the fuzzy objective, go to Step 6. Ifbo is given,
go to Step 8.

Step 6. Solve Problem 3 of Equation (3.39). A unique Werners solution is then provided.

Step 7. Is the solution of Problem 3 satisfying?

Consider the following cases:

IF THEN
1) solution is satisfied -> print out results and stop
2) the user has realized -> give the goal b o and go to
his/her goal step 8.
3) resource i, for some i -> decrease b i (and change Pi)
are idle and then go to step 1
4) tolerance i, for some i -> change Pi as desired and go
are not acceptable to step 3

Step 8. Is Po determined by the decision maker? If the decision maker would like to specify Po,
we should provide Table 3.14 to help the decision maker. It is noted that Po E [0, bo - z'1. Then
go to Step 9. If Po is not given, then go to Step 11.

Step 9. Solve Problem 4 of Equation (3.42). A unique Zimmermann solution is obtained.

Step 10. Is the solution of Problem 4 satisfying?

IF THEN
1) solution is satisfied -> print out results and stop
2) the user has realized -> give the goal b o (and Po)
better his/her goal and go to step 8
 139

(and its tolerance)

3) resource i, for some i -> decrease b i (and change Pi)
are idle and then go to step 1
4) tolerance i, for some i -> change Pi as desired and go
are not acceptable to step 3

Step 11. Solve Problem 5. That is: call Step 9 to solve Problem 4 for a set of Pos. Of course,
Po E [0, bo - z'1. Then the solutions are depicted in Table 3.14.

Step 12. Are the solutions satisfying? If yes, print out the solution and terminate the solution
procedure. Otherwise, go to Step 13.

Step 13. Ask the decision maker to specify the refined Po, and then go to Step 9. It is rather
reasonable to ask the decision maker Po at this step, because he has a good idea about it now.

For implementing the above interactive fuzzy linear programming, we need only two
solution-finding techniques: simplex method and parametric method. Therefore, the interactive
fuzzy linear programming approach proposed here can be easily programmed in a PC system for
its simplicity.
In order to understand the interactive fuzzy linear programming, it is necessary to have
a numeric example as follows.

3.2.3.3 Example: The Knox Production-Mix Selection Problem

Let us consider the above Knox Mix company problem by using our interactive fuzzy
linear programming system to obtain a satisfying solution.

Step 1. Solve the above linear programming problem by use of the simplex method. The
optimal solution is: x· = (5017,0,5517,0) = (7.14,0,7.86,0) and z· = $69517 = $99.29.
The actual resources used are 15, 73.57 and 100 units for man-weeks, material Y and material
Z, respectively.

Step 2. The decision maker feels that 73.57 units of material are actually enough to reach a
$99.25 profit. Thus, 120 - 73.57 = 46.43 pounds of material Y that are really idle resources.
The decision maker, therefore, would like to try to find the solution of a new linear programming
with the resources changed from (bI' b2, b3) = (15, 120, 100) to (15, 73.57, 100). Then, go
to Step 1.
 140

Step 1. Solve the crisp linear programming problem with b2 = 73.57. The optimal solution is
the same as previously indicated: x· = (7.14, 0, 7.86, 0), and z· = $99.29. The actual
resources used are IS, 73.57, and 100. Now there are no idle resources. Instead of just
optimizing a given LP system, the procedure has yielded a high-productivity system.

Step 2. If the decision maker is satisfied with this solution, print out the results and stop the
solution procedure. Sometimes a decision maker may want to further analyze his problem. Then
the following two cases may be considered.

c.aK..l.
Step 3-1. The decision maker may feel that a posterior analysis of the previous LP problem is
necessary before implementing its solution.
Let us assume that the decision maker provides that the maximum tolerances for man-
weeks, material Y and material Z are 3, 0, and 20 units, respectively. Then, the parametric
programming is:

max z
XEX

where X = {x I gl(x) S 15 + 38, g2(x) S 120, g3(x) S 100 + 208, x ~ 0, and 8 E [0, In
and 8 is a parameter. The optimal solution is: x· = (7.14 + 1.438,0, 7.86 + 1.578,0), and
z· = $(99.29 + 19.868). In order to more clearly display these results, Table 3.15 is provided
to the decision maker.

Tabla 3.15 Solutions of the parametric programming problem

resources actually used

(J z* man-weeks material Y material Z

0.0 99.29 15.00 73.57 100.00

0.1 101.28 15.30 75.04 102.00
0.2 103.27 15.60 76.51 104.00
0.3 105.26 15.90 77.98 106.00
0.4 107.25 16.20 79.45 108.00
0.5 109.24 16.50 80.92 110.00
0.6 111. 23 16.80 82.39 112.00
0.7 113.22 17.10 83.86 114.00
0.8 115.21 17.40 85.33 116.00
0.9 117.20 17.70 86.80 118.00
1.0 119.19 18.00 88.27 120.00
 141

Step 4-1. After an overview of the above table, the decision maker may pick out a satisfying
solution for implementation. If so, print out the results and terminate the solution procedures.
Otherwise, the following steps may be considered:
(i) Reduce the amount of material Y from 120 to 73.57 and give its maximum tolerance. Then
go to Step 1.
(ii) Decrease/increase the maximum tolerances PI and/or P3' and then go to Step 3-1. (Just
modify the results of the parametric process.)
The solution process will continue until the satisfying solution is reached.

~.

Step 3-2. The decision maker may consider that the traditional linear programming is not
enough to solve his problem because of the imprecise properties of the resources in nature. After
a detailed analysis, he feels that the consumption of man-weeks, material Y and material Z should
be "substantially less than or equal to" 15,80 and 100 units, respectively, with the corresponding
maximum tolerances of 5, 40 and 30 units. Thus the problem becomes a fuzzy linear
programming problem with a crisp objective and fuzzy constraints which are defined by their
membership function assumed to be linear. That is:

max z
xeX'

where X' = {x I gl(x) -

:s; 15, g2(x) :s; 80, g3(x) :s; 100, and x ~ O} and the membership
function I-'i for the ith fuzzy constraints, Vi, are:

~
if gl(x) < 15
"I(X) { : - [g1(x)-15J/5 if 15 :s; g I (x) :s; 20
if gl(x) > 20

~
if g2(x) < 80
,.,(x) { : - [g,(x)-80J/40 if 80 :s; g2(x) :s; 120
if g2(x) > 120

1 if g3(x) < 100

[
I-'3(x) = 10 - [g3(x)-100]/30 if 100 :s; g3(x) :s; 130
if g3(x) > 130.
 142

This nonsymmetric fuzzy linear programming is then equivalent to:

max z
XEX"

where X" = { x I gl(X) S 15 + 58, g2(x) :s 80 + 408, g3(x) S 100 + 308, x ~ 0, and 8 E

[0, In and 8 is a parameter. Therefore, we can use the parametric programming approach to
solve this problem and obtain its solutions as shown in Table 3.16.

Table 3.1' Solutions of a nonsymmetric fuzzy linear programming

problem with a crisp objective and fuzzy constraints

resources actually used

8 z* man-weeks material Y material Z

0.0 99.29 15.00 73.57 100.00

0.1 102.36 15.50 76.22 103.00
0.2 105.43 16.00 78.75 106.00
0.3 108.50 16.50 81.50 109.00
0.4 111.57 17.00 84.15 112.00
0.5 114.64 17.50 86.80 115.00
0.6 117.71 18.00 89.45 118.00
0.7 120.78 18.50 92.10 121. 00
0.8 123.85 19.00 94.75 124.00
0.9 126.92 19.50 97.40 127.00
1.0 130.00 20.00 100.00 130.00

Step 4-2. After referring to this solution table, the decision maker may choose a satisfying
solution for implementation, and then terminate the solution procedure. Otherwise, three
procedures (as those in the Algorithm) may be considered. Assume that a symmetric fuzzy linear
programming with a fuzzy objective and fuzzy constraints is considered, and then go to Step 5.

Step S. Present Table 3.16 to the decision maker and ask him to determine his subjective goal
boo If bo is given, go to Step 8. Otherwise, go to the next step.

Step 6. Solve Problem 3 given in (3.39). First, according to Table 3.16, let us identify:

zO = z.(8 = 0) = 99.29
zl = z·(8 = 1) = 130.

The symmetric linear programming problem can then be formulated as follows:

 143

max z
XEX'

where the membership functions of the fuzzy constraints are described in Step 3-2 and the
membership function 1-'0 of the fuzzy objective is defined as:

>
~~
if z 130
I'<tx) { - (13O-z)/(130-99.29) if 99.29 ::; z ::; 130
if z < 99.29.
This problem is then equivalent to:

min 0
XEXU

where Xu = { x I f(x) ~ 130 - 30.710, gl(x) ::; 15 + 50, g2(x) ::; 80 + 400, g3(x) ::; 100 +
300, x ~ 0 and 0 E [0, In. The solution is: x..... = (8.57,0, 8.93,0) and z ...... = $114.65 at 0
= 0.5, while the actually used resources are 17.5, 86.78 and 115.01 for man-weeks, material
Y and material Z, respectively.

Step 7. The decision maker may be satisfied with this solution. Print out the results and stop
the solution procedure. If he is not satisfied with this solution, three steps in Algorithm might
be considered. Let us assume that after recalling Table 3.16, determining bo is considered and
then go to Step 8.

Step 8. Presenting Table 3.16, ask the decision maker to determine Po. If Po is given, go to
Step 9. Otherwise, go to Step 11.

Step 9. Solve Problem 4. Now let us assume bo = 111.57 at 0 = 0.4 and Po = 10. It is noted
that Po should be in between 0 and 12.28. Our problem becomes:

find x
such that f(x) ~ 111.57 and x E X'

where the membership functions of the fuzzy constraints are described in Step 3-2 and the
membership function 1-'0 of the fuzzy objective is defined as:
 t-
144

ifz > 111.57

I'<J.x) = (1I1.57-z)/1O if 101.57 s z S 111.57
if z > 101.57.

The problem is then equivalent to:

min 8
XEXg'

where Xg' = { x I f(x) ~ 111.57 - 108, g.(x) s 15 + 58, g2(x) S 80 + 408, g3(x) S 100
+ 308, x ~ 0 and 8 E [0, In. The solution is: x·· = (8.01,0,8.50,0) and z·· = $108.54 at
8 = 0.30, while the resources actually used are 16.51, 81.57 and 109.03 for man-weeks,
material Y and material Z, respectively.

Step 10. Consider four steps shown in Algorithm. Here, let us assume that the solution is
satisfied. Then the solution procedure of our interactive fuzzy linear programming system is
completed. After such detailed communication between the system and the decision maker, the
solution will be quite good for implementation.

Step 11. Take a simulation by choosing a set of Pos. Let us take 5 possible values of Po E [0,
12.28]; 0, 3, 6, 9, 12.28. For each Po value, we call Step 9 to obtain the solution of Problem
4. The results are depicted in Table 3.17.

Step 12. After referring to Table 3.17, the decision maker may choose a satisfying solution and
terminate the solution procedure. Otherwise, he is asked to specify a refined Po. It is quite
reasonable here to obtain a Po which can really satisfy the decision maker's need. Table 3.17
provides very detailed information of the optimal solution of a symmetric fuzzy linear
programming problem with bo = 111.57 for each Po in the given set. Here, let us assume that
the optimal solution should be in between Po = 9 and 12.28.

Step 13. The decision maker finally indicates that Po = 10 will be perfect. Then go to Step 9.

Finally, it is necessary to compare the solutions of Problem 3 and Problem 4. Table 3.18
and Figure 3.20 show the differences.
 145

Table 3.17 The solutions of Problem 5

Po (}' z
.. x ..
0 0.400 111.570 (8.286, 0, 8.714, 0)
3 0.364 110.478 (8.184, 0, 8.638, 0)
6 0.335 109.562 (8.099, 0, 8.574, 0)
9 0.309 108.788 (8.027, 0, 8.520, 0)
12.28 0.286 108.057 (7.959, 0, 8.469, 0)
resources actually used
Po man-weeks material Y material Z

0 17.000 84.144 111.998

3 16.822 83.202 110.932
6 16.673 82.415 110.037
9 16.547 81. 749 109.281
12.28 16.428 81.120 108.567

Table 3.18 The comparison of the solutions of Problem 3

and Problem 4

Problem 3 Problem 4

(} = 0.5 0.3
z·· 114.65 108.54
x· = (8.57, 0, 8.93, 0) (8.01, 0, 8.50, 0)

resources used
man-week: 17.5 16.51
material Y: 86.78 81.57
material z: 115.01 109.03
 146

J.l.o

0.7 ----------

Problem

0.5 Problem 3

o
99.29 108.54 111.57 114.65 130.00
4x1 + 5x2 + 9x3 + 11x4

Figure 3.20 Comparison of the solutions of Problem 3 and Problem 4

3.2.4 Interactive Fuzzy Linear Programming - II

To develop a decision support system to solve all possible linear programming problems,
we first need to classify all possible situations. After a thorough literature survey, we classify
all possible situations into 11 cases and 7 problems as shown in Table 3.19.
In this section, we propose a revised version of our previous interactive fuzzy linear
programming. Interactive fuzzy linear programming - II (lFLP-II) integrates all possible
situations of fuzzy linear programming problems, and further provides a decision support system
for improving the flexibility and robustness of linear programming techniques (Lai and Hwang
[141)).
We have developed an interactive fuzzy linear programming for solving Problem 1 through
Problem 5, in the previous section. Thus we will only discuss Problems 6 and 7 in the
following.
 147

Table 3.19 Classification of fuzzy linear programming problems [141]

max/min cx fuzzy coefficients

s.t. AxS b fuzzy objective z c
x ~ 0
fuzzy resources b A
Case Problem Approach
1- Crisp LP Prob. 1 simplex method
2. b Prob. 2 Parametric programming
Verdegay's or Chanas's
- -
3. z & b Prob. 3 Werners's IFLP-I
Prob. 4 Zimmermann's
Prob. 5 Lai and Hwang's
4. c - Prob. 2 Parametric programming
5. -
A IFLP-II
6. b & c -
7. A & b Prob. 6 Carlsson and Korhonen's
8. A & c -
9. A, b & c -
10. z & A Prob. 7
- b)
Lai and Hwang's
11- z, A &
Noted: Case 4 is essentially the dual problem of Case 2.

Problem 6. (Carlsson and Korhonen's approach [31]) The problem is a general one where A,
band c might be fuzzy. Carlsson and Korhonen have provided an efficient approach to solve
the problem of fuzzy A, band c (Case 9). Their approach can also be applied to solve the
problems of fuzzy A (Case 5), fuzzy band c (Case 6), fuzzy A and b (Case 7), and fuzzy A and
c (Case 8).
Now, let us formulate the problems of Cases 5,6, 7, 8 and 9 as:
,~,

max z=cx
(-) '_J
S.t. Ax::;; b and x ~ 0 (3.44)

where the membership functions of fuzzy coefficients A, and/or b, and/or c are assumed linear
as follows (see Figure 3.21):
 148

< Il;jo
~
if Il;j

....<a,) = { -<
.....')1», if Il;,j0 :S Il;,j :S 1l;,j0+P;j
if Il;j > Il;jO+Pij

<
~
ifbi biG

...{b-) ={ - (b.-hi'll", if biG S; b i :S biG +Pi

if bi > biG +p;

if Cj > cjO
if CjO_Pj :S Cj S; cjO
if Cj < CjO-Pj.

It is noted that the membership functions of fuzzy C are for a maximization problem so that the
larger the coefficients, the more we prefer. On the other hand, the less the resources are
consumed, the better we feel, and so the technique coefficients. Carlsson and Korhonen pointed
out that if the full trade-off between A, b and Cis assumed, the optimal solution will always exist
at p. = P.A = IJ.t, = P.e, where the membership functions are non-increasing (or non-decreasing)
monotonic linear functions. Therefore, we obtain the following auxiliary problem:

I_...-__
P. if crisp coefficients

1 +-___a_jj_Cb_j).,..______ C_i_ _ _ _ __
I
I
I
I
I
I
I
I
I
I

Figura 3.21 Membership functions of Fuzzy A, band c

 149

max [1L-Ic(IL)]X
s_t. [1L-1A(IL)]X S IL-Ib(lL) and x ~ 0
or
max Ej [CjO - (1-IL)Pj]JS
s.t. Ej [ayO + (l-IL)Pij]JS S biG + (l-IL)Pi' Vi
JS ~ 0, Vj (3.45)

where CjO, 3;j o and biG are given critical values, and Pj' Pij and Pi' respectively, are given
tolerances, for all i and j. Obviously, Equation (3.45) is a linear programming problem if IL is
fixed first. Thus, for each given IL value, we can obtain the optimal solution of Equation (3.45).
In practice, we may set IL = 0.0, 0.1, ... , 0.9, 1. Here, IL is the ~verall degree of satisfaction
from the critical points (target points or goals). Finally, Table 3.20 is provided to decision
makers for further decisions.

Table 3.20 Solutions of Problem 6

Resources Used
bl ~

0.0
0.1

0.9
1.0

Problem 7. (Lai and Hwang's approach [141]) First let us formulate Cases 10 and 11 as
follows:

max z = cx
S.t. Ax s'b' and x ~ 0 (3.46)

where the membership functions of ~j and/or hi are assumed linear as in Problem 6. Based on
the concept of non-symmetrical models [BM64], the membership function of i can be established
by the analysis of fuzzy constraints. Here we consider that Werners's and Zimmermann's
membership functions are suitable for this situation. Thus we can solve Equation (3.46) by the
approaches provided by Carlsson and Korhonen (see Problem 6), and Werners (see Problem 3).
 150

That is: we first solve the following problem:

max I;j Cpt;

s.t. I;j [a;jO + (l-J.L)Pij]X; ~ biO + (I-J.L)Pi' Vi
X; ~ 0, Vj (3.47)

where bio and a;jo are given critical points, and Pi and Pij' respectively, are given tolerances. As
discussed in Problem 6, let I' = 0.0, 0.1, .. . ,0.9, 1.0. The solution table (as Table 3.20) and
the corresponding piece-wise linear (assumed) function (see Figure 3.22) between z· and I' are
generated. Thus, Wemers's membership function of the objective function (z) can be constructed
by ZO and Zl, where ZO = z·(J.L = 0) and Zl = z·(J.L = 1). The optimal solution Zw· is at the
intersection point of the membership function of the objective function and the assumed linear
segment of the relation functions between objective values (Z) and degree of satisfaction (I') as
shown in Figure 3.22.
On the other hand, the decision maker may like to provide his/her goal (bo) and its
tolerance <Po), after referring to Table 3.20. Then Zimmermann's membership function is more
meaningful. The solution (z..) can be found by solving two simultaneous equations -
Zimmermann's membership function and the assumed linear segment of the relation functions
between objective values (Z) and degree of satisfaction (I') as shown in Figure 3.22.

/J
1

.9

.8

.7

.6 I
I
I
.5 I
1 I
.4 -r-- :
I I I
.3 - - .. - J_ - - - -t- - - -
1 I I
I 1
.2 _ _ _ _ _ _ 1_ - ... - - __ ..J_ - - - - - - - -
I 1 I
.1 ------'--r----r---------
1

I
I

I
I

I
0
I I z·
Zo bo-Po

Figure 3.22 The solution of Problem 7

 151

3.2.4.1 The Algorithm of IFLP-U

The IFLP-II is basically a systematic integration of Zimmermann's [296], Wemers's [257],
Verdegay's [250, 251], Chanas's [34], Lai and Hwang's [137, 141], Lai [139] and Carlsson and
Korhonen's [31] fuzzy linear programming approaches, and additionally provdes a decision
support system for solving a specific domain of practical linear programming problems.
The algorithm (see Figure 3.23) is shown in the following:

Step 1. Solve a traditional linear programming problem of Problem I by.use of the simplex
method. The unique optimal solution with its corresponding consumed resources is presented to
the decision maker.

Step 2. Does this solution satisfy the decision maker? Consider the following cases:

IF THEN
1) solution is satisfied -> print out results and stop
2) resource i, for some i -> reduce resources available
are idle and go to step 1
3) available resources are -> make a parametric analysis
not precise and some and go to step 3
tolerances are possible

Step 3. Solve a parametric linear programming problem of Problem 2. The results are depicted
in Table 3.13. At the same time, let us identify zO = z*«(J = 0) and zl = z*«(J = 1).

Step 4. Do any of the solutions shown in Table 3.13 satisfy the decision maker? Consider the
following cases:

IF THEN
1) solution is satisfied -> print out results and stop
2) resource i, for some i -> decrease b i (and change Pi)
are idle and then go to step 1
3) tolerance i, for some i -> change Pi as desired and go
are not acceptable to step 3
4) the objective should be -> go to step 5
considered as imprecise
5) coefficients of the -> go to step 14
objective are imprecise
6) objective function and -> go to step 17
technique coefficients
(and/or resources available)
are imprecise
7) technique coefficients -> go to step 15
(and/or resources available
and/or coefficients of
objective function) are
imprecise
 ......
1•
w
.
N
W

I'Ij
15. Solve ....
Problem 6 ~
0
:r
I»
11
rt
Change aij (Pij)' b i (Pi)
0
(Pj)' for some i or j HI
1-4
N
III 0 ~
Ei I
0 '" 1-4
OJ
1-4
.... N .... ....
(J1
.Q 0 Po ...... I\)

'" III
01 ....&'
11. Solve ......., .... ~
Po Po AI
~ • I Problem 5 .<: I»
III u ::s
(Problem 4 Ilo
for n times) '0
~
nOI : AI
yes i
I .~ .Q
....
a ~
Po
~ III
AI OJ
I AI
........
III
I 5 N ...,..
U
III
yes _____ 1__ - ~ A

-_±--..-- I
I
I
I
I
L______________________________________________ - - ____________________ .2
 153

Step S. After referring to Table 3.13, the decision maker is then asked for his subjective goal
bo and its tolerance Po for solving a symmetric fuzzy linear programming problem. If the
decision maker does not like to give his goal for the fuzzy objective, go to Step 6. Ifbo is given,
go to Step 8.

Step 6. Solve Problem 3. A unique Wemers solution is then provided.

Step 7. Is the solution of Problem 3 satisfying? Consider the following cases:

IF THEN
1) solution is satisfied -> print out results and stop
2) the user has realized -> give the goal b o and go to
his/her goal step 8.
3) resource i, for some i -> decrease b i (and change p;)
are idle and then go to step 1
4) tolerance i, for some i -> change Pi as desired and go
are not acceptable to step 3

Step 8. Is Po determined by the decision maker? If the decision maker would like to specify Po,
we should provide Table 3.13 to help. It is noted that Po E [0, bo-zo:J. Then go to Step 9. Ifpo
is not given, then go to Step 11.

Step 9. Solve Problem 4. A unique Zimmermann solution is obtained.

Step 10. Is the solution of Problem 4 satisfying?

IF THEN
1) solution is satisfied -> print out results and stop

2) the user has better -> give the goal b o (and Po)
realized his/her goal and go to step 8
(and its tolerance)
3) resource i, for some i -> decrease b i (and change Pi)
are idle and then go to step 1
4) tolerance i, for some i -> change Pi as desired and go
are not acceptable to step 3

Step 11. Solve Problem 5. That is: call Step 9 to solve Problem 4 for a set of Pos. Of course,
Po E [0, bo-zo:J. Then the solutions are depicted in Table 3.14.

Step 12. Are the solutions satisfying? If yes, print out the solution and terminate the solution
procedure. Otherwise, go to Step 13.

Step 13. Ask the decision maker to specify the refined Po, and then go to Step 9.
 154

Step 14. Obtain the dual problem of the primal problem and then go to Step 3. The dual
problem now is a linear programming problem with fuzzy resources.

Step 15. Solve Problem 6. A solution table (fable 3.20) is then presented to the decision
maker.

Step 16. Does this solution table satisfy the decision maker? We can consider the following
cases:

IF THEN
1) solution is satisfied -> print out results and stop
2) resource i, for some i -> decrease b i (and change Pi)
are idle, and all the and then go to step 1
imprecise coefficients
are fixed
3) tolerance of resource i -> change Pi as desired and go
for some i are not to step 3
acceptable and all imprecise
coefficients,except
resources, are fixed
4) some of the technique -> change aij (Pij)' b i (Pi)'.
and/or objective and/or Cj (Pj)' for some 1
coefficients, and/or and/or j and go to step 15
resources should be
revised

Step 17. Solve Problem 7. A solution table (fable 3.20 or Figure 3.22) is presented. The
decision maker is then asked to choose hislher preferred membership function for the objective
function, either Werners's or Zimmermann's membership function. After that, a unique optimal
solution is obtained.

Step 18. Does the optimal solution or any solution in the solution table satisfy the decision
maker? We can consider the following cases:

IF THEN
1) solution is satisfied -> print out results and stop
2) resource i, for some i -> decrease b i (and change Pi)
are idle, and all the and then go to step 1
imprecise coefficients
are fixed
3) tolerance of resources -> change Pi as desired and go
i, for some i, are not to step 3
acceptable and all imprecise
coefficients,except
resources, are fixed
4) some of the technique -> change aij (Pij) and/or b~
coefficients, and/or (Pi)' for some i and/or J
resources should be and go to step 17
revised
 155

3.3 Some Extensions Of Fuzzy Linear Programming Problems

Beyond the above discussed problems and approaches, we would like to introduce some
extensions of fuzzy linear programming problems in the following four sub-sections. These
extensions could be applied to possibilistic linear programming, multiple objective linear
programming (MOLP), fuzzy multiple objective linear programming (FMOLP) and possibilistic
multiple objective linear programming (PMOLP) as in the following chapter.

3.3.1 Membership Functions

In this chapter, we have introduced some important fuzzy linear programming (FLP)
models based on linear, exponential and triangular membership functions, and max-min
operators. However, there are still other forms of membership functions in the literature such
as piecewise linear, hyperbolic, trapezoid, some specific power and exponential functions, and
so on. Some of the previous fuzzy linear programming models may be used with these types of
membership functions in appropriate situations. It cannot be over-emphasized that the shapes of
the membership functions must depend on the meanings of the particular problems. In this
chapter, we basically use continuous monotonically decreasing functions f<?r "fuzzy ~ "
constraints, continuous monotonically increasing functions for "fuzzy ~" constraints, and
triangular functions for "fuzzy =" constraints. For fuzzy parameters, we also use similar
membership functions to model our problems. In this monograph, we will use triangular and
trapezoid membership functions to model possibilistic linear programming (PLP) problems. But,
this does not mean that PLP have to use triangular and trapezoid membership functions and FLP
have to use continuous monotonically functions. The types of membership functions are
dependent on our problems and their meanings. In this sub-section, we will discuss piecewise
linear membership function as follows. Other considerations of membership functions as
discussed in Chapter 3 can also be applied in FLP problems.
Let us discuss Zimmermann's FLP model (Section 3.1.2) with piecewise membership
functions. Since the piecewise linear function can be used to approach many non-linear
functions, it is noteworthy to be especially introduced here. Furthermore, one way to obtain the
membership function is to ask the decision maker to scale his/her goal(s). For example, suppose
that the analyst provides a set of possible goals, say {45, 50, 55, 60, 65, 70}, to the decision
maker and asks the decision maker to provide his/her degree of satisfaction for each of the
numbers. Then we can obtain his/her subjective membership function of the goal as Figure 3.24,
for instance.
 156

1 ---------------------

0.8

0.6

0.4

0.2

o goal
o 45 50 55 60 65 70

Figure 3.24 A piecewise linear membership function

Hannan, Nakamura, and Inuiguchi, Ichihashi and Kume have provided their models to
solve FLP problems with piecewise linear membership functions. Here, let us discuss
Nakamura's [174] approach.
According to the fact that a fuzzy logical function which contains only A (conjunction) and
V (disjunction) can be reduced to polynomial form with respect to A whose monomials are in V,
Nakamura proposed piecewise linear membership functions which could act as approximations
of convex and/or concave curves. Suppose A to be a convex fuzzy set in Y with its membership
function expressed in a function form which combines only a finite number of linear functions,
q~(y), with A and V operators. Then, the membership function can be represented as:

. q
IJ.A(Y) = [{ A Vj(y)} AI] V 0, Y E Y (3.48)
j=l

rG)
where Vj(Y) = V oit(y).
k=1

Now, let us consider three examples, as shown in Figure 3.25:

 157

iJ.A (y)

y
(a) concave type

o y
(b) convex type

1 -------------- --- ----

y
o (c) convex-concave mixed type

Fiqura 3.25 Fuzzy logical representations of piecewise linear membership

functions
 158

Let us rewrite Zimmermann's model as follows:

Yo = cx ~ zo, Yo f Yo
Y; = (Ax); ~ b;, Y; f Y;, i = 1,2, ... , m
x ~ o. (3.49)

Suppose m + 1 convex fuzzy subset A; in Y; with the piecewise linear membership functions as
Equation (3.49). Then, Equation (3.13), becomes:

max a
s.t. a ~ f;«Bx);) = P-Aj(Y;), i = 0, 1, ... , m
x ~ 0, a f [0,1] (3.50)

or
max a

gl
s.t. a ~ P-Al(Yl) = [( A V1j(Yl)} A 1] V 0
j=1

g2
a ~ P-A2(Yz) = [( A V2j(Y2)} A 1] V 0
j=1

gm+l
a ~ P-Am+1(Ym> = [( A Vm+1j(Ym+l) } A 1] V 0
j=1

x ~ 0, a f [0,1] (3.51)
 159

riG)
where Vi/Yl) = V oiilc(Yi), i = 1,2, ... , m+l, and and JLAi«BX)i) may be in the form of
k=1
Figure 3.26. If riG) = 1 for all i and j, the Equation (3.51) becomes:

max a
s.t. a So; O\}(BX)I)
a So; ~I,I«BX)I)

a otm+ 1m+1 rm+l«BX)m+l)

°
So;

x ~ and a € [0, 1]. (3.52)

It is obvious that Equation (3.52) can be solved by linear programming as Equation (3.13).
In many situations, the convex form and/or the concave-convex mixed form of membership
function may occur. For example, JLAi(Yi) is convex; there then exists at least one Vij(Yi) =
Vk=l,fim oiilc(Yl) with riG) > 1. Thus we should consider the effective region, oijk*«Bx)J, with
the assumption that [d oiilc«Bx)i)]/[d«Bx)i)] =1= 0, all i and j. The solution procedure is described
as follows (see Figure 3.27):

Figure 3.26 The membership function of the objective or the constraint

 160

step o. Determine ~M
formulate (3.51)

step 1. Set 6 0

step 2. Solve (3.53)

and get solution a*

no

yes

step 6: optimal
decision is a*, x*

Figure 3.27 The computational procedure for the fuzzy linear programming
with piecewise linear membership function

Step O. Express ~M«(BX)J as a fuzzy logical function as Figure 3.25 for the membership
functions of the objective and constraints, and then, formate the fuzzy linear programming
problem as Equation (3.51).

Step 1. Set 0 = O.

Step 2. According to the value of 0, the effective regions of diJr.*(I)«(Bx)i) are determined. Then
 161

we solve the following subproblem:

max a
s. t. a ~ oJjk*(6)«Bx)i), ViJ
x ~ 0 and a E [0, I]. (3.53)

Step 3. If a· = a, then go to Step 4, otherwise go to Step 6.

Step 4. If there is no riG) > I, Equation (3.53) will be equivalent to Equation (3.51), and then
go to step 6. Otherwise, go to Step 5.

Step S. Set a = a·, then go to Step 2.

Step 6. a· is the optimal value of the original problem and the corresponding x· is the optimal
decision. STOP.
Finally, assume that aO is the optimal solution of Equation (3.51) and a· is the optimal
solution of Equation (3.53). Then for a E [0, a~ we have aO ~ a· ~a and aO =a iff a· =
a.
3.3.1.1 Example: A Truck Fleet Problem
Consider a truck fleet problem [296] and assume the fuzzy goal and constraints are
modelled by piecewise membership functions.
This problem is to find an optimal decision on the size and structure of the truck fleet of
a company at a minimal cost. The following requirements should be satisfied.
Four differently sized trucks xl' X2, x3 and x4 are considered. The objective is to
minimize cost and the constraints are to supply all customers (who have a strong seasonally
fluctuating demand). That is: certain quantities have to be moved (the quantity constraint) and
a minimum number of customers per day have to be contacted (routing constraint). For other
reasons, it is required that at least six of the smallest trucks be included in the fleet. The
manager wants to use quantitative analysis and agrees to the following suggested linear
programming:

min YI = 41,400xI +44,300x2+48,IOOx3+49,IOOx4 (cost)

s.t. Y2 = 0.84xI+1.44x2+2.16x3+2.40x4 ~ 170 (quantity)
Y3 = 16xI+16x2+16x3+16x4 ~ 1300 (customer)
Y4 = Xl ~ 6 (minimum level)
 162

(non-negative)

where $41,400, $44,300, $48,100, $49,100 are unit costs, and 0.84, 1.44,2.16,2.4 are the unit
loads (in tons) per truck for trucks Xl' X2 , X3 and ~ in the truck fleet, respectively. 170 is the
least total quantity needed to be moved (in tons). The number of customers contacted by each
truck is 6 and the least total number of customers needed to be contacted is 1300.
The solution of this crisp LP problem is x· = (6, 17.85,0,58.64) with the minimum cost
= $3,918,850.
Since the manager feels that (s)he is forced into giving precise constraints in spite of the
fact (s)he would rather have some given intervals, this leads to fuzzy linear programming
problems. Therefore, the following assumptions are agreed to by the manager:

a) Total cost should be between $3,700,000 and $4,200,000. Furthermore, the manager also
provides degrees of his/her satisfaction levels as follows:

1.0/3,700,000, 0.7/3,950,000,
0.25/4,075,000, 0.0/4,200,000.

b) The manager feels much better if there is some leeway on the constraints. After more
detailed study, the manager provided the following data for three constraints, respectively.

1.0/180, 0.9/176, 0.5/175, 0.11172, 0.0/170

1.0/1400, 0.7/1350, 0.4/1325, 0.9/1300
1.0/12, 0.9/11, 0.6/9, 0.25/8, 0.117, 0.0/6.

c) The linear approximations between available points are acceptable. Piecewise linear
membership functions are then assumed (see Figure 3.28).

d) There are no inter-dependencies between the constraints.

Finally we obtain the following fuzzy linear programming problem:

find X

such that Yl = (Bx)! ~ ZO

Yi = (BX)i ~ bi , i = 2,3,4
XI , Xz, X3, ~ ~ O.
 163

J.i.AI

1
I
I
I
______ LI ___________ _
0.7
I
I
I
I
I
I
I
I

0.25
I
------r------------~
I
I
I
_____ _
I I
I I

o 3700000 3950000 4200000

4075000

(a) AI: a l = 8.40 - O. 000002yp 14.92 - 0.0000036YI

a3 = 5.44 - O. 0000012YI

1 -------------------------.-=-.....- - -
0.9 - - - - - - - - - - _____ _

0.5

0.1

Y2
o 170 172 175 176 180

(b) ~: a4 = 0.05Y2 - 8.50, as = 0.13Y2 - 22.83

a6 = 0.40Y2 - 68.30, a7 = O. 025Y2 - 3.50

Pigure 3.28 Piecewise linear membership functions for the objective and
constraints of the truck fleet problem
 164

IJ.A3

1 ------------------------~---

0.7

0.4

~----~------~----+---------------~-------- Y3
o 1300 1325 1350 1400

(e) A3: 178 = 0.016Y3 -20.80, 0.04Y3 - 15.82

1710 = 0.006Y3 - 7.40

IJ.A4

1 ------------------------~---

0.9

0.6

0.25

0.1
--~~--~----r_--_+----------+_----~------ Y4
o 6 7 8 9 11 12

(d) ~: all = 0.10Y4 - 0.60, 1712 = 0.15Y4 - 0.95

0.35Y4 - 2.55, 1714 = 0.15Y4 - 0.75

0.10Y4 - 0.20

Figure 3.28 (Continolls)

 165

Let Ai be convex fuzzy set in Yi = {Yi I Yi ~ bi (hI = i') } for i = I, 2, 3, 4 and their
membership functions (see Figure 3.28) IlAi(YJ be given as follows:

Step O. IlAl(Yl) = [{ al V az } A a3 AI] V 0

1lA2(yiJ = [{ a4 V a5 V a6 } A a7 AI] V 0
IlA3(Y3) = [ as A a9 A alO AI] V 0
IlA4(Y~ = [{ all V alZ V al3 } A au A al6 AI] V O.

Step 1. Let (, = O.

Step 2. Solve the following subproblem:

max ex
s.t. ex S; 0"1 A 0"3 = al
ex S; 0"4 A 0"7 = 0"4
ex S; o"s A 0"9 A 0"10
ex S; O"ll A 0"14 A 0"11 A 0"16 = O"ll
ex E [0, I]

which is equivalent to:

max ex
s.t. ex S; 8.40 - 0.OOOO2YI
ex S; 0.05yz - 8.50
ex S; 0.016Y3 - 20.80
ex S; 0.04Y3 - 15.82
ex S; 0.OO6Y3 - 7.40
ex S; 0.IOY4 - 0.60
ex E [0, I]
or
max ex
s.t. ex S; -0.0828xI-0.0886xz-0.0962x3-0.0982~+8.400

ex S; -0.040XI +0.070Xz+0.I08x3+0.120~-8.500
ex S; 0.256(xl +xz+x3+~)-20.800
ex S; 0.640(xI+XZ+X3+~)-15.820
 166

cx ~ 0.096(xI+X2+X3+~)-7.400

XI' X2, X3, ~ ~ 0 and cx f [0, 1]

whose solution is: x* = (16.667,0,0, 70.323) with cx* = 0.445.

Step 3. cx .. = 0.445 > cx = O. Go to Step 4.

Step 4. r 2G) = r4 G) = 3. Go to Step 5.

Step S. Set () = 0.445. Go to Step 2.

Iteration 2.

Step 2. Solve the following subproblem:

max cx
S.t. cx ~ (12 A (13

cx ~ (1s A (17 = (1s

cx ~ (19 A (110

cx ~ (113 A (114 A (1IS

cx f [0, 1]
or
max cx
s.t. cx ~ 14.920-0.1490xcO.1595~-0.1730x3-0.1768x4

cx ~ 5 .440-0.0497xcO.0532~-0.0577x3-0.0589~
cx ~ 0.1094xI +0. 1872x2+0.2808x3 +0.3120~-22.830
cx ~ 0.640(XI+X2+X3+~)-l5.820

cx ~ 0.096(xI+X2+X3+~)-7.400

cx ~ 0.350x l -2.550, cx ~ 0.150xl -0.750

cx ~ 0.I00xcO.200
Xl> x2 , x3 , ~ ~ 0 and cx f [0, 1]

whose solution is: x* = (8.932, 5.651, 0, 66.503) with cx* = 0.576.

Step 3. cx* = 0.576 > cx = 0.445. Go to Step 4.

Step 4. r 4 G) = 2. Go to Step 5.
 167

Step 5. Set a = 0.576. Go to Step 2.

Iteration 3.

Step 2. Solve the following subproblem:

max a
s.t. a ~ (12 A (13

a ~ (16 A (17

a ~ (19 A (110

a ~ (113 A (114 A (115

a € [0, 1]
or
max a
s.t. a ~ 14.920-0.1490xI-0.1595x2-0.1732x3-0.1768x4
a ~ 5.440-0.0497xcO.0532x2-0.0577x3-0.0589~

a ~ 0.336xI +0.576x2+0.864x3+0.960~-68.300
a ~ 0.21OxI+0.360Xz+0.540X3+0.600~-3.500

a ~ 0.640(xl +X2+X3+~)-15.820
a ~ 0.096(xI+Xz+X3+~)-7.400

a ~ 0.350xl -2.550, a ~ 0.150xl -0.750

a ~ 0.I00xl -0.200
XI, X2 , X3 , ~ ~ 0, and a € [0, 1]

whose solution is : x· = (9.229, 14.715,0,59.747) with a· =0.634.

Step 3. a· =0.634 > a = 0.576. Go to Step 4.

Step 4. riG) = 1, all i. STOP. The optimal solution for the original problem is: x· = (9.229,
14.715,0,59.747) with a· = 0.634.

Test. Solve the following sub-problem:

max a
s.t. a ~ (12 A (13' a ~ (16 A (17

a ~ (19 A (110' a ~ (114 A (115

 168

cx E [0, 1]

whose solution is: x· = (9.229, 14.715,0,59.747) with cx· = 0.634 = 0 and the minimum cost
= 3,970,430.8, so the optimal solution is reached.
The differences between the two solutions are provided as follows:

Original/Crisp System Fuzzy System

··
Decision
Xl 6.00 9.23
17.85 14.72

··
X2
X3 0.00 0.00
x4 58.64 59.75

z
. 3,918,850 3,988,257

Constraints:
1-st 171. 50 172.35
2-nd 1,320.00 1,339.20
3-rd 6.00 9.23

3.3.2 Operators
In this chapter, all fuzzy linear programming approaches are based on the max-min
operator. However, the max-min operator is a non-compensatory operator. But in some
practical situations, a compensatory operator might be required to aggregate fuzzy
constraints/objective(s). For the FLP problem, Zimmermann and Zysno [299] suggested that an
acceptable compromise between empirical fit and computational efficiency seemed to be the
convex combination of the min operator and the max operator. Thus the "compensatory and"
operator was introduced as:

pdx) = r min ILj(X) + (1 - r) max ILj, r E [0, 1]. (3.54)

i

If we use this operator, the original FLP (Section 3.1.2) will become the following mix integer
programming problem [304]:
 169

m m
max {r min lJ.;(x) + (1 - r) max IJ.;}
XEX i=O i=O (3.55)

which is equivalent to :

max ra l + (1-r)az
s.t. lJ.;(x);;:: ai' i = 0, 1, 2, ... , m
lJ.;(x) ;;:: az, for at least one i E {O, ... , m}
XE X
or
max ra l + (1-r)az
s.t. /lj(x);;:: ai' i= 0,1, ... , m
/lj(X) ;;:: az + M,;, i = 0, 1, ... , m
EjYj ~ m
xEXandYjE {O, I}, Vi (3.56)

where Myj are very large real numbers.

Other practical operators are "fuzzy and" and "fuzzy or" proposed by Wemers [259].
"Fuzzy and" is defined as:

/lad = r miDj lJ.;(x) + (1-r)[ Ej /lj(x)]/(m + 1) (3.57)

with r E [0, 1]. And, "fuzzy or" is defined as:

/lor = r mllXj /lj(x) + (l-r)[ Ej /lj(x)]/(m+ 1) (3.58)

with r
.
E [0, 1] .
These two operators (connectives) are not inductive and associative, but they are
commutative, idempotent, strictly monotonic increasing in each component, continuous, and
compensatory. By use of the "fuzzy and" operator, the FLP model (in Section 3.l.2) will
become:

max {r miDj lJ.;(x) + (l-r)[ Ej /lj(x)]/(m + I)} (3.59)

XEX

which is equivalent to :
 170

max a + (l-r)[ I:i ai]/(m+l)

s.t. a + ai S #Li' i = 0, 1, ... , m
x E X, and a, ai' #Li' Vi, E [0, 1]. (3.60)

Finally, we would like to mention that some other operators discussed in Chapter 3 can
also be applied to FLP problems, but some of them will cause computational inefficiencies.

3.3.3 Sensitivity Analysis and Dual Theory

Speaking of linear programming, it seems important to discuss sensitivity analysis and dual
theory. For a fuzzy linear programming, we could study its sensitivity analysis and dual theory
after obtaining its auxiliary crisp linear programming. Thus usual sensitivity analysis and dual
approaches can be applied for fuzzy linear programming (or after some modifications if
necessary) .
Sensitivity analysis and dual theory of fuzzy linear programming problems have been
discussed in the literature of Hamacher, Leberling and Zimmermann [112], Rodder and
Zimmermann [215], Verdegay [250], Kabbara [122], Llena [153], Tanaka, Ichihashi and Asai
[242], and Ostermark [199]. However, for the space limitation of this monograph, we are not
going to discuss these topics. Interested people may refer to the references.

3.3.4 Fuzzy Non-Linear Programming

In this chapter, we have discussed several important techniques to solve linear
programming problems in a fuzzy environment. These techniques can also be applied on some
non-linear programming problems. Like fuzzy linear programming, a fuzzy non-linear
programming problem should be transferred into an equivalent crisp non-linear programming
problem. Then we can solve it by conventional solution techniques or packages of non-linear
programming, instead of linear programming.
As an example, let us discuss Zimmermann's symmetric model for a symmetric fuzzy
mathematical programming problem (see Section 3.1.2). Other models and fuzzy mathematical
programming problems can be applied in a similar way. Consider the following problems:

min ~(x)

s.t. gi(x) ~ bi, i = 1,2, ... , m

x~O (5.61)
 171

which, according to Zimmermann's concept, is equivalent to:

find x
such that 8o(x):S: bo
gi(X) ~ bi' Vi, and x ~ 0 (3.62)

where 80 and gi' may be nonlinear functions. Assume that membership functions are linear as
in Figure 3.6. We obtain the following equivalent problem:

max a
s.t. 8o(x):S: bo + (l-a)Po
gi(X) ~ bi - (l-a)Pi
a E [0, 1] and x ~ 0 (3.63)

where bo, bi, Po and Pi are given goals and their tolerances, respectively. This crisp non-linear
programming problem can then be solved by any non-linear programming solution package, for
example KSV-SUMT or GREG packages.

3.3.4.1 Example: A Fuzzy Machining Economics Problem

Let consider the following fuzzy machining economics (ME) problem [248]. A crisp cost
model for machining a Dw diameter, Iw length workpiece on a lathe, i.e. a single-pass turning
operation, can be written as:

(the restriction of cutting speed)

(the allowable feed rates)
(the maximum cutting force)
(the necessary machining power)
R. :s: R.,mal[ (the limited surface-roughness level)
1 :s: alf :s: 20 (depth of cut and feed rate)
vandf~O (non-negative)

w here the parameters and decision variables are explained in the following:
 172

Cb/Nb: unit set up and preparation cost for each workpiece

Cm: total machine and operator rate
It,: handling time for each workpiece
t;: idle time for each workpiece
~: machining time for each workpiece
t.,: tool change time
t.: tool life
C.: cost of each tool cutting edge
Cm~tJt.: unit tool change cost for each workpiece
Cm~t.,/(t.C.J: unit cost of a sharp tool edge for each workpiece
~ = Iw/[lOOOfv/(1I'Dw)]
ts = C' /(y"'Tp)

lOOOfv lOOOC'

Doubtlessly, defining a machining optimization problem, including the construction of an

objective equation and constraints, is a complicated decision making process. Depending on the
tool material, the workpiece material, and the type of turning process, constraint resources have
to be decided before the entire non-linear programming model can be constructed. Determining
the machining limitations and specifications can involve many human factors. Therefore, the
ME model can be represented by fuzzy non-linear programming to make the model more flexible
and adaptable to the human decision process. Furthermore, the fuzzy ME model emphasizes the
qualitative viewpoint of the decision maker, and thus is more human-oriented.
The fuzzy machining economics model is:

find f and v
such that Ko + K\Cmv-1fl + K2v",T.1F-l :::;; bo
-v :::;; - 1I'DwD.,miu/1OOO, v :::;; 1I'Dw;',max/1OOO
-f :::;; -fmin' f :::;; fmax
KpftiFarp :::;; Fc,max
-
Fcv /(6120o.J :::;; P max
 173

KRv'"¥IRarR :s; i..max

-aft :s; -1, aft :s; 20
fandv~O

where Ko = Cb/Nb +Cmt.. +Cmt;, K, = 'l/"DwVl000, ~ = K,(Cmt. + CJ/C , . If the goal bo, and
all the tolerances of the fuzzy goal and constraints are given, an equivalent crisp model will be
obtained. Let us consider the following numerical data for the fuzzy ME problem:

Tool: ISO SNMA 120XXX-P20

Holder: ISO PBSNR 2525
Material: SAE 1045 CO
Machine: ENGINE LATHE
Feed speed f: 0.05-2.5 mm/rev
Speed n.: 20-1699 rpm
at 80% efficiency
Cm : 0.25/min th: 1.35 min/pc
Co: 0.50/edge ~: 0.2min/pass
Cb : 7. 2/batch te: 1. 0 min/edge
Ow: 100 mm Iw: 250 mm
e: 55mm Nb: 25 parts/batch
P: 0.1 a : 3 mm
Fe: 250y-0012fOo75 a S 170 kp
Po: Fef/6120 S 6.0 kw
R.: 2. 43x10 4v1.52f S 1. 5 ",m
Tolerances for the constraints are 1, 4, 200, 0, 1,
10, 4, 1, 0 and 0, respectively.

The equivalent crisp non-linear programming problem is:

max a
s.t. 0.6755+ 19.635v·'f'+ 1. 178e"yo33 foll :s; bo+(I-a)1
-1000v/314 :s; -20 + (l-a)4
l000v/314 :s; 1600 + (l-a)200
-f :s; -0.05 + (l-a)O
f :s; 2.50 + (l-a)1
25Ov'()o'2fo7S(3) :s; 170 + (l-a)10
0.041vOoSSfo7S(3) :s; 6.0 + (l-a)4
2.43x10V1.S2f :s; 1.5 + (l-a)1
-3f' :s; -1 + (l-a)O
3f' :s; 10 + (l-a)O
 174

f and v ~ O.

Three aspiration costs, bo = 1.4, 1.6 and 1.7, are used for comparison. The solutions for these
aspirations are provided in the following:

Aspiration (Goal) 1.4 1.6 1.7

cost 1.4 1.6 1.65

a (satisfaction) 73.2% 93.3% 100%

v (m/min) 328.48 322.47 316.68
f (mm/r) 0.247 0.215 0.207

3.3.5 Fuzzy Integer Programming

Consider the following fuzzy integer programming problem:
max cx
s.t. (AX)i ~ bi, Vi
x ~ 0 and x integer (3.64)

where the tolerances Pi' Vi, are given for bi. Assume that the membership functions of the fuzzy
constraints are linear. To solve the above fuzzy constraints may be equivalent to the following
auxiliary crisp constraints:

(3.65)

where Wi> Vi, are binary variables attached to the constraints. To solve Equation (3.64), Fabian
and Stoica [83] proposed the following auxiliary crisp model:

max cx + Ei rJbi-(Ax)i]wi
S.t. (AX)i - PiWi ::;; bi, Wi E {O, I}, Vi
x ~ 0 and x integer (3.66)

where ri, Vi, are penalty coefficients. Obviously, Equation (3.66) becomes a non-linear
programming problem.
 175

Example. Consider the following integer linear programming problem:

max 2x. + 5Xz

s.t. 2x. - Xz ~ 9
2x. + 8Xz ~ 31
x., Xz ~ 0 and X"X2 integer.

Assume that r. = r2 = 1, p. = 3 and P2 = 4. Then we have:

max 2x. + 5Xz + (9-2x. +xJw. + (31-2x.-8xJw2

s.t. 2x. - Xz -3w. :s; 9
2x. + 8x2 - 4W2 :s; 31
W.,W2 E {O, I}, x., Xz ~ 0 and x.,Xz integer.

In Figure 3.29, the optimal solution of the crisp model (p. = P2 = 0) is x = (3, 3) and the
objective value is 21. The fuzzy model has the solution x = (4, 3) with the objective value 22.
For this result the second constraint is violated by 1 which means 3.3 % from the right-hand side.

4
19 ... _ 19
...
r-- r---
- 19 19
...... -1- ......
19 19 19 18 16

r--
-
22 _. 22
21 -
... 22 20 18

..
t:---..
3
r--- 1--_

2
18
~ 21 ,21 20

1~
I
18 15/ 15 15
1/

L "
1
I
I
o
o 1 2 3 4 5 6 7 8 9

Figure 3.29 Solution of the numerical example

 176

3.3.5.1 Fuzzy 0-1 Linear Programming

If the goal of the objective in Equation (3.64) is given as bo, then we will have the
following symmetrical model:

find x
such that cx ~ bo
(Ax); ~ b;, Vi
x ~ 0 and x integer.

Now, if ":i, j = 1, ... ,n, are either 0 or 1, the above problem will become the following
0-1 linear programming problem:

find (Xl' ~, ... , xJ

such that (Ax); ~ b;, Vi, and ":i = 0 or 1, Vj (3.67)

where the tolerance p; for b; is given, Vi, and the membership functions of the fuzzy constraints
are assumed to be:

~
if (Ax); < b;
",(x) ={ - [(Ax)rbJ/p; ifb; S (Ax); S b; + p;
if (Ax); > b; + Pi.

Based on the max-min operator, Zimmermann and Pollatschek [303] proposed the following
auxiliary crisp model for Equation (3.67):

max a
s.t. a S I';(x) = 1 - [(Ax);-bJ/p;, Vi
":i = 0 or 1, Vj and a e [0, 1]. (3.68)

For computational reasons, Zimmermann and Pollatschek reformulate Equation (3.68) as:

max {O, min[l, 1 + max min (b;'-(Ax);')]} (3.69)

x 1

where b;' = b/p; and (Ax);' = (Ax)/p;. Then, for a' = maxx mill; [b;'-(Ax);'], Equation (3.69)
 177

will be:

max {O, min[l, 1+ ex']}

s.t. x,; = 0 or 1, Vj. (3.70)

Since ex' = rna"" min; [b/-(Ax);'], the following basic inequality holds:

ex' :;:; min; m~ [b;'-(Ax)/]. (3.71)

The right-hand side of Equation (3.71) then serves as an upper bound for ex'. This equation is
important for the suggested branch and bound algorithm which is imbedded into a LIFO (last-in-
first-out) scanning procedure.
To use the branch and bound algorithm, let us first discuss the following definition.
Let N be the index set of variables. N = {I, 2, ... , n}. I is an ordered subset of N such
that each element of I has either a positive or a negative sign and may be underlined or not. The
sign of the elements in I indicates the values of the variables belonging to the indices in 1. For
example, I = {2, -I} means Xz = 1 and XI =0. The set N - I contains the indices of the
variables (free variables) to which numerical values have not yet been assigned. I is called a
node in a tree of the branch and bound algorithm and contains all relevant information about that
subset of variables which are no longer free variables at this stage.
Branching is done by computing values based on Equation (3.69) or (3.71), respectively,
for not completed nodes. If an increase of the bound can be achieved by adding indices of free
variables to I, one moves from node II to node 11+ 1 • At first II = (/). If a branch of the tree has
been terminated, a new node for further explorations is selected (backtracking). Which of the
nodes is selected is uniquely defined by the order of the elements of 1, considering also whether
they are underlined or not. The right-most element in I not underlined is underlined and all
elements to the right of it are excluded from 1.
Now, let us define:

(3.72)

as the upper bound of ex' for a complete II, where a;/ = a;/p; and A = [a;J This upper bound
can be easily calculated by:

(3.73)
 178

If BL(JJ is smaller than L = mil\ [hi' - Ejlij;K.i] of a known solution x (the so-called potential
solution with L = the potential value), then all II of this branch of the tree will be inferior to the
potential solution and one has to backtrack. On the other hand, if BL(IJ > L, an appropriate
variable for branching has to be selected. For a good selection, we need to calculate the
following two bounds for each free variable:

BL(II U {+j}) and BL(JI U {-j}). (3.74)

If for any variable both bounds are smaller than L, then backtracking will be started for the same
reason as before. If only one bound is smaller than L, the variable will be selected. For
instanc<?, if BL(II U {+j}) < L, then II is augmented by -j. If this condition holds for several
variables, they are all included into II. If both bounds are larger than L for any free variable,
the branching variable is chosen heuristically by using any of the following three rules:

(1) Select j € N - II for which BL(II U {+j}) - BL(II U {-j}) is minimal.

(2) Select j € N - II for which max [BL(JI U { +j}), BL(JI U {-j})] is minimal.
(3) Select j € N - II for which min [BL(II U {+j}), BL(JI U {-j})] is minimal.

By repeating the above procedures, we can obtain a solution cx· = 1 + L where L is initially set
to -1.
It is noted that Rules (1) and (2) are intended to reduce as much as possible the chances
that the alternative branch will have to be explored later, and that Rule (3) is intended to reduce
the chances of completions (termination) in the selected branch. Zimmermann and Pollatschek
have found that Rule (1) is the most effective, while Rule (3) is the least effective. Thus Rule
(1) is primarily used. If Rule (1) does not lead to a unique decision, Rule (2) will be used. And,
if a variable still has not been uniquely determined, Rule (3) will then be applied. In the case
of none of the rules leading to a unique solution, the minimal j is chosen.

Example. Consider the following simplified problem:

find XI' X2 , X3 € {O, I}
s.t. gl(x) = -10xl - 20x2 - 20X3 .§ -45
g2(X) = XI + X2 + X3 .§ 2.5

g3(X) = 2xI + Xz + 3X3 ~ 5

~(x) = O.5x l + 3x2 + X3 ~ 3
 179

where the tolerances (PI' Pl, Pl, pJ = (10, 0.75, 1.5, 2). The membership functions of the
fuzzy constrains are:

<
~
if gl(x) -45
.,(x) = { - [(g,(x)+4S)/IO] if -45 S gl(x) S -35
if gl(x) > -35

< 2.5
~
if gz(x)
,.,(x) = { - [(g,(x}-2.S)/O.75J if 2.5 S gz(x) S 3.25
if gz(x) > 3.25

<
~
if g3(X) 5
",(x) = { - [(g,(x}-S)/I.S] if 5 S g3(X) S 6.5
if g3(X) > 6.5

<3
~
if g..(x)

",(x) = { - [(g,(x)-3)12] if 3 S g..(x) S 5

if g..(x) > 5.

The algorithm starts with L = -1, t = 0 and Jo = (J). Thus the general upper bound of
a'is:

BL(Jo) = miDj [b/ - t j=I,2,3 min(O, 8;/)] = min [112, 10/3, 10/3, 312] = 112.

Since BL(Jo) > L, one (or several) appropriate variable has to be selected for branching.
Therefore, we must consider the following bounds:

BL({ + I}) = miDj [bi' - 8;1' - t j_Z,3 min(O, ~')]

= min [112, 2, 2, 514] = 112 (> L)

BL({-l}) = miDj [b/ - t j _Z,3 min(O, 8;/)]

= min [-112, 10/3, 1013, 312] = -112 (> L)

BL({ +2}) = miDj [bi' - Il;z' - t jzl,3 min(O, 8;/)]

= min [1/2, 2, 8/3, 0] = 0 (> L)
 180

= min [-3/2, 10/3, 10/3, 3/2] = -3/2 « L)

BL({ +3}) = mi~ [hi' - ~3' - Ej =1,2 min(O, ~j')]

= min [1/2, 2, 4/3, 1] = 1/2 (> L)

BL({-3}) = mi~ [hi' -Ej =1,2 min(O, ~j')]

= min [-3/2, 1013, 1013, 312] = -3/2. « L)

Since BL({-2}) < L < BL({ +2}) and BL({-3}) < L < BL({ +3}), 10 is augmented by +2 and
+3. II = {+2, +3}. Now, the only one free variable is xl. We calculate:

BL(JlU{ +I}) = mi~ [hi' -Ej =1,2,3 ay']

= min [112, -2/3, -2/3, -3/4] = 13/4

BL(JlU{-l}) = mi~ [hi' - Ej =2,3 ~/]

= min [-112, 2/3, 2/3, -112] = -112

and choose -I to complete the node II. For the potential solution (Xl' x2' X3) = (0, I, 1), we
obtain:

L = mi~ [h.'
I
- E·=l
J " 2 3 ~·'x.]
~ J

= min [-112, 2/3, 2/3, -112] = -112.

But from BL({-2}) = BL({-3}) = -3/2, it follows for any It including {-2} or {-3} that BL(JJ
~ -3/2 ~ -112 = L. Hence, min[1, I +a'] =I + a' ~ 112 and a* = max {O, min[1, 1 + a]}
= 0. It obviously results that all other branches are of no interest and X = (0, 1, 1) is the
optimal solution with a * = I + L = 1 -112 = 112.

3.3.5.1a Example: A Fuzzy Location Problem

For the network location problems, the p-center and p-median types of problems are most
concerned. The p-center problem is that p new facilities are located with respect to m locations
so as to minimize a maximum of weighted distance between each existing location and its closest
center. On the other hand, the objective function of the p-median problem is to minimize a sum
of weighted distances between each existing facility and its closest median. Both problems have
been transferred by Krarup and Pruzan into a mixed integer programming problem as shown in
Table 3.21.
 181

Table 3.21 Krarup and Pruzan's plant location model

min z = E E (Ci+tij)~ + E fiYi (cost)

i£I j£J i£I
s.t. E ~ ~ ~, j£J (demand)
ifI

kiYi - E ~ ~ 0, i f I (balance)
j£J
uij ~ 0, Vi,j, and Yi = 0 or 1
where
i £ I {l, ... , m}: potential locations for new facilities
j £ J {l, ... , n}: locations to be served by the new facilities
c i: the operating cost per unit for facilities i
tij: the cost of transporting a unit from i to j
f i: the fixed cost of opening the facility i
uij: the units to be transported from i to j
~: the demand by j
k i: a constant greater than the maximum production potential of i
(usually, ~ is replaced by the sum of ~'s)
Yi = 1 if a facility is located at ii = 0 otherwise.

Discrete locations models can also be formulated as 0-1 integer programming problems.
The location problem can then be stated as the optimal location of a number of facilities at some
points of a set of candidate points S so as to serve the entire membership of S.
Let S = {I, ... , i, ... , n} be the set of n potential locations and K = {{dij' jeJ}, ieI} be
a class of subsets of S where dij e K locates a facility at i to serve the rest of the points in the
subset (points in dij). Also, let

if du is in the cover
otherwise.

Then, Darzentas [50] proposed the following model:

 182

mm z = E E IljjYij
id jd

s.t. E Yij ~ 1, id, and Yij = 0 or 1. (3.75)

{dij,jEJ}

where Iljj' Vi,j, represent the cost of serving the group of points of the subset dij by a facility
located at i. The objective of the above problem is to minimize the cost of subsets of S
(members ofK) necessary to cover the members of S, and thus obtain the minimum cost needed,
according to certain criteria quantified by Iljj. The constraints of Equation (3.75) guarantee that
every i E S will be covered by at least one dij E K, since for every i there is a constraint forcing
at least one dij to be in the cover (Yij = 1).
Now, consider the fuzzy version of Equation (3.75) as:

find Yij' Vi,j

s.t. E E IljjYij ::;; Zo

id jd

E Yij ~ 1, iEI, and Yij = 0 or 1 (3.76)

{dij,jEJ}

where the goal Zo and its tolerance Po are given. Furthermore, the membership functions of the
fuzzy constraints are assumed to be:

if E IljjYij ~ 1
if I-Pi::;; E auYij ::;; 1
if E IljjYij ::;; I-pi

where Pi' Vi, are the tolerances of the fuzzy constraints. With the above assumptions, we can
then solve Equation (3.76) by using Zimmermann and Pollatschek's approach.
To illustrate the location problem, let us consider the numerical example of the road
network shown in Figure 3.30 [50] where the points 1, 2, 3, and 4 represent villages. The
population of these villages and the distances between the villages are also given. Our problem
is to optimally locate three facilities in order to serve (cover) each village by only one facility.
 183

Point Population

1 1,100
2 650
3 1,350
4 730

Distance (Km) 1 2 3 4

1 11 7 9
2 11 14
3 7
4 9 14

Figure 3.30 The numerical road network problem

First, let us construct the class K of subsets of the set S = {I, 2, 3, 4}, and assume that
each member of K should not have more than 2,000 total population and the distance between
villages should not be more than 15 Km. Table 3.22 gives the members of K for this example
and all other required information for the model of Equation (3.75).

Table 3.22 Numerical input data for the road network problem

(1) (2) (3) (4) (5) (6) (7)

Members Total Total
of K dj population distance Yij a ij=(4) / (3) aijx1000

1 1 1,100 0 Yll 0 0
2 1,2 1,750 11 Y12 0.006 6
3 1,4 1,830 9 Y13 0.005 5

4 2 650 0 Y24 0 0
5 2,1 1,750 11 Y25 0.006 6
6 2,4 1,380 14 Y26 0.010 10
7 3 1,350 0 Y37 0 0

8 4 730 0 Y48 0 0
9 4,1 1,830 9 Y49 0.005 5
10 4,2 1,380 14 Y410 0.010 10
 184

The crisp model of the location problem can then be formulated as the following 0-1
integer programming problem:

min z = 6YI2 + 5YI3 + 6Y25 + lOy26 + 5Y49 + lOY410

S.t. Yll + YI2 + YI3 + Y25 + Y49 = 1
YI2 + Y24 + Y25 + Y26 + Y410 = 1
YI3 + Y26 + Y48 + Y49 + Y410 = 1
Y37 = 1
Yll + YI2 + YI3 + Y24 + Y25 + Y26 + Y37 + Y48 + Y49 + Y410 = 3
Yij = 0 or 1, Vi,j.

Here, based on the enumeration algorithm, we can search for a solution from Yll to Y410 as in
the constraints of the above problem. When Yij = 1, then row i and columnj are removed, the
number of Yij :1= 0 in the j column are recorded. In this example, Y37 = 1 because Village 3 can
only be served by itself.
The enumeration gives the following feasible cover: (1) YI2 and Y48' (2) Y13 and Y24' and
(3) Y25 and Y48' The locations of the facilities and the villages they serve (cover) are given in the
following:

c:
® 4 ®4

3® 3 e
02

z = 6 z· = 5 z=6

The optimal solution is the cover of YI3 and Y24 and has the minimal cost of 5.
The fuzzy version of Equation (3.76) for this numerical example can also be formulated
and solved if the goal (ZQ) and its tolerance (pO), and the membership functions of the fuzzy
constraints are given (including the tolerances PJ.
Consider the following fuzzy version of the road network problem:
 185

find Yij' Yi,j

S.t. gl(Y) = 6Y12+5Y13+6Y25+ IOy26+5Y49+ IOy410 S; 5
g2(Y) = -(yu + YI2 + YI3 + Y25 + Y49) S; -1
g3(Y) = -(Y12 + Y24 + Y25 + Y26 + Y410) S; -1
~(x) = -(y13 + Y26 + Y48 + Y49 + Y41O) S; -1
Y37 =1
YII + YI2 + YI3 + Y24 + Y25 + Y26 + Y37 + Y4S + Y49 + Y410 =3
Yij = 0 or 1, Yi,j

where the tolerances for the fuzzy constraints are all equal to 1 (PI = P2 = P3 = P4 = 1). For
this particular example, Y37 should reserve one facility. Also, let us rename the variables YII -
Y410 as XI - XIO , respectively. Then, the above problem can be formulated as:

find'S, Yj (Note: X7 = 1)
S.t. gl(x) = 6~+5x3+6xs+ IO~+5~+ lOXIO S; 5
g2(X) = -(XI + ~ + X3 + Xs + ~) S; -1
g3(X) = -(~ + ~ + Xs + ~ + XIO) S; -1
~(x) = -(X3 + ~ + Xs + ~ + x S; -1 lO)

'S = 0 or 1, Yj

where XI + ~ + X3 + ~ + Xs + ~ + Xs + ~ + xlO = 2. With the assumption of linear

membership functions, we then use Zimmermann and Pollatschek's approach to solve this fuzzy
0-1 integer programming problem.
Let L = -1, t = 0 and Jo = (J). Then the general upper bound a' is:

BL(Jo) = millj [b/ - Ej min(O, Ilj/)] = min [5, -1, -1, -1] = -1.

Since BL(Jo) ~ L, one or several appropriate variables have to be selected for branching.
Therefore, we must consider the following bounds:

BL({ +I}) = millj [b/ -Iljl' - Ej=2->10 min (0, Ilj/)] = min [5, 4, 5, 5] = 4
BL({-I}) = millj [b/ - Ej=2_>IO min (0, Ilj/)] = min [5, 3, 5, 5] = 3
BL({ +2}) = -1, BL({-2}) = 3
BL({ +3}) = 0, BL({-3}) = 3
BL({ +4}) = 4, BL({-4}) =3
 186

BL({ +5}) = -1, BL({-5}) =3

BL({ +6}) = -5, BL({-6}) = 3
BL({ +S}) = 4, BL({-S}) = 3
BL({ +9}) = 0, BL({-9}) = 3
BL({ + 10}) = -5, BL({-lO}) = 3.

If BL(U) ~ -1 = L, then U will be inferior to the potential solution. Thus, the remaining
possible candidates for J I are XI' X3' ~, Xg and~. However, Jt , Vt, can only contain two
elements (or variables) because of XI + ~ + X3 + ~ + Xs + ~ + Xg + ~ + XIO = 2. Thus,
we have the following ten possible situations:

JI = {I, 3} => L = min [5, 0, -1, -1] = -1

JI = {I, 4} => L = min [5, 0, 0, -1] = -1
JI = {I, S} => L = min [5, 0, -1, 0] = -1
JI = {I, 9} => L = min [0, 0, -1, 0] -1
JI = {3, 4} => L = min [0, 0, 0, 0] = ° *
JI = {3, S} => L = min [0, 0, -1, 1] = -1
JI = {3, 9} => L = min [-5, 1, -1, 1] = -5
JI = {4, S} => L = min [5, -1, 0, 0] = -1
JI = {4, 9} => L = min [0, 0, 0, 0] = ° *
JI = {S, 9} => L = min [0, 0, -1, 1] = -1.

Obviously, cx· °
= 1 + L = 1 + = 1 at J = {3, 4} and {4, 9}, while cx· = 1 - 1 = at the
I °
other situations. Thus the optimal solutions are: Yl3 = Y24 = Y37 = 1 and Y24 = Y37 = Y49 = 1
where z· = 5. The locations of the facilities and the villages they serve (cover) are given in the
following:

and

.2
3$ 3 8
$2
4 POSSIBILISTIC PROGRAMMING

Stochastic programming has been used since the late 1950s for decision models where
input data (coefficients in LP problems) have been given probability distributions. The pioneer
works were done by Dantzig [47], Beale [9], Tintner [247], Simon [223], Charnes, Cooper and
Symonds [39], and Charnes and Cooper [40]. Since then, a number of stochastic programming
models have been formulated in inventory theory, system maintenance, micro-economics, and
banking and finance. The most recent summaries of the development of stochastic programming
methods are Stancu-Minasian [BM48] and Wets [260].
Even after nearly four decades of development of both theorem and applications, stochastic
programming approaches are still limited in solving practical problems. We believe that the main
problems of applying stochastic programming models are: (1) lack of computational efficiency,
and (2) inflexible probabilistic doctrines which might not be able to model the real imprecise
meaning of decision makers. In practice, the common probability distributions may not be
correct in the decision maker's mind. And the meanings of randomness for some
imprecise/uncertain situations may not be correct, either. Zadeh [284] stated:

"The pioneering work of Wiener and Shannon on the statistical theory of communication
has lead to a universal acceptance of the belief that information is intrinsically statistical in nature
and, as such, must be dealt with by the methods provided by probability theory.

Unquestionably, the statistical point of view has contributed deep insights into the
fundamental processes involved in the coding, transmission and reception of data, and played a
key role in the development of modem communication, detection and telemetering systems. In
recent years, however, a number of other important applications have come to the fore in which
the major issues center not on the transmission of information but on its meaning. In such
applications, what matters, is the ability to answer questions relating to information that is stored
in a database - as in nature language processing, knowledge representation, speech recognition,
robotics, medical diagnosis, analysis of rare events, decision-making under uncertainty, picture
analysis, information retrieval and related areas.

. .. our main concern is with the meanings of information - rather than with its measure -
the proper framework for information analysis is possibilistic rather than probabilistic in nature,
thus implying that what is needed for such an analysis is not probability theory but an analogous -
and yet different - theory which might be called the theory of possibility. "

Since Zadeh [284], there has been much research on the possibility theory. Possibilistic
decision making models have provided an important aspect in handling practical decision making
 188

problems. In this chapter we will discuss some important applications of the possibility theory
to linear programming problems with imprecise coefficients. We call linear programming
problems with imprecise coefficients restricted by possibilistic distributions Possibilistic Linear
Programming (PLP) problems. Unlike stochastic linear programming, PLP models provide
computational efficiency and flexible doctrines, as you will see. Yazenin [272] and Buckley [27]
have given some comments on the comparison of fuzzy/possibilistic and stochastic programming.
It is noted that the possibility measure of an event might be interpreted as the possibility
degree of its occurrence under a possibility distribution, 11"(.), (an analogous to a probability
distribution) in this study.
By considering all possible cases in a PLP problem, we proposed a classification of PLP
problems as shown in Table 4.1. In the next section we will discuss all PLP models. Some
other extensions of PLP will also be discussed in Section 4.2.

Table 4.1 The classification of possibilistic LP problems

(Lai [139])

max/min cx Imprecise coefficient

s.t. AxS b c
x ~ 0 Imprecise b A

Case Problem Approach

1- b- & .....A Prob. 1 Ramik & Rimanek, Tanaka,

Ichihashi & Asai, or Dubois
2. c - Prob. 2 Lai & Hwanq, Rommelfanqer,
Hanuscheck & Wolf, or
Delqado, Verdeqay & Vila
3. b (A dual problem of Problem 2)
4. c, b &A Prob. 3 Lai & Hwanq, Buckley, Neqi
or Fuller
5. c & A Prob. 4 Lai & Hwanq
-
6. q & b

7. A - 1
 189

4.1 Possibilistic Linear Programming Models

Consider the following LP problems:

c.-) c-l
s.t. Ax:S; b, and x ~ 0

where c, A and b may be imprecise with possibilistic distributions. Thus we have eight possible
cases. Except for the crisp LP problem, seven possible problems are listed in Table 4.1 (or in
Figure 1.4). Among those problems, let us first discuss the problem of imprecise resources only.

4.1.1 Linear Programming with Imprecise Resources and Technological Coefficients

Soyster [232] had developed an inexact programming that maximized an objective
function, cx, and was subject to x E X = {x I xIKjI + x21G2 + ... + XnKm C Bi , Vi, and x
~ OJ. Kjj and Bi, Vi,j, are crisp sets. Thus, if one can figure out all Kjj and Bi (as some kinds
of functions with parameters ), then the inexact linear programming may be solved by standard
linear programming or some existing nonlinear programming.
Consider the following PLP problem:

max cx
s.t. Ax:s; ii, and x ~ 0 (4.1)

where A and b are (L-R) fuzzy numbers with possibilistic distributions (11"). To solve Equation
(4.1), Soyster's concept of inclusion treats the inequality constraints as a set-inclusion
relationship. It is quite reasonable because the left and right hand sides are considered as fuzzy
numbers which are fuzzy sets.
In this sub-section we will discuss three approaches proposed by Ramik and Rimanek
[210], Tanaka, Ichihashi and Asai [240] and Dubois [72].

4.1.1.1 Ramik and Rimanek's Approach

Reconsider Equation (4.1) as:

max cx

x~O (4.2)
 190

where iii' ~2' ... ~n' bj, all i, are L-R fuzzy numbers here. Although there are many different
expressions for the L-R fuzzy number (trapezoid fuzzy numbers in this particular case), let us
use the tetrad expression, a = (m, n, a, (3) or a= (m +a Im, n In + 6) where m is the left main
value and n is the right main value with complete membership; a is the left spread and 6 is the
right spread (see Figure 4.1). By use of the set-inclusion concept, Ramik and Rimanek obtained
the following problem:

max cx
s.t. ~Ixi 6) lux2 6) .. •. 6) ~Xn €I bj, Vi
x~O (4.3)

where lilxI 6) ii2x2 6) •... 6) ~nXn = (EjmijXj' Ejlljjxj, Ejaijxj' Ej6jjxj). To solve this
possibilistic linear programming problem, the fuzzy inequality €I should be defined.
To define this fuzzy inequality, we may use any ranking approach discussed in Chapter
2. However, let us only introduce Ramik and Rimanek's approach here.
Consider two supposed fuzzy numbers ~j = (mjj' Iljj' aij' 6ij) and bj = (Pj, qj, Tj' OJ) (see
Figure 4.1). Ramik and Rimanek asserted:

Fiqure 4.1 The representation of fuzzy inequality of two L-R fuzzy

numbers a ij and b i
 191

Excluding the case of xI = x2 = ... = Xn = 0, we obtain the following crisp linear

programming problem from the facts of Equations (4.3) and (4.4):

max cx
s.t. Ej mijXj S Pi
Ej (mij - (Xij)Xj S Pi - Ti
Ej nijXj S ~
E·J (n"'J.. + JHx.
1J J s q.1 + a·
l'
Vi
X ~ O. (4.5)

The optimal solution of Equation (4.5) can be considered as the solution of the possibilistic linear
programming problem of Equation (4.2).

4.1.1.1a Example: A Profit Apportionment Problem

Consider the problem of profit apportionment in concern with mutual ownership [200].
The accounting problems of consolidated reporting can be divided into three broad categories:

a) elimination of reciprocal items, such as internal receivables and debts between the entities of
the concern;
b) consolidation of nonreciprocal items, that is, aggregation of the resources and debts of the
concern with respect to the outside world;
c) profit apportionment between concern majority and minority.

Ostermark assumed that points a) and b) have been solved according to some accepted
consolidation procedures. The profit apportionment problem for a concern with mutual
ownership is focused here.
Suppose' that the concern structure of mutual ownership is shown as Figure 4.2. The
figures along the arrows in Figure 4.2 represented the relative shares of th~ common stock owned
directly by the numbers of the concern.
The separate net income was as follows:
 192

Profit:
$30,000

PC

==> 50% 70%

Profit:
$10,000

Figure 4.2 Concern structure (PC = Parent Company, S = Subsidiary)

net income in PC: $30,000

Sl: 1,000
~: 2,000
S3: 3,000
S4: 4,000

consolidated net profit $40,000

As indicated in the figures, each subsidiary owns 12.5% of the common stock in the parent
company, whereas the parent company owns 70% of the common stock in each subsidiary. For
each subsidiary, there is an external minority holding 30 % of the shares. Since each subsidiary
Sj ( i = 1, ... , 4) owns 12.5% of the parent company, this fraction is affected by the mutual
ownership relation of the concern. We can then obtain the following information:

a) Mutual holding matrix:

(owned)
PC S

PC o 0.7 0
(owners) ==> Q [
S 4x.125 0 0.5
 193

b) Direct holdings:

(owned)
PC S

(owners)
CI
HI
0.5

0 0.3
0
==> R
[ 0.5

0
0

0.3 ]
where CI = Concern Interest and MI = Minority Interest.
c) External profits:

(owned)
PC S
PC 30000 0
(owners) ==> B o]
s o 10000 10000

First, Weil's basic formula for profit apportionment in concerns can be stated as:

where I = identity matrix, X = profit apportionment matrix (unknown), e = a column vector

with 1 in each position, and Q, R and B as indicated above. Obviously, we can easily obtain the
solution: (Xe)T = [28460, 11540]. That is: of the consolidated profit, 71.15% is apportioned
to the majority interest and 28.85% to the minority interest.
Wallin introduced Zimmermann's fuzzy linear programming concept into the profit
apportionment problem. The following model was then obtained:

max aK + aM
s.t. R(I-Qr1Be + Ca S Xe
aK and aM E [0, 1]

where the direction of the inequality in the first constraint depends on the form of the trade-off
between majority and minority, and
 194

The objective function in this linear programming contains two scalars, one for the concern
majority (aK) and one for the external minority (a~. The matrices R, Q, B and C are treated
as constants: the future ownership relations (matrix R and Q), the expected adjusted net profits
of the concern members (B) and profit aspirations (C) are assumed to be known. Thus, with aK
and aM confined in the unit interval, the solution to the problem can be interpreted as a profit
apportionment program, where majority and minority interests are simultaneously recognized.
Now, let us assume that A = R(l_Q)"I, W = Be and b = Xe. From the numeric data we
obtain:

A = (0.7692 0.5385)
0.2308 0.4616

and assume that:

bT = [ 25000, 12000].

The matrix C gives the trade-off intervals for the majority and minority interests, while the vector
b contains the starting point for the trading. The following resulting problem is then obtained:

max aK + aM

s.t. 0.7692w I + 0.5385W2 - 10000aK ~ 25000

0.2308w 1 + 0.4616w2 + 3000aM ~ 12000
aK and aM EO [0, 1] and WI and W2 ~ O.

The sign of the elements in C, together with the direction of the inequalities, reveal that the
problem formulation is based on the majority's (or concern management's) point of view. The
first element in vector b, 25000, represents the minimum acceptable level of majority's profit
share, whereas 12000 is the maximum tolerable level of minority's profit share. From majority's
point of view, the best conceivable solution would be one yielding aK = aM = 1; since this
would increase majority's profit share to the maximum value on the trade-off interval: 25000 +
 195

1()()()() = 35000; while simultaneously minimizing minority's profit share: 12000 - 3000 = 9000.
It is noted that the optimal shares of the majority and minority can be obtained directly by K =
0.7692wI + 0.5385w2 and M = 0.2308w I + 0.4616w2.
The management of the concern is uncertain of the maximum feasible profit share of the
minority (bM) and of the trade-off interval against the majority's profit share. Drawing upon data
from the economic situation of the concern and recognizing issues of business policy and
company law, management agrees upon the following estimation: from majority's point of view,
the best feasible ceiling for minority's profit share is in the interval bM E [12000, 2()()()()], with
left spread = 2000 and right spread = 3000. The best feasible trade-off against majority's
interest is achieved by a trade-off interval between 3000 and 8000, with left spread = 1000 and
right spread = 2000. Finally, we obtain the following possibilistic linear programming problem:

s.t. 0.7692wI + 0.5385w2 - l()()()()aK ~ 25000

(0.2308,0.2308 ,0,0)w 1+ (0 .4616,0.4616,0,0)w2 + (3000,8000, l000,2(00)aM

::::; (12000, 2()()()(), 2000,3(00)
aK and aM E [0, 1] and WI and W2 ~ O.

By using Equation (4.5), the above possibilistic linear programming can then be transferred as
follows:

s.t. 0.7692wI + 0.5385w2 - l()()()()aK ~ 25000

0.2308w I + 0.4616w2 + 3000aM ::::; 12000
0.2308w I + 0.4616w2 + 2000aM ::::; 1()()()()
0.2308w I + 0.4616w2 + 8000aM ::::; 2()()()()
0.2308w I + 0.4616w2 + 6000aM ::::; 17000
aK and aM E [0, 1] and WI and W2 ~ O.

The optimal solution is aK' = 0.1662 and aM' =1. According to this optimal program, the
parent company should only show profit. Thus the optimal profit apportionment to the problem
is: majority's profit share = 0.7692w I = 26662 and minority's profit share = 0.2308w I = 8000.
 196

4.1.1.2 Tanaka, Ichihashi and Asai's Approach

Tanaka, Ichihashi and Asai [240] assumed that the fuzzy coefficient (~j or hi) is a
symmetric triangular possibilistic distribution as follows (see Figure 4.3 also):

1 - 2(~r~l>I(~t-~l) if ~l s ~j s ~l + (~t-~l)12
"'aij(~j) =\ 1 - 2(~jU-~j)/(~t-~l) if ~l+(~t-~l>I2S~jS~t
o otherwise
or
"'aij(~j) ={ 1- I ~r(~t+~l> 1/(~t-~l) if ~l S ~j S ~t
o otherwise (4.6)

where ~t and ~l denote the upper and lower limits of the O-level set ~j' respectively. Let ~j
= (~l, ~jU) denote this type of symmetric triangular possibilistic distribution.
Consider the following fuzzy linear function:

The possibilistic distribution of Yi can be obtained as:

ifx=Oandy=O
ifx:l=O
if x = o andy :1= 0 (4.7)

Figure 4.3 A symmetric triangular possibilistic distribution

 197

From Equation (4.7), the a-level set of the fuzzy set Yi ('I/"yi(yJ ~ a) is the following
interval:

Since x ;;=: 0, we obtain:

(4.8)

Thus the possibility of Yi is easily calculated in the case of a linear function.

Based on the extension principle, Tanaka et al. defined "bi is greater than Yi" as "bi ;;=: Yi
< = > max [hi> ya = bi." Thus they considered the problem of Equation (4.1) as:

max cx
S.t. Yi = 8;x ~a bi> Vi (4.9)

where a is specified by the decision maker a priori and Yi = 8;x ::;;a bi if (YJhu Sa(bJhu and (yJhl
::;; (bJhl hold for hE [a, 1] as in Figure 4.4. The upper bound of (bi)a should be greater than or
equal to the upper bound of (Yi)a, and the lower bound of (bJa should be less than or equal to the
lower bound of (yJa. Here, level a corresponds to the degree of optimism of the decision maker.
By using the above results, Equation (4.1) might be equivalent to the following auxiliary
problem:

1 --------------

Pigure 4.4 Order relation between fuzzy sets Yi Sa bi

 198

max cx
s. t. {(a/2)at + (1-a/2)a/} x ~ (a/2)b;" + (1-a/2)b/, Vi
{(I-a/2)a;u+(a/2)a;I}x ~ (l-a/2)biu+(a/2)b/, Vi
x ~ 0 (4.10)

where a is given a priori by the decision maker.

Example. Consider the following simplified possibilistic linear programming problem:

max 25xI + 18~

s.t. (12, 18)xl + (32, 36)~ ~a (750, 850)

(19, 21)xI + (7, 13)x2 ~a (380, 480)

Assume that a = 0.4 is given. The equivalent crisp linear programming is then:

max 15xI + 18x2

s.t. 16.8xl + 35.2x2 ~ 830
13.2xI + 32.8x2 ~ 770
20.6x l + 11.8x2 ~ 460
19.4xI + 8.2x2 ~ 400
XI and X2 ~ O.

The optimal solution is: x· = (12.14, 17.78) and the optimal value is about 444. In this
problem, the upper values of b l and b2 on the right-hand side provide restrictions, while there
are leeways for the lower values of b l and ~.

4.1.1.3 Dubois's Approach

Dubois [72] expressed the fuzzy numbers (in Figure 4.1) as fuzzy intervals which carry
at once several possible ranges of values together with associated levels of confidence. The
interval of [lIl;j' Il;j] with possibility = I is the most uncertain interval and the interval of
[mij+aij, Il;j+l3iJ with possibility = 0 is the boundary of certain intervals. Thus, the degree of
the possibility can be interpreted as the compliment of the level of confidence. By use of the a-
level concept, (a;j)", has a level of confidence 1 - a: the smaller a, the more certain we are about
x taking its values in (a;j)",. Dubois [72] stated:
 199

"The use of fuzzy intervals does not presuppose the existence of statistical data (although
statistical interpretations do exist). They are manipulated by means of lattice operations, and do
not rely upon an additive structure. Fuzzy intervals are a nature model of soft restrictions of
parameter values, as specified by experts. The calculation with fuzzy intervals is a direct
multiple-valued extension of sensitivity analysis; hence we may expect results which are less
precise than what a probabilistic analysis would yield. But this loss in precision is balanced by
a gain in computational efficiency: the required amount of calculation for fuzzy optimization
problem-solving is often significantly less than that required for the same problem under a
stochastic approach. "

Let us consider a more general fuzzy constraint as follows (equality constraint):

f(x, a) = g(x, b) (4.11)

n n
where f(x, i) = r: ~X; + 8;.+1 and g(x, b) = r: biX; + b + 1•
D

i=1 i=1

a
The respective ranges A and Bof and b are supposed to be separable, i.e., are Cartesian
products of fuzzy intervals Ai and Bi, i = 1, 2, ... , n + 1, defined by:

n+l n+l
'irA = min 'lrAi and 'Irs = min (4.12)
i=1 i=1

This assumption means that the variables Il; (and b;) are not linked. The fuzzy intervals Ai and
Bi are represented by means of the L-R representation as in Section 4.l.1:

Now, g(x,B) can be expressed by means of fuzzy addition and scale multiplication:

where Xa+1 = l. Similarly, f(x, A) can be also expressed in the same way.
For the fuzzy constraints of Equation (4.11), Dubois provided two cases of "soft" and
"hard" equivalent constraints. First, let us discuss the "soft constraints."

(a) Soft constraints can be expressed by the a-,weak feasibility (a-WF) of x. That is:

WF(x) = Poss(f(x,a) = g(x,b» = "lI'"(Ax, ax) ~ a (4.13)

 200

where a is a required possibility threshold. Furthermore, Dubois obtained the following results
(see Figure 4.5):

x is a-weak feasible iff 'Ir[Bx,OO)·l(a) ~ 'Ir<.oo,Axl·1(a) and

1I'[Ax,OO)·1(a) ~ 'Ir<.oo,BxJ1(a) (4.14)

where 1I'[Ax,OO)·1(a) = Ej x;(mj-ajLj(.».

Similar expressions hold for 1I'[Bx,OO)·1(a), 1I'<.oo,Axl·1(a) and 'Ir<.oo,sxl·l(a). Thus, the weak
feasibility constraints WF(x) ~ a are equivalent to two linear inequality constraints which
express Equation (4.14):

n+1 n+l
E (pj - rjL'j·l(a»x; ~ E (11; - ,8jR;.l(a»x; (4.15)
i=l i=l

n+1 n+1
E (mj - ajLj·1(a»x; ~ E (q; - OjR'j·l(a»x;. (4.16)
i=l i=l

Bx Ax
1 ---------~--------~

al

o Ax/Bx
"[Ax,-)" (a) "<--,Ax]" (a)

"[Ax,.)" (al)="< __ ,BX] -, (al)

Figure 4.5 Soft constraint using possibility index

 201

However, we may have the following possibility linear programming problem:

min cx
s.t. Ax = b and x ~ 0 (4.17)

where b = (p, 00, r, 00) and Ax are the same as previously described. Equation (4.15) is only
meaningful and Equation (4.17) is equivalent to:

min cx
s.t. P - r-1L(a) ~ E j (Ilj + pj~-l(a»Xj.
x ~ o. (4.18)

It is noted that in PLP problem b is often (p, q, r, 0), instead of (p, 00, r, 00). Thus Equations
(4.15) and (4.16) are often considered at the same time. The original equality is then substituted
by two inequality constraints.

(b) A hard constraint is considered as a fuzzy extension of the set-inclusive constraint Ax cBx.
x is said to be a-hard feasible iff:

SF(x) = Nes(g(x,b) =f(x,ii) I a) = Nf(x,A)(g(x,B» ~ a

that is (see Figure 4.6):

(4.19)

Finally, Dubois obtained the following linear constraints:

n+1 n+1
E [Pj - rjL'j-l(a)]Xj ~ E [mj - a jL j-l(l-a)]Xj (4.20)
i=l i=1
n+1 n+1
E [Ilj - pj~-l(1-a)]Xj ~ E [~- OjR'j-l(a)]Xj. (4.21)
i=1 i=1

When aj = Pj = rj = OJ = 0, Soyster's set-inclusive constraints of the form xTA c xTS are

clearly recovered.
Consider the following problem:
 202

1
---,------
__ -\LO--
Ax Bx

0:1
\
\
0:2

0:

o
'Ir)AX,WI- 1 (1-0:)

J..'[BX,w) -1 (0:2) = J..'( -",Ax[ -1 (0:2)

J..')AX,WI- 1 (0:1) = J..'(-oo,Bx)-I(al)

Figure 4.6 Hard constraint using necessity index

min cx
s.t. Ax:::; b and x 2:= 0 (4.22)

where b = (p, 00, r, 00) and Ax is the same as previously described. By using hard constraints,
we obtain:

mm cx

n+l
S.t. P - r-1L(a):::; E (mj + a jLj-I(I-a»x;
i=1

x 2:= O. (4.23)

Obviously, if the decision maker provides a value, the problems of Equations (4.18) and
(4.23) will become crisp linear programming problems.
 203

4.1.2 Linear Programming with Imprecise Objective Coefficients

In practice, the objective coefficients are always imprecise, such as the unit costs/profits
of new products or new projects, the lending and borrowing interest rates and cash flows.
Assume that this imprecise nature can be modelled by the current possibility theory and that the
objective coefficients have their individual possibility distributions. Thus, we cannot directly
solve this kind of possibilistic linear programming problem by use of any existing linear
programming technique. Auxiliary programming should be first obtained.
This problem has been intensively discussed since 1987. In the following, we would like
to present the three approaches ofLai and Hwang [138], Rommelfanger, Hanuscheck and Wolf
[216], and Delgado, Verdegay and Vila [56].

4.1.2.1 Lai and Hwang's Approach

In this section, we will first discuss a linear programming problem with imprecise
coefficients which can be stated as (Lai and Hwang [138]):

max E j CjXj
s.t. x E X = {x I Ax S b and x ~ o} (4.24)

where Cj = (Cjm, cjP, CjO) , all i, are imprecise unit profits and have triangular possibility
distributions as in Figure 4.7. Cjm is the most possible value, CjP is the most pessimistic value
and e;.0 is the most optimistic value. When normalized, 'lrj(Cjm) = 1 and 'lrj(e;.P) = 'l\'"j(Cj~ = o.
Now, let us reconsider Equation (4.24) as:

max E j (CjmXj, CjPXj, CjOXj)

XEX

or
max (cmx, cPx, COx)
XEX (4.25)

where cm = (cl m, c"m, ... cnm) , cP = (cIP, c"P, ... cl) and CO = (cIO, c,,0, ... cnO). The
objective function is actually an imprecise objective function with a triangular possibility
distribution (cmx, cPx, COx).
To solve Equation (4.25), Tanaka, Ichihashi and Asai [240] provided the weighted average
of the upper and lower limits and then substituted this average into the problem. This results in
the following problem:
 204

Figure 4.7 The triangular possibility distribution of c i

max Ej [wjcr + (l-wj)cjO]X;.

xeX (4.26)

We have provided a most likely solution for a multiple objective linear programming
problem with imprecise objective coefficients. This most likely solution is found by solving the
following auxiliary linear programming problem of Equation (4.25):

max [(4cm +cp +co)/6]x.

xeX (4.27)

If the minimal acceptable possibility, a, is given (here, we do not imply the a-level set which
is a crisp set), we will obtain the following problem:

max [(4cm +cPa +coJ/6]x

xeX (4.28)

where CPa and COa are the most pessimistic and the most optimistic value of the acceptable events,
respectively. It is noted that the possibilities of the acceptable events (the events with the
possibility ~ a) are kept, because we do not like to lose any valuable information at this stage.
Keeping the possibility is important, especially when an interactive procedure is used. In regard
to the solution procedure, the acceptable events with their possibilities are correctly presented by
these three critical points. Therefore, the difference between Equations (4.26) and (4.27) is that
Equation (4.27) includes the most possible value. Thus, we will have different solutions for
 205

different possibilistic distributions, as in Figure 4.8. Equation (4.26) will provide the same
solution for the situation shown in Figure 4.8, where ex should be preferred to c'x if the minimal
acceptable degree is a and the possibility distributions are considered. Of course, we might also
like to change the weights among cm, cP" and co"' respectively.

1[.
....
cx
1

o c'x/cx
c'Px c'·x c·x

Figure 4.8 ex is preferred to c'x if the minimal accepted

grade of possibility is a

On the other hand, we may reconsider the fuzzy objective function with a triangular
distribution. This fuzzy objective is fully defined by three corner points (cmx, 1), (cPx, 0) and
(COx, 0), geometrically. Thus, maximizing the fuzzy objective can be obtained by pushing these
three critical points in the direction of the right- hand side. Fortunately, the vertical coordinates
of the critical points are fixed at either 1 or O. The only considerations then are the three
horizontal coordinates. Therefore, our problem is to solve:

max (cmx, cPx, COx)

xeX

where (cmx, cPx, COx) is the vector of three objective functions, cmx, cPx and cOx. In order to
keep the triangular shape (normal and convex), of the possibility distribution, it is necessary to
make a small change. Instead of maximizing these three objectives simultaneously, we are going
to maximize cmx, minimize [cmx - cPx] and maximize [(COx - cmx], where the last two objective
functions are actually relative measures from cmx, the first objective function (see Figure 4.9).
 206

The three new objectives also guarantee the previous argument of pushing the triangular
possibility distribution in the direction of the right-hand side.

I
I
I
(I) I (II)
I
I
0 cx
cPx c·x cox
Figure 4.9 The strategy to solve "max ex"

This leads us to the following auxiliary problem of Equation (4.25):

min Zl = (cm - cP)x

max ~ = cmx
max Z:! = (CO - cm)x
s.t. XE X. (4.29)

The crisp MOLP of Equation (4.29) is equivalent to maximizing the most possible value
of the imprecise profit (at the point of possibility degree = I). At the same time, we have
minimized the inferior side of the possibility distribution. This means minimizing the region (I)
which, in our opinion, is equivalent to "the risk of obtaining lower profit." And, we have also
maximized the region (II) of the possibility distribution, which is equivalent to "the possibility
of obtaining higher profit." As in Figure 4.9, we would prefer the possibility distribution ofB
to that of A. In portfolio theory, maximizing mean return, minimizing variance and maximizing
skewness are often considered [BM9], while the considered return has a probability distribution.
Clearly, our strategy is essentially analogous to that used in portfolio theory.
To solve Equation (4.29) we may use any MOLP technique (Hwang and Masud [BM19])
 207

such as utility theory, goal programming, fuzzy programming or interactive approaches.

However, in this study we suggest using Zimmermann's fuzzy programming method [274] with
our normalization process.
First we should get the following Positive Ideal Solutions (PIS) and Negative Ideal
Solutions (NIS) of the objective functions (Hwang and Yoon [BM20] and Lai and Hwang [143]):

ZlPIS = min (cm - cP)x, Z1NIS -- max (cm - cp)x

xeX xeX

~PIS = max cmx, ~NIS = min cmx

xeX xeX

zlIS = max (CO _ cm)x, Z3 NIS = min (CO _ cm)x.

xeX xeX (4.30)

The linear membership function of these objective functions can now be computed (see Figure
4.10) as:
if zl < ZlPIS

if zlPIS S zl S ZlNIS
if zl > ZlNIS (4.31)

>
~~_~NIS)/(~PIS_~NIS)
if~ ~PIS
JJ.z2 = { if ~PIS S ~ S ~NIS
. 0 if ~ < Z2N1S (4.32)

and JJ.z3 is similar to JJ.z2. This normalization has also been applied by Seo and Sakawa [BM45]

J.l. z, / J.l. z2

Figure 4.10 The membership functions of the objectives, zl and z2

 208

Finally, we solve the following Zimmermann's -equivalent single-objective linear

programming model:

max a
s.t. lJ.D(x) ~ a, i=I,2,3, and x E X. (4.33)

The optimal solution of Equation (4.33) will provide a satisfying solution under the previous
strategy of minimizing the risk of lower profit, and maximizing the most possible value and the
possibility of higher profit.
For minimization linear programming (minimize the total cost), we will have the following
equivalent problem:

max (cm- cP)x

min cmx
min (CO - cm)x
S.t. x EX (4.34)

which can be solved by using a similar approach (as previously described).

Now, let us look at the following simplified investment problem from Gould, Eppen and
Schmidt [BM13] with some modifications.

4.1.2.1a Example: A Winston-Salem Development Management Problem

Winston-Salem Development Management (WSDM) is trying to complete its investment
plans for the next 3 years. Currently, WSDM has $2 million available for investment. At 6-
month intervals over the next 3 years, WSDM expects the income stream from previous
investments (in millions), as shown in Table 4.2. In this particular case, $300,000 at the end of

the third year will not play any important role.

There are three development projects in which WSDM is considering participating. If

WSDM participated fully in the Foster City Development, it would have the projected cash flow

stream (in millions) at 6-month intervals over the next 3 years as shown in Table 4.2 where the

negative numbers represent investments, and positive numbers represent income. (5.5, 5.1, 6.2)

is an imprecise number and has a triangular possibility distribution (see Figure 4.11) with the
 209

Table 4.2 The multiple project and multiple period investment

problem (in million dollars)

period (year) 0 0.5 1 1.5 2 2.5 3

Income stream 2.0 0.5 0.4 0.38 0.36 0.34 0.3

Project 1 -3.0 -1.0 -1.8 0.4 1.8 1.8 (5.5,5.1,6.2)

Project 2 -2.0 -0.5 1.5 1.5 1.5 0.2 (-1,-1.2, .85)
Project 3 -2.0 -2.0 -1.8 1.0 1.0 1.0 (6.0,5.0,6.5)

Note. 1) Total outstanding principle ~ $2.0 million.

2) At most, $2.0 million can be borrowed at one time.
3) Invest surplus funds at 4% per half year.
4) The borrowing interest rate is 6% per half year.

most possible value = 5.5, the most pessimistic value = 5.1 and the most optimistic value =
6.2. A second project involves taking over the operation of some old, lower-middle income
housing on the condition that certain initial repairs to it be made and that it be demolished at the
end of 3 years. The third project is to invest in the Disney-Universe Hotel. The cash flow
streams for these two projects, if fully participated in, are also shown in Table 4.2. WSDM can
borrow money for half-year intervals at 6% interest per half-year. At most, $2 million can be
borrowed at one time; that is, the total outstanding principle can never exceed $2 million.
WSDM can invest surplus funds at 4% per half-year.
Consider the problem of maximizing WSDM's net worth at the end of 3 years. Disregard
taxes, and assume that if WSDM participates in a project at less than 100%, all the cash flows
of that project are reduced proportionately. Then this problem can be formulated as a linear
programming model. In order to do this, let:

F = fractional participation in Foster City

M = fractional participation in Lower-Middle
D = fractional participation in Disney-Universe
Bj = amount borrowed in period i, i = 1, ... , 6
Lj = amount loaned in period i, i = I, ... , 6
Z = net worth after the six periods
(without considering $300,000)

The linear programming model can be formulated as (in millions):

 210

max Z = (5.5,5.1,6.2)F+(-1.0,-1.2,-0.85)M+(6.0,5.0,6.5)D-1.06B6+1.04L6
s.t. 3F+2M+2D-B 1 +L 1 S 2
IF+0.5M+2D+1.06Bcl.04Ll-Bz+~ S 0.5
1.8F-1.5M+1.8D+1.06Bz-1.04~-~+~ S 0.4
-0.4F-1.5M-ID+ 1.06B3-1.04~-B4+L4 S 0.38
-1.8F-1.5M-ID+1.06B4-1.04L4-Bs +Ls S 0.36
-1.8F-O.2M-ID+ 1.06Bs-1.04Ls-B6+~ S 0.34
Bi E [0, 2], i = I, 2, 3, 4, 5, 6,
F, M and DE [0, I]. (4.35)

o Income of Project 1
5.1 5.5 6.2 at the end of the third
year

o Investment of Project 2
-1.2 -1.0 -0.85 at the end of the third
year

o Income of project 3
5.0 6.0 6.5 at the end of the
third year

Figure 4.11 The possibilistic distributions of the cash flows of three

projects at the end of the third year
 211

It is noted that the most pessimistic and optimistic values should be reversed when the
coefficients are negative for the maximization problem. Equation (4.35) can then be solved by
the following auxiliary problem:

min 0.4F-O.2M + D
max 5.5F-1.0M +6.0D-1.06B6+ 1.04L6
max 0.7F-O.15M+0.5D
s.t. xEX (4.36)

where X is defined as the constraint of Equation (4.35).

First, we obtain the PIS and NIS of Equation (4.36) which are (0.2, 9.00, 0.5S) and (0.S4,
3.12,0.15), respectively. Then, according to Equations (4.31), (4.32) and (4.33), the solution
is obtained and shown in Table 4.3. Furthermore, Figure 4.12 shows the possibilistic distribution
of the imprecise profit obtained. The upper and lower bounds of the profits of the problem
(S.72, 7.91), and the most likely profit (S.20) are obtained.

71 z
(4.61) (4.35) (4.57)
1

o z
profit

Figura 4.12 The possibilistic distributions of the optimal profits of

Equations (4.35), (4.57) and (4.61)

It is worthy noting that the possibilistic linear programming problem (PLP) studied here
is different from some fuzzy linear programming (FLP) problems in meaning. The FLP is based
on the preferred concept to establish membership functions. On the other hand, the PLP is
based on the degree of the occurrence of an event to obtain possibilistic distributions. In other
 212

Table 4.3 Solution of Equations (4.35), (4.57) and (4.61)

variable Solution of Solution of Solution of

Equation (4.35) Equation (4.57) Equation (4.61)
at P = 0.5 at P = 0.5
ex 0.8634 0.8861 0.9149

F 0.7642 0.7254 0.7935

M 0.0914 0.0365 0.1650
D 0.0000 0.0000 0.0000

0.4755 0.2491 0.7106

0.0000 0.0000 0.0000
2.0000 2.0000 1.1292
1.1860 1. 4924 0.0000

1. 7251 1. 4188 2.0000

0.0000 0.0000 0.0000
0.0000 0.0000 1.1851
0.2276 0.7736 0.0000

0.0000 0.0000 0.0000

2.1094 2.5280 0.7667

0.0000 2.0000 2.0000

3.9277 6.2983 4.5818

PIS for z, 0.20 0.23 0.23

z,· 0.287 0.291 0.287

NIS for z, 0.84 0.85 0.84

iJ. z, ( z, .) 86.3% 90.3% 91.5%

9.00 9.09 7.19

8.20 8.41 6.86

NIS for Z2 3.12 3.13 3.12

iJ. z2 (Z2·) 83.0% 88.6% 91.9%

0.58 0.63 0.64

0.521 0.577 0.521

NIS for Z3 0.15 0.16 0.16

iJ. z3(Z3*) 86.3% 88.6% 87.7%
The optimal profit is imprecise and has a triangular
possibility distribution of (Z2·' Z2*-Z,*, Z2*+Z3*).
That is -- Equation (4.35): (8.20, 7.91, 8.72)
Equation (4.57): (8.41, 8.12, 8.99)
Equation (4.61): (6.86, 6.58, 7.44).
 213

words, the FLP is an analogous of utility approaches, and the PLP is an analogous of probability
methods. However, after transferring the possibilistic linear programming problem to an
auxiliary crisp MOLP problem, we then use Zimmermann's fuzzy programming (preference-
based membership functions of the objective functions) to obtain the compromise solution.
Without using the l3-level set concept, we can easily solve the imprecise objective function and
obtain the whole possibilistic distribution of the satisfying objective value. Furthermore, we can
solve rather general, imprecise problems through an interactive process with the decision maker.
Three critical values (the most possible, the pessimistic and the optimistic values) are only
necessary for each imprecise object. It is very important that the satisfying objective value
should be imprecise since the unit profits/costs are imprecise and this nature is always existing.
Furthermore, our solution has the nature of minimizing the possibility of lower profit,
maximizing the most possible profit and maximizing the possibility of higher profit. As a matter
of fact, with little effort, we really obtain a lot of information about the solution. Thus we do
believe that our possibilistic linear programming technique is more efficient to solve the problems
with imprecise nature in the real world.

4.1.2.2 Rommelfanger, Hanuscheck and Wolf's Approach

The imprecise objective coefficients of the linear programming problem discussed here,
is assumed as follows:

max cx
s.t. x EX = {x I Ax s b and x ;;:: O} (4.37)

where c= [c L , C U] is an interval-valued coefficient vector. Analogously to the theory of Multiple

Objective Linear Programming (MOLP), the complete solution of the linear optimization problem
with interval-valued coefficients of Equation (4.37) is defined as:

E = {x E X I there is no x' E X such that cx' ;;:: ex for all c E [cL, cu] and
ex' > cx for at least one c E [cL , CU]}.

To solve Equation (4.37), let us first define:

which are two extreme cases. Rommelfanger, Hanuscheck and Wolf [216] reduced the infinite
 214

objective functions of Equation (4.37) to search a solution of the following MOLP problem:

max [z",;.,(x), z....x(x)].

XEX (4.38)

Rommelfanger et al. pointed out that this reduction of the objective function is also connected
to a reduction of the set of feasible compromise solutions since the complete solution of Equation
(4.38) is a mostly proper subset of E. By this procedure, the set of compromise solution,
however, is not restricted to the corner points of simplex X.
To solve Equation (4.38), we can then use any MODM technique. Here, we use
Zimmermann's [274] approach and first find:

Zmin· = zmin(x..w,') = max Zmin(X)

XEX

Zmox· = zmox(Xmox') = max zmox(x), and

xeX

(4.39)

Then, we can establish the membership functions of the objectives of Equation (4.38). These
are:

if ~(x) > ~.

if~' ~ ~(x) ~~.

if ~(x) < ~'

where k = min or max.

Finally, we compute:

max a
s.t. JLt(x) ~ a, k = min and max
XE X. (4.40)

Now, if the decision maker is capable of expressing his ideals about the imprecise
objective coefficients in the form of fuzzy sets:
 215

then the problem arises to take into account the additional information contained in possibility
distributions 1r(c). To handle this situation, let us first assume that the imprecise coefficients
have convex possibility distributions. Then, we may take a finite set of degrees of possibility
({I3I' 132, ... , I3r}) and consider r objectives with corresponding l3i-Ievel, i = 1,2, ... , r, specific
interval-valued parameters. Thus, our problem becomes:

max (CplX, c/l2x, ... , c/3rx)

XE:X

where CPi' Vi, are interval-valued as shown in Figure 4.12. By use of the concept of Equation
(4.38) for each objective function, we obtain the following auxiliary problem:

max a
s.t. l-'i,k(X) ~ a, k = min, max and i = 1, 2, ... r
x EX. (4.41)
where
1'f Zjot<x) >Zj,k•
if Zj,t' :::;Zj,t(x) :::;Zj,l:
if Zj,t<x)< Zi,t'
and
Zi,mio· = Zi,mio(Xj,min") = max Zi,mio(X)
XEX

Zi,mu.· = Zi,mu.(Xj,max) = max Zj,max(x)

XEX

Example. Consider the following linear programming with imprecise objective coefficients:

max ClX l + ~X2

S.t. xEX = {x I Xl + 4X2 :::; 100, Xl + 3x2 :::; 76

Xl + 2~ :::; 53, 3xl + 5x2 :::; 138
3x l + 4X2 :::; 120, 7Xl + 8X2 :::; 260
Xl + X2 :::; 36, 3xl + 2~ :::; 103
2Xl + X2 :::; 68 and Xl' X2 ~ O}
 216

where CI and ~ are imprecise and have their corresponding possibility distributions as shown in
Figure 4.13.

o
8 9 10

C',o = [0.5, 10]; C',O . 25 [1, 7]; C',O . 5 = [2, 5];

C"O.75 = [3.2, 4.5]

.75

.5

.25

o
1 2 3 4 5 6 7 8 9 10 11

C 2 ,O [1.4, 11]; C 2 ,O . 25 = [5, 10]; C 2 , O. 5 [6, 9];

C 2 ,O.75 = [6.5,7.5]

Figure 4.13 The possibility distributions of c, and c 2

To solve this problem, let {31 = 0.0, {32 = 0.25, {33 = 0.5 and {34 = 0.75. After
computing Equation (4.39) for each (3i' we then solve the following problem:

max ex
s.t. J'1,mia(X) = (O.5x I+ 1.4x2-25.2)/10.5 ~ ex
J'I,...,.(X) = (IOx I+ 11 x2-223)/1 45 ~ ex
 217

#'2,min(x) = (Xl +5xrI06)/19 ~ ex

#'2,max(X) = (7XI + IOxr259)/42 ~ ex

#'3,min(X) = (2XI +6xrI4S)/4 ~ ex

#'3,max(X) = (5XI +9X2-236)/S ~ ex
#'4,min(X) = (3.2xI +6.5x2-16S.2)/3.7 ~ ex
#'4,max(X) = (4.5xI +7.5xr204)/3 ~ ex
(Xl' X~ E X and ex E [0, I].

The solution of this problem is ex'" = 0.50 and X'" = (9, 22). This leads to the average objective
value:

4
z'" = (liS) E [Zi,min(X"') + Zi,max(X"')] = 192.45.
i=1

The corresponding realized degrees of satisfaction are:

.96 .75 .68 .79 .50 .87 .97 .50

4.1.2.3 Delgado, Verdegay and Vila's Approach

Delgado, Verdegay and Vila [52] considered the problem of Equation (4.37) and assume
that the imprecise objective coefficients ~, j = 1, 2, ... , n, have the following possibility
distributions (as shown in Figure 4.14):

o
hj(cj)
1I}Cj) = 1
glCj) (4.42)

Based on the following definition provided by Verdegay [250], a fuzzy objective to the
fuzzy mathematical programming problem is a fuzzy set of F(Rn), i.e., a function, cI>: F(Rn) - >
[0, I], where F(Rn) = {f I f:Rn-->R}. Let cI> E p[F(Rn)} be a fuzzy objective. We define the
following classical relation for any {3 E [0, I]:
 218

Figure 4.14 The possibility distributions of c j

Vx,y E RD, X > IJ Y < = = > f(x) ~ fey)

where Vf E F(RD) with cI>(f) ~ I - {3 and where > IJ must be read as "more preferred or
indifferent, with (3 grade, than." It is clear that cI>(f) ~ I - {3 is a preorder (reflexive and
transitive). Delgado et al. then defined:

c = (Cl> ~, ... , cJ and 1I'(c) = in~ 1I'j(c)

and obtained the following auxiliary problem of Equations (4.37) and (4.42):

s.t. T(C) ~ I - {3 and x E X. (4.43)

Taking into account that:

then, from Equation (4.42) it is obtained that:

Thus our problem becomes:

max {cx I x E X, C E r(l-{3), (3 E [0, In (4.44)

where r(l-{3) denotes the set of vectors c whose components cj are in the interval [h/(l-{3), g/(l-
 219

13)], Vj. Then with fixed 13 E [0, 1] the solution of Equation (4.44) is given by x· E X such that:

Vc E r(1-I3), Vx EX ==> cx· ~ x.

As the unity of that x· E X is not guaranteed, let SI_P denote the set of those points being
solutions of Equation (4.44). Thus, in accordance with the Decomposition Theorem for fuzzy
sets, it can be defined S = Up I3S1_P which is a fuzzy set giving the fuzzy solution to the former
problem.
Clearly, for any 13 E [0, 1], r(1-I3) is a convex set with extreme points defined by those
c E r(l-I3) such that its components are either in the lower hj-1(1-13) or upper bound gj-l(l-I3).
IfV(c) = {x I cx ~ 0, C E r(l-I3)}, it is clear that V(c) is a cone. Likewise, V Ir(l-I3) I = Uc
V(c), where c E r(1-I3), is a cone [the preferential cone of the solutions of Equation (4.44)].
Now, let E(I-I3) be the subset (of r(l-I3» constituted by vectors whose jth component is
equal to either the upper or the lower bound of Cj (hj-l(l_f3) or g{I(1-I3», Vj. That is:

(4.45)

Then it is evident that the maximum number of elements in E(I-I3) is 2B. On the other hand, it
is verified that V[r(l-I3)] = U c V(c) where c E E(l-I3). Therefore, we may use the upper or the
lower bound to solve Equation (4.44), rather than use the complete (infinite) family involved.
That is to solve:

max {cx I x E X, C E E(l-I3), 13 E [0, In

or
max (c1x, c?x, ... , CSx)
s.t. x E X, r} E E(I-I3), k = 1, 2, .•. , 2B
13 E [0, 1] (4.46)

where s =2 B, which is a conventional MOLP problem. Thus, we can solve Equation (4.45) by
using any MOLP technique.
Delgado et al. proposed two approaches. The first approach is the weighing objective
approach which is:
 220

max Et wtd'x
s.t. x E X, eft- E E(1-tn, k = I, 2, ... , 2D
fJ E [0, I], Wt O!: 0, Yk, and Et Wt = I. (4.47)

The secopd approach (restricting objectives) is to use the concept of Rommelfanger et al. [216]
as previously discussed. For an interval-valued objective Coefficient problem (fixed fJ) we have:

max a
s.t. IlmiD(x) = (z",;.,(x)-z",ia')/(Zmi:-z",;.,') ~ a
/lm.x(x) = (z",.,,(x)-z",.x')1(ZIIIO';-z",.,.') ~ a
x E X and a E [0, I] (4.48)

whose optimal solution can be obtained from:

max (z",."(x)-z",.,,')/(z,,,.,,·-z",.,,')
S.t. llmiD(X) = (z",ia(x)-Zmi;)/(z",iD·-Zmi;) ~ a
x E X and a E [0, I] (4.48')

where z",.,,(x) = (cfJU)x (the upper bound) and z",iD(X) = (c/)x (the lower bound), and other
notations are correspondingly defined. But the objective function of Equation (4.48') is
increasing on (CfJU)x and thus this problem is equivalent to:

max (cfJU)x
S.t. (c/)x ~ (Zmi:-Zmin')a + Zmi;
XE X. (4.49)

Similarly, to determine the compromise solution of Equation (4.46), the following problem is
solved:

max a
s.t. (eft-)x ~ a
eft- E E(I-fJ), k = I, 2, ... , 2D
X E X, a E Rand fJ E [0, I].

With fixed fJ (taking 1-(I-fJ», we then have the following equivalent problem:
 221

max a
s.t. (cfJ~x ~ a, k = 1, 2, ... , 2n
x E X, a E R and P E [0, 1]. (4.50)

However, Vx E X => (cl)x :s;; (cfJ~x :s;; (cl)x for k = 2, 3, ... , 2n-l. Thus the first
constraint of Equation (4.50) is equivalent to (cl)x ~ a. Our problem is then:

max (cfJU)x

s.t. (cl)x ~ a

x E X, a E R and P E [0, 1] (4.51)

where a is considered as a parameter.

Note that if x*(a) is the optimal solution to Equation (4.51), then the optimal solution to
Equation (4.50) willbe the particular solution ofx*(a), say x*(a), obtained for that a* such that:

a* = sup {al x*(a) is the solution of Equation (4.51)}.

Moreover, the optimal solution to Equation (4.50) for the particular value ao = (zmin*-Zmin')a
+ Zmin', will coincide with the corresponding optimal solution of Equation (4.49). Equations
(4.49) and (4.51) are considered equivalent in this way.
Finally, Delgado stated that if solution x* is the optimal one, then:

(4.52)

and because of the parametric nature of the optimal solution of Equation (4.51), say x(a), for
each fixed P E [0, 1], there exists ao such that:
(4.53)

From Equation (4.53) we can find the solution of Equation (4.48) without solving that problem.

Example. Consider the following problem:

max Z = CIXI + 2x2

s.t. x E X = {x I 2xI + 3x2 S; 13, Xl + 4x2 :s;; 16, 5xI + 4x2 S; 16,
Xl and x2 ~ O}
 222

where c, is imprecise cost and has the following possibility distribution:

if c, > 4 or c, < 112
if 112 :s c, :s 3
if 3 :s c, :s 4.

The interval representation is:

c"o = [112. 4], c"o.~ = [9/S, 30/S], c"o.' = [7/4, 14/4], C"O.1' = [19/S, 26/S].

In this situation, Equation (4.46) is written as:

max {[(6-5fj)/2]x,+ 2x2, (3+fj)xl +2x2}

S.t. x E X and {3 E [0, 1]. (4.54)

Suppose {3 = 0.75, we can then solve Equation (4.54) by weighing objectives as in Equation
(4.47). That is:

max w,[(19/S)x, +2xJ + w2[(26/S)XI+2~].

XEX

By taking w, = W 2 = 112, the optimal solution is (3, 0) with z· = S.43.

On the other hand, if Equation (4.54) is solved by restricting objectives with {3 = 0.75,
we will write Equation (4.51) as:

max 3.25x, + 2X2

S.t. 2.375x, + 2~ ~ a, x E X and a E R

whose solutions are: (60-8a, -71.25+ 10(3) with z· = 52.5 - 6a for a E [7.125, 7.5]; (3, 0) with
z· = 9.75 for a E [0, 7.125]. These solutions, for some particular values of a, reproduce the
corresponding solutions to Equation (4.48). In fact, ZmiD· = 7.5, Zmin' = 7.125, z....x. = 9.75,
z....x' = 7.5. When solving

it is found that O!o = 7.3125 and consequently, the solution is (1.5, 1.875) and a (of Equation
(4.48» is 0.5.
 223

4.1.3 Linear Programming with Imprecise Objective and Technological Coefficients

By recalling the investment problem of Equation (4.35) in the previous section, we may
consider a situation where the objective and technological coefficients are fuzzy and have
triangular possibility distributions with most possible and least possible values (Lai and Hwang
[138]). This is quite common, especially in the finance world. The interest rate is always not
fixed in current economic systems where the float interest rate policy is adapted. However, the
decision makers may know a possible interval, according to their experience. For example, the
interest rate will be around 8 % with no less than 6 % and no larger than 11 % in the next 2 years.
Therefore, it is important to study the case where the interest rate has a possibility distribution.
Consider the following problem:

max cx

s.t. x f X = {x I Ax ::;; b and x ~ O} (4.55)

where b is precise, but cand A are imprecise with triangUlar possibilistic distributions of (cID ,
cP, CO) and (AID, AP, AO), respectively. To solve this imprecise technological coefficient, we
may apply the concept of the most likely value which is (4AID jj + APjj + A°jj)/6 where {3 is a
minimal acceptable possibility (not (3-level set). Finally, we obtain [138]:

min ZI = (cm - cP)x

max ~ = cmx
max z3 = (CO - cID)x
S.t. {(4Amjj + APjj + AOjj)/6}x ::;; b
x ~ O. (4.56)

The weights between Amjj' APjj and A°jj can be changed subjectively. The reason for using most
likely values is that APjj is too pessimistic and A°jj' too optimistic. Of courses these two
boundary values provide boundary solutions.
Equation (4.56) is obviously a nonlinear programming problem. However, if {3 is initially
given by the decision maker, it will be a linear programming problem. Thus, we can provide
the decision maker a solution table with {3 = 0.0, 0.1, ... , 0.9, 1.0.
With the consideration of the floating interest rate, the coefficients of the objective and the
3rd through 6th constraints of Equation (4.35) should be considered as imprecise or fuzzy.
Interest rates of the beginning of the planning periods in the first and the second constraints are
 224

quite sure. However, interest rates will fluctuate in future years. Let us assume the following
problem:

max Z = (5.5,5.1,6.2)F+(-1.0,-1.2,-O.85)M+(6.0,5.0,6.5)D
+ (-1.065,-1.08,-1.058)B6+ (1.046, 1.040, 1.06)~

s.t. 3F+2M+2D-B1 +LI ~ 2

IF+0.5M+2D+1.06BI-1.04LI-~+~ ~ 0.5

1.8F-1.5M + 1.8D+(1.06, 1.055, 1.O65)~-(1.04, 1.035, 1.045)~-B3 + ~ ~ 0.4

-O.4F-1.5M-ID+(1.06,1.055,1.07)B3-(1.04,1.035,1.05)~-B4+L4 ~ 0.38

-1.8F-1.5M-ID+(1.065,1.06,1.07)B4-(1.044,1.038,1.05)L4-Bs +Ls S; 0.36

-1.8F-O.2M-ID+(I.065,1.058,1.075)Bs-(I.046,I.042,1.055)Ls-B6+~ S; 0.34

Bj E [0, 2], i = 1, 2, 3, 4, 5, 6,
F, M and DE [0, 1]. (4.57)

where (1.06, 1.055, 1.065) in the third constraint means: the most possible borrowing interest
rate = 6%; the most pessimistic rate = 6.5%; the most optimistic rate = 5.5%. Similarly, for
(1.04, 1.035, 1.045) the most possible lending interest rate = 4%, the most pessimistic rate =
3.5 % and the most optimistic rate = 4.5 %. Also, the possibility distributions of these interest
rates are changed as time goes on.
The auxiliary problem of Equation (4.57) at (3 = 0.5 is then:
min 0.4F-0.2M + D-O.015B6+O.OO6L6
max 5.5F-M + 6D-1.065B6+ 1.046L6
max O. 7F-O.15M +0.5D-O.OO7B6+O.014L6
s.t. 3F+2M+2D-B 1 +L1 S; 2
IF+0.5M+2D+1.06Br1.04LI-B2+~ ~ 0.5
1.8F-1.5M+1.8D+1.06~-1.04~-B3+~ ~ 0.4
-O.4F-l.5M-ID + 1.0604B3-1.0408~-B4 + L4 ~ 0.38
-1.8F-1.5M-ID+ 1.065B4-1.044L4-Bs + Ls S; 0.36
-1. 8F-O.2M-ID + 1.0653Bs-1.0464Ls -B6+L6 ~ 0.34
Bj E [0, 2], i = 1, 2, 3, 4, 5, 6,
F, M and D E [0, 1].
 225

The PIS is (0.23, 9.09, 0.63) and NIS (0.85, 3.13, 0.16). The final solution at P = 0.5 is
shown in Table 4.3 and the possibilistic distribution of the imprecise profit obtained is given in
Figure 4.12.

4.1.4 Linear Programming with Imprecise Coefficients

The income stream from previous investments may not be precise as time progresses. For
example, the interest incomes of money market accounts are changed with the market. Thus,
we may have the following possibilistic linear programming:

max cx -
s.t. Ax:s b, and x C?! 0 (4.58)

where A, band c are imprecise and possibility distributions. To solve Equation (4.58), we
present four approaches ofLai and Hwang [138], Buckley [24], Negi [175] and Fuller [97].

4.1.4.1 Lai and Hwang's Approach

In this section, we combine the fuzzy ranking concepts with our strategy for the imprecise
objective function, when the imprecise coefficients have triangular possibility distributions. We
obtain the following auxiliary model [138]:

min zl= (cm - cp)x

max ~ = c"'x

max ~ = (CO - cm)x

s.t. Ampx:S b mp, APpx :S bPp, AOpx :S bOp

xC?!O (4.59)

where the auxiliary inequality constraints have been proposed by Ramik and Rimanek as shown
in Section 4.1.1. If the minimal acceptable possibility, p, is given, Equation (4.59) will be a
MOLP problem.
Now, let us return to our investment problem with the modification of "income stream
from previous investment" (in millions): $0.5, ($0.4, $0.35, $0.5), ($0.38, $0.35, $0.4), ($0.36,
$0.34, $0.45), ($0.34, $0.3, $0.42) ($0.3, $0.28, $0.35). Other assumptions are the same as
previous ones. The problem becomes:
 226

max z = (S.S,S.I,6.2)F+(-1.0,-1.2,-0.8S)M+(6.0,S.0,6.S)D
+(-I.06S,-1.08,-1.0S8)B6+(1.046,I.04,I·06)4

1.8F-1.SM + 1.8D+(1.06, 1.05S,1.06S)Bz-(1.04,1.03S,1.04S)~-BJ + l.:J S(.4,.3S,.5)

-O.4F-l.SM-ID+(I.06, 1.0SS, 1.07)Br(1.04, 1.03S, 1.05)l.:J-B4+ L4 S (.38,.3S, .4)

-1.8F-l.5M-ID+(I.06S, 1.06, 1.07)Bdl.044, 1.038, 1.0S)L4-Bs+ Ls S (.36, .34, .4S)

-1.8F-0.2M-ID+(1.06S,I.058,1.07S)B,-(1.046,I.042,1.05S)Ls-B6+4S(.34,.3,.42)

Bj E [0, 2], i = 1, ... , 6

F, M and D E [0, 1]. (4.60)

Assume (3 = O.S. We obtain:

min 0.4F-0.2M + D-0.01SB6+0.006L6

max S .SF-M + 6D-1.06SB6+ 1.046L6
max O. 7F-0.1SM +0.5D-0.OO7B6 +0.014L I6
s.t. 3F+2M+2D-B1 +L1 S 2
1F+0.SM+2D+1.06Bc1.04LcBz+~ S 0.5
1.8F-1.SM+1.8D+1.06Bz-1.04~-~+l.:J S 0.4
1.8F-1.SM+1.8D+1.0S7SBz-1.037S~-~+l.:J S 0.375
1.8F-1.5M+1.8D+1.0625Bz-1.0425~-BJ+l.:J S O.4S
-0.4F-1.5M-D+ 1.06BJ-1.04l.:J-B4+ L4 S 0.38
-0.4F-1.5M-D+ 1.0S7SBJ-1.037Sl.:J-B4+ L4 S 0.365
-O.4F-1.5M-D+1.065BJ-1.045l.:J-B4+L4 S 0.39
-1.8F-1.SM-ID+ 1.065B4-1.044L4-Bs +Ls S 0.36
-1.8F-1.5M-ID+1.0625B4-1.041L4-Bs +Ls S 0.35
-1.8F-1.5M-ID+1.0675B4-1.047L4-Bs +Ls S 0.405
-1.8F-0.2M-ID+ 1.065Bs-1.046L,-B6+L6 S 0.34
-1.8F-0.2M-1D+ 1.061SBs-1.044Ls-B6+ 4 S 0.32
-1.8F-0.2M-1D+ 1.07Bs-1.0505L,-B6+L6 S 0.38
Bj E [0, 2], i = 1, 2, 3, 4, 5, 6
 227

F, M and D E [0, 1]. (4.61)

The PIS is (0.23, 7.19, 0.64) and NIS (0.84, 3.13, 0.16). The final solution of Equation (4.61)
is shown in Table 4.3 and the possibilistic distribution of the imprecise profit obtained is shown
in Figure 4.12.

4.1.4.2 Buckley's Approach

Buckley [24, 25] considered the problem of Equation (4.58) as:

max Z=cx
S.t. Ax S; b and x ~ 0 (4.62)

where c, b and A are all imprecise coefficient (fuzzy number) vectors and matrix. Assume that
the imprecise nature of the coefficients can be modelled by trapezoidal possibility distributions,
11"(.), (as shown in Figure 4.15). That denotes: C = (cl , ••• , cJ and Cj = (cj,1I cj,2' Cj,3I Cj,4); b =
(hI, ... , b,J and b; = (b;,db;,2' b;,3Ib;,4); and A = [8;j] and 8;j = (a;j,da;j,2' a;j,3I~j,.v.

Figure 4.15 The trapezoidal possibility distribution of c j

Buckley stated that the fuzzy numbers are the possibility distributions associated with the
fuzzy variables and hence place a restriction on the possible values the variables may assume.
He defined 1I"(Z = z) as being the possibility distribution of the objective function Z. Assume
that all the fuzzy variables 8;j' b; and C; are non-interactive. Then, the possibility that x satisfies
the ith constraint is specified as:
 228

r(x E FJ = SUp {rjoiat(ll;,bJ I Il;x :s; bi } (4.63)

Il;,b i
where

and Il; = (Il;), ... , a;..) and rjoiat(ll;,bi) is the joint distribution of Il;j' Vj, and bi . Therefore, for x
~ 0, we have the following possibility that x is feasible or satisfies all the constraints:

r(x E F) = min [r(x E F i )]. (4.64)

1 :s;i:S;m

Next, we construct the conditional possibility that Z = z under the given x as follows:

r(Z=z I x) = sUPc {rjOiDt(C) I cx = z) (4.65)

where c = (c)' ... , cn> and

Thus, the possibility distribution of Z can be defined as:

r)(Z=z) = r)(Z) = sup {min [r(Z=zl x), r(x E F)]}.

x~O (4.66)

Bucldey also provided another alternative definition of the possibility of Z which is to construct
the possibility distribution of the maximum values of Z. For given crisp values of c, b and A,
solve:

max Z = cx
s.t. Ax:S; b and x ~ 0 (4.67)

and let the maximum value of Z be z, and let rjOint(c, b, A) be the joint possibility distribution
of all the Cj' bj and Il;j. Then:

r2(Z) = sup {rjoiat(c, b, A) I Z = z} (4.68)

c,b,A

For Equation (4.62), the decision maker may like to obtain a compromise solution x·
 229

which has the following features: (I) the decision maker determines a value of a = a· E [0, I]
and a maximum value of Z = z(aj, both acceptable to him, so that r[Zc::z(aj] = a·; and (2)
the decision maker uses this a· to construct a crisp linear programming problem as Equation
(4.67), whose optimal solution for Z is z(aj, and an optimal value ofx is a compromise solution
x· to the possibilistic linear programming of Equation (4.62). This value of x solves Equation
(4.67) with max Z = z(aj and it will tum out that r[Zc::z(aj] = rl[Zc::z(aj] = r2[~c::z(aj]

= a·.
Buckley then employed a-cuts to develop his solution algorithm. In fact, the a-cuts of
normalized and convex fuzzy numbers are bounded, closed intervals. Thus, let us assume that:

" .. aU] b. a = [b.aL b.aU] and c. a = [c.aL c.aU] Vi J.

"-'J..a = ["-'J..aL'-U 'I 1'1' 1 J')' ,-

Then, solve:
max Z = Ej ctuJt.j
s.t. Ej IljtLJt.j S btU, Vi
Jt.j c:: 0, Vj, and a E [0, 1]. (4.69)

Let the solution z(a) and O(a) denote the feasible set to the problem of Equation (4.69). And,
define the possibility distribution of Z as:

if ZL S z S z(l)
if z = z(a) and 0 < a < 1
if z c:: z(O) (4.70)

where ZL = min {Ej cj ,2Jt.j I x E O(I)} and it is the lower bound of z(I). Obviously, z(o) is the
minimum value of the upper bound of the optimum value of Equation (4.69). Buckley proved
that if 0(1) :1= 0, then rl(Z=z) = 'Ir(z). Thus, the parametric linear programming of Equation
(4.69) gives the solution to rl(Z) because: 1) 0(1) is the largest set of x c:: 0 so that r(xEF) =
1; 2),0(a) steadily increases as a decreases from one to zero; and 3) 0(0) is the smallest set of
x c:: 0 so that r(xEF) = 0 whenever x is not in 0(0). Furthermore, Buckley proved that if 0(1)
:1= 0, 'lr1(Z) = r2(Z) for z c:: z(I).
Similarly, for a minimization problem:

min ex
s.t. Ax c:: b and x c:: 0
 230

we can obtain its compromise solution by solving the following auxiliary problem:

min Z = l:j Cj~

s.t. l:j BijGux,; ~ btL, Vi
x,; ~ 0, Vj, and a E [0, I]. (4.71)

Let the solution be z(a) and O(a) denote the feasible set to the problem of Equation (4.71). And,
define the possibility distribution of Z as:

if z(l) :s; z :s; ZU

if z = z(a) and 0 < a < 1

if z ~ z(O) (4.72)

where ZU = min {1:j Cj,3x,; 1 x E 0(1)} and it is the upper bound of z(I). Obviously, z(O) is the
maximum value of the lower bound of the optimum value of Equation (4.71).

4.1.4.28 Example: A Feed Mix (Diet) Problem

Consider the classical feed mix (diet) problem with two decision variables and the
information given in Table 4.4. The LLB Poultry Farm has 2000 chickens to feed and wishes
to plan their daily feeding schedule for the next few weeks. The feed consumption is around one
pound of feed per chicken per day. However, in order to fatten up the chickens for market, the
feed should meet certain nutritional needs. Also, these needs can be met by mixing different
feeds. This problem is then formulated as:

min Z = 0.15xl + 0.40~

s.t. O.Olxl + 0.025~ ~ 0.012(xl + xJ
O.09XI + 0.50~ :s; 0.22(xl + xJ

XI + ~ ~ 1000, and XI, ~ ~ 0

whose solution is X = (5200/3, 800/3) and Z· = $366.7.

Now, suppose that Cj' hi> 8;j' Vij, are fuzzy variables. ill = (.0051.01, .011·02), al2 =
(.011.02, .031.04), Izi = (.061.08, .11.12), in = (.41.46, .551.6), bl = (.0091·012,
.0121.015), ~ = (.181.2, .241.26), (;1 = (.11.13, .171.2), ~ (.31.4, .41.5) and N = (10001
2000, 2000 13000). The possibilistic linear programming is:
 231

Table 4.4 Data for feed mix problem

Nutritional Feeds (per lb) constraint

n eeds Corn Soybean

Calcium (lbs) a 11 =·Ol a I2 =·025 at least 1.2%

calcium per day =
p rote in (lbs) a 21 =·09 a 22= .50 at most 22% protei
per day = b 2

C ost c 1 =$·15 c 2 =$·40

Decision XI lbs x 2 lbs at least 1000 lbs

per day = N

min Z = CIX I + ~X2

s.t. anx l + al2 x2 ~ bl(X I + x~
8zIXI + ~X2 :::; b2(X I + X~
Xl + X2 ~ N and Xl and X2 ~ o.

To solve this problem, we first consider the following auxiliary:

min Z = C laLXI + ~aLX2

s.t. anauxl + a lt Ux2 ~ btL(X I +~
~laLXI + anaLX2 :::; b2aU (x l + X~
Xl +~ ~ NaL and Xl and X2 ~ o. (4.73)

Let the solution be x(a) with Z = z(a) and feasible set O(a), for a E [0, 1]. 0(1) is shown in
Figure 4.16. The first, second and third constraints of Equation (4.73) are labeled as A, Band
C, respectively. As a decreases from one to zero, we can see that: 1) the line A, always passing
through the origin, revolves clockwise to the xcaxis (a = 11113) and then below the xcaxis; 2)
the line B, also always passing through the origin, revolves counter-clockwise towards the X2-
axis; 3) the line C moves southwestward towards the origin; and 4) the solution x(a) first hits
xcaxis when a = 11113, and then moves to the left along the xcaxis as a approaches zero. It
is not difficult to see that as a decreases from one to zero, the point x(a) will give the correct
z value for 1r1(Z). Therefore,

if z ~ z(1) =314
if z = z(a)
if z :::; z(O) = 100.
 232

2000

2000

Figure 4.16 The feasible set nCl)

For a compromise solution, suppose that the decision maker selects the satisfying solution
at c/ = O.S. Then, the solution of Equation (4.73) is (1S00, 0) with z(O.S) = $223.2 and
1r(Z :S:223.2) = O.S.

4.1.4.3 Negi's Approach

Negi [175] used the concept of exceedance and strict exceedance possibilities, proposed
by Dubois and Prade [BMS] , to solve Equation (4.62) in the previous section.
The exceedance possibility [denoted by 1r(BL ~ RL)] is the possibility that Ii ~ R. By
referring to Figure 4.17 where E = (XI' ( 1), we can easily calculate this possibility as follows:

if b3 > rz
if b3 :S: r z and r I :S: b4
ifrl > b4 •

Similarly, we compute the strict exceedance possibility (1r(Bu~RU», point S, as follows:

 233

1
B
_--~t-----
R
. . --
I
i
/
i

o b, r

Figure 4.17 comparison of trapezoidal fuzzy number

if b3 > r4
ifb3 ~ r4 and r3 ~ b4
if r3 > b4 •

When fuzzy addition is used, LHS and RHS of the ith constraint in Equation (4.62) can be
considered as Rand B. If we use the exceedance possibility, we will have:

if bi ,3 > r i ,2
if bi,3 ~ r i,2 and ri,l ~ bi,4
if ri,l > bi,4 (4.74)

where 0i = (bi,4-ri,l)/[(bi,4-bi,3)+(ri,2-ri,l))' ri,l = Ej Il;j,lJe.j, ri,2 = Ej 1l;j,2Je.j and 1I"(xEF;) is defined in

the previous section. Then, we obtain:

(4.75)

with the decision space

 234

By referring to Equation (4.62), we have the imprecise objective function which is rewritten as:

Z= [~.J c·J, ,x.

-"] 1 ~.J c·J,2X . ~.J
-"]' C·J, 3X
-"]. I ~.J C·1,-"]
4X .] (4.76)

and its possibility distribution is:

8, = [z - ~j Cj ",,:;]/[ ~j Cj ,2":; - ~j cj,1":;] if ~j Cj,,":; s; z s; ~j Cj,2":;

1r(Z=zlx) = 1 if ~j cj ,2":; < z < ~j Cj,3":;

82 = [~j Cj ,4":; - z]/[ ~j Cj ,4":; - ~j cj,3":;] if ~j cj,3":; :s: z :s: ~j Cj ,4":;

o otherwise.

The largest objective value is at 1r(Z = z Ix) = 82 in the objective function space. Thus, we
obtain:

(4.77)

which satisfies:

Clearly, the equations of OJ and 82 are non-linear functions. However, if the decision maker has
a lower cut-off value (a), Equation (4.77) will be equivalent to:

max z
s.t. 151 , Oz, ... , Om' 82 ~ ex
b i ,3 s; ~j li;j,2":;' b i ,4 ~ ~j li;j,l ":;, Vi.
~j cj,3":; s; z s; ~j Cj ,4":;

z, ":;, Vj, ~ 0 (4.78)

which is obviously a crisp linear programming problem.

Example. Consider the following possibilistic linear programming problem:

max (112, 314)x, + (213, 416)Xz

S.t. (011, 213)x, + (112, 315)Xz s; (315,618)
(112, 316)x, + (011, 213)x2 s; (214,617)
x, and X2 ~ O.
 235

From Equation (4.74), we obtain:

ifxl+2x2 S 6
ifxl+2x2 ~ 6, x2 S 8
ifx2 ~ 8

if2xl+X2 S 6
if2xl+x2 ~ 6, xl ~ 7
if xl ~ 7

81 = (Z-Xr 2XV/(Xl +XV if xl +2X2SZS2xl +3x2

r(Z=z Ix) = 1 if 2xl +3x2Sz~3xl +4x2
82 = (4xl +6xrz)/(xl +2xV if 3xl +4x2Sz~4xl +6x2
o otherwise.

According to Equation (4.78), we obtain:

max Z
s.t. 61 ~ a [or 8-x2 ~ a(2+xl + xV]
~ ~ a [or 7-xl ~ a(1 +xl + xV]

82 ~ a [or 4xl +6xrz ~ a(xl +2x)]

xl+2x2 ~ 6, x2 S 8, 2xl+x2 ~ 6, Xl ~ 7
3xl +4x2 S Z S; 4xl +6x2'

The solution for a = 0.5 is Xl = 3.125, x2 = 3.625 and Z = 29.0625.

Similarly, if we adapt the strict exceedance possibility (62), then we will have:

max Z
S.t. 61, ~, ... , 6m, 82 ~ a
b·1,3 S E·] II'J,
.. 4X: b· 4 ~ E· II .. 3x, Vi.
--J' 1, ] 'J, J'

E·] c·J, 3x,J S z S E·J c·J, 4x,J

z, Xj' Vj, ~ 0 (4.79)

where 6i • Vi, are the strict exceedance possibilities for the constraints.
For a minimization problem, we can also use the (strict) exceedance possibility to solve
 236

it. However, we just have the Band R as the LHS and RHS, respectively. (This principle can
also be applied to a situation which includes both kinds of constraints and consider 81 instead of
82,) The remaining derivations are the same as previous descriptions.

4.1.4.4 Fuller's Approach

To solve Equation (4.62), .Fuller [97] proposed the following auxiliary crisp linear
programming:

max l:j djxj

S.t. l:j fijXj :s:; gi' Vi, and Xj ~ 0, Vj (4.80)

I:
where

dj = r[Cj(r) + Cj(r)]dr, Vj

fij = I: r[Bjj(r) + Aij(r)]dr, Vij

gi = I: r[bj(r) + Bi(r)]dr, Vi.

The fuzzy numbers of Cj' b i and Bjj' Vi,j, can be parametrized by {(cj(r), Cj(r), r)1 0 :s:; r ::::; I},
{(bier), Bi(r), r)1 0 ::::; r :s:; I} and ((Bjj(r), Aij(r), r)1 0 :s:; r :s:; I}, respectively.
Furthermore, Fuller proposed that if there exist real numbers Uj' Vij and Wi' Vi,j, such
that:

Uj - cj(r) = Cj(r) - Uj
Vij - Bjj(r) = ~j(r) - Vij
Wi - bier) = Bi(r) - Wi

for all r E [0, 1]. Then we will have the following auxiliary problem:

max l:j UjXj

s.t. l:j vijXj :s:; Wi' Xj ~ 0, V i and j (4.81)

for the Equation (4.62).

In our particular case of trapezoidal fuzzy numbers, we have:
 237

c·(r)
) = (C,). 1 - c·).21
-\r + c·). 2
C-(r)
J = (C,J3.-)c·. ~r + c·).4
llj.(r)
~
= (llj'~.~.1 - llj' :z)r + llj'~. 2
A··(r)
IJ
= (ll'J.•• 3 - llj'~.4I·\r + 1I'J.•• 4
b·(r)
1 = (b.1, 1 - b·1,21-\r + b·1,2
B·(r)
1 = (b.1, 3 - b·1,41·\r + b·1,4 (4.82)

for Vi,j. Finally, Equation (4.80) is obtained by substituting the following crisp numbers:

d·) = 1I3[c·)1.+)c·. 3 + (1I2)(c·). 2 + c·).41.\]

f··IJ = 1I3[llj' 1 + llj'J. 3 + (1I2)(a·). 2 + 'J. 4)]
~.
1I ••

g.1 = 113 [b.1, 1 + b·1, 3 + (1I2)(b·1, 2 + b·1,4'.\] (4.83)

for Vi,j. Thus, we can easily solve Equation (4.62) with trapezoidal fuzzy numbers by use of
Equations (4.80) and (4.83).

4.1.5 Other Problems

In Table 4.1, we have classified the possibilistic linear programming into seven cases.
These cases are further grouped into four main problems. In the previous sub-sections we have
discussed some approaches to solve these problems. There remains Cases 3, 6 and 7 to be
discussed.
Obviously. Case 3 is a dual problem of Case 2. Thus we can solve Case 3 by transferring
it into Case 2 and by using those approaches discussed in Section 4.1.2. On the other hand, Case
6 is very similar to Case 5, except the imprecise input data are in the right-hand side instead of
the left-hand side. Thus, we can solve it by using the approach discussed in Section 4.1.3.
Finally, we can solve Case 7, where only A is imprecise, by using the approach discussed in
Section 4.1.3, too. That is to obtain a crisp auxiliary problem by transferring the imprecise
technological coefficients A into crisp representive numbers by using the concept of the most
possible values.
In the next section, we are going to discuss some extensions of possibilistic linear
programming problems.
 238

4.2 Some Extensions of Possibilistic Linear Programming

In the previous section, we have discussed all the typical possibilistic linear programming
problems with imprecise coefficients. However, there are some other problems we call hybrid
problems such as these situations:

(1) The technological coefficients, available resources and the inequalities of the constraints are
imprecise. The imprecise technological coefficients and available resources are modelled by
possibilistic distributions and the imprecise/fuzzy (soft) inequalities are modelled by preference-
based membership functions.
(2) The coefficients of the objective function are imprecise and modelled by possibilistic
distributions. And, the fuzzy constraints involve fuzzy (soft) inequalities or fuzzy resources
available, which can be modelled by preference-based membership functions.
(3) The imprecise nature of the input data (c, b and/or A) should be modelled by both the
possibility theory and the probability theory.

First, let us discuss the linear programing problem with imprecise coefficients and
inequalities.

4.2.1 Linear Programming with Imprecise Coefficients and Fuzzy Inequalities

The first hybrid problem can be mathematically formulated as:

max cx
s.t. Ej 8;j~ ~ hi' and ~ ~ 0, Vi,j (4.84)

where a;j and hi' Vi,j, are triangular fuzzy numbers (fuzzy numbers with triangular possibility
distributions), and ~ is the soft inequality.
First, the soft inequality can be considered as the inequality with allowable tolerance to
accomplish the constraints. That is: it permits the constraints to be satisfied as well as possible.
Thus, we can model this imprecise situation by the preference-based membership function.
Consider the following constraint:

(4.85)

where the tolerance of inequality is given Pi. This inequality constraint indicates that the
consumption of resources ( Ej Il;j~) is at best less than or equal to bi , but no more than bi + Pi.
 239

Between b i and bi + Pi' there exists a non-increasing function representing the satisfaction of our
willingness concerning the consumption of' the resources. Thus the soft inequality can be
modelled by the membership function (as discussed in Chapter 3), fJ.i(X). Assume it is linear such
as:

,,(x) ~ { i- ~,~
(l:, - bJ/]\
if Ej a;j'S S bi
if bi S Ej ll;j'S S bi+Pi
if E·J .)
lI··X· ~ b· +p.
J 1 1·
(4.86)

Let a = mill; fJ.;(x) be the minimal acceptable level of satisfaction. That is: fJ.i(X) ;;:: a, Vi. Then
the constraint of Equation (4.85) will be equivalent to:

(4.87)

Next, let us consider the following constraint:

(4.88)

where 3:;j = (a;t, a;/, a;t) and bi = (him, bt, biO), Vj and have triangular possibility distributions.
This notation has been used in Section 4.1.2 (see Figure 4.6 also). Equation (4.88) is actually
similar to the problem discussed in Section 4.1.1. Thus, all the approaches discussed in Section
4.1.1 can by applied here.
In order to solve the constraints of Equation (4.84), Delgado, Verdegay and Vila [55]
proposed a relatione between fuzzy numbers, which preserves the ranking when fuzzy numbers
are multiplied by positive scalars. e can be any reasonable inequality operator. However,
Delgado et al. proposed the following two operations:

-
X elY - < => xm S ym, (4.89)

where X - = (Xm, XP, XO) and Y- = (ym, YP, yO). Thus, Equation (4.84) involves fuzzy ranking
problems and may be equivalent to the following auxiliary problem:

max cx
s.t. Ej 8;j'S e bi + p;(l-a) (4.90)

where 'i;j' bi and Pi are fuzzy numbers (have possibility distributions).

 240

Furthermore, for Equation (4.90) Campos [29], and Campos and Verdegay [30] provided
the following equivalent crisp linear programming models:

(1) By use of Yager's first index (see [30]), we can obtain the equivalent model as:

max cx
s.t. (a;+a/+at)x ~ (bj+bjl+bt) + (t;+t;I+t;U)(1-a)
x ~ 0 and ex e [0, I], Vi.

(2) By use of Yager's second index (see [30]), we can obtain the equivalent model as:

max cx
s.t. (2a;+a;I+a;u)x ~ (2bj+b/+bjU) + (2t;+t;I+t;u)(1-ex)
x ~ 0 and ex e [0, I], Vi.

(3) By use of Campo's k-preference index [29], we can obtain the equivalent model as:

max cx
s.t. (kll;+(I-k)ll;U)x ~ (kbj+(l-k)bt) + (kt;+(I-k)t;U)(1-ex)
x ~ 0 and ex e [0, I], Vi

where k e (0, I] is fixed by the decision maker a priori.

(4) By use of Dubois and Prade's k-weak feasibility concept [66], we can obtain the equivalent
model as:

max cx
s.t. <ka;+(I-k)ll;I)X ~ (kbj+(1-k)bjU) + (kt;+(I-k)t;U)(1-a)
x ~ 0 and ex e [0, 1], Vi

where k e (0, 1] is fixed by the decision maker a priori.

(5) By use of Dubois and Prade's k-hard feasibility concept [66], we can obtain the equivalent
model as:
 241

max cx
s.t. <kai+(l-k)abx:s; (kbj+(I-k)bb + Octi+(I-k)tb(l-a)
x ~ 0 and a E [0, I], Vi

where k E (0, I] is fixed by the decision maker a priori.

(6) By use of Tanaka, Ichihashi and Asai's [240] inequality relation, we can obtain the
equivalent model as:

max cx
s.t. <kai+(l-k)abx:s; (kbj+(I-k)bj1) + Octi+(I-k)tb(l-a)
<kai+(I-k)IljU)x :s; (kbj+(I-k)bjU) + Octi+(I-k)1jU)(I-a)
x ~ 0 and a E [0, I], Vi

where k E (0, I] is fixed by the decision maker a priori.

The above six possible auxiliary models of Equation (4.84) are only crisp linear
programming models and can be solved by any LP package. Of course, we can also use some
other ranking methods discussed in Chapter 3 to obtain crisp auxiliary models. However, some
of them will cause computational difficulties, such as nonlinearity, in solving auxiliary crisp
models.

4.2.1a Example: A Fuzzy Matrix Game Problem

Consider a fuzzy matrix game [29] defined by the following matrix [Iljj]:

Yl Y2 ••• Yn

Xl
X2 [ 8a2lu 8" ... 8,.
a22 ••• a2n
1
Xu. amI 8m2 ••• a mn (4.91)

As is well known, the optimal strategies x· and y. of a classical game will be the solutions of
the following linear programming problems:

max v
s.t. Ej IljjXj ~ = I, 2, ... n
v, j
Xj ~ 0 and Ej Xj = I, i = I, ... , m
and
 242

min w
s.t. Ej ~jYj S w, i = 1, 2, ... , m
Yj ~ ° and E j Yj = 1, j = 1, ••. , n.

If we let Uj = x/v and Sj = y/w, Vi,j, we will have:

min Ej ~

s.t. Ej ~jUj ~ 1, Vj
Xj ~ 0, Vi (4.92)
and
max Ej Sj

s.t. Ej ~jSj S 1, Vi
Yj ~ 0, Vj. (4.93)

Now, if ~j , Vi,j, are fuzzy numbers and the inequalities are soft, then Equations (4.92)
and (4.93) might be considered as:

min Ej Uj

s.t. Ej ~jUj ~ 1 - lip-a), Vj

Xj ~ 0, Vi (4.94)
and
max Ej Sj

s.t. Ej ~jSi S 1 + 'i(I-a), Vi

Yj ~ 0, Vj (4.95)

where Pi and ~, Vi,j, are imprecise tolerances for the soft inequalities. Equations (4.94) and
(4.95) can then be solved by any of the previous approaches.
Consider the following fuzzy game:

-
where 180 -
= (180, 175, 190), 156 = (156, 150, 158),90 - = (90, 80, 100). Assume that players
I and II have their imprecise margins (tolerances): PI = iiz = (0.10,0.08, 0.11) and ql = ~ =
(0.15,0.14,0.17), respectively.
 243

By use of Yager's second Index (2), we obtain the following crisp parametric linear
programming problems:

min ul + u2
s.t. 725ul + 36Ou2 C!: 4 - 0.39(I-a)
620u 1 + 725u2 C!: 4 - 0.39(I-a)
ul' u2 C!: 0 and a E [0,1 ]
and
max sl + s2
S.t. 725s 1 + 620~ s 4 + 0.61(1-a)
360s 1 + 725s2 S 4 + 0.6I(l-a)
sl' s2 C!: 0 and a E [0, I].

The solution is: x· = (0.77,0.23) with v(a) = 643.45/[4-0.39(I-a)], and y. = (0.23, 0.77) with
w(a) = 643.45/[4+0.61(l-a)], where a = (0. I].

4.2.2 Linear Programming with Imprecise Objective Coefficients and Fuzzy Resources
In practice, the unit costs/profits of new products or new projects, the lending and
borrowing interest rates, and cash flows are always imprecise. This imprecise nature has long
been studied with the help of the probability theory. However, the probability theory might not
give us the correct meaning to solve some practical decision-making problems. In addition,
applying the probability theory to some optimization problems has a negative effect on the
computational efficiency. In these cases, the possibility theory might be helpful.
In this subsection, let us first discuss the following particular possibilistic linear
programming problem:

max Ej Xj/(1 +r~

s.t. x E X = {x I Ax S b and x C!: O} (4.96)

where r = (rn, rP, t» which indicates the triangular possibilistic distribution, 1f(.) as in Figure
4.7. rn, rP and t> are the most possible value, the most pessimistic value and the most optimistic
value, respectively. If 11"(') is normalized, 1I"(rn) = 1 and 1I"(rP) = 1I"(t» = O.
In order to solve Equation (4.96), we first represent the triangular possibilities of r by the
confidence interval of its Jj-cut (Kaufmann [131, 132]). That is: r can be represented by [rL(p),
 244

rUom or [r"+(I""'-r")fj, r"-(r"-I""')fj] where fj E [0, I]. Thus, the problem becomes:

max Ej Jt.i/(I + [r(p),ru(p)]y

XEX
or
max Ej Jt.i[1/(l +ru({J)y, 1/(1 +r(p)y]. (4.97)
XEX

Since the summation of interval values will be an interval value, the objective function of
Equation (4.97) is obviously a crisp interval value when fj is given. Thus we have the following
equivalent problem:

max [ZL(X, p), ZU(x, P)] (4.98)

XEX

where ZL(X, p) = Ej Jt.i/(I +ru(p)y, zU(x, (3) = Ej Jt.i/(l +r(p)y. To solve Equation (4.98), the
simplest way is to maximize wzL(x,P) + (l-w)zu(x,(3) where w E [0, I] indicates the possible
weight. On the other hand, the objective of maximizing an interval value can also be solved by
simultaneously maximizing the lower and upper bounds of the interval value. Thus we [140]
propose the following bi-objective auxiliary problem:

max ZL(X,fj) = Ej Jt.i/(I +rU«(3)y

max zU(x,(3) = Ej Jt.i/(I +r(p)y
s.t. x E X (4.99)

where r«(3) = r"+(I""'-r")(3, r U«(3) = rO-(r"-I""')(3 and (3 E [0, I]. If (3 is first given, then r(p) and
rU(p) will be constants and the problem of Equation (4.99) will be a crisp bi-objective linear
programming problem which can be solved by any MODM approach. However, in this study
we will propose an augmented max-min approach which is an extended version of Zimmermann's
approach for solving a crisp multiple objective programming problem.
The problem of Equation (4.99) may be solved by the following two steps. The first step
is to solve the following max-min problem (Zimmermann [298]):
max r
s.t. I-'1(ZL(X, fj» ~ r
I'2(ZU(x, (3» ~ r

xEX (4.100)
 245

whose solution is rO and xo. Note that in the above equation r is min [ILl(zL(x,p», l'2(zU(x,
P))].
There may be multiple decisions (x) which will end up with the same rO. In this case,
the solution of Equation (4.100) might not be unique nor efficient, because Equation (4.100)
randomly selected a solution which might be dominated by another solution with the same
satisfaction level rOo Moreover, we do not know whether or not multiple solutions exist. To
resolve this problem, we proposed the second step which is:

max wIILl(zL(x, P» + W2J'2(Zu(X, P»

s.t. ILl(zL(x, P» ~ r<>
l'2(zu(x, P» ~ r<>
xeX (4.101)

where wI and w2 are the weights (relative importance) among the corresponding objective
functions, and wI + w2 = 1. This second step is essentially trying to use the weighted
averaging operator to allow for possible compensation among objectives, which is similar to that
proposed by Li [151]. Equation (4.101) also guarantees that the overall satisfactory degree of
the compromise of the objectives is at least at rO.
It is noted that for a bi-objective programming problem, the solution of Equation (4.101)
may be equal to the solution of Equation (4.100). For a problem with a large number objective
functions, Equation (4.101) should be formed and solved for obtaining efficient solutions.
Furthermore, by observing the solution procedure of the simplex method, the second step
will automatically be solved after solving the first step, if we unify both objectives of Equations
(4.100) and (4.101) as follows:

max r + l) 11c wkILk(fk(x»

s.t. ILI(zL(x, P» ~ r
IL2(ZU(X, P» ~ r
xeX (4.102)

where l) is a sufficiently small positive number. We considered Equation (4.102) an augmented

max-min approach.
With equal weights among objectives, the augmented max-min approl!.ch is, in some points
of view, a dual model of the augmented minimax programming model proposed by Sakawa and
 246

Yano [220]. However, if there are not equal weights among objectives, the proposed augmented
max-min approach will not be a dual problem of augmented minimax approach at all.
From the aspect of the fuzzy sets theory, it should be noted that the proposed max-min
approach allows for compensation among objectives, which cannot be reached by using the max-
min operator.
Besides, the max-min model (when equal weight) is a dual model of the Tchebycheff
distance criterion which has the largest measurement unit, and the average model is basically the
city block distance criterion which has the smallest measurement unit among Minkowski's I;,
metric. Thus, the augmented max-min approach is first to reach compromise solutions at a large
unit, and then reevaluates these compromise solutions and obtains an efficient compromise
solution at a smaller unit (a smaller measurement unit will provide more precise measures).
Through this strategy, we will obtain an efficient compromise solution which provides more
uniform achieved rates (realized degrees of satisfaction) for each objective than the solution
directly from the average model. The individually realized satisfaction levels of the objectives
are always more consistent from the max-min model than from the average model.
It is worth noting that the membership functions (1-'10 and 1-'20), which are elicited on the
basis of preference concept, can be any non-decreasing functions such as linear, piecewise linear,
exponential, hyperbolic, inverse hyperbolic or some specific power functions as discussed in
Chapter 2. In this study, linear membership functions are assumed, because the other types of
membership functions can be transferred into equivalent linear forms by use of the variable
transformation concept. That is ({3 is given):

> z!"o
~ ~(X)-z'--l/[z'-.-z'--l
if z!"(x)
.,(x) { if z!"- :::; z!"(x) :::; z!..o
if z!"(x) < zl..-

where z!"o and zl..- are the positive and negative ideal solutions (individual maximum and
minimum), respectively. Similarly, we can establish l-'2(x). By substituting these membership
functions into Equation (4.102), we will obtain a crisp linear programming problem under given
{3 and o.
When the interest rates (rj ) of different time periods have different possibilistic
distributions, we will have the following problem:
 247

K
max x/(1 +r1U@» + ... + XK/{ 1T [1 +rt U@)]}
k=1

K
max x/(1 +r 1L@» + ... + XK/{ IT [1 +rlL@)]}
k=1

s.t. xf X

where fL@) = rp+(rn-rI'){j, rU@) = t>-(t>-rn){j, Vi, and {j f [0, 1]. Similarly, The above
equation can be solved by the augmented max-min approach.
For the imprecise resources available, we have the following constraints of (Ax)j ~ bj,
Vi, where the preference-based membership functions of bj, Vi, are assumed to be the same as
Equation (4.86). Thus we have the crisp auxiliary constraints of (Ax)j ~ bj + Pj{1 - a), Vi, as
shown in the previous section.
In order to demonstrate the approach, let us consider the following bank hedging decision
problem.

4.2.2a Example: A Bank Hedging Decision Problem

How to manage the interest rate problem in bank balance sheet management has been
widely discussed by Booth and Koveos [17]. The probability theory is always a unique way to
handle the uncertain interest rate problem in traditional financial analysis. However, in this
study, we will assume that the interest rate is imprecise and has a possibilistic distribution,
instead of a probabilistic distribution. The considered basic bank hedging decision problem is
based on Booth and Koveos' s model. We further modify this model by considering
fuzzy/imprecise interest rates, price of futures contract, loan demand, deposit supply and ratio
of desired loan to deposit. In this problem, the fuzziness of interest rates and price of futures
should be modelled by (triangular) possibilistic distributions. On the other hand, the fuzziness
of loan demand, deposit supply and ratio of desired loan to deposit should be modelled by
preference-based membership functions because they involve the nature of the subjective
satisfaction.
This fuzzy bank hedging decision model can then be modelled as:
 248

max E (4. 103 a)

s.t. R ~ aD + mS F (4.103b)
SSTS :s;; (l - bSTS)STS· (4.103c)
SLTS :s;; (1 - bLTS)LTS· (4.103d)
SF :s;; (1 - bLn>LTS· (4.103e)
'"
P :s;; Id (4.103f)
L

D:S;; d (4.103g)
L:s;; cD (4.103h)
STS = (l - bs-rJSTS· + P STS - SSTS (4.103i)
LTS = (l - bLTS)LTS· + P LTS - SLTS (4.103j)
L = (l - bJL· + P L (4.103k)
R + STS + LTS + L = D + E (4.1031)

where

E = E· + r STS·(I-bSTS)STS· + rLTS·(l-bLTS)LTS· + r L·(I-bJL· + rSTSPSTS + rLTSPLTS + rLPL

(1) (2a) (2b) (2c) (3a) (3b) (3c)

and

(.51bs-rJ+ .5
kSTS = 1 - {rSTS·[ E (1 +rsTSY1 + (1 +rSTSY<oSIbSTSl+oSn
t=1

(.5IbLn> + .5
kLTS = 1 - {rLTS·[ E (l +rLTSY1 + (1 +rLn>-(.SIbLTSl+oS}
t=1

and where "*" indicates historical data, and where ~F' rD, rSTS and "i"LTS are fuzzy and have
triangular possibilistic distributions, and id, d and c are fuzzy and have preference-based
membership functions. The specific variables and parameters of Equation (4.103) are shown in
Table 4.5.
 249

Table 4.5 Variables and parameters

Balance sheet variables:

R Reserves
STS Short-term securities
LTS Long-term securities
L Loans
o Deposits
E Equity
Decision variables:

First stage:
PSTS Purchase short-term securities (STS)
P LTS Purchase long-term securities (LTS)
PL Purchase Loans (L)
SF Sell futures
Second stage:
SSTS Sell short-term securities (STS)
SLTS Sell long-term securities (LTS)
Parameters and other variables:

a Portion of deposits (D) associated with reserves

m Margin per dollar of futures sold (SF)
b STS Portion of short-term securities (STS) maturing
b LTS Portion of long-term securities (LTS) maturing
bL Portion of loans (L) maturing
ld Maximum new loan demand
d Maximum new deposit supply
c Maximum desired loan (L) to deposits (D) ratio
r STS Short-term security (STS) interest rate
r LTS Long-term security (LTS) interest rate
rL Loan (L) interest rate
PSF Price of futures contract sold
PPF Price of futures contract purchased
k STS Capital gains/losses of short-term security sales
k LTS capital gains/losses of long-term security sales
 250

The objective of Equation (4. 103 a) is to maximize the expected ending balance sheet value
of stockholders' equity or to maximize expected profit. Stockholders' equity at the end of the
period E equals: (1) beginning equity (E.); plus the net revenues associated with (2) non-maturing
and (3) newly acquired securities and loans; minus (4) the net revenues corresponding to
securities sold; plus (or minus) (5) capital gains or losses (book value minus anticipated ending
price); plus (or minus) (6) the gain (or loss) on the hedging transaction; minus (7) the cost of
deposits.
It should be noted that the futures purchase transaction is assumed to exactly offset the sell
position. Thus, per dollar profit of selling futures is [l-<i)pF/PSF)] which reflects the observation
that futures and spot prices are not mechanically related and permits the bank to be subject to
basis risk.
Constraint (4.103b) requires that at the period's end the bank's cash and reserve holdings
are at least a certain portion of deposits plus a margin requirement associated with the sale of
futures contracts, according to mandated and managerial reserve requirements. Constraint
(4.103c) indicates that the bank can sell short-term, non-maturing securities but cannot sell
securities purchased. Constraint (4.103i) specifies the ending balance sheet value for short-term
securities to equal short-term securities that have not yet matured plus securities purchased during
the period minus securities sold. Similarly, we have constraints (4.103d) and (4.103j) that
pertain for long-term securities. Constraint (4.103e) is constructed so that only the interest rate
risk associated with long-term securities is hedged and the magnitude of this hedge is determined
by the maximum dollar volume of long-term securities that could be sold. Constraint (4.103f)
specifies that the bank cannot make loans in excess of that supplied. Constraint (4.103k)
indicates that the ending balance sheet value for loans equals those loans not maturing during the
planning period plus loans made. Constraint (4.103g) indicates that a bank cannot obtain more
deposits than are supplied. Constraint (4.1 03h) brings together constraints (4.1 03g) and (4.1 03k)
by defining a managerially determined relationship between loans and deposits where the loan
to deposit ratio indicates the flexibility for managing liquidity. Finally, constraint (4.1031)
requires that total assets equal deposits plus equity.
Now, let us assume the numerical data as in Table 4.6 where the interest rates and price
of futures contract purchased are imprecise and have triangular possibilistic distributions such as
"about 8%" = (0.08,0.075,0.085), "about 9%" = (0.09,0.085,0.10), "about 4%" = (0.04,
0.035, 0.05) and "about $92" = $(92, 91, 94). In the above equation, the maximum new loan
demand (ld), the maximum new deposit supply (d) and the maximum desired loan to deposit ratio
 251

(c) are subjectively determined. Thus they should be modelled by fuzzy sets. Let us assume that
the membership functions of the fuzzy sets id, d and ~ are linearly decreasing functions in [400,
480], [950, 1070] and [0.8, 0.9], respectively. More precisely, we define the membership
function of ld as:

{~
ifld < 400
pOd) = - (ld-400)/80 if 400 ~ ld ~ 480
if ld > 480.

Similarly, we can also define the membership functions of d and cas the above function.

Table 4.6 Numerical data of the bank hedging decision problem

variable/ Historical Current Future

Parameter value value value
Balance Sheet

R $200
STS 100
LTS 300
L 400
D 900
E 100

Interest Rate (annual)

r STS .07 .08 "about .08"
r LTS .08 .09 "about .09"
rL .11 .14
ro "about .04 "
Futures
PSF $92
PPF "about $92"
m .025

Market Size
ld $400 (80)
d $950 (120)

Others
a .20
b sTS .50
b LTS .125
bL .125
c .8 ( .1)

Note: (1) interest rates and futures prices are expressed

in annual terms, and other data are quarterly;
(2) the numbers inside the parenthesis indicate the
corresponding tolerance values.
 252

Assume a = min (P(ld), ,,(d), ,,(c». Then we have ,,(ld) C!: a, ,,(d) C!: a and ,,(c) C!:
a. That is: Id S; 400 + (l-a)80, d S; 950 + (l-a)120 and c S; [0.8 + (l-a)O.I], where a E

[0, I]. Here, a is the minimum satisfaction level of the fuzzy constraints or the degree of
feasibility .
Finally, it is noted that the annual data should be transferred into quarterly units before
entering Equation (4.103). The numerical fuzzy bank hedging decision is then shown in follows:

max E
s.t. R C!: 0.2D + O.25SF
SSTS S; 50, SLTS S; 262.5, SF:S: 262.5
PL S; 400 + (l-a)80
o S; 950 + (l-a)120
L S; [0.8 + (l-a)O.I]O
STS = (50 + PSTS - SSTS)
LTS = (262.5 + PLTS - SLTS)
L = 350 + PL
R + STS + LTS + L =0 + E
where
E = 115.75 + O.02PSTS + O.022SPLTS + 0.035PL - (0.0175+ksTS>SSTS
- (0.02+kLTS>SLTS + (0,-.022,.011)SF - (0.01,0.00875,0.0125)0
kSTS = 1- [0.01751(1.02,1.01875,1.02125) + 1.01751(1.02,1.01875,1.02125)1.~
kLTS = 1 - [0.02/(1.0225,1.02125,1.025) + 0.021(1.0225,1.02125,1.025)2 +
0.02/(1.0225,1.02125,1.025)3 + 0.02/(1.0225,1.02125,1.025)4 +
1.021(1.0225,1.02125,1.025)4.5 .

To solve this problem, let us first assume that the acceptable degree of the feasibility of the
constraints is O. 7 (or a= 0.7) and the acceptable degree of the possibilistic occurrence of the
fuzzy objective is 0.5 (or fJ = 0.5).
Next, we should obtain the PIS and the NIS (or individual maximum and minimum)
solutions for both objectives, zL(x, 0.5) and zU(x, 0.5). The solution is then zL* = 133.6698,
zu* = 136.6401, zL- = 94.0643 and zU- = 100.9776.
With the PIS and NIS, linear membership functions of both objectives of Equation (4.98)
can be obtained. And the remaining problem is to solve Equation (4.101) which is (assume a
= 0.7, fJ = 0.5, a = 0.01 and wl = w2 = 0.5):
 253

max r + 0.OOO126EL + O.OOOl40Eu

s.t. EL - 39.6055r ~ 94.0643
EU - 35.6625r ~ 100.9776
R - 0.2D - O.25SF ~ 0
SSTS s 50, SLTS S 262.5, SF s 262.5
P L s 424, D S 986, L - 0.83D S 0
STS - P STS + SSTS = 50
LTS - P LTS + SLTS = 50
L - P L = 350
R + STS + LTS - D -E =0
EL S E S E U
EL - O.02PSTS - 0.0225PLTS - 0.035PL + 0.01354SSTS + 0.026776SLTS +
O.l1SF+ O.0l125D = 115.75
EU - 0.02PSTS - 0.0225PLTS - 0.035PL + 0.011704SSTS + 0.018832SLTS -
0.0055S F+ 0.009375D= 115.75
r e [0, 1]

where all the decision variables are non-negative. The numerical solution is provided in Table
4.7. Table 4.8 presents the solution of the balance sheet and the corresponding decisions.
Furthermore, we also solve this problem with various possibility degrees. Their solutions are
provided in Table 4.7 and Table 4.8. The possibility distribution of the optimal equity is
presented in Figure 4.18. Obviously. it is a non-linear, bell-shaped distribution.

~able 4.7 Solutions of the bank hedging decision problem

(a = 0.7)

~ PIS NIS Compromise solution

z L· z U· ZL- zu- ZL ZU r
1.0 135.2494 135.2494 1.0000
0.9 134.7159 135.5271 98.5064 99.8898 134.7159 135.5240 0.9999
0.8 134.4540 135.8050 97.3944 100.1614 134.2779 135.6356 0.9952
0.7 134.1916 136.0824 96.2821 100.4325 133.8897 135.7985 0.9920

0.6 133.9312 136.3600 95.1745 100.7039 133.5161 135.9781 0.9893

0.5 133.6698 136.6401 94.0643 100.9776 133.1435 136.1662 0.9867
0.4 133.4085 137.1803 92.9546 101.2498 132.6211 136.4809 0.9805
0.3 133.1466 137.7533 91.8455 101.5211 132.0675 136.8067 0.9739

0.2 132.8852 138.3268 90.7369 101. 7930 131.5050 137.1304 0.9673

0.1 132.6236 138.9025 89.6280 102.0670 130.9343 137.4550 0.9607
0.0 132.3619 139.4760 88.5194 102.3389 130.3572 137.7779 0.9543
 254

Table 4.8 The optimal balance sheet and decision (a 0.7)

1.0 0.9 0.8 0.7 0.6 0.5

Balance sheet

R 197.2000 197.2000 197.2000 197.2000 197.0000 197.2001

STS 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000

LTS 924.0494 924.1552 924.4357 924.5985 924.7780 924.9662

L 774.0000 774.0000 774.0000 774.0000 774.0000 774.0000

D 986.0000 986.0000 986.0000 986.0000 986.0000 985.9999
E 135.2494 135.5240 135.6356 135.7985 135.9781 136.1662

Decision

1st stage
PSTS 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
P lTS 924.0493 924.1552 844.5420 817.8739 805.5091 798.6815
Pl 424.0000 424.0000 424.0000 424.400 424.0000 424.0000
SF 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000

2nd stage

SSTS 50.0000 50.0000 50.0000 50.0000 50.0000 50.0000

SlTS 262.5000 262.5000 182.6064 155.7754 143.2311 136.2154

(Continuous)

{3 0.4 0.3 0.2 0.1 0.0

Balance sheet

R 197.2001 197.1999 197.2001 197.2000 197.2000

STS 0.0000 0.0000 0.0000 0.0000 0.0000

LTS 925.2808 925.6066 925.9304 926.2550 926.2578

L 774.0000 774.0000 774.0000 774.0000 774.0000

D 986.0000 986.0000 986.0000 986.0000 986.0000
E 136.4809 136.8067 137.1304 137.4550 137.7779

Decision

1st stage
PSTS 0.0000 0.0000 0.0000 0.0000 0.0000
P lTS 823.4657 845.1398 862.0480 875.7864 887.2448
Pl 424.0000 424.0000 424.0000 424.0000 424.0000
SF 0.0000 0.0000 0.0000 0.0000 0.0000

2nd stage
SSTS 50.0000 50.0000 50.0000 50.0000 50.0000
SlTS 160.6861 182.0331 198.6176 212.0314 223.1670
 255

"e (or (j)

1.0 -------------------

0.9 ------------------

0.8

0.7

0.6 - - - - - - - - - - - - - - J-lo-+-+- ..,...."-t-\

O. 5 - - - - - - - - - - - - - ++-T--;'-+---t",""~

0.4

0.3

0.2

0.1

0.0 E
130 131 132 133 134 135 136 137 138

Figure 4.18 The possibility distribution of the optimal equity

From the above solution tables, the bank should obtain the funds through the maturation
of its existing short-tenn security, long-tenn security and loan balance, and accept all the deposits
supplied to it. Reserve balances should be reduced to the lowest level possible as indicated by
the reserve and potential margin requirements. Thus, the reserve balance is reduced to 197.2
from 200. All of the non-maturing short-tenn securities should be sold. The amount of selling
long-tenn securities is significantly varied with the acceptable possibility degree. Therefore,
controlling the long-tenn securities is one of the most important decisions in order to maximize
the equity. The bank should, at the same time, use some of these funds to purchase long-tenn
securities and loans. Again, purchase of the long-tenn securities is very sensitive to the degree
of possibility, and thus the bank should pay more attention to controlling long-tenn securities.
On the other hand, the bank should always purchase all possible loans which will greatly increase
the equity.
In this particular case, the bank should not purchase short-tenn securities and futures
which will not increase its equity. Although we have considered the possible hedging strategy,
no hedging is required. However, if the purchase price of futures is reduced, the potential
hedging profit will increase; thereby we will purchase futures to hedge the interest-rate risk.
Similarly, we can obtain another distribution of the optimal equity and the corresponding
balance sheet for a different a value ( degree of allowable feasibility).
 256

In this subsection, we have discussed a linear programming problem with a maximization

objective of an investment discounted value where the interest rate is imprecise and has a
triangular possibility distribution. An auxiliary bi-objective linear programming model is
proposed to resolve this possibilistic nature. Furthermore, we have also proposed an augmented
max-min· approach to solve this crisp bi-objective (multiple objective, also) decision-making
problem.
Through solving the numerical bank hedging decision problem, we have realized that the
distribution of the objective function is a non-linear bell shape, even though the interest rates and
the price of futures contract purchased have simple triangular possibility distributions. Moreover,
the solution may be very sensitive to the allowable degree of the possibility of the interest rates.
Therefore, a rather detailed solution table should be presented to the decision maker, in order to
explain the structure of such a complex problem. In our particular example, we also find that
long-term securities are key factors in maximizing equity, for they are very sensitive to the
possibility degree of the corresponding interest rate.
In practice, the possibility distributions of the interest rates should be different as time goes
on. To solve this dynamic system, the solution procedure discussed here can be applied also,
except we need to build an interactive system to allow for the dynamic change.
Finally, we should emphasize that the desired possibilities (P) for the interest rates should
not be confused with the desired degrees (ex) of the feasibility. Because, the membership
functions of the feasibility are elicited on the basis of the preference concept. Thus, the desired
degree of feasibility may be strongly dependent upon the psychological value system. On the
other hand, the possibility distributions are built on the basis of the possible occurrence of an
event which is analogous to the probability concept. The possibility distribution is then
determined and influenced by the (considered) physical environment. Of course, both can be
subjectively evaluated by the decision makers.

4.2.3 Stochastic Possibilistic Linear Programming

In the decision-making process we might face a hybrid environment where both linguistic
and frequent imprecise nature exists. For instance, the interest rate for the coming year might
have a 60 % chance of being "around 8 %" and a 40% chance of being "around 9.5 %," when a
particular new government policy might either be executed (60% chance) or aborted (40%
chance). The problem of frequent imprecision has been solved by using the probability theory.
On the other hand, the problem of linguistic imprecision should be handled by using the fuzzy
 257

sets theory. Thus, to solve this hybrid problem, it is necessary to use both theories
simultaneously. (Lai and Hwang [142])
The hybrid problems have been discussed in the literature since Zadeh in 1968 [275]. The
most detailed discussions of vague statistical problems were provided by Kruse and Meyer
[BM36] , Negoita and Ralescu [BM41] , Jones, Kaufmann and Zimmermann [BM23] and
Kacprzyk and Fedrizzi [BM27]. For a linear programming problem with hybrid environment,
Luhandjula [156] used the concepts of fuzzy and non-fuzzy probabilities of a fuzzy event
proposed by Zadeh [275], and the concepts of the expected values and min-operator to handle
the hybrid imprecision on the constraint coefficients. Wierzchon [264a] considered the problem
that the right-hand side of the constraint inequalities are independently distributed normal random
variables with fuzzy mean values and fuzzy standard deviations.
Here, we will discuss the problem of maximizing investment discounted value. That is:

max Ej x/(l +~)i

S.t. x EX (4.104)

where i = {is I 'is = (rsm, rl, rsO) with probability ps and s is the state index} (see Figure 4.19
when s = 1 and 2 only). As shown in Figure 4.19, rsm, rl and rso are the most possible,
pessimistic and optimistic values, respectively.

7T(r)
I
1

o r
rIP r 1• rl rIo r 2• r 20
(8.2%) (st) (9%) (9.2%) (9.5%) (10%)
Pl(60%) P2(40%)

Figure 4.19 The probabilistic and possibilistic natures

of the interest rate (Lai and Hwang [142])
 258

To solve Equation (4.104), we must first resolve the stochastic and possibilistic natures.
For computational efficiency, we may use the concept of expected value to resolve the
probabilistic nature. That is to maximize the expected value of the stochastic and possibilistic
objective. Of course, we may also minimize the variance and/or maximize the skewness of this
objective function. In this study, we first consider the case of maximizing the expected value.
Thus, with the assumption of the independence property we obtain the following possibilistic
linear programming:

max E5 { Ej ".i/(1 +rJ }P5

s.t. x E X
or
max Ej Xj { E5 p/(1 +rJ }
s.t. x EX (4.105)

where -r5 -_ (rsm ,rsp ' rs°).

Obviously, our next step is to solve the possibilistic nature of Equation (4.105). To do
that, let us rewrite the possibilistic distribution of rs = (rsm, rl, rso) , Vs, as the following
function form (see Figure 4.19):

ifrl~rs~rsm

if rsm~rs~rso
otherwise. (4.106)

Then for each p-level cut, we have the confidence interval of [rsL@), r su @)] = [rl+(rsm-rl)p,
rsO-(rsO-rsm)p], P E [0, 1] as in the previous section. Then, Equation (4.105) will be equivalent
to:

max
XEX

= max Ej Xj { Es p/[l +r5L@), 1 +rsU@)V }

XEX

= max Es { Ej ".iPs([1I(1 +rsU(p»' 11(1 +rsL(p»V }

XEX
 259

= max Es Ej [(x,;p.)lIj /(1 + rsU(fJ» , (xjPs) l/j /(1 +rsL(fJ»p

XEX

= max [zL(x,~), zU(x,~)]. (4.107)

XEX

The objective function of Equation (4.107) is obviously a crisp interval value when ~ is given,
since the summation of interval values will be an interval value. Therefore, we may use the
weighted average concept. That is to maximize wzL(x,~) + (l-w)zu(x,~), where w E [0, 1]
indicates the relative weight (importance). However, to maximize an interval value can also be
obtained by simultaneously maximizing the lower and upper bound of the interval value. Thus
we proposed the following bi-objective auxiliary problem for Equation (4.104):

max zL(x,~)
max zU(x,~)
s.t. x EX (4.108)

where ~ E [0, 1]. If ~ is first given, then rL(fJ) and rU(fJ) will be constants and the problem of
Equation (4.108) will be a crisp bi-objective linear programming problem which can be solved
by the previous augmented max-min approach. That is:

max r + a ~ wkl'k(fk(x»
s.t. I-'1(zL(x, ~» ~ r
JI.2(zu(x, ~» ~ r
XE X (4.109)

where ais a sufficiently small positive number. With the membership functions 1-'1(.) and 1J.2(.)
as discussed in the previous section, Equation (4.109) will be a crisp linear programming
problem under given ~ and a.
When the interest rates (ry of different time periods have different possibilistic
distributions, we will have the following problem:

n _
max x1/(l +rl) + ... + "n/[ TT (l +~)]
j=1
s.t. x E X (4.110)
 260

where i=j = {rjS I TjS = (rjsm, rjl, rjso) with probability Ps and s is the state index} for j = 1, ... ,
n. Similarly, Equation (4.110) can be solved by the augmented max-min approach.
In order to demonstrate the proposed approach, a real-world problem should be considered
and solved.

4.2.3a Example: A Bank Hedging Decision Problem

Consider the bank hedging decision problem as discussed in Section 4.2.2a. Here we
assumed the following environment (see Table 4.9 also) (Lai and Hwang [142]):

(1) Some interest rates are assumed, for instance, to have a 40% chance of being "around 8%"
and a 60% chance of being "around 9.5%." That is: the decision maker can only provide
linguistic values, which are then modelled by possibilistic distributions, in each (probability)
state. Similarly, the price of the futures contract also has the same nature.
(2) The maximum desired values of loan demand and ratio of desired loan to deposit are often
imprecise/fuzzy and have preference-based membership functions.
(3) Deposit supply has the nature of both randomness and fuzziness. The randomness is
modelled by multiple states as in (1). The fuzziness is modelled by preference-based membership
function as in (2).

This stochastic possibilistic bank hedging decision model is then:

max Es PsEs
S.t. ~ ~ aDs + mSF' Vs
Ss,STS S (1 - bSTS)STS", Vs
S5,LTS S (1 - bLTg)LTS", Vs
SF S (1 - bLTS)LTS"

-
Os S d5 , Vs
L S cDs, Vs
STSs = (1 - bSTg)STS"+ PSTS - Ss,STS' Vs
LTSs = (1 - bLTS)LTS" + P LTS - S5,LTS' Vs
L = (1 - b0L" + PL
~ + STS5 + LTSs + L = Os + Es' Vs (4.111)
where
 261

Es = E· + rSTs·(1-b STS)STS· + rLTS·(I-bLT,s)LTS· + rL·(1-bVL• + rSTSPSTS + rLTSPLTS

+ rLPL - (rSTS· + k.,ST,s)Ss,STS + {rLTS· + k.,LT,s)Ss,LTS + [1-<Ps, PF/Ps,SF)]SF - rs,nDs

(.SIb. ']'- \+.5

(SIbLTS)+ S
ks,LTS = 1 - {rLTS• [ ... 81
E
- _ -
(1 +rs,LT,s) ~ + [1 +rs,LTsJ . . }
t=1

where "*" indicates historical data; Ps,PF' fs,D' fs,STS and fs,LTS, Vs, are fuzzy and have
triangular possibilistic distributions; and id, ds, Vs, and ~ are fuzzy and have preference-based
membership functions. The variables and parameters of the above equation are defined in Table
4.5. The explanation of the above equation has been given in Section 4.2.2a. However, it
should be noted that the constraints in the above equation involve every state at the same time.
Thus, it is not equivalent to solve the problem in Section 4.2.2a for each individual state and then
to find their expected values.
Now, assume that the input data are either precisely numerical or linguistic (see Table
4.9). For simplicity, two stages (s = 1,2) are considered and have equal probabilities for each
stage. In Table 4.9, while other data are quarterly, interest rates and futures prices are expressed
in annual terms and must be transferred into quarterly values before substituting them into
Equation (4.111).

have been discussed in the previous section.

- ...
The triangular possibilistic distributions of interest rates and price of futures purchased
As to the fuzzy sets ld, d l , d2 and
- c, their
membership functions are assumed to be linear decreasing functions within the considered ranges.
'" is defined as:
For example, the membership function of the fuzzy set ld

<
~
ifld 400

!'!lei) - [ - (Id-4OO)/80 if 400 s ld S 480

ifld > 480.

Similarly, we can also define the membership functions of d l , d2 and ~ as the above function.
Now, let a = min (JLOd), ",(d l ), ",(dV, ",(c». We have ",(ld) ~ a, ",(d l ) ~ a, ",(dV ~

a and ",(c) ~ a. That is:

 262

PL S 400 + (l-al)80,
01 S 950 + (1-aV2OO, 02 S 1050 + (1-a3)300
L S [0.8+0.1(l-a~]01' L S [0.8+0. I (l-aS)]02.

Here, a is the minimum satisfaction level of the fuzzy constraints or the degree of feasibility.

Table 4.9 Numerical data of bank balance sheet problem [142]

variable/ Historical Present -- Future values --

Parameter value value state 1 state 2
(Prob.=.5) (Prob.=.5)
Balance Sheet
R $200
STS 100
LTS 300
L 400
D 900
E 100
Interest Rate (annual)
r~s .07 .08 "about .08" "about .09"
r LTS .08 .09 "about .09" "about .10"
rL .11 .14
rD "about .04" "about .05"
Futures
PSF $92
PPF $91 $90
m .025
Market Size
ld $400 (80)
d $950 (200) $1050 (300)
others
a .20
b sTS .50
b LTS .125
bL .125
C .8 (.1)

Note: (1) interest rates and futures prices are expressed in

annual terms, and other data are quarterly;
(2) the numbers inside the parenthesis indicate the
corresponding tolerances;
(3) "about .08" = (.08, .075, .085); "about .09" =
(.09, .085, .10); "about .10" = (.10, .09, .12);
"about .04" = (.04, .035, .05); "about .05" =
( • OS, .04, .06).
 263

The numerical model of the bank hedging decision under fuzziness and randomness is
then:

max E = .5E1 + .S~

s.t. Rl ~ .201 + .000SF, R2 ~ .202 + .000SF

SlSTS :S 50, S2STS :s 50
SILTS :s 262.5, ~LTS :s 262.5
SF s 262.5
PL :s 400 + (l-al)80
01 :s 950 + (l-a~200, 02 S 1050 + (l-a3)300
L :s [.8+(1-a~.1]01' L :s [.8+(I-a5).1]02
STS 1= 50 + PSTS - SISTS, STS2 = 50 + P STS - S2STS
LTS1 = 262,S+PLTS-SILTS' LTS2 = 262.5+PLTS-S2LTS
L = 350 + P L
R 1+STS 1+LTS 1+L = 01+El' R2+STS2+LTS2+L = °2+~
where
El = l1S.7S+.02PSTS+·0225PLTS+·03SPL-(·017S+klSTS>SlSTS
-(.02+k1LTS>SILTS+ .01087SF"(.01,.OO87S, .012S)0IS
~ = l1S.7S+.02PSTS+·0225PLTS+·03SPL-(·017S+k2STS)~STS
-(.02+k2LTS>~LTS+.02174SF"(.0125,.01, .015)025
k lSTS = 1-[.01751(1.02,1.01875,1.02125)+ 1.01751(1.02,1.01875,1.02125)1.5]
k2STS = 1-[.01751(1.0225,1.02125,1.025)+ 1.0175/(1.0225, 1.02125, 1.025)1.5]
kILTS = 1-{ I:j=1_>4.02I(1.0225,1.02125,1.025)i+ 1.02/«1.0225,1.02125,1.025)4.5
k2LTS = 1-{ I:j = 1-> 4. 02/(1.025, 1.0225, 1.03)i + 1.02/«1.025,1.0225,1.03)4.5.
To solve the problem presented in Table 4.9, we first assume that the acceptable degree
of feasibility for the constraint is 0.7 (a = 0.7). Also, we must elicit the membership functions
of the upper and lower bounds of the expected equity from the decision maker. Here, assume
these membership functions are:

ifEk > Ek•

if Ek- :S Ek :S Ek•
ifEk < Ek-
 264

where Ek* (Ek-) is the individual maximum (minimum) for each objective, and k = U and L.
With the assumption of ~ = 0.001 and wI = w2 = 0.5, we can obtain Ek* ( = 0.5E 1k*
+ 0.5~k), and Ek- (= 0.5Et + 0.5~k-), for k = L and U, and the corresponding
compromise solutions, which are presented in Table 4.10, for fJ = 1.0,0.9, ... ,0.1,0.0.

~ab1. 4.10 solutions of the bank hedging decision problem

(a = 0.7)

PIS NIS Compromise solution

EI - - E2 - - - r
EL" E U- EL- E U' ElL E IU El E 2U

1.0 136.683 136.683 137.406 137.406 1.000

0.9 136.704 137.290 95.196 96.311 136.322 136.874 137.112 137.705 0.998
0.8 136.364 137.535 94.502 96.732 135.963 137.065 136.818 138.003 0.999
0.7 136.024 137.779 93.809 97.153 135.603 137.257 136.524 138.301 1.000
0.6 135.684 138.024 93.117 97.575 135.242 137.449 136.230 138.560 0.999
0.5 135.344 138.269 92.425 97.997 134.882 137.640 135.936 138.898 0.999
0.4 135.047 148.595 91. 735 98.418 134.477 137.869 135.453 139.168 0.998
0.3 134.776 138.975 91.046 98.841 134.081 138.086 135.038 139.466 0.995
0.2 134.056 139.355 90.357 99.263 133.683 138.303 134.632 139.775 0.992
0.1 134.235 139.735 89.670 99.685 133.283 138.521 134.228 140.091 0.989
0.0 133.964 140.115 88.983 100.11 132.882 138.738 133.829 140.412 0.986

For simplicity, we only show the case of fJ = 0.5 in the following. First, for fJ = 0.5
we obtain E L* = 135.3438, EU * = 138.2693, E L- = 92.4254 and EU - = 97.9966. Then the
above linear membership functions of both objectives can be easily obtained. Finally, we solve
the following linear programming problem of Equation (4.109):

max r + 0.OOOO1l7EL + 0.OOOO124Eu

S.t. EL - 42.9184r ~ 92.4254
EU - 40.2727r ~ 97.9966
R r O.2D 1-O.25S F ~ °
Rr O.2D2-O·25S F ~ °
SISTS :$; 50
S2STS :s; 50
SILTS :s; 262.5
S2LTS :s; 262.5
SF :$; 262.5
P L :s; 424
D :s; 1010
 265

L-O.83D I:s 0
L-O.83D2 :s 0
STSrPSTS+SISTS = 50
ST~-PSTS+S2STS = 50
LTSrPLTs+SILTS = 262.5
LT~-PLTS+~TS = 262.5
L - P L = 350
RI +STS I +LTSI-DI-E I =0
R2+ST~+LTS2-D2-~ = 0
ElL :S EI :S EIU, ~L :S ~ :S ~U
EIL-O.02PSTS-O·0225PLTS-0.035PL +0.01354S ISTS +
0.026776SILTS+0.01087SF+0.01125DI = 115.75

~L-O.02PSTS-0.0225PLTS-0.035PL +0.018107S2STS +
0.042431~LTS+0.02174SF+0.1375D2 = 115.75

EI u-0.02PSTS-O·0225PLTS-O·035PL +0.011703S ISTS +

0.018831SILTS-O.01087SF+0.009375DI = 115.75

~U-O.02PSTs-O.0225PLTS-O.035PL +0.015371~TS+
0.026776S2LTS-0.02174SF+0.01125D2 = 115.75

EL = 0.5EIL+0.5~L
E U = 0.5EIU+ ~u
r E [0, 1]

where all the decision variables are non-negative. The numerical solution is provided in Table
4.11. Furthermore, the possibility distributions of the optimal equities for each state are
presented in Figure 4.20. Obviously, the possibility distributions of expected equity and equities
of both states are non-linearly bell shaped. The optimal balance-sheet and two-stage decisions
at (3 = 0.5 are provided in Table 4.11.
If the acceptable possibility level is 0.5, then the bank should obtain the funds through the
maturation of its existing short-tertn security, long-term security and loan balance, and accept all
the deposits supplied to it (see Table 4.11).
Reserve balances should be reduced to the lowest level possible as indicated by the reserve
and potential margin requirements. Thus, reserve balance is increased from 200 to 267.625 and
291.625. All of the non-maturing short-term securities should be sold. Some of long-term
 266

securities (97.258) are sold in State 1. On the other hand, no long-term security should be sold
in State 2. The bank should, at the same time, use some of these funds to purchase long-term
securities and loans. However, the bank should not purchase short-term securities at all. Selling
futures should proceed (as much as possible) for hedging the interest rate risk. By taking the
above actions at fJ = 0.5, the expected equity will be between $135.40 and $138.269.
For simplicity, we will not present all the lengthy results in all cases of fJ = 1.0,0.9, ... ,
0.1, 0.0. A brief conclusion is that the amount of purchasing and selling long-term securities
varies with the change of the possibility (fJ values), in both states. Therefore, the bank should
pay more attention to controlling long-term securities. Other decisions are stable with the change
of the possibility.
Similarly, we can also obtain another distribution of the optimal equity and the
corresponding balance sheet for a different ex value ( degree of desirable feasibility).

Table 4.11 The optimal balance sheet and decisions

(a = 0.7 and P = 0.5)

Initial Ending Values

Values state 1 state 2
Balance sheet
Reserves (R) 200 267.625 291. 625
Short-term securities (STS) 100 0.000 0.000
Long-term securities (LTS) 300 880.015 977.273
Loan (L) 400 774.000 774.000
Deposits (D) 900 1010.000 1130.000
Equity (E) 100 137.640 138.898
Decision
First stage
Purchase short-term securities (P STS ) 0.000 0.000
Purchase long-term securities (P LTS ) 714.773 714.773
Purchase loans (PL) 424.000 424.000
Sell futures (SF) 262.500 262.500
Second stage
Sell short-term securities (SSTS) 50.000 50.000
Sell long-term securities (SLTS) 97.258 0.000
 267

"e (or (j) 50%

state 2
1.0

0.9

0.8

0.7

0.6

0.5 --------------7-~~------T_~~
0.4

0.3

0.2

0.1 ---------T--t~~------------------~--_7~

0.0 E
131 132 133 134 135 136 137 138 139 140 141

Figure 4.20 The possibil i ty distr i butions of the optimal equity

5 CONCLUDING REMARKS

During the last 25 years, the fuzzy set theory has been applied in many disciplines such
as operations research, management science, artificial intelligence/expert system, control theory,
statistics, etc. In this study, we concentrate on fuzzy mathematical programming which is one
of the most important fields in operations research and management science. This study provides
readers and researchers with a capsule look into the existing methods, their characteristics and
their applicability to analysis of single-objective mathematical programming under fuzziness/
imprecision.
Based on the membership function and the possibilistic distribution, we have systematically
classified fuzzy mathematical programming problems into two major categories: fuzzy and
possibilistic mathematical programming problems. In our point of view, it is rather important
to distinguish preference-based membership functions from possibility distributions. Preference-
based membership concerns the degree of the decision maker's satisfaction, while possibility is
the degree of the occurrence of an event. Eleven approaches of fuzzy linear programming and
fourteen methods of possibilistic programming have been systematically and thoroughly classified
to solve all possible existing problems. Fifteen practical problems have also been discussed in
this study. Most of these methods and applications have been proposed by various researchers
in the last decade, but here for the first time they are presented all together.
It is important to compare fuzzy and possibilistic approaches with stochastic or
probabilistic ones for solving fuzzy/imprecise/uncertainty nature. In Section 5.1, we will
summarize some essential reasons why fuzzy set theory is preferred in this study. Furthermore,
we concisely present the advantages of fuzzy and possibilistic programming compared to
stochastic programming in Section 5.2. Section 5.3 will present some directions of future
studies. We finally present a short introduction of our following volume of Fuzzy Multiple
Objective Decision Making: Methods and Applications, and the previous volume of Fuzzy
Multiple Attribute Decision Making: Methods and Applications in Sections 5.4 and 5.5,
respectively.

5.1 Probability Theory versus Fuzzy Set Theory

Since Fortuna, the fickle, wheel-toting goddess of chance, ruled a large and growing
empire in the sciences, the law of the realm has been probability theory and statistics [BM12].
 269

From the beginning of the mid-seventeenth century, probability theory spread in the eighteenth
century from gambling problems to jurisprudence, data analysis, inductive inference (induction
may be traced back at least as far as Bacon's Novum Organum in 1620s [BM5]), and insurance;
and from there to sociology, to physics, to biology, and to psychology in the nineteenth century;
and on to agronomy, polling, medical testing, baseball, and innumerable other practical (and not
so practical) matters in the twentieth [BMI2].
Thus, the concept and techniques of the probability theory are often used to deal
quantitatively with imprecise goals, constraints and consequences of possible actions. Imprecision
- whatever its nature - is tacitly accepted to be equated with randomness [12]. However, the
imprecise nature may be associated with classes in which there is no sharp transition from
membership to nonmembership. From the view of psychology, computer engineering and
artificial intelligence and risk analysis, a rough but fairly functional classification of objections
(of the probability theory) includes: (I) phenomenological, philosophical and semantic arguments
for the existence of non-probabilistic uncertainty; (2) meta-rationalist claim that the probability
theory is untenable [BM47].
For the first objection, we may have certainty about a fuzzy event (such as "is it raining")
or a graded attribute (such as "John is mil"). It seems counter-intuitive to describe the vagueness
inherent in the above events or attributes by subjective probabilities [288]. In Zadeh's terms, this
objection is not to the axioms of the probability theory itself, but instead to its limited capacities
for expression. Thus, Zadeh claimed that the probability theory may be unable to represent the
meanings of propositions involving imprecision/fuzziness of predicates (mWl, )!QWl&, ••• ),

quantifiers~, 2&n, ... ), probabilities~, lml&,~, ...), truth values (very true, ... )
and modifiers ~, somewhat, ... ).
For the second objection, meta-rationality arguments against probability refer primarily
to its computational complexity and the amount of information required, especially in the context
of applying Bayes's Rule. Despite computational complexity, we must frequently add or delete
individual pieces of knowledge in actually constructing a knowledge base. This cannot be done
without revising the dependency information on all relevant subgroups of evidence. As to the
required sufficient information, even sophisticated records cannot provide that kind of associative
information. This information unavailability criticism includes (I) the absence of such
information and (2) the changeable, unreliable or irreducibly contextual nature of information.
To distinguish fuzziness from frequentist's probability seems obvious and has largely been
conceded by many probabilists. However, subjectivists consider probability as a degree of belief
 270

which is close to, but not equal to the fuzzy set theory. In this aspect, the fuzzy set theory does
provide the attribution of a non-probabilistic meaning to the concept of graded sets. For
example, we may be entirely certain that a particular object is "dark." In this case, there is no
probabilistic uncertainty associated with this proposition, and the only source of uncertainty is
the graded nature of the concept of darkness. Another argument is based on the concept of the
possibility theory. Dubois and Prade [BM7] pointed out that a fuzzy concept like "tall" conveys
a possibility distribution over the set of all heights, so that the possibility grade reflects the degree
to which a given object might possibly belong to the set of tall people given hislher height. They
also argued that possibility is a looser kind of uncertainty than probability, since anything that
is impossible must be improbable but whatever is possible need not be probable.
As pointed out by Smithson [BM47], the majority of arguments· for and against fuzziness
rest primarily on appeals to axiomatics, meaningfulness, ease of expression and use, semantics,
andlor philosophical distinctions. As a result, probabilists argue that we need not invent a new
theory to handle uncertainty about probability, instead we may use meta-probabilities (higher-
order probability)]; or that any self-consistent system for quantifying degree of belief with a
single number must be probabilistic. Even such apparently non-probabilistic concepts as
possibility may also be represented by probability. For example, a possibility distribution is
merely subjective probabilities associated with possibilistic statements. Thus we only induce a
second-order probability distribution over the first -order subjective probabilities. However, some
evidence shows that high-order imprecision has little practical importance, or that human judges
do not usually resort to higher-order. Furthermore, if the new frameworks deliver the goods in
applications, then they will continue to be used regardless of what the probabilists are saying;
unless the probabilists can show similar capabilities for their frameworks. In conclusion, the
fuzzy set theory abandons the laws of the excluded middle (or additivity requirement) and
contradiction, and thus abandons the standard probability calculus altogether, since it destroys
the de Morgan relations. Studies on fuzzy set theory are not going to block the search for reality
nor restrict the development of probability, but are going to impel humanity to develop its full
power.

5.2 Stochastic versus Possibilistic Programming

Some major arguments about the differences between probability and fuzzy set theories
have been discussed in Section 5.1. Here, we would like to discuss the advantages of
possibilistic programming compared to stochastic programming.
 271

First, stochastic programming deals with situations where input data (c, b and/or A) of
Equation (5.1) are imprecise and described by random variables. The basic idea is to convert
the random/probabilistic nature of the problem into an equivalent deterministic situation. For
more than three decades, there were numerous stochastic programming models in the literature.
Interested readers may refer to Wets [260] and Stancu-Minasian [BM48] for details.
Among various stochastic models, we like to introduce the chance-constrained
programming developed by Charnes, Cooper and Symonds [39], and Charnes and Cooper [40].
A chance-constrained programming model can be defined as:

max z = Ej CjXj
s.t. P{ Ej ~jXj ~ bi } ~ 1 - ai' Vi
Xj ~ 0, Vj (5.1)

where Cj' bi and ~j are all random variables and P{} is the degree of probability. The name
"chance-constrained" follows from each constraint Ej ~jXj ~ b i being realized with a minimum
probability of 1 - ai, 0 < ai < 1.
For illustrative purposes, let us assume hi and ~j are normally distributed with known
means and variances. If Cj is a random variable, it could always be replaced by its expected
value. Thus we have the following chance-constrained programming problem:

max z = Ej CjXj

s.t. P{ Ej ~jXj - bi ~ 0 } ~ 1 - ai' Vi

Xj ~ 0, Vj (5.2)

where hi and ~j are normally distributed as shown in Figure 5.1.

To solve Equation (5.2), let gi = Ej ~jXj - bi. Then gi is also normally distributed with:

where XT = (Xl' X2' ... , Xn, 1) and Di = the ith covariance matrix =

Var{ail}

Cov{~n,~d

Cov{bi,~d
 272

Figure 5.1 Probability density function f(b i ) of b i where

f(b i ) = (2WVar{bi})-1/2exp[-(bi-E{bi})/2var{bi}] for -~<bi<~

Now, the constraints of Equation (5.2) may be treated as:

gj-E{gj} O-E{gj}
P{gj :S; O} = P{ - - - - : S ; . } ~ I - (Xj
Var{gj}112 Var{gj}]112

or

where fII represents the CDF of the standard normal distribution.

Let Sod be the standard normal value such that fII(Saj) = I - (Xj. Then P{gj :S; O} ~ I-
(Xj is realized, iff:

Finally, we have the following deterministic model:

max z = Ej CjXj

s.t. Ej E{lljj}xj- E{bj} + Saj(XTDjX)1I2, Vi

Xj ~ 0, Vj (5.3)

which can then be solved by any constrained non-linear programming algorithm.

It is noted that Equation (5.3) is actually a non-linear programming problem, but not the
 273

original linear programming problem. Obviously, it is not easy to solve a non-linear

programming problem. On the other hand, possibilistic programming provides more efficient
techniques to solve the imprecise nature of c, b and/or A and also preserves the original linear
model. Besides, membership functions/possibility distributions provide more flexible and
meaningful representation of imprecision/uncertainty.
Consider the following possibilistic programming problem:

max z = I:j CjXj

s.t. I:j Bijxj :S hi' Vi
Xj ~ 0, Vj (5.4)

where biand Bij , Vi,j, are imprecise and have their corresponding triangular possibilistic
distributions 1I"bi = (hi' biL, biU) and 1I"aij = (~j' ~jL, ~jU) as shown in Figure 5.2. 11". is the
degree of possibility .

fl.

Figure 5.2 Illustration of a triangular possibilistic distribution

To solve Equation (5.4), the following auxiliary model (see Chapter 4 for details) might
be used:

max z = I:j CjXj

s.t. ( I:j ~jXj' I:j ~jLXj' I:j ~jUXj ) :S (hi' biL, b j u), Vi
Xj ~ 0, Vj. (5.5)
 274

The imprecise constraints of the above problem can be resolved by considering the following

inequalities:

(S.6)

which preserve that the triangular possibilistic distribution of hi must greater than or equal to the
one of Ej iijxj as indicated in Figure S.2.
Finally, we have the following deterministic model:

max z = Ej CjXj
s.t. Ej ~jXj ~ bi, Vi
Ej ~jLXj ~ b iL, Vi
Ej ~jUXj ~ biU, Vi
Xj ~ 0, Vj. (S.7)

Equation (S.7) is still a linear programming problem. Of course, the number of constraints has
been extended to around three times. In our opinion, this increase of the number of constraints
might not cause as many trouble as a non-linear programming problem of Equation (S.3).
The question as to what is a better approach encompasses the following criteria:

(I) Execution time required.

(2) Size of the considered problem (dimensionality, number of constraints).
(3) Accuracy of the solution with respect to the optimal decision variable and/or the objective
function and constraints.
(4) Simplicity of use.
(S) Simplicity of computer program to execute the algorithm.
(6) Applicability to real-world (large scale) problems.

According to these criteria, Equation (S.7) (linear) is obviously better than the Equation (S.3)
(non-linear). That is, possibilistic programming is better (more efficient) than stochastic
programming in the sense of methodology.
Furthermore, the assumed randomness for the input coefficients may not be good enough
to represent the inherent impreciseness or uncertainty. With assumed normal distributions, for
example, negative values of bi/~j are possible, this result is clearly counter-intuitive and
 275

incorrect. Therefore, bounded distributions are better tools to represent a mathematical

programming problem with imprecise input. As to the meaning of possibility and probability
theories, let us consider the following example:

"Tomorrow it will probably rain QJQ1. "

Usually, probability theory is used to model probably and possibility theory is used to model Q.
JQJ.. The imprecise nature of c, b and/or A may be described by the linguistic term of "around
34.5," for instance, than by the chance concept of "70% to be 34.5." Thus, possibilistic
programming techniques are considered to be more efficient and meaningful than stochastic
programming techniques.

5.3 Future Research

In the last decade, fuzzy set theory has provided a new research direction on both concepts
and methodologies to formulate and solve mathematical programming and multiple objective
decision making problems. However, a number of problems still remain to be solved in the
future research. These problems are summarized in the following:
(1) In this study, we often assume that membership functions and possibility distributions are
reasonably given. However, in practice, we should generate and obtain membership functions
and possibility distributions from decision makers and/or historical resources. Even though there
have already been some efforts on generating membership functions and possibility distributions,
more research is urgently needed.
(2) To solve mathematical programming problems in various fuzzy/imprecise environments, we
often use the max-min operator, although many other operators already exist. Thus, we should
extend these existing operators or develop new operators to fit our specific mathematical
programming problems in future studies.
(3) As indicated before, ranking approaches are very important to resolve fuzzy/imprecise
constraints. However, most of the existing ranking approaches are not perfect (see Chen and
Hwang [BM3]). Searching for better ranking methods is an urgent and momentous need to
resolve fuzzy/imprecise constraints and other problems.
(4) In the literature, few fuzzy multiple objective decision making and possibilistic multiple
objective decision making problems have been solved. At the same time, the efficiency of some
existing approaches should be further improved.
 276

(5) It frequently happens that fuzziness/imprecision and randomness simultaneously exist in a

decision process. However, little research has been done on solving stochastic fuzzy/possibilistic
mathematical programming and multiple objective decision-making problems.
(6) Up to date fuzzy problems are first transferred into crisp problems and then solved by
traditional (crisp) computer packages. The efficiency of fuzzy methodology is reduced. Thus,
research on developing fuzzy algorithm and coding existing fuzzy approaches in computer
packages is urgently needed.
(7) The most important area for further research would probably involve real applications with
real decision makers. Very few such applications have been reported in the literature of fuzzy
mathematical programming. It should be noted once more that all concepts and techniques are
developed for solving real-world problems.

5.4 Introduction of the Following Volume

Single-objective mathematical programming, however, cannot solve all decision-making
problems of selecting a possible course of action from available alternatives. In most such
problems, the multiplicity of criteria for judging the alternatives is pervasive. That is, for many
such problems the decision maker wants to attain more than one objective or goal in selecting the
course of action, while satisfying the constraints dictated by environment, processes and
resources. For example, the objectives of a water quality management problem may be to raise
the dissolved oxygen levels, to keep a high percentage return on industrial investment, and to
retain low tax rates for cities. For a nutrition allocation problem, the objectives may be to
maximize carbohydrate intake, to minimize cholesterol intake, and to minimize cost. For a bank
balance sheet management problem, the objectives may be to maximize net after-tax profits, and
to minimize the capital adequacy and risk assets.
Mathematically, these multiple objective mathematical programming problems (or multiple
objective decision making MODM) can be formulated as:

max Z = [f1(x), f2(x), ... , fK(x)]

s.t. gj(x) {s,=,~} 0, Vi and x ~ 0 (5.8)

where fk and gj(x) may be either linear or non-linear functions. The basic characteristic of
MODM is that the objectives are apparently non-commensurable and conflict with each other.
To resolve these problems, Hwang and Masud [BM19] have already provided a systematical and
 277

thorough survey. Since then, some new approaches, especially fuzzy programming techniques,
have been published in the literature. These new approaches will be presented in our second
volume of Fuzzy Multiple Objective Decision Moldng: Methods and Applications which is
considered as a sequel of Hwang and Masud's Multiple Objective Decision Moldng: Methods and
Applications [BM19].
Furthermore, the input data of the MODM problem may be fuzzy/imprecise because of
incomplete or non-obtainable information. To solve these fuzzy MODM problems, we have also
carried out a complete survey of all existing approaches. The classification of fuzzy MODM is
similar to fuzzy mathematical programming problems as shown in Figures 1.3 and 1.4. We also
distinguish fuzzy MODM from possibilistic MODM for the same reasons as discussed in this
volume.
It should be noted that as with this volume, we will again make a strong effort to present
practical problems. After all, all concepts and methods are developed for solving real-world
problems. In presenting practical problems, we believe that the next volume can also fill the gap
between researchers and practitioners (or real decision makers). Thereby, practitioners can really
use the concepts and methods developed by theoretical researchers to solve real-world problems
with real data.

5.5 Fuzzy Multiple Attribute Decision Making

In the study of decision making in complex environments, such terms as "multiple
objectives," "multiple attributes," "multiple criteria," or "multiple dimensions" are used to
describe decision situations. Often these terms are used interchangeably, and there are no
universal definitions for them. Multiple criteria decision making (MCDM) has become the
accepted designation for all methodologies dealing with MODM and/or MADM. In our survey,
the problems of MCDM are broadly classified into two categories - MODM and MADM. The
definitions to distinguish MODM from MADM are presented in Hwang and Yoon [BM20].
Basically, MADM methods are for selecting an alternative from a small, explicit list of
alternatives, while MODM methods are used for an infinite set of alternatives implicitly defined
by the constraints. Thus, MADM methods are useful for solving choice problems while MODM
methods address design problems.
If the input data of MADM problems are fuzzy/imprecise, then we will have fuzzy
MADM (FMADM) problems. Literature on FMADM methods and applications has been
extensively reviewed by Chen and Hwang [BM3]. FMADM methods were systematically
 278

classified; a taxonomy of the methods is presented in Table 5.1. Detailed discussions of these
methodolgies are provided in [BM3].

Table 5.1 A taxonomy of FMADM methods

Problem Data type Corresponding Technique Approach

size MADM methods involved

FMADM n < 10 All fuzzy Simple a-cut Baas and Kwakernaak's

m < 10 Additive Kwakernaak's
Weighting Dubois and Prade's
Cheng and Mcinnis's

Fuzzy Bonissone
arithmetic

AHP Eigenvector Saaty's

Weight Laarhoven
assessing and Pedrycz's
arithmetic Buckley

Crisp/fuzzy Conjunction/ Possibility/ Dubois and Prade's

Disjunction necessity

MAUF Human Efstathiou

intuition and Rajkovic's

General MADM Fuzzy ranking Negi's

and fuzzy
arithmatic

Outranking Fuzzy Roy

outranking Siskos et a1.
relation Brans et a1.
Takeda
n < 350 All fuzzy Maximin max and min Bellman and Zadeh's
m, any (singleton) operators Yager's
number
crisp/fuzzy General MADM Linguistic -> Chen and Hwang
fuzzy set ->
crisp score
BIBLIOGRAPHY

Books, Monographs and Conference Proceedings:

[BMl] Bezdek, J.C., Analysis of Fuzzy I'!formation (Vol.III): Applications in Engineering and
Sciences (CRC Press, Boca Raton, 1987).

[BM2] Bouchon B., L. Saitta and R.R. Yager (eds.), Uncertainty and Intelligent Systems
(Springer-Verlag, Heidelberg, 1988).

[BM3] Chen, S.J. and C.L. Hwang, Fuzzy Multiple Attribute Decision Making: Methods and
Applications (Springer-Verlag, Heidelberg, 1991).

[BM4] Chvatal, V., Linear Programming (W.H. Freeman and Company, New York, 1983).

[BM5] Cohen, L.J., An Introduction to the Philosophy of Induction and Probability

(Clarendon, Oxford, 1989).

[BM6] Di Nola, A., S. Sessa, W. Pedrycz and E. Sanchez, Fuzzy Relation Equations and Their
Applications to Knowledge Engineering (Kluwer Academic, Dordrecht, 1989).

[BM7] Dubois, D. and H. Prade, Fuzzy Sets and Systems -- Theory and Application (Academic,
New York, 1980).

[BM8] Dubois, D. and H. Prade, Possibility Theory: An Approach to Computerized Processing

of Uncertainty (plenum, New York, 1988).

[BM9] Elton, E.J. and M.J. Gruber, Modem Portfolio Theory and Investment Analysis, Third
Edition (Wiley, New York, 1987).

[BMlO] Evans, G.W., W. Karwowski and M.R. Wilhelm (eds.), Applications of Fuzzy Set
Methodologies in Industrial Engineering (Elsevier Science, Amsterdam, 1989).

[BM11] Gal, T., Postoptimal Analysis, Parametric Programming and Related Topics (McGraw-
Hill, New York, 1981).

[BMI2] Gigerenzer, G., Z. Swijtink, T. Porter, L. Daston, J. Beatty and L. Kruger, The Empire
of Chance (Cambridge University, Cambridge, 1989).

[BM13] Gould, P.J., G.D. Eppen and C. Schmidt, Quantitative Concepts for Management:
Decision Making without Algorithm, Third Edition (prentice-Hall, Englewood Cliffs,
1988).

[BMI4] Gupta, M.M. and E. Sanchez (eds.), Fuzzy Information and Decision Process (North-
Holland, Amsterdam, 1982).
 280

[BM15] Gupta, M.M. and E. Sanchez (eds.), Approximate Reasoning in Decision Analysis
(North-Holland, Amsterdam, 1982).

[BM16] Gupta, M.M. and T. Yamakawa (eds.), Fuzzy Logic in Knowledge-Based Systems,
Decision and Control (North-Holland, Amsterdam, 1988).

[BM17] Hall, A.D. m, Metasystems Methodology: A New Synthesis and Unification (pergamon,
Oxford, 1989).

[BM18] Hollnagel, E., G. Mancini and D.D. Woods (eds.), Intelligent Decision Support in
Process Environments (Springer-Verlag, Heidelberg, 1985).

[BM19] Hwang, C.L. and A.S.M. Masud, Multiple Objective Decision Making: Methods and
Applications (Springer-Verlag, Heidelberg, 1979).

[BM20] Hwang, C.L. and K. Yoon, Multiple Attribute Decision Making: Methods and
Applications (Springer-Verlag, Heidelberg, 1981).

[BM21] Hwang, C.L. and Lin M.-J., Group Decision Making under Multiple Criteria: Methods
and Applications (Springer-Verlag, Heidelberg, 1987).

[BM22] Janko, W.H., M. Roubens and H.-J. Zimmermann (eds.), Progressive in Fuzzy Sets and
Systems (Kluwer Academic, Dordrecht, 1989).

[BM23] Jones, A., A. Kaufmann and H.-J. Zimmermann (eds.), Fuzzy Sets Theory and
Applications (D. Reidel, Dordrecht, 1985).

[BM24] Kacprzyk, J., Multistage Decision-Making under Fuzziness: Theory and Applications
(Verlag TUV Rheinland, New York, 1981).

[BM25] Kacprzyk, J. and R.R. Yager (eds.), Management Decision Support Systems Using Fuzzy
Sets and Possibility Theory (Verlag TUV Rheinland, New York, 1985).

[BM26] Kacprzyk, J. and S.A. Orlovski (eds.), Optimization Models Using Fuzzy Sets and
Possibility Theory (D. Reidel, Dordrecht, 1987).

[BM27] Kacprzyk, J. and M. Fedrizzi (eds.), Combining Fuzzy Imprecision with Probabilistic
Uncertainty in Decision Making (Springer-Verlag, Heidelberg, 1988).

[BM28] Kacprzyk, J. and M. Roubens (eds.), Non-Conventional Preference Relations inDecision

Making (Springer-Verlag, Heidelberg, 1988).

[BM29] Kandel, A., Fuzzy Mathematical Techniques with Applications (Addison-Wesley,

Reading, 1987).

[BM30] Karwowski, W. and A. Mital (eds.), Applications ofFuzzy Set Theory in Human Factors
(Elsevier Science, Amsterdam, 1986).
 281

[BM31] Kaufmann, A. and M.M. Gupta, Introduction to Fuzzy Arithmetic: Theory and
Applications (Van Nostrand Reinhold, New York, 1985).

[BM32] Kaufmann, A. and M.M. Gupta, Fuzzy Mathematical Models in Engineering and
Management Science (North-Holland, Ansterdam, 1988).

[BM33] Kickert, W.J.M., Fuzzy Theories on Decision-Making (Martinus Nijhoff, Leiden, 1978).

[BM34] Klir, G.J. (ed.), Applied General Systems Research (Wiley, New York, 1978).

[BM35] Klir, G.J. and T.A. Folger, Fuzzy Sets Uncenainty, and Information (prentice-Hall,
Englewood Cliff, 1988).

[BM36] Kruse, R. and K.D. Meyer, Statistics with Vague Data (D. Reidel, Dordrecht, 1987).

[BM37] Lasker, G.E. (ed.), Applied Systems and Cybernetics Vol. VI (pergamon, New York,
1981).

[BM38] Leung, Y., Spatial Analysis and Planning under Imprecision (North-Holland,
Amsterdam, 1988).

[BM39] Mitra, G., H.J. Greenberg, F.A. Lootsma, M.J. RijckaertandH.-J. Zimmermann (eds.),
Mathematical Models for Decision Suppon (Springer-Verlag, Heidelberg, 1988).

[BM40] Negoita, C.V., Management Applications of System Theory (Birkhauser Verlag,

Bucharest, 1979).

[BM41] Negoita, C.V. and D. Ralescu, Simulation, Knowledge-Based, Computing, and Fuzzy
Statistics (Van Nostrand Reinhold, New York, 1987).

[BM42] Nemhauser, G.L., A.H.G. Rinnooy Kan and M.J. Todd (eds.), Handbooks in Operations
Research and Management Science (Vol. I): Optimization (North-Holland, Amsterdam,
1989).

[BM43] Sanchez, E.(ed.), Fuzzy Information, Knowledge Representation and Decision Analysis,
A Proceedings of the IFAC Symposium, Marseille, France (Oxford, New York, 1983).

[BM44] Schmucker, K.J., Fuzzy Sets, Natural Language, Computations, and Risk Analysis
(Computer Science, Rockville, 1984).

[BM45] Seo, F. and M. Sakawa, Multiple Criteria Decision Analysis in Regional Planning -
Concepts, Methods and Applications (D. Reidel, Dordrecht, 1988).

[BM46] Smithson, M., Fuzzy Set Analysis for Behavioral and Social Sciences (Springer-Verlag,
Heidelberg, 1987).

[BM47] Smithson, M., Ignorance and Uncenainty (Springer-Verlag, Heidelberg, 1989).

 282

[BM48] Stancu-Minasian, I.M., Stochastic Programming with Multiple Objective Functions (D.
Reidel, Dordrecht, 1984).

[BM49] Trappl R., G.J. Klir and L. Ricciardi (eds.), Progress in Cybernetics and Systems
Research: General System Methodology, Fuzzy Mathematics, Fuzzy Systems,
Biocybernetics and Theoretical Neurobiology, vol. III (Hemisphere, Washington, 1978).

[BM50] Trappl R., G.J. Klir and F.R. Pichler (eds.), Progress in Cybernetics and Systems
Research: General System Methodology, Mathematical System Theory, and Fuzzy Sets
vol. VIII (Hemisphere, Washington, 1982).

[BM51] Trappl, R. (ed.), Cybernetics and Systems Research (North-Holland, Amsterdam, 1982).

[BM52] Trappl, R. (ed.), Cybernetics and Systems Research (North-Holland, Amsterdam, 1984).

[BM53] Trappl, R. (ed.), Cybernetics and Systems (D. Reidel, Dordrecht, 1986).

[BM54] Wagner, H.M., Principles of Operations Research with Applications to Managerial

Decisions (Prentice-Hall, Englewood Cliffs, 1969).

[BM55] Wang, P.P. (ed.), Advance in Fuzzy Sets, Possibility Theory and Applications (plenum,
New York, 1983).

[BM56] Yager, R.R. (ed.), Fuzzy Set and Possibility Theory: Recent Developments (pergamon,
New York, 1982).

[BM57] Yager, R.R., S. Ovchinnikov, R.M. Tong and H.T. Nguyen (eds.), Fuzzy Sets and
Applications: Selected Papers by L.A. Zadeh (Wiley, New York, 1987).

[BM58] Yu, P.L., Multiple Criteria Decision Making: Concepts, Techniques and Extensions
(plenum, New York, 1985).

[BM59] Yu, P .L., Forming Winning Strategies: An Integrated Theory of Habitual Domains
(Springer-Verlag, Heidelberg, 1990).

[BM60] Zadeh, L.A., K.S. Fu, K. Tanaka and M. Shirmura (eds.), Fuzzy Sets and Their
Applications to Cognitive and Decision Processes (Academic, New York, 1975).

[BM61] Zetenyi, T. (ed.), Fuzzy Sets in Psychology (North-Holland, Amsterdam, 1988).

[BM62] Zimmermann, H.-J., L.A. Zadeh and B.R. Gaines (eds.), Fuzzy Sets and Decision
Analysis (Elsevier Science, Amsterdam, 1984).

[BM63] Zimmermann, H.-J., Fuzzy Set Theory and Its Applications (Kluwer-Nijhoff, Hinghum,
1985).

[BM64] Zimmermann, H.-J., Fuzzy Sets, Decision Making and Expert Systems (Kluwer
Academic, Norwell, 1987).
 283

Journal Articles, Technical Reports, and Theses:

[1] Ackoff, R.L., The future of operational research is past, Journal of Operational Research
Society 30 (1979) 93-104.

[2] Ackoff, R.L., Resurrecting the future of operational research, Journal of Operational
Research Society 30 (1979) 189-199.

[3] Ammar, S., Determining the 'best' decision in the presence of imprecise information, Fuzzy
Sets and Systems 29 (1989) 293-302.

[4] Asai, K., H. Tanaka and T. Okuda, On discrimination of fuzzy states in probability space,
Kybernetes 6 (1977) 185-192.

[5] Aubin, J.P., Cooperative fuzzy games: the static and dynamic points of view, in [BM62]
(1984) 407-428.

[6] Baldwin, J.F. and N.C.F. Guild, Comparison of fuzzy sets on the same decision space, Fuzzy
Sets and Systems 2 (1979) 213-231.

[7] Bandemer, H., From fuzzy data to functional relationships, Mathematical Modelling 9 (1987)
419-426.

[8] Baptistella, L.F.B. and A. Ollero, Fuzzy methodologies for interactive multicriteria
optimization, IEEE Transactions on Systems, Man, and Cybernetics 10 (1980) 355-365.

[9] Beale, E.M.L., On minimizing a convex function subject to linear inequalities, Journal of
Royal Statistist Society B17 (1955) 173-184.

[10] Bedrosian, S.D., A role for fuzzy concepts in interactive decision-making, in [BM25] (1985)
38-47.

[11] Behringer, F. A., A simplex method based algorithm for the lexicographically extended linear
maxmin problem, European Journal of Operational Research 7 (1981) 274-283.

[12] Bellman, R.E. and L.A. Zadeh, Decision-making in a fuzzy environment, Management
Science 17 (1970) BI41-BI64.

[13] Bellman, R. and M. Giertz, On the analytic formalism of the theory of the fuzzy sets,
Information Sciences 5 (1973) 149-156.

[14] Bensoussan, A. and J.M. Proth, Planning horizons for production planning models in the
case of concave costs, in [BM43] (1983) 393-398.

[15] Bezdek, J.C. and R.J. Hathaway, Clustering with relational c-means partitions from pairwise
distance data, Mathematical Modelling 9 (1987) 435-439.
 284

[16] Booth, G.G. and G.H. Dash, Ir., Alternative programming structures for bank portfolios,
Journal of Banking and Finance 3 (1979) 67-82.

[17] Booth, G.G. and P.E. Koveos, A programming model for bank hedging decisions, The
Journal of Financial Research IX (1986) 271-279.

[18] Booth, G.G. and W. Bessler, Goal programming models for managing interest-rate risk,
OMEGA International Journal of Management Science 17 (1989) 81-89.

[19] Booth, G.G., W. Bessler and W.G. Foote, Managing interest rate risk in banking
institutions, European Journal of Operational Research 41 (1989) 302-313.

[20] Bortoian, G. and R. Degani, A review of some methods for ranking fuzzy subsets, Fuzzy
Sets and Systems 15 (1985) 17-19.

[21] Bouchon, B., Preferences deduced from fuzzy questions, in [BM26] (1987) 110-120.

[22] Buckley, I.I., Possibility and necessity in optimization, Fuzzy Sets and Systems 25 (1988)
1-13.

[23] Buckley, I.I., Generalized and extended fuzzy sets with application, Fuzzy Sets and Systems
25 (1988) 159-174.

[24] Buckley, I.I., Possibilistic linear programming with triangular fuzzy numbers, Fuzzy Sets
and Systems 26 (1988) 135-138.

[25] Buckley, I.I., Solving possibilistic linear programming problems, Fuzzy Sets and Systems
31 (1989) 329-341.

[26] Buckley, I.I., Fuzzy PERT, in [BMlO] (1989) 103-114.

[27] Buckley, I.I., Stochastic versus possibilistic programming, Fuzzy Sets and Systems 34 (1990)
173-177.

[28] Buckley, I.I. and Y. Qu, Solving linear and quadratic fuzzy equations, Fuzzy Sets and
Systems 38 (1990) 43-59.

[29] Campos, L., Fuzzy linear programming models to solve fuzzy matrix games, Fuzzy Sets and
Systems 32 (1989) 271-289.

[30] Campos, L. and I.L. Verdegay, Linear programming problems and ranking of fuzzy
numbers, Fuzzy Sets and Systems 32 (1989) 1-11. .

[31] Carlsson, C. and P. Korhonen, A parametric approach to fuzzy linear programming, Fuzzy
Sets and Systems 20 (1986) 17-30.

[32] Chanas, S. and I. Kamburowski, The use of fuzzy variables in PERT, Fuzzy Sets and
Systems 5 (1981) 11-19.
 285

[33] Chanas, S. and W. Kolodziejczyk, Maximum flow in a network with fuzzy arc capacities,
Fuzzy Sets and Systems 8 (1982) 165-173.

[34] Chanas, S., The use of parametric programming in FLP, Fuzzy Sets and Systems 11 (1983)
243-251.

[35] Chanas, S. and M. Kulej, A fuzzy linear programming problem with equality constraints,
Control and Cybernetics 13 (1984) 195-201.

[36] Chanas, S., Fuzzy optimization in networks, in [BM26] (1987) 303-327.

[37] Chanas, S., Fuzzy programming in multiobjective linear programming - a parametric

approach, Fuzzy Sets and Systems 29 (1989) 303-313.

[38] Chang, T.C. and K.C. Crandall, An algorithm for solving expected possibility and its
application in construction resource allocation, Fuzzy Sets and Systems 34 (1990) 157-171.

[39] Charnes, A., W. W. Cooper and G .H. Symonds, Cost horizons and certainty equivalents: an
approach to stochastic programming of heating oil, Management Science 4 (1958) 235-263.

[40] Charnes, A. and W.W. Cooper, Chance-constrainted programming, Management Science

5 (1959) 73-79.

[41] Chen, G.Q., S.C. Lee and E.S.H Yu, Application of fuzzy set theory to economics, in
[BM55] (1983) 277-305.

[42] Chen, S.H., Ranking fuzzy numbers with maximizing set and minimizing set, Fuzzy Sets
and Systems 17 (1985) 113-129.

[43] Chen, Y.C., Fuzzy linear programming and fuzzy multiple objective linear programming,
Master's Thesis (1984), Dept. of I.E., Kansas State University.

[44] Chien, T.Q., Medium distances of probability fuzzy-points and an application to linear
programming, Kybernetika 25 (1989) 494-504.

[45] Cooley, J.W., and J.D. Hicks, Jr., A fuzzy set approach to aggregating internal control
judgments, Management Science 29 (1983) 317-334.

[46] Cooley, J.W. and J.D. Hicks, Jr., The use of fuzzy sets in the accounting spectrum: the
evaluation of internal control systems, in [BM25] (1985) 334-342.

[47] Dantzig, G.B., Linear programming under uncertainty, Management Science 1 (1955) 197-
204.

[48] Darby-Dowman, K, C. Lucas, G. Mitra and J. Yadegar, Linear, integer, separable and fuzzy
programming problems: a unified approach towards reformulation, Journal of Operational
Research Society 29 (1988) 161-171.
 286

[49] Darzentas, J., A discrete location model with fuzzy accessibility measures, Fuzzy Sets and
Systems 23 (1987) 149-154.

[50] Darzentas, J., On fuzzy location models, in [BM26] (1987) 328-341.

[51] de Campos Ibanez, L.M., A sUbjective approach for ranking fuzzy numbers, Fuzzy Sets and
Systems 29 (1989) 145-153.

[52] Delgado, M., J.L. Verdegay, and M.A. Vila, Imprecise costs in mathematical programming
problems, Control and Cybernetics 16 (1987) 113-121.

[53] Delgado, M., J.L. Verdegay, and M.A. Vila, Fuzzy transportation problems: a general
analysis, in [BM26] (1987) 342-358.

[54] Delgado, M., A procedure for ranking fuzzy number using fuzzy relations, Fuzzy Sets and
Systems 26 (1988) 49-62.

[55] Delgado, M., J.L. Verdegay, and M.A. Vila, A general model for fuzzy linear
programming, Fuzzy Sets and Systems 29 (1989) 21-29.

[56] Delgado, M., J.L. Verdegay, and M.A. Vila, Relating different approaches to solve linear
programming problems with imprecise costs, Fuzzy Sets and Systems 37 (1990) 33-42.

[57] Delgado, M., J.L. Verdegay, M. A. Vila, On valuation and optimization problems in fuzzy
graphs: a general approach and some particular cases, ORSA Journal on Computing 2 (1990)
74-83.

[58] Dias, D.P., Jr., The R&D project selection problem with fuzzy coefficients, Fuzzy Sets and
Systems 26 (1988) 299-316.

[59] Dishkant, H.C., About membership-function estimation, Fuzzy Sets and Systems 5 (1981)
141-148.

[60] Dombi, J., A general class of fuzzy operators, the Demorgan class of fuzzy operators, and
fuzziness measures induced by fuzzy operators, Fuzzy Sets and Systems 8 (1982) 149-163.

[61] Dombi, J., Membership function as an evaluation, Fuzzy Sets and Systems 35 (1990) 1-21.

[62] Dowlatshahi, F. and L.J. Kohout, Intentional possibilistic approach to dealing with
incompleteness, vagueness and uncertainty, Fuzzy Sets and Systems 25 (1988) 277-295.

[63] Dubois, D., A fuzzy, heuristic, interactive approach to the optimal network problem, in
[BM55] (1983) 253-275.

[64] Dubois, D. and H. Prade, System of linear fuzzy constraints, Fuzzy Sets and Systems 3
(1980) 37-48.

[65] Dubois, D. and H. Prade, The use of fuzzy numbers in decision making, in [BMI5] (1982)
309-321.
 287

[66] Dubois, D. and H. Prade, Ranking fuzzy numbers in the setting of possibility theory,
Information Sciences 30 (1983) 183-224.

[67] Dubois, D. and H. Prade, Criteria aggregation and ranking of alternatives in the framework
of fuzzy set theory, in [BM62] (1984) 209-240.

[68] Dubois, D. and H. Prade, The advantages of fuzzy approach in OR/MS demonstrated on two
examples of resources allocation problems, in [BM52] (1984) 491-497.

[69] Dubois, D. and H. Prade, Fuzzy-theoretic difference and inclusions and their use in analysis
of fuzzy equations, Control and Cybernetics 13 (1984) 129-145.

[70] Dubois, D. and H. Prade, A review of fuzzy set aggregation connectives, Information
Sciences 36 (1985) 85-121.

[71] Dubois, D., An application of fuzzy arithmetic to the optimization of industrial machining
processes, Mathematical Modelling 9 (1987) 461-475.

[72] Dubois, D., Linear programming with fuzzy data, in [BMl] (1987) 241-263.

[73] Dubois, D. and H. Prade, Properties of measurer of information in evidence and possibility
theories, Fuzzy Sets and Systems 24 (1987) 161-182.

[74] Dubois, D. and H. Prade, Necessity measures and the resolution principle, IEEE
Transactions on Systems, Man, and Cybernetics 17 (1987) 474-478.

[75] Dubois, D. and H. Prade, Incomplete conjunctive information, Computers and Mathematics
with Applications 15 (1988) 797-810.

[76] Dubois, D., Fuzzy knowledge in an artificial intelligence system for job shop scheduling,
in [BMI0] (1989) 73-89.

[77] Dumitru, V. and F. Luban, Membership functions, some mathematical programming models
and production scheduling, Fuzzy Sets and Systems 8 (1982) 19-33.

[78] Dumitru, V. and F. Luban, On some optimization problems under uncertainty, Fuzzy Sets
and Systems 18 (1986) 257-272.

[79] Dyson, R.G., Maximin programming, Fuzzy linear programming and multicriteria decision
making, Journal ofOperationa/ Research Society 31 (1981) 263-267.

[80] Esogbue, A.O. and Z.M. Ahipo, Fuzzy sets and water resources planning, in [BM56] (1982)
450-465.

[81] Evans, G.W., M.R. Wilhelm and W. Karwowski, A layout design heuristic employing the
theory of fuzzy sets, International Journal of Production Research 25 (1987) 1431-1450.

[82] Evans, G.W., W. Karwowski and M.R. Wilhelm, An introduction to fuzzy set
methodologies for industrial and systems engineering, in [BM 10] (1989) 3-11.
 288

[83] Fabian, Cs. and M. Stoica, Fuzzy integer programming, in [BM62] (1984) 123-131.

[84] Fabian, Cs., Gh. Ciobanu and M. Stoica, Interactive polyoptimization for fuzzy mathematical
programming, in [BM26] (1987) 272-291.

[85] Falk, J.E., Exact solutions of inexact linear programs, Operations Research 22 (1974) 783-
787.

[86] Fedrizzi, M., J. Kacprzyk and S. Zadrozny, An interactive multi-user decision support
system for consensus reaching processes using fuzzy logic with linguistic quantifiers,
Decision Support Systems 4 (1988) 313-327.

[87] Feng, Y.J., A method using fuzzy mathematics to solve the vector maximum problem, Fuzzy
Sets and Systems 9 (1983) 129-136.

[88] Feng, Y.J., Fuzzy programming - a new model of optimization, in [BM26] (1987) 216-225.

[89] Feng, Y.-J., M.-Z. Lin and C. Jiang, Application of fuzzy decision-making in earthquake
research, Fuzzy Sets and Systems 36 (1990) 15-26.

[90] Fiksel, J., Applications of fuzzy sets and possibility theory to systems management, in
[BM56] (1982) 494-503.

[91] Flachs, J. and M.A. Pollatschek, Further results on fuzzy mathematical programming,
Information and Control 38 (1978) 241-257.

[92] Freeling, A.N.S., Fuzzy sets and decision analysis, IEEE Transactions on Systems, Man,
and Cybernetics 10 (1980) 341-354.

[93] Freeling, A.N.S., Possibilities versus fuzzy probabilities - two alternative decision aids, in
[BM62] (1984) 67-81.

[94] French, S., Fuzzy decision analysis: some criticisms, in [BM62] (1984) 29-44.

[95] Friedman, M., M. Schneider and A. Kandel, The use of weighted fuzzy expected value
(WFEV) in fuzzy expert systems, Fuzzy Sets and Systems 31 (1989) 37-45.

[96] Fu, K.S., M. Ishlzukaand J.T.P. Yao, Applications of fuzzy sets in earthquake engineering,
in [BM56] (1982) 504-523.

[97] Fuller, R., On a special type of fuzzy linear programming, Colloquia Mathematica Societatis
Janos Bolyai 50 (1986) 511-519.

[98] Fuller, R., On stability in fuzzy linear programming problems - a short communication,
Fuzzy Sets and Systems 30 (1989) 339-344.

[99] Fuller, R., On stability in possibilistic linear equality systems with Lipschitzian fuzzy
numbers, Fuzzy Sets and Systems 34 (1990) 347-353.
 289

[100] Furuta, H., K. Furukawa, and N. Shiraishi, Application of fuzzy optimality criterion to
earthquake-resistant design, in [BMl] (1987) 17-27.

[101] Gaines, B.R. and L.J. Kohout, The fuzzy decade: a bibliography of fuzzy systems and
closely related topics, International Journal of Man-Machine Studies 9 (1977) 1-68.

[102] Gaines, B.R., Precise past - fuzzy future, International Journal of Man-Machine Studies
19 (1983) 117-134.

[103] Gaines, B.R., Fundamentals of decision: probabilistic, possibilistic and other forms of
uncertainty in decision analysis, in [BM62] (1984) 47-65.

[104] Gaines, B.R., New paradigms in systems engineering: from hard to soft approaches, in
[BM26] (1987) 3-12.

[105] Gambarelli, G., J. Holubiec and J. Kacprzyk, Modeling and optimization of international
economic cooperation via fuzzy mathematical programming and cooperative games, Control
and Cybernetics 17 (1988) 325-335.

[106] Giles, R., The concept of grade of membership, Fuzzy Sets and Systems 25 (1988) 297-323.

[107] Ghassan, K., New utilization of fuzzy optimization method, in [BMI5] (1982) 239-246.

[108] Goguen, J.A., L-fuzzy sets, Journal of Mathematical Analysis and Its Applications 18
(1967) 145-174.

[109] Grobelny, J., The fuzzy approach to facilities layout problems, Fuzzy Sets and Systems 23
(1987) 175-190.

[110] Grobelny, J., On one possible 'fuzzy' approach to facilities layout problems, International
Journal of Production Research 25 (1987) 1123-1141.

[111] Grobelny, J., The linguistic pattern method for a workstation layout analysis, International
Journal of Production Research 26 (1988) 1779-1798.

[112] Hamacher, H., H. Leberling and H.-J. Zimmermann, Sensitivity analysis in fuzzy linear
programming, Fuzzy Sets and Systems 1 (1978) 269-281.

[113] Hammerbacher, I.M. and R.R. Yager, Predicting television revenues using fuzzy subsets,
in [BM62] (1984) 469-477.

[114] Hersh, H.M. and A. Caramazzn, A fuzzy set approach to modifiers and vagueness in
natural language, Journal of Experimental Psychology: General 105 (1976) 254-276.

[115] Hessel, M. and M. Zeleny, Optimal system design: towards new interpretation of shadow
prices in linear programming, Computers and Operations Research 14 (1987) 265-27l.

[116] Hidding, G.J., DSS conceptualizes in O.R.: should you apply AI?, in [BM39] (1988) 665-
694.
 290

[117] Hisdal, E., Are grades of membership probabilities?, Fuzzy Sets and Systems 25 (1988)
325-348.

[118] Hisdal, E., The philosophical issues raised by fuzzy set theory, Fuzzy Sets and Systems 25
(1988) 349-356.

[119] Hisdal, E., A flexible classification structure, in [BMI6] (1988) 11-67.

[120] Inuiguchi, M., H. Ichihashi and Y. Kume, A solution algorithm for fuzzy linear
programming with piecewise linear membership functions, Fuzzy Sets and Systems 34
(1990) 15-31.

[121] Inuiguchi, M. and H. Ichihashi, Relative modalities and their use in possibilistic linear
programming, Fuzzy Sets and Systems 35 (1990) 303-323.

[122] Kabbara, G., New utilization of fuzzy optimization method, in [BMI5] (1982) 239-245.

[123] Kacprzyk, I. and A. Straszak, A fuzzy approach to the stability of integrated regional
development, in [BM37] (1981) 2997-3004.

[124] Kacprzyk, I., Multistage decision process in a fuzzy environment: a survey, in [BMI4]
(1982) 251-263.

[125] Kacprzyk, I. and A. Straszak, Determination of stable regional development trajectories via
fuzzy decision-making model, in [BM56] (1982) 531-541.

[126] Kacprzyk, I. and P. Staniewski, Long-term inventory policy-making through fuzzy decision-
making models, Fuzzy Sets and Systems 8 (1982) 117-132.

[127] Kacprzyk, I., Towards human-consistent multistage decision making and control models
using fuzzy sets and fuzzy logic, Fuzzy Sets and Systems 18 (1986) 299-314.

[128] Kacprzyk, I. and S.A. Orlovski, Fuzzy optimization and mathematical programming: a
brief introduction and survey, in [BM26] (1987) 50-72.

[129] Karwowski, W., T. Marek, C. Noworol and K. Ostaszewski, Fuzzy modeling of risk
factors for industrial accident prevention: some empirical results, in [BMlO] (1989) 141-
153.

[130] Kaufmann, A., Fuzzy zero-base budgeting, in [BM62] (1984) 479-487.

[131] Kaufmann, A., Hybrid data - various associations between fuzzy subsets and random
variables, in [BM23] (1985) 171-211.

[132] Kaufmann, A., Fuzzy subsets applications in O.R. and management, in [BM29] (1985) 257-
300.

[133] Kim, I.B., A. Baartmans and N.S. Sahadin, Determinant theory for fuzzy matrices, Fuzzy
Sets and Systems 29 (1989) 349-356.
 291

[134] Kim, K. and F.W. Roush, Fuzzy flow on networks, Fuzzy Sets and Systems 8 (1982) 35-38.

[135] Kim, K. and K.S. Park, Ranking fuzzy numbers with index of optimism, Fuzzy Sets and
Systems 35 (1990) 143-150.

[136] Kuzmin, V .B., A parametric approach to description of linguistic values of variables and
hedges, Fuzzy Sets and Systems 6 (1981) 27-41.

[137] Lai, Y.]. and C.L. Hwang, Interactive fuzzy linear programming, Fuzzy Sets and Systems
45 (1992) 169-183.

[138] Lai, Y.]. and C.L. Hwang, A new approach to some possibilistic linear programming
problem, Fuzzy Sets and Systems 49 (1992).

[139] Lai, Y.]., Fuzzy mathematical programming and fuzzy multiple objective decision making,
Ph.D. Dissertation (1991), Dept. of I.E., Kansas State University.

[140] Lai, Y.]. and C.L. Hwang, Possibilistic linear programming for managing interest rate risk,
Fuzzy Sets and Systems (forthcoming).

[141] Lai, Y.]. and C.L. Hwang, IFLP-I1: a decision support system, Fuzzy Sets and Systems
(forthcoming) .

[142] Lai, Y.]. and C.L. Hwang, Stochastic possibilistic linear programming (submitted).

[143] Lai, Y.]., T.Y. Liu and C.L. Hwang, TOPSIS for MODM, European Journal of
Operational Research (forthcoming).

[144] Lee, E.S. and R.-]. Li, Comparison of fuzzy numbers based on probability measure of
fuzzy events, Computers and Mathematics with Applications 15 (1988) 887-896.

[145] Lee, Y.Y., Fuzzy Sets Theory Approach to Aggregate Production Planning and Inventory
Control, Ph.D. Dissertation (1990), Dept. of I.E., Kansas State University.

[146] Lee, Y.Y., B.A. Kramer and C.L. Hwang, Part-period balancing with uncertainty: a fuzzy
sets theory approach, International Journal of Production Research 28 (1990) 1771-1778.

[147] Lehtimaki, A.K., An approach for solving decision problems of master scheduling by
utilizing theory of fuzzy sets, International Journal ofProduction Research 25 (1987) 1781-
1793.

[148] Leung, Y., Interregional equilibrium and fuzzy linear programming - 1, Environment and
Planning A 20 (1988) 25-40.

[149] Leung, Y., Interregional equilibrium and fuzzy linear programming - 2, Environment and
Planning A 20 (1988) 219-230.

[150] Li Calzi, M., Towards a general setting for the fuzzy mathematics of finance, Fuzzy Sets
and Systems 35 (1990) 265-280.
 292

[151] Li, R.I., Multiple Objective Decision Making in A Fuzzy Environment (1990), Ph.D.
Dissertation, Dept. of I.E., Kansas State University.

[152] Liu, X., M. Wang and P. Wang, A fuzzy mathematical method for earthquake disaster
prediction, in [BMl] (1987) 55-65.

[153] Llena, I., On fuzzy linear programming, European Journal of Operational Research 22
(1985) 216-213.

[154] Lokwick, W .A., Analysis of structure in fuzzy linear programs, Fuzzy Sets and Systems 38
(1990) 15-26.

[155] Luhandjula, M.K., Compensatory operators in fuzzy linear programming with multiple
objectives, Fuzzy Sets and Systems 8 (1982) 245-252.

[156] Luhandjula, M.K., Linear programming under randomness and fuzziness, Fuzzy Sets and
Systems 10 (1983) 45-55.

[157] Luhandjula, M.K., On possibilistic linear programming, Fuzzy Sets and Systems 18 (1986)
15-30.

[158] Luhandjula, M.K., Satisfying solution for a possibilistic linear program, Information
Sciences 40 (1986) 247-265.

[159] Luhandjula, M.K., Linear programming with a possibilistic objective function, European
Journal of Operational Research 31 (1987) 110-117.

[160] Luhandjula, M.K., Multiple objective programming problems with possibilistic coefficients,
Fuzzy Sets and Systems 21 (1987) 135-145.

[161] Luhandjula, M.K., Fuzzy optimization an appraisal, Fuzzy Sets and Systems 30 (1989) 257-
282.

[162] Mabuchi, S., An approach to the comparison of fuzzy subsets with a-cut dependent index,
IEEE Transactions on Systems, Man, and Cybernetics 18 (1988) 264-272.

[163] Marek, T. and C. Noworol, Fuzzy rational aggregation for mental load and fatigue
assessment, in [BMlO] (1989) 155-165.

[164] Mares, M., Network analysis of fuzzy technologies, in',[BMI0] (1989) 115-125.

[165] Martine, I.A., Some geometrical properties of trapezoidal functions for their association
with linguistic labels leading to a reduced 2-D representation, in [BM2] (1988) 47-54.

[166] Mccahon, C.S., Fuzzy Set Theory Applied to Production and Inventory Control (1987),
Ph.D. Dissertation, Dept. of I.E., Kansas State University.

[167] Mccahon, C.S. and E.S. Lee, Project network analysis with fuzzy activity times, Computers
and Mathematics with Applications 15 (1988) 829-838.
 293

[168] Mital, A. and W. Karwowski, A framework of the fuzzy linguistic approach to facilities
location problem, in [BMlO] (1989) 323-330.

[169] Mitra, G., Models for decision making: an overview of problems, tools and major issues,
in [BM39] (1989) 17-53.

[170] Mizumoto, M., Fuzzy sets and their operators - I, Information Sciences 48 (1981) 30-48.

[171] Mizumoto, M., Fuzzy sets and their operators - II, Information Sciences 50 (1981) 160-
174.

[172] Mjelde, K.M., Fuzzy resource allocation, Fuzzy Sets and Systems 19 (1986) 239-250.

[173] Nahmias, S., Fuzzy variables, Fuzzy Sets and Systems 1 (1978) 97-110.

[174] Nakamura, K., Some extensions of fuzzy linear programming, Fuzzy Sets and Systems 14
(1984) 211-229.

[175] Negi, D.S., Fuzzy Analysis and Optimization (1989), Ph.D. Dissertation, Dept. of I.E.,
Kansas State University.

[176] Negoita, C.V. and M. Sularia, On fuzzy mathematical programming and tolerances in
planning, Economic Computer and Economic Cybernetic Studies and Researches 1 (1976)
3-15.

[177] Negoita, C. V., P. Flondor and M. Sularia, On fuzzy environment in optimization problems,
Economic Computer and Economic Cybernetic Studies and Researches 2 (1977) 13-24.

[178] Negoita, C.V. and D.A. Ralescu, On fuzzy optimization, Kybernetes 6 (1977) 193-196.

[179] Negoita, C.V., The current interest in fuzzy optimization, Fuzzy Sets and Systems 6 (1981)
261-269.

[180] Negoita, C.V. and A.C. Stefanesca, On fuzzy optimization, in [BMI5] (1982) 247-250.

[181] Negoita, C.V., Structure and logic in optimization, Control and Cybernetics 13 (1984) 121-
127.

[182] Norwich, A.M. and I.B. Turksen, The construction of membership functions, in [BM56]
(1982) 61-67.

[183] Norwich, A.M. and I.B. Turksen, Meaningfulness in fuzzy set theory, in [BM56] (1982)
68-74.

[184] Oden, G.C., Integration of fuzzy logical information, Journal oJExperimental Psychology:
Human Perception and Performance 3 (1976) 565-575.

[185] OhEigeartaigh, M., A fuzzy transportation algorithm, Fuzzy Sets and Systems 8 (1982) 235-
243.
 294

[186] Okuda, T., H. Tanaka and K. Asai, A formulation of fuzzy decision problems with fuzzy
information using probability measures of fuzzy events, lriformanon and Control 38 (1978)
135-147.

[187] Okuda, T., H. Tanaka and K. Asai, Discrimination problem with fuzzy states and fuzzy
information, in [BM62] (1984) 97-106.

[188] Onisawa, T., A representation of human reliability using fuzzy concepts, lriformanon
Sciences 45 (1988) 153-173.

[189] Onisawa, T., M. Sugeno, Y. Nishiwak, H. Kawai and Y. Harima, Fuzzy measure analysis
of public attitude towards the use of nuclear energy, Fuzzy Sets and Systems 20 (1988) 259-
289.

[190] Onisawa, T., Fuzzy concepts in human reliability, in [BMI6] (1988) 317-335.

[191] Onisawa, T., An application of fuzzy concepts to modelling of reliability analysis, Fuzzy
Sets and Systems 37 (1990) 267-286.

[192] Orlovsky, S.A., On programming with fuzzy constraint sets, Kybernetes 6 (1977) 197-202.

[193] Orlovsky, S.A., Decision-making with a Fuzzy Preference Relation, Fuzzy Sets and Systems
1 (1978) 155-167.

[194] Orlovsky, S.A., On formalization of a general fuzzy mathematical problem, Fuzzy Sets and
Systems 3 (1980) 311-321.

[195] Orlovsky, S.A., Effective alternatives for multiple fuzzy preference relations, in [BM51]
(1982) 185-189.

[196] Orlovsky, S.A., Multiobjective programming problems with fuzzy parameters, Control and
Cybernetics 13 (1984) 175-183.

[197] Orlovsky, S.A., Mathematical programming problems with fuzzy parameters, in [BM25]
(1985) 136-145.

[198] Ostasiewicz, W., A new approach to fuzzy programming, Fuzzy Sets and Systems 7 (1982)
139-152.

[199] Ostermark, R., Sensitivity analysis of fuzzy linear programs: an approach to parametric
interdependence, Kybemetes 16 (1987) 113-120.

[200] Ostermark, R., Profit apportionment in concerns with mutual ownership - an application
of fuzzy inequalities, Fuzzy Sets and Systems 26 (1988) 283-297.

[201] Ostermark, R., Fuzzy linear constraints in the capital asset pricing model, Fuzzy Sets and
Systems 30 (1989) 93-102. .
 295

[202] Owsinski, J.W., S. Zadrozny and J. Kacprzyk, Analysis of water use and needs in
agriculture through a fuzzy programming model, in [BM26] (1987) 328-341.

[203] Pedrycz, W., Fuzzy models and relational equations, Mathematical Modelling 9 (1987) 427-
434.

[204] Pedrycz, W., Methodological and applicational aspects of fuzzy models for systems
engineering, in [BMlO] (1989) 13-26.

[205] Pence, J.A. and A.L. Soyster, Relationship between fuzzy set programming and semi
infinite linear programming, in [BMlO] (1989) 237-252.

[206] Ponsard, C., Fuzzy sets in economics: foundations of soft decision theory, in [BM25]
(1985) 25-37.

[207] Prade, H., Using fuzzy set theory in a scheduling problem: a case study, Fuzzy Sets and
Systems 2 (1979) 153-165.

[208] Ralescu, D., Optimization in fuzzy environment, in [BM43] (1983) 301-305.

[209] Ramik, J., Extension principle and fuzzy mathematical programming, Kybemetica 19
(1983) 516-525.

[210] Ramik, J. and J. Rimanek, Inequality relation between fuzzy numbers and its use in fuzzy
optimization, Fuzzy Sets and Systems 16 (1985) 123-138.

[211] Ramik, J., Extension principle in fuzzy optimization, Fuzzy Sets and Systems 19 (1986) 29-
35.

[212] Ramik, J. and J. Rimanek, Fuzzy parameters in optimal allocation of resources, in [BM26]
(1987) 359-374.

[213] Rapoport, A., T.S. Wallsten and J.A. Cox, Direct and indirect scaling of membership
functions of probability phrases, Mathematical Modelling 9 (1987) 397-417.

[214] Rinks, D.B., A heuristic approach to aggregate production scheduling using linguistic
variables: Methodology and application, in [BM56] (1982) 562-581.

[215] Rodder, W. and H.-J. Zimmermann, Duality in fuzzy linear programming, in Fiacco, A. V.
and K.D. Kortanek (eds.), External Methods and Systems Analysis (Springer-Verlag,
Heidelberg, 1980) 415-427.

[216] Rommelfanger, H., R. Hanuscheck and J. Wolf, Linear programming with fuzzy
objectives, Fuzzy Sets and Systems 29 (1989) 31-48.

[217] Rommelfanger, H., Interactive decision making in fuzzy linear optimization problems,
European Journal of Operational Research 41 (1989) 210-217.
 296

[218] Roubens, M. and Ph. Vincke, Fuzzy preferences in an optimization perspective, in [BM26]
(1987) 77-90.

[219] Roventa, E., On the degree of fuzziness of a fuzzy set,

Fuzzy Sets and Systems 36 (1990) 259-264.

[220] Sakawa, M. and H. Yano, An interactive fuzzy satisfying method using augmented minimax
problems and its application to environmental systems, IEEE Transactions on Systems,
Man, and Cybernetics 5 (1985) 720-729.

[221] Shafai, B. and G. Sotirov, Uniqueness of solution in FLP under parameter perturbations,
Fuzzy Sets and Systems 34 (1990) 179-186.

[222] Sher, A.P., Solution of the mathematical programming problem with a linear objective
function and fuzzy constraints, Automata (1981) 997-1002.

[223] Simon, H., Dynamic programming under uncertainty with a quadratic criterion function,
Econometrica 24 (1956) 74-81.

[224] Simon, H., Two heads are better then one: the collaboration between AI and OR, Interfaces
17 (1987) 8-15.

[225] Slowinski, R., Multiobjective network scheduling with efficient use of renewable and
nonrenewable resources, European Journal of Operational Research 7 (19~2) 265-273.

[226] Slowinski, R., An interactive method for multi objective linear programming with fuzzy
parameters and its application to water supply planning, in [BM26] (1987) 396-414.

[227] Slowinski, R., A multicriteria fuzzy linear programming method for water supply system
development planning, Fuzzy Sets and Systems 19 (1986) 217-237.

[228] Slowinski, R. and J. Teghem, Jr., Fuzzy versus stochastic approaches to multicriteria linear
programming under uncertainty, Naval Research Logistics 35 (1988) 673-695.

[229] Smets, P. and P. Magrez, The measure of the degree of truth and of the grade of
membership, Fuzzy Sets and Systems 25 (1988) 67-72.

[230] Smithson, M., Fuzzy set theory and the social sciences: the scope for application, Fuzzy
Sets and Systems 26 (1988) 1-21.

[231] Sommer, G. and M.A. Pollatschek, A fuzzy programming approach to an air pollution
regulation problem, in [BM49] (1978) 303-313.

[232] Soyster, A.L., Convex programming with set-inclusive constraints and applications to
inexact linear programming, Operations Research 21 (1973) 1154-1157.

[233] Soyster, A.L., A duality theory for convex programming with set-inclusive constraints,
Operations Research 22 (1974) 892-898.
 297

[234] Soyster, A.L., Convex programming with set inclusive constraints application to inexact
programming, Operations Research 6 (1981) 333-348.

[235] Stohr, E.A., Automated support for formulating linear programs, in [BM39] (1988) 519-
538.

[236] Tanaka, H., T. Okuda and K. Asai, On fuzzy mathematical programming, Journal of
Cybernetics 3 (1974) 37-46.

[237] Tanaka, H., T. Okuda and K. Asai, A formulation of fuzzy decision problems and its
application to an investment problem, Kybernetes 5 (1976) 25-30.

[238] Tanaka, H. and K. Asai, Fuzzy solution in fuzzy linear programming problems, IEEE
Transactions on Systems, Man, and Cybernetics 14 (1984) 325-328.

[239] Tanaka, H. and K. Asai, Fuzzy linear programming problems with fuzzy numbers, Fuzzy
Sets and Systems 13 (1984) 1-10.

[240] Tanaka, H., H. Ichihashi and K. Asai, A formulation of fuzzy linear programming
problems based on comparison of fuzzy numbers, Control and Cybernetics 13 (1984) 186-
194.

[241] Tanaka, H., H. Ichihashi and K. Asai, Fuzzy decision in linear programming problems with
trapezoid fuzzy parameters, in [BM25] (1985) 146-154.

[242] Tanaka, H., H. Ichihashi and K. Asai, A value of information in FLP problem via
sensitivity analysis, Fuzzy Sets and Systems 18 (1986) 119-129.

[243] Tanaka, H., I. Watada and K. Asai, Linear regression analysis by possibilistic models, in
[BM26] (1987) 186-199.

[244] Tanaka, H., Fuzzy data analysis by possibilistic linear models, Fuzzy Sets and Systems 24
(1987) 363-375.

[245] Thole, U., H.-I. Zimmermann and P. Zysno, On the suitability of minimum and product
operators for the intersection of fuzzy sets, Fuzzy Sets and Systems 2 (1979) 167-180.

[246] Thompson, S.D. and W.I. Davis, An integrated approach for modeling uncertainty in
aggregate production planning, IEEE Transactions on Systems, Man, and Cybernetics 20
(1990) 1()()()-1012. .

[247] Tintner, G, Stochastic linear programming with application to argricultural economics, in

H.A. Antonosiewicz (ed.), Proceedings of the Second Symposium in Linear Programming
(National Bureau Standard, Washington, 1955) 197-228.

[248] Trappey, I.F.C., C., R. Liu and T.C. Chang, Fuzzy non-linear programming: theory and
application in manufacturing, International Journal of Production Research 26 (1988) 957-
985.
 298

[249] Triantaphyllou, E. and S.H. Mann, An evaluation of the eigenvalue approach for
determining the membership values in fuzzy sets, Fuzzy Sets and Systems 35 (1990) 295-
301.

[250] Verdegay, J.L., Fuzzy mathematical programming, in [BM15] (1982) 231-236.

[251] Verdegay, J.L., A dual approach to solve the fuzzy linear programming problem, Fuzzy
Sets and Systems 14 (1984) 131-141.

[252] Verdegay, J .L., Application of fuzzy optimization in operational research, Control and
Cybernetics 13 (1984) 229-239.

[253] Wagenknecht, M. and K. Hartmann, On the existence of minimal solutions for fuzzy
equations with tolerances, Fuzzy Sets and Systems 34 (1990) 237-244.

[254] Wang, S. and Y.Z. Lu, Knowledge-based successive linear programming for linearly
constraint optimization, International Journal of Man-Machine Studies 29 (1988) 625-636.

[255] Wang, Z. and Y. Dang, Application of fuzzy decision making and fuzzy linear
programming in prediction and planning of animal husbandry system in farming region, in
[BM52] (1984) 563-566.

[256] Ward, T.L., Fuzzy discounted cash flow analysis, in [BM14] (1989) 91-102.

[257] Werners, B., Interactive multiple objective programming subject to flexible constraints,
European Journal of Operational Research 31 (1987) 342-349.

[258] Werners, B., An interactive fuzzy programming system, Fuzzy Sets and Systems 23 (1987)
131-147.

[259] Werners, B., Aggregation models in mathematical programming, in [BM39] (1988) 295-
305.

[260] Wets, R.J.B., Stochasticprogl'amming, inG.L. Nemhauser, A.H.G. RinnooyKan and M.J.
Todd (eds.), Handbooks in Operations Research and Management Science: Optimization
(North-holland, Amsterdam, 1989) 573-629.

[261] Whalen, Th., Introduction to decision making under various kinds of uncertainty, in
[BM26] (1987) 27-49.

[262] White, D.J., A bibliography on the applications of mathematical programming multiple-

objective methods, Journal of Operational Research Society 41 (1990) 669-691.

[263] Wiedey, G. and H.-J. Zimmermann, Media selection and fuzzy linear programming,
Journal of Operational Research Society 29 (1978) 1071-1084.

[264] Wierzchon, S. T., Linear programming with fuzzy sets: a general approach, Mathematical
Modelling 9 (1987) 447-459.
 299

[264a] Wierzchon, S. T., Randomness and fuzziness in a linear programming problem, in [BM27]
(1988) 227-239.

[265] Wu, J. -Y., V. Van Brunt, W. -R. Zhang and J. C. Bezdek, Tower packing evaluation using
linguistic variables, Computers and Mathematics with Applications 15 (1988) 863-869.

[266] Yager, R.R., Mathematical programming with fuzzy constraints and a preference on the
objective, Kybernetes 8 (1979) 285-292.

[267] Yager, R.R., A measurement informational discussion of fuzzy union and intersection,
International Journal of Man-Machine Studies 11 (1979) 189-200.

[268] Yager, R.R., Possibilistic decision making, IEEE Transactions on Systems, Man, and
Cybernetics 9 (1979) 388-392.

[269] Yager, R.R., A foundation for a theory of possibility, Journal of Cybernetics 10 (1980)
177-204.

[270] Yager, R.R., Quantifiers in the formulation of multiple objective decision functions,
Information Sciences 31 (1983) 107-139.

[271] Yazenin, A.V., Fuzzy and stochastic programming, Fuzzy Sets and Systems 22 (1987) 171-
180.

[272] Yu, P.L., Dissolution of fuzziness for better decision - perspective and techniques, in
[BM62] 171-207.

[273] Zadeh, L.A., From circuit theory to system theory, Proceedings of Institute of Radio
Engineering 50 (1962) 856-865.

[274] Zadeh, L.A., Fuzzy sets, Information and Control 8 (1965) 338-353.

[275] Zadeh, L.A., Probability measures of fuzzy events, Mathematical Analysis and Applications
23 (1968) 421-427.

[276] Zadeh, L.A., Fuzzy algorithms, Information and Control 12 (1968) 94-102.

[277] Zadeh, L.A., Similarity relations and fuzzy orderings, Information Sciences 3 (1971) 177-
200.

[278] Zadeh, L.A., A fuzzy-set-theoretic interpretation oflinguistic hedges, Journal of

Cybernetics 2 (1972) 4-34.

[279] Zadeh, L.A., Outline of a new approach to the analysis of complex systems and decision
processes, IEEE Transactions on Systems, Man, and Cybernetics 3 (1973) 28-44.

[280] Zadeh, L.A., The concept of a linguistic variable and its application to approximate
reasoning - part 1, Information Sciences 8 (1975) 199-249.
 300

[281] Zadeh, L.A., The concept of a linguistic variable and its application to approximate
reasoning - part 2, Information Sciences 8 (1975) 301-357.

[282] Zadeh, L.A., The concept of a linguistic variable and its application to approximate
reasoning - part 3, Information Sciences 9 (1976) 43-80.

[283] Zadeh, L.A., A fuzzy-algorithmic approach to the definition of complex or imprecise

concepts, International Journal of Man-Machine Studies 8 (1976) 249-291.

[284] Zadeh, L.A., Fuzzy sets as a basis for a theory of possibility, Fuzzy Sets and Systems 1
(1978) 3-28.

[285] Zadeh, L.A., PRUF - a meaning representation language for natural languages,
International Journal of Man-Machine Studies 10 (1978) 395-460.

[286] Zadeh, L.A., A theory of approximate reasoning, Machine Intelligence 9 (1979) 149-194.

[287] Zadeh, L.A., The role of fuzzy logic in the management of uncertainty in expert systems,
Fuzzy Sets and Systems 11 (1983) 199-227.

[288] Zadeh, L.A., A computational approach to fuzzy quantifiers in natural language, Computers
and Mathematics with Applications 9 (1983) 149-184.

[289] Zadeh, L.A., Syllogistic reasoning in fuzzy logic and its application to usuality and
reasoning with dispositions, IEEE Transactions on Systems, Man, and Cybernetics 15
(1985) 754-762.

[290] Zadeh, L.A., Test-score semantics as a basis for a computational approach to the
representation of meaning, Literary and Linguistic Computing 1 (1986) 24-35.

[291] Zadeh, L.A., Outline of a theory ofusuality based on fuzzy logic, in [BM34], 1988, 79-97.

[292] Zeleny, M., Multiobjective design of high-productivity systems, Proceedings of Joint

Automatic Control Conference, Paper Appl9-4 (ASME, New York, 1976) 297-300.

[293] Zeleny, M., Optimal system design with multiple criteria: De Novo programming approach,
Engineering Cost and Production Economics 10 (1986) 89-94.

[294] Zeleny, M., An external reconstruction approach (ERA) to linear programming, Computers
and Operations Research 13 (1986) 95-100.

[295] Zeleny, M., Parallelism, integration, autocoordination and ambiguity in human support
systems, in [BMI6] (1988) 107-121.

[296] Zimmermann, H.-J., Description and optimization of fuzzy system, International Journal
of General System 2 (1976) 209-216.

[297] Zimmermann, H.-J., Results of empirical studies in fuzzy set theory, in [BM34] (1978)
303-312.
 301

[298] Zimmermann, H.-J., Fuzzy programming and linear programming with several objective
functions, Fuzzy Sets and Systems I (1978) 45-55.

[299] Zimmermann, H.-J. and P. Zysno, Latent connectives in human decision making, Fuzzy
Sets and Systems 4 (1980) 37-51.

[300] Zimmermann, H.-J., Fuzzy mathematical programming, Computers and Operations

Research 10 (1983) 291-298.

[301] Zimmermann, H.-J., Using fuzzy sets in operational research, European Journal of
Operational Research l3 (1983) 201-216.

[302] Zimmermann, H.-J. and P. Zysno, Decisions and evaluations by hierarchical aggregation
of information, Fuzzy Sets and Systems 10 (1983) 244-260.

[303] Zimmermann, H.-J. and M.A. Pollatschek, Fuzzy 0-1 linear programming, in [BM62]
(1984) l33-145.

[304] Zimmermann, H.-J., Application of fuzzy set theory to mathematical programming,

Information Sciences 36 (1985) 29-58.

[305] Zimmermann, H.-J. and P. Zysno, Quantifying vagueness in decision models, European
Journal of Operational Research 22 (1985) 148-158.

[306] Zimmermann, H.-J., Fuzzy sets theory and mathematical programming, in [BM23] (1985)
99-114.

[307] Zimmermann, H.-J., Address by first EURO gold metal recipient, European Journal of
Operational Research 31 (1987) 173-176.

[308] Zimmermann, H.-J., Modeling and solving ill-structured problems in operations research,
in [BM1] (1987) 217-239.

[309] Zimmermann, H. -J., Interactive decision support for semi-structured mathematical

programming problems, in [BM39] (1988) 307-319.

[310] Zysno, P., The integration of concepts within judgmental and evaluative processes, in
[BM50] (1982) 509-514.
Lecture Notes in Economics
and Mathematical Systems
For information about Vols. 1-210
please contact your bookseller or Springer-Verlag

Vol. 211: P. van den Heuvel, The Stability of a Macroeconomic Vol. 233: G. Wagenhals, The World Copper Market. XI, 190
System with Quantity Constraints. VII, 169 pages. 1983. pages. 1984.
Vol. 212: R. Sato and T. Nona, Invariance Principles and the Vol. 234: B.C. Eaves, A Course in Triangulations for Solving
Structure of Technology. V, 94 pages. 1983. Equations with Deformations. III, 302 pages. 1984.
Vol. 213: Aspiration Levels in Bargaining and Economic Vol. 235: Stochastic Models in Reliability Theory Proceedings,
Decision Making. Proceedings, 1982. Edited by R. Tietz. VIII, 1984. Edited by S. Osaki and Y. Hatoyama. VII, 212 pages.
406 pages. 1983. 1984.
Vol. 214: M. Faber, H. Niemes und G. Stephan, Entropie, Um- Vol. 236: G. Gandolfo, P.c. Padoan, A Disequilibrium Model
weltschutz und Rohstoffverbrauch. IX, 181 Seiten. 1983. of Real and Financial Accumulation in an Open Economy. VI,
Vol. 215: Semi-Infinite Programming and Applications. Pro- 172 pages. 1984.
ceedings, 1981. Edited by A.V. Fiacco and K.O. Kortanek. XI, Vol. 237: Misspecification Analysis. Proceedings, 1983. Edited
322 pages. 1983. by T.K. Dijkstra. V, 129 pages. 1984.
Vol. 216: H.H. Muller, Fiscal Policies in a General Equilibrium Vol. 238: W. Domschke, A. Drexl, Location and Layout
Model with Persistent Unemployment. VI, 92 pages. 1983. Planning. IV, 134 pages. 1985.
Vol. 217: Ch. Grootaert, The Relation Between Final Demand Vol. 239: Microeconomic Models of Housing Markets. Edited
and Income Distribution. XIV, 105 pages. 1983. by K. Stahl. VII, 197 pages. 1985.
Vol 218: P. van Loon, A Dynamic Theory of the Firm: Vol. 240: Contributions to Operations Research. Proceedings,
Production, Finance and Investment. VII, 191 pages. 1983. 1984. Edited by K. Neumann and D. Pallaschke. V, 190 pages.
Vol. 219: E. van Damme, Refinements of the Nash Equilibrium 1985.
Concept. VI. lSI pages. 1983. Vol. 241: U. Wittmann, Das Konzept rationaler Preiser-
Vol. 220: M. Aoki, Notes on Economic Time Series Analysis: wartungen. XI, 310 Seiten. 1985.
System Theoretic Perspectives. IX, 249 pages. 1983. Vol. 242: Decision Making with Multiple Objectives.
Vol. 221: S. Nakamura, An Inter-Industry Translog Model of Proceedings, 1984. Edited by Y.Y. Haimes and V. Chankong.
Prices and Technical Change for the West German Economy. XI, 571 pages. 1985.
XIV, 290 pages. 1984. Vol. 243: Integer Programming and Related Areas. A Classified
Vol. 222: P. Meier, Energy Systems Analysis for Developing Bibliography 1981-1984. Edited by R. von Randow. XX, 386
Countries. VI, 344 pages. 1984. pages. 1985.

Vol. 223: W. Trockel, Market Demand. VIII, 205 pages. 1984. Vol. 244: Advances in Equilibrium Theory, Proceedings, 1984.
Edited by C.D, Aliprantis, O. Burkinshaw and N.J. Rothman.
Vol. 224: M. Kiy, Ein disaggregiertes Prognosesystem fur die 11,235 pages. 1985.
Bundesrepublik Deutschland. XVIII, 276 Seiten. 1984.
Vol. 245: J.E.M. Wilhelm, Arbitrage Theory. VII, 114 pages.
Vol. 225: T.R. von Ungern-Sternberg, Zur Analyse von Mark- 1985.
ten mit unvollstandiger Nachfragerinformaton. IX, 125 Seiten,
1984. Vol. 246: P.W, Otter, Dynamic Feature Space Modelling,
Filtering and Self-Tuning Control of Stochastic Systems. XIV,
Vol. 226: Selected Topics in Operations Research and Math- 177 pages.1985.
ematical Economics. Proceedings, 1963. Edited by G. Hammer
and D. Pallaschke IX, 478 pages. 1984. Vol. 247: Optimization and Discrete Choice in Urban Systems.
Proceedings, 1983. Edited by B.G. Hutchinson, P. Nijkamp and
Vol. 227: Risk and Capital. Proceedings, 1983. Edited by G. M. Batty VI, 371 pages. 1985.
Bamberg and K. Spremann VII, 306 pages. 1984.
Vol. 248: Pural Rationality and Interactive Decision Processes.
Vol. 228: Nonlinear Models of Fluctuating Growth. Proceedings, Proceedings, 1984. Edited by M. Grauer, M. Thompson and A.P.
1983. Edited by R.M. Goodwin, M. Kruger and A. Vercelli. Wierzbicki. VI, 354 pages. 1985.
XVII, 277 pages. 1984.
Vol. 249: Spatial Price Equilibrium: Advances in Theory,
Vol. 229: Interactive Decision Analysis. Proceedings, 1983. Ed- Computation and Application. Proceedings, 1984. Edited by P.T.
ited by M. Grauer and A.P. Wierzbicki. VIII, 269 pages. 1984. Harker. VII, 277 pages. 1985.
Vol. 230: Macro-Economic Planning with Conflicting Goals. Vol. 250: M. Roubens, Ph. Vincke, Preference Modelling. VIII,
Proceedings, 1982. Edited by M. Despontin, P. Nijkamp and J. 94 pages. 1985.
Spronk. VI, 297 pages. 1984.
Vol. 251: Input-Output Modeling. Proceedings, 1984. Edited
Vol. 231: G.F. Newell, The M/M/8 Service System with Ranked by A. Smyshlyaev. VI, 261 pages. 1985.
Servers in Heavy Traffic. XI, 126 pages. 1984.
Vol. 252: A. Birolini, On the Use of Stochastic Processes in
Vol. 232: L. Bauwens, Bayesian Full Information Analysis of Modeling Reliability Problems. VI, 105 pages. 1985.
Simultaneous Equation Models Using Integration by Monte
Carlo. VI, 114 pages. 1984.
Vol. 253: C. Withagen, Economic Theory and International Vol. 281: Ch.-L. Hwang, M.-J. Lin, Group Decision Making
Trade in Natural Exhaustible Resources. VI, 172 pages. 1985. under Multiple Criteria. XI, 400 pages. 1987.
Vol. 254: S. MUlier, Arbitrage Pricing of Contingent Claims. Vol. 282: K. Schittkowski, More Test Examples for Nonlinear
VIII, 151 pages. 1985. Programming Codes. V, 261 pages. 1987.
Vol. 255: Nondifferentiable Optimization: Motivations and Ap- Vol. 283: G. Gabisch, H.-W. Lorenz, Business Cycle Theory.
plications. Proceedings, 1984. Edited by V.F. Demyanov and VII, 229 pages. 1987.
D. Pallaschke. VI, 350 pages. 1985. Vol. 284: H. LUtkepohl, Forecasting Aggregated Vector ARMA
Vol. 256: Convexity and Duality in Optimization. Proceedings, Processes. X, 323 pages. 1987.
1984. Edited by J. Ponstein. V, 142 pages. 1985. Vol. 285: Toward Interactive and Intelligent Decision Support
Vol. 257: Dynamics of Macrosystems. Proceedings, 1984. Edited Systems. Volume I. Proceedings, 1986. Edited by Y. Sawaragi,
by J.-P. Aubin, D. Saari and K. Sigmund. VI, 280 pages. 1985. K. Inoue and H. Nakayama. XII, 445 pages. 1987.
Vol. 258: H. Funke, Eine allgemeine Theorie der Polypol- und Vol. 286: Toward Interactive and Intelligent Decision Support
Oligopolpreisbildung. III, 237 pages. 1985. Systems. Volume 2. Proceedings, 1986. Edited by Y. Sawaragi,
Vol. 259: Infinite Programming. Proceedings, 1984. Edited by K. Inoue and H. Nakayama. XII, 450 pages. 1987.
E.J. Anderson and A.B. Philpott. XIV, 244 pages. 1985. Vol. 287: Dynamical Systems. Proceedings, 1985. Edited by
Vol. 260: H.-J. Kruse, Degeneracy Graphs and the A.B. Kurzhanski and K. Sigmund. VI, 215 pages. 1987.
Neighbourhood Problem. VIII, 128 pages. 1986. Vol. 288: G.D. Rudebusch, The Estimation of Macroeconomic
Vol. 261: Th.R. Gulledge, Jr., N.K. Womer, The Economics of Disequilibrium Models with Regime Classification Informati-
Made-to-Order Production. VI, 134 pages. 1986. on. VII,I28 pages. 1987.
Vol. 262: H.U. Buhl, A Neo-Classical Theory of Distribution Vol. 289: B.R. Meijboom, Planning in Decentralized Firms. X,
and Wealth. V, 146 pages. 1986. 168 pages. 1987.
Vol. 263: M. Schafer, Resource Extraction and Market Vol. 290: D.A. Carlson, A. Haurie, Infinite Horizon Optimal
Struucture. XI, 154 pages. 1986. Control. XI, 254 pages. 1987.
Vol. 264: Models of Economic Dynamics. Proceedings, 1983. Vol. 291: N. Takahashi, Design of Adaptive Organizations. VI,
Edited by H.F. Sonnenschein. VII, 212 pages. 1986. 140 pages. 1987.
Vol. 265: Dynamic Games and Applications in Economics. Vol. 292: I. Tchijov, L. Tomaszewicz (Eds.), Input-Output
Edited by T. Basar. IX, 288 pages. 1986. Modeling. Proceedings, 1985. VI, 195 pages. 1987.
Vol. 266: Multi-Stage Production Planning and Inventory Vol. 293: D. Batten, J. Casti, B. Johansson (Eds.), Economic
Control. Edited by S. Axsater, Ch. Schnee weiss and E. Silver. Evolution and Structural Adjustment. Proceedings, 1985. VI,
V, 264 pages. 1986. 382 pages.
Vol. 267: R. Bemelmans, The Capacity Aspect of Inventories. Vol. 294: J. Jahn, W. Knabs (Eds.), Recent Advances and
IX, 165 pages. 1986. Historical Development of Vector Optimization. VII, 405 pages.
1987.
Vol. 268: V. Firchau, Information Evaluation in Capital Markets.
VII, 103 pages. 1986. Vol. 295. H. Meister, The Purification Problem for Constrained
Games with Incomplete Information. X, 127 pages. 1987.
Vol. 269: A. Borglin, H. Keiding, Optimality in Infinite Horizon
Economies. VI, 180 pages. 1986. Vol. 296: A. Borsch-Supan, Econometric Analysis of Discrete
Choice. VIII, 211 pages. 1987.
Vol. 270: Technological Change, Employment and Spatial Dy-
namics. Proceedings, 1985. Edited by P. Nijkamp. VII, 466 Vol. 297: V. Fedorov, H. Lauter (Eds.), Model-Oriented Data
pages. 1986. Analysis. Proceedings, 1987. VI, 239 pages. 1988.
Vol. 271: C. Hildreth, The Cowles Commission in Chicago, Vol. 298: S.H. Chew, Q. Zheng, Integral Global Optimization.
1939-1955. V, 176 pages. 1986. VII, 179 pages. 1988.

Vol. 272: G. Clemenz, Credit Markets with Asymmetric Infor- Vol. 299: K. Marti, Descent Directions and Efficient Solutions
mation. VIII,212 pages. 1986. in Discretely Distributed Stochastic Programs. XIV, 178 pages.
1988.
Vol. 273: Large-Scale Modelling and Interactive Decision
Analysis. Proceedings, 1985. Edited by G. Fandel, M. Grauer, Vol. 300: U. Derigs, Programming in Networks and Graphs.
A. Kurzhanski and A.P. Wierzbicki. VII, 363 pages. 1986. XI, 315 pages. 1988.

Vol. 274: W.K. Klein Haneveld, Duality in Stochastic Linear Vol. 301: J. Kacprzyk, M. Roubens (Eds.), Non-Conventional
and Dynamic Programming. VII, 295 pages. 1986. Preference Relations in Decision Making. VII, 155 pages. 1988.
Vol. 275: Competition, Instability, and Nonlinear Cycles. Pro- Vol. 302: H.A. Eiselt, G. Pederzoli (Eds.), Advances in
ceedings, 1985. Edited by W. Semmler. XII, 340 pages. 1986. Optimization and Control. Proceedings, 1986. VIII, 372 pages.
1988.
Vol. 276: M.R. Baye, D.A. Black, Consumer Behavior, Cost of
Living Measures, and the Income Tax. VII, 119 pages. 1986. Vol. 303: F.x. Diebold, Empirical Modeling of Exchange Rate
Dynamics. VII, 143 pages. 1988.
Vol. 277: Studies in Austrian Capital Theory, Investment and
Time. Edited by M. Faber. VI, 317 pages. 1986. Vol. 304: A. Kurzhanski, K. Neumann, D. Pallaschke (Eds.),
Optimization, Parallel Processing and Applications.
Vol. 278: W.E. Diewert, The Measurement of the Economic Proceedings, 1987. VI, 292 pages. 1988.
Benefits of Infrastructure Services. V, 202 pages. 1986.
Vol. 305: G.-J.C.Th. van Schijndel, Dynamic Firm and Investor
Vol. 279: H.-J. BUttler, G. Frei and B. Schips, Estimation of Behaviour under Progressive Personal Taxation. X. 215
Disequilibrium Modes. VI, 114 pages. 1986. pages. 1988.
Vol. 280: H.T. Lau, Combinatorial Heuristic Algorithms with Vol. 306: Ch. Klein, A Static Microeconomic Model of Pure
FORTRAN. VII, 126 pages. 1986. Competition. VIII, 139 pages. 1988.
Vol. 307: T.K. Dijkstra (Ed.), On Model Uncertainty and its Vol. 333: N. Dellaert, Production to Order. VII, 158 pages. 1989.
Statistical Implications. VII, 138 pages. 1988. Vol. 334: H.-W. Lorenz, Nonlinear Dynamical Economics and
Vol. 308: J.R. Daduna, A. Wren (Eds.), Computer-Aided Tran- Chaotic Motion. XI, 248 pages. 1989.
sit Scheduling. VIII, 339 pages. 1988. Vol. 335: A.G. Lockett, G. Islei (Eds.), Improving Decision
Vol. 309: G. Ricci, K. Velupillai (Eds.), Growth Cycles and Making in Organisations. Proceedings. IX, 606 pages. 1989.
Multisectoral Economics: the Goodwin Tradition. III, 126 pages. Vol. 336: T. Puu, Nonlinear Economic Dynamics. VII, 119
1988. pages. 1989.
Vol. 310: J. Kacprzyk, M. Fedrizzi (Eds.), Combining Fuzzy Vol. 337: A. Lewandowski, I. Stanchev (Eds.), Methodology
Imprecision with Probabilistic Uncertainty in Decision Making. and Software for Interactive Decision Support. VIII, 309 pages.
IX, 399 pages. 1988. 1989.
Vol. 311: R. Fare, Fundamentals of Production Theory. IX, 163 Vol. 338: 1.K. Ho, R.P. Sundarraj, DECOMP: an ImplemeIitation
pages. 1988. of Dantzig-Wolfe Decomposition for Linear Programming. VI,
Vol. 312: J. Krishnakumar, Estimation of Simultaneous Equation 206 pages.
Models with Error Components Structure. X, 357 pages. 1988. Vol. 339: 1. Terceiro Lomba, Estimation of Dynamic
Vol. 313: W. Jammernegg, Sequential Binary Investment Deci- Econometric Models with Errors in Variables. VIII, 116 pages.
sions. VI, 156 pages. 1988. 1990.
Vol. 314: R. Tietz, W. Albers, R. Selten (Eds.), Bounded Ratio- Vol. 340: T. Vasko, R. Ayres, L. Fontvieille (Eds.), Life Cycles
nal Behavior in Experimental Games and Markets. VI, 368 and Long Waves. XIV, 293 pages. 1990.
pages. 1988. Vol. 341: G.R. Uhlich, Descriptive Theories of Bargaining. IX,
Vol. 315: I. Orishimo, G.1.D. Hewings, P. Nijkamp (Eds ), In- 165 pages. 1990.
formation Technology: Social and Spatial Perspectives. Vol. 342: K. Okuguchi, F. Szidarovszky, The Theory of
Proceedings 1986. VI, 268 pages. 1988. Oligopoly with Multi-Product Firms. V, 167 pages. 1990.
Vol. 316: R.L. Basmann, DJ. Slottje, K. Hayes, J.D. Johnson, Vol. 343: C. Chiarella, The Elements of a Nonlinear Theory of
D.J. Molina, The Generalized Fechner-Thurstone Direct Utility Economic Dynamics. IX, 149 pages. 1990.
Function and Some of its Uses. VIII, 159 pages. 1988.
Vol. 344: K. Neumann, Stochastic Project Networks. XI, 237
Vol. 317: L. Bianco, A. La Bella (Eds.), Freight Transport pages. 1990.
Planning and Logistics. Proceedings, 1987. X, 568 pages. 1988.
Vol. 345: A. Cambini, E. Castagnoli, L. Martein, P Mazzoleni,
Vol. 318: T. Doup, Simplicial Algorithms on the Simplotope. S. Schaible (Eds.), Generalized Convexity and Fractional
VIII, 262 pages. 1988. Programming with Economic Applications. Proceedings, 1988.
Vol. 319: D.T. Luc, Theory of Vector Optimization. VIII, 173 VII, 361 pages. 1990.
pages. 1989. Vol. 346: R. von Randow (Ed.), Integer Programming and
Vol. 320: D. van der Wijst, Financial Structure in Small Related Areas. A Classified Bibliography 1984-1987. XIII, 514
Business. VII, 181 pages. 1989. pages. 1990.
Vol. 321: M. Di Matteo, R.M. Goodwin, A. Vercelli (Eds.), Vol. 347: D. Rios Insua, Sensitivity Analysis in Multi-objective
Technological and Social Factors in Long Term Fluctuations. Decision Making. XI, 193 pages. 1990.
Proceedings. IX, 442 pages. 1989. Vol. 348: H. StOrmer, Binary Functions and their Applications.
Vol. 322: T. Kollintzas (Ed.), The Rational Expectations VIII, 151 pages. 1990.
Equilibrium Inventory Model. XI, 269 pages. 1989. Vol. 349: G.A. Pfann, Dynamic Modelling of Stochastic Demand
Vol. 323: M.B.M. de Koster, Capacity Oriented Analysis and for Manufacturing Employment. VI, 158 pages. 1990.
Design of Production Systems. XII, 245 pages. 1989. Vol. 350: W.-B. Zhang, Economic Dynamics. X, 232 pages.
Vol. 324: I.M. Bomze, B.M. Potscher, Game Theoretical Foun- 1990.
dations of Evolutionary Stability. VI, 145 pages. 1989. Vol. 35 I: A. Lewandowski, V. Volkovich (Eds.), Multiobjective
Vol. 325: P. Ferri, E. Greenberg, The Labor Market and Business Problems of Mathematical Programming. Proceedings, 1988.
Cycle Theories. X, 183 pages. 1989. VII, 315 pages. 1991.
Vol. 326: Ch. Sauer, Alternative Theories of Output, Unem- Vol. 352: O. van Hilten,Optimal Firm Behaviour in the Context
ployment, and Inflation in Germany: 1960-1985. XIII, 206 of Technological Progress and a Business Cycle. XII, 229 pages.
pages. 1989. 1991.
Vol. 327: M. Tawada, Production Structure and International Vol. 353: G. Ricci (Ed.), Decision Processes in Economics. Pro-
Trade. V, 132 pages. 1989. ceedings, 1989. III, 209 pages 1991.
Vol. 328: W. Giith, B. Kalkofen, Unique Solutions for Strategic Vol. 354: M. Ivaldi, A Structural Analysis of Expectation For-
Games. VII, 200 pages. 1989. mation. XII, 230 pages. 1991.
Vol. 329: G. Tillmann, Equity, Incentives, and Taxation. VI, Vol. 355: M. Salomon. Deterministic Lotsizing Models for Pro-
132 pages. 1989. duction Planning. VII, 158 pages. 1991.
Vol. 330: P.M. Kort, Optimal Dynamic Investment Policies of Vol. 356: P. Korhonen, A. Lewandowski, J . Wallenius (Eds.),
a Value Maximizing Firm. VII, 185 pages. 1989. Multiple Criteria Decision Support. Proceedings, 1989. XII, 393
Vol. 331: A. Lewandowski, A.P. Wierzbicki (Eds.), Aspiration pages. 1991.
Based Decision Support Systems. X, 400 pages. 1989. Vol. 357: P. Zornig, Degeneracy Graphs and Simplex Cycling.
Vol. 332: T.R. Gulledge, Jr., L.A. Litteral (Eds.), Cost Analysis XV, 194 pages. 1991.
Applications of Economics and Operations Research. Vol. 358: P. Knottnerus, Linear Models with Correlated Distur-
Proceedings. VII, 422 pages. 1989. bances. VIII, 196 pages. 1991.
Vol. 359: E. de Jong, Exchange Rate Detennination and Opti- Vol. 384: A. Martina, Lectures on the Economic Theory of
mal Economic Policy Under Various Exchange Rate Regimes. Taxation. XII, 3 I3 pages. 1992.
VII, 270 pages. 1991. Vol. 385: J. Gardeazabal, M. Regulez, The Monetary Model of
Vol. 360: P. Stalder, Regime Translations, Spillovers and Buffer Exchange Rates and Cointegration. X, 194 pages. 1992.
Stocks. VI, 193 pages. 1991. Vol. 386: M. Desrochers, J.-M. Rousseau (Eds.), Computer-
Vol. 361: C. F. Daganzo, Logistics Systems Analysis. X, 321 Aided Transit Scheduling. Proceedings, 1990. XIII, 432 pages.
pages. 1991. 1992.
Vol. 362: F. Gehrels, Essays In Macroeconomics of an Open Vol. 387: W. Gaertner, M. Klemisch-Ahlert, Social Choice and
Economy. VII, 183 pages. 1991. Bargaining Perspectives on Distributive Justice. VIII, 131 pages.
Vol. 363: C. Puppe, Distorted Probabilities and Choice under 1992.
Risk. VIII, 100 pages. 1991 Vol. 388: D. Bartmann, M. J. Beckmann, Inventory Control.
Vol. 364: B. Horvath, Are Policy Variables Exogenous? XII, XV, 252 pages. 1992.
162 pages. 1991. Vol. 389: B. Dutta, D. Mookherjee, T. Parthasarathy, T.
Vol. 365: G. A. Heuer, U. Leopold-Wildburger. Balanced Raghavan, D. Ray, S. Tijs (Eds.), Game Theory and Economic
Silvennan Games on General Discrete Sets. V, 140 pages. 1991. Applications. Proceedings, 1990. ??, ?? pages. 1992.

Vol. 366: J. Gruber (Ed.), Econometric Decision Models. Pro- Vol. 390: G. Sorger, Minimum Impatience Theorem for
ceedings, 1989. VIII, 636 pages. 1991. Recursive Economic Models. X, 162 pages. 1992.
Vol. 367: M. Grauer, D. B. Pressmar (Eds.), Parallel Computing Vol. 391: C. Keser, Experimental Duopoly Markets with
and Mathematical Optimization. Proceedings. V, 208 pages. Demand Inertia. X, 150 pages. 1992.
1991. Vol. 392: K. Frauendorfer, Stochastic Two-Stage Programming.
Vol. 368: M. Fedrizzi, J. Kacprzyk, M. Roubens (Eds.), VIII, 228 pages. 1992.
Interactive Fuzzy Optimization. VII, 216 pages. 1991. Vol. 393: B. Lucke, Price Stabilization on World Agricultural
Vol. 369: R. Koblo, The Visible Hand. VIII, 131 pages.199 I. Markets. XI, 274 pages. 1992.

Vol. 370: M. J. Beckmann, M. N. Gopalan, R. Subramanian Vol. 394: Y.-I. Lai, c.-L. Hwang, Fuzzy Mathematical
(Eds.), Stochastic Processes and their Applications. Proceedings, Programming. XIII, 301 pages. 1992.
1990. XLI, 292 pages. 1991. Vol. 395: G. Haag, U. Mueller, K. G. Troitzsch (Eds.), Economic
Vol. 371: A. Schmutzler, Flexibility and Adjustment to Infor- Evolution and Demographic Change. XVI, 409 pages. 1992.
mation in Sequential Decision Problems. VIII, 198 pages. 1991.
Vol. 372: J. Esteban, The Social Viability of Money. X, 202
pages. 1991.
Vol. 373: A. Billot, Economic Theory of Fuzzy Equilibria. XIII,
164 pages. 1992.
Vol. 374: G. Pflug, U. Dieter (Eds.), Simulation and Optimi-
zation. Proceedings, 1990. X, 162 pages. 1992.
Vol. 375: S.-I. Chen, Ch.-L. Hwang, Fuzzy Multiple Attribute
Decision Making. XII, 536 pages. 1992.
Vol. 376: K.-H. IOckel, G. Rothe, W. Sendler (Eds.), Bootstrap-
ping and Related Techniques. Proceedings, 1990. VIII, 247
pages. 1992.
Vol. 377: A. Villar, Operator Theorems with Applications to
Distributive Problems and Equilibrium Models. XVI, 160 pages.
1992.
Vol. 378: W. Krabs, J. Zowe (Eds.), Modern Methods of
Optimization. Proceedings, 1990. VIII, 348 pages. 1992.
Vol. 379: K. Marti (Ed.), Stochastic Optimization. Proceedings,
1990. VII, 182 pages. 1992.
Vol. 380: J. Odelstad, Invariance and Structural Dependence.
XII, 245 pages. 1992.
Vol. 381: C. Giannini, Topics in Structural VAR Econometrics.
XI, 131 pages. 1992.
Vol. 382: W. Oettli, D. Pallaschke (Eds.), Advances in
Optimization. Proceedings, 1991. X, 527 pages. 1992.
Vol. 383: I. Vartiainen, Capital Accumulation in a Corporatist
Economy. VII, 177 pages. 1992.