1992 Fuzzy Multiple Attribute Decision Making Methods and Applications Compress

Lecture Notes in Economics
and Mathematical Systems 375

Editorial Board:
H. Albach, M. Beckmann (Managing Editor)
P. Dhrymes, G. Fandel, G. Feichinger, W. Hildenbrand
W. Krelle (Managing Editor)
H. P. Ktinzi, K. Ritter, U. Schittko, P. Schonfeld, R. Selten, W. Trockel
Managing Editors:
Prof. Dr. M. Beckmann
Brown University
Providence, RI 02912, USA
Prof. Dr. W. Krelle
Institut flir Gesellschafts- und Wirtschaftswissenschaften
der Universitat Bonn
Adenauerallee 24-42, W-5300 Bonn, FRG
Shu-Jen Chen Chin-Lai Hwang
In Collaboration with Frank P. Hwang
Fuzzy Multiple Attribute

Decision Making
Methods and Applications
Springer-Verlag
Berlin Heidelberg New York
London Paris Tokyo
Hong Kong Barcelona
Budapest
Authors
Shu-Jen Chen
HTX International Inc.
115 North 4th Street
Manhattan, KS 66502, USA
Ching-Lai Hwang
Department of Industrial Engineering
Kansas State University
Manhattan, KS 66506, USA
ISBN-13: 978-3-540-54998-7 e-ISBN-13: 978-3-642-46768-4

DOl: 10.1007/978-3-642-46768-4
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
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 this rapidly growing field of operations research.
This is a sequel to our previous works: "Multiple Objective Decision
Making--Methods and Applications: A state-of-the-Art Survey" (No.164
of the Lecture Notes); "Multiple Attribute Decision Making--Methods
and Applications: A State-of-the-Art Survey" (No.186 of the Lecture
Notes); and "Group Decision Making under Multiple Criteria--Methods
and Applications" (No.281 of the Lecture Notes).
In this monograph, the literature on methods of fuzzy Multiple
Attribute Decision Making (MADM) has been reviewed thoroughly and
critically, and classified systematically. This study provides
readers with a capsule look into the existing methods, their
characteristics, and applicability to the analysis of fuzzy MADM
problems.
The basic concepts and algorithms from the classical MADM methods
have been used in the development of the fuzzy MADM methods. We give
an overview of the classical MADM in Chapter II.
Chapter III presents the basic concepts and mathematical
operations of fuzzy set theory with simple numerical examples in
a easy-to-read and easy-to-follow manner.
Fuzzy MADM methods basically consist of two phases: (1) the
aggregation of the performance scores with respect to all the
attributes for each alternative, and (2) the rank ordering of the
alternatives according to the aggregated scores.
Chapter IV deals with the fuzzy ranking methods. Fuzzy ranking
methods are concerned with comparing fuzzy numbers and can be applied
to all aspects of fuzzy applications. A systematic classification of

about two dozen existing ranking methods is presented. Each method is
critically reviewed. The basic concepts, the computational procedures,
and the characteristics of each method are concisely discussed.
Numerical examples are used to illustrate the computational procedures.
A system for classifying over one dozen fuzzy MADM methods is
presented in Chapter V. The basic concept, the algorithm, and the
characteristics of each method are discussed. The computational
procedure of each method is illustrated by solving a simple numerical
example.
Most of the real-world MADM problems contain a mixture of fuzzy
and crisp data and may have a large number of alternatives and up to
hundreds of attributes. The existing fuzzy MADM methods are cumbersome

to use and difficult to understand. They cannot effectively solve
most real-world problems of more than ten alternatives and ten
attributes. We present a new and practical approach which can solve

any size real-world problems. This approach is conceptually easy to
understand and operationally easy to use.
An up-to-date bibliographical listing of more than 400 references
is presented.
We wish to acknowledge the pioneering and outstanding work on
fuzzy set theory done by Dr. L.A. Zadeh. We are indebted to all the
fine scholars listed in the references who have carried out and
published their research results. Special thanks are due to
Professors D. Dubois, E.S. Lee, H. Prade, B. Roy, T.L. Saaty,
H. Tanaka, K. Yoon, P.L. Yu, and H.J. Zimmermann. We thank Mei-Hua
Chen for typing and Trenetta Jones for editing.
Shu-Jen Chen C.L. Hwang

HTX International, Inc. Kansas State University
Manhattan, Kansas Manhattan, Kansas
spring 1991 spring 1991
TABLE OF CONTENTS
I. INTRODUCTION 1
II. MULTIPLE ATTRIBUTE DECISION MAKING -- AN OVERVIEW 16

2.1 Basics and Concepts 16
2.2 Classifications of MACM Methods 19
2.2.1 Classification by Information 20
2.2.2 Classification by Solution Aimed At 24
2.2.3 Classification by Data Type 24
2.3 Description of MACM Methods 24
Method (1): DOMINANCE 27
Method (2): MAXIMIN 28
Method (3): MAXlMAX 29
Method (4): CONJUNCTIVE METHOD 30
Method (5): DISJUNCTIVE METHOD 31
Method (6): LEXICOGRAPHIC METHOD 32
Method (7): LEXICOGRAPHIC SEMIORDER METHOD 33
Method (8): ELIMINATION BY ASPECTS (EBA) 34
Method (9): LINEAR ASSIGNMENT METHOD (LAM) 35
Method (10): SIMPLE ADDITIVE WEIGHTING METHOD (SAW) 36
Method (11): ELECTRE (Elimination et Choice 37
Translating Reality)
Method (12): TOPSIS (Technique for Order Preference 38
by Similarity to Ideal Solution)
Method (13): WEIGHTED PRODUCT METHOD 40
Method (14): DISTANCE FROM TARGET METHOD 41
III. FUZZY SETS AND THEIR OPERATIONS 42

3.1 Introduction 42
3.2 Basics of Fuzzy Sets 43
3.2.1 Definition of a Fuzzy Set 43
VIII
3.2.2 Basic Concepts of Fuzzy Sets 45
3.2.2.1 Complement of a Fuzzy Set 45
3.2.2.2 Support of a Fuzzy Set 46
3.2.2.3 a-cut of a Fuzzy Set 46
3.2.2.4 convexity of a Fuzzy Set 47
3.2.2.5 Normality of a Fuzzy Set 48
3.2.2.6 Cardinality of a Fuzzy Set 48
3.2.2.7 The mth Power of a Fuzzy Set 49
3.3 Set-Theoretic operations with Fuzzy Sets 49
3.3.1 No compensation operators 52

3.3.1.1 The Min Operator 52
3.3.2 compensation-Min Operators 53
3.3.2.1 Algebraic Product 53
3.3.2.2 Bounded Product 53
3.3.2.3 Hamacher's Min Operator 53
3.3.2.4 Yager's Min Operator 54

3.3.2.5 Dubois and Prade's Min Operator 54
3.3.3 Full Compensation Operators 57
3.3.3.1 The Max Operator 57
3.3.4 compensation-Max Operators 57
3.3.4.1 Algebraic Sum 57
3.3.4.2 Bounded Sum 58
3.3.4.3 Hamacher's Max Operator 58
3.3.4.4 Yager's Max Operator 58
3.3.4.5 Dubois and Prade's Max Operator 59
3.3.5 General Compensation Operators 60
3.3.5.1 Zimmermann and Zysno's 8 Operator 61

3.3.6 Selecting Appropriate Operators 62
3.4 The Extension Principle and Fuzzy Arithmetics 63
3.4.1 The Extension Principle 64
3.4.2 Fuzzy Arithmetics 66
3.4.2.1 Fuzzy Number 66
3.4.2.2 Addition of Fuzzy Numbers 69
3.4.2.3 Subtraction of Fuzzy Numbers 73
3.4.2.4 Multiplication of Fuzzy Numbers 76
3.4.2.5 Division of Fuzzy Numbers 79
3.4.2.6 Fuzzy Max and Fuzzy Min 82
3.4.3 Special Fuzzy Numbers 86
3.4.3.1 L-R Fuzzy Number 87
3.4.3.2 Triangular (or Trapezoidal) Fuzzy Number 88
3.4.3.3 Proof of Formulas 89
3.4.3.3.1 The Image of Fuzzy Number N 94
3.4.3.3.2 The Inverse of Fuzzy Number N 95
3.4.3.3.3 Addition and Subtraction 96
3.4.3.3.4 Multiplication and Division 97
3.5 Conclusions 99
IV. FUZZY RANKING METHODS 101

4.1 Introduction 101
4.2 Ranking Using Degree of optimality 114
4.2.1 Baas and Kwakernaak's Approach 114
4.2.2 Watson et al.'s Approach 125
4.2.3 Baldwin and Guild's Approach 127
4.3 Ranking using Hamming Distance 134
4.3.1 Yager's Approach 137
4.3.2 Kerre's Approach 143
4.3.3 Nakamura's Approach 147
4.3.4 Kolodziejczyk's Approach 158
x
4.4 Ranking using a-cuts 165
4.4.1 Adamo's Approach 166

4.4.2 Buckley and Chanas' Approach 168
4.4.3 Mabuchi's Approach 169
4.5 Ranking using comparison Function 182
4.5.1 Dubois and Prade's Approach 183

4.5.2 Tsukamoto et al.'s Approach 199
4.5.3 Delgado et al.'s Approach 203
4.6 Ranking using Fuzzy Mean and Spread 217
4.6.1 Lee and Li's Approach 217
4.7 Ranking using Proportion to The Ideal 225
4.7.1 McCahone's Approach 225
4.8 Ranking using Left and Right Scores 233
4.8.1 Jain's Approach 234
4.8.2 Chen's Approach 238
4.8.3 Chen and Hwang's Approach 246
4.9 Ranking with Centroid Index 252
4.9.1 Yager's centroid Index 252
4.9.2 Murakami et al.'s Approach 255
4.10 Ranking Using Area Measurement 259
4.11 Linguistic Ranking Methods 265
4.11.1 Efstathiou and Tong's Approach 266
4.11.2 Tong and Bonissone's Approach 274
V. FUZZY MULTIPLE ATTRIBUTE DECISION MAKING METHODS 289
5.1 Introduction 289
5.2 Fuzzy Simple Additive Weighting Methods 292

5.2.1 Baas and Kwakernaak's Approach 293
5.2.2 Kwakernaak's Approach 301

5.2.3 Dubois and Prade's Approach 306
5.2.4 Cheng and McInnis's Approach 315
5.2.5 Bonissone's Approach 323
5.3 Analytic Hierarchical Process (AHP) Methods 329
5.3.1 Saaty's AHP Approach 331
5.3.2 Laarhoven and Pedrycz's Approach 339
5.3.3 Buckley's Approach 351
5.4 Fuzzy Conjunctive/Disjunctive Method 371
5.4.1 Dubois, Prade, and Testemale's Approach 371
5.5 Heuristic MAUF Approach 385
5.6 Negi's Approach 395
5.7 Fuzzy outranking Methods 407
5.7.1 Roy's Approach 409
5.7.2 siskos et al.'s Approach 419
5.7.3 Brans et al.'s Approach 432
5.7.4 Takeda's Approach 450
5.8 Maximin Methods 454
5.8.1 Bellman and Zadeh's Approach 455
5.9 A New Approach to Fuzzy MACM Problems 465
5.9.1 Converting Linguistic Terms to Fuzzy Numbers 466
5.9.2 converting Fuzzy Numbers to crisp Scores 474
5.9.3 The Algorithm 476
VI. CONCLUDING REMARKS 487

6.1 MADM Problems and Fuzzy Sets 487
6.2 On Existing MACM Solution Methods 488
6.2.1 Classical Methods for MACM Problems 488
6.2.2 Fuzzy Methods for MACM Problems 488
6.2.2.1 Fuzzy Ranking Methods 489
XII
6.2.2.2 Fuzzy MACM Methods 489

6.3 Critiques of the Existing Fuzzy Methods 489
6.3.1 Size of Problem 490
6.3.2 Fuzzy vs. Crisp Data 490
6.4 A New Approach to Fuzzy MACM Problem Solving 491
6.4.1 Semantic Modeling of Linguistic Terms 491
6.4.2 Fuzzy Scoring System 492
6.4.3 The Solution 492
6.4.4 The Advantages of the New Approach 493
6.5 Other Multiple Criteria Decision Making Methods 493
6.5.1 Multiple Objective Decision Making Methods 493
6.5.2 Methods of Group Decision Making under
Multiple criteria 494
6.5.2.1 Social Choice Theory 494
6.5.2.2 Experts Judgement/Group Participation 498
6.5.2.3 Game Theory 498
6.6 On Future Studies 501
6.6.1 Semantics of Linguistic Terms 501
6.6.2 Fuzzy Ranking Methods 502
6.6.3 Fuzzy MACM Methods 503
6.6.4 MACM Expert Decision Support Systems 503
VII. BIBLIOGRAPHY 505

I. INTRODUCTION
Making decisions is a part of our daily lives. The major concern

is that almost all decision problems have multiple, usually
conflicting, criteria. Research on how to solve such problems has
been enormous. Methodologies, as well as their applications, appear
in professional journals of different disciplines. Diversified as
such problems may be, they are broadly classified into two categories:
(1) Multiple Attribute Decision Making (MADM) and (2) Multiple
Objective Decis,ion Making (MODM). From a practical viewpoint, MADM is
associated with problems whose number of alternatives has been
predetermined. The Decision Maker (DM) is to select/prioritize/rank
a finite number of courses of action. On the other hand, MODM is
not associated with problems in which the alternatives have been
predetermined. The DM's primary concern is to design a "most"
promising alternative with respect to limited resources.
Methods and applications of MADM and MODM regarding a single
decision maker have been thoroughly and critically reviewed, and
systematically classified by Hwang and Yoon [H13], and Hwang and Masud
[H12], respectively. Complexity arises when there is more than one
decision maker. The preferred solution must be agreed on by all
interest groups. The analysis must be extended to account for the
conflicts among different groups who, normally, have different goals.
Hwang and Lin [H14] conducted a critical review and systematical
classification on Group Decision Making under Multiple criteria
(GDMMC). This study--multiple attribute decision making under fuzzy
environment--is a sequel to the above works.
It is not surprising to see that uncertainty always exists in the
human world. Research that attempt to model uncertainty into decision
analysis is done basically through probability theory and/or fuzzy set
2
theory . The former pr esents the stochastic nature of decision

analysis while the latter captures the subjectivity of human behavior.
It is suggested by Efstathiou [E2] and Dubois and Prade [031] that a
stochastic decision method such as statistical decision analysis does
not measure the imprecision in human behavior; rather, this method is
a way to model i ncomplete knowl edge about the external environment
surrounding human beings . Fuzzy set theory, on the other hand, is a
perfect means for modeling uncertainty (or imprecision) arising from
mental phenomena which are neither random nor stochastic. The two
different problem domains are shown in Fig. 1.1, where area C
represents the problem domain we will be dealing with . Human beings
are heavily involved in the process of decision analysis. A rational
approach toward decision making should take into account human
subjectivity, rather than employing only objective probability
measures . This attitude towards the uncertainty of human behavior
led to the study of a new decision analysis field--fuzzy decision
making .
Probability problem Fuzzy problem

Domain Domain
(Random and Fuzzy)
Fig. 1 . 1 Probability and fuzzy problem domains .

3
Objectives of This study
This study is an introduction to the application of fuzzy set
theory toward Multiple Attribute Decision Making. It gives a
state-of-the-art survey of the existing methods which solve fuzzy MADM
problems and their applications. Many diversified methods are
reviewed thoroughly and critically and classified systematically.
Some basic concepts and terminologies will be defined so that we can
explain the literature in a consistent manner. This study offers its
readers a capsule look into the existing methods, their characteris-
tics, and applicability. We will also present our new fuzzy MADM
approach.
MADM Problems and Fuzzy Sets
Multiple Attribute Decision Making (MADM) refers to making
selections among some courses of action in the presence of multiple,
usually conflicting, attributes.
Problems dealing with multiple attribute decision making are
common occurrences in everyday life. For example, one may choose a
job (out of several offers) depending on salary, work location,
promotion opportunity, colleagues, etc. One may choose a car (out of
several cars) depending on cost, safety, comfort, gas mileage, etc.
The water resources development plans for a community should be
evaluated in terms of cost, possibility of water shortage, energy
(reuse factor), flood protection, water quality, etc. The selection
criteria for an assistant professor can be based on research ability,
teaching ability, communication skill, and maturity. The examples can
go on forever. Any individual, organization, society, and even a
whole nation must face problems of this type, i.e., MADM problems.
A MADM problem can be concisely expressed in matrix format as:

4
Xl X2 Xn
Al xII x 12 X 1n (1.1)
[
A2 X 21 X 22 x 2n
0
Am x m1 x m2 xmn 1
where Ai' i = 1, ... , m are possible course of actions (referred to as
alternatives); Xj , j = 1, ... , n are attributes with which alternative

performances are measured; x .. is the performance (or rating) of
. 1J
alternative Ai with respect to attribute Xj . Studies on solving MADM
problems have been numerous. Hwang and Yoon [H13] presented a
thorough and systematic survey of the classical MADM solution methods.
The taxonomy of the classical MAOM methods is shown in Fig. 1.2. It
is worth emphasizing that many of the basic concepts of these
classical MAOM methods are used in fuzzy MAOM methods.
It is not uncommon that, at times, the x ij value (or rating)
cannot be assessed precisely. The imprecision may come from different
sources:
1. Unquantifiable information. The price of a new car can be
easily determined while the safety or comfort of a car is not
quantifiable. Safety and comfort are usually expressed in linguistic
terms such as good, fair, poor, etc. They are qualitative data
(subjective judgment by an individual).
2. Incomplete information. The speed of a fast moving object can
be measured by some equipment as "about 90 mph" but not "exactly 90
mph." Such data may be represented as a fuzzy set because of
incomplete information.
3. Nonobtainable information. Sometimes crisp data is obtainable
but the cost is too high, and the OM may wish to get an "approxima-
tion" of that crisp data. When the data is very sensitive (i.e.,
government top secret, an individual's bank account, or a young

I. Type of Information II. Salient Feature III. Major Classes
From the Decision of Information of Methods
Maker
11.1.1 Dominance
) 1. No Information 1 - - - - - - - - - - - - - 1 · 1 . 1 . 2 Max im in
L -_ _ _
1.1.3 _________________
Maximax ~
2.1.1 Conjunctive Method

2.1 Standard Level (Satisficing Method)
2.1.2 Disjunctive Method
I LI 2 . 2 . 1 Lexicographic Method
2.2 Ordinal I 2.2.2 Elimination by Aspects
2.2.3 Permutation Method
Multiple
Attribute 2. Information 2.3.1 Linear Assignment Method
Decision on Attribute 2.3.2 Simple Additive
Weighting Method (SAW)
Making 2.3 Cardinal
2.3.3 Analytic Hierarchical
Process (AHP)
2.3.4 ELECTRE UI
2.3.5 TOPSIS
2.4 Marginal Rate 2.4.1 Hierarchical Trade-Ofts

of Substitution
3.
3.1 Pairwise 3.1.1 LlNMAP
Preference 3.1.2 Interactive SPW Method
3.2 Order of Pair- 13.2.1 MDS With Ideal Point

wise Proximity
Fig. 1.2 A taxonomy of methods for classical MADM problems

[Hwang and Yoon, H131.
6
woman's age, etc.), some "approximated" data or linguistic descriptions
are used. The information is fuzzy because of its unavailability.
4. Partial Ignorance. Some fuzziness is attributed to partial
ignorance of the phenomenon since one knows only part of the facts.
The classical MADM methods cannot effectively handle problems
with such imprecise information. To resolve this difficulty, fuzzy
set theory, first introduced in 1965 by Zadeh [Zl], is used. A fuzzy
set is defined by Zadeh as the following:
Let U be a set of objects (or elements) denoted by x. Thus, U
{x}. A fuzzy set A in U is characterized by a set of ordered pairs A
= {(x) '~A(x)}, V x e U, where ~A(x) is the grade or degree of
membership of x in A. ~A(x) assumes its value in [0,1]. When ~ = 0,

we know x does not belong to the set A; while ~ 1 indicates x is
surely an element in A. The use of a numerical scale [0,1] allows a
convenient representation of gradation. Note, however, that a precise
~ value does not exist; rather, it is subjectively assigned by some
individual(s).
Literature Surveys
The first attempt at applying fuzzy set theory to multi-attribute
analysis was done by Bellman and Zadeh [B8], who outlined one possible
route toward fuzzy decision making. Another important approach was by
Zadeh [Z5] who outlined the possibility of using the max-min rule to
combine relational matrices. Pappies [P1] followed and modeled the
decision matrix using a singleton, i.e., a fuzzy set with only one
element in it (see also Zadeh [Z5]). In this approach, the best
alternative is chosen based on the max-min rule.
In 1978, Kickert [K18] summarized fuzzy set theory application in
relation to MADM problems. Efstathiou [E2] critically reviewed
methods developed prior to 1979. The studies of Yager and Basson [Y1],
7
Yager [Y4], Jain [J2,J5], Baas and Kwakernaak [B1], and Baldwin and
Guild [B3] were among the ones discussed. An in-depth summary of
fuzzy set theory and its application was done by Dubois and Prade
[D26]. They classified the fuzzy MADM into a fuzzy rating phase, in
which the fuzzy utility of each alternative is obtained, and a fuzzy
ranking phase, in which the fuzzy utilities are compared. In
addition, both fuzziness and randomness were accounted for as one of
the possible fuzzy applications to decision analysis. Kaufmann and
Gupta [K15] give an easy-to-read introduction to fuzzy arithmetic,
which is crucial to our application of fuzzy algebraic operations.
The most updated summary of the fuzzy MADM study was done by
Zimmermann [Z30,Z31]. This study also treats fuzzy MADM problems as a
two-phase process. The first phase derives the fuzzy utilities which
are then compared in the second phase using a fuzzy ranking method.
Zimmermann's first book entitled "Fuzzy set Theory--and Its
Applications" [Z30] places more emphasis on fuzzy set theory and its
development rather than its application. The second book entitled
"Fuzzy Sets, Decision Making, and Expert Systems" [Z31] devotes itself
fully to fuzzy decision making and expert systems. It is a good
source of existing fuzzy decision making studies.
In addition to the aforementioned books, a few good articles on
fuzzy decision analysis can be found in books of collected papers.
Table 1.1 summarizes the books and proceedings on this topic.
Classification of Methods
The rapid growth of fuzzy set theory as applied to MADM in the
last decade makes a thorough review of the existing literature and a
systematic classification of methods necessary.
Fuzzy MADM methods basically consist of two phases (Dubois and
Prade [D26], Zimmermann [Z31]): (1) the aggregation of the performance

8
Table 1.1 Books, Monographs, and Proceedings
Class Reference Year

Basic fuzzy set theory and Dinola and Ventre [013] 1986
its operations Dubois and Prade [026] 1980
Kandel [K12] 1986
Kaufmann [K13] 1975
Kaufmann and Gupta [K15] 1985
Zimmermann [Z30] 1985
Fuzzy set theory and Gupta and Sanchez [G12,G13] 1982,1982

decision analysis Kacprzyk and Yager (eds.) [K5] 1985
Kacprzyk and Orlovsky
(eds.) [K8] 1987
Negoita and Ralescu [N9] 1975
Sanchez [S6] 1983
Wang and Chang [W1a] 1980
Zimmermann, Zadeh, and Gaines
(eds.) [Z28] 1984
Zimmermann [Z31] 1987
Fuzzy set theory and Dubois and Prade [036] 1988
its applications in Gupta, Saridies, and Gaines
general (eds.) [G10] 1977
Gupta, Ragade, and Yager
(eds.) [Gll] 1979
Kaufmann and Gupta (eds.) [K16] 1988
Mamdani and Gaines (eds.) [M5] 1981
Negoita [N12] 1979
Zadeh, Fu, Tanaka, and Shimura
(eds.) [Z9] 1975
Decision analysis in Hwang and Masud [H12] 1979

general Hwang and Yoon [H13] 1980
Hwang and Lin [H14] 1987
Kickert [K18] 1978
scores with respect to all the attributes for each alternative, and
(2) the rank ordering of the alternatives according to the aggregated

9
scores. We will refer to the results of the first and second phase
using the terms "final rating" and "ranking order," respectively.
For a crisp MADM problem, the final ratings are expressed as real
numbers. The ranking order can be easily obtained by comparing these
real numbers. In this case, the main focus of MADM problem solving is
the first phase. In a fuzzy MADM problem, the performance scores of
an alternative with respect to all attributes may be expressed by
fuzzy sets. As a result, the final ratings are expressed by fuzzy
sets. obtaining the ranking order of these fuzzy sets is not a
trivial task. In this case, both phase one and phase two are
important in solving the MADM problem. That is, algorithms for
computing and comparing fuzzy final ratings must be used in sequence
to solve the fuzzy MADM problem.
As indicated by Zimmermann [Z31] fuzzy MADM methods are different
from each other in that they either consider phase one, phase two, or
both phases of MADM. It seems appropriate, therefore, to classify
fuzzy MADM methods into two categories, (1) one that focuses on the
second phase, and (2) another that focuses on either the first phase
or both the first and the second phases together. We shall refer to
the methods in the first category as "fuzzy ranking methods," and
those in the second category as "fuzzy MADM methods." The reason for
such a classification scheme is explained below. Fuzzy ranking
methods are concerned with comparing fuzzy numbers, and can be applied
to virtually all aspects of fuzzy applications (not just fuzzy MADM).
Fuzzy MADM methods are specifically designed to solve MADM problems
which contain fuzzy data. The nature of the general applicability of
fuzzy ranking methods makes it appropriate to separate them from fuzzy
MADM methods. Fuzzy ranking methods and fuzzy MADM methods are
systematically presented in Chapters IV and V, respectively.
10
Classification of Fuzzy Ranking Methods
We classify fuzzy ranking methods based on two factors:
(1) the comparison medium used, and (2) the technique (fuzzy or
nonfuzzy) needed to develop the comparison medium. A taxonomy of
fuzzy ranking methods is shown in Fig. 1.3. The classification has
been made in three stages: stage I, the type of comparison medium used
for comparing fuzzy sets, (such as preference relation, mean and
spread of fuzzy set, fuzzy scoring, and linguistic expression); stage
II, the technique needed to develop the comparison medium, (such as

the degree of optimality, the Hamming distance, a-cut, comparison
function, probability distribution, proportion to optimal, left and
right scores, centroid index, area measurement, intuition, and
linguistic approximation); stage III, the major approaches in any
branch formed from stages I and II.
Classification of Fuzzy MADM Methods
A taxonomy of fuzzy MADM methods is shown in Fig. 1.4. The

classification has been made in five stages:
stage I. The size of a problem that a fuzzy MADM method is
suitable to solve. The size of a MADM problem is characterized by the
number of attributes and the number of alternatives. Fuzzy MADM
methods are suitable for solving a problem that has either less than
ten alternatives and ten attributes, or any number of alternatives and
less than 350 attributes.
stage II. The data type allowed by each method can be: (1) all
fuzzy, (2) all fuzzy singleton, (3) all crisp, or (4) a mixture of
fuzzy and crisp. Real world MADM problems contain a mixture of fuzzy
and crisp data.
stage III. The basic concepts of fuzzy MADM methods were derived
mainly from classical MADM methods. The classical MADM methods whose
I.Comparison Medium II.Technique Involved Ill.Approaches
Baas and Kwakernaak [Bl]

Degree of
optimality Watson et al. [W3]
Baldwin and Guild [B3]
Hamming Yager [Yll]

distance Kerre [K17]
Preference Nakamura [ N2]
relation a-cut Kolodzijezyk [K27]
Comparison
function
Probability Du bois and Prade [029]

Fuzzy distribution Tsu kamoto et al. [T15]
ranking Fuzzy mean Delgado et al. [04]
and spread
Proportion to Lee and Li (L3]
optimal
Left/right McCahone [Mel]

scores
Fuzzy scoring Jain (J2,J5]
Centroid index Chen (C12]
Chen and Hwang (C13)
Area Yager (Y5)

measurement Murakami et a!. (M23)
Yager [Y18)
Intuition
linguistic Efstathiou and Tong [E5]
expression Linguistic
approximation Tong and Bonissone (T11)
Fig. 1.3 A taxonomy of fuzzy ranking methods.

Problem Corresponding Technique
Data Type Approaches
Size MADM Methods Involved
~imple Additive Baas and Kwakernaak [B1)

l All fuzzy a-cut J Kwakernaak [K32)
eighting method
H Dubois and Prade [D28)
Cheng and Mcinnis [C17)
l Fuzzy Arithmetics 1
AHP Bonissone [B26,B27)
r---
y 1Eigenvector method I
n c 10
Saaty [S1)
I m ( 10
~ Weight asaessing .~
arithmetic operatio Laarhoven and
Pedrycz [L 1)
ICrisp/fuzzy I- Conjunction/
Disjunction Buckley [B38,B39)
~ Posaibility and I
Fuzzy method necessity measures J
Multiple Dubois et al. [D37)
Attribute r--- I
Human
Decision intuition Efstathiou and
Making ~
H MAUF r Rajkovic [E3) I\)
All crisp I
I Ranking methods I
General MADM Fuzzy arithmetic r Negi [N7)
method
I n c 350 I Outranking Fuzzy outranking I- Siskos et al. [S20)

m • any num ber method relation Brans et al. [B35)
J
All fuzzy
(singleton) I
I Maximin rl Max and min Bellman and Zadeh [B8)
I operators Yager [Y4)
y Crisp/fuzzy General MADnLinguistiC-.fUZZ Y setl I

methods I
Fig. 1.4 A taxonomy of fuzzy MADM methods.

13
basic concepts were adopted include Simple Additive weighting (SAW)
method, Analytic Hierarchical Process (AHP) method, Conjunctive method,
Disjunctive method, Multiple Attribute utility Function (MAUF) theory,
Outranking method, Maximin, TOPSIS, and general classical MADM
methods.
Stage IV. The technique required to apply each fuzzy MADM
method. The techniques include a-cut, fuzzy arithmetic operations,
weight assessing method (e.g., Eigenvector method), possibility and
necessity measures, human intuition, fuzzy outranking relation,
maximum and minimum operators, and semantic modeling (linguistic data
-> fuzzy data -> crisp number).
Stage V. The major approaches in any branch formed from the
previous four stages are listed here.
Because of the broad interdisciplinary character of this research
field, the literature is diversified in many journals as shown in
Table 1.2. Journals in which fuzzy decision analysis articles appear
frequently are indicated by an asterisk *.
Although we have tried to give a reasonably complete survey, some
papers may have been inadvertently overlooked or were not considered
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.
Note on Bibliography of Fuzzy Set Theory and Applications
There are many bibliographies of fuzzy set theories and their
applications. We shall list them for the readers who may be
interested in applications other than MADM (see Table 1.3). They are:
(1) Gaines and Kohout [G2] list 1150 articles and/or books.
(2) Kandel and Yager [K11] list some 1799 entries which overlap
somewhat with the first survey.
(3) Kaufmann [K14] collected some important articles published

14
Table 1.2 List of Journals

1. Automatica
2. Advances in control
3. Automation and Remote Control
4. Applied System and Cybernetics
5.* BUSEFAL (Bulletins for Studies and Exchanges or
Fuzziness and its Application) (in French)
6. Computer and Mathematical Applications
7. Control and Cybernetics
8. Cybernetica
9. Cybernetics and Systems
10. Decision Science
11. Economic Compo Econ. Cybern. Stud. Res.
12. European Journal of Operations Research
13. Finnish Journal of Business Economics
14.* Fuzzy Sets and Systems
15.* IEEE Symposiums on Multi-Valued Logic
16. IEEE Trans. On Automatics and Control
17. IEEE Trans. On Computers
18. IEEE Trans. On Power Apparatus and Systems
19.* IEEE Trans. On Systems, Man, and Cybernetics
20. Information and Control
21. Information Sciences
22.* International Journal of Approximate Reasoning
23. International Journal of Control
24. International Journal of Computing Information Sci.
25. International Journal of General Systems
26. International Journal of Intelligent System
27.* International Journal of Man-Machine Studies
28. International Journal of Math. Ed. Sci. Tech.
29. International Journal of System Science
30. Journal of Cybernetics
31. Journal of Experimental Psychology
32. Journal of JAACE (in Japanese)
33.* Journal of Mathematical Analysis and Applications
34. Journal of Optimization Theory and Applications
35. Journal of Operational Research Society
36.* Journal of Fuzzy Mathematics (in Chinese)
37. Kybernetes
38. Management Science
39. Philosophy of Science
40. Stochastica
41. Theory and Decision
42. water Resources Research
15
Table 1.3 Collection of Bibliographies in Chronological Order
1. Gaines and Kohout [G2] (1977)

2. Kandel and Yager [K11] (1979)
3. Kaufmann [K14] (1980)
4. Maiers and Sherif [M4] (1985)
5. Kandel [K12] (1986)
6. Zimmermann [Z31] (1987)
7. Kaufmann and Gupta (eds.) [K16] (1988)
8. Dubois and Prade [034,035] (1987, 1988)
before 1980 which deal with fuzzy set theory and its applications.
(4) Maiers and Sherif [M4] list some 450 entries. They are
systematically classified into different application areas such as:
Automatic Control, Medical, Economics, General Engineering,
Environmental Topics, Psychology, Reliability, Pattern Recognition,
General Science, Operations Research, and Decision Making. There are
about 87 papers which are directly related to Decision Analysis.
(5) Kandel [K12] lists about 1000 references that he believes
important on fuzzy set theory and its applications.
(6) Zimmermann [Z31] collects many important references
specifically for fuzzy decision analysis and fuzzy expert systems.
(7) Kaufmann and Gupta [K16] give 57 books pertaining to fuzzy set
theory and its applications. They also list the current major sources
where one may look in for important articles: (a) BUSEFUL, (b) Journal
of Fuzzy Mathematics, (c) Fuzzy Sets and Systems, (d) Conferences
sponsored by North American Fuzzy Information Processing society,
(e) Japanese Working Group in Fuzzy Systems, (f) Chinese Working Group
in Fuzzy Sets, and (g) Indian Working Group in Fuzzy Sets.
(8) Dubois and Prade [034,035] collect many important references
regarding fuzzy set theory and its applications which can be seen in
the section "RECENT Literature" of Fuzzy Sets and Systems.

II. MULTIPLE ATTRIBUTE DECISION MAKING -- AN OVERVIEW
Multiple Attribute Decision Making (MADK) refers to making

decisions in the presence of multiple, usually conflicting,
attributes. Problems for multiple attributes decision making are
commmon occurrences in every aspect of life. For example:
In a personal context, the job one chooses may depend upon its
prestige, location, salary, advancement opportunities, working
conditions, and so on. The car one buys may be characterized in terms
of price, gas mileage, style, safety, comfort, etc.
In a business context, a business executive's choice of corporate
strategy may depend on the company's earnings over a period of time,
its stock price, share of the market, goodwill, labor relations,
corporate image, obligation to society, and so forth.
In an academic context, a private college evaluates candidates
for admission based on IQ, leadership, high school grades, rank in
senior class, etc.
In a public context, the water resources development plan for a
community should be evaluated in terms of cost, probability of water
shortage, energy (reuse factor), recreation, flood protection, land
and forest use, water quality, etc.
In a goverment context, the choice of missile systems for the Air
Force could be based on speed, yield, accuracy, range, vulnerability,
reliabililty, etc.
2.1 Basics and Concepts

The problems of MADK are widely diverse. However, even with the
diversity, all the problems which are considered here share the
following common characteristics:
Alternatives. A finite number of alternatives, from several to
17
thousands, are to be screened, prioritized, selected and ranked. For
example, the number of missile systems the Air Force can choose from
may be less than five; while an elite college may have over thousands
of applicants for admission each year. The term "alternative" may be
referred to as "cause of action," or "candidates," among others.

Multiple attributes. Each problem has multiple attributes. A
decision maker must generate relevant attributes for each problem
setting. The term "attributes" may be referred to as "goals" or
"criteria." The number of attributes can be between several to around
350. For example, one may use price, gas mileage, safety, leg room,
workmanship, and style to evaluate cars; while there may be over 100
factors which must be considered when selecting a site for a plant.

In most cases where the number of attributes is large, the
attributes are in a hierachical structure. There may be several major
attributes; each major attibute will have several subattributes; and
each subattribute may have several sub-subattributes. For example,
a hierachical tree for about 350 attributes may look like this:
Sub Sub-Sub
X111
Xl12
X11
X12
X117
Xl
X17
X2
Attributes--- X3
X4
X5
X6
X7
Fig. 2.1a A hierarchy tree of attributes.

18
Note that the number seven appears in the major and subattribute
categories. The number seven is based on Miller's [M16] theory that
seven plus or minus two represents the greatest amount of information
an observer can give us about an object on the basis of an absolute
judgement.
Conflict among attributes. Multiple attributes usually conflict
with each other. For example, in selecting a car, the higher gas
mileage might reduce the comfort rating because of the smaller
passenger space.
Incommensurable units. Each attribute has a different unit of
measurement. In the car selection case, gas mileage is expressed by
miles per gallon (MPG), comfort is expressed by cubic feet if it is
measured by passenger space, cost is indicated by dollars, safety may
be indicated in a nonnumerical way, etc.
Decision Weights. Almost all methods and/or MACM problems
require information regarding the relative importance of each
attribute. The relative importance is usually given by a set of
weights which are normalized to sum to one. In the case of n
attributes, a weight set is
and 1.
The weights can be assigned by the decision maker directly, or

calculated using the eigenvector method or the weighted least square
method (see Sasty [S1], Chu, Kalaba, and spingarn [C21], Hwang and
Yoon [H13]).
Decision matrix. ~ MACM problem can be concisely expressed in a

matrix format. A decision matrix D is a (m x n) matrix whose element
x ij indicates the performance rating of alternative i, Ai' with
19
respect to attribute j, Xj • Hence Ai' i 1,2, ... ,m is denoted by
and the column vector,
shows the contrast of each alternative with respect to attribute Xj .
Numerical Example (Hwang and Yoon [H13]):
A country decided to purchase a fleet of jet fighters from the
u.s. The Pentagon officials offered the characteristic information of
four models which may be sold to that country. The Air Force analyst
team of that country agreed that six characteristics (attributes)
should be considered. They are: maximum speed (Xl)' ferry range

(X 2 ), maximum payload (X 3 ), purchasing cost (X 4 ), reliability (X s ),
and maneuverability (X 6 ). The measurement units for the attributes
are: mach, miles, pounds, dollars (in millions), high-low scale, and
high-low scale, respectively. The decision matrix for the fighter
aircraft selection problem, then, is:
Xl X2 X3 X4 Xs X6
A1 2.0 lS00 20000 S.S average very high
[
A2 2.S 2700 18000 6.S low average
1
0 A3 1.8 2000 21000 4.S high high
A4 2.2 1800 20000 S.O average average
2.2 Classifications of MADM Methods
There are dozens of MADM methods available. Each method has its
own characteristics and applicability. By examining each method's

20
characteristics and applicability, one may be able to categorize MADM

methods into different groups. The result of this classification
provides readers with a systematic and overall view of the MADM
research field. Various classification schemes have been proposed
during the past two decades. We shall present the most dominant ones.
2.2.1 Classification by Information

Hwang and Yoon [H13] classified a group of 17 MADM methods
according to the type of information from the decision maker and the
salient features of the information. A taxonomy of the methods is
shown in Fig. 2.1.
In this classification, the methods were first categorized by the
type of information received from the decision maker: no information,
information on attributes, or information on alternatives. If no
information was given, the methods in this category are dominance,
maximin and maximax. If information was given, a subcategory, the
salient feature of the received information from the decision maker,
was used to further group the methods. The information given may be a
standard level of each attribute, such as in the conjunctive or
disjunctive method; it may be the ordinal preference of attributes of
which, for example, the lexicographic method and Elimination by
Aspects (EAB) would apply; it may be a cardinal preference of the
attributes of which, for example, the Simple Additive weighting (SAW)
method, Elimination et Choice Translating Reality (ELECTRE) and
Technique for Order Preference by Similarity to Ideal Solution
(TOPSIS) apply; or the information may be of the marginal rate of
substitution between the attributes where the hierarchical tradeoffs
method applies. If the information was given on alternatives, the
methods were further subclassified by whether the information was of
pairwise preference or the order of pairwise proximity.
I. Type of Information II. Salient Feature III. Major Classes
From the Decision of Information of Methods
Maker
1.1.1 Dominance
)1. No Information i1.1.2 Max~min
1.1.3 Maxlmax
2.1.1 Conjunctive Method

2.1 Standard Level (Sat isf icing Method)
2.1.2 Disjunctive Method
I ~12.2.1 Lexicographic Method

2.2 Ordinal 2.2.2 Elimination by Aspects
2.2.3 Permutation Method
Multiple
Attribute 2. Information 2.3.1 Linear Assignment Method
Decision on Attribute 2.3.2 Simple Additive
Weighting Method (SAW)
Making 2.3 Cardinal
2.3.3 Analytic Hierarchical
Process (AHP)
2.3.4 ELECTRE ~
2.3.5 TOPSIS
2.4 Marginal Rate !-2~4.1 Hierarchical Trade-Ofts

of Substitution
3. Information
3.1 Pairwise 3.1.1 LlNMAP
Preference 3.1.2 Interactive SAW Method
3.2 Order of Pair- 13.2.1 MDS With Ide;1 P~i~t .- I

wise Proximity
Fig. 2.1 A taxonomy of methods for classical MADM problems

(Hwang and Yoon, H131.
22
The taxonomy of MAOM methods by Hwang and Yoon [H13] was modified
by Hwang [H15] to Fig. 2.2. Six methods were removed and three new
methods were added. The three methods added were the lexicographic
semiorder method, the weighted product method, and the distance from
target method. The methods removed were the permutation method, the
Analytic Hierarchical Process (AHP) method, the Linear Programming
Techniques for Multidimensional Analysis of Preference (LINMAP)
method, the interactive SAW method and the Multidimensional scaling

(MOS) with ideal point method.
The permutation method was removed because its logic is more
difficult for the decision maker to understand, and because the
computations are quite involved and not practical for more than about
four or five alternatives. The AHP method is not a new technique, but
a combination of the eigenvector method and the simple additive
weighting method. The hierarchical tradeoffs method requires
extensive information from the decision maker, and the approach is
more useful for designing an alternative rather than selecting one.
The other three methods removed were LINMAP, the interactive SAW
method, and the MOS with ideal point method. These three all belong
to the third major branch of the original taxonomy in which the type
of information from the OM consists of the information on
alternatives. These methods require that the OM indicate his/her
preference between two alternatives. This kind of information is far
more demanding to assess than the information on attributes, which is
one reason why these three methods were removed. The problems which
LINMAP and the MOS with ideal point method were designed for involve
market research or consumer preference. These methods try to find the
reason people buy a particular car; that is, which attributes are most
important in influencing the decision. This is a different type of
decision making problem.

Type of Information
From the Decision Salient Feature Major Classes
Maker of Information of Methods
Dominance
Maximin
No Information Maximax
Conjunctive Method
Standard level (Satisficing Method)
Disjunctive Method
Multiple
Attribute
Decision Lexicographic Method
Elimination By Aspect I\:)
Making Col
Ordinal Lexicographic
Semiorder
Information on
Linear Assignment
Attribute
Method
Simple Additive
Cardinal Weighting Method
ELECTRE
TOPSIS
Weighted Product
Distance from Target
Fig. 2.2 A taxonomy of MADM methods (Hwang [H15]).
24
2.2.2 Classification by Solution Aimed At

In addition to classifying MADM methods by type of information
received from the decision makers, and the salient feature of the
information (as in Fig. 2.2), other classification schemes are
possible. Another approach (Hwang [H15]) is to classify the methods
according to the solution aimed at, as shown in Fig. 2.3. In this
classification, if the solution aimed at by the decision maker is to
screen, then the dominance method, conjunctive method, or disjunctive
method is appropriate. If the solution aimed at is to evaluate,

prioritize and select, then maximin, SAW, ELECTRE, or TOPSIS are among
the appropriate methods. In some situations, the solution aimed at
may be to first screen then evaluate, prioritize and select. In this

case, one of the methods for screening can be used for the screening
stage, and one of the other methods can be used for the evaluation,
prioritization, and selection stage.
2.2.3 Classification by Data Type
One more way of classifying methods is by the data type which the
problem contains, as shown in Fig. 2.4 (Hwang [H15]). In this
classification, the methods are grouped according to whether they are
of data type yes-no where only the dominance, lexicographic,
lexicographic semiorder, and EBA method apply; of data type rank where
only the dominance, lexicographic, lexicographic semiorder and Linear
Assignment Method (LAM) method apply; or of data type numeric where
the conjunctive method, SAW, ELECTRE and TOPSIS, for example, apply.
2.3 Description of MADM Methods
The MADM methods to be included are those which are easy to
understand and/or easy to apply to real world large size problems.
Only the essential ideas will be presented. The details of the

25
Solution aimed at MADM Methods
Dominance
Screen Conjunctive method
Disjunctive method
Maximin
[MADM Maximax
Lexicographic
Lexicographic
Semiorder
EBA
Evaluate! LAM
Prioritize! SAW
Select ELECTRE
TOPSIS
Weighted Product
Fig. 2.3 MADM methods classified by solution aimed at

(Hwang [H15)).
Data Type MADM Methods

Dominance
Lexicographic
Lexicographic
Semiorder
EBA
Dominance
Lexicographic
Lexicographic
Semiorder
LAM
Dominance
Maximin
Maximax
Conjunctive
Disjunctive
Lexicographic
Lexicographic
SAW Semiorder
ELECTRE
TOPSIS
Weighted Product
Fig. 2.4 MADM methods classified by data type

(Hwang [H15)).
26
methods can be seen in the references mentioned for each method.
The methods are presented in a format which can be used to
explain the methods to the user. In the format used, the
characteristics of each method are described in the following order:
the logic of the method, its basic principle, the step-by-step
procedure, any requirements, when it is applicable, its advantages and

disadvantages, and its reference.
The methods to be described are listed below in the order of
presentation.
1) Dominance
2) Maximin
3) Maximax
4) conjunctive method (satisficing method)
5) Disjunctive method
6) Lexicographic method
7) Lexicographic semi corder
8) Elimination by Aspects (EBA)
9) Linear Assignment Method (LAM)
10) simple Additive Weighting method (SAW)
11) ELECTRE
12) TOPSIS
13) Weighted Product

14) Distance from Target
27
Method (1): DOMINANCE
Logic and Basic Principle: An alternative is dominated if there is
another alternative which excels it in one or more attributes and
equals it in the remaining attributes.
Procedure:
1) Compare the first two alternatives. If one is dominated by
the other, discard the dominated one.
2) Next, compare the undiscarded alternatives with the third
alternative. Discard any dominated alternatives.
3) Then, introduce the fourth alternative and so on.
4) After (m-1) stages, the nondominated set is determined.
Requirement: None
Applicable when: the solution aimed at is to screen out dominated
alternatives.
Advantages: simple, easy to use and understand.
Disadvantages: Some dominated alternatives, which would get discarded,
may actually be better overall than some of the nondominated
alternatives.
References: Hwang and Yoon [H13].

28
Method (2): MAXIMIN
Logic: A chain is only as strong as its weakest link.
Example: An astronaut's life or death in orbit may depend upon his/her
worst vital organ.
Basic Principle: The overall performance of an alternative is
determined by its weakest or poorest attribute.
Procedure:
1) For each alternative, determine its poorest attribute value.
2) Select the alternative with the best value on the poorest
attribute. In mathematical notation, an alternative. A+, is selected
such that
{A;
.
I max min x;)'}, j
i j •
1,2, ••. ,n; i 1,2, ..• ,m.
Requirement: All attributes must be measured on a common scale, i.e.,
attributes should have commensurable units.
Applicable when: the OM is assumed to have a pessimistic nature
about the decision making situation.
Advantages: Simple, easy to use and understand.
Disadvantages: Only one attribute is used to represent an alternative.
All other (n-l) attributes for a particular alternative are ignored.
In other words, the tradeoff among attributes is noncompensatory.
References: Hwang and Yoon [H13], MacCrimmon [M2].

29
Method (3): MAXIMAX
Logic: An alternative is selected by its best attribute value.
Example: Professional football players are selected based on their
best talent: passing, running, kicking, etc.
Basic principle: The overall performance of an alternative is
determined by its best attribute.
Procedure:
1) For each alternative, identify its best attribute value.
2) Select the alternative with the maximum overall best value.
In mathematical notation, an alternative, A+, is selected such that
1,2, •.. ,n; i 1,2, ..• ,m.
Requirement: All attributes must be measured on a common scale,
i.e., all attributes should have commensurable units.

Applicable when: the DM is assumed to have an optimistic nature about
the decision making situation.
Disadvantages: Only one attribute is used to represent an alternative.
All other (n-l) attributes for a particular alternative are ignored.
In other words, the tradeoff among attributes is noncompensatory.
References: Hwang and Yoon [HI3], MacCrimmon [M2].

30
Method (4): CONJUNCTIVE METHOD

Logic: An alternative which does not meet the minimal acceptable level
for all attributes is rejected.
Example: To obtain a driver's license, one must get an acceptable
score on all tests.
Basic princlple: The minimal acceptable levels for each attribute are
used to screen out unacceptable alternatives.
Procedure:
1) The DM specifies a minimal acceptable level (cutoff score)
for each attribute.
2) For each alternative, determine if the value of each of its
attributes equals or exceeds the minimal acceptable level.
3) If so, this alternative is acceptable. Oherwise, it is
rejected. Mathematically, Ai is an acceptable alternative only if
1,2, ••• ,n,
where xj is the minimal acceptable level for Xj •

Requirement: A minimal acceptable level on each attribute must be
specified.
Applicable when: the solution aimed at is to screen out unacceptable
alternatives.
Disadvantages: A candidate with just one unacceptable attribute will
be rejected, even if that candidate has high values for all other
attributes. In other words, the tradeoff among attributes is
noncompensatory.
References: Hwang and Yoon [H13], Dawes [DIa].
31
Method (5): DISJUNCTIVE METHOD
Logic: A candidate is selected who has an extreme talent in anyone

attribute.
Example: Professional football players are selected who have an
extreme talent in passing, running, kicking etc.
Basic Principle: Desirable levels for each attribute are used to
select alternatives which equal or exceed those levels in anyone
attribute. An alternative is evaluated based on the greatest value
(or talent) of an attribute.
Procedure:
1) The DM specifies a desirable level for each attribute.
2) For each alternative, determine if any of its attribute
values equals or exceeds the desirable level.
3) If any do, the alternative is acceptable. Otherwise, it is
rejected. Mathematically, Ai is an acceptable alternative only if
1, or 2 or ... or n,
where xj is a desirable level of x j .
Requirement: A minimal acceptable (desirable) level for each attribute
must be specified.
Applicable when: the solution aimed at is to screen out unacceptable
alternatives.
Disadvantages: Alternatives who are good in all attributes but lack
an exceptional one will not be selected. The tradeoff among
attributes is noncompensatory.
Reference: Hwang and Yoon [H13), Dawes [D1a).

32
Method (6): LEXICOGRAPHIC METHOD

Logic: In some decision making situations a single attribute seems to
predominate.
Example: The "buy the cheapest" rule is one in which price is the
most important attribute to the OM.
Basic Principle: To compare the alternatives in the order of the
important attributes.
Procedure:
1) Compare all alternatives with respect to the most important
attribute. Select the alternative with the highest value on that
attribute.
2) If there are several alternatives with the highest value,
compare those tied alternatives with respect to the next most
important attribute and select the alternative with the highest
value in that attribute.
3) Proceed in this manner until only one alternative is left or
until all attributes have been considered.
Requirement: The attributes must be ranked in terms of importance.
Applicable when: attributes have a dominating relationship such that
W1 »>W 2 »> ... »>Wn where wi is the weight of the ith most important
attribute.
Disadvantages: The tradeoff among attributes is noncompensatory.
33
Method (7): LEXICOGRAPHIC SEMIORDER METHOD
Logic: In some decision making situations a single attribute seems to
predominate but allow bands of imperfect discrimination so that one
alternative is not judged better just because it has a slightly higher
value on the predominated attribute.
Procedure:
1) Compare all alternatives with respect to the most important
attribute. Select the alternative(s) with the highest value on that
attribute or with a value not significantly lower than the highest
value.
2) If more than one alternative is selected, compare those
tied alternatives with respect to the next most important attribute
and select the alternative(s) with the highest or near highest value
in that attribute.
3) Proceed in this manner until only one alternative is left or
Requirement: The attributes must be ranked in terms of importance.
Also a tolerance value must be specified on each attribute,
indicating the amount of difference from the best value which is
not considered significant.
Applicable when: attributes have a dominating relationship such that
w1 »>w 2 »> •.. »>w n ' where wi is the weight of the ith most important
attribute.
Disadvantages: The tradeoff among attributes is noncompensatory.
References: Hwang and Yoon [H13], Luce [L9], Tversky [T14].

34
Method (8): ELIMINATION BY ASPECTS (EBA)

Logic and Basic Principle: The elimination process is governed by the
successive selection of aspects (attributes). Alternatives are
compared one attribute at a time and eliminated from consideration if
they do not pass a yes-no or minimum acceptabale level.
Procedure:
1) The DM specifies minimum cutoffs for each attribute.
2) starting with the attribute that has the most discrimination
power in a probabilistic mode, eliminate all alternatives which do not
pass the yes-no or minimum cutoff for that attribute.
3) Proceed attribute by attribute in order of the discrimination
power in a probabililstic mode until only one alternative is left or
Requirment: The attributes are ranked in terms of their
discrimination power in a probabilistic mode.
Disadvantages: A candidate with just one unacceptable attribute will
be rejected even if that candidate has high values for all other
attributes. In other words, the "tradeoff among attributes is
noncompensatory.
References: Hwang and Yoon [H13], Tversky [T14].
35
Method (9): LINEAR ASSIGNMENT METHOD (LAM)
Logic and Basic Principle: An alternative which has many high ranked
attributes should be ranked high.
Procedure:
1) Rank the alternatives for each attribute.
2) Assign an importance weight to each attribute.
3) Create a square (m x m) nonnegative matrix IT whose element
ITik represents the score of alternative Ai on the kth attributewise
ranking. The score ITik is the summation of the weights of all
attributes where Ai is ranked k.
4) Use the linear assignment method to assign a rank to each
alternative such that the summation of the scores for that assignment
is maximized.
Requirement: None
Applicable when: ordinal data is given as the score of the alternative
on each attribute.
Advantages: Simple, easy to use and understand. It requires less
effort to collect data as compared to methods requiring cardinal
data.
Disadvantages: The actual cardinal difference between alternatives on
each attribute is not considered. Thus, an alternative ranked first
on an attribute could have a cardinal score of 100, and one ranked
second could have a high score of 99. Yet on a different attribute,
an alternative ranked first may have a score of 100, while one ranked
second only a score of 50.
References: Hwang and Yoon [H13], Bernardo and Blin [B9b].

36
Method (10): SIMPLE ADDITIVE WEIGHTING METHOD (SAW)

Logic and Basic Principle: The overall score of an alternative is
computed as the weighted sum of the attribute values.
Procedure:
1) For each alternative, compute a score by multiplying the scale
rating of each attribute by its importance weight and summing these
products over all attributes.
2) Select the alternative with the highest score. Mathematically,
the most preferred alternative, A-, is selected such that
A-
where x ij is the outcome of the ith alternative about the jth attribute
with a numericallY comparable scale, Wj is the importance weight of
the jth attribute.
Requirement: The attributes must be both numerical and comparable.
The decision maker assigns importance weights to attributes.
Advantages: The best known and most widely used method. Simple.
easy to use and understand. The tradeoff among attributes is
compensatory.
Disadvantages: If attributes are complementary (a high score on
one attribute always occurs with a high score on another attribute),
the computed score violates the assumption of separable utility of
each attribute.
References: Hwang and Yoon [H13], MacCrimmon [M2].
37
Method (11): ELECTRE (Elimination et Choice Translating Reality)
Logic: The concept of an outranking relationship is used, which says
that even though two alternatives ~ and Al do not dominate each

other mathematically, the DM accepts the risk of regarding ~ as
almost surely better than AI.
Procedure:
Because the procedure is rather lengthy, it will not be presented
here. Those interested can refer to the references shown below,
specifically, Hwang and Yoon [H13].

Advantages: The tradeoff among attributes is compensatory. It fully
utilizes the information contained in the decision matrix.
Disadvantages: Only a partial prioritization of alternatives is
computed. As the number of alternatives increases, the amount of
calculations rises quite rapidly. computational procedures are quite
elaborate.
References: Hwang and Yoon [H13], Benayoun, Roy, and Sussman [Bga],
Roy [R10].
38
Method (12): TOPSIS (Technique for Order Preference by Similarity
to Ideal Solution)
Logic and Basic Principle: The chosen alternative should have the
shortest distance from the ideal solution and the farthest distance
from the negative-ideal solution.
Procedure:
1) Calculate the normalized decision matrix. The normalized
value r ij is calculated as:
i 1,2, ... ,m; j 1,2, ... ,n.
/
2) Calculate the weighted normalized decision matrix. The
weighted normalized value v ij is calculated as:
1,2, ... ,m; j 1,2, ... ,n,
n
where wJ' is the weight of the jth attribute and L w. = 1.
j=1 J
3) Determine the ideal and negative-ideal solution .
A
• {(max vijl j E J) , (min vijl j E J') I i =1,2, ... ,m}
.
i
•
{V 1 ' v 2 ' ... , .
vj '
i
... , v
n
} ,
A - { (min vijl j E J) , (max vijl j E J') I i =1,2, ... ,m}

i i
{V~, v 2 ' ... , vj , ... , vn },
where J = {j=1,2, ... ,nlj associated with benefit criteria}
where J' = {j=1,2, ... ,nlj associated with cost criteria}

39
4) Calculate the separation measures. The separation between
each alternative can be measured by the n-dimensional Euclidean
distance. The separation of each alternative from the ideal one is
then given as:
Similarly, the separation from the negative-ideal solution
is given as:
S ,-
~ I ~
J!;'l
(V IJ - V-)2
J ' i=l,2, ... ,m.
5) Calculate the relative closeness to the ideal solution. The
relative closeness of Ai with respect to A• is defined as:
6) Rank the preference order.
Advantages: Simple, easy to use and understand. The tradeoff among
attributes is compensatory.

40
Method (13): WEIGHTED PRODUCT METHOD
Logic: In order to penalize alternatives with poor attribute values
more heavily, a product instead of a sum of the values is made
across the attributes.
Procedure:
1) For each alternative, raise the scale rating of each
attribute to a power equal to the importance weight of the attribute.
Then multiply the resulting values over all attributes.
2) Select the alternative with the highest product.
Mathematically, the most preferred alternative, A·, is selected
such that
A• max
i
where x ij is the outcome of the ith alternative about the jth
attribute, with a numerically comparable scale, and Wj is the
normalized importance weight of the jth attribute.
Applicable when: the OM wishes to avoid alternatives with poor
attribute values.
References: Easton [E1].

41
Method (14): DISTANCE FROM TARGET METHOD
Logic: For some attributes, the best value may be located in the
middle of the attribute range.
Example: In buying a house the number of rooms should be neither too
many nor too few.
Basic Principle: The alternative which has the shortest distance
from the target alternative is selected.
Procedure:
1) For each alternative, compute the deviation from the target.
2) Select the alternative with the shortest distance value.
Mathematically, an alternative, A; is selected with the shortest
distance
I i:
j =1
W2
j
(x
i j
- t)2
j'
i=1,2, ... ,m,
where x .. is the outcome of the ith alternative about the jth attribute
~]
with a numerically comparable scale; tj and Wj are the target level
and the normalized weight of the jth attribute, respectively.
Requirement: A target level must be specified on each attribute.
The attributes must be both numerical and comparable.
Applicable when: the DM has in mind a set of target levels on each
attribute. Also, the attributes do not necessarily have a
monotonically increasing or monotonically decreasing utility. That
is, for "benefit" attributes, bigger is not necessarily better or for
"cost" attributes, smaller better. A value in the middle may be
preferred here.
References: Easton [E1].

III. FUZZY SETS AND THEIR OPERATIONS
3.1 Introduction
Fuzzy set theory is developed for solving problems in which
descriptions of activities and observations are imprecise, vague, and
uncertain. The term "fuzzy" refers to the situation in which there
are no well-defined boundaries of the set of activities or
observations to which the descriptions apply. For example, one can
easily assign a person seven feet tall to the "class of tall men".
But it would be difficult to justify the inclusion or exclusion of a
six-foot tall person to that class, because the term "tall" does not
constitute a well-defined boundary. This notion of fuzziness exists
almost everywhere in our daily life, such as the "class of red
flowers," the "class of good kickers," the "class of expensive cars,"
or "numbers close to 10," etc. These classes of objects cannot be
well represented by classical set theory. In classical set theory,
an object is either in a set or not in a set. An object cannot

partially belong to a set.
To cope with this difficulty, Zadeh [Zl] proposed the fuzzy set
theory in 1965. A fuzzy set is a class of objects with a continuum of
membership grades. A membership function, which assigns to each
object a grade of membership, is associated with each fuzzy set.
Usually, the membership grades are in [0,1]. When the grade of
membership for an object in a set is one, this object is absolutely in
that set; when the grade of membership is zero, the object is
absolutely not in that set. Borderline cases are assigned numbers
between zero and one. Precise membership grades do not convey any
absolute significance. They are context-dependent and can be
subjectively assessed.
In the following sections, we will present some basic definitions
of fuzzy set and operations on fuzzy sets from mathematical aspects.

43
Subsequently, the extension principle and fuzzy number operations,
which are important to subsequent discussions, will be introduced.
special fuzzy numbers such as triangular numbers, L-R fuzzy numbers,

and trapezoidal numbers and their arithmetic operations are also
presented. Numerical and graphical examples are used to make the
contents more understandable.
3.2 Basics of Fuzzy Sets
In this section we will review the definition of a fuzzy set as
well as some of its basic concepts as they apply to later chapters.
3.2.1 Definition of a Fuzzy Set
Let U be a classical (or ordinary) set of objects, called the
universe, whose generic elements are denoted by x. That is, U = {x}.

A fuzzy set A in U is characterized by a membership function ~A(X)
which associates with each element in U a real number in the interval
[0,1]. The fuzzy set, A, is usually denoted by the set of pairs
A {(X'~A(X», x e U}. (3.1)
For an ordinary set, A,
1, iff x e A,
{ (3.2)
0, iff x E A.
When U is a finite set {Xl' ... , Xn }, the fuzzy set on U may also
be represented as (Zadeh [Z4], Dubois and Prade [026]):
n
A [ xi/~A(xi)· (3.3)
i=l
44
When U is an infinite set, the fuzzy set may be represented as:
A
(3.4)
Example 1. Let U = {Ken, John, Allen, Peter}, which is a finite
set. Evaluated by a girl, the fuzzy set "handsome boys", may be
characterized as:
A {(Ken, 0.7), (John, 0.2), (Allen, O.S), (Peter 0.6)},

or
A Ken/0.7 + John/0.2 + Allen/O.S + Peter/0.6.
Example 2 (zimmermann [Z26]). Let U = {10, 20, 30, 40, 50, 60,
70, SO, 90, 100}, the possible speed (mph) at which cars can cruise
over a long distance. Then the fuzzy set "comfortable speed for long
distance travel" may be defined by an individual as:
A {(JO, .7), (40, .75), (50, .SO), (60, .SO),

(70,1.0), (SO, .SO), (90, .30)}.
Notice that x = 10, 20, and 100 are considered as "absolutely
uncomfortable cruising speed", i.e., the grade of comfort is zero.
They are omitted from the fuzzy set.
Example 3 (Dubois and Prade [026]). Let U = {positive real
numbers}, which is an infinite set. Then, the fuzzy set A = "real
numbers close to 10" (see Fig.3.1) may be defined as A = {(x'~A(x»}
where
1 / {1 + [ 1/5 (x - 10)]2}.
45
real numbers not close to 10

IL
1 .(X) ", / rea/~~rs close to 10
.9 "
.8
.7
.6
.5
.4
.3
.2
.1
0
0 2 4 6 8 10 12 14 16 18 20 x
Fig. 3.1 The fuzzy set "real numbers close to 10".
3.2.2 Basic Concepts of Fuzzy sets
The basic concepts presented here include complement, support,
a-cut, convexity, normality, cardinality, and the mth power of a
fuzzy set A.
3.2.2.1 Complement of a Fuzzy Set: The complement of fuzzy set
A, denoted by A, is defined as:
(3.5)
Example (finite fuzzy set case): Consider Example 2 again.
The complement of fuzzy set A, "comfortable speed for long distance
travel," A, is computed as:
A {(10,1.0), (20,1.0), (30,.30), (40,.25), (50,.20), (60,.20),

(80,.20), (90,.70)}.
For example, the membership grade ~A(X) at x 30 is computed as:
A is the fuzzy set "uncomfortable speed for long distance travel."

46
Since one has the most comfortable feeling at 70 mph, i.e., ~A(70) 1,
it is not possible to include x = 70 in A.

Example (infinite fuzzy set case): Consider the fuzzy set "real
numbers close to 10" (see Fig. 3.1). Its complement set is
represented by the dashed curve in Fig.3.1. This complement set may
be interpreted as "real numbers not close to 10."
3.2.2.2 Support of a Fuzzy Set: It is often necesary to
consider those elements in a fuzzy set which have nonzero membership
grades. These elements are the support of that fuzzy set. That is,
given a fuzzy set A, its support is an ordinary set on U defined as:
SeA) (3.6)
3.2.2.3 a-cut of a Fuzzy Set: The a-cut of fuzzy set A is an

ordinary set whose elements belong to fuzzy set A -- at least to the
degree of a. That is, for fuzzy set A its a-cut is defined as:
(3.7)
The a-cut is a more general case of the support of a fuzzy set. When
S (A) •
Example (finite fuzzy set case): Consider Example 2 again. The
support of fuzzy set A, "comfortable cruising speed for long
distance", is given as:
SeA) {30, 40, 50, 60, 70, 80, 90},
where all these x values have their corresponding ~A(x) > O.

47
/L(x) convex nonconvex

1. -
Fig. 3.2 Convex and nonconvex fuzzy sets.
By setting a 0.50, we obtain
AO.S {30, 40, 50, 60, 70, aO} ,
where x = 90 is discarded since /lA(90) < 0.5. If we set a 0.9,
AO.9 = {70}.
Example (infinite fuzzy set case): Consider again the fuzzy set
"real numbers close to 10" used in Example 3. Its support is any
real number between [4,16]. And its a-cut at degree of 0.55 is any
real number between [7,12]. That is, the set of real numbers that
have at least 0.55 membership value is between 7 and 12 (inclusive).
3.2.2.4 Convexity of a Fuzzy set: The convexity of a fuzzy set
is an important property from the application aspect. A fuzzy set A
is convex if
(3. a)
where xl' x 2 e U, and A e [0,1]. Fig. 3.2 gives a convex fuzzy set
and a nonconvex fuzzy set. All the fuzzy sets in the following
48
chapters are assumed convex. For simplicity, we will use the term
"fuzzy set" to denote a convex fuzzy set.
3.2.2.5 Normality of a Fuzzy set: A fuzzy set A is normal if
and only if there are one or more x, values such that ~A(x') = 1. All
fuzzy sets in Figs. 3.1 and 3.2 are normal fuzzy sets. This property
guarantees that at least one element in a fuzzy set fully satisfies

\
the phenomenon that the fuzzy set applies to. Unless otherwise
stated, all the fuzzy sets in the following chapters are assumed
normal.
3.2.2.6 cardinality of a Fuzzy Set: The cardinality of fuzzy
set A evaluates the proportion of elements of U having the property A.
When U is finite, it is defined as:
L ~A(x), x E U. (3.9)
For infinite U the cardinality is defined as:
J ~A(x) dx. (3.10)

x
The relative cardinality of A is defined as:
IIAII (3.11)
The relative cardinality can be interpreted as the proportion of
elements of U being in A weighted by their degree of membership in A.
Example (finite fuzzy set case): For the fuzzy set A,
"comfortable cruising speed for long distance travel", in Example 2,
its cardinality IAI and relative cardinality I IAI I are computed as:
49
.7 + .75 + .80 + .80 + 1.0 + .80 + .30 5.15,
5.15
IIAII 10
0.515.
3.2.2.7 The mth Power of a Fuzzy Set: The mth power of fuzzy
set A is defined by Zadeh [Z4] as:
(3.12)
It is very useful in modeling linguistic modifiers into fuzzy sets.
For example, the second power of a fuzzy set, "good", is interpreted
as "very good" where "very" is the linguistic modifier used to modify
fuzzy set "good."
Example: Let fuzzy set A be
x 3 4 5 6 7 8 9 10
o o .2 .4 .6 .8 1.0 o
The second power of A is computed as:
x 3 4 5 6 7 8 9 10
o o .04 .16 .36 .64 1.0 o
For instance, ~A2(7) .36.
3.3 Set-Theoretic operations with Fuzzy Sets
In classical set theory, intersection and union operations are
frequently used on ordinary sets. The notion of intersection and
union can be carried over to fuzzy sets, too. Zadeh [Zl] was the
first to extend intersection and union to fuzzy sets. He used the min
50
operator and the max operator to model intersection and union,
respectively. A justification of the choice of min and max was given
by Bellman and Giertz [B9]. They also pointed out that from a logical
point of view, we may interpret the intersection as "logical and," the
union as "logical or."
Set-theoretic operators are not limited to the min operator and
the max operator. Many set-theoretic operators have been developed
and published in the past. They have been systematically presented in
the works of Oubois and Prade [030,032], Mizumoto [M19,M20],
Zimmermann [Z30,Z31], Oombi [014,015], Oombi and Vas [017], Klement
[K20], Czogala and Zimmermann [C25], Alsina, Trillas, and Valverde
[AS], Alsina [A6], and others.
In this section, we do not intend to exhaust all operators in
this field. Rather, only important operators which are frequently
cited will be addressed. These set-theoretic operators were broadly
classified into the following five categories: no compensation,
compensation-min, full compensation, compensation-max, and general
compensation operators. The classification is illustrated by Fig.
3.3. This taxonomy is designed from a decision-making view point.
Given several attributes in a MAOM problem, an alternative is selected
solely on its best attribute, even if all of its other attributes are
below average. This is interpreted as a fully-compensated situation.
The max operator is the most suitable operator for that case.
similarly, an alternative is rejected simply because it has a very
poor attribute value, even when all its other attributes are well
above average. This is a situation of no compensation. The basic
operator in this class is the min operator. The compensation-min
operators allow the OM to express how strong s/he means "and" for a
decision problem. similarly, the compensation-max operators provides
the OM the flexibility to show how strong s/he means "or" for a
~.3.1 No Compensation H3.3.1.1 The Min Operator
Operators
~.3.2.1Algebraic Product
3.3.2.2 Bounded Product
3.3.2.3 Hamacher's Min
~.3.2 Compensation-Min ~ Operator
Operators 3.3.2.4 Yager's Min Operator
3.3.2.5 Dubois and Prade's Min
Operator
Full Compensation
----f.3.3.1 The max Operator
IFuzzy Operators
V~.3.3
-
Operators
~.3.4.1 Algebraic Sum ~

3.3.4.2 Bounded Sum
~il.3.4 Compensation-Max 3.3.4.3 Hamacher's Max
Operators - Operator
3.3.4.4 Yager's Max Operator
3.3.4.5 Dubois and Prade's Max
Operator
3.3.5 General Compensatio 3.3.5.1 Zimmermann and Zysno's

Operators eOperator
-
Fig. 3.3 A taxonomy of set theoretic operators.
52
decision problem. The general compensation operators avoid both
extremes by setting certain parameter values. Such parameter values
will determine the actual operator's position between "logical and"
and "logical or."

Throughout this section, we have not given the proof or
mathematical properties of each fuzzy operator in order to avoid
lengthy discussion of each operator and involved mathematical text.
The cited references provide detailed discussions. Also, for the

sake of simplicity, binary operation will be used instead of n-ary
operation.
The following operators are used on fuzzy sets A and B. For
simplicity, we will denote ~A(x) and ~B(x) by ~A and ~B' respectively,

throughout this chapter.
3.3.1 No compensation operators
The min operator is the only no compensation operator. It will
be briefly discussed in this section.
3.3.1.1 The Min Operator: For fuzzy sets A and B, the min
operator determines a fuzzy set {(x'~AnB)} where ~AnB is defined by:
(3.13 )
or, in abbreviated form
(3.14)
~AnB may be viewed as the intersection of fuzzy sets A and B
(Zadeh [Zl]). It is interpreted as the "largest" fuzzy set which is
contained in both fuzzy sets A and B.

53
3.3.2 compensation-Min operators
The compensation-min operators allow some degrees of compensation
when taking the intersection of fuzzy sets in the decision space. The
operators included in this category are the algebraic product, the
bounded product, Hamacher's min operator, Yager's min operator, and
Dubois and Prade's min operator. Each will be briefly discussed in
this section.
3.3.2.1 Algebraic Product: The algebraic product of fuzzy sets
A and B is given by {(x'~AoB)} where
(3.15)
This is a very probabilistic-like operator.
3.3.2.2 Bounded Product: The bounded product of fuzzy sets A
and B is given as {(X'~A(o)B)} where ~A(o)B is defined as:
max (0, ~A + ~B - 1). (3.16 )
A complete interpretation of the operator (0) has not been provided.
3.3.2.3 Hamacher's Min Operator: Hamacher's (Dubois and Prade
[030], Zimmermann [Z30]) min operator is a parameterized intersection
written as A7B, which is characterized by:
For a given (~A'~B) pair, we have (Dubois and Prade [030]):
(3.18 )
where
54
IlA, i f IlS 1,
TW(IlA'IlS)
{ IlS'
0,
if IlA
otherwise.
1, (3.19)
When 7 = 1, IlA7B = (IlA) (Il S ). If any of the pair (IlA,IlS ) is 0, IlA7S

o. If IlA = 1 and Ils ~ 0, IlA7S = Ils' and vice versa. The IlA7S
approaches its upper bound when 7 = o.
3.3.2.4 Yager's Min operator: Yager's [Y21] min operator is

defined as:
IlAqS (3.20)
IlAqS increases with respect to q. Sy setting different values for q,

various fuzzy intersection operators can be derived. special cases
for q = 1 and m are summarized below.
q IlAqB
1 max (0, IlA+ Il s - 1)

min (Il A ,IlB )
The parameter q is inversely related to the strength of the

"and." That is, the lower the q the stronger the "and." This implies
that q is a measure of how strong we mean "and" in the decision space.
For example, when q is approaching zero, the OM's demand for a fast
"and" cheap car is stronger than the case where q is approaching m.
3.3.2.5 Dubois and Prade's Min Operator: Dubois and Prade's

[030] min operator is defined as:
IlAAB (3.21)
55
It decreases with respect to A. When A is at its extreme, i.e., A

is ° or 1, we have:
IlAAB
° min (IlA,Il B )
1 (IlA) (Il B )
When A = ° and IlA ~ Il B , then IlAAB = Il B , i.e., min (IlA,Il B ) Il B •

Similar remarks can be applied to the case when IlB ~ IlA.
Note:
The min operator, the algebraic product, and the bounded product
may be seen as three basic operators for intersection. Each measures
different degree of "and" in the decision space. Sometimes, it is
not so obvious as to which one to use to combine fuzzy sets. This
difficulty may be alleviated by using Hamacher's, Yager's, or Dubois
and Prade's min operators because they have parameters which allows
the DM to define how strong s/he means "and" in the decision space.
By setting parameter values, the DM automatically determines an
intersection operator that may very well represent how s/he wants
to combine fuzzy sets in the decision space. The DM specified
operator may not be any of the basic intersection operators.
Example: We shall use an example to illustrate the computational

process of each intersection operator. The physical meaning of the
fuzzy sets and operational results will not be assumed. Let
A {(4,0), (5,.2), (6,.4), (7,.6), (8,.8), (9,1.0), (10,0)}

and
B {(3,0), (4,.5), (5,.7), (6,1.0), (7,.7), (8,.5), (9,0)}.
Assume 7 = q = A = .5. The various intersection results of A and B

are summarized below.
56
x 3 4 5 6 7 8 9 10
IlA 0 0 .2 .4 .6 .8 1.0 0
IlS 0 .5 .7 1.0 .7 .5 0 0
min (IlA, Il S ) 0 0 .2 .4 .6 .5 0 0
IlAoS 0 0 .14 .4 .42 .4 0 0
IlA( o)S 0 0 0 .4 .3 .3 0 0
IlAl'S 0 0 .16 .4 .45 .42 0 0
Il AqS 0 0 0 .4 0 0 0 0
IlAAS 0 0 .2 .4 .6 .5 0 0
For example, at x = 7, we have:
min (IlA (7), IlS (7» = • 6;
Il AoS (7) = IlAIlS = (.6)(.7) .42;
IlA (o)s(7) = max (0, IlA (7) + IlS (7) - 1) = max [0, .3] .3;
IlA1'S (7) IlA (7)IlS (7)/[.5 + .5(IlA (7) + Il S (7) - IlA (7)IlS (7»] .45;
IlAqS (7) 1 - min(l,[(l - IlA (7»·5 + (1 -IlS (7».5]1/.5)

1 - min [1,1.39] = 0;
IlA (7)Il S (7)/[maX (IlA (7),IlS (7),A)]

• 42 / [max (. 6, . 7 , . 5)] = •6•
Throughout this book, we will frequently use this table format to

present fuzzy sets. It is much easier to read than Eqs.(3.1), (3.3),
and (3.4).
57
3.3.3 Full Compensation operators

The max operator is the only operator in this category. This
operator models the "logical or" in the decision process. It will be
briefly discussed in this section.
3.3.3.1 The Max Operator: For fuzzy sets A and B, the max
operator determines a fuzzy set {(x'~AvB)} where ~AvB is defined as:
~AvB (x) (3.22)
or
~AVB(X) ~A(X) V ~B(X). (3.23)
~AvB may be seen as the union of fuzzy sets A and B (Zadeh [Zl). It
is interpreted as the "smallest" fuzzy set containing both fuzzy sets
A and B.
3.3.4 compensation-Max Operators
The compensation-max operators allow some degrees of (but not
full) compensation when taking the union of fuzzy sets in the decision
space. The operators in this category are the algebraic sum, the
bounded sum, Hamacher's max operator, Yager's max operator, and Dubois
and Prade's max operator. Each will be briefly discussed in this
section.
3.3.4.1 Algebraic Sum: The algebraic sum of fuzzy sets A
and B is characterized by:
~A+B (3.24)
This is a probabilistic-like union operator. Note that Eq.(3.24) is
meaningful only when ~A + ~B s 1.

58
3.3.4.2 Bounded Sum: The bounded sum of fuzzy sets A and B is

characterized by:
/.LA(+)B min (1, /.LA + /.LB). (3.25)
This is another way of representing the union of fuzzy sets. So far,

no physical meaning has been defined.
3.3.4.3 Hamacher's Max Operator: Hamacher's max operator

(Zimmermann [Z30]) on fuzzy sets A and B is defined as:
/.LA7'B (3.26)
For any pair of (/.LA,/.LB), we have
/.LA + /.L B
(3.27)
and its value increases as '1' approaches o.
3.3.4.4 Yager's Max Operator: Yager's max operator (Yager

[Y21]) on fuzzy sets A and B is characterized by:
q' + /.Lq' ) l/q' ], q'i!! 1.

min [1, (/.LA (3.28)
/.LAq'B B
special cases for q' 1 and CD are:
q' /.LAq'B
1 min (1, /.LA + /.LB)
CD max (/.LA,/.LB)
59
The membership function ~Aq'B is a general form of the union

operator. By setting different q' values, various fuzzy union
operators can be derived. As the parameter q' approaches one, we have
a very soft demanding "or," and as q' approaches co, we have the
strongest demanding "or."
3.3.4.5 Dubois and Prade's Max Operator: Dubois and Prade's

max operator [D30] on fuzzy sets A and B is defined as:
~A + ~B - (~A) (~B) - min [l-i\"~A'~B]

~Ai\'B , i\'e [0,1].
max [i\', l-~A' l-~B]
(3.29)
For i\' = 1, ~Ai\'B becomes the algebraic sum ~A+B. The membership
value ~Ai\'B decreases when i\' approaches 0.
Note:
The max operator, the product sum, and bounded sum can be seen as
the three basic forms of the union operator. Each measures different
degree of "or" in the decision space. It is sometimes difficult to
decide which basic form to use to combine fuzzy sets. The
parameterized operators such as Hamacher's max operator provides the
DM the flexibility to specify how strong s/he means "or" for a given
decision problem. The DM specified union operator may not be any of
the basic union operators.
Example: Let fuzzy sets A and B be
A {(4,0), (5,.2), (6,.4), (7,.6), (8,.8), (9,1.0), (10,0)}

and
B {(3,0), (4,.5), (5,.7), (6,1.0), (7,.7), (8,.5), (9,0)}.
Assume that 7' = q' = i\' = .5. The results of various union operators
can be summarized as follows.
60
x 3 4 5 6 7 8 9 10
o o .2 .4 .6 .8 1.0 o
o .5 .7 1.0 .7 .5 o o
max (JlA,Jl B) o .5 .7 1.0 .7 .8 1.0 o

JlA+B o .5 .76 1.0 .88 .90 1.0 o
JlA(+)B o .5 .9 1.0 1.0 1.0 1.0 o
JlA'1'B o .5 .96 1.0 .93 .94 1.0 o
JlAq'B o .5 1.0 1.0 1.0 1.0 1.0 o
JlAA'B o .5 .7 1.0 .76 .8 1.0 o
For example, at x = 7, we have:
(0.5) JlA (7) Jl B (7) + (0.5) (JlA (7) + JlB (7»

.93;
• 5 + JlA(7) JlB (7)
1.0;
JlA (7) + Jl B (7) - JlA (7)Jl B (7) - [.5 A JlA (7) A Jl B (7)]
max [.5, 1 - JlA (7) , 1 - Jl B (7)]
.76.
3.3.5 General compensation Operators

All the operators discussed so far have modeled either the
"logical and" or the Illogical or." In a decision making model, it is
as if we are picking an appropriate course of action based on either
61
the best or the worst attribute value alone. However, when the DM
wants a compromised solution (i.e., a course of action that lies
between the results of "logical and" and "logical or"), the operators
proposed in the previous sections are not applicable. Therefore, some
general operators must be used to encompass the tradeoff information.
This is a new breed of operators which aggregate fuzzy sets in a more
"generalized" format.
3.3.5.1 Zimmermann and Zysno's a Operator: This operator [Z24]

does not distinguish between intersection and union. The a parameter
determines the tradeoff between intersection and union. For a pair
(~A'~B)' the a operator is generally defined as:
~AaB (3.30)
One may use the algebraic product and the algebraic sum for
intersection and union, respectively. Eq.(3.30) then becomes:
(3.31)
The parameter a indicates where the actual operator is located between

the "logical and" and "logical or."
Example: Let
A {(4,0), (5,.2), (6,.4), (7,.6), (8,.8), (9,1.0), (10,0)}

and
B = {(3,0), (4,.5), (5,.7), (6,1.0), (7,.7), (8,.5), (9,0)}.
Assume that the algebraic product and the algebraic sum are used as the
intersection and union operations, respectively. When we set a = 0.5,
(i.e., the actual operator is located exactly in between "logical and"
and "logical or"), we obtain:

62
x 3 4 5 6 7 8 9 10
~A 0 0 .2 .4 .6 .8 1.0 0
~B 0 .5 .7 1.0 .7 .5 0 0
~A8B 0 0 .33 .63 .61 .60 0 0
For instance, we calculate ~A8B(7) as:
.61.
Note that ~A8B(7) lies in the interval [~A.B(7)'~A+B(7)] = [.40,.88]

for 058 5 1. Similar remarks can be made regarding ~A8B(x), x =
5,6,8, as well.
3.3.6 Selecting Appropriate Operators
So far, we have about 15 different fuzzy operators available.
Which one to use in combining fuzzy sets is not a question with a
simple answer. Yager [Y21], Dubois and Prade [030], and Zimmermann
[Z30,Z31] provide some useful rules that can be used as decision aids
in selecting appropriate operators. Following are the eight general

guidelines proposed by Zimmermann [Z30,Z31]:
1. Axiomatic strength: An operator with less axiomatic
restriction is better.
2, Empirical fit: The operator must be an appropriate model of
real system behavior which can only be proven by empirical testing.
3. Adaptability: The aggregation operator should be dependent
on the context and the semantic interpretation. No operators are
general enough for all contexts.

63
4. Numerical efficiency: The operator should be reasonably easy
to compute.
5. Compensation: Logical "and" and logical "or" are too
restrictive on a decision model. We should have operators that can
aggregate fuzzy sets within both extremes.
6. Range of Compensation: In general, the larger the range of
compensation, the better the compensatory operator.
7. Aggregating Behavior: considering normal or subnormal fuzzy

sets, the degree of membership in the aggregated set depends very
frequently on the number of sets combined. For instance, combining
fuzzy sets by the product operator will decrease the membership value
each time a new fuzzy set is introduced. Such behavior mayor may not
be desirable.
8. Required scale level of membership function: Different
operators require different scale levels. For instance, the min
operator is admissible for ordinal information but the product
operator is not. In general, all others being equal, the operator
which requires the lowest scale level is the most preferred one.
3.4 The Extension Principle and Fuzzy Arithmetics
As noted by Dubois and Prade [026], the extension principle
introduced by Zadeh [Z4,Z6,Z7] and others is one of the most basic
ideas of fuzzy set theory. It is used to generalize nonfuzzy (crisp)
mathematical concepts into fuzzy quantities. An important field of
applications for the extension principle is given by algebraic
operations such as addition and multiplication. We shall give the
definition of the extension principle first and extend from it to
fuzzy algebraic operations.

64
3.4.1 The Extension Principle (Dubois and Prade [026], Zadeh [Z6])
Before introducing the extension principle, we have to define the
concept of cartesian product first. Let U be a cartesian product of
universe, U = U1 x •.• x Un' and A1 , •.• , An be n fuzzy sets in U1 '

... , Un' respectively. The cartesian product of A1 , ... , An is
defined as:
Example: Let fuzzy sets A and B be
x 3 4 5 6 7 8 9 10
o o .2 .4 .6 .8 1.0 o
o .5 .7 1.0 .7 .5 o o
The cartesian product of A and B is:
AxB {[(5;4),.2], [(5;6),.2], [(5;8),.2], [(6;4),.4], ... ,
[(8;8),.8], [(9;4),.5], [(9;6),.7], [(9;8), 1.0]).
There is a total of 15 elements which are pairs of each element in A
and each element in B. As a demonstration, we derive
min [.4,.5] .4.
It follows that the extension principle can be defined as
follows. Given a function f that maps from U1 x ... , x Un to a
universe Y such that y = f(x 1 , ... , Xn ) where y E Y and xi E Ui ' vi,
i. e. ,
(3.33)
65
Then a fuzzy set M in Y can be defined as:
M {(y,~(y» I y (3.34)
where
-1
SUp min (~A (x1)'···'~A (x n », if f (y)~O,
{
Y = f(x 1 , .•. x n ) 1 n
~(y)
0, otherwise. (3.35)
Note that Eq.(3.35) is true only when the inverse of f is not zero,
-1 -1
i.e., f (y) ~ o. When f (y) = 0, ~(y) = o. ~(y) is the greatest
among the membership values ~A xA (xl' ... , xn) of the realization

1 x ••• n
of y using n-tuples (xl' .•• , x n )·
The special case for n = 1 gives:
{
~A(f-1(y», if f-1(y) ~ 0,
~(y) (3.36)
0, otherwise.
Example: This example shows how we can use the extension

principle to extend a crisp algebraic operation into a fuzzy one.
Let fuzzy sets A1 and A2 be
2 3 4 5 6 7
o .4 1 .7 o o
o .1 .8 1 .3 o
Based on the crisp algebraic function, f(x) = 2X 1 + x 2 ' the composi-
tion of A1 and A2 is completed using the extension principle as:

66
6 789 10 11 12 13 14 15 16 17
o o .1 .1 .4 .4 .8 1 .7 .7 .3 o
For instance, to get ~B(12) we know the possible (x 1 ,x 2 ) pairs such
that 12 = 2x 1 + x 2 are:
4 5
4 2
The corresponding ~A (xl) and ~A (X 2 ) and their minimums are:

1 2
~A (Xl) .4 1 .7
1
~A (X 2 ) .3 .8 0
2
~A (Xl)" ~A (X 2 ) .3 .8 0
1 2
Thus, ~B(Y=12) max [.3,.8,0] .8.
3.4.2 Fuzzy Arithmetics
Fuzzy arithmetics is a direct application of the extension
principle, and is used on fuzzy numbers. Some works related to fuzzy
number operations are from Jain [J2], Mizumoto and Tanaka [M17,M18],
Baas and Kwakernaak [B1], Dubois and Prade [022,026], Dijkman,
Haeringen, and DeLange [09], Gupta [G14], Kaufmann and Gupta [K15]
among others have been
3.4.2.1 Fuzzy Number
The term fuzzy number is used to handle imprecise numerical
quantities, such as "close to 10," "about 7," "several," etc. A
general definition of a fuzzy number is given by Dubois and Prade

67
[022,026]: any fuzzy subset M = {(x,~(x»} where x takes its number

on the real line R and ~(x) e [0,1]. The membership function denotes
the degree of truth that M takes a specific number x'. Two fuzzy
numbers are equal if and only if they have the same membership
functions. This membership function ~(x) can be described by the
following (see Fig.3.4):
P.(x)
1................. .------,.
Fig. 3.4 A special fuzzy number M with continuous membership

function.
1. A continuous mapping from R to the closed interval [0,1];
=
2. Constant on (-CD,a]: ~(X)
° vx e (-CD,a];
3. strictly increasing on [a,b] ;

4. Constant on [b,c] : ~(X) = 1, vx e [b,c];
5. strictly decreasing on [c,d] ;
6. Constant on [d,+CD) : ~M(x) = 0, vx e [d,+CD) .
The terms a, b, c, and dare real numbers.

Note that we can have a -CD, or b = c, or a = b, or c = d, or
d = +CD. If a = band c = d, M is an interval [b,c]. If b = c, M
is a representation of the fuzzy number, "approximately b."
A fuzzy number may be represented in discrete or continuous form.
68
",(x)
1.
.9
.8
.7
.6
.5
.4
.3-
.2
.1
0 x
0 10 20 30 40 50 60 70 80 90 100
x 30 40 50 60 70 80 90
~(x) .3 .5 .8 1.0 .S .5 .3
Fig. 3.5 A discrete fuzzy number M.
",(x)
1.
.9-
.8
.7
.6
.5
.4-
.3
.2
.1
o ~~~~~~--~~--~-T--~~
o 10 20 30 40 50 60 70 80 90
Fig. 3.6 A continuous fuzzy number M.
For example, let M be the fuzzy number "about 60" which may be given
as either one of the following:
(i) Discrete membership function: Given the universe
U = {10,20,30,40,SO,60,70,SO,90,100},
the fuzzy number M may be represented as shown in Fig. 3.5.
(ii) Continuous membership function: Given the universe

69
U ={ real numbers }, the continuous membership function for M may be
represented as (see Fig. 3.6):
~(x)
3.4.2.2 Addition of Fuzzy Numbers
The addition of two fuzzy numbers M and N may be done in two
different ways.
(i) Use of a-cut: Let's define the a level sets for M and N
using the intervals of confidence as:
(3.37)
and
(3.38)
The addition of M and N may be rewritten as:
(3.39)
This is equivalent to adding two intervals of confidence level by
level (Kaufmann and Gupta ([K15]).
(ii) Use of max-min convolution: Let V x, y, z, e R. Then the
addition of M and N equals
(3.40)
One can see that Eq. (3.40) is an example of the extension principle.
For fuzzy numbers similar to the one in Fig. 3.5, we would use
max-min convolution to get the sum of their addition. For fuzzy
numbers similar to the one in Fig. 3.6, we would use a-cut to get
70
the sum of their addition. Note, however, that different addition
operations may be used interchangeably.

It has been proved by Kaufmann and Gupta [K1S] that Eqs.(3.39)
and (3.40) describe the same operation. Let x, y, z e R, then the
addition of M and N can be computed using
~ (+)N (z) max (~(x) A ~N (y». (3.41)

a a z=x+y a a
Assume that ~(x) = 1, if x e [m 1 ,m 2 ]· Otherwise, ~(x) = o.

Similarly, ~N(Y) = 1, if Y e [n 1 ,n 2 ]· Otherwise ~N(Y) = o. Thus, for
all x and y such that ~M(x) = 1 and ~N(Y) = 1, the right side of
Eq. (3.41) gives 1. If not, Eq.(3.41) gives o. And since z = x + y,
we write
(3.42)
Eq.(3.42) may be regarded as another form of Eq.(3.39).
The properties of fuzzy addition can be summarized as follows:
a. Commutative: M (+) N = N (+) M

b. Associative: (M (+) N) (+) K M (+) (N (+) K)
c. If a neutral exists at the left and the right, it

is the real number o. Thus M (+) 0 = 0 (+) M = M.
d. Nonsymmetric: M (+) (-N) = (-N) (+) M * 0 where -N

is the image of N with membership function
We shall use the following examples to show the computational

procedure of each fuzzy addition operation.
71
Example ( discrete case): Let M represent "integers close to
3" and N represent "integers close to 2," as shown in Table 3.1.
Table 3.1 Fuzzy Numbers M and N
x,y o 1 2 3 4 5 6
~(x) o .3 .8 1. .5 .1 o
o .6 1. .9 .4 o
Their addition is summarized below.
z = x+y 1 2 3 4 5 6 7 8 9 10
o .3 .6 .8 1 .9 .5 .4 .1 o
For instance, to get ~(+)N(Z=6), the possible (x,y) pairs, their

corresponding membership values, and (~M(X)A~N(Y» are:
x 0 1 2 3 4 5 6
Y 6 5 4 3 2 1 0
~M(x) 0 .3 .8 1 .5 .1
°
~N(Y) 0 .4 .9 1 .6
°
~M(X)A~N(Y) 0
° .4 .9 .5 .1
°
Thus, ~(+)N(Z=6) = max [0,0,.4,.9,.5,.1,0] = .9.
Example (continuous case): Let M represent "real numbers close
to 2" and N represent "real numbers close to 8" (see Fig. 3.7), where
72
~(X) 0, x :s 0,
x/2, 0 < X :s 2,
(4-x)/2, 2 < X :s 4,
0, X > 4,
J1 N (y) 0, y :s 3,
[ (y-3)/5, 3 < Y :s 8,
(11-y)/3, 8 < Y :s 11,
0, Y > 11.
JL(X) M N
1.
.9 '.M{+)N
.8
.7
.6
.5
.4
.3-
.2
.1
0 x
0 2 4 6 8 10 12 14 16 18 20
Fig. 3.7 The fuzzy numbers M, N, and M(+)N.
The addition of M and N is illustrated as follows. The a-cut of
M and N are:
That is, at some a level, the X can be either m1 or m2 , and y can take
either n 1 or n 2 • Thus, if we set a = x/2 for ~(X), we have a = m1 /2,

i.e., m1 = 2a. Similarly, we can obtain other a cut values as:
2« -2a+4 5a+3 -3a+11

73
The addition of M and N at a level is computed based on Eq.(3.40) as:
M + N [2a+(5a + 3),(-2a + 4)+(-3a + 11)] [7a + 3, -5a + 15].
[7a + 3, -5a + 15], then
a = (zl - 3)/7, and a = (-Z2 + 15)/5.
Consequently, we have Z {(z'~Z(z»} (see Fig. 3.7) where
0, z :s 3
(z-3)/7, 3 < z :s 10
(-z+15)/5, 10< z :s 15
0, z > 15
From the example shown above, monotonicity, convexity, and

normality are well preserved by addition. The proof can be seen in
Kaufmann and Gupta [K15].
3.4.2.3 Subtraction of Fuzzy Numbers

The definition of subtraction can also be defined by either a-cut
or max-min convolution.
(i) a-cut:
(3.43)
(ii) max-min convolution:
(3.44)
Since the image of fuzzy number N is given by
(3.45 )
74
Eq.(3.44) may be rewritten as:
~M(-)N(Z) = max (~(x) A ~N(-Y» = max (~M(x) A ~-N(Y»'

z=x+(-y) z=x+y
(3.46)
The subtraction, M(-)N, is equivalent to the addition of the image
of N to M, M(+) (-N).
Because a negative number may appear as a result of subtraction,
the commutative and associative properties cannot be preserved.
However, since M and N are fuzzy numbers, M(-)N must be a fuzzy
number (Dubois and Prade [022,026]).
Example (discrete case): Let M and N be fuzzy numbers presented
in Table 3.1. The result of M(-)N, Z, is computed as:
z -5 -4 -3 -2 -1 o 1 2 3 4 5
o o .3 .4 .8 .9 1 .6 .5 .1 o
Note that (-5,0) may be dropped from the fuzzy set since, by the
definition of a fuzzy number, any number smaller than -4 must have a
membership value of O. The computational procedure for subtraction
is the same as for addition. For example, to get ~M(_)N(z=-l), the

possible (x,y) pairs, their corresponding membership values, and
x 0 1 2 3 4 5
Y 1 2 3 4 5 6
~M(Y) 0 .3 .8 1 .5 .1
~N(Y) .6 1 .9 .4 0
~M(x)A~N(Y) 0 .3 .8 .1 0
Thus, ~M(-)N(-l) max [0, .3, .8, .4, 0, 0] .8.

75
Example (continuous case): Let M and N be fuzzy numbers

presented in Fig. 3.7. They are the same fuzzy numbers we used for
addition. The a-cut of M and N are:
2a -2a+4 5a+3 -3a+11
Based on Eq.(3.43), we have
[5a-11, -7a+1].
Consequently, the membership function ~(_)N(Z) is (see Fig.3.8):
0, x ~ -11,
(z+11)/5, -11 < x ~ -6,
(1-z)/7, -6 < x ~ 1,
0, x > 1.
lUx)
1.
-12-10 -8 -6 -4 -2 0 2 4 6
Fig. 3.8 The fuzzy number M(-)N.

76
3.4.2.4 Multiplication of Fuzzy Numbers
The multiplication of fuzzy numbers is a bit complicated because
the signs of fuzzy numbers must be considered. We shall consider the
case in which both M and N are positive fuzzy numbers, i.e.,
~(x) = 0, V x < 0
and
~N(Y) = 0, V Y < O.
Let Z be the product of the multiplication of M and N. since
gz(z) increases 'monotonically to the left of the peak (~z(z) = 1) and

decreases monotonically to the right of the peak, the multiplication
is done in the following manner (Kaufmann and Gupta [K15]):
1. At the left, we take into account all pairs (x,y) such that
xy ~ z. That is, the left leg of ~z(z) is defined as:
= max (3.47a)
xy~z
2. At the right, we take into account all pairs (x,y) such that
xy ~ z. That is, the right leg of ~z(z) is defined as:
= max (3.47b)
xy~z
3. To simplify the process, omit from consideration any (x,y)
pair where either ~(x) or ~N(Y) is zero. Conversely, we compute z
for Which ~(.)N(Z) = 1. This will show what value of z occurs when
we pass from the left to the right of the peak.
When both M and N are continuous membership functions, their
multiplication is defined as:
(3.48)
77
Eqs.(3.47) and (3.48) are equivalent. This can be easily proved (as
in the case of addition).
The properties of fuzzy multiplication can be summarized as below:
a. When both M and N have the same sign, M(o)N can also be a
positive fuzzy number.
b. Since (-M) (o)N = -(M(o)N), we know M and N can take different
signs (Dubois and Prade [022,026]).
c. The multiplication of fuzzy numbers M and N is commutative
and associative, i.e.,
M(o)N = N(o)M,
and
(M(o)N) (o)K = M(o) (N(o)K).
d. If a neutral exists at the left and at the right, it is the
real number 1 (Kaufmann and Gupta [K15]), i.e., M(o)l = l(o)M = M.

e. The inverse of M is M- 1 and M(o)M- 1 ~ 1 where
1 1
[--m-' -m--]·
2 1
Example (discrete case): Let M and N be the fuzzy numbers shown
in Table 3.1. By applying Eqs. (3.47a) and (3.47b), we can obtain
Z = M(o)N:
z 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
o .3 .6 .6 .8.8 1.9.9.9 .5 .5 .5 .4 .4 .4 .4 0
For instance, ~(o)N(6) is calculated as:
1.
78
There are no other z values where ~M(.)N(Z) = 1. Thus, for a z value

less than 6, say 4, the membership function ~M(.)N(4) can be derived as
follows. For the (x,y) pairs where xy ~ 4, we have (~M(X)A~N(Y» as:
x\y 1 2 3 4
1 .3 .3 .3 .3
2 .6 .8
3 .6
4 .S
Thus, ~M ( . ) N ( 4) = max [ . 3, .3, . 3 , .3, . 6, .8, . 6, . S] = .8. Note

that (~M(2), ~N(2», (~M(l), ~N(4», and (~M(4), ~N(l» are not the
only pairs being evaluated.
For a z value greater than 6, say 14, the membership value
~M ( . ) N (14) is derived as follows. For the (x,y) pairs where xy ~ 14,
we have:
x 4 S S
y 4 4 3
.4 .1 .1
Thus, ~M(.)N(14) =max [.4,.1,.1] = .4. Note that the pairs such as
14:Sxy
(~M(3) A ~N(S» are dropped from evaluation because ~N(S) = o.
Example (continuous case): Let M and N be presented as in Fig.

3.7. The a level sets for M and N are the same as in the addition case:
Ma [2a, -2a + 4],

and
Na [Sa + 3, -3a + 11].
79
According to Eq. (3.48), we can obtain
[(2a) (5a + 3), (-2a + 4) (-3a + 11)]
[10a 2 + 6a, 6a 2 - 34a + 44].
We now solve the following two equations,
10a 2 + 6a - z 0, (3.49)
and
6a 2 - 34a + 44 - z 0. (3.50)
The roots for Eqs. (3.49) and (3.50) are:
a = (-6 + (36 +40z) .5)/20,
and
a = (34 - (100 + 24Z)0.5)/12,
respectively. Thus, we have (see Fig. 3.9)
0, z :s 0,
(-6 + (36 + 40Z)0.5)/20, °< z :s 16,

J.lM(·)N(Z)
{ (34 - (100 + 24Z)0.5)/12, 16 < z :s 44,
0, z > 44. (3.51)
Clearly, ~(')N(Z) is still a fuzzy number even though its left and
right "legs" are no longer linear.
3.4.2.5 Division of Fuzzy Numbers
Division of two positive fuzzy numbers M and N can be defined as
follows: For the left leg of M(:)N, we have
max (~(x) A J.lN(y» Y X,y,z. (3.52a)

z?x/y
80
P.(X)M N
1.
.9
.8
.7
.6- 'M(.)N
.5
.4
.3
.2
.1 ""
00 5 10 15 20 25 30 35 40 45 50 X
Fig. 3.9 Fuzzy number M(·)N.
For the right leg of M(:)N, we have
max (~(x) A ~N(Y» V x,y,z. (3.52b)

z~x/y
If M and N are continuous membership functions, we define
(3.53)
The division operation is an extension of multiplication, i.e.,

-1
M (:) N = M(·)N- 1 where N- 1 is the inverse of N. Recall that N can
be written as:
~N(1/Y) (Dubois and Prade [022,026]), (3.54 )
or
[~, ___
1 __ ] (Kaufmann and Gupta [K15]). (3.55)
2 n1
Thus, Eqs.(3.52a) and (3.52b) can be easily revised to

81
~(o)N-l(Z) = max ('!M(x) 1\ i-LN(l/y)) = max (~(x) 1\ i-LN-1(y))

z~x(l/y) z~xy
(3.56)
and
= max max (~(x) 1\ i-LN-1(y)).

z:sxy z:sxy
(3.57)
And Eq.(3.54) can easily be written as Eq.(3.53):
1 1
[m 1 (-n--)' m2 (-n--) ] .
2 1
The computation of division is identical to that of multiplication.
Because of this, we shall omit numerical examples of division
altogether.
Generally, the real number algebraic distributivity is preserved
pretty well in fuzzy cases. It has been noted by Dubois and Prade
[022,026] and Zimmermann [Z30] that given three positive fuzzy numbers
M, Nand K,
(M (+) N) (0) K (M (0) K) (+) (N (0) K). (3.58 )
In fact, the left side of Eq.(3.58) is equivalent to (Kaufmann and
Gupta [K15]):
and the right side of Eq.(3.58) is equivalent to

82
3.4.2.6 Fuzzy Max and Fuzzy Min
Dubois and Prade [022,026]] pointed out that the fuzzy max is the
dual operation with respect to union, while the fuzzy min is the dual
operation with respect to intersection. It is easy to derive from the
foregoing statement the fuzzy max and the fuzzy min as:
(3.59)
(3.60 )
respectively, or
(3.61)
(3.62 )
respectively. Graphically, the fuzzy max and the fuzzy min are
presented in Figs. 3.10 and 3.11.
p.(x)
M N
1.
Fig. 3.10 An example of the fuzzy max.

83
JJ.(X)
M N
1.
Fig. 3.11 An example of the fuzzy min.
The properties of the fuzzy max and the fuzzy min are summarized
as follows:
a. The fuzzy max and min are commutative and associative
operations.
b. Distributive: Let M, N, and K be fuzzy numbers, then
min [M, max(N,K)) max [min(M,N),min(M,K)),

and
max [M, min(N,K)) min [max(M,N) ,max(M,K)).
c. Absorption: Given fuzzy numbers M and N,
max [M, min(M,N)) M,

and
min [M, max(M,N)) M.
d. De Morgan's Law: Given fuzzy numbers M and N, then
1 - min (M,N) max [l(-)M, l(-)N),
1 - max (M,N) min [l(-)M, l(-)N).

84
e. Idempotence:
max (M,M) = M min (M,M)
f. Given fuzzy numbers M, N, and K,
M (+) max(N,K) = max [M(+)N,M(+)K).
The same property holds true for the fuzzy min.
g. max(M,N) (+) min(M,N) M(+)N.
Example (discrete case): Let M and N be fuzzy numbers presented

in Table 3.2.
Table 3.2 Fuzzy Numbers M and N
x,y 1 2 3 4 5 6
~(x) 0 .7 1 .4 .2 0
JlN(y) 0 .3 1 .6 0
The fuzzy max of M and N is computed as:
z = xvY 1 2 3 4 5 6
Jlmax(Z) 0 .3 1 .7 .2 0
For instance, to obtain Jlmax (Z=3), the (x, y) pairs that satisfy
3 = xvY are:
x 1 2 3 3 3
y 3 3 3 2 1
85
The corresponding ~ and ~N' and their minimums are:
~(x) 0 .7 1 1 1
~N(Y) 1 1 1 .3 0
~(x) " ~N(Y) 0 .7 1 .3 0
Thus, ~max(3) max [0,.7,1,.3,0] 1.
The fuzzy min of M and N is computed as:
z = x"y 1 2 3 4 5 6
o .7 1 .4 o
The computional procedure for ~min(z) is the same as that of ~max(z)
except that the fuzzy min uses z = x"y but the fuzzy max uses z = xVY.
Example (continuous case): Let M and N be fuzzy numbers
presented in Fig. 3.12. By taking the a-cut, we have
Ma [4a + 1, 8 - 3a],
and
Na [a + 2, 9 - 6a].
Based on Eq.(3.59), their maximum is defined as:
[(4a+1) V (a+2), (8-3a) V (9-6a)].
By changing the a value, Eq.(3.59) may yield different results. That
is, when 0 s a s .33 , we get Ma (V) Na = [a+2, 9-6a].

86
''(x)
1.
.9-
.8
.7
.6
-\" N
0
0
M
'\" TjfUZZY max

".
.5 "'-
....
.4 ""
.3- "
.2
.1
0 x
0 2 3 4 5 6 7 8 9 10
Fig. 3.12 Fuzzy numbers M and N.
Similarly, when .33 ~ a S I, we get Ma (v) Na [4a+l, 8-3a).
Thus, the computed membership function is
~(V)N(Z) 0, z ~ 2
(z-2)/l, 2 < z S 2.33
(z-1)/4, 2.33 < z ~ 5
(8-z)/3, 5 < z S 7
(9-z)/6, 7 < z ~ 9
0, z > 9.
This membership function is illustrated in Fig. 3.12 by the dashed

line. Similarly, the fuzzy min can be obtained and illustrated as the
dotted line in Fig. 3.12.
3.4.3 Special Fuzzy Numbers

In many cases, the use of extension principle operations on fuzzy
numbers tends to be cumbersome. This argument is clearly supported by
the numerical examples presented earlier. (Remember examples
87
presented are binary operations). Thus, special fuzzy numbers are

proposed to reduce the amount of computational effort. So far,
triangular numbers (Laarhoven and Pedrycz [Ll]), trapezoidal numbers
(Buckley [B36,B39]), L-R triangular fuzzy numbers (Dubois and Prade
[026]), and L-R trapezoidal fuzzy numbers (Dubois and Prade [026],
Bonissone [B26,B27]) have been applied to various decision models.
Figs.3.l3a and 3.l3b present some special fuzzy numbers. We can
interpret the fuzzy number M with a unique peak as a fuzzy quantity
"approximately m," and a trapezoidal number may be seen as a fuzzy
quantity "approximately in the interval of [m l ,m2 ]."
3.4.3.1 L-R Fuzzy Number (Dubois and Prade [026])

A function denoted by L or R is a reference function of fuzzy
numbers iff (1) L (x) = L (-x); (2) L (0) = 1; (3) L is nonincreasing

on [0,+111).
A fuzzy number M is said to be L-R type iff
L«m-x/a),xsm,a>O,
~(x) { (3.63)
R«x-m/~),x~m,~>O,
where m is the "mean" of fuzzy number M and a, ~ are the left and
right "spreads", respectively. When a = ~ = 0, M is considered a
crisp number m. It is often written as (see Fig.3.l3a)
M (m, a, ~).
If the peak is not unique, the L-R number M has a flat region.
It can be written as (see Fig. 3.l3a):
88
P.lx) . M M'
1....................................... .
Fig. S.l3a L-R type fuzzy numbers.
P.lx) M M'
1.' ..................................... .
Fig. S.l3b Triangular and trepezoidal fuzzy numbers.
3.4.3.2 Triangular (or Trapezoidal) FUzzy Number (Kaufmann and
Gupta [KlS], Laarhoven and Pedryzy [Ll])
Let x, 1, m, u e R. A triangular fuzzy number M is defined as
(see Fig. 3.l3b):
~(x) = 0, X :s 1,
(x - 1) / (m - 1), 1 < x s m, (3.64)
(u - x) / (u - m), m < x:s u,
0, x > u.
89
In Fig. 3.13b, M = (1, m, u) with 1 and u being the lower and upper
bounds of fuzzy number M.
When there are multiple peaks, fuzzy number M is represented by
M (a, b, c, d)
with the [b,c] interval being the most likely values for M and any
value below a and above d being totally impossible. The membership
value decreases gradually (or linearly) from b to a and from c to d.
It is clear that the triangular (or trapezoidal) number is a
more restricted form than the L-R fuzzy number, in that all "legs"
must be linear. Furthermore, we find that
M (l,m,u) (m,ex,/3)
where ex m - 1 and /3 u - m. Similarly, we find that
M' (a,b,c,d) (b,c,ex,/3)
where ex = b - a and /3 = d - c. The characteristics of M and M' in
Figs. 3.13a and 3.13b remain the same.
We have four different special forms of fuzzy numbers. Each has
its own algebraic operation formulas. Table 3.3 summarizes the
algebraic operations for L-R triangular numbers. Table 3.4.
summarizes the algebraic operations for L-R trapezoidal numbers.
Tables 3.5 and 3.6 summarize the algebraic operations for triangular
and trapezoidal fuzzy numbers, respectively.
3.4.3.3 Proof of Formulas (Dubois and Prade [026])
Since the L-R fuzzy number (m,ex,/3) is a special case of the L-R
trapezoidal number (m 1 ,m 2 ,ex,/3) (when m1 = m2 ) and the triangular (or
trapezoidal) numbers are a special case of L-R type fuzzy numbers

90
Table 3.3. Algebraic Operation: M =(m, a, ~), N (n, '1, 0)
Image of N : -N = (-n, 0, '1) (3.65 )
Inverse of N: N- 1 (n -1 , on -2 , '1n
-2
) (3.66 )
Addition: M (+) N (m+n, a+'1, ~+o) (3.67)
subtraction: M (-) N = (m-n, a+o, ~+'1) (3.68)
Multiplication:
M > 0, N > 0: M ( . ) N (mn, m'1+na, mo+n~) (3.69)
M < 0, N > 0: M ( . ) N (mn, na-mo, n~-m'1) (3.70)
M < 0, N < 0: M ( . ) N (mn, -n~-mo, -na-n'1) (3.71)
Scalar Multiplication:
k > 0, k e R: k ( . ) M (km, ka, k~) (3.72)
k < 0, k e R: k ( . ) M (km, -k~, -ka) (3.73)

Division:
(!!! mo + na m'1 + X~)
M > 0, N > 0: M ( : ) N n (3.74)
n2 n2
m na - m'1 n~ - mo )
M < 0, N > 0: M ( : ) N (ii' (3.75)
n2 n2
M < 0, N < 0: M ( : ) N
m
(ii'
-n~ - m'1 -na - mo)
) (3.76)
2
n2 n
91
Table 3.4. Fuzzy Operations for M = (a,b,a,~), N (c,d,7,05)

(Bonissone [B26,B27])
Image of N: -N = (-d, -c, 05, 7) (3.77)

-1 1 1 05 7
Inverse of N: N = (a' C' d(d+05)' C(C-7)') (3.78)
Addition: M (+)N = (a+c, b+d, a+7, ~+05) (3.79)
subtraction: M (-) N = (a-d, b-c, a+05, ~+7) (3.80)

Multiplication:
M > 0, N > 0: M ( . ) N (ac,bd,a7+ca-a7,b05+d~+~05) (3.81)

M < 0, M > 0: M ( . ) N (ad,bc,da-a05+a05,-b7+C~-~7) (3.82)
M < 0, N < 0: M ( . ) N (bd,ac,-b05-d~-~o,-a7-ca+a7) (3.83)
Division:
a b ao + da b7 + c~
M > 0, N > 0: M ( : ) N (d' C' d(d+05)' C(C-7» (3.84 )
a b ca - a7 d~ - b05
M < 0, N > 0: M ( : ) N (c' d' (3.85)
C(C-7)' d(d+o»
b a -b7 - c~ , -a05 - da
M < 0, N < 0: M ( : ) N (c' d' d(d+05) ) (3.86)
C(C-7)
92
Table 3.5 Fuzzy Operations for M =(l,m,u), N (a,b,c)
Image of N: -N = (-c, -b, -a) (3.87)

1 1
Inverse of N: N- 1 (c' h' !)
a (3.88)
Addition: M (+) N (l+a, m+b, u+c) (3.89)

Subtraction: M (-) N = (l-c, m-b, u-a) (3.90 )
Scalar Multiplications:
'" k > 0, k e R: k ( . ) M (kl, km, ku) (3.91)
'" k < 0, k e R: k ( . ) M (ku, km, kl) (3.92)

Multiplications:
M > 0, N > 0: M ( . ) N (la, mb, uc) (3.93)
M < 0, N > 0: M ( . ) N (lc, mb, ua) (3.94 )
M < 0, N < 0: M ( . ) N (uc, mb, la) (3.95 )

Division:
I m
M > 0, N > 0: M ( : ) N (c' h' ~)
a (3.96)
u m
M < 0, N > 0: M ( : ) N (c' h' .!)
a (3.97)
(a' h' .!)

u m
M < 0, N < 0: M ( : ) N c (3.98)
93
Table 3.6 Fuzzy operations for M
(3.99 )
Inverse of N: N- 1 (3.100)
Scalar Multiplication:
'<J k > 0, k E R: k ( . ) M (ka 1 ,kb 1 ,kc 1 ,kd 1 ) (3.103)
'<J k < 0, k E R: k ( . ) M (kd 1 ,kc 1 ,kb 1 ,ka 1 ) (3.104)
Multiplication:
M > 0, N > 0: M ( . ) N (a1b1,a2b2,a3b3,a4b4) (3.105)
M < 0, N > 0: M ( . ) N (a2d1,b2c1,c2b1,d2a1) (3.106)
M < 0, N < 0: M ( . ) N (d1d2,c1c2,b1b2,a1a2) (3.107)
Division:
a1 b1 c1 d1
M > 0, N > 0: M ( : ) N (d' c 2 , b 2 , -
a 2)
(3.108)
2
d1 c1 b1 a1
M < 0, N > 0: M ( : ) N (d' c 2 , b 2 , -
a2
) (3.109)
2
d1 c1 b1 a1
M < 0, N < 0: M ( : ) N (a' b 2 , c 2 , -
d2
) (3.110)
2
94
(when L, R reference functions are linear), we shall prove the

legitimacy of the formulas used for the L-R trapezoidal number only
without losing information. Also, the properties that are preserved
by the L-R trapezoidal number must be valid for other special fuzzy
numbers.
We shall prove Eqs. (3. 77), (3. 78), (3, 79), (3. 80), (3. 81) and
(3.84) for positive fuzzy numbers as follows. (The other formulas in
Table 3.4 may be proved in a similar manner.)
Let fuzzy numbers, M = (a,b,a,~) and N = (c,d,7,~), be
S{{a-x) / a), x s a, a > 0

"'M (x) { S ({x-b) / ~), x ~ b, (3 > 0
1 , otherwise
S{(c-y) / 7), y s c, 7 > 0

{ S«y-d) / ~), y ~ d, ~ > 0
1 , otherwise
where S(·) is the reference function that represents either L or R

function.
3.4.3.3.1 The Image of Fuzzy Number N

We know that ~-N(Y) = ~(-y). Referring to Fig. 3.14, we know
the right slope is defined as:
S «y- (-c» / 7), Y ~ -c, 7 > O.
Using the same figure, we can define the left slope as:
S «(-d) -y) / ~), -d ~ y, ~ > o.
The result is -N = (-d,-c,~,7). Note that Nand -N are symmetrical

with respect to y = o.
95
PIx)
-M 1. M
-d -b--c -a
Fig. 3.14 The image of fuzzy number M, -M.
3.4.3.3.2 The Inverse of Fuzzy Number N

-1
Since ~N-1(Y) = ~N(l/Y), the equation of right slope of N is:
Again, this is the left slope of N (a, d, 7, 5). Similarly, the
left slope of -N is:
S ( dy 5y 1.
- 1) , Y s -d
-1
Notice that N is no longer a fuzzy number. If we consider only the
neighborhood of lie, i.e.,
and the neighborhood of lid, i.e.,
dy - 1 ~ 1 5
5y (y - d I d(d+5»'
96
-1
then N can be approximated as
1 1 c'l 7
(a' ~, d(d+c'l)' C(C-7))·
3.4.3.3.3 Addition and Subtraction
The addition of M and N may be obtained as follows: Considering
left slopes first, let x and y be two unique real numbers, such that
S«a-x)/a) w S«c-y)/-r)
where w is a fixed value in [0,1]. This is equivalent to
x a - a S-l(W),
which implies
-1
z = x + y a + c - (a+7) S (w)
where S( (a+c) - z) = w. The same reasoning holds true for the right
a + 7
slopes of M and N, where
S(z - (b+d) w.
(Hc'l
The result is a new fuzzy number, M(+)N, with
IlM (+)N(Z) ! S(
S(
1,
(a+c)-z)
a+7 '
z ~ a+c, a+7 > 0,
otherwise.
97
which makes M(+)N = (a+c,b+d,a+r,~+5).
Using the identity M (-) N = M (+) -N where -N is defined in part
(a). We can use the same reasoning to obtain the formula for
subtraction:
(a,b,a,/3) (-) (c, d, r, 5) (a-d, b-c, a+r, /3+r).
3.4.3.3.4 Multiplication and Division
Using the same reasoning as for addition, when M > 0 and N > 0,
we determine the left slope to be:
-1
z = xy ac - (ar + ca) S (w) + ar (S -1 (w» 2
.
One may neglect the second-order equation in S-l(w) when a and rare
relatively small (compared with a and c), and/or w is in the
neighborhood of one, thus giving an approximation formula
(a,b,a,/3) (0) (c,d,r,5) (ac,bd,ar+ca ,b5+d/3)
However, since we cannot guarantee that z and r are actually very
small and that w is close to one, we need to use another approximation
formula to give the rough shape of ~M(o)N(z). This approximated

formula may be defined as:
M(o)N _ (ac, bd, ar + ca - ar, b5 + d/3 - 5/3).
The membership function defined on the right hand side of Eq. (3.81)
coincides with ~M(o)N(Z) in at least three points: (ac,l),
«a-a) (c-r) ,S(l», and «b+/3) (dH), S(l». Similarly, when M < 0,
98
N > 0 and M < 0, N < 0 the membership functions ~(')N(Z) may be
approximated as those shown in Table 3.4.
The approximate formulas for M (+) N can also be derived using
the identity
M(:) N
where lIN is the inverse of (N defined in part (b». The formula for M
> 0 and N > 0 is then given as:
a b acS + do: bl' + c(3

M(: )N
(d' c' d(d+cS)' c(c-l'»
The formulas for M < 0, N > 0 and M < 0, N < 0 can all be derived in
a similar manner.
Remarks
It is interesting to note that Eqs.(3.77) through (3.86) are
very consistent and preserve many useful properties. Some equations
may preserve certain properties very well but some other equations may
preserve other properties very well. The equations are summarized as:
Properties Equations
M (-) N = M + (-N) Eqs. (3.77), (3.79)
1
M (+) N M (.) N Eqs. (3.78), (3.81), (3.82), (3.83)
-1 1
M = M
Eq.(3.78)
M2 =M(')N Eq. (3.80)

(M"l) (.) (M"2) = M"1+H2 Eq.(3.79)
( M"l ) n2 = M"lxn2 Eqs. (3.81) , (3.82) , (3.83)

99
However, since the inverse elements for addition and
multiplication are missing, the following identities are only valid
for fuzzy numbers:
M (+) (-M) (a-b, b-a, a+~, ~+a) * 0 (3.111)
a b a~+ba
, a~+ba
M > 0
(
b' a' b(b+~) a(a-a»'
M (.) (!)
M { a' b a
h'
-a~-b
a(a-a)
, -a~-b
b(b+~»'
M < 0
(3.112)
3.5 Conclusions
Fuzzy sets are used to model imprecise or vague concepts in the
human world in the mathematical domain. with the help of fuzzy
set-theoretic operators, one can combine various fuzzy sets and derive
some specific answers which cannot be solved using classical
mathematical models.
A fuzzy number is a fuzzy set with its elements confined in real
line~. It is a natural way to incorporate classical mathematical
concepts with fuzzy (vague) quantities using Zadeh's [Z4,Z6,Z7]
extension principle. However, we find it extremely cumbersome to
apply it to algebraic operations (even with the help of the a-cut
technique). Thus, a new breed of fuzzy numbers, with some restrictive
properties such as normality and convexity have been introduced by
various researchers. In general, these special fuzzy numbers satisfy
our needs of modeling decision problems. For each type of special
fuzzy number, a number of algebraic operation formulas are provided.
These algebraic formulas generate results which in many cases, are
only approximations. Thus, they add more uncertainty (or fuzziness)
into the original problem. The question of tradeoff between accuracy

100
and simplicity is no easy question to answer. The DM must choose

between accuracy, i.e., using the extension principle and regular
fuzzy numbers, and simplicity, i.e., using the special fuzzy number
and approximation formula.
However, we feel that from a practical viewpoint, simplicity
may carry more weight since most of the real world problems are large
in size and complex computational procedures cannot handle them at a
reasonable cost.
IV. FUZZY RANKING METHODS
4.1 Introduction
Recall that in Chapter 2 we define the Multiple Attribute
Decision Making (MADM) problem as:
Xl X2 Xn
A1 x 11 x 12 x 1n
[
A2 x 21 x 22 x 2n
D
Am x m1 x m2 x~ 1
~ (w 1 , w2 '···, Wn )
where Ai' i 1, ... , m are possible courses of action (or alterna-
tives); Xj , j = 1, ... , n are attributes with which alternative
performances are measured; x ij is the performance (or rating) of
alternative Ai with respect to attribute Xj ; wj , j = l, .•• ,n are the
relative importance of attributes.
Traditionally, MADM solution methods assume all x ij ' Wj values

are crisp numbers. A utility function U(x 1 ,x 2 , ... ,xm) is implicitly
or explicitly defined by the decision maker. For Ai' the utility
function aggregates its performance ratings x ij ' vj, into a final
rating, Ui . Such a final rating represents how well one alternative
satisfies the decision maker's utility. The alternatives with higher
final ratings are said to be preferred by the decision maker. Since
the final ratings are real numbers, it is straightforward to compare
final ratings.
In reality, the alternative performance rating x ij can be crisp,
fuzzy, and/or linguistic. For example, three candidates are
considered for a professor position. The attributes used are
creativity (Xl)' maturity (X 2 ), communication skill (X 3 ), and number

of publications (X 4 ). The performance rating for the first three
102
attributes cannot be numbers. They are represented by linguistic

terms such as "good," "average," "poor," and so on. The performance
rating for the fourth attribute can be some integer numbers. This
MADM problem contains a mixture of fuzzy and crisp data. Most of the
real world problems are of this type.
When fuzzy data are incorporated into the MACM problem, the final
ratings are no longer crisp numbers; they are fuzzy numbers. Since a
fuzzy number represents many possible real numbers that have different
membership values, it is not easy to compare the final ratings to
determine which alternatives are preferred. In other words, fuzzy
numbers do not always yield a totally ordered set as real numbers
do. In MACM applications, when the final ratings are fuzzy, it is
very difficult to distinguish the best possible course of action from
the mediocre ones, or even the worst one.
To resolve this problem, many (Lee and Li [L3,L7], Nakamura [N2],
Dubois and Prade [D27,D29], Jain [J2,J5], and others) have proposed
fuzzy ranking methods that can be used to compare fuzzy numbers. This
is one of the promising research fields which fuzzy set theory offers.
The study of fuzzy ranking began in the early '70s. Since then over
two dozen methods or theories have been proposed. In 1980, Freeling
[F1] systematically classified existing fuzzy ranking methods
according to the techniques used for comparison. Freeling's five
basic categories are: (a) methods using extended maximum; (b) methods
using implication logic; (c) methods using preference relations; (d)
methods of direct comparison and; (e) linguistic approaches. Nakamura
[N2] classifies existing methods as either fuzzy implication methods
or direct comparison methods. This classification tends to be broad
but indicates at least part of the developing trend in fuzzy ranking
methods. Bortoland and Degani [B34] provided not only a systematical
review of fuzzy ranking research but also gave results of comparisons
103
among various ranking methods. Lee and Li [L3,L7] and McCahone [Mc1]
broadly classified ranking methods into mathematical approaches versus
linguistic approaches. The mathematical approaches are further sub-
divided into methods of possibility theory and methods of probability
theory. Similar remarks can also be found in Dubois and Prade [D29].
Recently, Tseng, Klein, and Leonard [T13] classified fuzzy ranking
methods based on the technique(s) used in each method. There are
methods using Hamming distance, fuzzy boundaries, centroid index,
possibility dominance, or probability proportions. Tseng et al.'s
classification combined the advantages of Freeling's [F1] and Lee and
Li's [L3] classifications. Tseng et al. [T13] also gave some general
guidelines as to when to use which method(s) to generate better
ranking results.
In this chapter, some 20 ranking methods are classified into
three major classes according to the means (or media) each method
uses. The readers should refer to the taxonomy shown in Fig. 4.1.
There are preference relation methods, a fuzzy mean and spread method,
fuzzy sC0ring (or direct comparison) methods, and linguistic methods.
Each main class is further divided according to the techniques used.
For instance, methods using degree of optimality is a subclass of

preference relation, methods using centroid index is a subclass of
fuzzy scoring, and methods using linguistic approximation is a
subclass of linguistic methods. We shall review each method in the
following sections. Numerical examples are also given at the end of
each section.
Basic Notations
1. Ai' i = l,2, ... ,m: There will always be m number of

alternatives in a MADM problem.
2. Xj , j l,2, ... ,n: There will always be n number of
attributes in a MADM problem.

I.Comparison Medium II.Technique Involved III.Approaches
Baas and Kwakernaak [B1J

Degree of
optimality Watson et al. [W3J
Baldwin and Guild [B3J
Hamming Yager [Y11J

Preference distance Kerre IK17J
Nakamura I N2J
relation K Kolodzijezyk [K27J
a-cut
Comparison
function
Probability Dubois and Prade [029J

Fuzzy distribution Tsukamoto et al. [T1SJ
ranking
rl Fuzzy mean Delgado et al. (04)
and spread
Proportion to ~ Lee and Li [L3J
~ ~
optimal
~
Left/right "'-I McCahone [Mc1J
scores
Fuzzy scoring I~ Jain IJ2.JSJ
~ Chen [C12J
Centroid Indexl
Chen and Hwang [C13J
Area - Yager [YSJ

measurement Murakami et al. [M23J
Yager [Y18J
Intuition
linguistic Efstathiou and Tong [ESJ
expression Linguistic
approximation Tong and Bonissone IT11J
L-
Fig. 4.1 A taxonomy of fuzzy ranking methods.
105
3. X ij ' Vi,j: The performance rating for alternative Ai with
respect to attribute Xj . x ij can be crisp or fuzzy.
4. wj , j = 1,2, .•. ,n: The relative importance of attributes.
Wj can be unknown, crisp, or fuzzy.

5. Ui , i = 1,2, ... ,m: There will be m number of final ratings
for comparison in a MADM problem. Ui can be a crisp number, a fuzzy
number, or a linguistic expression.
All the fuzzy ranking methods presented in this chapter assume
that fuzzy final ratings Ui , vi, are readily available for comparison.
The manner in which the fuzzy final ratings were derived shall be
presented in Chapter 5.
Example Set
We have systematically collected 15 ranking problems. They are
used to demonstrate the computational procedure of each method, to
show the merits and/or weaknesses of each method, or help explain
hard-to-understand concepts or notations. There are 13 examples
involving continuous fuzzy sets and two examples of discrete fuzzy
sets. Each example is illustrated with a figure.
(1) Example 1: There are two fuzzy numbers, M and N. All
ranking methods favor Mover N. This is an obvious fact. We use
it primarily to show the definition of the Hamming distance and to
illustrate a fuzzy maximum, and a fuzzy intersection.
(2) Example 2: There are two fuzzy numbers, M and N. This
example demonstrates a situation in which fuzzy ranking becomes an
important issue because it is not clear whether "M > Nil or liN > M."
(3) Example 3: This example is another case in which it is not
clear which fuzzy set is better. Based on our intuition, it seems
that N is better than M. However, some methods, such as Baas and
Kwakernaak's [B1] method, may result in a counter-intuitive solution.

106
P.(x)
N M
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.1a Example 1.
p.(x)
N' M
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.1b Example 2.
p.(x)
M
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.1c Example 3.

107
(4) Example 4: This is an example in which most of the methods
cannot discriminate between M and N. However, Lee and Li's [L3] and
McCahone's [Mel] methods give N > M, based on the assumption that our
intuition favors fuzzy sets with higher means and smaller spreads.
/L(X)
N M
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 ~~--~--~-+~~~~~--~--~~x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.ld Example 4.
(5) Example 5: When two fuzzy sets share the same "right leg",
some methods, such as Jain's [J2,J5] method, cannot tell whether M
or N is better. Yet, it is obvious that N is better than M.
/L(X)
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 ~~--~--~~--~--~~--~--~~ x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.le Example 5.

108
(6) Example 6: There are three fuzzy sets, M1 , M2 , M3 , in this

example. Fuzzy set M3 is crisp number 1 and by common sense should be
ranked first. However, some methods, such as Yager's [Y10] method,
ranks M3 as the worst. This ranking result is anti-intuitive.
/L(x) M1 M2 M3
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.1f Example 6.
(7) Example 7: This is another example which shows that the

methods which use Hamming distance measurement to determine the
ranking order of fuzzy sets may generate counter-intuitive results.
/L(x) M1 M3
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.1g Example 7.

109
(8) Example 8: This example compares three fuzzy sets. It is
used primarily to demonstrate the computational procedures of some
ranking methods.
/L(x) U2 U3 U1
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.1h Example 8.
(9) Example 9: This is a case where all fuzzy sets have the same
spread.
/L(x)
M1 M2 M3
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0
0 .1 .2 .3 .4 .5 .6 .7 .8 .9
Fig. 4.li Example 9.

110
(10) Example 10: The fuzzy sets U1 ' U2 , and U3 are from
Laarhoven and Pedrycz [L1]. This example is primarily used to
demonstrate the computational procedures of some ranking methods.
p,(x)
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 ~~~~~~~--~~~-4--~--~~x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.1j Example 10.
(11) Example 11: In this case, four fuzzy sets are compared.
This example is from Bonissone [B27] and is used to demonstrate the
algorithm of Bonissone's [B27] method.
P,(x)
Ml
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.1k Example 11.

111
(12) Example 12: This example is used to demonstrate the logical
problem of McCahone's [Mc1] method. The fuzzy set M3 is, obviously,
better than M2 but McCahone's method cannot distingush between them.
P,(x)
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 ~-+--~~~~--4---~-+--~~~~ x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.11 Example 12.
(13) Example 13: This is an example to show that ranking
methods, such as Adamo [A1] and Buckley and Chanas's [B40) methods are
best for screening. Any fuzzy set whose peak is lower than the a
level set by the DM should be eliminated. The remaining ones are then
ranked by other methods.
p,(x)
~ M3 M5
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.1m Example 13.

112
(14) Example 14: The fuzzy sets in this example are discrete.
This example is used to demonstrate the computational procedure for
computing the Hamming distance.
P.(x)
1. M •
.9 N *
.8
.7
.6
.5
.4
.3
I
.2
.1
0 ~-+--~~~-+--~--~-+--~--r-~x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.1n Example 14.
(15) Example 15: This is an example which demonstrates the
comparison of discrete fuzzy sets using Mabuchi's [M1] method.
p.(x)
1. M 0
.9 N *
.8
.7
.6
.5 *0
.4
.3
.2
.1
0 x
0 2 3 4 5 6 7 8 9 10
Fig. 4.10 Example 15.

113
Remark
Many of the methods presented here are not perfect. But they
have shown the process of human efforts to find ways to solve
problems. There are always some good points coming out of each
method. We wish to recognize the effort and process of improvement.
A flawless ranking method may possibly be obtained by combining some
of these good ideas into one algorithm.
114
4.2 Ranking Using Degree of Optimality

This group of methods computes an index of preference for a set
of fuzzy sets. A fuzzy set with higher preference index value is
preferred. Baas and Kwakernaak [B1] first determines the best fuzzy
set and then establishes a preference index which denotes the degree
of preferability of this fuzzy set over other fuzzy sets. Baldwin and
Guild [B3] proposed the use of a two-dimensional preference relation
for each pair of fuzzy sets in question. The preference relation is

then used to derive the degree of optimality for each fuzzy set. The
fuzzy set with the highest degree of optimality is ranked first.
utilizing fuzzy implications, watson, Weiss, and Donnell [W3]
developed an index of strict preference for each pair of fuzzy sets.
4.2.1 Baas and Kwakernaak's Approach
Baas and Kwakernaak [B1] proposed a ranking method that first
identifies a set of preferred fuzzy sets. A preference score, P, is
derived for each preferred fuzzy set. The preference score denotes
how much better a particular fuzzy set is in relation to all other
fuzzy sets. A fuzzy set with a higher preference score is considered
better than one with a lower score.
Algorithm
Given normal fuzzy sets M.1 = {(r.,~_

1 'M i
(r.»},
1
i = l, ... ,m, where
r 1. takes its number from the real line ~ and ~_

'M i
(r.)
1
E [0,1], Baas and
Kwakernaak use the following steps for ranking:
step 1. A conditional fuzzy set rlu = {(i'~rlu(ilr1, ... ,rm»} is

defined first, where r = {l, ... ,m} and U = {real number}. The
membership function is defined as:
115
1, if r i ~ r j , Vj E I,
{ (4.1)
0, otherwise.
This conditional membership function illustrates that fuzzy set Mi is

considered the best if and only if r i ~ r j , Vj E I. Note that this
conditional fuzzy set is nonfuzzy because ~IIU is either 0 or 1.
step 2. The set of preferred fuzzy sets, 0, is defined as:
(4.2)
or Eq. (4.2) may be rewritten as:
sup { min [~ (r.)] }. (4.3)

r 1 ,···rm j J
ri~rj,VjEI
This membership function indicates the degree to which the ith fuzzy
set is ranked first. For example, in Fig. 4.2 ~0(1) = 0.5, the number
0.5 characterizes the extent to which M1 is ranked first while ~0(2)
1, the number 1 characterizes the extent to which M2 is the best.

Obviously, M2 is better than M1 •
If there is a unique Mi for which ~o(i) = 1, then Mi dominates
crisply all other fuzzy sets and the algorithm terminates. However,
when either or both the following situation(s) occur, step 3 must be
performed to derive the preference score(s). The conditions are:
(1) the OM wants to know how much better Mi is over other Mj'
Ifj ~ i,
(2) which fuzzy set is better when there is more than one fuzzy
116
IL(X) M M
1 ............................ t ......... 2
. 1Lo(2)
.9
.8
.7
.6 IL. (1)
.5 ···0···- ................. .
.4
.3
.2
.1
o ~~--~-4--~~~~--+_~--4_~x
o .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.2 Example of ~O(i), i = 1,2.
step 3. The preference score, P i = {(v.,~p (v.»}, of M. is

1 i 1 1
defined as:
~p (v.) sup min (4.4)

i 1
r l ' . . . , rm j
v i =h(r 1 ,···,rm)
where
m
L (4.5)
j=l
j"i
The membership function ~p (v.) characterizes the preferability of Mi

i 1
over all other fuzzy sets. When the number of fuzzy sets to be
compared is small, (less than 10), step 3.1 may be used. Otherwise,
step 3.2 is recommended. Note that either step 3.1 or step 3.2 should
give the same result.
step 3.1
The membership function ~p (v.) is not directly obtainable when

i 1
~M. (r i ), vi, are piecewise continuously differentiable functions and
1
117
h(r1, ••• ,rm) is also continuously differentiable. This difficulty may

be resolved by using the a-cut.
Let a o be a predetermined membership value of ~p (v.), i.e.,
i l.
~p (v.) = a. The corresponding greatest lower bound and smallest
i l. 0
upper bound of vl." [v. ,v. ], are then computed as follows. We

l.min l.max
are to find the appropriate ~l' ~2' ... '~m values such that
(4.6)
For fuzzy set Mi , there may be more than one ~i value that satisfies
Eq.(4.6). For example, for Ml in Fig.4.3, the ~l values that satisfy
~ (r 1 ) a o are 4.5 and 5.5; while when ~2 = 6.5 and 7.5, we have
1
~M (r 2 ) ao ·
2
To obtain the greatest lower bound and the smallest upper bound
[v. ,v. ], we simply apply Eq.(4.5) on all possible
l.min l.max
P.lx)
1.
.9
.8 ~Q •••....•••••••.••••••
.7
.6
.5
.4
.3
.2
.1
o x
o .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
•
Fl.g. 4.3 The rA 1 , rA 2 values at a o .
118
(~l' •.• '~m) combinations. This enables us to identify the lowest

value and the highest value as V imin and v imax ' respectively. The
process of identifying [v. ,v. can be easily performed on a

l.min l.max
computer for cases of no more than 10 fuzzy sets.
step 3.2 The advantage of using a computer can quickly vanish as
the number of fuzzy sets to be compared increases to a larger number,
such a"s 20. The number of possible (~1'··· '~m) combinations
. 220 , Le., 1,048,576. To avoid such tedious computations, Baas and
l.S
Kwakernaak [B1] proposed the following. Let
1lM. (r j ) d~M
j
(r.) /dr. , Vj .. i,
J J
J
and
~. (~i) d~
i
(r.),
l.
l.
If we can find a (~1' ... '~ ) combination such that ~ (~.), Vj .. i,

m j )
"-
all have the same signs, while ~u (r.) has the opposite sign. Then
i l.
that set of combination (~1' ... '~ ) will give either v. or v.
m l.min l.max
By setting various ~o values, we can construct an approximated
membership function ~p (v.). Each ~p (v.) denotes the preference
i l. i l.
score of fuzzy set Mi.
Numerical Example
The described algorithm will be applied to Example 10. The
fuzzy final ratings to be compared are: U1 = (.20,.30,.50), U2 = (.17,
.32, .58), and U 3 = (.25, .40, .70).
step 1. The conditional fuzzy set is computed based on Eq.(4.1)

as:
119
P.(x) U, U2 U3
1.
0(3)
.9
.8
.7
0(1)
.6
.5
.4-
.3
.2
.1
0 X
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.4 Three final fuzzy ratings.
i 1 2 3
1 1 1
For example, in Fig. 4.4, the variable r 1 may take values from the
interval [0.20,0.50], Le., r 1 may be .20, .25, .30, .35, .40, .45,
.50, or other real numbers in [0.20,0.50]. Similarly, we know r 2 can
take any real number in [0.17,0.58] and r3 can take any real number in
[0.25,0.70]. Since there is at least one r 1 that is greater than some
r 2 and some r 3 , e.g., r 1 = 0.50, r 2 = 0.20, and r3 = 0.17, we know that
~Ilu(l) = 1 (based on Eq. (4.1».

Step 2. Using Eq.(4.3), the set of preferred fuzzy final ratings
is obtained as:
i 1 2 3
.71 .81 .1
For fuzzy final ratings U1 ' Eq.(4.3) gives

120
Jl o (1) sup .71.

r 1 ,r 2 ,r 3
r1~r2,r3
In Fig. 4.4, the dashed line indicates the intersection of ~ (r.), Vj.
j J
That is, the dashed line represents
min [~(r.)l, j=1,2,3,

j j J
which satisfies r 1 ~ r 2 ,r 3 . The highest point of the dashed line
gives Jl O (1) = .71. Jl O (1) can also be obtained by taking the
intersection of the right leg of ~ (r 1 ) and the left leg of ~ (r 3 ).

1 3
Based on Jlo(i) values, we obtain the ranking order: U3 > U2 > U1 .
step 3. Assume that the DM is interested in determining how much
better U3 is over other fuzzy final ratings. This may be answered by
deriving fuzzy set (v 3 ,Jl p (V 3 )). This fuzzy set may be seen as the
3
preference score of U3 .
since all three fuzzy final ratings are piecewise linear and the
number of fuzzy sets to be compared is only 3, the fuzzy preference
score shall be obtained using step 3.1. We shall set a o = .50 as the
first a level, i.e., set Jl p (v 3 ) = 0.5. We are to find some proper

3
1\ 1\ 1\
(r 1 , r 2 , r 3 ) combinations such that v 3 . and v3 can be obtained.
m~n max
The possible combinations of ~1' ~2' and ~3 are shown in Fig. 4.5.
The v3 is identified as:

max
.55 - [.25 + .245]/2 .3025.
Similarly, we can identify v = -.1050.

3 min
121
A A
r2 r3 V3
.25
[.245 [
.325
.55
.0775
.3025 (= V3
max
)
.325 .0200
.46 [ .55 .2450
.40
[.245 [
.325
.55
.0025
.2275
.325 -.1050 (= v3 . )
.46 [ .55 .1200
m1n
Fig. 4.5 The possible (~1'~2'~3) combinations and v3 at a o =0.5.
The next a level is set at a o = 1. Again, we want to find

A A
two combinations of (r 1 , r 2 , r 3 ) such that v 3 . and v3 can be

m1n max
obtained. Since (.30, .32, .40) is the only combination at a o = 1,
the v 3 . and v3 happen to be on the same point. Based on
m1n max
1
Eq.(4.5), we obtain v3 = .40 - 2 (.30 + .32) = .09.
r-~--~--r-~--4---r-~--~--r--4 x
-.5 -.4 -.3 -.2 -.1 0 .1 .2 .3 .4 .5
Fig. 4.6 The gp (V 3 ) function.

3
122
Given the v3 values at «0 = 0.5 and «0 = 1, we can approximate

~p (v 3 ) as shown in Fig.4.6. Normally, we would have to set more than
3
two « levels to derive the approximate ~p. since all ~u. (r i ) are
~
piecewise linear in this numerical example, ~p (v.) should preserve

i ~
that property. Hence, the three v values can be used to determine
~p (v 3 ). This graph may be interpreted as "U 3 is generally but not

3
universally preferred over other final ratings." (Note: to obtain a
natural language expression for a fuzzy set is an important research
topic. Interested readers should see Bonissone [B24].)
For demonstration purposes, we calculate v3 values at «0 = 0.5,

using step 3.2. That is, instead of computing all possible v3 values
first and finding the minimum and maximum among them, we want to
identify those (~1' ~2' ~3) combinations with which
and ~U(~l) and ~U(~2) have the same sign and ~U(~3) has the opposite
1 2 3
sign.
For example, when (r 1 , r 2 , r 3 )

1\ 1\ 1\
= (.250,.245,.325) we have
~U (r 1 ) d~u (r 1 )/dr 1 10,

1 1
~U (r 2 ) 6.6,
2
~U (r 3 ) 6.6.
3
since ~U (r 3 ) has the same sign as ~U (r 1 ) and ~U (r 2 ), this

3 1 2
(~1'~2'~3) combination is not an appropriate one for deriving either
v3 . or v3
m~n max
(.25,.245,.55), we then have

123
10 6.6 -3.3
since ~u (r 3 ) does not have the same sign as ~u (r 1 ) and ~u (r 2 ), this

3 1 2
(~1'~2'~3) combination is an appropriate one for deriving one of the
v3 extremes. It is computed as:
V3 = .55 - [.25 + .245]/2 = .3025.
step 3.2 is completed when both v 3 . and v3 have been found.

ml.n max
Note
1. McCahone [Mc1] found that when ~o(i), Vi, are piecewise
linear, ~o(i) can be calculated by finding the highest membership
function value associated with the points on the x-axis where a fuzzy
set's membership function values exceeded the others. For example, in
Fig. 4.2, ~o(l) = 1 because the value 1 is the highest membership
function value associated with the points on the x-axis where MI'S
membership values exceed the membership values of M2 . The ~o(i) index
compares fuzzy sets based solely on the locations of their peaks.
This may cause an illogical result (See Note 3).
2. In the numerical example presented earlier, all ~U (x.) are
i l.
piecewise linear. The ~o(i) index, vi, can be calculated by taking
the intersection of the "left leg" of one alternative with the "right
leg" of another. The result is U3 > U2 > U1 • This is the same as
completing Eqs.(4.I) through (4.3).
3. Baldwin and Guild [B3] pointed out that Baas and Kwakernaak's
method may generate counter-intuitive results. For example, in Fig.
4.7a MI is ranked first by this method but our intuition would favor
M2 because M2 's mean value is greater than the mean value of MI.
Baas and Kwakernaak's method also does not have high
124
discriminating ability. Lee and Li [L3] pointed out (see Fig. 4.7b)
that Ml and M2 (which share the same peak), are considered equivalent
when Baas and Kwakernaak's method is used. However, Lee and Li
suggest that our intuition would favor the fuzzy set which has a
larger mean and a smaller spread. In this case, M2 would be preferred
because it has a smaller spread. Similar criticism about Baas and
Kwakernaak's method may be found in Baldwin and Guild [B3], Bortolan

and Degani [B34], and McCahone [Mel].
/L(X)
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 ~-+--~--~-+--~~~~--~--r-~x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.7a Counter-intuitive case of Baas and Kwakernaak's method.
/L(X)
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 ~~--~--~~--~--r--T--~--r--+x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.7b Indiscrimination case of Baas and Kwakernaak's method.

125
4.2.2 watson et al.'s Approach
watson et al. [W3] envisioned the ranking problem as the modeling
of the implication:
x => Y (4.7)
where
X: " Fuzzy set 1 is M1 and fuzzy set 2 is M2 " with
min
Y: " M1 is strictly preferred to M2 " with
1,
{
0, otherwise.
Eq. (4.7) is equivalent to the following. Given fuzzy sets M1 =
{(X1'~M (xl»} and M2 = {(x2'~ (x 2 »}, the strict dominance of M1

1 2
over M2 is defined as:
~(X -> Y) min {1 - min [~M (Xl)' ~M (x 2 )]), (4.8)

x1:sx2 1 2
or
~(X -> Y) 1 - max min [~M (Xl)' ~M (x 2 )]· (4.9)

x1:sx2 1 2
Numerical Example
watson et al.'s method is applied to Example 10. The
computational procedure is described as follows. For U1 and U2 ' we
obtain
~(X -> Y) 1 - .81 .19,
where
126
X: " FUzzy set 1 is U1 and fuzzy set 2 is U2 " with
min
Y: " U1 is strictly preferred to U2 " with
1,
{
0, otherwise.
Similarly, we obtain all the pairwise comparisons using Eq.(4.9) as:
X1 \X 2 1 2 3
1 .19 o
ll(X - > Y) 2
3
[ .29
.29 .19
o
1
Let Zij e [0,1] be an element of the matrix. It denotes the strict
dominance of Ui over Uj. When Zij = 0, this means that Ui would never
dominate Uj' while Zij = 1 shows that Ui dominates Uj' crisply. Based
on Zij' Vi,j, the ranking order is determined to be: U3 > U2 > U1 .
Note
Watson et al.'s method is just another form of Baas and
Kwakernaak's [B1] method, since
ll(X - > Y) (4.10)
where X : "fuzzy set M. and fuzzy set M.", Y: "fuzzy set M; strictly
~ J ~
dominates Mj ", and llo(j), (from Eq.(4.3» represents the degree of Mj
being ranked first. Because of the similarity of the two methods,
they share the same problems discussed in section 4.2.1.

127
4.2.3 Baldwin and Guild's Approach

Baldwin and Guild [B2] pointed out that Baas and Kwakernaak's
method results in an anti-intuitive result (as discussed in Section
4.2.1). To avoid such difficulty, they defined a two-dimensional
fuzzy preference relation P ij to indicate the degree to which a real
number xi is preferred over another real number x j , Vj ~ i, as:
{ « xi' x J' ), IJ. p (x. , x . ) ) }, vj ~ i, (4.11)

ij 1 J
where IJ. p (x.,x.) varies according to the OM's attitude toward risk.
ij 1 J
When the OM is risk-averse, we have
IJ. p (x. , x.) (4.12)

ij 1 J
When the OM is risk-neutral, we have a linear function
IJ. p (x. , x.) (4.13 )

ij 1 J
When the OM is risk-prone, we have
IJ. p (x. , x.) (4.14 )

ij 1 J
For simplicity, we shall assume the OM is risk-neutral, i.e.,

Eq. (4.13) will be used throughout this section.
Given fuzzy sets M.= {(x.,IJ.M (x.»} and MJ.
1 1 i 1
preferability of Mi over Mj is:
IJ.O.(i) = sup min (4. 15)

J xi ,x j
128
which is similar to that of Baas and Kwakernaak's ~o(i) defined in

Eq. (4.3).
When there are m fuzzy sets M1 , M2 , •.• , Mm' Eq.(4.15) will be used
on each (Mi,M j ) pair, and then combined using the min operator. Thus
m
min ~o. (i) , (4.16)
j 1 )
j '" i
which represents the preferability of Mi over its best rival.

By combining Eqs.(4.15) and (4.16), we have:
m
~O(i) min
j 1
j '" i
Eq.(4.17) may be rewritten as:
~ (x.)] 1\ (4.18 )
j )
The membership function ~o(i) is not directly obtainable when

~ (x.) and ~ (Xi,x).) are piecewise continuously differentiable
i ~ Pij
functions. To resolve the computational difficulty, Baldwin and Guild
[B3] proposed the following. In general, the evaluation of ~o(i)
requires the solution of (m-1) sets of simultaneous equations:
z ~. (Xi) ,
~
Z ~.(Xj)'
)
Z ~P .. (Xi,X j ),
~)
129
and then
J1 0 ( i) min (Z).
j
If the membership functions J1M (x.), ~ (x.), and J1 p (x.,x.) are all
i ~ j ) ij ~ )
piecewise linear functions which may be characterized by Fig. 4.8,
then the simultaneous equations become
0 - x.~
Z
0 - '1
J1 M • (xi»
~
,
Xj - a
Z J1 M (x.»,
f3 - a j ]
Z x. - x. J1 p (x. , x . ) ) ,
~ ) ij ~ )
which gives the solution
Z
o - a (4.19 )
1 + (o-'r) + (f3-a) ,
and then
min (Z). (4.20)

j
/L(X)
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0
0 Ci
Fig. 4.8 Characterization of linear fuzzy sets Mi and Mj .

Numerical Example
Baldwin and Guild's method is applied to Example 10. The three
fuzzy final ratings are U1 = [.20,.30,.50], U2 = [.17,.32,.58], and
U3 = [.25, .40, .70]. Their ranking order is determined as follows.
Assume that the DM is risk-neutral, i.e., ~p (x.,x.) is a

ij ~ )
linear function (defined in Eq.(4.13». The fuzzy ratings are
piecewise linear as well. Since all membership functions are piecewise
linear, we can use Eqs.(3.19) and (3.20) to derive ~o(i), i = 1,2,3,

as:
i 1 2 3
.18 .28 .36
The ranking order is: U3 > U2 > U1 .
We shall detail the computation of ~O(l) to show the computa-

tional procedure. For fuzzy final ratings U1 and U2 , Eq.(4.19) will
give (see Fig. 4.9a):
6 - a .5 - .17
.24.
1 + (6-7) + (~-a) 1 + .2 + .15
For U1 and U3 , Eq.(4.19) will give (see Fig. 4.9b):
.5 - .25
1 + .2 + .15 .18.
Based on Eq. (4.20) ~o(l) is determined to be:
min (Zl)') .18.

j=2,3
That is, the degree to which U1 is ranked first is only 0.18.

131
P.(x) U1 U2
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
OJ-....,..-'"-I----;...:.....-........---'I--'-r--r----r-........-...x
a 'Y{3
Fig. 4.9a The fuzzy final ratings U1 and u 2 .
P.(x) U1 U3
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1 .
0 x
Fig. 4.9b The fuzzy final ratings U1 and u3 •
Note
1. The Up (x.,x.) function represents the OM's attitude toward
ij ~ J
risk. Recall that when the OM is risk-neutral, Up (x.,x.) is
ij ~ J
defined by Eq.(4.13). Subsequently, the solution Uo(i), Vi, is
computed based on Eqs.(4.19) and (4.20).
When the OM takes a risk-averse attitude, Up (x.,x.) is defined
ij ~ J
by Eq.(4.12). The solution Uo(i), Vi, is computed using
132
Z (& - Z(& - 1»1/2 - (a + Z(~ _ a»1/2, (4.21)
which may be solved using Newton's method.
When the DM takes a risk-prone attitude, ~p (x.,x.) is defined

ij 1 J
by Eq.(4.14). The solution ~o(i), vi, is computed using
z2[(~_a)2 _ (&_1)2] + [Z(l + 2&(&-1) + 2a(~-a)] + a 2 ~ &2 - 0,

(4.22)
which is a quadratic equation in Z and can be solved analytically.
The ranking order for the same set of fuzzy sets may be
different, depending on the DM's attitude toward risk. Generally, a
risk-averse DM tends to fa~or fuzzy sets whose supports do not extend

to very low values of x even though very high values may also be
present. A risk-prone DM will favor fuzzy sets whose supports contain
very high values even though very low x values are present.
2. Recall that, in Baas and Kwakernaak's method the construction
of ~p (v.) (the preference index) is done using the a-cut and by trial
i 1
and error to find appropriate ~i' vi combinations such that v. and

1min
v are located on the horizontal-axis. By setting different a o
i max
values, we can approximate the ~p (v.) function.

i 1
In Baldwin and Guild's method, however, trial and error is not
used. The solution to ~o(i) is obtained by solving (m-1) sets of
simultaneous equations, thus eliminating the arbitrariness found in
Baas and Kwakernaak's method.
3. In Fig. 4.7a, Baas and Kwakernaak's method gives M1 > M2 .
This does not comply with our intuition. The anti-intuition case can
be resolved by Baldwin and Guild's method which computes ~o(i) as:

133
i 1 2
.10 .31
The ranking order is M2 > M1 • This complies with our intuition.

In Fig. 4.7b, we would favor M2 over M1 because Lee and Li [L3]
suggest that human intuition would favor fuzzy sets of higher means
and smaller spreads. Using Baas and Kwakernaak's method, M1 = M2 •
This shows the method's lack of discriminating ability. Baldwin
and Guild's method results in:
i 1 2
.33 .28
The ranking order is M1 > M2 . This result is counter-intuitive.

Thus, Baldwin and Guild's method is not logically sound either.
134
4.3 Ranking using Hamming Distance

This family of ranking methods involves the measurement of the
Hamming distance between two fuzzy sets. The Hamming distance between
fuzzy sets M and N is defined as:
+..,
d(M,N) .f I ~(x) - Il N(X) I dx (4.23)
-..,
for continuous functions, and
(4.24)
for discrete functions. If M and N are continuous functions as shown

in Fig. 4.10, then the shaded areas represent the Hamming distance
between M and N.
:!.l...._ _ _ _ x
x·
Fig. 4.10 The Hamming distance between fuzzy sets M and N.
Yager [Y11] defines a fuzzy maximum first and then computes the
Hamming distance between each fuzzy set and the fuzzy maximum. The
fuzzy set with the smallest distance from the fuzzy max is considered
the best. Kerre [K17] follows Yager's ranking concept -- except that
he defines a fuzzy max differently from Yager. Nakamura [N2] derives
135
fuzzy minimums from the greatest lower set and the greatest upper set,
respectively. The Hamming distances for a fuzzy min from the greatest
upper set and a fuzzy min from the greatest lower set are computed. A
fuzzy order relation, p, between fuzzy sets M and N is then
constructed. The membership value ~p(M,N) denotes the weighted sum of
M's unique superiority over N for the best and worst possible
situations. Kolodziejczyk [K27] extends Orlovsky's fuzzy preference
relation [02] and the notion of Hamming distance to construct several
preference indices, P1' P 2 , and P 3 • These indices follow very closely
with Kerre's index and Nakamura's preference relation.
Notice that in Fig. 4.10 the Hamming distance is the summation
of the areas not commonly occupied by both M and N. The methods in
this category compare fuzzy sets with the fuzzy max (the fuzzy min)
using the Hamming distance and the fuzzy set that has shortest
(longest) distance to the fuzzy max (the fuzzy min) is considered
better. The ideas of comparison are good. However, since the
comparison is based solely on area measurement and the fuzzy set's
relative location on the x-axis is ignored, the logic of the methods
in this category is not sound.
For example, let us consider fuzzy sets M1 , M2 , and M3 in Fig.
4.11a. Keen observation and common sense indicate M3 > M2 > M1 .
Yager's method, which defines a unique fuzzy max for all comparison
cases, would result in: d(max,M 3 ) = 0.5, d(max,M 2 ) .452, and
d(max,M 1 ) = .436. The ranking order is M1 > M2 > M3 . This ranking
order is against human intuition. The problems with Yager's method
are that crisp numbers will always be ranked the lowest, and a fuzzy
set that overlaps more with Yager's fuzzy max will get a higher rank,
regardless of their relative locations on the x-axis.
Kerre's [K17] method defines a fuzzy max which is problem-
dependent. The fuzzy max in Fig 4.11a is M3 (by Kerre's definition).
136
P.(x) M1 M2 M3
1.
.9
.8
\
Yager's fuzzy max .'
..........
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.11a A demonstration of weakness of Yager's method.
p.(x)
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0
0 .1 .2 .3 .4 .5 .6 .7 .8 .9
Fig. 4.11b A demonstration of weakness of Kerre's method.
The Hamming distances are: d(max,M 3 ) = 0; d(max,M 2 ) = area occupied by
M2 ; and d(max,M 1 ) = area occupied by M1 . The ranking order is M3 > M2

> M1 , which complies with human intuition. Kerre's method for this
example looks better than Yager's approach. However, let us consider
another set of fuzzy numbers as shown in Fig. 4.11b. Human intuition
would favor M2 over M1 . Kerre's method gives the result Ml > M2 .
This result is against our intuition. This counter-intuition case
shows that Kerre's method would favor a fuzzy set with smaller area
137
measurement, regardless of its relative location on the x-axis.
The illogicality observed in Yager's and Kerre's methods, i.e.,
ignoring fuzzy sets' relative location on x-axis, can be addressed to
Nakamura's and Kolodziejczyk's methods as well.
4.3.1 Yager's Approach
Yager [Y11] proposed a ranking procedure for fuzzy sets of unit
interval. First of all, a fuzzy maximum of the fuzzy sets to be
ranked is determined. Then, each fuzzy set is compared with the fuzzy
maximum using the Hamming distance measurement. The fuzzy set(s) that
have the smallest Hamming distance to the fuzzy max are ranked as the
best.
Yager's fuzzy max is defined as {(x'~max(x»} with membership

function
x, (4.25)
where x and ~max(x) are in [0,1]. Fig. 4.12a illustrates Yager's
fuzzy max for discrete functions, and Fig. 4.12b shows Yager's fuzzy
max for continuous functions.
Once the fuzzy max is defined, we shall use the Hamming distance
to measure the closeness of each fuzzy set to the fuzzy max. For
example, given two fuzzy sets M and N as shown in Fig. 4.13, the
Hamming distance between the fuzzy max and M is calculated as:
d(max,M) 10 - 01 + 1.1 - 01 + 1.2 - 01 + 1.3 - .31 + 1.4 - .Sl

+ 1. 5 - 11 + 1.6 - .SI + 1. 7 - .31 + I·S - 01
+ 1.9 - 01 + 11 - 01 = 4.5.
similarly, the Hamming distance between the fuzzy max and N is
d(max,N) = 6.0. Since d(max,M) is smaller than d(max,N), M > N.

138
p.(x)
1.
.9
.8 Yager's fuzzy max
.7
.6
.5
.4
.3
:~o ,I
t--+-+--+--4-~-+-~'---f-~~-f. X
o .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
,0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.0
.0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.0
Fig. 4.12a Yager's fuzzy maximum for discrete functions.
P.(x)
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0
0 .1 .2 .3 .4 .5 .6 .7 .8 .9
Fig. 4.12b Yager's fuzzy maximum for continuous functions.

139
p.(x)
1. M •
.9 N *
.8
.7
.6
.5
.4
.3
.2
.1
0 f x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
x, 0 .1 .2 .3 .4 • 5 .6 .7 .8 .9 1
J-LM(X) 0 0 0 .3 .8 1 .8 .3 0 0 0
J-LN(y) 0 .2 .8 1 .8 .3 0 0 0 0 0
Fig. 4.13 Discrete fuzzy sets M and N.
This ranking procedure can be applied to fuzzy sets with a
piecewise linear membership function. Given fuzzy set M and Yager's

fuzzy max as depicted in Fig. 4.14a, the Hamming distance between
Yager's fuzzy max and M can be seen as designated by the shaded areas.
It is calculated using Eq.(4.23) as:
1
d(max,M) f I J-Lmax(x) - ~(x) I dx
o
.375 .375 3 .5 3
f
o
x dx - f
.3.
(~) dx + f
2
«~)
.375 .2
- x) dx
.587 .70 7
+ f «~) - x) dx + f (x - (~» dx
.50. 2 .58.2
1.0
+ f x dx .433
.7
similarly, the Hamming distance between Yager's fuzzy max and fuzzy
set N in Fig. 4.14b is d(max,N) = .50. The result is M > N.
140
JJ.(X)
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.14a The Hamming distance between Yager's fuzzy max and M.
JJ.(x)
N M yager'\~uzzY m~~
1.
.9 " ,,,,, ,,,, .
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.14b Yager's fuzzy max and fuzzy sets M and N.
JJ.(X)
N' M
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 ~~--~~~~--~~~~--~--~-Tx
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.14c Yager's fuzzy max, M, and N'.

141
When N becomes N' as shown in Fig. 4.14c, the Hamming distance of
N' to Yager's fuzzy max is d(max,N'} = .403. comparing d(max,M} =

.433 with d(max,N'} = .403, the smaller Hamming distance to Yager's
fuzzy max indicates that N' > M.
Numerical Example
Given three fuzzy final ratings u1 ' u2 ' and U3 as shown in Fig.
4.15, the Hamming distances between Yager's fuzzy max and u 1 ' u 2 ' and
U3 are calculated using Eq. (4.23) as:
d(max,U 1 } = .48, d(max,U 2 } = .47, and .42.

The resulting ranking order is: U3 > U2 > U1 .
P,(x)
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1 -
0 ~~~~~--~~~T-~--T--T--~X
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.15 Fuzzy final ratings U1 , U2 , and U3 •
Note
1. Yager's approach is against common sense. Let us consider
fuzzy sets with only a single element in them. For example, there are
two fuzzy sets M and N defined as:
M {(.2,.2}} N {(.2,.8}}.
Our intuition suggests that N > M, because N represents a higher

degree of confidence for x = 0.2.
142
If Yager's method is used, the Hamming distances from Yager's

fuzzy max to M and N, respectively, are:
d(max,M) = 5.3, d(max,N) 5.9.
The fuzzy set M is preferred. This result is against our intuition.

2. Yager's index is not logically sound. For example, let us
consider three fuzzy numbers M1 , M2 , and M3 as shown in Fig. 4.16.
The fuzzy number M3 is a crisp number 1.0. The Hamming distance from
M3 to Yager's fuzzy max is d(max,M 3 ) 0.5. For the fuzzy numbers M1
and M2 , we get d(maX,M 1 ) < d(maX,M 2 ) < 0.5. Based on Yager's index,
the resulting ranking order is: M1 > M2 > M3 • This is contrary to the
obvious fact that M3 > M2 > M1 •
Lee and Li [L3] pointed out that human intuition would favor a
fuzzy number with the following characteristics: higher mean value and
at the same time lower spread. The fuzzy final rating M1 in Fig. 4.16
possesses none of the two characteristics. Thus, the ranking order
M1 > M2 > M3 is against human intuition.
P.(x)
M1 M2 M3
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.16 Illustration of the logical weakness of Yager's method.

143
4.3.2 Kerre's Approach

Kerre's [K17] ranking procedure is the same as Yager's except for
the definition of the fuzzy max. Let M and N be two fuzzy sets, their
fuzzy max is defined as {(z'~max(z»} where
~max(z) = sup [~(x) A ~(y)]. (4.26)

z=(Xyy)
We shall use the fuzzy numbers M and N presented in Fig. 4.13 to show
the derivation of ~max(z). For z = 0.6, the possible (x,y) pairs that
satisfy 0.6 = (x y y) are
x .6 .6 .6 .6 .6 .6 .6 .5 .4 .3 .2 .1 0
Y 0 .1 .2 .3 .4 .5 .6 .6 .6 .6 .6 .6 .6
The corresponding ~(x), ~N(Y)' and ~(x) A ~N(Y) are summarized as:
~(X) .8 .8 .8 .8 .8 .8 .8 1 .8 .3 0 0 0
~N(Y) 0 .2 .8 1 .8 .3 0 0 0 0 0 0 0
~(x) A ~N(Y) 0 .2 .8 .8 .8 .3 0 0 0 0 0 0 0
Thus, ~max(Z = .6) = sup [ (~(x) A ~N(Y»] 0.8.
We can follow the same process to get ~max(Z) for other z values.
The resulting ~max(z) is (see Fig. 4.17) :
x,y,z .1 .2 .3 .4 .5 .6 .7 .8 .9 1.0
~max(z) 0 0 .3 .8 1 .8 .3 0 0 0
~(x) 0 0 .3 .8 1 .8 .3 0 0 0
~N(Y) .2 .8 1 .8 .3 0 0 0 0 0
144
p.(z)
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 ~~--~~--~--~-+--4-~--~~z
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.17 The fuzzy max for fuzzy sets M and N in Fig. 4.13.
The Hamming distance between the fuzzy max and M is calculated

using Eq.(4.24) as d(max,M) = o. similarly, we can obtain
d(max,N)= 3.5. The resulting ranking order is M > N.
When ~(x) and ~N(x) are piecewise linear, their fuzzy max can be
shown as illustrated by the dotted line in Fig. 4.18. The fuzzy max
~max(Z) can be obtained in the following manner. For z = 0.4, the

possible (x,y) pairs that satisfy 0.4 = (x V y) are summarized as:
x .4 [.3, .4]
y [.1,.4] .4
z =x V y .4 .4
Consequently, we obtain
~(x) .5 [0, .5]
~N(Y) [0,1] .5
s up ( (x) 1\ ~ (y» .5 .5
N
That is, sup (~M(.4) A ~N(Y» = 0.5, and sup (~N(.4) A ~(x» 0.5.
Thus, ~max(Z=.4) = max {.5,.5} = .5.

145
P,(x)
N M
1.
.9 The fuzzy max 000
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.18 The fuzzy max for continuous fuzzy sets M and N.
For z = 0.45 and z = 0.55, respectively, the possible (x,y) pairs

are summarized as:
7-1 .45
[.10, .45]
1[.30, .45]
.45 7-1
.55
[.10,.50]
Their corresponding membership values are:
/olM(X) .75 [0,.75] /olM(X) .75
/olNey) [0,1] .25 IlN (y) [0,1]
sup (IlM(x) "/olN(Y» .75 .25 sup (~(X) " Il N (y» .75
Thus, Ilmax (Z=.45) = .75 and Ilmax (Z=.55) = .75, respectively. We can
use the same procedure to get /olmax(Z) for other z values. In this
case, the fuzzy max is the fuzzy set M.
The Hamming distance between the fuzzy max and fuzzy set M,
d(max,M), is 0; while the Hamming distance between the fuzzy max
and fuzzy set N, d(max,N), is represented by the shaded area in Fig.
4.10. It is obvious that M > N.
146
P,(x)
N' M The fuzzy max 000
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 "'YI'---.---r X
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.19 The fuzzy max for continuous fuzzy sets M and N'.
When N becomes N', the fuzzy max (see Fig 4.19) is determined
according to the same procedure described above. The Hamming distance
of M to the fuzzy max is defined by the shaded areas in Fig. 4.19. It
is calculated using Eq.(4.23) as
.4 3 .7 8 7 .8 8
d(max,M) J (x-.) dx + J (. -x - . -x) dx + J (. -x) dx
.3. 2 .6. 4 •2 .7 .4
.05
Similarly, we can obtain d(max,N') = .03. The Hamming distance from
N' to the fuzzy max is smaller, hence, N' > M.
Numerical Example
Kerre's ranking method is applied to Example 10 (see Fig. 4.1j).
The fuzzy max for u 1 ' u 2 ' and U3 is the fuzzy final rating u 3 . Since
d(max,U 3 ) = 0, U3 is the best choice.
Note
Referring to Fig. 4.16, the fuzzy max is the fuzzy set M3 . The
Hamming distance from M3 to the fuzzy max is 0; the Hamming distances
from M3 to M1 and M2 are the areas occupied by M1 and M2 , respectively.

147
Kerre's method gives a ranking order M3 > M2 > M1 • Recall that

Yager's method gives a counter-intuitive ranking order Ml > M2 > M3 •
Kerre's method seems better than Yager's method.
However, Kerre's method is not logically sound either. Let us
consider the fuzzy numbers presented in Fiqure 4.11b. In this case,
the fuzzy max is the fuzzy set M3 • Since the area occupied by Ml is
smaller than M2 , Kerre's method would result in a smaller Hamming
distance between Ml and the fuzzy max. Therefore, Ml > M2 , which is
against the obvious fact that M2 > M1 •
4.3.3 Nakamura's Approach

Nakamura [N2) utilizes the concepts of a fuzzy minimum and the
Hamming distance measurement to define a preference index for a pair
of fuzzy numbers (M,N). The fuzzy numbers which have a longer
distance to the fuzzy min are considered better. Nakamura defines two
fuzzy min and four Hamming distance measurements. The aggregation of
four Hamming distance measurements gives a unique preference index
~p(M,N), which characterizes the degree of M > N.
Definitions
To understand the derivation of the two fuzzy min and the four
Hamming distances, the following definitions are given.
Fuzzy Min: Given two fuzzy numbers, M and N, in a universe, U,
the fuzzy min is defined as
sup [~(x) A ~N(Y»)' x,y,z e U. (4.27)

z=xAy
The computational procedures for ~min (Eq.(4.27» and ~max (Eq.(4.26»

are the same except that z = x A y for ~min and z = x v y for ~max.
When both ~M and ~N are continuous membership functions, ~min can be

depicted as in Fig. 4.20.
148
lL(x)
N M The fuzzy min 000
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.20 The fuzzy min for continuous fuzzy sets Hand N.
Greatest Lower Set (GL) and Greatest Upper Set (GU):
Let H be a fuzzy set in universe U. A fuzzy set H- is said to be

the Greatest Lower set, (GL), of H, iff
~-(x) sup ~(y), V x,y e U. (4.28)

y~x
We shall use Fig. 4.21 to demonstrate how Eq.(4.28) works. Given a
point Xo = 0.37 in Fig. 4.21, there will be some y values on the

x-axis, such that y ~ x o , (i.e., y = [.37,1.0] ~ xo). Each y value
determines a ~(y) value. Among the ~(y) values, their maximum gives
~-(Xo). In this case,
[.37,1.0]) 1.0.
By setting different XO points, we can obtain ~-(x).
A fuzzy set H+ is said to be the Greatest Upper set, (GU) , of H,
iff
149
~+(x) sup ~(y), V x,y E U. (4.29)

y:sX
Consider the point XO in Fig. 4.21 again, there will be some y values
such that y :s xo. In this case Xo 0.37, Y = [0,.37] :s xo. Each y
determines a ~(y) value. Among the ~(y) values, their maximum gives
~+(Xo). In this case,
[0,.37]) .33.
By setting different Xo points, we can obtain ~+(x).
p.(x)
M
1.
.9 GU of M - M+ _"''''''''
.8
.7
GLofM-M - 000
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.21 GL and GU of fuzzy set M (i.e., M+ and M-).
The fuzzy sets M- and M+ are presented in Fig. 4.21. The fuzzy
set M- may be seen as the right leg of M plus some ~-(x) = 1 for
x :s x*; while the fuzzy set M+ consists of the left leg of M plus some
~+(x) = 1 for x ~ x*. The intersection of M+ and M- gives the fuzzy
set M itself. The GU and GL will be used in measuring the Hamming
distance.
150
Fuzzy Min and Hamming Distance
In order to obtain Nakamura's [N2] preference index, two fuzzy
mins and four Hamming distances are defined first. Given two fuzzy
sets, M and N (see Fig. 4.20), we can get M , M , N , and N- by using

+ - +
Eqs. (4.28) and (4.29). Nakamura defines two fuzzy mins: min(M+,N+)
(see Fig. 4.22a) and min(M-,N-) (see Fig. 4.22b). The four Hamming
distance measurements are:

(a) d(M-,min(M-,N-». This is interpreted as the unique advantage
of Mover N, given M at its best possible state.

(b) d(M+,min(M+,N+». This is seen as the unique advantage of M
over N, given M at its worst possible state.
(c) d(N-,min(M-,N-». This is seen as the unique advantage of N
over M, given N at its best possible state.
(d) d(N+,min(M+,N+». This is interpreted as the unique advantage
of N over M, given N at its worst possible state.
The results of the four Hamming distance measurements are (see
Fig. 4.22c):
Fuzzy Preference Relation

A fuzzy relation for two fuzzy numbers M and N is characterized
as:
~p(M, N ) (4.30)
1
2 , if Kw = 0,
151
where
w [d(M+,min(M+,N+» + d(N+,min(M+,N+») + (4.31)
(l-W)[d(M-,min(M-,N-» + d(N-,min(M-,N-»).
M+_ N+ ... mln(M+,N+) 000

,.,,(x)
'i9•• O••Q _ 9 •• 0 .•ct_ 't.O _.D. _ 9 ••Q _ .D•• Po_
1.
d
.9
.8
.7
.6
.5
.4
.3
.2
.1
o ~~--~--~-T--~--~~--~--~~x
Fig. 4.22a The fuzzy min of M+ and N+.
,.,,(x)
1. m"1nl"1ITlnnnnl"1nl"'--'
.9
.8 d,
.7
.6
.5
.4
.3
.2
.1
o ~~--~--~~--~--~-+--~--~~x
Fig. 4.22b The fuzzy min of M- and N-.
The membership function ~p(M,N) will represent the proportion of the
weighted combination of the M's unique advantages over N for the best
and worst state, to the sum of such weighted combinations of the M's
advantages over Nand N's advantages over M. It is an aggregated

152
J.L(x)
N M
1.
.9·
.8
.7
.6
.5
.4
.3
.2-
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.22c The Hamming distance measurements.
index which combines all four Hamming distance measurements to show
the degree of M > N.
Referring to Fig. 4.22c again, the degree of preference of Mover
N is determined to be
The decision maker is to supply a w value to show the relative weight
of "the M and N comparison based on their best states" and "the M and
N comparison based on their worst states." When w is set to 1, the
preference relation, P, is related to the MAXIMIN selection criterion.
If w = 0, the preference relation is related to the MAXIMAX selection

criterion.
Theoretically, we would be required to calculate ~p(N,M) as well.
When ~p(M,N) 2: ~p(N,M), we can conclude that "M 2: N"; otherwise, "M <
N." Fortunately, Nakamura's [N2] ~p index has an important property:
which reduces the calculation requirements by half.

153
Even with this property, the calculation of ~p(M,N) is still very

time consuming which makes Nakamura's method less practical. To
resolve this difficulty, Nakamura has proved that, given two normal
and convex fuzzy numbers, M and N, whose membership functions are
piecewise continuous, we can obtain
d(M,min(M,N» + d(N,min(M,N» d(M,N) . (4.32)
Subsequently, Eq (4.32) is used to modify Eq. (4.31) as
K'w (4.33)
which in turn simplifies Eq.(4.30) to
[w d(M+,min(M+,N+) + (l-w) d(M-,min(M-,N-»] /
[w d(M+,N+) + (l-w) d(M-,N-)], K~ ~ 0,
1 K' O. (4.34)
2"' w
Note that fuzzy numbers M and N must be normal and convex in
Eq.(4.34).
Numerical Example
Three fuzzy final ratings in Example 10 are to be compared. They are
summarized in Table 4.1.
Table 4.1 Fuzzy Final Ratings
Final Ratings
Fuzzy Numbers (.20,.30,.50) (.17,.32,.58) (.25,.40,.70)

154
since all final ratings are fuzzy triangular numbers, we can
apply Eq.(4.34) to calculate ~P(U1,U2)' ~P(U1,U3)' and ~P(U2,U3).
As for ~P(U2,U1)' ~P(U3,U1)' ~P(U3,U2) we shall apply the property
When Ui = Uj' the Hamming distance measurement becomes zero. This

1
will make ~P(Ui,Ui) or ~p(Uj' Uj) = 2. Assuming the w value is set at
0.5 (i.e., equal importance of the worst and the best possible state),
we shall compute ~P(U1,U2) in detail to show how Eq.(4.34) works. By

observing Figs. 4.23, we know that
o.
p.(x) U
1.
.9 <12'-
.8
.7
.6
.5
.4
.3
.2
.1
Of--~-'-+-":""""~-.----'\--':""'--.,....---..--.---,.x
Fig. 4.23 The Hamming distances in Nakamura's method.
To compute the Hamming distance, we calculate 51' 52' 53' and
54' respectively, as
(.2 - .17) (Xl

.009
2
155
where a 1 = .6 is determined by taking the ~ value of the intersection
of x - .17 (left leg of U 2 ), and x - .2 (left leg of U1 ).

.15 .1
a b
(area under U1 ) - ( J x - .2 dx + J
X-.17 dX
.1 .15
.2 a
.5 .5 - x
+ J dx)
b .2
.15 - (.018 + .039 +.090) .038
where a = .26 is the x value of the intersection of the left leg of U2
and the left leg of U1 , b = .31 is the x value of the intersection of
the left leg of U2 and the right leg of U1 = (.~;x).
(.32 - .30) (1 - .95) .0005

2
.084
Given 51' 52' 53' and 54' we know that
(.5) (51) + (1-.5) 0

(.5) (5 1 +5 2 +5 3 ) + (1-.5) (5 3 +5 4 )
.0045 I (.02375 + .04225) = .068.
Similarly, we can obtain ~P(U1,U3) = 0 and ~P(U2,U3) = o.
As for ~P(U2,U1)' ~P(U3,U1)' ~P(U3,U2) we can apply the property
1.
156
As a result,
1 - .068 .932, 1, 1.
To present the result, we construct a preference relation matrix on
U1 U2 U3
0.5 .068 o
[ .932
1
.5
1
o
0.5
1
Notice that U3 dominates all the others because
Furthermore, ~p{U2,U1) is greater than ~p{U1,U2). Thus, we conclude

that U3 > U2 > U1 .
Note
1. The difference between the previous two methods in this group
and Nakamura's method is that Yager's and Kerre's methods define fuzzy
maximums and measure the Hamming distance, but Nakamura's method uses
fuzzy minimums and measures various Hamming distances with respect to
the GL and the GU.
Even though the equations for computing the Hamming distance are
different, we find that Kerre's and Nakamura's indices basically
measure the same differences. This may be demonstrated by the
following derivations. Eq.{4.33) may be extended (according to Lee
and Li [L3]) to:
d{M,min{M,N» + d{N,min{M,N» (4.35)
d{M,max{M,N» + d{N,max{M,N»
d{M,N) .
157
When we set w 0.5, Eq.(4.34) may be modified as:
~p(M,N) [d(M+,min(M+,N+» + d(M-,min(M-,N-»] / [d(M,N)]
[d(M+ ,max(M + ,N + » + d(M - ,max(M-,N

- »] / [d(M,N)]
= [d(M,max(M,N»] / [d(M,N)], (4.36)
where d(M,max(M,N» is Kerre's preference index. Eq. (4.36), however,

guarantees that the resulting ~p is in [0,1], i.e., normalized.
2. Lee and Li [L3] states that our intuition favors a fuzzy set
with a larger mean and a smaller spread. Nakamura's method fails to
comply with this intuition rule in some cases. For example, in Fig.
4.16, Nakamura's method favors M2 over M1 . However, as M2 is shifting

toward the left (as shown in Fig. 4.24), Nakamura's method cannot
distinguish between them (intuitively, one would favor M1 over M2 ).
p,(x)
1.
.9
.8 M1
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.24 The case of indiscrimination of Nakamura's method.

158
4.3.4 Kolodziejczyk's Approach
This approach [K27) is quite similar to that of Nakamura [N2).
It defines the Greatest Upper set (GU) and the Greatest Lower set
(GL). Then a fuzzy max is derived from the GU and the GL instead of a
fuzzy min (as in Nakamura's method). Some preference indices are then
defined using the GL, the GU, the fuzzy max, and the Hamming distance.
This approach defines three preference relations: Pi' i = 1,2,3.
The fuzzy numbers which are operated on by the preference relations
are convex and piecewise linear. All three preference relations
Pi(M,N) measure the degree of "M s N." When Pi(M,N) ~ Pi(N,M), we
conclude that M is worse than N because M is farther from the fuzzy
max than N is. otherwise, N is said to be worse than M. One or more
preference relations can be used to derive ranking order.
Preference Relations P 1 , P 2 and P 3
The relation P 1 is given as:
P 1 (M,N) (d«MJ"IN) ,0) + d(max(M,N) ,M) )/[d(M,O) + d(N,O») (4.37)
where d(',') is the Hamming distance defined by Eq.(4.23), and MAN and
max(M,N) are shown in Figs.4.25 and 4.26, respectively, with their
membership functions defined as:
(4.38)
Ilmax (Z) sup [IlM(X) A Il N (y»). (4.39)

z=xvy
The computation of P 1 (M,N) can be best explained by Fig. 4.26.
The Hamming distance d(MJ"IN,O) = S5 is the measurement area of the

159
"(x) N M Fuzzy Inter8ectlon 000

1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.25 Intersection of fuzzy sets M and N, (MAN).
,,(x)
N M The fuzzy max 000
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.26 The fuzzy max of fuzzy sets M and N.
intersection of M and Ni d(max(M,N),M) = 51 + 54 is the measurement

area of the difference between the fuzzy max and Mi d(N,O) = 52 + 54 +
55 is the area covered by Ni and d(M,O) = 51 + 53 + 55 is the area
covered by M. P1(M,N) is then given as:
P 1 (M,N) (4.40)
160
It follows that since
P 1 (N,M)
the property,
1, (4.41)
must hold for any pair of (M,N).
The second preference relation is an aggregated index that takes
into account the comparison of M and its fuzzy max at the worst and
the best possible states. The higher P 2 indicates that M is worse
than N. The P 2 index is given as:
P 2 (M,N) [d(M-,max(M-,N-» + d(M+,maX(M+,N+»]/ (4.42)

[d(M-,N-) + d(M+,N+)]
where M- and M+ are defined as in Nakamura's method. It is assumed
that the case when ~+(x) ~N-(Y) implies x < y is of great

importance when comparing the fuzzy numbers M and N. This means that
the greatest value of M is less than the lowest value of N at a given
membership level, presumably at the peak level. Eq.(4.42) gives

1
P(M,N) =2 directly when M =N while Nakamura's method gives P(M,N)
i (when M = N) with an additional condition (see Eq.(4.30».
Referring to Fig. 4.26 again, we have
d(M+,N+) = 5 1 + 5 2'
+ + +
d(M ,max(M ,N » = 51 + 52.
The index P 2 (M,N) is calculated as

161
The P 3 index is a modification of Eq.(4.37) with the same

assumption as that of the P 2 index.
[d(M- ,max(M- ,N- » + d(M+ ,max(M+ ,N+ ) + d(MnN,O)]

P 3 (M,N)
(4.43)
where M- and M+ are defined as before. 5ince all the Hamming distance
measurements in Eq. (4.43) have been calculated when deriving PI and
P 2 , the P 3 index is calculated as:
P3 (M,N)
One or more of the fuzzy preference relations can be used to

derive the ranking order.
Numerical Example
We have three fuzzy final ratings as U1 : (.20, .30, .50), U2
(.17, .32, .58), and U3 = (.25, .40, .70) which may be seen in
Fig. 4.1j. We will rank them using the PI index for its simplicity.
To obtain the P 1 (U 1 ,U 2 ) index, refer to Fig. 4.27. 5ince
d(U1nU 2 ,0) = 53' d(max(u 1 ,U2 ),U1 ) : 52 + 54' d(U1 ,0) = 52 + 53'
and d(U 2 ,0) = 51 + 53 + 54' we get:
.562.
The areas of 5 i , vi, are computed directly if 5 i is triangular or

through integration if 5 i is an irregular shape. For example, 51 is
directly obtained using formula ~ (hb) = .009 where h = .6 (the
162
P.(x) U, U2
1.
.9-
.8
.7
.6
.5
.4
.3-
.2
.1 -
0 X
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
intersection of the left legs of U1 and U2 ) is the height of the
triangle Sl and b .03 is the bottom of the triangle Sl' On the

other hand, using integration, S3 is computed as:
.26 .31 17 .50 5

f (X-.2) dx + f (~)d + f (~)dx .1464 .
.2 .1 . 26 .15 .31. 2
similarly, we can get S2 = .0036 and S4 = .0496. Thus, P1 = .562.

By property (4.41), we know P 1 (U 1 ,U 2 ) + P 1 (U 2 ,U 1 ) = 1, hence
P 1 (U 2 ,U1 ) = 1 - P 1 (U 1 ,U 2 ) = .438. Similarly, we obtain
.767, .333,
The preference relation resulting from the P 1 index is
U1 U2 U3
U1 .50 .562 .767
U2
U3 [ .438
.333
.50
.309
.691
.50
1
163
Recall that the higher P1(Ui ,U j ) value indicates a strong degree of

"U i :s Uj", thus U3 > U2 > U1 is the ranking order.
Note
1. The preference index P 1 is the same as Kerre's index. Recall
that P 1 (M,N) (Eq.(4.37» is given as:
P 1 (M,N) [d«MnN),O) + d(max(M,N),M)]/[d(M,O) + d(N,O)],
where d«MnN),O), d(M,O), and d(N,O) are constant for both P 1 (M,N) and
P1 (N,M). That is the P 1 (M,N) preference index can be determined using
d(max(M,N),M) which resembles Kerre's index. The denominator
guarantees P 1 falls in [0,1].
2. The preference indices P 2 and P 3 resemble Nakamura's index in
the following way. First of all, the property
1, i 1,2,3 (4.44)
holds for both methods. Secondly, the notions of the greatest lower
and upper sets and Hamming distance measurement are used.
The differences between these two methods may be summarized as
follows. .In Nakamura's method, P(M,N) denotes the proposition of "M ~
Nil with IIp representing the degree of truth of such an argument. But
in Kolodziejczyk's method, Pi(M,N), i 1,2,3, denotes the proposition
"M is not better than N," hence, larger Pi(M,N) indicates a stronger
degree of truth of "M :s N." Secondly, Nakamura's index defined by
Eq.(4.30) or Eq.(4.34) explicitly sets IIp = 21 where M N. That
notion is implicitly preserved by introducing Hamming distance d(',O)
such that even when M = N the denominator will not be zero and IIp,
1.
i, vi. Thirdly, in Nakamura's method the OM is allowed to express the

relative importance of the proposition "M ~ Nil in the best and worst
164
possible states. .
In Kolodziejczyk's method, equal importance of the
proposition "M :S Nil is assumed in the best and worst possible states.
Lastly, the fuzzy min is explicitly used in Nakamura's method while it
is only implicitly used in Kolodziejczyk's method.
3. One or more of the fuzzy preference relations can be used to
derive the ranking order. Kolodziejczyk has given various numerical
examples to show that there is no conflict among the results generated
by the three preference relations. If all three preference indices
always give the same ranking order, the usefulness of P 2 and P 3 (from
a practical viewpoint) is in doubt. Indeed, if one can easily apply
P 1 to derive ranking order, why bother with P 2 and P 3 at all?
165
4.4 Ranking Using a-cuts

This group of methods utilizes the a-cut to compare fuzzy sets
(or fuzzy numbers). An a-level set of a fuzzy set, M, is defined as:
{ x e U I ~(x) ~ a }, a e [0,1]. (4.45)
For example, given fuzzy set M as
x 123 456
.2 .5 .8 1 .7 .3
its a-level set at a .6 is
M.6 = p, 4, 5}.
If ~ is a continuous function, Ma may be shown as the interval [a,b]

in Fig. 4.28. The a-cut on the fuzzy set M includes the points
between a and b on the horizontal axis, given that ~(x) is convex.
P.(x)
M
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0
Fig. 4.28 a-level set, Ma [a,b] .
Adamo [A1] presents an a-preference index. The OM is to specify

the minimum acceptable degree, a, for a group of fuzzy sets. The
166
fuzzy sets with the higher a-cut values are considered better.
Similar notions can be seen in Buckley and Chanas [B40], with minor
modifications. Mabuchi [M1] uses the a-cut to derive the degree of
dominance of one fuzzy set over another. For each pair of fuzzy sets,
one can have either a functional type index or a single value index.
The single value index which is derived directly from a functional
type index is used for quick reference to the conclusion of
comparison. Linguistic descriptions of the comparison results are
also given.
/L(x)
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0
0 .1 .2 .5 .7 .8 .9 1. x
Fig. 4.29 Illustration of Adamo's method.
4.4.1 Adamo's Approach
Adamo [A1] uses the concept of a-cut to obtain an a-preference
index for fuzzy set Mas:
Fa(M) max { x E ul ~(x) ~ a}, a E [0,1], (4.46 )
where a is given by the OM as an acceptance threshold. Given two
fuzzy sets, M1 and M2 , as shown in Fig. 4.29, Adamo's preference index
may be derived (at a = 0.9) as:

167
.44 .52
The optimal solution is:
(4.47)
Hence M2 is the best when a 0.9.
Note
The use of the a-cut in Adamo's approach does not give reliable
ranking order. For example, referring to Fig. 4.29, when a > 0.5,
M2 > M1 ; when a = 0.5, M1 = M2 ; and when a < .5, M1 > M2 . Due to

this inconsistency, this method should not be taken seriously as a
ranking method. It should be considered as a prescreening
procedure. Given fuzzy sets as shown in Fig. 4.30 where not all
membership functions assume their peaks at 1, any fuzzy set that has
its peak lower than the a level will not be accepted. In this case,
M1 and M4 will be rejected at the a = 0.8 level, while only M2 and M3

are acceptable at the a = 1.0 level.
p,(x)
1.
.9
MS
,---------,
.8
.7
.6
.5
.4
.3
.2
.1
0 ~~~~~~~--~--~~-L4-~~~ x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.30 An example of using a-cut as a screening procedure.

168
4.4.2 Buckley and Chanas' Approach
Buckley and Chanas [B40] place a more restricted condition for
the proposition "M > M2 .11 Given fuzzy numbers Ml and M2 and their
\x ex
corresponding ex level sets,
The ranking order is determined by the condition,
(4.48)
For example (see Fig. 4.29), if we set ex 0.9 then
[.40,.44] and M2 [.48, .52].

0.9
Based on Eq.(4.48), we know M2 > Ml because a 2 .44.
Note
This method suffers the same problem of inconsistency as Adamo's
method, and should only be used for screening rather than ranking
purposes.
In addition, it lacks discriminating ability in many cases.
Referring to Fig. 4.29, at ex = 0.7, we obtain
[.40,.52] and M2 [.44, .56].

0.7
since a l = .40 < b2 = .56 and a 2 = .44 < bl = .52, we cannot tell
which fuzzy set is better by using this method. When ex = 0.3, we
obtain
[.40,.68] and M2 [.36,.64].

0.3
Again, using this method, we cannot tell which fuzzy set is better.
169
4.4.3 Mabuchi's Approach

Mabuchi [M1] realized that using a single a-cut value gives
an inconsistent result (as seen in sections 4.4.1 and 4.4.2). He then
proposed a ranking procedure which compares fuzzy sets using multiple
a-cut values.
First of all, one must transform the compared fuzzy sets into a
functional type index Jij(a) which characterizes the degree of
dominance of fuzzy set Mi over Mj • The a e [0,1] is the notion of
a-cut. Each Jij(a) function can be expressed linguistically by
assigning a linguistic statement to a curve shape of a versus J ij •
Fig. 4.31 presents two examples of such curves. Curve 1 indicates "M.~
is strongly dominant with moderate complication." Curve 2 shows "M.~
is definitely dominant with much complication." Secondly, a single
index, J~j' is derived from Jij(a). J~j is the weighted average of
the dominance of Mi over Mj . The weights are the a values. J~j gives
a quick reference to the comparison result •
.----.----.--...1-.......---4--+-.......-----.--..----. J..
-1. -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1. II
Fig. 4.31 Examples of a-J ij curves.

170
Algorithm
To compare fuzzy sets Hi and Hj , three steps are used.

step 1. The difference of Hi and Hj (i.e., Hi - Hj ) is a fuzzy set
D = {((V'~D(v»} where
~D(v) sup min(~.(x), ~.(y». (4.49)

v=x-y ~ J
The introduction of D reduces the comparison of fuzzy sets Hi and Hj

to a comparison of a fuzzy set D with crisp number o.
step 2. Let Da = [v - ,v * ] denote the a level set of D where
v* sup v (4.50)
veD
a
v inf v. (4.51)
veD
a
Also define L* and L- as the length of positive and negative regions

of the .
~nterval (v * , v - ], .
respect~vely. Thus,
(4.52)
If the length of the interval is zero, Jij(a) is defined as:
1, if v * v > 0,
Jij(a)
{ -1, i f v *
0, if v *
v
v
< 0,
O.
(4.53)
Jij(a) is then interpreted as the degree of dominance of Hi over Hj at

a level. The higher Jij(a) value indicates a stronger degree of
"M.~ ~ M.".
)
Let
~M.>M.(a). (4.54)
~ )
It follows that
a = ~M.>M.(J). (4.55)
~ )
The membership function in Eq. (4.55) shows that the statement "Mi > Mj
is at J" has a degree of truth of a.
When the curve of a versus J is mostly in the positive region of
J, (particularly at higher a's), it can be said that Mi is strongly
dominant over Mj , whereas the curve widely distributed over both
positive and negative regions reveals that no definite conclusion can
be made.
Since we may set several a values, Jij(a) shall be treated as a

multi-valued index. When a simplified rating regarding the fuzzy
comparison is needed, Step 3 is conducted.
Step 3. The single index JO is defined by
h(D)
2 J a J(a) da, JO e [-1,1], (4.56)
o
where h(D) is the upper bound of ~D(v). This index is considered the
weighted average of the dominance of Mi over Mj . The weights are the
a values. (Recall that a is viewed as a membership function value in
Eq.(4.55).) The derivation of JO can best be depicted by Fig.4.32.
When Sl and S2 are on the opposite sides of the a-axis, their

172
1.0 ---- .. ------.

51
.--....,.....~.__-+-----.-_+___.-_.____.,.-_.___, J ..
-1. -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1. II
Fig. 4.32 The derivation of JO.
difference determines JOj when both 51 and 52 are on the positive side,
the summation of 51 and 52 gives JOj and when 51 and 52 are on the
negative side, -(5 1 +5 2 ) gives JO. Thus, JO can be a positive or
negative number.
Eq.(4.56) may be too cumbersome to compute. An approximate
calculation of JO is defined as follows: discretize the interval [0,
h(D)] of a into K subintervals and rewrite Eq.(4.56) as:
K . *h
2 [ L i*J(~ ) ~*J(h)], (4.57)
K2 i=o K
where h = h(D) and JO E [-1,1]. The selection of the number for
subinterval K is rather arbitrary. The larger the K value, the

more calculation effort is required, but a close-to-real JO is
achieved.
Numerical Example
Given two fuzzy sets M and N (see Fig. 4.33) ,we can compare them
using the following steps.

173
P.(x)
1. M •
.9 N *
.8
.7
.6
.5 *•
.4
.3
.2
.1
0 * 10 X
0 2 3 4 5 6 7 8 9
x,y 3 4 5 6 7 8
/..LM(X) 0 1 .75 .50 .25 0
/..LN(Y) 0 .50 1 .50 0 0
Fig. 4.33 Fuzzy sets M and N.
J.b(v)
1.0
0.5
~-r--~·-4--4---~~--+--4--~~V
-5 -4 -3 -2 -1 0 2 3 4 5
Fig •. 4.34 Discrete difference set.
step 1. The fuzzy difference set D is constructed according to
Eq. (4.49) (see Fig. 4.34):
v -3 -2 -1 o 1 2 3 4
o .5 1 .75 .5 .5 .25 o
174
The element v = 5 is dropped from the difference set, O. Because at v
= 4, we have a bound where gO(v) = 0.
The element (-2,.5) is obtained as follows. We know that the
possible (x,y) combinations such that v = x - y = -2 are
-;--1---:----:----:----:
The corresponding membership values and (~(x) A gN(y» are
~(x) ° 1 .75 .50

gN(y) 1 .50 ° °
° .50 ° °
According to Eq.(4.49), gO(-2) = max [0,.5,0,0] .50.
step 2. Set a = 0, we have
0a=o [-3, -2, -1, 0, 1, 2, 3, 4],
v* sup v = 4,
veOa=o
and
v inf v = -3.
veoa=o
Consequently, L* = 4 and L- 4-3

3. Thus, J(a = 0) = ~ = 0.14,
which indicates the degree of "M > Nil is only 0.14 at the a = ° level.
Similarly, if we set a = 0.25, a .50, a = .75, and a = 1.0, we can
have different J(a) values. They are summarized below •
a
° .25 .50 . 75 1.0
J(a) .14 .20

° -1 -1
175
Ci
-------------------------1.0
.---.-~.___.____.-+_.J.......,.-.....____._-....___. J ..
-1. -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1. II
Fig. 4.35 a-J ij curve for fuzzy sets M and N in Fig. 4.34.
Note that J(a 1.) = -1 can also be directly obtained from Eq.(4.53)
.
s~nce v* v < o. The curve shape of a versus J ij is approximated in
Fig. 4.35. It shows that "M is strongly dominated (by N) with much
complication" (see Note 3 for details).
step 3. At different a levels, the preference orders vary.
Thus, a unified index, JO, is needed. It is computed using Eq.(4.57)
(given h = 1 and K = 4):
4
~ [E i*J({) - i*J(1)] -0.775.
(4) i=o
A negative JO indicates that M can never be better than N.
Note
1. Although the Jij(a) index shown previously is for two fuzzy
sets, problems with more than two fuzzy sets can also be solved using
Mabuchi's method. If there are U1 , ... , Um fuzzy final ratings, we
can use Jij(a) for each pair of (Ui,U j ) and construct a preference
relation matrix. The ranking order can then be obtained.
Or we can define a fuzzy min as
176
(4.58)
and compare each ~u{ (xi)' Vi, with ~min(v). For each ~U (Xi)' Vi, we
... i
get a Jmi(a). The final ratings with larger Jmi(a) indicate higher
ranking position.
2. The proposed algorithm is meant to rank fuzzy sets with
discrete membership functions. When there are fuzzy sets with
continuous membership functions, it cannot be applied. To resolve the
difficulty, Mabuchi [M1] developed equations that can be used to
derive Jij(a) for fuzzy sets with continuous membership functions.
There are two fuzzy sets, Mi and Mj , (as seen in Fig. 4.36). For
an arbitrary a level, we have
M{~
... ~
= [a.,b.] and M.
J
~ ~ a
There are different equations for calculating Jij(a) under differnt
situations.
(a) When b i * a j , we have
Ib i -
ajl - la i - b·1
J
Jij(a) (b i - a.)
J
- (a.
~
-
b. )
J
(4.59 )
(b) When b.~ aj (Le., zero interval), we have
1, bi > a j ,
Jij(a)
{ 0, b.
~
aj ,
-1, b i < a ..
J
(4.60)
(c) When b i > a. and b.> ai' we have

J J
(a i + bi) - (a j + b j )
J .. (a) (b. (4.61)
~J
~
a i ) + (b j aj )
177
JL(X)
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
o ~~--~~--~~~~-+--~~~ x
o aja j b j b j 1.
Fig. 4.36 Fuzzy sets and their a-cuts.
1
Given a difference set, D, that has a continuous membership
function, we can have an a-level set [a,b] where a ~ b. Jij(a) can be

computed using the following formulas.
a < b then L * b - a and L - o. (4.62)

If 0
'"
If a < 0 < b then L * band L - a. (4.63 )
If a < b 0 then L * 0 and L

- b - a. (4.64)
'"
In any situation, we have
(4.65 )
where L * - L- = Ibl - lal, and L * + L- = b - a.

Eqs. (4.59), (4.60), (4.61), and (4.65) are proposed for fuzzy
sets with continuous membership functions. For example, given fuzzy
numbers M. and M. (see Fig 4.36) and a = 0.7, we can obtain

1 J
[.44,.56] and M. [.40,.52]

J. 7
178
since b. > a. and b. > a., we can use Eq.(4.61) to compute

) l. l. )
Jij(CX) [(.44+.56) - (.40+.52)]/[(.56-.44) + (.52-.40)] 0.33.
The difference set, 0, can be determined by using the subtraction
procedure presented in Section 3.4.2 (Fuzzy Arithmetics), as shown in
Fig. 4.37. At cx = 0.7 level, we get an interval [a,b] = [-.08,.16].
By applying Eq.(4.65), Jij(CX) is calculated as:
[1.161 - 1-.081]/[.16 - (-.08)] .33.
-.5 -.08 0 .16 .5
Fig. 4.37 The fuzzy difference set.
Note that the results from both Eqs.(4.61) and (4.65) are the
same. This verifies that the comparison ofM. and M. can indeed be
l. )
replaced by comparing the fuzzy difference set, 0, with o.

3. The multi-valued index Jij(CX) reveals some preference
relationship between fuzzy sets Mi and Mj . This relationship may be
linguistically expressed by assigning a linguistic statement to the
curve shapes of ex versus J ij . There are three factors to consider in
assigning linguistic statements: (a) the distribution of the curve

179
over the J ij axis, (b) the average position along the J ij axis, and
(c) the height of the curve. Table 4.2 gives the correspondence
between a-J .. curve and the linguistically described conclusions.

~]
This process of assigning linguistic terms to fuzzy sets (based on
Table 4.2) is by no means a standard. As a matter of fact, it is a
preliminary finding of Mabuchi [Ml]. Further research on assigning
linguistic terms to fuzzy sets is needed.
Table 4.2 Relationship Between Curve and Linguistic statement
a-J ij curve Conclusion
I. Distribution Degree of complication
Impulse type No complication (nonfuzzy)

One-sided Slight complication
Narrow one-sided Very slight complication
Both sides but biased Moderate complication
on one side
Equally distributed Much complication
on both sides
II. Average position Degree of dominance
Nearly 1 Definite dominance

Over 0.5 Strong dominance
About 0.5 Moderate dominance
Under 0.5 Slight dominance
Nearly 0 No dominance
(Negative positions denote "dominated" relationship)
III. h (heig:ht) Credibilit~
Nearly 1 Credible
Over 0.5 Moderate credibility
Under 0.5 Little credibility
Nearly 0 No credibility
180
When a curve takes the form of an impulse whose value is zero

everywhere but at one point where its value is h, this, with h = 1,
corresponds to a nonfuzzy conclusion. A one-sided distribution of a
curve (either on the positive or neqative reqion) qives rather clear
conclusions. But the distribution of a curve over both reqions
complicates the matter.
The interpretation of the averaqe position of a curve is obvious:
it is the deqree of dominance. Since the siqnificance of the Jij(a)
index is considered to increase as the a values approach 1, it would
be better to assume this averaqe to be the a-weiqhted averaqe.
The value of h can be related to the credibility of result. A
smaller h indicates a less credible result. This is supposed to be a
cautionary factor, and the remarks concerninq this factor should be
applied only when h is smaller.
In Fiq. 4.38, several example curves are drawn. Curve 1 has an
impulse type distribution, an averaqe position of 0.5, and a heiqht of
1. Thus, we assiqn the linquistic statement "Mi is moderately
dominant with no complication". The credibility is omitted because h
is not small. Curve 2 has a one-sided narrow shape, and an averaqe
position of nearly one, but a very low h value, thus the linquistic
statement "Mi is definitely dominant with no complication and little
credibility" can be made. similarly, curve 3 and curve 4 are assiqned
statements "Mi is moderately dominant with much complication," and "Mi
is sliqhtly dominant with sliqht complication," respectively.
4. When special membership functions, such as trapezoidal and
trianqular fuzzy numbers, are presented, different equations must be
used. Mabuchi [M1] derived expressions of Jij(a) and J oij for
trapezoidal and trianqular fuzzy numbers. The readers may refer to
the oriqinal paper for detailed formulas.
181
ex curve 1
-1. -.8 -.6 -.4 -.2 0 .2 .4 .6 .8
Fig. 4.38 Examples of a-J .. curve.

1.]
182
4.S Ranking using Comparison Function

The methods summarized in this category are Dubois and Prade
[D29], Tsukamoto, Nikiforuk, and Gupta [T1S], and Delgado, Verdegay,
and Vila [D4]. Given a fuzzy number M as shown in Fig.4.39, we may
have two possible comparison functions: (a) "all the crisp numbers
possibly greater than M" which is characterized by the dashed line,
and (b) "all the crisp numbers possibly less than M" which is
characterized by the dotted line.
Different methods may define their comparison functions
differently. Dubois and Prade [D29] and Tsukamoto et al. [T1S]
utilize Zadeh's possibility theory to derive comparison functions.
Delgado et al. [D4] uses Sugeno's A-measure to define comparison
functions.
Fig. 4.39 Two comparison functions on fuzzy number M.
Dubois and Prade [D29] defined the comparison functions using

possibility theory. The various comparison functions are then used to
derive four ranking indices: (a) possibility of Dominance (PD), (b)
Possibility of strict Dominance (PSD), (c) Necessity of Dominance
(ND), and (d) Necessity of strict Dominance (NSD). The ranking order
may be determined using one or more indices.
183
Tsukamoto et al.'s method [T15] is basically similar to Dubois
and Prade's method. The comparison functions are determined first
using possibility theory. For fuzzy sets M and N, the comparison
functions are used to derive IT (M<N), IT(M=N), and IT(M>N) which denote
the possibility of "M < N," "M = N," and "M > N," respectively. The
three indices are compared, and the index with the highest value
determines the relationship between M and N.
Delgado et al. [D4] use Sugeno's A-measure to define comparison
functions. Such functions are then used to define two ranking indices
for fuzzy sets M and N. The first one measures the proposition "M ~
Nil while the second one measures the proposition "M :s N." When the
first index value is greater than zero, the second index value must be
zero, and vice versa. The non-zero index determines the relationship
between M and N.
4.5.1 Dubois and Prade's Approach
Dubois and Prade [D29] proposed a set of four ranking indices
in the framework of Zadeh's [ZlO] possibility theory. Each index can
be used for comparing two fuzzy sets or be extended for cases of
more than two fuzzy sets.
Definitions
Let U be a set of elementary events. Any subset of U is called
an event. An event A ~ U is said to occur when some elementary event
in A occurs. For example, let U = {2,5,8}. An event A can be any
element in the set {~,{2},{5},{8},{2,5},{2,8},{5,8},{2,5,8}}. Note
that U and A are not fuzzy.
Possibility Measure
A possibility measure IT on U can be built from a set function IT
from the power set of U, 2U, to the unit interval [0,1]. That is,
184
II.. 2 U ---+ [0,1]. (4.66 )
U
Again, let U = {2,5,8}. 2 = {0, {2}, {5}, {8}, {2,5}, {2,8}, {5,8},
{2,5,8}}. An event A can be any element in 2 U. When A = 0, the
possibility measure is 0; when A = U the possibility measure is 1.

Any element between 0 and U shall have a possibility measure between 0
and 1. That is,
II (0) 0, II (U) 1, (4.67)
VA,B E 2U, II (AuB) max (II(A),II(B». (4.68)
Given a normalized fuzzy set F (i.e., some ~F(x) = 1), the
possibility measure IIF(A) can be derived from the membership function
~F(x) as:
sup ~F(X), VA S U, (4.69)

xEA
where ~F(X) may be seen as the possibility distribution function
underlying II F . IIF(A) is interpreted as the possibility of realizing
event A when the possibilities of elementary events of U are known.
Given the possibility distribution F = {(x'~F(X»} = {(2,.4), (5,.7),
(8,1)} and assuming A = {2,5}, the possibility measure of realizing
A = {2,5} is:
sup ~F(x) sup [.4,.7] .7.

xE{2,5}
Notice that A is a crisp subset of U in Eq. (4.69). When both F
and A are fuzzy sets, Eq.(4.69) is extended to

185
sup min (~F(x), ~A(X)). (4.70)

x
For example, given F = {(2,.4), (S,.7), (8,1)} and A {(2,.S), (S,.6)},
the possibility of realizing fuzzy event A is
sup min [~F(X), ~A(X)] sup [.4,.6] .6.

xe{2,S}
Such an extension is the only possible one if we require Eq. (4.70) to
be interpreted in terms of the intersection of the level cuts of F and
A. That is, Eq. (4.70) is equivalent to
(4.71)
To show Eqs. (4.70) and (4.71) give the same possibility measure,
F = {( 2 , .4), (S,. 7), (8, 1)} and A = {( 2, • S) , (5, . 6)} are used again.
Assuming that a 1 = .5 and a 2 = .5, F.5 = {5,8}, A.5 = {2,S}. Since
F.S is equivalent to {(2,O), (5,1), (8,1)}, we can calculate llF. 5 (A. s )
using Eq. (4.69) as:
sup ~F (x) sup [0,1] 1.

xe{2,5} .5
By setting different a 1 and a 2 levels, we can get llF (Aa) as the

a1 2
elements in the following matrix.

186
<Xl
.1 .2 .3 .4 .5 .6 .7 .8 .9 1.
.1 1 1 1 1 1 1 1 0 0 0
.2 1 1 1 1 1 1 1 0 0 0
.3 1 1 1 1 1 1 1 0 0 0
.4 1 1 1 1 1 1 1 0 0 0
<X2 .5 1 1 1 1 1 1 1 0 0 0
.6 1 1 1 1 1 1 1 0 0 0
.7 0 0 0 0 0 0 0 0 0 0
.8 0 0 0 0 0 0 0 0 0 0
.9 0 0 0 0 0 0 0 0 0 0
1. 0 0 0 0 0 0 0 0 0 0
The possibility measure ITF(A) is calculated based on Eq.(4.71) as:
ITF(A) = max [.1, .2, ... , .6, 0, ... , 0, 0] = .6
Given the same F and A, Eqs.(4.70) and (4.71) will result in the same
possibility measure.
When Eq. (4.71) holds for fuzzy event A and possibility
distribution F, we have ITF(A) = ITA(F). The possibility of a fuzzy

event with respect to a fuzzy set of elementary events is a
symmetrical concept expressing consistency or partial matching between
fuzzy sets. The concept of possibility refers to set intersection.
Necessity Measure
A necessity measure on U can be built similar to the possibility
measure such that
N: 2 U --+ [0,1],
and the following properties hold:
N (fII) 0, N(U) 1, (4.72)
and
N(MB) min (N(A),N(B)), VA,B ~ u. (4.73)

187
The degree of necessity of realizing event A is defined as:
V A ~ U, N (A) 1 - Il(A), (4.74)
where A is the complementary set of A and Il(A) is the possibility
measure of realizing event A. That is, the degree of necessity of
realizing event A equals 1 minus the possibility of not realizing
event A, or the degree of necessity of realizing an event A equals
the degree of impossibility of realizing the opposite event. For
example, let U = {win, loss} and A = {win}. The complement of A is
A = {loss}. Consequently, N(win) = 1 - Il(loss). similarly, the
possibility measure can be defined as:
Il(A) 1 - N (A) • (4.75)
If IlF is defined as Eq. (4.69), it is obvious that
VA, NF(A) inf (1 - ~F(x». (4.76)

xeA
When both A and F are fuzzy, Eq.(4.76) is equivalent to
NF(A) 1 - sup min (~F(X) , (1 - ~A(x») (4.77)

x
inf max (1 - ~F(X)'~A(X».

x
When event A is crisp, Eq.(4.77) becomes
inf ~F(x) . (4.78)

xeA
The reason for using both possibility and necessity measures is
as follows. Given a crisp event, A, then

188
max (Il(A),Il(A» 1. (4.79 )
Intuitively, this means at least one of two contrary events must be

possible. Consequently, Il(A) and N(A) serve as nonredundant
information which must be considered simultaneously. The following
relationships hold between them:
N(A) :s Il(A) (4.80)
N(A) > 0 ~ N(A) = 0 ~ Il(A) = 1 (4.81)
Il(A) < 1 ~ Il(A) = 1 ~ N(A) = 0 (4.82)
Eq. (4.80) can be intuitively interpreted: what is necessary must be
possible, but not conversely so. The relationships shown in Eqs.
(4.81) and (4.82) make explicit the complementarity of indices N(A)
and II (A) .
If A is fuzzy, Eq. (4.79) does not hold, because max (Il(A),
Il(A» is not 1 in general. However, since
sup ~AuA(x) ~ 0.5, (4.83 )

xeU
the following must be true:
0.5, iff sup ~F(x) ~ 0.5. (4.84)

xeU
When F is normalized, we have
(4.85)
Assessment of Fuzzy Sets having Fuzzy Number M as a Fuzzy Bound
Given a fuzzy number M, one can derive four sets of numbers
having M as a fuzzy bound: (i) (M,+co), (ii) (-co,M], (iii) ]M,co), and
189
(iv) (-oo,M[.
(i) The set of numbers possibly greater than or equal to M is
[M,+oo) (see Fig. 4.40a) with membership function
Il[M,oo) (r) sup IJ.M(X) (4.86)

ri!:X
In possibility theory, the universe is U = {all real numbers}. The

possibility distribution is F = Il M(x), and the fuzzy event is A
{xlxe(-oo,r)}. The possibility of r i!: M is ITM«-oo,r]).
(ii) The set of numbers possibly smaller than or equal to M is
(-oo,M] (see Fig. 4.40b) with membership function
1l(-oo,M] (r) sup 11M (x) (4.87)

r:5x
In possibility theory, the fuzzy event in this case is {xlxe[r,oo)}.

The possibility of r :5 M is ITM([r,oo».
(iii) The set of numbers necessarily greater than M is ]M,oo)
(see Fig. 4.40a) with membership function
inf (1 - Il M(x» NM ( ( -00, r [ ) . (4.88 )

r:5X
The grade of necessity of realizing the event A = {xlxe(-oo,r[} equals

1 minus the possibility of realizing the opposite event A=
{xlxe[r,oo)}. That is,
(iv) The set of numbers necessarily smaller than M is (-oo,M[

(see Fig. 4.40b) with membership function
1l(-oo,M[ (r) (4.89)

190
The grade of necessity of realizing the event A = {xlxe]r,m)} equals 1
minus the possibility of realizing the opposite event A=

{xlxe(-m,r]}. That is,
The fuzz'iset IM,oo)

p,(x) /
1. ----------------------""""""""""""""'''il'~o''ifo''I>''8'lt'8'lt'ij'lt'o
o
o
/\
o
The fuzzy set )M,oo)
Fig. 4.40a The fuzzy sets [M,m) and ]M,m).
P,(x) j
The fuzzy set (-oo,M[
1. "\T"o"'o"'V"'(F~" """""':' -, - - - - - - - - -
! "I""Y ," (~,Ml
Fig. 4.40b The fuzzy sets (-m,M] and (-m,M[.
Assessment of Degree of "M '" R"
The previous section pertained to the comparison between a fuzzy
set M and a crisp number r. In this section, however, the comparison

191
involves two fuzzy numbers M and R. The fuzzy sets [R,+~) (the set of
numbers possibly greater than R) and ]R,~) (the set of numbers
necessarily greater than R) are used to derive four indices which
access the proposition "M ~ R." The four indices are IIM([R,+ao»,
IIM(]R,+ao», NM(]R,+ao) and NM([R,+ao» (see Fig.4.41).
JL(X) M R
Fig. 4.41 The ranking indices for fuzzy numbers M and R.
(i) since M and [R,+ao) are both fuzzy, Eq.(4.70) is used to

derive the first index, which shows the degree of possibility that
"M is possibly greater than R" as:
IIM( [R,+ao» sup min (IlM(U), Il[R,+ao) (u» (4.90)

u
sup min (IlM(U) , sup IlR (v»
u u~v
sup min (IlM(U) , IlR(V».

u~v
Note that, according to Eq.(4.86), Il[R +ao) (u) = sup IlR(V).

, u~v
(ii) By applying Eqs.(4.70) and (4.88) we get the second index,

which shows the degree of possibility that "M is necessarily greater
192
than R" as:
IIM(]R,+oo) ) sup min (~M(U), ~]R,+oo) (u» (4.91)

u
sup min (~M(U), inf (1 - ~R(v»)
u v~u
sup inf min (~(u) , (1 - ~R (v) )) .

u v~u
Note that, according to Eq.(4.88), ~]N,+oo) (u) = inf (1 - ~R(v».

v~u
(iii) By applying Eqs. (4.77) and (4.86), we get the third index,
which characterizes the degree of necessity of the proposition "M is
possibly greater than R" as:
inf max «1 - ~M(u», sup ~R(v» (4.92)

u V::5U
inf sup max «1 - ~M(u», ~R(v».

u V::5U
(iv) By applying Eqs. (4.77) and (4.88) we get the fourth index,
which characterizes the degree of necessity of the proposition "M is
necessarily greater than R" as:
inf max «1 - ~M(u», inf (1 - ~R(V») (4.93)

u v~u
1 - sup min (~M(u), ~R(V».

U::5V
Given that M and R are fuzzy intervals M = [m1 ,m 2 ], R = [r 1 ,r 2 ],

the degree of truth of the proposition "M ~ R" is computed using the
proposed indices. Table 4.3 clearly indicates that the four indices
are necessary and sufficient to characterize the respective locations
of M and R. Any three indices cannot discriminate all the configura-
tions shown in Table 4.3. Thus, the four indices must work in a group
such that full discrimination ability can be preserved.

193
Table 4.3 Comparison of Dubois and Prade's Indices
Configuration TIM ( [R,oo» TIM (] R, 00) ) NM( [R,oo» NM (] R, 00) )
IMI
I
I
R I 1 1 1 1
---I: 1 1 1 1 0
! I
--r--r-i
1 1 1 0 0
1 0 1 0
1 :I--! 1 0 0 0
'---I
I 0 0 0 0
I
comparison Indices for m Fuzzy Numbers
Let M1 , ... , Mm be fuzzy quantities that are to be compared. We
shall extend the pairwise indices (as given by Eqs. (4.90), (4.91),
(4.92), and (4.93) to m-ary versions by introducing the fuzzy maximum
defined by Eq. (4.94) as:
sup min IlM (u.). (4.94)

u 1 ' . . . ,um i i 1.
v=max(u 1 ,···,um)
This fuzzy set is very easily obtained from the shapes of M1 , ... , Mm.
That is,
n
max U [( (\ [M. ,+00» (\ M.], (4.95 )
i=l j .. i J 1.
which can be interpreted as: Is Mi greater than or equal to the

194
greatest of the Mj'S (j ~ i)? The problem of comparing m fuzzy

numbers is now converted to comparing each Mi with max M.•
j~i J
The four indices for the dominance of Mi in the set Ml , •.• , ~
are given by:
(a) Possibility of dominance
sup min ("M. (u), '"'max (v) ) (4.96)

Ui!:V 1.
(b) Possibility of strict dominance
PSD(M i ) lIM. (]l!'a',' Mj , +a>)) sup inf min("M. (u),l - '"'max(v»

1. J~1. u Vi!:U 1.
(4.97)
(c) Necessity of dominance
NM. ([l!'a',' Mj , +a>)) inf sup max (l-"M. (U)''"'max(v»

1. J~1. u v~u 1.
(4.98 )
(d) Necessity of strict dominance
NSD(M i ) 1 - lIM. ([l!'a',' Mj,+a>)) 1 - sup min ("M. (u)''"'max(V»

1. J~1. u:sv 1.
(4.99)
It is recommended by Dubois and Prade [D29] that the four

comparison indices should all be used in deriving ranking order.
These indices give exactly four ranking orders. If all are identical,
the corresponding ranking is validated. otherwise, the DM is required
to make the decision on which index (or indices) to use in deriving
ranking order.
195
Numerical Example
Dubois and Prade's ranking indices will be applied to Example 10.
First, we shall derive the fuzzy maximum of U1 and U2 using
Eq. (4.94) as:
~max (v) sup min [~u (u 1 ), ~u (u 2 )]·

12 u 1 ,u 2 1 2
max(u 1 ,u 2 )=v
similarly, using Eq.(4.94), we can get max 13 and max 23 , respectively.

For fuzzy final rating U1 ' the four indices ( see Fig. 4.42) are
listed below.
PD(U 1 ) sup min (u 1 ), .714

(~U
1
~
max 23 (v»
u1~v
PSD(U 1 ) sup inf min (~U (u 1 ), 1 - ~

max 23 (v» .200
u v~u1 1
inf sup max (1 - ~U (u 1 ), ~max (v» .200

u V:5U1 1 23
1 - sup min (~U (U 1 ), ~max (v» o

U:5V 1 23
/L(X) U1 max 23
1. ----------- ----- - -- ----------1-----------------
I'
--.~-
.9 \, ~
~
I.'.'
.8
.7 .'.'.'
,.'
~
.6 I
.5 .'
.4
.3 NO(U)
.2 . __ ...1.. ... __ .................... ~~~~~1
.1
o ~~--~~--~~--~~--~-T~X
o .1 .2 .3 .4 .5 .6 .7 1.
Fig. 4.42 PD, PSD, ND, NSD indices for U1 •

196
Notice that these four indices happen to be the intersections of
different membership functions with fuzzy maximum, max 23 . In this
case, max 23 is simply U3 itself.
Similarly, we can get the four indices for U2 and the four
indices for u 3 . The results are summarized in Table 4.4. All four
indices give the identical ranking order: A3 > A2 > Ai
Table 4.4 Summary of Numerical Example

Fuzzy
Ratings PD PSD ND NSD
U1 .714 .200 .200 0
U2 .805 .327 .233 0
U3 1 .767 .678 1
Note
1. Dubois and Prade [D29] gave physical meaning to the four
indices as follows.
(i) ITM([R,oo» is used to compare the worst part of R (the left
leg of R) with the best part of M (the right leg of M). The index
value will be high when the left leg of R is smaller than or equal to
the right leg of M.
(ii) ITM(]R,oo» compares the right legs of both M and R. This
index will be large when the right leg of R is smaller than the right
leg of M.
(iii) NM([R,oo» compares the left legs of both M and R. This
index will be large if the left leg of R is smaller than the left leg
of M.
(iv) NM(]N,oo» compares the right leg of R with the left leg of
M. The index value will be large if the right leg of R is smaller
than the left leg of M.
2. The equations for deriving PD, PSD, ND, and NSD are very
197
cumbersome to use. Indeed, using Eqs.(4.96) through (4.99) to rank

discrete fuzzy sets is a very time consuming process. But for
continuous membership functions that are convex and normal, it is just
a matter of finding the intersections between continuous membership
functions.
Two fuzzy triangular numbers M and N are presented in Fig. 4.43.
Their membership functions are given as:
L (m~u), u:!5 m
JIM(u) {
R (u;m), u ~ m
L (n~u), U:!5 n
{
R (u~n), u ~ n
,.,,(x)
M N
1.
a· m- a
~. b - m
1· n - c
o· d - n
Fig. 4.43 Fuzzy triangular numbers M and N.
The ranking problem has come down to solving the following equations.
(1) find u such that
and the resulting u value gives rrM([N,oo»i

198
(2) find u' such that
and the resulting u' value gives NH([N,m»i and
(3) find u" such that
u"-m u"-n
1 - R(-{3-) R(-c')-)
and the resulting u" value gives IIH(]N,m».
3. Dubois and Prade [029] also suggested the construction of
fuzzy outranking relations on the fuzzy sets H1 , ... , Hm through
pairwise comparison of the fuzzy sets. The outranking relations are
then used to get ranking order. We feel that this alternative
approach is very difficult to use. In addition, it does not guarantee
a better result than previous approaches. This outranking relations
approach is not recommended.
4. One should be reminded that the PO index is the same as Baas
and Kwakernaak's [B1] ranking index, while the NSD index is identical
to watson et al.'s [W3] ranking index. Dubois and Prade [029] pointed
out that unless all four index values are used, counter-intuitive
results cannot be prevented.
5. It is worth noting that the IIH[R,m) = 1 index measures the
extent that "H ~ R." The IIR[H,m) index measures the extent that "R ~
H." When one of them is 1, another is not necessarily o. For
example, in Fig. 4.44, when IIR[H,m) = 1, IIH[R,m) = O. On the other
hand, when H becomes H', we have IIR[H',m) = 1 and IIH,[R,m) = ~1 ~ o.

similar remarks can be made to other indices in Dubois and Prade's
method.
199
P,(X) M
1. -------
M' R
4.5,2 Tsukamoto et al.'s Approach
Tsukamoto et al. [T15] propose three ranking indices within the
possibility theory. For a pair of fuzzy numbers M and N, the three
indices are IT (N<M), IT(N=N), and IT(N)M). They denote the possibilities
of the three events liN < M," liN M," and liN > M," respectively, given
that M is the reference set. The possibility measure with the highest
value determines the relation of M and N.
The three ranking indices are derived through two comparison
functions: (a) the set of numbers possibly ~ x (with x being
restricted by M), ~M (u); and (b) the set of numbers possibly ~ x,

L
~~(U). Their membership functions (see Fig. 4.45) are
~ (u) sup ~(x) (4.100)

L u~x
~(U) sup ~M(x). (4.101)

U:5X
Notice that if Dubois and Prade's [029] notations are used, ML = [M,m)
and ~ = (-m,M]; ML ~ ~ = M. This is easily observed in Fig. 4.45.

200
The three indices on Nand M are then defined (see Fig. 4.46) as:
JI(N < M) sup min (1 - IlM (u) , IlN (u» , (4.102)

u L
JI(N M) sup min (IlM(U) , IlN(u», (4.103)

u
JI(N > M) sup min (1 - 1l~(U), IlN(U» . (4.104)

u
The highest index value determines the relationship of M and N.
Fig. 4.45 Tsukamoto et aI's comparison functions.
/L(x) 1 - to;,. M N 1- ~
1. +----.;;;....,
JI(N"ML_
-mlil>Mr---
ll(N<M)
Fig. 4.46 Tsukamoto et al.'s indices on fuzzy sets M and N.

201
p.(x) U1 1 - U1R
1. -!-__.,----
,
,,,,,
· , r ; , - - - - - - - - - l ' f f I I f J l l l l l l l f l " " ' l I f l f l l r t . I I ! I I I I I f { l f f l l l l l f I f I l I lJ
.9
.8
.7 '-. f
.6
.5
.4
.3
.2
.1 1- I1L
0 ~~=W~~~~~~~~~~~x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.47 Tsukamoto et al.'s indices for U1 and U2 .
Numerical Example
Example 10 is solved again using Tsukamoto et al.'s method. For
U1 and U2 , Tsukamoto et al.'s indices (referring to Fig. 4.47) are
We can conclude that U1 and U2 are equivalent to each other.
Similarly, we obtain
We conclude that U3 is greater than U1 with high possibility.
Combining the above answers, we know U3 > U1 ~ U2 .
Note
1. Eq. (4.102) may be treated as the degree of nonemptiness of
the fuzzy set, N~(-oo,M], of numbers smaller than or equal to M. Since
the complement of (-oo,M] is equivalent to ]M,oo), Eq. (4.102) is equiva-
lent to the possibility measure rrM(]M,oo) which indicates the degree of
possibility of "N < M." Similarly, rr(N > M) is equivalent to
rrN«-oo,M[) which indicates the degree of possibility of "N > M."
Index rr(N = M) is clearly an index for assessing the possibility of

202
equality. Obviously, TI(N)M) and TI(N<M) have the same characteristics

of the PSD index described in Dubois and Prade's [D29] method.
2. Tsukamoto et al. [TIS] mentioned that there is a set of dual
indices defined as:
N(N<M) inf max (1 - ~(U),1 - ~(u», (4.105)

u
N(N=M) inf max (~(U), 1 - IlN(U», (4.106)

u
and
N(N)M) inf max (1 - ~ (u) , 1 - IlN(U». (4.107)
u
They may be referred to as the necessity measures of the propositions

liN < M, II liN M, II and liN > M, II respectively.
However, it has been proven by Dubois and Prade [D29] that
N(N=M) NN(M) 1 - TIN(H) (4.108)

1 - max (TIN ( (-m,MO, TIN ()M,m) »
1 - max (TI(N<M),TI(N>M».
Moreover, since [M,m) H ]M,m) ,

"
N(N<M) 1 - max (TI(N>M) ,TI(N=M»,
(4.109)
and similarly
N(N)M) = 1 - max (TI(N<M), TI(N=M». (4.110)
The dual indices N(N<M), N(N=M), and N(N)M) contain redundant

information with respect to TI(N<M), TI(N=M), and TI(N)M). Thus, it is
concluded that Tsukamoto et al.'s method basically resembles Dubois
and Prade's method except that Tsukamoto et al.'s method considers
only the PSD and NSD indices.
203
3. This method can easily be extended to cases with more than

two fuzzy sets. One way is to make a pairwise comparison for each
(Mi,M j ) and then construct a ranking order as shown in the numerical
example. An alternative approach, possibly a more promising one, is
to define the fuzzy max, max, as that in Dubois and Prade [029], and
compute all IT index values for each Mi versus the fuzzy max. The
ranking order is determined as follows. If a smaller IT(Mi< max) is
found, Mi is considered better. If a larger ITCMi> max) is found, Mi
is also considered better. When a larger IT(M i = max) is found, Mi has
higher degree of matching with the fuzzy max, i.e., Mi is better.
When all three indices give identical ranking orders, we can stop the
process. Otherwise, the OM is required to choose one or more ranking
index to find the ranking order.
4.5.3 Delgado et al.'s Approach

Delgado et al. [04] use Sugeno's [S28] A-measure to define
comparison functions. Such functions are then used to define two
ranking indices for fuzzy sets M and N. The first one measures the
proposition "M ~ N" while the second one measures the proposition "M s
N." When the first index value is greater than zero, the second index
value must be zero, and vice versa. The non-zero index determines the
relationship between M and N.
Definitions
The Fuzzy Sets "numbers possibly ~ M" and "numbers possibly s M"
A fuzzy number, M, defined by 4-tuple (a,b,c,d) is shown in Fig.
4.48. The membership function is defined (Dubois and Prade [026]) as:
S (b-x I b-a), x s b,
~(X) { S (x-c I d-c), x ~ c, (4.111)
1, otherwise.
204
#L(X)
1. -............... ,.------.
Fig. 4.48 Fuzzy number M (a,b,c,d) •
This is an L-R type fuzzy number. The set of numbers possibly
greater than M, f M, and the set of numbers possibly less than M, gM'
are characterized as (see Fig. 4.48a and 4.48b):
0, if x a,
{
:5
fM(x) 1, if x ?!c, (4.112)

/.lM(X) , i f c > x > a.
1, if x b,
{
:5
gM(x) 0, if x ?! d, (4.113)
/.lM(X) , i f d > x > b.
Note that fM is a possibility distribution function pertaining to the
property "?! M." Other possibility distributions such as f M' and fM"
may exit. Similary, gM is not the only possibility distribution
pertaining to ":5 M." Other possibility distributions such as gM' and
gM" may exist. The choice of a possibility distribution is a subjec-
tive matter.
205
JL(X)
1.
Fig. 4.48a The possible distribution functions: fM' fM" and f M".
JL(X)
1.' """".,","
Fig. 4.48b The possible distribution functions: gM' gM', and gM"·
8ugeno's A Measure and Comparison Functions
A fuzzy measure denotes the possibility of an unallocated element
belonging to a (fuzzy or crisp) set. For example, given
N = {(x'~N(x)) I x E [10,20]}, and a subset of N, A = {xiX is about 15},
the possibility of an unallocated element, say 13, in A is a fuzzy
measure.
8ugeno's [828] A-measure is a fuzzy measure. Consider a fuzzy
number M = (a,b,c,d). Its A measure is defined as:

206
[~(y) - ~(x)] / [1 + A ~(X)], A > -1, (4.114)
where (x,y] is a interval on R and ~(x) is a possibility distribution
function of M given as:
0, if x s a,
~(x)
{ 1, if x ~ c, (4.115)
~(x), if c > x > a.
The dual of mA(o) is defined as:
(4.116)
where A is the complement of A.
From m~(o) we can determine a comparison function for the

property II~ Mil:
\I x e R, f M , A (x) (4.117)
or
\I x e R, f M , A (x) [(A+1)~(X)] / [1 + A~(X)]. (4.118)
When A e (-1,0], fM A(o) can be seen as the membership function of

,
lithe set of numbers believably ~ M." Note also that when A = 0
fM,A(o) = ~(o). Thus fM,o(o) is lithe set of numbers possibly ~ M."
Since fM A(o) decreases with A in (a,b], A may be considered a

,
requirement level for the accomplishment degree of the property II~ Mil.
similarly, we can define a function PM(X) as:
1, if x ~ d,
{ 0, if x s c,
1 - gM(x), if c > x > d.
(4.119)
207
which has the same properties as those of ~(X). Let nACo) be the
A-measure corresponding to PM(X). The associated comparison function
is
(4.120)
where A e [O,m). When A = 0, gM,A(o) may be seen as the membership

function of the "set of numbers possibly :s M". Again, A plays the
role of requirement level, but the gM,A(x) value decreases when A

increases on [O,m).
Fuzzy Relation (zimmermann [Z30])
Let X,Y ~ ~ be universal sets then
i = {(x,y)'Ui(X,y» I (x,y) ~ X x Y} (4.121a)
is called a fuzzy relation on X x Y.
Max-Min composition (Zimmermann [Z30])

Let R1 (X,y), (x,y) ~ X x Y and R2 (y,Z), (y,z) ~ Y x Z be two
fuzzy relations. The max-min composition of R1 and R2 denoted by R10
R2 is the fuzzy set
{[(x,z),max{min {UR (x,y)'UR (y,z)}}]lxeX,yeY,zeZ}

y 1 2 (4.121b)
For example, given two relations R1 and R2
Y1 Y2 Y3 zl z2 z3 z4
xl .3 .7 .1 Y1 .9 0 .4 .4
R1 x2 .6 .5 .2 R2 Y2 .4 .2 .3 0
x3 .8 1 .1 Y3 0 1 .7 .6
the result of max-min composition R1 ° R2 is

208
zl z2 z3 Z4
xl .4 .2 .3 .3
x2 .6 .2 .4 .4
x3 .8 .2 .7 .4
We shall use ~R 0 R (X 2 ,z3) to demonstrate the calculation procedure.

1 2
since
min {~R (X 2 Y1)'~R (Y1 z3)} min { .6, .4 } .4
l' 2'
min {~R (x 2 Y2)'~R (Y2 z3)} min { .5, .3 } .3
l' 2'
min {~R (x 2 Y3) '~R (Y3 z3)} min .2, .7 } .2,
l' 2'
max { . 4 , . 3 , . 2 } .4
Max-min Transitive
A fuzzy relation R in X x X is called max-min transitive if and
only if R 0 R~ R (i.e., for all pairs of (x,y), ~RoR(X,y) s ~R(X,y».
Reflexive (Zadeh [Z3])
A fuzzy relation R in X x X is called reflexive iff ~R(x,x) 1 for
all x E X.
symmetrical and Antisymmetrical
A fuzzy relation R in X x X is called symmetrical iff ~R(x,y)
~R(Y'x) for all pairs of (x,y).
A fuzzy relation R in X x X is called antisymmetrical if and only
if for all (x,y) pairs where x ~ y, we have either ~R(x,y) = ~R(Y'x) =

o or ~R(X,y) ~ ~R(Y'X).
A fuzzy relation R in X x X is called perfectly antisymmetrical
if for x ~ y, whenever ~R(x,y) > 0 then ~R(Y'x) = 0 for all (x,y)
pairs.
strict Fuzzy Order Relation (Kaufmann [K12], Zimmermann [Z30])
A fuzzy relation in X x X that is transitive, reflexive, and

209
perfectly antisymmetrical is called a strict fuzzy order relation. A
strict fuzzy order relation is used to structure and order the
elements of a given set (i.e., the preferred elements can be crisply
identified).
comparison Relations: A comparison relation is a fuzzy order
relation which helps identify the preferred elements in a given set.

To access the degree of "M ~ N," a fuzzy relation is given as:
sup (fN(x) A ~M(x» (4.122)

XE~
where ~ is the set of all real numbers. Two properties are observed
1 which is said to be reflexive,
(b) if c 1 < c 2 then aT(N,M) = sup (fN(x) A ~M(X» and

XE [C1,C21
there exists ~ E [C1,C21 such that aT(N,M) 5 min (~M(~) ,fN(~»'
This is said to be not antisymmetrical.
The property (b) suggests that a T (·,·) is not a strict fuzzy order
relation, and it is not adequate to serve as a comparison relation.
To make a comparison relation between M and N, a simple
modification on Eq. (4.122) is conducted. The resulting fuzzy relation

is given as:
1 - sup (fN(x) A ~M(X»' (4.123)

XE~
where ~ is the set of all real numbers. The fuzzy relation ~T is

perfectly antisymmetrical. Suppose M = (a 1 ,b 1 ,c 1 ,d 1 ) and N =
(a 2 ,b 2 ,c 2 ,d 2 ), when ~T(N,M) ~ 0 then c 1 < c 2 and ~T(M,N) = 0; ~T is

210
max-min transitive (the reader should refer to the original paper for
proof); and ~T("') is obviously irreflexive since aT(N,M) = 1 when c 1
~ c2. These properties make ~T a strict fuzzy order relation.
similarly, for the case of ":5 N," we define a fuzzy relation
sup (gN(x) A ~(x», (4.124)

xeR
which again is not a strict fuzzy order relation. We then define
another fuzzy relation
(4.125)
which is a strict fuzzy order relation.
Note that if ~T = 0 then 0T > 0, and vice versa. When ~T > 0, we
know "N ~ M" is true. When 0T > 0, we know "N :5 M" is true.
Algorithm
For comparing M = (a 1 ,b 1 ,c 1 ,d 1 ) and N

following steps are used.
step 1. For the ,,~ N" case, set A' e (-1,0]. And for the ":5 N"
case, set A" e [0,00).

step 2. Compute for the ,,~ N" case,
fN , A' (x) «A'+1) ~(x» / (1+",,-' hN(x», (4.126)
where
0, if x a2 ,
{
:5
1, if x ~ c2, (4.127 )
~N(x), if c 2 > x > a 2 ·
similarly, for the ":5 N" case, we obtain
(4.128)
211
where
0, if x s C2 '
{ 1, if x
1 -
~
~N(X),
d2,
if d 2 > x > c 2 •
(4.129)
step 3. Get comparison relations.

For the liN ~ Mil case, we compute
1 - sup (fN,A'(X) A ~(X». (4.130)

xeR
For the liN :s Mil case, we compute
1 - sup (gN A"(X) A ~(X». (4.131)

xeR '
If i3 T = 0 then cST > 0, and vice versa. When i3 T > 0, we know liN ~ Mil
is true. When cST > 0 we know liN s Mil is true.
The algorithm must be repeated for every pair of fuzzy numbers in
question. It should be noted that the comparison relations are more
"crisp" (i.e., i3 T !!E 1) when A' is the smallest, and cST !!E 1 when A" is
the greatest.
Numerical Example
Delgado et al.'s method will be applied to Example 10 in which
the final fuzzy ratings are: u 1 = (.20,.30,.50), u 2 = (.17,.32,.58),
and u 3 = (.25, .40, .70).
We shall give the detailed computational procedure on the
comparison of u 2 and u 3 to demonstrate the algorithm.
step 1. Set A' = 0 and A" = 0 as the requirement levels for the
accomplishment degree of the property ,,~ u 3 " and "s u 3 ," respectively.
Step 2. Compute comparison functions, fu A'(x), and gu A"(X),
3' 3'
respectively. Given the membership function of u 3 '
212
P.lx)
1.
.9
.8-
.7
.6
.5-
.4
.3
.2
.1
0 ~~~~~~--~~~--~~~x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.49 The functions hu (x) and Pu (x).

3 3
x - .25 ,
.40 .25 - .25 :s x :s .40,
Jlu (x) .70 - x ,

.70 - .40 .40 < x :s .70,
fl
3
0, otherwise,
we can derive hu (x) and P u (x) (see Fig. 4.49) as:

3 3
0, ifx :s .25,
hu (x)
3 { 1, ifx
Jl u
'"
(x), if x e (.25,.40),
.40,
3
and
1, if x '" .70,
Pu (x)
3 { 0, ifx :s .40,
1 - Jl u (x), if x e (.40,.70).
3
Given functions hu (x) and Pu (x), we can derive fu A'(x) and

3 3 3'
213
gu A" (x) (see Fig. 4.50) as:

3'
tu A' (x) «A' + 1) hu (x» I (1 + A' hU (x»

3' 3 3
gu A" (x) ( (1 - P u (x» I ( 1 + A" P u (x»

3' 3 3
or
0, if x .70
'"
gu
3'
A" (x)
{ 1, ifx
IlU (x),
;!;
if x
.40
E (.40,.70)
3
pIx) f!;h ).'

1. ~---..;;a:.;~-'l\:1""''''''''''''''''''''''''''''''''''''' 1",,,,,,,,,,,,,,,,,,,,,
.9 j
~
,8 I
I
,7 I
~
,6
,I'
I'
.5
.4 ,/'
.3
,"
~
.2
.1
f
~'
o """"'''''''''''''''''''''' x
o ,1 ,2 ,3 .4 .5 .6 .7 ,8 .9 1.
Fig. 4.50 The fu A'(x) and gu A"(x) functions.

3' 3'
step 3. The comparison relation for the "", u 3 " case is computed
as:
1 - .82 = .18,
where
214
x ~1517 , .17 ::s X ::s .32,
~u (x) .58 - X
, .32 < X ::s .58,
2 .26
0, otherwise.
Similarly, we can compute the comparison relation for the case of "::s
1 - sup (9U A"(X) /\ ~U (x»

xeR 3' 2
1 - 1 = o.
The a T (u 3 ,U 2 ) and '1 T (U 3 ,U 2 ) indices happen to be the intersections of

some continuous membership functions that are convex and normal (see
Fig. 4.51).
.7
.6
.5
.4-
.3
.2
.1
0 ~~~=L~~~-i~~~~~~x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Since 0T(U 3 ,U 2 ) = 0 and ~T(U3,U2) = 0.18, we conclude that the

statement "U 3 i1! U2 " has a degree of truth of 0.18.
We reapply the algorithm twice with the same A values on

215
.18 .28 .07
o o o
Thus, we conclude that: U3 > U2 > U1 •
Note
1. Recall that the comparison relations are obtained through
sup (fN(x) A ~(X» (4.132)

xeR
and
sup (gN(x) A ~M(x» (4.133)

xeR
where aT(o,o) and 7 T (o,o) denote the relationships associated to the
minimum as T-norm (or triangular norm). Eqs.(4.132) and (4.133) may
be generalized as:
a T= sup T(fN(x)'~M(X» (4.134)

xeR
sup T(gN(X),~(X» (4.135)

xeR
where T is a triangular norm (Schweizer and Sklar [S21]). That is, if
a and ~M(x) = b, then T(fN(x),~(X» can be replaced by any of

the forms in Eqs.(4.136a) through (4.136f). Similar remarks can be
T(a,b) min (a,b), (4.136a)
T(a,b) max (a,b), (4.136b)
T(a,b) Max (a+b-l,O), (4.136c)

216
T(a,b) a·b, (4.136d)
T(a,b) a+b-ab, (4.136e)
T(a,b) min (a+b,l). (4.136f)
Each T-norm satisfies the following properties. Let [x,y] be a
closed interval. Then
(a) T(a 1 ,b1 ) ~ T(a 2 ,b 2 ), if a 1 , a 2 , b 1 , b 2 e [x,y] and

a1 ~ a 2 and b 1 ~ b2 ,
(b) V a e [x,y], T(a,y) = T(y,a) = a,
(c) T(a,b) T(b,a), V a,b e [x,y],
(d) T(a,l) a for all a in I.
2. We suggest a few rules that would further reduce the
computation efforts:
a. Always compute ~T first. Once ~T~ 0 is found, 0T is quickly

assigned a number of zero.
b. Always compute one ~T on each pair of alternatives, i.e.,
compute either ~T(N,M) or ~T(M,N) but not both. Once ~T(N,M) ~ 0 is
found, ~T(M,N) = o.
c. The computed ~T value tends to be close to 1 when A' is set
to its minimum, e.g., A' = -.999. However, the ranking order will not
be altered due to different A' values. We suggest the use of A' = 0
as well as A" = 0 to preserve the simplicity.
3. The algorithm cannot be easily understood without reasonable
background knowledge on T-norm, fuzzy measures, and fuzzy relation.
The computation is, however, not so tedious - especially when A' = 0
and A" = o. Referring to Fig. 4.51, aT(N,M) can be found by taking
the rightmost intersection of ~(x) and fN(x) and 7 T (N,M) can be
obtained in a similar manner.

217
4.6 Ranking Using Fuzzy Mean and Spread
The method in this category compares the fuzzy numbers using two
criteria: the mean value and the spread of a fuzzy number. The
underlying assumption is that human intuition would favor a fuzzy
number with a higher mean value and a lower spread, i.e., higher
overall score with less variation.
4.6.1 Lee and Li's Approach
Lee and Li [L3] propose the use of generalized mean and standard
deviation based on the probability measures of fuzzy events to rank
fuzzy numbers. They assume two kinds of probability distributuion
for fuzzy events and derive corresponding indices as follows:
(1) Uniform distribution: f(M) =~ and M E U.
Given a fuzzy number M, its generalized mean value is calculated as:
J
SCM) x ~M(x) dx
(4.137)
The denominator measures the area under fuzzy number M (see Fig.
4.52). The standard deviation is defined as:
CJ'U(M) (4.138)
where SCM) is the support of fuzzy number M (see Section 3.2 Basics of
Fuzzy Sets).
218
P.(x)
M
1.
.9
.8-
.7
.6
.5
.4
.3
.2
.1
0
Fig. 4.52 The derivation of J ~(x) dx.

SCM)
P.(x) M
1.
.9
.8
.7-
.6
.5
.4
.3
.2
.1
o !--........-.-~-..---i-m-..----.--+n-.....-----.. x
Fig. 4.53 Triangular fuzzy number M.
When the fuzzy number M is triangular (see Fig 4.53), Eqs.(4.137)

and (4.138) may be rewritten as following:
1
'3 (1 + m + n), (4.139)
(]"U(M) (4.140)
where 1 inf SCM), ~(m) 1, and n sup SCM).

219
(2) Proportional distribution: f(M) k·~(x), Me U, where k is

the proportional constant. We have
S x 2 JlM(X) dx
SCM)
Xp(M) (4.141)
S [~(X)]2 dx
SCM)
S x 2 [JlM(x) ] 2dx
1/2
SCM)
- [x"
(jp(M)
[ S
SCM)
[ ~(X)]2dX
(M)) ,] (4.142)
When M is a fuzzy triangular number, Eq.(4.141) and (4.142) can be
rewritten as:
i (1 + 2m + n) (4.141a)
(jp(M) (4.142a)
The subscripts u and p denote the uniform and proportional
distributions, respectively.
When based on the proportional distribution, the fuzzy mean is
closer to m where JlM(m) = 1 and the standard deviation is smaller.

For example, assume that we have 1 = 2, m = 3, and n = 5. By applying
Eqs.(4.139) and (4.140), we get Xu(M) = 3.33 and (ju(M) 0.39,
respectively. By applying Eqs. (4.141a) and (4.142a), we get Xp(M)
0.24, respectively. This phenomenon suggests that a
stronger central tendency exists in the proportional distribution
case.
220
The choice of the uniform or proportional distribution is

arbitrary. One can use Eqs.(4.137) and (4.138) or Eqs.(4.141) and
(4.142) to calculate the fuzzy mean and standard deviation. Assume
that the mean values and spreads are calculated for the fuzzy numbers
Mi and Mj • The rules for ranking are the following:
Relation of Relation of
X(M i ) and X(M j ) (T (M i ) and (T(M j ) Ranking order
X(M i ) > X(M j ) Mi > Mj
X(M i ) = X(M j ) (T (M i ) < CT(M j ) Mi > Mj
Notice that the spreads, CT, are used only when the mean values are
equal; a smaller spread indicates a better fuzzy number.
Numerical Example
There are three investment projects to be evaluated according to
attributes X1 (cost) , X2 (environmental impact), X3 (estimated profit),
and X4 (maintenance cost/yr). The fuzzy final ratings are given
(see Fig. 4.54) as:
( 2, 3, 5), and U3 (1,4,4).
Assume the probability function is a uniform density function.

We then apply Eqs.(4.139) and (4.140) on Ui , i = 1,2,3 as:
i 1 2 3
6.47 3.33 3.00

0.51 0.39 0.50
221
#,(X)
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
o X
o 2 3 4 5 6 7 8 9 10
Fig. 4.S4 Fuzzy final ratings.
since X(U 1 ) > X(U 2 ) > X(U 3 ), the ranking order is U1 > U2 > U3 .
To show the computational process, we calculate the mean and variance
of U1 as:
1
"31 (l+m+n) "3 (S + 6 + 8.4) 6.47,
.S1.
Note
1. This method ranks fuzzy numbers based on two different
criteria, namely, the fuzzy mean and the fuzzy spread of the fuzzy
numbers. It is probably the most logical ranking method. The
difficulty of it is that the choice of the uniform or proportional
distribution is arbitrary. In addition, the results from Eqs. (4.137)
and (4.138) can sometimes be conflicting with those from Eqs. (4.141)
and (4.142). Lee and Li suggest that when conflict does occur, the
proportional distribution seems reasonable to use.
To justify their preference for the proportional distribution,
Lee and Li [L3] give the following example. Referring to Fig. 4.SSa,
222
we have M2 > M1 , regardless of which probability distribution is used.

That is, the result of using Eqs.(4.137) and (4.138) or Eqs.(4.141)
and (4.142) to rank Ml and M2 will not make any difference. As Ml
moves toward its right, the preference structure changes. In Fig.
M2 when the uniform distribution is
assumed, and M2 > Ml when the proportional distribution is assumed.
This is due to the fact that the x values with higher possibility are
more heavily weighted. In Fig. 4.55c d 1 = d 2 , we get Ml = M2 in the
proportional distribution case while Ml > M2 in the uniform
distribution case. When Ml moves further to its right, Ml > M2 for
both distributions.
2. Given a trapezoidal fuzzy number, M = (a,b,c,d). If we use
the uniform distribution as a probability function, then Eqs. (4.137)
and (4.138) can be rewritten as:
X(M) (4.143)
and
1 b4 ab 4 a4
<reM) { [b-a (4 -3- + 12) + !3 (c 3 _b 3 ) + (4.144)
1 d4 c 3d c4
d-c (12 -3- + 4 )]/[~ (-a-b+c+d)]} -
{(-a 2 -b 2 +c 2 +d 2 -ab+cd)/[3(-a-b+c+d)]}2,
repectively. The proof of conversion from Eq.(4.137) to Eq.(4.143)

and Eq.(4.138) to Eq.(4.144) can be found in Negi [N7, pp.403-407].
223
"(x)
1.
Fig. 4.55a Case of M2 > Ml for both uniform and proportional

distributions (no conflict).
"Ix)
1.
Fig. 4.55b Case of M2 = Ml for uniform distribution and M2 > Ml

for proportional distribution (conflicting results).
224
-I
p.(x) M,
1.
Fig. 4.55c Case of M2 < Ml for uniform distribution and M2 = Ml

for proportional distribution (conflicting results).
225
4.7 Ranking using Proportion to The Ideal

The method in this category compares fuzzy numbers to some
predetermined fuzzy ideals : the fuzzy maximum and the fuzzy minimum.
A fuzzy number outranks the others when its area overlaps the fuzzy
maximum the most, but overlaps the fuzzy minimum the least .
4.7.1 McCahone's Approach

McCahone [Mc1] developed a method that compares the fuzzy numbers
to some specified fuzzy ideals. The fuzzy ideals in this approach are
the fuzzy max and the fuzzy min. The ranking is determined by
measuring how much the proportional area of a fuzzy number contributes
to the fuzzy max and the fuzzy min. The higher percentage a fuzzy
number contributes to the fuzzy max, the higher it is ranked, and the
higher percentage it contributes to the fuzzy min, the lower it is
ranked .
P.(x) Mj
1. The f uzzy max 000
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig . 4 . 56 The contribution of Mi toward the fuzzy max.
Given two fuzzy numbers Mi and Mj (as seen in Fig. 4.56), the
contribution of fuzzy number Mi to the fuzzy max is determined by:
226
(4.145)
where
In Fig. 4.56, the shaded area denotes the contribution of Mi toward

the fuzzy max. The denominator of Eq.(4.145) guarantees that
P.lx) Mj
~
1. The fuzzy min 000
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.57 The contribution of Mi toward the fuzzy min.
Similarly, the proportion of fuzzy number Mi which contributed to

the fuzzy min is defined by (see Fig. 4.57):
~min(x) A ~M. (x) ] dx

J.
(4.146)
where
~min(x)
227
The shaded area of Fig.4.57 denotes the contribution of Mi toward the

fuzzy min.
After P(M i ) and N(M i ) , vi, are calculated, the fuzzy numbers
are ranked according to the first rule set.
(1) Rule set 1:
Relation I and Relation II Ranking order
P(M i ) > P (M.) N(M i ) < N(M.) Mi > Mj

] ]
N(M i ) = N(M j ) Mi > Mj

N(M i ) > N(M j ) unknown
P(M i ) = P(M j ) N(M i ) > N(M.)

] Mi < Mj
N(M i ) = N(M j ) Mi = Mj
N(M i ) < N(M j ) Mi > Mj
P(M i ) < P (M.) N(M i ) > N(M j ) Mi < Mj

]
N(M i ) = N(M j ) Mi < Mj

N(M i ) < N(M j ) unknown
For example, If P(M i ) > P(M j ) and N(M i ) < N(M j ), then Mi > Mj .
In case of unknown ranking order, McCahone proposed two composite
indices to discriminate Mi and Mj . The two composite indices are
defined as:
(4.147)
•
and
(4.148)
These indices are used to compare Mi and Mj based on the following

rules.
228
(2) Rule set 2:
Relation I and Relation II Ranking order

CP(M i ) > CP(M j ) CN(M i ) < CN(M j ) Mi > Mj
CP(M i ) < CP(M i ) CN(M i ) > CN(M i ) Mi < Mj
CP(M i ) CP(Mj) CN(M i ) CN(M j ) unknown
In case of unknown ranking order in the second rule set, the

following rules are proposed to distinguish Mi and Mj .
(3) Rule set 3:
Relation Ranking order

P(M i ) + N(M i ) > P(M j ) + N(M j ) Mi > Mj
P(M i ) + N(M i ) P(M j ) + N(M j ) Mi Mj
P(M i ) + N(M i ) < P(M j ) + N(M j ) Mi < Mj
The equations and comparison rule sets are good for ranking two
fuzzy numbers. When more than two fuzzy numbers are compared, the
following steps must be taken:
step 1. Find the fuzzy max and the fuzzy min based on all fuzzy
numbers in question (see Fig. 4.58).
step 2. Compute P(M i ) and N(M i ), vi.
step 2.1. Rank the fuzzy numbers according to P(M i ) alone (the
bigger the better).
step 2.2. Rank the fuzzy numbers according to N(M i ) alone (the
smaller the better).
step 3. Compare the two ranking orders found in steps 2.1 and
2.2.
229
The fuzzy max ••• The fuzzy min 000
#L(X) M1 M3 M2
1. o -,
'. . .,
, ,
.9 ~ ~;. *...
. . ."
"
,
.8 "
'
.7
.6 .".'
.5
.4
.'"
"
.3
.2
..
'
.1
0 ~~~~--~-+--~~~~--~--~~
.'. x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.58 The fuzzy max and the fuzzy min for more than
two fuzzy numbers.
step 3.1 If the two ranking orders are different, we shall
pick those fuzzy numbers that contribute to the conflict and perform
pairwise comparisons on them using the rules described earlier.
step 3.2 If the two ranking orders are identical, the
algorithm stops.
Numerical Example
One wants to select one out of three investment tools for
investing a large sum of capital. The tools are commodity, stock, and
gold/diamonds. The attributes used are risk of losing capital (Xl)'
inflation effect (X 2 ), and cash availability (X 3 ). The data is fuzzy.
The final ratings are given as in Figure 4.59. We shall apply
McCahone's [Mc1] method to pick the best investment tool.
Using Eqs. (4.145) and (4.146), we obtain P(M i ) and N(M i ), i
1,2,3, as follows:
i 1 2 3
.610 .875 .875
.670 1 .560
230
The fuzzy max ...

P.(x)
1.
13 The fuzzy min 000
.9
.8
u
.7
.6
.5
.4
.3
.2
.1
o ~~--~--~-+--~--~~--~~~~x
o 2 3 4 5 6 7 8 9 10
Fig. 4.59 Fuzzy final ratings.
We shall compute P(U 1 ) and N(U 1 ) to demonstrate the computational
procedure. They are calculated as:
9
S ~max(x) A ~U (x) )dx
3 1
9
.61
S ~U (x) ) dx
3 1
9
S [~min(x) A ~U (x)) dx
3 1
.67.
9
S [~U (x)) dx
3 1
Based on the P(U i ) index alone, U3 = U2 > U1i while on the N(U i )
index alone, we have U3 > U1 > U2 . The two ranking orders are
different. The alternatives which contribute to this conflict are U1
and U2 . Further investigation is needed. Since P(U 1 ) = .61 < P(U 2 )
.67 < N(U 2 ) = 1, rule set 1 cannot distinguish
between U1 and U2 . Therefore, the composite indices must be used for
further comparisons.
using Eqs.(4.147) and (4.148), we obtain

231
.48
and
.52.
Similarly, CP(U 2 ) = .47 and CN(U2 ) = .53. Since CP(U 1 ) > CP(U 2 ) and
CN(U 1 ) < CN(U 2 ), we conclude that U1 > U2 , based on rule set 2. We
have completed the algorithm and have come to the conclusion that
U3 > U1 > U2 ·
Note
1. In McCahone's method, the use of some predetermined fuzzy
ideals, (the fuzzy max and the fuzzy min), as the basis for comparison
is similar to those methods in section 4.3 (ranking using Hamming
distance). The uniqueness of McCahone's method is that it measures

the areas that overlap with the fuzzy ideals rather than the areas
that are not overlapping with the fuzzy ideals, as seen in measuring
the Hamming distance. A fuzzy number that overlaps more with the
fuzzy max and less with the fuzzy min is considered better in
McCahone's method.
2. The problem of McCahone's method is the same as that found in
section 4.3. Namely, the locations of fuzzy numbers are not
considered when comparing fuzzy numbers. For example, given four
fuzzy numbers Ml , M2 , M3 , and M4 as seen in Fig. 4.60, the ranking
order is obvious: M4 > M3 > M2 > MI' However, using McCahone's

method, P(M 4 ) 1 and P(M 3 ) = P(M 2 ) = P(M l ) = 0; while N(M l ) = land
N(M 2 ) = N(M 3 ) N(M 4 ) = o. By applying rule set 1, we obtain M4 > M3
= M2 > Ml · McCahone's method fails to discriminate M2 and M3 , even
though M3 > M2 is clear by observation.
The indiscrimination case can be generalized as follows. If
there is a set of fuzzy numbers {M l , M2 , ... , Mn} where Ml is the fuzzy

232
min and Mn is the fuzzy max, and Ml and Mn do not overlap with other
fuzzy numbers, then M2 , ... ,Mn _ 1 cannot be distinguished by McCahone's
method.
JL(X) The\fUZZY min

1. 0
.9 0
o
.8
.7
.6
.5
.4
.3
.2
.1
oo .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
x
Fig. 4.60 Indiscrimination case of McCahone's method.

233
4.8 Ranking Using Left and Right Scores
Given fuzzy numbers M1 , ... , Mn' the left and right scores refer
to the intersections of a fuzzy number Mi with the fuzzy min and the
fuzzy max, respectively. Fig. 4.61 illustrates the aforementioned
notions (using Chen's [C12] fuzzy max and fuzzy min) where R i , i 1,
2, is the right score and L i , i = 1, 2, is the left score. The
methods in this category utilize either the right score or both scores
to derive the total score for each fuzzy number. The fuzzy number
with a higher total score is considered better.
\
IL(X)Chen's fuzzy min Chen's fuzzy max
1, :',
' ,
/
,,
.,
,,
Fig. 4.61 Illustration of the left score and the right score.
Jain [J2,J5] defines a fuzzy max as:
(xI x max ) k ' k > 0 ,

{ (4.149)
0, otherwise,
where x is the support of Jain's fuzzy max and x max is the biggest
value of the supports. The intersection of the nonincreasing part of
a fuzzy number and Jain's fuzzy max is the crisp score of the fuzzy
number. The fuzzy number with the highest score is the best choice.
234
Chen (C12] found that Jain's method is not sufficient to
discriminate fuzzy numbers in some cases. For example, when the
nonincreasing parts of two fuzzy numbers overlap, Jain's method will
rank them equally, even though one is obviously better than the other.
Thus, he proposed the use of both nonincreasing and nondecreasing
parts of a fuzzy number to obtain right and left scores, respectively.

The scores are obtained in the same manner as those in Jain's method
except that a fuzzy min is used in addition to a fuzzy max. Note also
that the fuzzy max defined in Chen's method is slightly different from
that in Jain's method.
Chen and Hwang (C13] pointed out that Chen's method considered
only the relative locations of fuzzy numbers on the horizontal axis
and ignored the absolute location of fuzzy numbers on the horizontal
axis. As a result, Chen's method may result in counter-intuitive
answers. Thus, they redefined a fuzzy max and a fuzzy min (which are
different from those found in Jain's and Chen's methods). The
procedure for obtaining the left and right scores is the same as that
used in Chen's method.
4.8.1 Jain's Approach

Jain (J2,J5] proposes a ranking method which computes for each
fuzzy number a crisp score. The fuzzy numbers with higher scores are
considered better. The crisp score is computed by taking the inter-
section of the nonincreasing part of a fuzzy number with the fuzzy max.
Jain points out that, to properly compare the fuzzy numbers, we
should consider both the fuzzy max associated with various fuzzy
numbers and the fuzzy numbers themselves. For that concept, a
fuzzy max for fuzzy numbers M1 , ... , Mn is defined as:
k
i-Lmax(X) (x / X max ] , k > 0, (4.150)
235
where k is an integer indicating the OM's attitude toward risk, and

xmax is the maximum x value of Mi'S support set. Note that x max and
the support set of Mi' SeMi)' are defined as:
n
sup [ V S (M . ) ] (4.151)
i=l 1
and
{x I ~.(x) > O}. (4.152)

1
Jain's fuzzy max suggests that when x is approaching x max '

~max(x) will be closer to 1. When x is far away from x max ' ~max(x)
will be closer to o. The value k indicates the OM's attitude toward

risk. When some k < 1 is chosen, the OM is said to have a risk-averse
attitude. When some k > 1 is chosen, the OM is said to have a
risk-taking attitude. The OM is said to have a risk-neutral attitude
when k = 1 is used, as shown in Eq.(4.150). The choice of the k value
is problem- and OM-dependent.
After Jain's fuzzy max is determined, an optimal set (also fuzzy)
is formed for each fuzzy number Mi as O(i), whose membership function
is defined as:
sup [~max(X) A ~M. (x)]. (4.153)

x 1
~O(i) represents the degree of optimality of fuzzy number Mi. Since

this is a single real number e [0,1], we can use it as the score of
the fuzzy number Mi. Graphically, ~o(i) is presented in Fig. 4.62.
~o(l) and ~o(2) are the crisp scores of M1 and M2 , respectively. They
are the intersections of the nonincreasing parts of Mi , i = 1, 2, and
Jain's fuzzy max. Since ~o(l) > ~O(2), we know M1 is better than M2 •
236
P.(x)
Jain's fuzzy max
1.
1'0(1)
00:::0(0::)0:::::::::::::: : ::::0 ;
1'0 2 ./
Fig. 4.62 Jain's crisp scores of fuzzy numbers H1 and H2 •
Numerical Example
Jain's method will be applied to Example 10. We are to use
Eq.(4.153) to find ~o(i). The results are summarized (see Fig.4.63)

as:
i 1 2 3
.56 .60 .70
The ranking order is U3 > U2 > U1 .
Notice that ~o(i) is simply the intersection of the right leg
(nonincreasing part) of ~u.(x) and ~max(x) as demonstrated in Fig.

~
4.63. We shall compute ~o(l) by taking the intersection of the right
leg of ~U1 (x) and ~max(x) to illustrate this point. Bo.th U1 and
Jain's fuzzy max are triangular in shape, the nonincreasing parts of
~u (x) and Jain's fuzzy max can be represented as:

1
.50 - x • 30:5 :5 . 50 .
~U (x) . 20
X
1
x
~max(x) = 7i
The intersection of ~U (x) and ~max(x) determines ~O(1) .56.

1
237
P,(x)
Jain's fuzzy max
1.
.9
.8
0(3)
.7 --~(2r-------
.6
.5 0(1)
.4
.3
.2-
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.63 Jain's crisp scores of U1 , U2 , and U3 .
Note
1. It is pointed out by many (Baldwin and Guild [B2], Bortolan
and Degani [B34], Chen [C12], Zimmermann [Z31], etc.) that this method
gives counter-intuitive results in some cases. For example, given M1
and M2 (see Fig. 4.64) which share the same "right leg," Jain's method
assigns the same score to both, i.e., M1 = M2 , even though M2 is

clearly better than M1 . This counter-intuitive result is attributed
to the fact that only partial information contained in a fuzzy number
(i.e., the nonincreasing part) is considered. Since the nondecreasing
part of a fuzzy number is ignored, it is not surprising that some
counter-intuitive results may occur.
2. As noted by Chen [C12], if some of the fuzzy numbers Mi
contain negative support (i.e., for some x < 0 in {xl gM. (x) > O}),
~
then the membership function gmax(x) becomes negative when k is an odd
integer (see Eq.(4.150)). This contradicts the definition of a
membership function. (Recall that any membership function value E
[0,1].) Furthermore, when k is an even number and gM. (x) contains

1
negative support, gmax(x) = (x / Xmax)k may be greater than 1. Again,
this contradicts the definition of a membership function. These

238
remarks are particularly fatal to Jain's method when the supports of a
fuzzy number Mi are not confined to positive numbers.
IL(X)
Jain's fuzzy max
1. , .. :
.9
.8
.7
.6
.5
.4
.3
.2
.1 .
0
0 .1 .2 .3 .4 .5 .6 .7
Fig. 4.64 Counter-intuitive example of Jain's method.
4.8.2 Chen's Approach

Chen [C12] identified the inconsistency caused by Jain's method
and suggested some modifications to it. First of all, a fuzzy max
should include the minimum and maximum x points associated with the
fuzzy numbers in question. As a result, he redefined a fuzzy max
(for M1 , ... , Mn fuzzy numbers) as:
Ilmax (x) (4.154)
where k is an integer indicating the OM's attitude toward risk; and
x max and xmin are the maximum and minimum numbers in the support set
of Mi' S(M i ). They are defined as:
n
sup [V S(M i )], (4.155 )
i=l
239
n
inf [V SeMi)]' (4.156)
i=l
and
{x I ~,(x) > O}. (4.157)
1.
Secondly, a fuzzy min is introduced to incorporate the information

contained in the left part of a fuzzy number into the ranking process:
x - x
max ]k, (4.158)
~min(x)
xmin - x max xm1.'n s X ::s xmax '
where xmax ' xmin' and k are defined as those in Eq.(4.154). Chen's
fuzzy max and fuzzy min are presented in Fig. 4.65.
P.(X)Chen's fuzzy min M1

\
M2 Chen's fuzzy max
1, !'-"/ ,;
, I+R(2)
"-------------:------
min
Fig. 4.65 The left and right scores by Chen's method.
There are two scores associated with each fuzzy number Mi'
namely, the right score ~R(i) and the left score ~L(i) (see Fig.4.65).
They are defined as:
sup [ ~max(x) A ~,(X)] (4.159)

X 1.
240
tlL (i) sup [ tlmin(x) A ~. (X)]. (4.160)

X ~
Note that when Mi is a continuous, convex, and normal fuzzy number,

tlR(i) can be determined by taking the intersection of the
non increasing part of ~. (x) and tlmax(X) and tlL(i) is the intersection
~
of the nondecreasing part of ~_ (x) and tl . (x) .

. l'li m~n
tlR(i) and tlL(i) together guarantee the full utilization of

information contained in Mi' Since the higher tlR(i) values indicate
better fuzzy numbers and higher tlL(i) values indicate worse fuzzy
numbers, the total score of Mi can be defined as:
(4.161)
where we use (1 - tl L ) to reverse the effect of the tlL(i) value. The

division of 2 makes the total score of Mi fall in [0,1]. The higher
tlT(i) value determines the preferred fuzzy number Mi'
Numerical Example
Chen's method will be applied to Example 10. Assuming k 1,
Chen's fuzzy max and fuzzy min are defined as:
x - x .
m~n x - .17
tlmax(X) .53 ], .17:s x:s .7,
x - .7
tlmin(x) x. - xmax --~--.'5~3~--]' .17 :s x :s .7.
m~n
The fuzzy number U1 can be represented by:

241
x - .2 .2 :s x :s .3 (left leg),
{
.1
J.l U (x) .5 - x
1 • 3 < x :s .5 (right leg) .
.2
The intersection of the left leg and J.lmin(X) is given as J.l L (l) = .79
The intersection of the right leg and J.lmax(X) is J.lR (l) = .45. The
total score of U1 is then calculated as:
We can obtain J.lT(i), i 2,3, in a similar manner. They are
summarized as:
i 1 2 3
J.lR{i) .45 .52 .64
J.lL{i) .79 .78 .66
J.l T (i) .33 .37 .49
By comparing the three J.l T values, we conclude that U3 > U2 > U1 •
Note
1. It is noted by Chen [C12] that the J.lT(i) index alone may not
provide discrimination ability for some fuzzy comparison cases. For
example, given M1 and M2 as shown in Fig. 4.66 where both fuzzy
numbers assume the same total score, Chen's method cannot discriminate
between them. To resolve this difficulty, he suggested that we can
locate points m1 and m2 , such that J.lM (m 1 ) = 1

1
x-axis and the fuzzy number which has a higher m value is preferred.
In this case, M2 is preferred.
2. In most ranking procedures, the membership function J.lM. (x) is

~
restricted to the normal form, i.e., there exists at least one x,
point such that J.lM. (x') = 1. But in this method, Chen removes that
~
242
L,:
Chen's fuzzy max
M1~
JL(x)
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
o ~~ __ ~-L~~__~__~~~~__~~ X
o .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.66 Fuzzy numbers with equal total scores by Chen's method.
JL(x)
M1
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.67 Two fuzzy numbers of unequal heights.
restriction by assuming ~M. (x) e [O,h] with 0 < h s 1. The reason for
~
that relaxation is explained as follows. We have two fuzzy numbers M1
and M2 (as shown in Fig. 4.67), where M1 is a normal fuzzy number and
M2 is not. M1 and M2 have the same meaning, eg., about 0.4. These
fuzzy numbers may be obtained independently from two persons
evaluating the same character. The person who expresses M1 has more
confidence in describing this character than the person who expresses
M2 does.
243
JL(X) Mn
1. -_. -- ;~- ----------- --------------------------------
x X
max
Fig. 4.68 Case of n fuzzy numbers of unequal height.
without consideration of the confidence levels, we can say that M1 and
M2 are the same.

If we accept the foregoing argument, the following formulas may
be useful in calculating the total scores for general fuzzy numbers.
Given fuzzy numbers Mi , i = 1, ... ,n, each has three vertices with
coordinates (ai,O), (mi,h i ), (bi,O) as shown in Fig. 4.68. The
membership functions of Mi , Chen's fuzzy max and fuzzy min are defined
(set k = 1) as:
!
hi{x - ai)/{m i - a i ), ai ::s x ::s mi ,
(x) (4.162)
~i hi{x - bi)/{m i - b i ) , mi ::s x ::s bi ,
0 otherwise,
h (x - x min )
(xmax - x min ) (4.163)
~max{X)
o , otherwise,
244
h (x - xmax)
(xmin - x max ) (4.164)
/.Lmin(x)
o , otherwise.
The left, right, and total scores of fuzzy number Mi , Vi, are then
given as:
h h.
1
(b i - x min )
/.LR (i) (4.165)
hi (Xmax - Xmin ) - h (m. - b i )
1
h h.
1
( x max - ai )
/.LL (i) (4.166)
hi (Xmax - x min ) + h (m i - ai )
(4.167)
Similar to the normalized case, Chen [C12] provides detailed
computation formulas for the k = 2 and k = 21 cases. Again, k = 2
represents the risk-prone membership function while k = i represents
the risk-averse membership function. The formulas for computing
/.LT(i), however, are much more complex Since they don't provide
significant advantages over the k = 1 case, we will not discuss them

here. The interested reader should refer to Chen's [C12] paper.
3. Recall that Jain's method cannot tell the difference between
M1 and M2 in Fig 4.64. This indiscrimination situation can be
resolved by Chen's method. However, Chen's method cannot distinguish
the following case.
considering the case of Fig. 4.69, both M1 and M2 not only have
the same crisp scores but also the same m = .5 value such that
245
./!
Chen's fuzzy min Chen's fuzzy max
:'" \ \ M, M2
.8
.7
.6
.5
.4
.3
.2
.1
0 ~~--~--4---~~--~--+---r-~--~ X
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.69 The indiscrimination case by Chen's method.
gA (.5) = gA (.5) = 1. Thus, M1 and M2 are considered equivalent by

1 2
this method. This phenomenon has been noted by Li and Lee [L7] as
the indiscrimination between symmetrical fuzzy numbers with equal
means.
Another difficulty may arise when applying Chen's method. That
is the ignorance of the absolute locations of fuzzy numbers on the
x-axis by Chen's method. For instance, we have M1 and M2 in Fig.
4.70a and Mi and M2 in Fig. 4.70b. We find that gT(M 1 ) = gT(Mi)
.417 and gT(M 2 ) = gT(M 2 ) = .545. Since M1 and Mi are obviously

different fuzzy numbers, their crisp scores should not be the same.
The equality is attributed to the fact that this method considers only
the relative locations of fuzzy numbers but not their absolute
locations on real line R.

246
P.(x) M2
M1
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig 4.70a Illogicality of Chen's method - part (a).
p.(x)
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 ~~--,---~~--~--~-+--,-~~~x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.70b Illogicality of Chen's method - part (b).
4.8.3 Chen and Hwang's Approach
Chen and Hwang [C13] found that Jain's method may result in
anti-intuitive results while, Chen's method ignores the absolute
locations of the fuzzy numbers in question.
To eliminate the above problems, they redefined a fuzzy max and a
fuzzy min (see Fig. 4.71) as:

247
".(x) M1
1 Chen and Hwang's ~Chen and Hwang'
. ""fuzzy min fuzzy max . /
..... .)
L"/~R(2)
<. - •••••••• -.iLR(1i .
.5
~~--~--~~--~--r_~~~~~~ X
o .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.71 The left and right scores by Chen and Hwang's method.
J.Lmax(x) { x, o :s X :s 1,
0, otherwise,
(4.168)
{
1 - x, O:sx:s 1,
J.Lmin (x) 0, otherwise.
(4.169)
Chen and Hwang's fuzzy max is the same as Yager's fuzzy max defined in
Section 4.3.1. Their fuzzy max and fuzzy min are defined in a manner
such that absolute locations of fuzzy numbers can be automatically
incorporated in the comparison process. Note that their fuzzy max and
fuzzy min are used for all fuzzy comparison cases.
The computational procedure is the same as that of Chen's except
that the definition of J.Lmax(x) and J.Lmin(X) is different. The left
utility score of each fuzzy number Mi is defined (see Fig. 4.71) as:
sup min [J.Lmin(X), ~. (x)]. (4.170)

x 1
The J.LL(i) score is a unique, crisp, real number in [0,1]. It is the

248
maximum membership value of the intersection of fuzzy number Mi and

the fuzzy min. The right utility score may be obtained in a similar
manner (See Fig. 4.71).
sup min [~max(x), ~. (x)]. (4.171)

x 1
Again, ~R(i) is a crisp number in [0,1].

Given the left and right scores, the total score of a fuzzy
number Mi is defined as:
(4.172)
Numerical Example
We shall apply Chen and Hwang's method to Example 10. The left
and right scores are presented in Fig. 4.72. The total score of
fuzzy number U1 is calculated as follows:
P.(x)
1. ' .
.9
.8 .73
.7
.6 , .................. _.. ,9.4~
.5
.4
.3
.2
.1 .' ••.......
o ~'-'--.--L.,"--'---r----.--+--:...-+---.--....-~ X
o .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig 4.72 The left and right scores for the fuzzy final ratings.
249
J,LL(U 1 ) sup min (J,Lmin(x) , J,LU (x» .73,

x 1
J,LR(U 1 ) sup min (J,Lmax(x), J,LU (x» .42,

x 1
J,LT (U 1 ) [J,LR (U 1 ) + 1 - J,LL(U 1 )] / 2 .35.
We can obtain J,LT(i) , i 2,3, in a similar manner. The results are

summarized as:
i 1 2 3
J,LR (i) .42 .46 .54
J,LL (i) .73 .72 .65

J,LT (i) .35 .37 .44
By comparing the J,LT scores, we conclude that U3 > U2 > U1 •
Note
1. We shall compare the results of the three methods. Given
three fuzzy numbers M1 , M2 , and M3 as shown in Fig.4.73a, the ranking
order should be M3 > M2 > M1 , based on human intuition. Table 4.5a
summarizes the crisp scores obtained by the three approaches.
Table 4.5a Crisp Scores for Mi' i = 1, 2, 3

Fuzzy numbers Jain Chen Chen and Hwang
M1 .557 .375 .335

M2 .557 .415 .360
M3 .874 .786 .550
Jain's method considers only the right utility of the fuzzy

number. The ranking order is M3 > M2 = M1 which is against human
intuition. On the other hand, Chen [C12] and Chen and Hwang [C13]
250
give ranking orders that comply with human intuition.

Although Table 4.5a shows no significant difference between the
last two methods, the problem caused by Chen's method can be
demonstrated by the following example. By shifting M1 , M2 , and MJ to
their right as shown in Fig. 4.7Jb, we get three fuzzy numbers M1 ',
M2 ', and MJ '. Table 4.5b summaries the results generated by Chen's as
well as Chen and Hwang's methods. It shows that using Chen's method,
the total scores for M1 ', M2 ', and MJ ' remained the same as those in
Table 4.5a. This is not logical because the absolute location of the
fuzzy number is not considered at all. On the other hand, Chen and
Hwang's method not only preserves the ranking order but also considers
the absolute location of each fuzzy number.
Table 4.5b Crisp Scores for Mi , i = 1, 2, 3

Fuzzy numbers Chen Chen and Hwang
Ml .J75 .400
M2 .415 .440
MJ .786 .777
P.(x)
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 ~-+--T-~--~~~~--+--T--~-TX
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 4.7Ja comparison among the fuzzy score methods - part (a).
251
~~
1.
.9
~
~
~
.5
A
~
.2
.1
0 x
0 .1 .2 ~ .4 .5 .6 ~ .8 ~ 1.
Fig. 4.73b comparison of fuzzy scoring methids - part (b).
2. Since the crisp scores may, in a way, be treated as the mean
values of fuzzy numbers by Chen and Hwang's method as well as Chen's
method, the symmetrical fuzzy numbers of equal means and different
spreads will be rated equivalent (see Fig. 4.69). In other words, the
indiscriminated cases where n fuzzy numbers share the same XO value
where ~M. (x o ) = 1, i = 1, ... ,n, still cannot be resolved by Chen and

1
Hwang's method.
252
4.9 Ranking with centroid Index

The methods in this category find the geometric center of a fuzzy
number, Mi. Each geometric center corresponds to an x value on the
horizontal axis and a y value on the vertical axis (see Fig. 4.74),
i.e., (xo'Yo). The ordering of the fuzzy number is conducted either
on Xo values alone or with the help of both Xo and Yo values. The
methods in this category include Yager's [Y5] method which calculates
for each fuzzy number a Xo value, and Murakami, Maeda, and Imamura's
[M23] method which calculates both Xo and Yo for each fuzzy number.
P.(x)
M
1.
Yo ----------------- ---------,
Fig. 4.74 Centroid point of fuzzy number M.
4.9.1 Yager's centroid Index

Yager [Y5] proposed a ranking index
1 1
J g(x) ~. (x) dx / J ~M. (x) dx, (4.173)
o ~ 0 ~
where g(x) is treated as a weight function measuring the importance of

the value Xi while the denominator serves as a normalizing factor
253
whose value is equal to the area under the membership function ~.'
1
When g(x) x (linear weight), Eq.(4.173) gives the xo of the
geometric center (see Fig. 4.74). The value of xo may be seen as the
weighted mean value of fuzzy number Mi' The fuzzy numbers with higher
xo values are considered better.
Numerical Example
Yager's centroid index is applied to Example 10. Assume that
g(x) = x in Eq. (4.173). The fuzzy number U1 is characterized by the

membership function
x - .2 0.2 x
,; ,; 0.3,
{
.1
lJ. u (x) =
.5 - x
1 0.3 < x ,; 0.5.
.2
The centroid X01 is computed by:
1 1
f x IJ. U (x) dx / f lJ. u (x) dx
o 1 o 1
.30 2 .50 5
f (x * ~) dx + f (x * ~) dx
.20 .1 .30 .2
0.3 .50 5
f (x-.2) dx + f (~) dx
0.2 .1 .30 .2
.013 + .037
.333 .
. 150
similarly, we obtain x o2 .37 and x03 .45. Thus, the ranking
order is u 3 > u 2 > u1 .
Note
Yager's Xo index measures the general mean of the fuzzy numbers.
It is not surprising to see that the Xo index alone provides very poor
discrimination ability. For example, we have M1 and M2 in Fig. 4.75a

254
P.(x)
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
00 2 3 4 5 6 7 8 9 10 x
Fig. 4.75a symmetrical fuzzy numbers of equal mean and

different spreads.
P.(x)
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 ~~--~--~~~4-~r-~--~--r-~x
0 2 3 4 5 6 7 8 9 10
Fig. 4.75b Nonsymmetrical fuzzy numbers with the same spread.
and both have the same fuzzy mean but different fuzzy spreads. Since
both fuzzy numbers Ml and M2 have the same Xo value, Ml = M2 .

According to Lee and Li [L3], a fuzzy number with larger mean and
smaller spread should be ranked higher. If one follows Lee and Li's
criteria, M2 should be better than M1 .
Let's consider another example as shown in Fig. 4.75b, where both
Ml and M2 have the same spread. By using Eq.(4.173), where g(x) = x

255
is assumed, we obtain the same Xo value for both M1 and M2 , i.e., M1

M2 • In this case, Xo index alone cannot discriminate M1 and M2 •
4.9.2 Murakami et al.'s Approach

Murakami et al. [M23] propose two ranking methods. One is the
a-cut method which is the same as Adamo's [A1] index and will not be
discussed in this section. The second one is to find the geometric
center (xo'Yo) for each fuzzy number. The ordering of fuzzy numbers
is performed using both Xo and Yo values. The higher Xo and/or Yo
values determine the preferred fuzzy number. The centroid point
(Xo'Yo) for fuzzy number M (see Fig. 4.74) is defined as:
1 1
J x ~M(x) dx / J ~M(X) dx, (4.174)
o 0
1 1
J x ~(X) d~(X) / J ~(X) dx. (4.175)
o 0
The denominator in Eqs.(4.174) and (4.175) serves as a

normalizing factor whose value is determined by the area under the
membership function, ~M(x). The derivation of Xo and Yo is illutrated
in Figs. 4.76a and 4.76b.
According to Murakami et al., the optimal choice is the fuzzy
number that attains the maximum value on either axis. For example,
when n fuzzy numbers Ml , ... , Mn are compared based on x o . and Yo.
1 1
indices, the fuzzy number ~ is said to be optimal if x ok= max [Xoi ]'
i
256
/L(X) M
1........... --... --_. _......... -
Fig. 4.76a The derivation of xo.
/L(X)
1... _..... __ ._-_._-_._- --- M
Fig. 4.76b The derivation of Yo.
There is not always a unique optimal choice. For example, Mk may have
a maximum xoi value but Ml has a maximum Yoi value. Thus the choice
of Mk or Ml is not clearly defined. Murakami et al. suggested that
the OM may make a decision based on his subjective judgment of which
index is more important. From there, the choice of an optimal fuzzy
number solely depends on the most important index.

257
Numerical Example
Murakami et al.'s method is applied to Example 10.
Fuzzy number U1 is characterized by membership function
x - .2 0.2 ::s X ::s 0.3,
{
.1
Jl u (x)
.5 - x 0.3 < x ::s 0.5.
1
.2
The centroid point (x 01 'Y01) is computed by:
1 1
x o1 ~ x JlU1 (X) dx / ~ JlU1 (x) dx
.30 .50
f (x X-.2) dx + f (x .5-X) dx
.1 .2
.20 .30
0.3 2 .50 5
f (x-.) dx + f (. -x) dx
0.2 .1 .30 .2
.013 + .037
.333 •
. 150
1
Since U1 is a fuzzy triangular number (see Note 1), Y01 = 3·
We can obtain x oi ' Yoi' i = 2,3, in a similar manner. They can
be summarized as:
i 1 2 3
.33 .37 .45
.33 .33 .33
since all Yoi' i = 1, 2, 3 are the same, the ranking order depends
solely on the Xo index as: U3 > U2 > U1 .

Note
1. Notice that the Yoi values in our example are the same, i.e.,
Y01 = Yo2 = Y03 1/3. This is not an unexpected result. Indeed, the
centroid point Yo e [0,1] is always 1/3 for a triangle and is always
1/2 for a rectangle. Thus, the ordering of fuzzy numbers is simply a
matter of comparing xo values. This argument is confirmed by Bortolan
and oelgani [B34] and Lee and Li [L3], who agree that xo seems to be
the only rational index for comparing fuzzy numbers.
2. Yager's Xo index may be seen as a general formula for
calculating both Xo and Yo given different g(x) function. When g(x)
x, Murakami et al.'s Xo can be calculated. When g(x) = ~M(x)/2,
Murakami et al.'s Yo can be derived (Lee and Li [L3]). Since Yager's

Xo index is the same as Murakami et al.'s Xo index, the problem of
Murakami et al.'s Xo index is the same as that of Yager's Xo index.
Thus, Murakami et al.'s method is not logically sound either.
3. When fuzzy numbers Mi , i = 1, ... ,n, take irregular shape, Xo
and Yo may be combined as:
(4.176)
where w1 and w2 are the relative importance of Xo and Yo indices

expressed by the OM. This index Ci takes into account the OM's
subjective judgements and resolve the conflicting solutions possibly
observed when Xo and Yo act individually. The higher Ci value
indicates a better fuzzy number Mi.
259
4.10 Ranking Using Area Measurement

Yager [Y18] proposes a ranking index, F, which measures the area
between the mean of the membership function to the crisp number o.
Given a fuzzy number M with the membership function ~M shown in Fig.
4.77, the ranking index measures the shaded area between the dashed
line and the vertical-axis . The point (x,a o ) on the dashed line
represents the average value of the elements on the x-axis having at
least a o degree of membership . The dashed line represents all the
average values of the x elements at various a levels. The fuzzy
number with a higher index value is considered the preferred one.
Given fuzzy number M, Yager's index is defined as:
F(M) f (4.177)
o
where a max = sup ~M(x) and X(Ma) represents the average value of the
x
elements having at least a degree of membership .
/L(x)
1.
Fig. 4.77 Illustration of Yager's F ranking index .

260
p.(x)
1.
.9-
.8
.7
.6
.5
.4
.3
1
.2
.1
0 1
~-+--4---~-+--4---~-T--~--~-T
.3 .5 .6 .7 .8 .9 1.
X
0 .1 .2 .4
Fig. 4.78 Discrete fuzzy number M.
The computation of X(Ma> can best be explained by the following
example. Given a fuzzy number M as shown in Fig. 4.78, the various
a level sets are:
M {.1, .2, .3, .4, .5, . 6}, 0 < a s .2,

a
Ma {.3, .4, .5, .6}, .2 < a s .4,
M {.4, .5, • 6}, .4 < a s .6,

a
M {.5, • 6}, .6 < a s .8,

a
Ma {.6}, . 8 < a s 1.0 .
From this set of Ma' we can compute X(Ma> as:
X(Ma> 2.1 / 6 .35, 0 < a s .2,
X(Ma> 1.8 / 4 .45, .2 < a s .4,
X(Ma> 1.5 / 3 .50, .4 < a s .6,
X(Ma> 1.1 / 2 .55, .6 < a s .8,
X(Ma> 0.6 / 1 .60, . 8 < a s 1.0 .

261
The F(x) index is computed as:
.2 .4 1
F(M) I .35 da + I .45 da + ..• + I .60 da
o .2 .8
(.35)(.20) + (.45)(.20) + ••• + (.60)(.20) .49.
When the fuzzy number M is piecewise linear, as shown in Fig.

4.79, the X(Ma) can be easily determined as the line passing two
points, (ml,o) and (m 2 ,amax ) where ml is the mean of the support of M
and m2 is the mean of the values having a max membership value.
can be obtained using the a-cut. Referring to Fig. 4.79, where M is
characterized by:
x - 3, 3 :S X :S 4,
IlM(x)
{ 1
9 - x
-3-
4 < X
6 < X
:S
:S
6,
9.
The a-cut of fuzzy number M is Ma [a+3, 9-3a]. Consequently, we

can compute
X(Ma) [a+3+9-3a] / 2 -a + 6
1
F(M) I X(Ma) da 5.5.
o
The shaded area between X(Ma) and the vertical axis in Fig. 4.79
denotes F(M). The higher F(M) index value denotes a better fuzzy
number M.
262
P.(x)
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 -......:..,-'-~-'>------.. X
0 2 3 4 5 8 9 10
Fig. 4.79 Continuous fuzzy number M and its X(M a ).
Numerical Example
We shall rank the fuzzy numbers shown in Example 10 using Y"ager' s
F ranking index.
For U1 ' it is described by the membership function as:
x - .2 x
.2 .3,
{
~ ~
.1
IlU (x)
1
.5 - x
.3 < x ~ .5.
.2
Its a-cut and corresponding X(U 1a ) is calculated as:
U1a = [.la+.2, .5-.2a],
Hence,
1
J (.35 - .05a) da .325.
o
We can obtain F(U i ), i 2,3, in a similar manner. F(U i ), vi, can be
summarized as:
i 1 2 3
.325 .350 .440
263
since the higher F index value indicates a better fuzzy number, the
Note
1 Yager [Y18] points out that when M is a normal, trapezoidal
fuzzy number, F(M) can be easily computed by
F(M) (4.178)
where m1 is the mean of the support of M, and m2 is the mean of the
x values having a = 1.
2. The counter-intuitive example of the F(M) index is demonstrated
in Fig. 4.80. Given three symmetrical fuzzy numbers M1 , M2 , and M3
that share the same peak value (i.e., (5) = ~M (5)

~M = ~ (5) = 1),
12 M3
Yager's F index will result in M1 = M2 = M3 . That is because all
three fuzzy numbers have the same dashed line and consequently the
same F index value. Yager [Y18] interprets the three fuzzy numbers as
"approximately 5," "about 5," and "close to 5." Furthermore, he
suggests that since all three fuzzy numbers represent the same fuzzy
event, they should have the same F value.
This argument is not shared by Lee and Li [L3] who favored the
assumption that human intuition would choose a fuzzy number which has
a higher mean and a smaller spread. In this case, the ranking order
264
/L(x)
1. --- ----------- --------.-- ---
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 2 3 4 5 6 7 8 9 10
Fig. 4.80 Case of indiscrimination by Yager's F index.

265
4.11 Linguistic Ranking Methods

The most common criticism (Freeling [F1], Efstathiou and Tong
[ES], and Tong and Bonissone [T10]) on mathematical ranking procedures
is that they tend to defuzzify the problem by calculating a single
real number for each fuzzy set, thereby transforming the fuzzy sets
into a precise scale. But defuzzification may generate counter-
intuitive results because it ignores too much information contained in
the fuzzy set. Linguistic ranking methods which suffer less from the
stated difficulty are proposed. The idea is that since subjective
data are given linguistically, the outputs should also be in verbal
form so that we can make better use of the original information in
obtaining accurate results.
Zadeh [Z5] was the first to introduce such an idea. Freeling
[F1] praises the concept as ideal for fuzzy ranking. He then suggests
two ways in which linguistic ranking may be achieved. The first one
states that, for a comparison between two fuzzy sets M and N, a
preference relation such as "M is better than N" can be combined with
linguistic modifiers (Zadeh [ZS]) such as "very," "more or less," etc,
to form various linguistic expressions for comparison results. The
second approach is to use linguistic approximation to generate
decision statements of the format "(M is preferred to N) is "["" where "["
is a linguistic truth value (Zadeh [Z5]) such as "very true" or "more
or less true". However, in the early 70's, how to put the proposed
concepts into a workable algorithm was not clear.
Efstathiou and Tong [ES] demonstrated that mathematically
specified relations fail to capture the imprecision inherent in
preference ordering. They defined a linguistic preference relation
which preserves the properties of fuzzy preference relations (defined
by Zadeh [Z3]) to rank alternatives. The entries in the linguistic
preference relation are obtained through interactions with the OM
using an interactive algorithm.
200
Tong and Bonissone [T10] explore the second idea of Freeling, and
propose the use of context-free grammar and linguistic approximation
to generate decision statements such as "fuzzy number M is preferred

to others is more or less true." The grammar is used to generate
various decision statements that take the format of "M is P over all
other alternatives is ~." The purpose of linguistic approximation is
to match the fuzzy set with an appropriate decision statement which

has a predefined fuzzy set associated with it.
4.11.1 Efstathiou and Tong's Approach
Efstathiou and Tong [ES] propose a linguistic approach to rank
fuzzy final ratings of alternatives. The idea behind this method is
that since the decision maker's attitude towards risk, e.g.,

risk-averse, or risk-seeking, plays an important role in making
choices, the decision maker must participate in the rank ordering
process. This is done through an interactive decision aid which
follows a question/answer format. The decision maker is required to
reveal his/her preference between two fuzzy final ratings. The
interactive decision aid is used repetitively for every pair of
alternatives. Eventually, a fuzzy relation matrix is constructed.

The fuzzy relation is checked to see if it is a fuzzy preference
relation. If it is, a linear order among alternatives can be
determined; otherwise, the fuzzy relation must be reconstructed if a
linear order is desired.
Linguistic Preference Relation
A linguistic preference relation L on a set of objects X is a
fuzzy relation in which the preference of one element over another is
expressed linguistically rather than by a real number in [0,1]. For

267
example, given fuzzy numbers M and N, when the decision maker feels
that M is slightly better than N, a linguistic preference relation
L(M,N) = 'slightly' is obtained. The relation L preserves the
following properties (Zadeh [Z3]).
1. Antireflexive: A fuzzy relation on object x is antireflexive
if /lR(X,x) = 0, 'Vx. This means for a fuzzy number M, L(M,M) = "no
preference."
2. Antisymmetric: A fuzzy relation on objects x and y is
antisymmetric if /lR(x,y) > 0 and /lR(Y'X) > 0 then x = y. This

means for fuzzy numbers M and N, if L(M,N) * "no preference" then
L(N,M) equals "no preference."
3. Transitive: A fuzzy relation on a set of objects x, y, z is
transitive if
/lR(X,Z) ~ min max (/lR(X,y), /lR(Y'Z». (4.179)

y
This means that for fuzzy numbers K, M, and N, if M ~ Nand N ~ K then
the strength of preference for Mover K is at least as great as the
larger of Mover Nand N over K.
Given a set of fuzzy numbers K, M, and N, these properties imply
(a) no preference for Mover Mj (b) if M is preferred to N, then N
is not preferred to Mj and (c) if M is preferred to Nand N is
preferred to K, then M is preferred to K.
Obtainment of Linguistic Preference Relation, L
To construct a linguistic preference relation is equivalent to
determining the decision maker's preferences. A set of sequential
questions is used to extract the linguistic preference relation
between a pair of fuzzy final ratings. The complete relation L is
obtained by repeated pairwise comparisons within the set of fuzzy
final ratings.
The questions are asked surrounding the following five factors:
(1) overlap of support sets;

(2) separation of peak values;
(3) tradeoff between better and worse outcomes;
(4) proximity to threshold values;
(5) relative height of fuzzy sets.
The discriminating ability of these factors decreases from the top

down. The questions are designed so that the first questions are the
most discriminatory and lead to firm statements of preference. Thus
nonoverlapping support sets indicate a definite rank order and no more
questions need be asked. As the analysis proceeds in less clear cut
cases, the resulting preference is less strong. For fuzzy numbers
with overlapping support sets and/or fuzzy numbers with peak values
which are not well separated, one can only access marginal preference
among fuzzy numbers.
When the top three factors fail to discriminate among fuzzy final
ratings, the decision maker's attitude toward risk comes into play.
If the decision maker is risk-seeking, s/he will gamble for the
possibility of a good or a very good outcome despite the great chance
of a very poor outcome. On the other hand, a risk-averse decision
maker will favor those alternatives which offer only the moderate
outcomes with a greater degree of certainty.
The decision maker's attitude may change considerably on either
side of a threshold level which may be treated as a break-even point.
The decision maker is strongly averse to outcomes falling on the lower
side of the threshold, and less discriminating towards possibilities
falling on the upper side. This suggests that the preference ordering
of a pair of fuzzy final ratings may depend on their closeness to the
269
threshhold. In short, a linguistic preference relation is constructed

through careful evaluation against five factors. The factors are used
one by one. The top ones have better discriminating ability while the
bottom ones have poor discriminating ability.
Algorithm
Given fuzzy numbers M1 , ... ,Mm, the major steps of this algorithm
are:
step 1. Define a set of terms that contains the various degrees
of preference the OM requires, such as {strongly, slightly, very much,
no preference, •.• }. This term set is problem- as well as
OM-dependent.
step 2. Ask the decision maker to express his preferences,

linguistically, for each pair of fuzzy numbers. This results in a
pairwise preference matrix, L, for m fuzzy numbers:
M1 Mi Mm
M1 r 11 r 1i r 1m
L Mi r i1 r ii rim
Mm r m1 rmi r mm
where r ij , V i,j, are linguistic preference relations; r ~~

.. = 'no
preference,' vi. If r ij is a preference term other than 'no
difference,' then r ji = 'no difference.' Note that r ij is obtained
through a screening process in which all linguistic terms can be
expressed by their corresponding fuzzy sets. Fig. 4.81 presents the
fuzzy sets of linguistic terms such as 'no difference,' 'slightly,'
'somewhat,' and 'very much.' These fuzzy sets may be treated as the
meaning of the linguistic terms.
270
~(xl
1.0 ,...,
NO SL SO VE
-4 -2 0 2 4 6 8 10 12 14
NO: No difference SL: Slightly SO: Somewhat VE: Very much
Fig. 4.81 Meaning of linguistic terms.
Step 3. To have a linear order among Mi , vi, the L matrix must
be:
(i) Antireflexive. This implies L(Mi,M i ) = 'no difference.'

Observe that in Step 2, we specifically set r ii to "no difference" to
preserve this property.
(ii) Antisymmetric. This implies if L(Mi,M i ) is any preference,
then L(Mj,M i ) = 'no difference.' This property is also preserved
when constructing the L matrix.
(iii) Transitive. A linguistic preference relation L is
transitive if
r ij l!: (ror) ij
where
(ror) .. min (4.180)

1)
k
Note that r ik and r kj are linguistic terms whose meaning can be
represented by fuzzy sets. The summation of r ik and r kj is performed
using the fuzzy arithmetics presented in section 3.4.2.

271
If L satisfies all three requirements, we have linear order amonq

the fuzzy numbers, Mi' Vi. We can only obtain partial order where two
or more alternatives cannot be ranked linearly. The decision maker
may be asked to repeat the process aqain if a linear order is desired.
Numerical Example
An example from Efstathiou and Tong [ES] is used to illustrate
some of the features of this approach. Three fuzzy final ratings for
alternative A1 , A2 , and A3 are given linguistically U1 'Slightly
below medium,' U2 = 'Poor,' and U3 = 'Good but not very good' (see
Fig. 4.82). Assume the decision maker's preference is obtained as:
U1 U2 U3
L NO SL NO
NO NO NO
SO VE NO
1
where NO, SL, SO, and VE represent 'no preference,' 'slightly,'
'somewhat,' and 'very much,' respectively. Fig. 4.81 shows a possible
way to interpret the above linguistic terms.
/L(X} ·slightly below medium·

1. U1
.9
.8 •goO$! but not very good·
.7 ~",
l/~U3
~ ~
.6 ~~ "
.5 ~
"
I
I
I
"
.4 I
""I
.3 \
I
.2 \
I
I
.1 I
\
\
0 ~~--~--~~--~--~~--~\--~~x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.

272
We shall briefly outline the obtainment of L(U 3 ,U l ) to
demonstrate the process. Referring to Fig. 4.82, we see that U3 is
the rightmost fuzzy number, i.e., U3 has a more favorable support.
similarly, the peaks are well separated between Ul and U3 . U3 has the
best peak position. When we analyze the trade-off between the good
outcome vs. the bad outcome, we see that Ul has a low possibility of
good outcome and a high possibility of medium outcome; U3 has a high
possibility of good outcome and a low possibility of bad outcome.
Assume the decision maker has decided that the alternative must at
least be a medium outcome. All phenomenon suggest that alternative A3
is a favorable choice. The relative height of U3 is less important
here since the peaks are well separated. However, the lower peak
value of U3 makes its preference over the others less strong. Thus,
we conclude that alternative A3 is somewhat preferred to Ai' i.e.,
L(U 3 ,U l ) = 'somewhat.'
The obtained relation L must be checked if it preserves the

properties needed to construct the linear order of alternatives.
We know that L is antireflexive since all L(Ui,U i ) = 'no preference.'
L is antisyrnmetrical because when L(Ui,U j ) * 'no difference,' L(Uj,U i )

'no preference.' To check if L is transitive, the relation L is
composed with itself using Eq.(4.180) as:
1
[
NO+NO SL+NO
NO+NO
LoL NO+NO NO+NO NO+NO
SO+NO SO+SL NO+NO
The test for transitivity ought to be that r ij ~ (ror)ij.
Clearly this notion of inclusion is not well defined when
combining linguistic terms. Efstathiou and Tong [ES] assume that
adding "NO" to any term will not alter the meaning of that term.
273
LoL then becomes
[
NO SL NO
NO NO NO
SO SL+SO NO 1
All the elements in LoL equal their corresponding elements in L except
(r or)32 = 'SL+SO.' In Fig. 4.83, it is easy to see that
r 32 = 'VE' > (r or)32 = 'SL+SO.'
Since the condition r ij ~ (ror)ij' vi,j is satisfied, we can conclude
that the linguistic preference relation L obtained from the OM
preserves all properties needed to form a linear ranking order. The
ranking order is
Somewhat> Slightly>
~(x) SO+SL VE
1.0
-4 -2 0 2 4 6 8 10 12 14
Fig. 4.83 The fuzzy sets "VE" and"SO+SL".
Note
1. We feel that this approach is still in its infancy and should
not be taken seriously as a ranking procedure. For example, if the

274
OM's preference can be assessed through a sequence of questioning, we
really don't have to go through any ranking procedure at all. The
only merit of this algorithm is its use of the antireflexive, the
antisymmetrical,.and the transitivity properties of fuzzy relations to
guarantee the consistency of the OM's assignment of preference.
4.11.2 Tong and Bonissone's Approach
Tong and Bonissone [T10] assume that the final ratings for m
alternatives are obtained in the form of fuzzy numbers, Ui ' i
1, ... ,m. To rank U i ' Vi, a dominance set 0 is first derived. A
ranking order can be derived from the dominance set. This dominance
set alone may result in several fuzzy numbers being ranked the same.
That is, we may see some indiscrimination case when using the
dominance set for ranking. To resolve this ranking difficulty, a
preference set is constructed for each of the equally ranked fuzzy
numbers. The preference sets are used to further distinguish the
equally ranked fuzzy numbers. The ranking result is given in natural
language.
Dominance Set
A dominance set, 0, is defined as:
max [min {J..L su . (x), J..L U . (x) }] (4.181)

j=l, ... ,n 1 J
j"i
where
{
1, if x s xo,
J..L s U. (x) o (4.182)
1
J..L U . (x), if x > x ,
1
with Xo being the lowest value of x for which J..L u . (x) 1; and liS U."
1
1
275
is the notion for "less than or equal to Ui " formed from Ui • In Fig.
4.84a where the peak of Ui is to the left of Uj' Eq.(4.181) will
result in the intersection of ":s U." and U.; while in Fig. 4. 84b,
~ J
Eq.(4.181) will always give a value of 1.
Fig. 4.84a The fuzzy set ":s U."

~
- case 1.
The fuzzy set "c U'-

,.,,(x) \ - I
1. '!'!'!'!'!'!'!'!'!'!'!'!'!'!'!'!'!'!'!'!"'!'!'!'!'!'!'!"""""""""""""""
Fig. 4.84b The fuzzy number ":s Ui " - case 2.
The definition of a dominance set indicates the dominance of Ui

over Uj' vi, j. However, the discrimination ability of Eq.(4.181) is
not very high. For example, referring to Fig. 4.85 and Table 4.6,
276
~
3
~
.5
A
~
~
.1
0
0 2 3 4 5 6 7 8 9 10
Fig. 4.85 Four fuzzy numbers.
Table 4.6 Dominance Matrix for Fig. 4.84 Data
U1 U2 U3 U4
~D U1 1 .50 0 0
[
U2 1 1 .50 .80
U3 1 1 1 1
U4 1 1 1 1
1
Eq.(4.181) cannot tell if U3 is better than U4 or the other way
around, because ~D(U3,U4) ~D(U4,U3) = 1. This ranking difficulty is

attributed to the fact that the dominance set does not consider the
shapes of fuzzy numbers.
Obtainment of preference set Z
The ranking difficulty can be resolved by introducing the

preference set z. To define the preference set, a vector V must
first be defined as:
(4.183)
277
which indicates the overall degree to which alternative Ai dominates

the others and may be thought of as defining a fuzzy set of
nondominated alternatives. For example, referring to Table 4.6, V(U 2 )
is calculated as:
This indicates that the overall degree to which U2 dominates all other
alternatives is 0.5. We can obtain V(U 1 ), V(U 2 ), and V(U 3 ) in a
similar manner. In this case, since V(U 3 ) = V(U 4 ) = 1, A3 and A4 are

nondominated alternatives.
Given V(U i ) vector, an n-dimensional function is defined as:
Xk - [ (4.184)
where the index k corresponds to a position in V(U i ) where V(Uk ) 1.
The fuzzy preference set, Zk' is then induced by gk when xi in

Eq. (4.184) is replaced by Ui. The Extension Principle (see Section
3.4.1) defines the fuzzy preference set as:
n
~Z (z) max [ min ~U. (xi)]' (4.185)
k (x 1 , . . . ,x n ) i=l 1
where z = gk(X 1 , ... , x n ). These fuzzy preference sets Zk' k = 3,4,

are now used as an aid to further distinguish among nondominated
alternatives.
One can easily obtain Z3 and Z4 using Eq. (4.185) if ~U (x.), ~i,
i 1
278
are discrete membership functions. When "u. (xi) are piecewise linear
1.
and continuous, the computation of /J.z (z) is extremely difficult, i f

k
not impossible. Tong and Bonissone [T10] assumed that any piecewise
linear and continuous membership function M can be represented by
(a,b,a,~) (see Fig. 4.86). It follows that the arithmetic operations
of this special type of fuzzy number can be conducted using
Bonissone's [B25] formulas.
Bonissone's arithmetic operations (Bonissone [B25])

Given M1 = (a1,b1,a1'~1) and M2
arithmetic operations are given as:
(4.186)
(4.187)
a1~2+b2a1 b1a2+a2~1 (4.189)

b 2 (b2+~2)' a 2 (a 2 -a 2 ) ).
P.(x)
1.' ........ . M
Fig. 4.86 Special fuzzy number M = (a,b,a,~).

279
The fuzzy numbers U3 and U4 in Fig. 4.85 cannot be distinguished

by using Eq.(4.181). However, with the help of Bonissone's [B25]
arithmetic formulas, we can obtain preference sets Z3 and Z4 using
n
L V(U i ) Ui
i=1
iook , k 3,4., i 1,2,3,4, (4.190)
Zk Uk - n
L V(U i )
i=1
iook
where Uk = (a k , ~, ak , ~k)' Ui = (ai' b i , ai'

and Zk = (a z ' b z '
~i)
k k
a zk ' ~~). For example, the preference set Z3 is calculated as:
(9,9,2,2) - 0*(1,1,.5,2)+.5*(3,6,2,4)+1*(7,10,1,.5)
o + .5 + 1
(9,9,2,2) - (8.5,13,2,2.5)/1.5
(.33,3.33,3.67,3.33),
where the summation, subtraction, multiplication, and division between

two fuzzy sets is conducted according to Bonissone's arithmetic
formulas. The resulting Z3 and z4 is displayed in Fig. 4.87. Since
Z3 is to the right of Z4' we can conclude that final rating U3 is
better than U4 .
It is believed that the OM can better understand a linguistic
statement characterizing the decision set than a numerical membership
function. The linguistic structure suggested is given as:
"~ is P over all other alternatives is II."
For instance, we may assert that "alternative ~ is marginally

preferred to other alternatives is very true". To translate Zk
280
P.(z)
1
-6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6
Fig. 4.87 The fuzzy preference sets Z3 and Z4.
meaningfully in the proposed linguistic structure, P is a term of the
linguistic variable "preference" which takes its value in the term

set: {marginally better, absolutely better, indifferent ... }, and IT is
a term of the linguistic variable "true" which takes its value in the
term set = {true, not true, very true, ... false, not false, very
false, ... , not very true and not very false, ... } (Zadeh [Z5]).
The next step is to find terms in the term sets that approximate
the unlabelled Zk as closely as possible and then use the selected
linguistic terms to express Zk in natural language. The tools for
linguistic approximation one may use are pattern recognition by
Bonissone [B22,B23] and a context-free grammar by Zadeh [Z5]. This
step will be elaborated in the Note at the end of the section.
Algorithm
The algorithm can now be summarized by the following steps.
step 1. Identify the dominance matrix, D.
step 2. Find the dominance vector, v, using
min
j
281
step 3. For each V(Uk ) 1, Vk, we construct a fuzzy decision

set, Zk.
step 4. Express Zk in linguistic structure as "~ is P over all
other alternatives is IT" where P and IT are linguistic variables whose
values are linguistic terms predefined in term sets. The matching is
done using linguistic approximation and context-free grammar.
Numerical Example
Three investment tools, commodity (A 1 ), stock (A 2 ), and real
estate (A 3 ) are to be evaluated according to four attributes: risk of
losing capital (Xl)' inflation impact (X 2 ), interest received (X 3 ),
and cash availability (X 4 ). The fuzzy final ratings are given as:
U1 = (1.26, 1.34, .62, .40), U2 = (1.46, 1.46, .86, .80), and U3 =
(2.32, 2.42, .94, .58) (see Fig. 4.88).
Given the fuzzy final ratings, we shall use the stated algorithm
to find the best solution and corresponding linguistic expression.
step 1. Obtain the dominance matrix using
max [~< U. (x) A ~u.(x)]

j=l, ... ,n - 1 J
j~i
as
U2 U3
[
o .90 .27
1 .51
1 1 1
The element ~D(U1,U2) is obtained as follows. The fuzzy set ~<U (x)
- 1
is obtained using Eq. (4.182) as (see Fig. 4.89):
~<U
- 1
(x)
{ 1, if
~U
1
x s 1.26
(x), if x > 1.26
282
p.(x) U1
1.
LLa
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 2 3
P.(x)
1.0000000000
.9
.8
.7
.6
.5
.4
.3
.2
.1
o ~~~~~--~--~~~~~--~~x
o 2 3
Fig. 4.89 Illustration of the fuzzy set ~~ U (x).

1
Consequently, ~D(U1,U2) is calculated as:
The dashed line in Fig. 4.89 represents the result of [~<u (x) A
- 1
~U (x)]. The peak point of the dashed line is ~D(U1,U2).

2
step 2. Find the dominance vector using Eq. (4.183):
v = {.27, .51, 1}.

283
For example, V(U 2 ) = .51 is obtained as:
min {~D(U2,U1)'~D(U2,U2)'~D(U2,U3) } min {l,1, .51}

.51
step 3. Observing that there is only V(U3 ) =1 in the dominance

vector, we construct the Z3 decision set using Eq.(4.190) (see
Fig. 4.90):
]
V(U 1 ) U1 + V(U 2 ) U2
V(U 1 ) + V(U 2 ) (0.90,1.03,1.60,1.36)
where fuzzy arithmetic operations follow Bonissone's [B25] formulas.
-1 o 2 2.5
Fig. 4.90 The preference set Z3 for the numerical example.
step 4. Based on Z3' we can approximate the result by linguistic

terms such as,
A3 = real estate;
P highly preferable;
II = true.
The conclusion is then given as: "real estate is highly preferable to

284
all other alternatives is true." See the Note for the complete
algorithm in deriving the linguistic conclusion.
Note
1. Step 4 of this method is actually an important research topic
called linguistic approximation, which intends to attach a linguistic
term to a given unlabeled membership function. The specific technique
used in this step is called pattern recognition which is proposed by

Bonissone [B22,B23]. We shall treat the procedure of pattern
recognition technique as substeps of 5tep 4. A brief discussion of
the substeps follows.
Step 4.1. Generate two sets of possible terms which shall be
referred to as term sets. An efficient way to generate the term
set is to define some primary terms for P and IT, respectively, and
then use the context-free grammar of Zadeh [Z5] to obtain the term
sets, one by one.

The grammar, G, which is a 4-tuple (VN, VT, 5, R) provides a
systematic way to generate a term set.
(1) VN = {A,B,C,V,H,T} is the set of linguistic variables which
take their values from VTj
(2) VT = {Indifferent, Better, Worse, Not, And, Or, Marginal,

More or less, Sort of, Very, Highly}. This set of terminals
(vocabulary) includes the primary terms, linguistic hedges, and
relations (see Table 4.7a)j
(3) 5 is the starting symbolj and
(4) R is the production system (see Table 4.7b).
With the primary terms and context-free grammar, we can get a term set
= {Indifferent, Better, ... , Marginally Better, ... , Not Better, Not

Indifferent And Not Better, ... }. Another term set = { more or less
285
Table 4.7a Overview of VT
Primary term Hedge Relation
Indifferent More or less Not

Better Sort of And
Worse Marginal Or
Very
Highly
Table 4.7b Overview of the production system R
S A
S A C A
S U
A B
A NB
B T
B HT
B VVT
N Not
H Marginally
H More Or Less
H Sort of
V Very
V Highly
T Indifferent
T Better
T Worse
C Or
C And
286
true, true, very true, ... , not true, not very true, ... } can be
obtained in a similar manner. Keep in mind that the exact form of
the term set is problem-dependent, but it should be large enough to
facilitate a wide range of decision situations.
step 4.2. with two term sets, one for P and one for IT, we can
combine them and obtain a composite set, L, which contains decision
statements, such as "Marginally better (than other alternatives) is
very true."
step 4.3. For each decision statement in the composite set L,
there is a membership function. We need to find the function that is
closest to our ~z(z) function and attach that decision statement to
the unlabelled Z. However, it is extremely inefficient to make
comparisons for each statement in L with ~z(z). We shall conduct a
preliminary screening so as to reduce the number of comparisons.
step 4.3.1. We shall represent each decision statement, L i , by a
pattern, LP i , of characteristic features. The pattern LP i is
represented by 4-tuple (LP i1 , LP i2 , LP i3 , LP i4 ). The four elements
are defined as
LP i l f ~L. (x) , (power of fuzzy set L i ), (4.191)

x ~
LP i2 f S (~L. (x) ) , (entropy of fuzzy set L i ), (4.192)

x ~
where S(y) -y In(y) - (l-y) In[ (y) (l-y) ],
~A(x)
LPiJ f [f(x) ], (centroid point of fuzzy set L i ) ,
LP i l
x
(4.193)
where f(x) is the probability distribution function of x,

287
J [f(x) - LP i3 ] 3 , (skewness of fuzzy set Li ), (4.194)

x
which measures the asymmetry of function ~L.(x). For simplicity of

~
computation, Tong and Bonissone [T10] assume the universe of discourse

U is partitioned into K, a finite and discrete number of points.
Hence, Eqs.(4.191) through (4.194) may be rewritten as:
K
LP i l E ~L. (X j ), (4.195)
j=1 ~
K
LP i2 E S (ILL (x.», (4.196)
j=1 i J
K ILL. (X j )
~
LPiJ E (j-1) LP i l (4.197)
j=1
K ILL. (X j )
3
LP i4 E {[ (j-1) - (LPiJ)] [
~
LP i l ]}. (4.198)
j=1
since several decision statements may share the same pattern,

LP i , the number of patterns is much smaller than the number of actual
decision statements, Li
step 4.3.2. The obtained pattern is compared with the pattern of

Il Z ' i.e., unlabelled decision set Z, using
(4.199)
where A is the relative importance of a characteristic feature and is

288
defined by:
(4.200)
where Wj may be obtained using Saaty's [S1] method; and
Dj = max [LP ij ] - min [LP ij ].
The computed d 1 (LP i ,LP Z ) is compared with c which is the tolerance in
judging the similarity between patterns. Once d 1 < c is found, a
small set of decision statements, L:, which are semantically similar

to z and denoted by LA[Z], are identified.
step 4.3.3. To determine the unique decision statement for our
preference set, Z, we need to compare each decision statement in LA[Z]
with Z using
K
L
j=1
* (4.201)
where LP i1 is the power of Li and LPZ1is the power of fuzzy set Z.

However, since d 2 (·,·) does not satisfy triangular inequality, a
function is defined as
d * (L *
i , Z) (4.202)
such that triangular inequality is preserved. We then identify the Li*
which results in minimum d* value, and we shall use that decision
statement as the label of Z. The algorithm stops here and the ranking
result is given in natural language form, "(Ak is P) is II," rather
than in numerical numbers.

v. FUZZY MULTIPLE ATTRIBUTE DECISION MAKING METHODS
5.1 Introduction
A MADM problem is given as:
Xl X2 X
n
Al xII x 12 X ln
[
D A2 X 2l X 22 x 2n
A
m
x ml x m2 xmn 1
~ (WI' w2 '···, wn )
where Ai' i = 1, ... , m, are possible courses of action (candidates,

alternatives); Xj , j = l, ... ,n, are attributes with which alternative
performances are measured; x ij is the performance score (or rating) of
alternative Ai with respect to attribute Xj ; wj , j = l, ... ,n are the
relative importance of attributes.
The classical MADM solution methods assume all x .. , w. values are

1J J
crisp numbers. A utility function U(x I ,x 2 , ... ,xm) is implicitly or
explicitly defined by the decision maker. For Ai' the utility
function aggregates its performance ratings x ij ' vj, into a final
utility, Ui . such a final utility represents how well one alternative
satisfies the decision maker's utility. The alternatives with higher
final utilities are said to be preferred by the decision maker. Since
the final utilities are real numbers, the preferred alternatives are
those with higher final utilities (ratings).
In reality, alternative performance rating x ij can be crisp,
fuzzy, and/or linguistic. For example, three candidates are being
considered for a professor position. The attributes used are
creativity (Xl)' maturity (X 2 ), communication skill (X 3 ), and number
of publications (X 4 ). The performance scores for the first three

290
attributes are not quantifiable, rather they are represented by
linguistic terms such as "good," "average," "poor," etc. The fourth
attribute can be some integer numbers. This MADM problem contains a
mixture of fuzzy and crisp data. Most of the real world MADM problems
are of this type.
Fuzzy MADM methods are proposed to solve problems which involve
fuzzy data. Bellman and Zadeh [BS] were the first to relate fuzzy set
theory to decision making problems. In 1977, Baas and Kwakernaak [B1]
proposed a fuzzy MADM method that is widely regarded as the classic
work of fuzzy MADM method. During the past 10 years, several fuzzy
MADM methods have been proposed. The only systematic reviews of fuzzy
MADM methods have been conducted by Kickert [K1S] and Zimmermann [Z30,
Z31]. Zimmermann [Z31] among others treated the fuzzy MADM method as
a two-phase process. The first phase requires finding the fuzzy
utilities (fuzzy final ratings) of alternatives. The second phase
requires applying fuzzy ranking methodes) to determine the ranking

order of alternatives.
In this chapter, we will provide a thorough, systematic review of
the existing fuzzy MADM methods. There are a total of 1S fuzzy MADM
methods. They are systematically classified into eight categories
(see Fig. 5.1). The classification is based on four factors: (1)
their capability of solving large-size problems, (2) the type of data
allowed, (3) the classical MADM method each fuzzy MADM method relates
to, and (4) the technique each method uses. Theoretical backgound as
well as the algorithm are presented for each method. Numerical
examples are also given so that each method is easily understood. The
advantages and disadvantages of each method will also be brought out
when possible. Finally, a new approach to the fuzzy MADM problems is
proposed.
Problem Correspond ing Technique
Data Type Approaches
Size MADM Methods Involved
~~imple Additive Baas and Kwakernaak (B1)

I All fuzzy H a-cut i Kwakernaak (K32)
eighting method
Dubois and Prade (028)
Cheng and Mcinnis (C171
I Fuzzy Arithmetics 1
AHP Bonissone (B28,B27)
n c 10 J-- -- l Eigenvector method I

I Saaty (S1)
I m c 10
~ Weight assessing +-
arithmetic operatio Laarhoven and
Pedrycz (L 1)
ICrisplfuzzy - Conjunctionl
Disjunction Buckley (B36,B39)
Possibility and I
method
Fuzzy
Hnecessity measures r
Multiple Dubois et al. (037)
Attribute J--
Decision Human
in tu ition Efstathiou and
Making Rajkovic (E3)
H MAUF J ~
All crisp I
I Ranking methods
General MADM Fuzzy arithmetic JI Negi (N7)
method
I n c 350 Outranking Fuzzy outranking Siskos et aJ. (S20)

m • any number method
I I relation
I Brans et al. (B35)
I All fuzzy
(singleton)
Maximin Max and min Bellman and Zadeh (B8)
r 1 operators Yager (Y4)
~ Crisp/fuzzy General MADnLingUistiC-.fUZZY setl I

methods
Fig. 5.1 A taxonomy of fuzzy MADM methods.

292
5.2 Fuzzy Simple Additive weighting Methods
The classical Simple Additive weighting method (Churchman and
Ackoff [C22], MacCrimmon [M2], Hwang and Yoon [H13]) is mathematically
defined as follows. Suppose the OM assigns a set of weights, ~ =

(w 1 , ... ,Wn ), to the attributes, Xj , j = 1, ... ,n. The performance of
alternative, Ai' is calculated as:
n n
U.=
1 r w). r i )· / r w). (5.1)
j=l j=l
where r .. is the rating of the ith alternative under the jth attribute
1)
with a numerically comparable scale. This is the simplest form in
Multiple Attribute utility Theory (MAUT). The most preferred

.
alternat1ve, A*, .1S then selected such that
A * (5.2)
When both Wj and r ij are fuzzy sets defined as:
and
r i ). {(x.1)·,11 r .. »}, lJi,j,
(x 1)
ij
where y). and xi)' take their numbers on the real line ~ and 11 (y.)
Wj )
and 11 (x .. ) take values in [0,1], the utility of alternative Ai'
r ij 1)
U. = {(u',llu (u.»}, can be calculated as follows. The variable u 1'

1 1 i 1
takes its value on the real line ~ and can be obtained using
n n
r y). xi)' / r y).. (5.3)
j=l j=l
293
The membership function ~U. (u i ) can be calculated using

~
n n
sup {[ 1\ ~w. (YJ')] 1\ [ 1\ ~ (x .. )]} (5.4)
v j=l J j=l
r ij ~J
where v = (Y1'···'Yn' x i1 ' · · · ' x in )·

The membership function ~U (u.) is not directly obtainable when
i ~
~ (y.) and ~ (x .. ) are piecewise continuously differentiable

Wj J r ij ~J
functions. To resolve this difficulty and preserve the simplicity of
the SAW method, several approaches have been proposed by Baas and
Kwakernaak [B1], Kwakernaak [K32], Dubois and Prade [028], Cheng and
McInnis [C17], and Bonissone [B27]. The first four approaches utilize
the a-cut to approximate the ~U (u.). On the other hand, Bonissone

i ~
assumes that all piecewise continuously differentiable fuzzy numbers
can be approximated by L-R type trapezoidal numbers. Special fuzzy
arithmetic functions are then applied to calculate the fuzzy utility
Ui for alternative Ai' Bonissone's approach is much easier to use

than the other approaches in this category.
5.2.1 Baas and Kwakernaak's Approach
Baas and Kwakernaak [B1] were the first to identify the computa-
tional problem in calculating ~u (u.). To resolve this difficulty,

i ~
they proposed the use of the a-cut to obtain the fuzzy utility Ui . An
a value for ~U (u.) is assumed first, i.e., ~U.(u~) = ao ' The

o i ~ ~ ...
corresponding u i value(s) are then calculated using Eq.(5.3). By
setting different a o values and repeating Baas and Kwakernaak's
algorithm, an approximated fuzzy utility Ui can be obtained.

294
Algorithm
It is assumed that ~ (y.) and ~ (x .. ) are normalized
Wj ) r ij 1)
membership functions. We are to determine the approximated fuzzy
utility Ui for alternative Ai' vi, using the following steps.
step 1. set an a o level for ~U. (u i ).

1
step 2. After setting ~U. (u i ) a o ' identify the Yj and x ij

1
values that satisfy
~ (x .. ) a o ' Vi,j. (5.5)

r ij 1)
There may be more than one Yj value for ~w. (Yj) a o and more than
)
onex .. valuefor~ (x .. ) = a o '

1) r ij 1)
For example, we have two fuzzy sets, Wj and r ij , as shown in Fig.

5.2. Assume that a o = 0.7, (y1,y2) (3,5) are values which make
and (x 1 ,x 2 ) = (7,9) are values which make ~ (x .. )

r ij 1)
JL(x)
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
O~I-",,--T-.....--+-....--i-l.-.---i-..,..........,.....l......
o 6 7 8 9 10 11 1l
Fig. 5.2 a-level set for Wj and r ij at a o 0.7.
295
J.l.(x)
1.
Fig. 5.3 Case of two fuzzy attributes and two fuzzy weights.
step 3. There are many u i values such that gu i (u.) = a. We

1 0
want to know the extreme ones, u. and u For example, given

1min i max
two fuzzy attributes and two fuzzy weights as shown in Fig. 5.3, there
will be a total of 24 = 16 u i values, i.e., 16 possible combinations
of Yj and x ij ' i,j 1,2. We simply pick the highest u 1" u and
i max '
the lowest u i ' u i , ' and drop all other ui's.
m1n
If the size of the problem increases, such as five attributes and
five weights, there will be 210 = 1024 u 1' values. To identify u
i max
and u, is a bit tedious without the help of a computer. When
1min
problem size increases to 10 attributes and 10 weights, there will be
over a million u, values. Again, to identify u and u. is

1 i max 1min
impossible without the help of a computer.
To avoid calculating all the u i values such that gu i (u.) = a ,

1 0
Baas and Kwakernaak suggest the following. Given a set of real

A A A A A
numbers (Yl' ... , Yn' x 1'I'···' X in ) such that g r '(x 1J
.. ) and
i
[g' (~.)/(~.,-U,)], l;Ii,j, where
Wj J 1J 1
296
Ji' (x .. ) dJi r (x 1)
.. ) / dX 1.). (5.6)
r i 1) i
and
(5.7)
have the same sign, the resulting u i will either be u imax or u.

1m in
step 3 is complete when both u. and u i have been found.
1max min
To check if Jir:(~ij) and [Ji~j(~j)/(~ij-Ui)]' Vi,j, have the same
A A A
sign, one would have to use the set of numbers (x il ' ... , x in ' Yl ,
... , ~n) to compute its corresponding u i value using Eq. (5.3).

A A A A
Consequently, the values (x il , ... , x in ' Yl , ... , Yn ) are used in
Eqs. (5.6) and (5.7). If Jir:(~ij) and [Ji~j(~j)/(~ij-Ui)] have the same
sign, i.e., both are either positive or negative, the u i value

A A A A.
determined by (X il , ... , x in ' Yl , ... , Yn ) w1ll be u i or u. If
max 1min
the signs are not the same, the corresponding u i would be dropped.
The algorithm loops back to step 1 for another a value. One must
give several a values in order to get an approximated Ji U (u.)
i 1
function. The number of a values needed to construct Ji u . (u i ) is a

1
subjective matter. If more a values are given, the approximated
function will be closer to the real one but will require much more
computational effort.
Numerical Example
To illustrate the computational procedure, the 2 x 2 decision
matrix from Baas and Kwakernaak [Bl] will be used.
Xl X2
o good
fair
fair
good ]
297
P.(x)
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 ~~~~~~-+--~~~~~4-~~-+ X
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 5.4 Fuzzy representation of linguistic terms.
where [r 11 ,r 12 ] = [good, fair] and [r 21 ,r 22 ] = [fair,90od]. The weight
set is [W 1 'W 2 ] = [very important, rather unimportant]. Fig. 5.4

presents the meaning of these linguistic terms.
The fuzzy utility of A1 is calculated using the following steps.
step 1. set (xo 0.75.
step 2. " "

Identify x 11 , x 12 , Y1' Y2 values such that " " ~
r 11 "
(x 11 ) =
~r
12
"
(x 12 ) ~w (Y1)
1
" ~w (Y2)
2
" = 0.75. The values are summarized as:
.85 .65 1.0 .25

.75 .55 .95 .15
step 3. There are a total of 24 = 16 possible combinations of
" " "

(x 11 ,x 12 , Y1' Y2)· " By applying Eq. (5.3) on all x 1j and
Yj combinations, we obtain 16 u 1 values. They are summarized in
Fig. 5.5 where u 1 and u 1 . are easily identified. It follows that

max mln
.75 and .75

298
II II II
XII X12 Y1 ~2 U1
0.25 .8100
1 . 0 - L 0.15 .8239 (= u 1
0.25 .8083 max
. 95-L 0.15 .8227
'85_["65_[
0.25 .7900
1 . 0 - L 0.15 .8109
.55-[ 0.25 .7875
. 95-L 0.15 .8091
0.25 .7300
1 . 0 - L 0.15 .7370
'75_["6~[ 0.25
. 95-L 0.15
0.25
.7292
.7364
.7100
1 . 0 - L 0.15 .7239
.55-[ 0.25 .7083 (= u 1 . )
. 95-L 0.15 .7227 m1n
Fig. 5.5 The possible combinations of YJ. and x .. and

1J
their corresponding u i values.
We shall detail the derivation of u 1 = .8100 to demonstrate the
calculation procedure. Given the combination (~11'~12'~1'~2)

(.85,.65,1.0,.25), u 1 can be obtained using Eq.(5.3) as:
u
1
= .8100
To verify if ~r:(~ij) and [~~j(~j)/(~ij - u i )] have the same sign

II II II II
for (x11,x12'Y1'Y2) = (.85,.65,1.0,.25), we compute
r 11 (.85)
-5, -5,
~' ~'
r 12(.65)
±oo/ • 04 ±oo,
-5/[-.16] 31.25.
299
In this case ~ I(~ .. ) and [~I (~.)/(~ .. - ui)],Vi,j, do not h~ve the
r i 1) Wj ) 1)
same sign.
either u 1 nor u 1 . .
max m1n
On the other hand,
a u i = .8239 which is the maximum among all possible u i values, because
~' (.85) -5, ~' (.65) -5,

r 11 r 12
±00/.0261 too,
and
5/[-.1731] -28.88,
all take the negative sign. (Note that ~' (1) ±oo can be considered
wI
as either +00 or -00. Here we take -oo.) We can identify u = .7083

Imin
in a similar manner.
The first iteration (where a o = 0.75) has been completed. We can
set a o at 0, 0.5, and 1, respectively, and carry out the algorithm
three more times. The results can be summarized as:
~u (u 1 ) = ao 0 .50 .75 1.0

1
u1 1.0 .8820 .8239 .7667
max
u1 .5333 .7083
.6500 .7667
min
Given various u 1 values at different a levels, the approximated
fuzzy utility U1 can be graphically represented in Fig. 5.6.
Similarly, we can get the fuzzy utility U2 using the same algorithm.
The result is also seen in Fig. 5.6.

300
p.(x) U2 U1
1.......................................... .
.9
.8
.7
.6
.5
.4
.3
.2
.1
0
0 .1 .2 .3 .4 .5 .6 .7 .8 .9
Fig. 5.6 The fuzzy utilities.
Ranking of Alternatives
The ranking of U1 and U2 can be carried out by using any proper
ranking methodes) presented in Chapter 4. Here by simple observation,
Note
1. According to Baas and Kwakernaak, to identify u and u.

i max 1min
is a trial-and-error process. One way to identify u and u. is
i max 1min
A
to select a set of Yj and x ij combinations (x i1 , ... , x in ' Y1' ... ,
Yn ) and check if ~ r '(~ 1)

i
.. ) and [~~).(~)·)/(~i).-Ui))' ~i,j, have the same
A
sign. If they do, we can then compute u i using Eq. (5.3) with (X i1 ,
A
.•• , X in ' Y1' ... , y n )· Otherwise, we can proceed with another set of
combinations.
We feel that with modern computers the trial-and-error process
can be replaced by systematically computing all the possible u i values
and then picking the maximum and minimum from among these values.
The example presented earlier has two attributes. Together with the
two weights, they make a total of 16 combinations. We can compute all

301
16 u~ values within a few seconds. The~. and u. can be easily

... ~max ~min
obtained. Even when there are, say, five attributes and five weights
(which makes 2 10 = 1,024 combinations), the computation of u i values
and the identification of their maximum and minimum are trivial.
As the number of attributes increases, the number of u i that needs

to be computed increases dramatically. For example, when there are 10
attributes, a total of 220 (= 1,048,576) u i values must be calculated,
and for 20 attributes, we must calculate 2 40 (= 1,099,511,600,000) u i
values. As a general rule, we do not recommend Baas and Kwakernaak's
[B1] approach when the number of attributes exceeds 10.
5.2.2 Kwakernaak's Approach
Kwakernaak's [K32] approach is a modification of Baas and
Kwakernaak's approach. Kwakernaak pointed out that the use of
trial-and-error to identify the desired u i values was not efficient.
Therefore, an improved algorithm was proposed.
Given fuzzy weights Wj = {(Yj'~w.(Yj»} and fuzzy attribute

)
r i ). = {(x ..
1.)
,~
r ij
(x .. »} for alternative Ai' the following steps are
1.)
used to derive fuzzy utility Ui = {(u"~U

1. i
(u.»}.
1.
step 1. Choose an a o level.
step 2. For alternative Ai' determine the following real numbers

(see Fig. 5.7):
-
x ij min { X ij E R I ~ (x .. ) ii!: ao }, Ifj , (5.8)
r ij 1.)
*
X ij max { X ij E R I ~r .. (X ij ) ii!: ao }, Ifj, (5.9)
1.)
Yj
- min { Yj E R I ~w.(Yj) ii!: ao } , Ifj , (5.10)
)
302
max { Yj e R I gW.(Yj) ~ a o }, vj. (5.11)

J
step 3. - *ij ].
At the a o level, the r ij may be represented by [Xij'X
Put X-ij ' vj, in an order such that
-
m1 s m2 s ... s mn , - (5.12)
-
where m1 min x ij and - The corresponding Yj- values
j
will also be put in an order such that
(5.13)
Similarly, x *
ij are rearranged in an order such that
(5.14)
where m*
1 min x *
ij and *
mn * .
max Xij The corresponding Yj* will
j j
be put in an order such that
(5.15)
lUx)
1.
Fig. 5.7 a-level sets for Wj and r ij .

303
I'(x)
1.
Fig. 5.8 Fuzzy numbers r 11 , r 12 , and r 13 .
For example, in Fig. 5.8, there are three fuzzy numbers r 11 , r 12
and r 13 . Since x~2 ~ x 13 ~ x 11 , we set
m1 x 12 , m2 x 13 , m3
-
x 11 ·
*
Similarly, since x 13 ~ *
x 12 ~ * , we set
x 11
m*1 * ,
x 13 m*2 * ,
x 12 m3* * ·
x 11
Step 4. Let fuzzy utility U; at the a o level be C

~ ao
aa and b a are defined as:
o o
j *
L zk m + L
n
z- m
k=l k k=j+1 k k
min j * n ],
(5.16)
o ~ j ~ n
L zk + L zk
k=l k=j+1
j n
L zk mk* + L z * mk*
k=l k=j+1 k
ba max j ] . (5.17)
0 0 ~ j ~ n
L
-
zk + L
n
z*
k=l k=j+1 k
304
The lower bound aaois computed according to the following
concept. Since x ij appears only in the numerator of the function,
n n
E YJ" x ij / E YJ", (S.18)
j=l j=l
the minimum of u i is guaranteed when x ij ' Vj, take x ij as their
values. However, u" cannot be guaranteed, even when YJ", Vj, take
~min
Yj as their values. Thus, we need only be concerned with the
combinations of Y; and Yj such that u i is minimum. Similar remarks
can be applied to Eq.(S.17).

The algorithm may be applied several times to different a
levels. Eventually, an approximated fuzzy utility Ui can be obtained.
Numerical Example
The same example solved by Baas and Kwakernaak's approach (B1]
will be solved again.
Step 1. Set a o 0.7S.
Step 2. The intervals of fuzzy attributes r ij , - ij* ], and

(Xij,X
- * are (see Fig. S.4):
the interval of fuzzy weights wj , (Yj'Yj]'
[.7S,.8S] [.SS,.6S] [.9S,1.0] [.lS,.2S]
- ll
The interval, [Xll,X * ], is obtained as:
min { x ll e R I ~r (x ll ) ~ 0.75} 0.75

11
max { x ll e R I ~r (x ll ) ~ 0.7S} 0.8S.

11
305
step 3. * as:
Rearrange x ij and x ij
m1 - m2 m1* m2*
.55 .75 .65 .85
similarly, the order of Yj and Yj* are summarized as:
zl
- z2 zl* z2*
.15 .95 .25 1.0
step 4. At 0: 0 0.75, we have ao:=.75 and bo:=.75 as:
zl m1 + z2 m2 zl* m1 + z2 m2 zl* m1 + z2*

min [
0::5j::52 z~ + z; zl* + z2 * + z2*
zl
min [ .7227, .7083, .7100] = .7083
zl m1* + z2* m*
2
zl + z2*
max [ • 8100 , . 8239, . 8227] = . 8239 •
Therefore, C. 75 = [ a. 75 ,b. 75 ] = [.7083,.8239] such that ~U1 (.7083)
~u (.8239) = .75. The result is identical to that of Baas and

1
Kwakernaak's [B1] approach.
By setting different 0: levels we can obtain the fuzzy utilities
U1 and U2 as shown in Fig. 5.6.
ranking method(s) presented in Chapter 4. Here by simple observation,

Note
1. Kwakernaak finds that by taking the maximum values of x ij ' we
are guaranteed to have a maximum u i . Similarly, by taking the minimum
of x ij ' we get a minimum u i • However, since Wj appears both in the
numerator and denominator, a maximum Wj does not guarantee a maximum
u i and a minimum Wj does not guarantee a minimum u i • Hence, the number
of possible combinations one needs to test is Cn2n instead of 22n as in
Baas and Kwakernaak's approach where n is the number of weights.
For example, for a problem with five attributes and five weights,
Baas and Kwakernaak's approach requires 2 10 = 1024 tries, while
Kwakernaak's approach needs only c;O = 252 tries. For a problem of 10
attributes and 10 weights, Baas and Kwakernaak's approach requires a
maximum of 1,048,576 tries while Kwakernaak's approach requires
184,756 tries. The improvement is obvious.
5.2.3 Dubois and Prade's Approach

Dubois and Prade [D28] pointed out that Baas and Kwakernaak's
approach can only effectively solve two-attribute problems. Since the
trial-and-error technique, which was used in Baas and Kwakernaak's
[B1] approach to derive fuzzy utilities, was not efficient, an
alternative approach was proposed.
Dubois and Prade's approach also uses the a-cut technique, but
provides a more efficient search procedure for obtaining u i values.
The algorithm assumes that all fuzzy weights Wj and fuzzy rating r ij
are normalized fuzzy numbers. Given an a level, we can obtain an
a-level set for each fuzzy weight and fuzzy rating The a level sets
are used to derive fuzzy utilities based on the classical SAW method.
307
Algorithm
step 1. set an a level and determine a-level sets for Wj and r ij
to be:
(5.19 )
(5.20)
Recall that the actual computation of Eq. (5.1) is tedious (or even
impossible) for continuous membership functions. This step allows
us to focus on two points of each fuzzy number at any given time, thus
simplifying the computation for the later steps.
step 2. Compute normalized fuzzy weights, P j , vj. Given the
a-level sets of wj , - *
[Yj'Yj]' j = 1, ... ,n, we can obtain n a-level sets
- * of the normalized fuzzy weights P j , vj, as:
[Pj'Pj]
Pj* Yj* / (Yj* + L Y~) (5.21)

k"j
and
Pj Yj / (Yj + L Yk* ) . (5.22)

k .. j
1, (5.23)
must hold.
step 3. For alternative, Ai' the rating r ij may be represented
by an a-level set as in Eq.(5.20). That is,

308
We are going to order x -ij and x ij

* ' ~j, respectively, as:
m1 - ~
-
m2 ~ ~
mn (5.24)
in which m-1 min x ij - and mn

-
max x ij ' and
j j
m*
1 ~ m2* s ... s m*
n (5.25)
in which m*1 *
min x ij and mn* = max x ij
* • Eqs.(5.24) and (5.25)
j j
facilitate the construction of Eqs.(5.24) and (5.25) in the later
steps.
step 4. The smallest upper and the largest lower bound of Ui ,
[u. and u ], are computed as:

l.min' i max
u. (
d-1
E Pj*
-
mj ) + [1
d-1
- E p.* -
n
Ep. - -
md +
n
E - -
p. mj
l.min j=1 j=1 J j=d+1 J j=d+1 J
(5.26)
u imax
e-1
(E - e-1
Pj mj* ) + [1 - L p: - E
n n
* me* + E p.* mj*
p.]
j=1 j=1 J j=e+1 J j=e+1 J
(5.27)
The only unknowns in Eqs.(5.26) and (5.27) are parameters d and e.
The parameter d can be determined when condition,
d-1 * n
1 - E
j=1
PJ' - E
j=d+1 J
p: Zd e [P~ , P~ ], (5.28)
is satisfied. The search process is carried out in the following
manner. By substituting d with 1, the value of z1 can be obtained.

- * is true.
We can easily determine if z1 e [P1,P1] If the answer is yes,
309
we can set d = 1 and compute u. using Eq. (5.26); otherwise, we need

~min
to sUbstitute d with 2 and compute a z2 value. Again, we need to
- *
determine if z2 e [P2,P2]. The search process goes on with the value
of d increasing by one each time until the condition (5.28) is met.
The resulting zd will be the value assumed by weight wd ' and
d-1 * n
r
j=l
PJ' + r p.
j=d+1 J
+ zd 1
which satisfies Eq. (5.23). Dubois and Prade [028] have shown that
there is only one value of d such that condition (5.28) is satisfied.

similarly, we can determine the value of e when the condition,
e-1 n *
1 - r
j=l
p. -
] j=e+1
r Pj (5.29)
is satisfied. The resulting e value is used in Eq. (5.27) to compute
The values assumed by wj ' Vj, must satisfy Eq. (5.23), i.e.,
e-1 n *
r
j=l
PJ' +
j=e+1
r Pj + z e 1.
step 4. At any a level, the fuzzy utility Ui can be represented
by the interval [u. ,u. ]. The OM can set several a levels and
~min ~max
repeat the algorithm several times to derive an approximated fuzzy
utility Ui .
Numerical Example ( Modified from Bonissone [B27])
Three possible investment alternatives--commodity (A 1 ), stock
(A 2 ), and real estate (A 3 )--are to be evaluated according to four

310
attributes: risk of losing capital (Xl)' inflation impact (X 2 ),

interest received (X 3 ), and cash realizability (X 4 ). The decision
matrix is given as:
Xl X2 X3 X4
[
A1 high mol high v. high fair
D A2 fair fair fair mol good
A3 low v. low mol high poor
1
The weight vector assigned by the DM is
mol mol very mol

important important important unimportant]·
Each linguistic term may be represented by a trapezoidal fuzzy number

(a,b,c,d) as shown in Fig. 5.9a. Table 5.1 contains the fuzzy numbers
for our decision problem.
Table 5.l. The Decision Matrix with Fuzzy Numbers
Xl X2 X3 X4
A1 (0,0,.1,.3) (0, .2, .2, .4) (.8,1,1,1) (.3, .5, .5, .7)
D A2 (.3,.5,.5,.7) (.3, .5, .5, .7) (.3, .5, .5, .7) (.6,.8,.8,1)
A3 (.7,.9,1,1) (.8,1,1,1) (.6,.8,.8,1) (0,0, .1, ~3)
W (.6,.8,.8,1) (.6, .8, .8,1) (.8,.8,1,1) (0, .2, .2, .4)
We shall use the algorithm to compute the fuzzy utility of A1 .

step 1. Let's assume the ~ level is set at 0.75. And the
~-level sets Wj~ = - * and r 1ja
[Yj'Yj] = - ,X 1j
[X 1j * ], j = 1,2,3,4, are:
311
w1a w2a w3a W4a
[.75,.85] [.75,.85] [.95,1. ] [.15,.25]
r 11a r 12a r 13a r 14a
[0,.15] [.15,.25] [.95,1.] [.45,.55]
I·/,(x)
1.
Fig. 5.9a Trapezoidal fuzzy number (a,b,c,d).
Step 2. Normalizing fuzzy weights wj , Vj. The a-level sets of
the ..
normal~zed we~ghts
- * are computed using Eq. (5.21) and
P j , [Pj'Pj]'
(5.22). For example, the a-level set of P 1 is computed as
Y1* .85
.315
Y1* + Y2 + Y3 + Y4 .85 + .75 + .95 + .15
Y1 .75
-
Y1 3 + Y4*
+ Y2* + Y* .75 + .85 + 1. + .25
.263.
We can obtain a-level sets for P j , j 2,3,4, in a similar manner.

They are summarized as:
312
[.263,.315] [.263,.315] [.328,.377] [.053,.093]
Step 3. Given the a-level sets, we can order the four lower
bounds such that
o $ .15 $ .45 $ .95.~ m1- $ m2
and the four upper bounds such that
step 4. Find values for the parameters d and e such that
Eqs.(5.28) and (5.29) are satisfied.
First of all, set d = 1 and compute zd=1 as:
.355
Hence, we reject the proposition that d 1. Next, we set d 2 and
compute zd=2 as:
1 - p~ - (p; + p~) .304
Here, we conclude that ~= 2. The values which weights wj , Vj, assume
are determined as:
.315 .304 .328 .053
Set d 2 and compute u 1 . using Eq.(5.26) as:

m1n
313
.244
where m-1 = 0, m2 - -
.15, m3 .45, and m4 = .95 as obtained in step 3.
similarly, we can set e 1, and compute ze=l to check if

condition (5.29) is satisfied. The result is
.215
We then set e = 2 and compute ze=2 as:
.267
Hence, we conclude that e 2. The values which w., Vj, assume can be
J
summarized as:
.263 .267 .377 .093
Setting e 2, we can compute u. using Eq. (5.27) as:

lmax
.407
where m1* = .15, m2* .25, m3* = .55, and m*

4 = 1.0 are from Step 3.
The a-level set of u1 ' [U 1 . ' u1 ], at a = 0.75 is
mln max
U1a =.75 = [.244, .407].
step 5. By setting a levels at 0 and 1, we can obtain a-level
sets U1a =0 and U1a =1 in a manner similar to the way we obtained
U1a=.75. We can summarize these a-level sets as:

314
P.(x) U1 U2 U3
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 5.9b The fuzzy utilities.
[.075,.620] [.244,.407] [.289,.337]
The fuzzy utility U1 can be depicted as shown in Fig. 5.9b.
The fuzzy utilities U2 and U3 can also be obtained in a
similar manner (see Fig. 5.9b).
The ranking of U1 ' U2 , and U3 can be accomplished by using any

proper ranking methodes) presented in Chapter 4. Here by simple
observation, we see U3 > U2 > U1 .
Note
Recall in step 4 it was necessary to find the values d and e in
order to directly compute u. and u 1· max using Eqs.(5.26) and (5.27),

lmin
respectively. To check if Eq. (5.28) is satisfied, one needs to test
at most n different values (where n is the number of attributes), in
order to find an appropriate d value. Similar remarks hold true for
the e value. The total number of testing for this algorithm is 2n at

315
most. Comparing this number with (2)2n (in Baas and Kwakernaak's
approach) and c~n (in Kwakernaak's approach), we conclude that this

algorithm is the least time-consuming one. For example, we need to
test 1024 combinations for a five-attributes, five-weights problem
using the Baas and Kwakernaak algorithm. It takes 252 tries using the
Kwakernaak algorithm, while only 10 tries are needed using the Dubois
and Prade algorithm.
5.2.4 Cheng and McInnis's Approach
Cheng and McInnis [C17] pointed out that continuous membership
functions of r ij and Wj are the cause of the complexity of obtaining
fuzzy utilities. To avoid such difficulty, they suggested to first
convert the continuous membership functions to discrete ones and then
compute the fuzzy utilities using the following algorithm.
Algorithm
The following steps are taken for deriving fuzzy utilities.
step 1. The continuous membership function is converted to a
discrete one. This is done by having the OM specify the number of
a levels sjhe wants to use. The width of intervals is determined
according to the OM's preference. For example, Fig. 5.10a gives a
continuous membership function. It may be represented by Fig.5.10b,
Fig.5.10c, or any other shape with different numbers of a levels and
widths of intervals.
The OM may specify different numbers of a levels and widths of
intervals for different membership functions in an MAOM problem. For
example, the membership functions w1 and r 11 in Fig. 5.4 may be
redrawn as shown in Figs. 5.11a and 5.llb, respectively. The number
of a levels for w1 is four while the number of a levels for r 11 is
only two. The width of the interval for w1 is uniform, i.e.,

316
P.(x) M
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 5.10a A continuous membership function ~(X).
P.(x) M
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 5.10b A possible discretized ~(x).
p.(x) M
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 5.10c Another possible discretized ~(x).

317
",(x) W1
Ci ---------------------------------------------------
/3 ---------------------------------------------
'Y ----------------------------------------
l) --------- -------------------------
Fig. 5.iia The discretized wi of Fig. 5.4.
"'(x) r11
Ci --------------------------------------- r - - - - - ,
/3 ----- ----- --- -- -------- -----
~~--~--~~--~--~~--~--~-+ x
o .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 5.iib The discretized r ii of Fig. 5.4.
a2 - ai = a3 - a2 = a4 - a3 = a 5 - a 4 , while the width of the inter-

vals for r ii is different. These variations are rather arbitary.
step 2. For each a-level, we need to perform steps 3 and 4.
The first a level to be considered is the largest one among all the Wj
and r ij graphs.
step 3. Given a o ' we can obtain the a-level set for each r ij and
each Wj as:
318
(5.30)
and
(5.31)
- ij
That is, at a o ' r ij can take any value in the interval [Xij'X * ] and Wj
- *
can take any value in [Yj'Yj].
step 4. Given the upper and lower bounds of r .. and w. at the

~] ]
a o level as shown in Eqs. (5.30) and (5.31), we can compute the upper
and lower bounds of the fuzzy utility at a o ' U [u. ,u. ],
ia o ~min ~max
using the following process.
step 4.1. Compute u imax To obtain the upper bound of Ui at the
a o level, U *
imaX ' the upper bound of r ij , vj, i.e. , x ij ' must be used.
Taking the equation
(5.33)
since the x ij value appears only in the denominator, a higher x ij
value will guarantee a larger u i value; on the other hand, since Yj'
Vj, appear in both numerator and denominator, increasing Yj may not
give a larger u i value. Thus to maximize u i ' we must decide whether
Yj- or Yj* should be used.

Cheng and McInnis proposed a search process to test whether Yj* or
Yj should be used by comparing the maximum values of all r ij . First

* , V],
of all, x ij . are rearranged as:
m1* :S *
m2:s :S *
mn (5.34)
in which m1* min x *ij and m* max x *... Assume m*1 x*

ik ' the
j n ~]
j
319
- as its value.
corresponding wk should take Yk *
Assume mn xiI' the
corresponding wI should take YI* as its value. * such that
For some mp
m1* < m* < mn* , if the condition

p
n
n L w.
j=l J
jE1Wj r ij j .. p
n n
L w. L w. + (w +A) (5.35)
j=l J j=l J P
j .. p
where A is any positive real number, holds, then the upper bound of
.
1.e., * should be selected.
yp' otherwise, y~ is selected.
Given the right combinations of y; and yj, Vj, we can easily
compute u imax using Eq. (5.33).
step 4.2. After finding u. , u. can be easily identified.

lmax lmin
First of all, we will use X 1j for all r ij . Secondly, for those Wj
whose upper bounds were used for deriving u imax ' we will use their
lower bounds in computing u. and vice versa.

lmin
Steps 3 and 4 are used for the next largest a-level until all a
levels are exhausted. The resulting fuzzy utilities are also discrete
and have several "steps" in it.
Numerical Example
We shall use the same numerical example solved in Section 5.2.1
to demonstrate this algorithm. For alternative A1 , the following

steps are performed.
Step 1. Assume that the membership functions in Fig. 5.4 are
redrawn as shown in Figs. 5.12a and 5.12b, based on the OM's
specifications.
Step 2. The highest a level in this case is 1, i.e., a o 1.

320
p.(x)
1.
.9
.8
.7
.6
.5 r-- =-
.4
.3
.2
.1
o x
o .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 5.12a Discretized r 11 and r 12 of Fig. 5.4.
P.(X) w2 w1
1. r--
.9
.8
.7
.6
.5 r--- '--
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 5.12b Discretized w1 and w2 of Fig. 5.4.
step 3. The lower and upper bounds for r ij and Wj at a o 1 can

be obtained as:
1 [.7,.9] [.9,1.]
2 [.5,.7] [.1,.3]
step 4. To obtain U1a [U 1 . ' u 1 ], the following steps

o ml.n max
will be taken.
321
step 4.1- Since a higher r ij value guarantees a larger u i value
in Eq.(5.33), r 11 and r 12 .
w~ll * and x 12
take x 11 * ' .
respect~vely, as
their values in computing u 1
max
As for the appropriate wj ' the following steps must be taken.
.
S~nce there are two "
attr~butes, ~.e., n = * > x 12
2, and x 11 * ' we know
where m*
1 * and m2*
x 12 * ·
x 11 Based on Cheng and McInnis's search
process, w1 should take Y1* as its value while w2 should take y; as its
value.
Having identified the appropriate r 1j and wj ' j 1,2, values,
u1 is easily computed as:

max
.882.
step 4.2. In this step, r l l and r 12 will assume x l l and x 12 as
their values, respectivly. Furthermore, since w1 and w2 take Y1* and

-
Y2' respectively, as their values in computing u 1 ,we will use Y1 -
max
and Y2* . ..
~n der~v~ng u1 . as:
m~n
.650.
We conclude that fuzzy utility Ul at a o = 1 is any real number in the
interval [.650,.882]. These two points should be plotted as extreme
points of the interval with an a o = 1 in Fig. 5.13.

322
,,(x)
1. U2 U1
.9
.8
.7
.6
.5
.4
.3
~ ~
- '----
.2
.1
o x
o .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 5.13 The fuzzy utilities of the discrete form.
steps 3 and 4 will be used repetitively until all a o levels are

exhausted. In this case a o is set to 1, .5 and 0 in sequence. The
fuzzy utility U1 is shown as in Fig. 5.13. The algorithm is repeated
for alternative A2 . Its fuzzy utility, u 2 ' can also be seen in
Fig.5.13.
ranking methodes) presented in Chapter 4. Here by simple observation,
Note
The process of finding u imax and u ioutperforms those in
min
sections 5.2.1, 5.2.2, and 5.2.3 because it requires less
computational efforts. It is doubtful, though, that Cheng and
McInnis's stepwise discrete membership functions can represent fuzzy
information better than the simple piecewise linear membership
functions as many ( Tong and Bonissone [T11], Laarhoven and Prdeycz
[L1], Buckley [B39], Godo and Sierra [G7]) pointed out that fuzzy data
323
can be best described by either trapezoidal or triangular shaped fuzzy
numbers. Also many approximated arithmetic operations can be easily
applied to obtain approximated fuzzy utilities. The computational

•
requirements are even less. This is demonstrated in Bonissone's
approach in the next section.
5.2.5 Bonissone's Approach
Bonissone [B27] assumed that fuzzy/crisp information in decision
problems can be approximated by a parameter-based representation. It
is called the L-R type trapezoidal number (a,b,a,~) (see Fig 5.14).
It is this family of fuzzy numbers which allows approximated arithmetic
operations on fuzzy numbers (see Section 3.4).
Fuzzy Arithmetic Operations
The fuzzy arithmetic operations have been presented in Table 3.4

of Chapter 3. Here, we shall list only the formulas regarding the
addition, subtraction, multiplication, and division operations.
Let fuzzy numbers M = (a,b,a,~) and N = (c,d,7,~), and M > 0 and N >
o. Their arithmetic operations can be dispayed as:
M(+)N (a+c, b+d, a+7. ~+~) (5.36)
M(-)N (a-d, b-c, a+~. ~+7) (5.37)
M(o)N (5.38)
a b a~+da b7+C~ )
M(+)N
(d' c' d(d+~)' C(C-7) (5.39)
Note that only Eq. (5.36) is exact. The remaining equations are only
approximate in that the left and right spreads of the calculation

324
P.(x)
1.
Fig. 5.14 L-R type trapezoidal fuzzy number (a,b,a,~).
results are not exact. However, they introduce very little error.
with the help of the approximated algebraic operations, we can
quickly compute the performance of alternative Ai with respect to
attributes, x j ' j=l, ... , n, using
n
E w. r .. (5.40)
j=l J 1.J
where Wj and r ij may be crisp or fuzzy numbers represented in the L-R

trapezoidal number format.
Numerical Example (Modified from Bonissone [B27])

Three possible investment alternatives--commodity market, stock
market, and real estate--are to be evaluated according to four
attributes: risk of losing capital (Xl)' inflation impact (X 2 ),
interest received (X 3 ), and cash realizability of the capital (X 4 ).
The decision matrix is given as:
325
xl X2 X3 X4
A1 high mol high v. high fair
D A2 fair fair fair mol good
A3 low v. low mol high bad
The weight vector is given as:
~ = [mol important, mol important, v.important, mol unimportant].
Table 5.2 presents the fuzzy set associated with each linguistic term
(see Fig. 5.15).
The fuzzy utility for alternative A1 is computed using Eg. (5.40)
as:
4
E w)' Xl)'
j=l
(1.26,1.34, .62, .64)
Table 5.2 Meaning of Linguistic Terms in Trapezoidal Numbers
Interpretation when used with

Shape Fuzzy Numbers X X4 Weights
1
1 (0,0,0,.2) v.high v.high v.low v.bad v.unimp.

2 (0, .1,0, .2) high high low bad unimp.
mol mol mol mol mol

3 (.2, .2, .2, .2)
high high low bad unimp.
4 (.5, .5, .2, .2) fair fair fair fair indif.
mol mol mol mol mol

5 (.8, .8, .2, .2)
low low high good imp.
6 (.9,1, .2,0) low low high good imp.
7 (1,1,.2,0) v.low v.low v.high v.good v.imp.
Note: Unimportant unimp., Indifferent = indif.,

Important = imp., Very = v., More or less = mol.
326
~~
1.
~
B
.7
~
.5
A
~
2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 B .9 1.
Fig. 5.15 Fuzzy representation of linguistic terms.
~(x)
1.
o
where additions and multiplications are conducted using Eqs. (5.36) and
(5.38), respectively.
We can obtain U2 and U3 in a similar manner. The fuzzy utilities
can be summarized as (see Fig. 5.16):
(1.26,1.34,.62,.64) (1.46,1.46,.86,.80) (2.32,2.42,.94,.52)

327
The ranking of U1 ' U2 , and U3 can be carried out by using any
proper ranking methodes) presented in Chapter 4. Here by simple
observation, we see U3 > U2 e U1 •
Note
1. Bonissone's approach is much simpler to use than other fuzzy
SAW methods. It is applicable only when fuzzy concepts are
represented by trapezoidal or triangular numbers. If that assumption
does not hold, then other approaches discussed earlier may be
appropriate. In addition, Bonissone's approach generates less precise
fuzzy utility, i.e., the spreads generated by Bonissone's approach
are larger than those which were ~enerated by other fuzzy SAW methods.
Since L-R trepezoidal fuzzy number provides satisfactory
explanations to fuzzy concepts, Bonissone's approach may be an
appropriate method to use when larger spreads are tolerable. We
recommend this approach for its simplicity.
To show that this approximation approach is indeed simpler than
the ~-cut approaches discussed in previous subsections, we shall solve
the numerical example presented in section 5.2.1.
Given the decision matrix as:
Xl X2
o
good
fair
fair
good ]
! Very important, rather unimportant]
The linguistic terms are converted to 4-tuple representation as (see
Fig. 5.4):
WI very important = (1.0,1.0,.2,.0),
rather unimportant (.2, .2, .2, .2),

328
good (.8, .8, .2, .2),
fair (.6, .6, .2, .2).
The fuzzy utility of A1 is computed using Eq.(5.40) as:
(.92, .92, .44, .40).
Similarly, we can obtain
U2 = (.76,.76,.44,.36).
Fig. 5.17 presents U1 and U2 . Comparing Fig. 5.6 with Fig. 5.17, we
find the ordering for A1 and A2 remains the same but the spreads of Ui
in Fig. 5.17 tend to be larger. That is attributed to the fact that
Bonissone's arithmetic formulas are only approximations of the real
formulas.
,.,.(x)
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
Ol-.....-L....r--'-...--........-,......L...,........l....,-...-........--.-x
o 2

329
5.3 Analytic Hierarchical Process (AHP) Methods
In the classical AHP method, the DM is asked to supply ratios a ..

~J
for each pairwise comparison between alternatives A1 , A2 , ... , Am for
each attribute in a hierarchy, and also between attributes. This
results in n+1 positive pairwise comparison matrices, where n is the
number of attributes. Each matrix is represented as:
all a 12 a 1n w1/w1 w1/w2 w1/wn

A a 21 a 22 a 2n w2/w1 w2/w2 w2/wn
a n1 a n2 a wn/w1 wn/w2 wn/wn

nn
(5.41)
which is a 'reciprocal matrix' with all a ij being positive. The ratio
a ij represents, for the DM, the relative importance of Ai over Aj .
For example, when the DM considers A1 more important than A5 , a 15
might equal 3/1, or 5/1, or 7/1, or 9/1. Since the numbers for the
ratio are usually taken from the set {1, 2, ... , 9}, a 15 could be sl/s2
where sl' s2 E {1,2, ... ,9}. Note that if a 15 = 3/1, then a 51 must be
1/3. This is why matrix A is called a 'reciprocal matrix'.
The AHP method uses the pairwise comparison matrices for each
attribute to compute the performance score of alternative Ai with
respect to attribute Xj , r ij . The pairwise comparison matrix for the
attributes is used to compute the weights of the attributes. The
performance scores and weight set are organized as:
Xl X2 X
n
A1 r 11 r 12 r 1n
D A2
A
m
[ r 21
r m1
r 22
r m2
r 2n
r
mn 1
(5.42)
w (w 1 , w2 ' ... , wn )
330
where r ij and wj , Vi,j, take their numbers on the real line R. The
classical SAW method is used to find the utilities of Ai' Ui , vi:
n n
L wJ' riJ' / L wJ'. (5.43)
j=l j=l
The Analytic Hierarchical Process (AHP) method was first proposed
by saaty [Sl,S2]. In saaty's approach, the pairwise comparison ratios
a ij , Vi,j, are real numbers. Each pairwise comparison matrix is

solved using the eigenvector method. The resulting weights and
performance scores are also crisp, real numbers. The classical SAW
method is used to calculate the alternatives' utilities.
saaty's AHP method was extended by Laarhoven and Pedrycz [L1] in

1983. They argue that if a person considers A1 more important than
A5 , then the ratio a 15 might be "approximated 3 to 1," or "about 5 to
1," or "between 5 to 1 and 7 to 1." These linguistic expressions are

expressed by triangular fuzzy numbers. That is, a ij is a triangular
fuzzy number. In addition, Laarhoven and Pedrycz allow several OMs to
express their ratios on the same pair of alternatives (or attributes).
In this case, the pairwise comparison ratio may be represented by a, 'k

~J
(k = O,l, ... ,Pij). The term Pij denotes the number of persons who
expressed their comparison ratios.
There are many methods one can use to derive performance scores
and attributes' weight. According to Laarhoven and Pedrycz, because
of the presence of fuzzy, multiple comparison ratios for the same pair
of alternatives (or attributes), the most suitable method for their
approach is Lootsma's [L1] logarithmic least square method. Once the
fuzzy performance scores r ij and the fuzzy weights Wj have been
derived, fuzzy arithmetic operations that are suitable for triangular
fuzzy numbers are used to obtain the fuzzy utilities, Ui , Vi' where
r ij , wj , and Ui are triangular fuzzy numbers.

331
P.(x)
1.
Fig. 5.18 Trapezoidal fuzzy number (aij'~ij'7ij'~ij).
Buckley [B36,B39] also extends Saaty's AHP method to the case in

which the OMs can express their preference in fuzzy ratios instead of
crisp ratios. The fuzzy ratios a ij are given as a trapezoidal number
(aij'~ij,7ij'~ij) where 0 ~ a ij ~ ~ij ~ 7 ij ~ ~ij (see Fig.S.1S).

The geometric mean method is employed to calculate the fuzzy weights
Wj and the fuzzy performance scores r ij . Note that the derived Wj

and r ij may not be trapezoidal fuzzy numbers anymore. In this case,
special fuzzy arithmetic formulas are needed in order to add and/or
multiply them. Buckley [B36] has develped some special
fuzzy arithmetic formulas for that purpose.
5.3.1 Saaty's AHP Approach
Saaty [Sl,S2] states that there are two types of fuzziness:
fuzziness in perception and fuzziness in meaning. The first one is
caused by complexity of objects or ideas which cannot be apprehended
at once. The second one is attributed to relativism of meaning, i.e.,
the meaning of objects is tied to what function those objects perform
in the fulfillment of different purposes. When we decompose the

332
objects, they appear fuzzy because they have different meanings
according to the context of the decomposition.
A method is proposed to give meaning to both kinds of fuzziness.
This method measures the relativity of fuzziness by structuring the
functions of a system hierarchically in a multiple-attribute
framework.
Before introducing the algorithm, we shall briefly introduce a
method --the eigenvector method~- that can be used to generate the
relative importance of attributes and the performance scores.
Eigenvector Method
Let the positive reciprocal Matrix A be
(5.44)
where
1 / a ji , 'v'i,j, (5.45)
a .. (5.46)
1)
(5.47)
Matrix A is called a 'reciprocal matrix'.

T
Multiplying A by ~ = (Wi' W2 ' ... , Wn ) yields
nnT
A W Wi /W 2 nw
1[~:l
Wi/Wi Wi/W n
[ W2 /W i
Wn/W i
W2 /W 2
Wn /W 2
W2 /W n
Wn/Wn n
(5.48)
or
(A - n I) W o. (5.49)
Due to the consistency property of Eq.(5.46), the system of
homogeneous linear equations, Eq.(5.49), has only trivial solutions.
In general, the precise values of wi/w j are unknown and must be
estimated. In other words, human judgments cannot be so accurate that
Eq.(5.46) be satisfied completely. We know that in any matrix, small
permutations in the coefficients imply small permutations in the
eigenvalues. If we define A' as the DM's estimate of A and w'
corresponds to A', then
A' w' max w'

A (5.50)
where Amax is the largest eigenvalue of A'. w' can be obtained by

solving the system of linear equations, Eq.(5.50). The consistency of
the estimates in the matrix, A, is guaranteed when Amax s n. When
Amax is not close to n, we must modify the estimates in A so that
consistency is preserved.
The comparison scale uses ranges 1 to 9, each representing fuzzy
entries as follows:
1: equally important
3: weakly more important
5: strongly more important
7: demonstratively more important
9: absolutely more important
2, 4, 6, and 8 are intermediate values between two adjacent judgments.

Given any a ij , the reciprocal value a ji = l/a ij . Obviously, a ii = 1
is always true.
Hierarchical Decision structure

A hierarchy structure can be best described by the following
example. Three job offers are considered by a new Ph.D. The
attributes considered are research, growth, benefits, colleagues,
location, and reputation. Since the entries of this MADM problem are
only vaguely known, we cannot form a MADM decision matrix. To resolve
this problem, a three-level hierarchy is constructed (see Fig. 5.19).
The data of this MADM problem can be derived from this hierarchy
decision structure.
The first level consists of a single objective, to have a
successful career. The importance of it is assumed unity. The second
level consists of six attributes. Their relative importance is
determined using the eigenvector method with respect to the objective
in the first level. The third level consists of the three jobs being
considered. Their relative performances are derived using the
eigenvector method with respect to each attribute. The weights and
performance scores are then combined using the SAW method. The
results are the final ratings of the three jobs. The job with the
highest final rating can best fulfill the ultimate goal (successful
career) .
1 st level Successful Career
2 nd level
3 rd level
Res.= Research. G.= Growth. B.= Benefit.

C.= Colleagues. L.= Location. Rep.= Reputation.
Fig. 5.19 Selection of job hierarchy.

335
Formal Hierarchy (Saaty [Sl]): It is essentially a formalization in

terms of partially ordered sets of our intuitive understanding of the
idea. It has levels: the top level consists of a single element and
each element of a given level dominates or covers (serves as a property
of a purpose for) some or all of the elements in the level immediately
below. The pairwise comparison matrix approach may then be applied to
compare elements in a single level with respect to a purpose from the
adjacent higher level. The process is repeated up the hierarchy and
the problem is to compose the resulting priorities (obtained by the
eigenvector method) in such a way as to obtain one overall priority
vector of the impact of the lowest elements on the top element of the
hierarchy by successive weighting and composition.
Let the symbol Lk represent the kth level of a hierarchy of h

levels. Assume that y = (Y1' Y2' ... , Yk) e Lk and that X = (xl' x 2 '
..• , xk +1 ) e ~+1. Also assume that there is an element ze~_l such
that Y is covered by z. We can then consider the priority functions
wz : Y -> [0,1] and Wy: X -> [0,1]. (5.51)
We construct the "priority function of the elements in X with respect

to z" denoted w, w: X -> [0,1], by
k
r
j=l
w (x.) Wz(YJ.)' i= 1,2, .•• ,k+1.
Yj ~
(5.52)
It is obvious that this is no more than the process of weighting the

influence of the element Yj on the priority of xi by multiplying it
with the importance of Yj with respect to z.
The algorithms involved will be simplified if one combines the
wy.(X i ) i~to a matrix B by setting b i ]. =w (x.). If we further set
] Yj 1
Wi = W(X i ) and Wj = Wz(Yj)' then the above formula becomes
k
L bi]'W]~' i 1,2, ..• ,k+l. (5.53)
j=l
Thus, we may speak of the priority vector wand, indeed, of the

priority matrix B; this gives the final formulation W = BW'.
A hierarchy is complete if all xe~ are dominated by every
element in ~-1' k = 2, ... , h. Let H be a complete hierarchy with
lowest element band h levels. Let Bk be the priority matrix of kth
level, k = 1, 2, ••• , h. If W' is the priority vector of the pth
level with respect to some element z in the (P_l)st level, then the
priority vector W of the qth level (p < q) with respect to z is given
as:
W Bq Bq- 1 ••. BP+l W'. (5.54)
Thus, the priority vector of the lowest level with respect to the
element b is given as:
W (5.55)
If Ll has a single element, as usual, W' is just a scalar; if it has

more elements, it is a vector.
Numerical Example (Choosing a Job - saaty [Sl])

A student who has just received his Ph.D. degree has been offered
three jobs. Six attributes have been selected for the comparison.
337
They are: research, growth, benefits, colleagues, location, and
reputation. His criteria for selecting the jobs and their pairwise
comparison matrix are given in Table 5.3. Due to the vague nature of
the criteria, he constructs the pairwise comparison matrices of the
jobs with respect to each criteria, rather than the decision matrix.
They are given in Table 5.4.
The eigenvalue of the matrix of Table 5.3 is Amax 6.35 and the
corresponding eigenvector is
Res G B C L Rep
[ .16 .19 .19 .05 .12 .30]T
The eigenvalue and the eigenvectors of the remaining matrices are
given as:
Amax 3.02 3.02 3.56 3.05 3. 3.21
Res G B C L Rep
[
B3 A .14 .10 .32 .28 .47 .77
1
B .63 .33 .22 .65 .47 .17
C .24 .57 .46 .07 .07 .05
The composite vector for the job with h 3 is given by
A
B
[ .40
.34
1
C .26
The differences were sufficiently large for the candidate to accept
the offer of job A.

338
Table 5.3 Pairwise Comparison for Attributes
Res. G. B. C. L. Rep.
1
Research 1 1 1 4 1
2
1
Growth 1 1 2 4 1
2
1 1
Benefits 1 1 5 3
2 2
1 1 1 1 1
Colleagues 4 4 5 1
3 3
1
Location 1 1 3 1 1
3
Reputation 2 2 2 3 1 1
Table 5.4 comEarison of Jobs with Respect to six Attributes
Research (Res) Growth (G)

A B C A B C
1 1 1 1
A 1 A 1
4 5
4 2
1
B 4 1 3 B 4 1
2
1
C 2 1 C 5 2 1
3
Benefits (B) Colleagues (C)
A B C A B C
1 1
A 1 3 A 1 5
3 3
1
B 1 1 B 3 1 7
3
1 1
C 3 1 1 C 5 7 1
Location (L) Reputation (Rep)

A B C A B C
A 1 1 7 A 1 7 9
1
B 1 1 7 B 1 5
7
1 1 1 1
C 1 C 1
7 7 9 5
Note
saaty's AHP method does not directly use fuzzy numbers or
membership functions to express fuzzy information. Rather, it uses

339
the estimation of an underlying ratio scale, together with the measure

of consistency to measure the fuzziness of a MADM problem. The
algorithm provides a good way to represent fuzziness which reveals the
properties of consistency, stability and pareto optimality.
Saaty's AHP method results in a cardinal order which can be used
to select or rank alternatives. When a hierarchy has only three levels,
the AHP method is equivalent to the classical SAW method. We feel that
the AHP method is really worth utilizing when there are more than three
levels in a hierarchy.
5.3.2 Laarhoven and Pedrycz's Approach

Laarhoven and Pedrycz [Ll] proposed an algorithm which is a
direct extension of Saaty's AHP method. Recall that in the AHP
method, fuzziness is not directly represented by fuzzy notations but
is indirectly modeled in a decision problem by forming a reciprocal
matrix. In this extended version of AHP, the elements in the
reciprocal matrix, A, are represented by triangular fuzzy numbers (see
Fig. 5.20). The computation steps are the same as those in AHP. The
Lootsma's logarithmic least square method is used to derive fuzzy
weights and fuzzy performance scores. The arithmetic operations for
fuzzy triangular numbers are applied to compute fuzzy utilities. The
opinions of multiple decision makers can also be modeled in the
reciprocal matrix.
Lootsma's Logarithmic Least Square Method

This weight-assessing method was chosen because it is suitable
for handling multiple decision maker's opinions and is easily extended
to the fuzzy case.
Let the positive reciprocal matrix A be represented as:
340
all a l2 a ln
[ [
A a 21 a 22 a 2n
anI a n2 ann
where a ij are real numbers. The estimated vector ~
is derived by minimizing
(5.56)
p.(x)
1.
Fig. 5.20 Triangular fuzzy number (l,m,u).
When there are multiple OMs, the weight vector ~ is derived by

minimizing
L (5.57)
i<j
where a ijk , k = 1,2, ... ,Pij' are Pij estimates for wi/w j . Note that
Pij can be 0 (if no comparison ratios are expressed), equal to one, or
341
greater than one, in which case there are several DMs who have
expressed their comparison ratios.
If we put Yijk = ln a ijk , zi ln wi' and Zj ln wj , we can
minimize
(5.58)
by solving the associated normal equations
p
n n n ij
z.1 E Pij - E Pij Zj E E Yijk' vi, (5.59)
j=l j=l j=l k=l
j .. i j"i j .. i
for zi. Taking the exponentials of the zi and normalizing them, we

can obtain estimates for the weight vector! =. (w 1 , w2 '···, Wn ).
Arithmetic Operations for Triangular Fuzzy Numbers

Triangular fuzzy numbers (see Fig.5.20) are assumed throughout
Laarhoven and Pedrycz's approach. Some arithmetic operations
performed on triangular fuzzy numbers, such as addition and multipli-
cation are presented here to facilitate the computation of fuzzy
utilities.
Assume that we have two triangular fuzzy numbers M1 = (11,m 1 ,u1 )
and M2 = (12,m 2 ,u 2 ). Their multiplication and addition are defined as:
(5.60)
(5.61)
342
Algorithm
The algorithm is shown in the following steps.
step 1. Consult with the DMs and obtain n+1 fuzzy reciprocal
matrices that take the following form (see Table 5.5).
Table 5.5 Fuzzy Reciprocal Matrix of Multiple DMs
D
(1,1,1)
(1,1,1)
a a
--~~~~!-------------------------~~~~~---
.
----------------------------------------
(1,1,1)
where a. . are fuzzy ratios estimated by multiple OMs. Note that

~JPij
Pij may be 0 when no OMs express their comparison ratios or greater

than 1 when more than one OM expresses his/her comparison ratios.
step 2. Let zi = (li,mi,u i ). Solve the following linear
equations:
n n P ij
l·(EP .. ) - E E [lnliJ·k],Vi. (5.62)
~ j=l ~J j=l k=l
j .. i j .. i
343
P ..
n n n 1J
m.( I: P .. ) I: P .. m. I: I: [In mijk ] , vi. (5.63)
1 j=l 1J j=l 1J J j=l k=l
j"i j"i j"i
P ij
n n n
u.( I: P .. ) - I: P . . 1. I: I: [In U ijk ], vi. (5.64)
1 j=l 1J j=l 1J J j=l k=l
j"i j"i j"i
As In(lijk) and In(u ijk ) are lower and upper values of In(a ijk )
-In (a jik ), the following must hold true.
0, Vi,j,k.
Thus Eqs. (5.62) and (5.64) are linear dependent. The same holds for
Eq.(5.63). Generally, a solution for Eqs.(5.62), (5.63), and (5.64)
is given as
(5.65)
where tl and t2 can be chosen arbitrarily.

step 3. Recall that in the above linear system, all right hand
sides have taken logarithmic operations. We have to take exponentials
on I i ' mi , and u i and compute the fuzzy weight, wi' as follows:
(5.66)
n -1
n -1
where Al ( I: exp(u i » , A2 ( I: exp(m i » ,
i=l i=l
n -1
A3 ( I: exp(li» .
i=l
344
Eq.(5.66) can also be used to determine the performance score r ij •

step 4. steps 1 through 3 are repeated several times until all
reciprocal matrices have been solved. with the fuzzy weights and
performance scores, we can compute the fuzzy utility for alternative
n
r
j=l
wJ' riJ' (5.67)
Note that the multiplication and summation of two triangular fuzzy

numbers are based on Eq.(5.60) and (5.61), respectively.
Numerical Example (modified from Laarhoven and Pedrycz [L1])

There are three candidates being considered for an assistant
professor opening. Four attributes are being considered--mathematic
creativity (Xl)' scholarship (X 2 ), research ability (X 3 ), and
communication skill (X 4 ). The problem is presented to a three-member
serach board, i.e., three decision makers. The following steps are
taken to solve the problem:
step 1. The three OMs' opinions about the relative importance
of a pair of alternatives (or attibutes) are presented in Tables 5.6
through 5.10. Notice that these matrices have missing elements, which
indicates that the decision makers need not express their opinions
for every pairwise comparison. Each element in the reciprocal matrix
is fuzzy. For example, a ij = (~, 1, ~) is equivalent to "about
equally important"; or a ij = (~, 2, ~) is equivalent to "attribute
i is about twice as important as attribute j."
step 2.
(5.64) on Table 5.6:
345
Table 5.6 Pairwise comparison of Candidates under Attribute 1
Al A2 A3
2 2
(1, 1, 1) (3' 1, 2~) (3' 1, 2~)
2 2 1 2
(3' 1, 2~) (5' 2' 3)
----2-----3--------------------2--1--2----
(3' 1, 2) (1, 1, 1) (5' 2' 3)
2 3
(3' 1, 2)
----2-----3------3-----5------------------
(3' 1, 2) (2' 2, 2) (1, 1, 1)
(~, 2, ~)
Table 5.7 Pairwise Comparison of Candidates under Attribute 2
(1, 1, 1)
(1, 1, 1)
(1, 1, 1)
Al A2 A3
(1, 1, 1)
5
(2' 3, 2)
2
5
(2' 3 2)
2
5
(2' 3, 2)
2
3
(2' 2, 2)
2
3' 2)
2 1 2
(7' (1, 1, 1) (3' 1, ~)
2 2
2 1
(7' 3' 5~)
2 1
(5' 2' 3~)
2 1 2
(7' 3' 5) (1, 1, 1)
346
A1 A2 A3
3
A1 (1, 1, 1) (2' 2, ~)
2
2 1 ~)
(5' 2' 3
(1, 1, 1)
(2 1 ~) (1, 1, 1)
5' 2' 3
(3 2 ~)
2' '2
Table 5.10 pairwise Comparison of Attributes
Xl X2 X3 X4
2 1 ~) 2 2 1
(1, 1, 1) (3' 2' 2 (3' 1, ~) (7' 3' ~)
2 5
2 1 2 1
Xl (5' 2' ~)
3 (7' 3' ~)
5
3 ~) 2 1
(2' 2. 2 (5' 2' ~)
3
------------------------------------------------
(~, 1, ~) (1, 1, 1) (5 3 Z) (~, 1, ~)
2' '2
(~, 2, ~) (~, 3, ~) (~, 1, ~)
(2 1 ~)
(~, 2, ~)
5' 2' 3
------------------------------------------------
5
(2' 3, Z)
2
2
(3' 1, 2~)
3
(2' 2, ~)
2
(1, 1, 1)
X4
5
(2' 3, Z)
2
2
(3' 1, 2~)
3 2 1
(2' 2, ~) (5' 2' ~)
2 3
347
3 3 3 Pij
11 E P1j E P 1j U j E E In (1 1jk ) (5.68)
j=2 j=2 j=2 k=1
3 3 3 Pij
12 E P2j - E P 2j U j E E In(1 2jk ) (5.69)
j=1 j=1 j=1 k=1
j"2 j"2 j"2
2 2 2 Pij
13 E P3 j - E P 3j U j E E In(1 3jk ) (5.70)
j=1 j=1 j=1 k=1
3 3 3 Pij
In1 E P1 j - E P 1j In j E E In(In 1jk ) (5.71)
j=2 j=2 j=2 k=1
3 3 3 Pij
In2 E P2 j - E P 2j In j E E In(m 2jk ) (5.72)
j=l j=1 j=l k=l
j"2 j"2 j"2
2 2 2 Pij
m3 E P3j - E P3jmj E E In(m 3jk ) (5.73)
j=1 j=1 j=1 k=l
3 3 3 Pij
U1 E P1 j - E P 2j l j L E In(u 1jk ) (5.74)
j=2 j=2 j=2 k=1
3 3 3 Pij
U2 E P2 j - E P 2j l j E E In(u 2jk ) (5.75)
j=1 j=1 j=1 k=1
j"2 j"2 jOO2
2 2 2 Pij
u3 E P3 j - E P3jlj E E In(u 3jk ) (5.76)
j=1 j=l j=1 k=1
Eqs. (5.68) through (5.76) are a set of simultaneous linear equations
and can be represented as:

348
411 - 2U 2 - 2U3 -2.133 (5.77)

31 2 - 2U 1 - u3 -1.727
31 3 - 2U 1 - u2 .405
4m 1 - 2m 2 - 2m3 - .693
3m2 - 2m 1 - m3 - .693
3m 3 - 2m 1 - m2 1.386
4U 1 - 212 - 21 3 .811
3U 2 - 211 - 13 .405
3U 3 - 211 - 12 2.238
The solutions to Eq. (5.77) are:
i Ii mi ui
1 0 .087 .4812
2 -.0063 0 .3228
3 .5634 .520 .7439
step 3. We shall take the exponentials of Ii' mi , and u i ' i 1,

2, 3, respectively, as:
1 1.000 1.0909 1. 6181
2 0.9937 1. 0000, 1.3810
3 1. 7566 1. 6820, 2.1040
with the exponential numbers, we can compute the fuzzy performance
scores r 11 using Eq.(5.66) as:
r 11 (A 1 exp(1 1 ), A2 exp(m 1 ), A3 exp(u 1 »
( .1959,.2890,.4310 )
349
The term A1 is calculated as:
3 -1
( 1: exp(u.» .1959,
i=l 1
Similarly, we can obtain A2 and A3 by replacing u i with mi and Ii'
respectively. The Aj , j = 1,2,3, are summarized as:
.1959 .2650 .2666
The fuzzy performance scores r 12 and r 13 may be determined in a manner
similar to that of r 11 . The r 1j , j = 1,2,3, can be summarized as:
.1959, .2890, .4310) (.1951, .2650, .3681) (.3441,.4457,.5609
Step 4. Steps 1 through 3 are applied to Tables 5.7, 5.8, 5.9,
and 5.10, respectively. The results are aggregated as:
Xl X2 X3 X4
A1 [(.196, .289, .431) (.405, .546, .714) ( .54, .579, .603) (.162, .25, .394)]
A2 (.195,.265,.368) (.162,.182,.204) (.163,.217,.292) (.313, .5,.763)
A3 (.344,.446,.561)(.277,.273, .34)(.158,.205,.267)(.209, .25,.305)
w [(.149,.194,.256),(.235,.319,.431),(.112,.140,.180),
(.263,.347,.451)]
This matrix is a fuzzy MADM problem. We can apply Eq. (5.67) to
calculate fuzzy utilities U1 (see Fig. 5.21) as:
3
1: w). r 1)' (.227, .398, .705)
j=l
350
p.(x)
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 ~~~~----~~~~--~----~--Tx
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
We can obtain U2 and UJ in a similar manner. They are summarized as:
(.227, .J98, .705) (168, .J1J, .579) (188, .289, .504)
Ra~king of Fuzzy utilities

The fuzzy utilities can be ranked by any appropriate fuzzy
ranking methodes) presented in Chapter 4. In this case, by simple
observation, we know U1 > U2 ~ UJ.
Note
Laarhoven and Pedrycz's approach suffers the following drawbacks.
1. There is not always a solution to Eqs.(5.62), (5.6J), and
(5.64). Indeed, as In(lijk) and In(u ijk ) are lower and upper values
of In(a ijk ) = -In (a jik ), the following must hold:
ln (u.~J·k) + ln (u.J~·k) 0, '" i,j,k.
Thus Eqs.(5.62) and (5.64) are linear dependent. The same holds true
351
for Eq. (5.63). There may not be any solution to Eqs.(5.62) through
(5.64) .
2. Even if there is a solution zi = (l i ,m i ,u i ) to simultaneous
Eqs. (5.62) through (5.64), zi may not be a triangular fuzzy number.
For example, in the numerical example presented in this section,
z3 = (.5634,.5200,.7439)
where 13 is no longer smaller than m3 . That is, z3 is no longer a
fuzzy triangular number. If so, the applicability of Eqs. (5.60) and
(5.61) (which are for multiplication and addition of fuzzy triangular
numbers) toward derivation of fuzzy utilities is in serious doubt.
3. The computational requirement for this approach is
tremendous, even for a problem as small as our numerical example.
This reduces the approach's applicability to MADM problems in which
there are more than 10 attributes and more than 10 alternatives. We
feel that it is not practical to attempt to solve any problems with
dimensions larger than the numerical example presented earlier.
5.3.3 Buckley's Approach
Buckley [B39] also extended Saaty's AHP method to incorporate
fuzzy comparison ratios a ij . He pointed out that Laarhoven and
Pedrycz's [L1] method is subject to two problems. First, the linear
equations of Eqs.(5.62), (5.63), and (5.64) do not always have a
unique solution. Secondly, they insist on obtaining triangular fuzzy
numbers for their weights. Since algebraic operations on triangular
fuzzy numbers do not necessarily produce a triangular fuzzy number,
Laarhoven and Pedrycz are forced to employ approximate methods to
preserve the shape of the fuzzy number.
To overcome these difficulties, Buckley uses the geometric mean
method to derive fuzzy weights and performance scores. This method is

352
used because it is easy to extend to the fuzzy case and guarantees a
unique solution to the reciprocal comparison matrix. Instead of using
a triangular fuzzy number, Buckley uses a trapezoidal fuzzy number
(a,b,c,d) (see Fig.5.22a) to represent the fuzzy ratio expressed by
the OMs. For example, in Fig. 5.22b (5,6,8,9) represents the fuzzy
ratio "between 6 to 1 and 8 to 1" and (4,5,5,6) represents the fuzzy
ratio "about 5 to 1". The fuzzy utilities, however, are not
restricted to trapezoidal shape. Buckley believed that his approach
avoids all the problems found in Laarhoven and Pedrycz's approach.
Geometric Mean Method

This weight assessing method is chosen for its simplicity and
ease in its application to the fuzzy case.
Given the positive comparison matrix as:
The geometric mean of each row is calculated as:
[
~ a .. ] lin
j=l 1.)
The weight wi is calculated as:

353
",(x)
1." ---------------~---"'"
Fig. 5.22a Trapezoidal fuzzy number (a,b,c,d).
",(x)
about 5 to 1
1.
.9
.8
.7
.6
.5-
.4
.3
.2
.1
00 2 3 4 5 6 7 8 9 10 X
Fig. 5.22b Examples of fuzzy ratios.
To facilitate the calculation of fuzzy weights, fuzzy performance
scores, and fuzzy utilities, the following arithmetic operations are
presented.
Fuzzy Arithmetic on Trapezoidal Fuzzy Numbers
Assume that the comparison ratios aij , Vi,j, take trapezoidal
fuzzy numbers (a .. ,b .. ,c .. ,d .. ). Their addition and multiplication

1J 1J 1J 1J
are defined in this section.
354
Let M1 = (a 1 , b 1 , c 1 , d 1 ) and M2
trapezoidal fuzzy numbers.
(1) Addition
Q (5.83)
~Q(X) is still a trapezoidal fuzzy number.
(2) Multiplication
Q (5.84)
where
~Q(x) is no longer a trapezoidal fuzzy number and is defined as
follows. For any unique x on the horizontal axis, ~Q(x) can be:
x ~Q(x)
:s a 0
d
'"
b:sx:sc
0
1
a:sx:sb ex E [0,1]
c:sx:sd ex E [0,1]
When a :s X :s b, x is defined as follows. Given xl
[a 2 ,b 2 ] where
1,2, (5.85)
355
the product x
x (5.86)
Similarly, when c ~ x ~ d, we can define
(5.87)
Fuzzy addition involving fuzzy numbers taking the form of
(ai[Li1,Li2],bi,Ci,di[Ri1,Ri2]) can be determined as follows.

Let two fuzzy numbers Q1 and Q2 be:
The addition of Q1 and Q2 is defined as:
Q' {{a 1 +a 2 ) [L11+L21,L12+L22]' (b 1 +b 2 ), (c 1 +c 2 ),
(d 1 +d 2 ) [R11+R21,R12+R22])} (5.88)
The membership function ~Q'{x) is defined as follows. Let x be some
real numbers on the horizontal axis. ~Q'(x) is defined as:
x ~Q' (x)
~ (a 1 +a 2 ) 0
z:: (d 1 +d 2 ) 0
(b 1 +b 2 ) :S x ~ (c 1 +c 2 ) 1
(a 1 +a 2 ) ~ x ~ (b 1 +b 2 ) a E [0,1]
(c 1 +c 2 ) ~ x ~ (c 1 +c 2 ) a E [0,1]
the addition x xl + x 2 will take the form
x (5.89)
x (5.90)
Algorithm
The algorithm may be applied to single or multiple OMs. A single
OM is assumed for the following steps. The case of multiple OMs shall
be explained in the Note section.
step 1. Consult the OM and obtain the comparison matrix A whose

-
elements a ij = (aij,bij,bij,dij)' Vi,j, are trapezoidal fuzzy numbers.
step 2. The fuzzy weights wi can be calculated as follows.
The geometric mean for each row is determined as:
(5.91)
where the sign 0 represents fuzzy multiplication. The fuzzy weight wi
is given as:
(5.92)
where the sign ® is for fuzzy addition.

357
The following will detail the derivation of fuzzy weight wi. Let
the left leg and right leg of a- ij be defined as:
n l/n
[ II «b" - a 1'J') 0: + a 1'J')] , 0: e [0,1], (5.93)
j=l 1J
n
g,(o:) = [ II «c .. - d 1,),) 0: + b, ,)]l/n, 0: e [0,1], (5.94)
1 j=l 1) 1)
respectively. Furthermore, let
n
[ II a .. ] l/n (5.95)
j=l 1)
and
m
a = L ai · (5.96)
i=l
similarly, we can define b i and b, c i and c, and d i and d. The fuzzy
weight wi is determined as:
b, c, di
1 1
c:,~, a ), vi, (5.97)
where the membership function g (x) is defind as follows. Let x be a

wi
real number on the horizontal axis. The gw, (x) can be summarized as:
1
x gw, (x)
1
~ (ai/d) ) 0
;,: (di/a) 0
[bi/c, ci/b] 1
[ai/d, bi/c] 0: e [0,1]
[ci/b, di/a] 0: e [0,1]

When x e [ai/d, bi/c], the x is calculated as:
x (5.98)
and when x e [Ci/b, di/a], the x is determined as:
x (5.99)
where
m
f(a) E fiCa) (5.100)
i=l
m
g(a) E gi(a) (5.101)
i=l
step 2 is repeated until the fuzzy performance scores r ij , Vi,j,

are obtained in a similar manner.
step 3. The fuzzy weights and fuzzy performance scores are
aggregated as a fuzzy MADM problem. The fuzzy utilities Ui , Vi, are
obtained based on
n
E wJ" r iJ", vi. (5.102)
j=l
The multiplication and addition of fuzzy numbers is done according to

the equations presented earlier.
Numerical Example (Buckley [B39])

An agency wants to rank chemicals A1 , A2 , and A3 from most
harmful to the environment to least harmful to the environment, with
respect to three criteria, xl = aquatic life, x 2 = agricultural impact,
and x3 = timber impact.
359
step 1. The expert estimates the relative weight ratios for each
pair of alternatives under every attribute as well as the relative
weight ratios for the attributes. This results in four reciprocal
comparison matrices:
Xl A1 A2 A3
1 1 1 1 1 1
A1 1 (4' 3' 3' 1)
2 (2' 2' 2' 1)
2
A2 (2, 3, 3, 4) 1 (1, 1, 2, 2)
1 1
A3 (2, 2, 2, 2) (2' 2' 1, 1) 1
X2 A1 A2 A3
A1 1 ( 6, 6, 6, 7) (2, 2, 4, 4)
1 1 1 1) 1 1
1 (2' 1, 1)
A2 (7' 6' 6' 6 2'
1 1 1 1)
A3 (4' 4' 2' 2 (1, 1, 2, 2) 1
X3 A1 A2 A3
A1 1 (1, 2, 2, 3) (7, 8, 8, 8)
1 1 1
A2 (3' 2' 2' 1) 1 (3, 3, 4, 4)
1 1 1 1 1 1
A3 (8' 8' 8' 1)
7 (4' 4' 3' 1)
3
1
Xl X2 X3
1 1 1 1 1 1
Xl 1 (7' 6' 6' 1)
5 (3' 2' 2' 1)
X2 (5, 6, 6, 7) 1 (3, 3, 3, 3)
1 1 1
X3 (1, 2, 2, 3) (3' 3' 3' 1)
3
1
360
step 2. For the first reciprocal matrix, the geometric mean is:
3
( II a .)1/3 )1/3 (1* 1* 1)1/3 .5.
a1 (all * a 12 * a 13 4 2
j=l 1J
and
3
a2 ( II a .)1/3 1. 2599
j=l 2J
3
a3 ( II a .)1/3 1. 0000.
j=l 3J
Hence,
3
a = E a. 2.7599
i=l 1
Similarly, we can get b i and b, c i and c, and d i and d. They are
summarized as:
i 1 2 3
ai .500 1. 2599 1.0000
bi .5503 1. 4422 1.0000
ci .5503 1.8171 1.2599
di .6300 2.0000 1.2599
and consequently, (a,b,c,d) = (2.7599, 2.9925, 3.6273, 3.8899).
The performance scores r 1j , vj, can be obtained as:
a1 b1 c1 d1
r 11 «(i' c
, "'1) , -a) (.1285, .1517, .1839, .2283),
a2 b2 c2 d2
r 21 «(i' c
, "'1)' -a) (.3239, .3976, .6072, .7247),
a3 b3 c3 d3
r 31 «(i, -C' "'1) , -a) (.2571, .2757, .4210, .4565) •
361
We repeat step 2 on the other reciprocal matrices one by one.
The results are the fuzzy numbers r i2 , r i3 , and wj , Vi,j.
steE 3. All the fuzzy numbers are aggregated as a fuzzy MADM
problem. I t is given as:
Xl X2
Al[
(.1285, .1517, .1839, .2283) (.4991, .5162, .8594, .9107 )
A2 (.3239, .3976, .6072, .7247) (.0904, .0985, .1640, .1650)
A3 (.2571, .2757, .4210, .4565) (.1373, .1420, .2980, .2999)
X3
(.3957, .6106, .6495, .8936) A1
(.2069, .2774, .3248, .4918) A2
(.-652, .0763, .0894, .1123) A3
[(.0834, .1111, .1111, .1660), (.5678, .6667, .6667, .7833),
W3
(.1596, .2222, .2222, .2839)]
We shall give the detailed computation of U1 to demonstrate the
computational procedure used in step 2. First of all, the product
form w1 r 11 is computed based on Eqs. (5.84) as:
{(a 1 a 2 )[L 1 ,L 2 ], b 1 b 2 , c 1 c 2 ' (d 1 d 2 )[R1 ,R2 ])}
{.0107[.00064,.00549],.0168,.0204,.0379[.00244,-.0199]}
where r 11 = (.1285,.1517,.1839,.2283) = (aI' b l , c l ' d 1 ),
W1 = (.0834, .1111, .1111, .1660) = (a 2 , b 2 , c 2 ' d 2 ),

362
Rl (d 1 - c 1 ) (d 2 - c 2 )
R2 - [d 2 (d 1 - c 1 ) + d 1 (d 2 - c 2 )]·
Similarly, we can obtain w2 r 12 and w3 r 13 . The products wj r 1j are
summarized as:
1 {.0107[.00064,.00549],.0168,.0204,.0379[.00244,-.01990]}
2 {.2834[.00169,.05907],.3441,.5730,.7134[.00598,-.14637]}
3 {.0632[.01345,.05907],.1357,.1443,.2537[.01506,-.12444]}
By summing up the three fuzzy numbers based on Eq.(5.88), we can

obtain U1 (see Fig. 5.23) as:
U1 = {.3573[.0158,.1236],.4966,.7377,1.0050[.0235,-.2907]}.
The' membership function value of ~u (x) may be summaried as:

1
x ~U (x)
1
s .3573 0
~ 1.005 0
.4966 s x s .7377 1
.3573 s x s .4966 a e [0,1]
.7377 s x s 1.005 a e [0,1]
When x e [.3573,.4966], it is defined as:

363
x (.01578)a 2 + (.12363)a + .3573;
and when x E [.7377,1.0050], it is defined as:
x (.02348)a 2 + (-.29071)a + 1.0050.
The fuzzy utilities U2 and U3 can be obtained in a similar
manner. They are also presented in Fig. 5.23 .
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 ~~~~--~~--~--~~--~--r-~x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Note
The proposed approach can be easily extended to incorporate
multiple experts in a MADM problem. We shall use an example (Buckley
[B36]) to demonstrate how it works in a group multiattribute decision
making case.
There are three alternatives, Ai' i = 1, 2, 3 to be ranked using

two attributes, x j , j = 1, 2, by five experts, J k , k = 1, ... , 5.
Each expert judges the relative importance of the attributes as well
as the relative preference of the alternatives. The comparison ratios
for two attributes by five experts are:

364
Jl J2 J3 J4 J5
[
Xl (7,8,8,9) (5,5,6,6) (6,7,7,8) (8,9,9,9) (5,5,5,5']
X2 (4,4,5,6) (7,7,7,7) (0,0,0,1) (2,2,3,3) (5,6,6,7)
For attribute Xl' the comparison ratios of the alternatives are:
Jl J2 J3 J4 J5
Al [ (5,5,5,5, (2,2,2,2) (7,8,9,9) (4,4,5,5) (0,0,0,"]
A2 (5,5,5,5) (7,8,9,9) (5,6,6,7) (9,9,9,9) (7,8,8,9)
A3 (6,7,7,8) (4,4,5,5) (0,1,2,3) (5,5,7,7) (2,3,4,4)
For attribute X2 , the comparison ratios of the alternatives are:
Al
A2
r Jl
4 '4'S'S'
(5,5,5,5)
J2
(5,5,6,6)
(5,6,6,7)
J3
(5,5,6,7)
(5,5,5,7)
J4
(4,4,6,6)
(9,9,9,9)
J5
(4,5,6,7']
(5,6,6,7)
A3 (6,7,7,7) (6,7,7,8) (8,8,8,8) (0,0,0,2) (7,7,7,7)
The problem is solved according to the following steps.

step 1. Average across the experts' judgment of the relative
importance of the attributes. They are summarized as:
(6.2, 6.8, 7.0, 7.4) (3.6, 3.8, 4.2, 4.8)
The calculation procedure is demonstrated by the following:
wl [(7,8,8,9) + (5,5,6,6) + ••• + (5,5,5,5)]/5
(6.2, 6.8, 7.0, 7.4).

365
Clearly, attribute Xl is more important than attribute X2 .
step 2. Average across the experts judgment of the comparison
ratio of the alternatives. Let r 11 be the average across the experts
for A1 with respect to Xl. It is computed as:
r 11 [(5,5,5,5) + (2,2,2,2) + ••• + (0,0,0,1)]/ 5
(3.6, 3.8, 4.2, 4.4).
similarly, all the r ij , vi,j, can be obtained. They are summarized
as:
i r il r i2
1 (3.6, 3.8, 4.2, 4.4) (4.4, 4.6, 5.8, 6.2)
2 (6.6, 7.2, 7.4, 7.8) (5.8, 6.2, 6.2, 7.0)
3 (3.4, 4.0, 5.0, 5.4) (5.4, 5.8, 5.8, 6.4)
step 3. The fuzzy utility of Ai' Ui , is computed as follows:
To insure the fuzzy weights have their support in [O,L], we have to
divide by L. Since the supports of wj , vj, are in [0,10], we
divide in this example by 10. Thus, the fuzzy utility Ui is
determined using
n Wj
[,r (--yo- ) riJ,]/n (5.103)
J=l
where n denotes the number of attributes. Ui may be seen as the
average weights across the attribute.
Base on Eg. (5.103), we compute U1 as:
{1.908 [.008,.25] 2.166, 2.688, 3.116 [.016,-.444]}.

The multiplication and addition of fuzzy numbers is the same as that
used in a single decision maker case.
similarly, we obtain U2 and U3 :
U2 {3.09 [.022,.514],3.626,3.892,4.566 [.032,-.706]}
U3 {2.026 [.022,.414],2.462,2.968,3.534 [.026,-.592]}.
Step 4. The ranking order of Ui , i 1,2,3, is determined as

follows. Let's define
(5.104)
which indicates the degree of Ai ~ Aj . On the other hand, e ji
indicates the degree of Aj ~ Ai.
We can conclude that Ai is better than Aj , if e ij =1 and e ji < U

where u e [0,1] is some real number predetermined by the OM. The u
value may be seen as a nonsignificant threshhold. It is recommended
by Buckley that values like 0.7, 0.8, 0.9 might be appropriate for u.
If e ij = 1 and e ji > u, Ai does not dominate Aj . When Ai does not
dominate Aj and Aj does not dominate Ai' we can conclude that Ai ~ Aj .

For a pair of alternatives Ai and Aj , either e ij = 1, or e ji = 1, or
e ij e ji = 1.
We shall calculate e 23 and e 32 to demonstrate the computational
procedure of Eq.(5.104). They are (see Fig. 5.24):
e 23 max [min (~U (x 3 ), ~U (x 2 »] 1

x2~x3 3 2
e 32 max [min (~U (x 2 ), ~U (x 3 )] .402.

x3~x2 2 3
367
JL(x)
1.
~~------~--~----------~r-~~x
2.968 3.080 3.534 3.828
Fig. 5.24 e 23 and e 32 .
We can obtain e ij and e ji , Vi,j, in a similar manner. They are
summarized as:
i\j 1 2 3
1 .112
e ..
1.J
2
3
[ 1
1 .402
Let's assume that ~ 0.9. The rank of the fuzzy utilities can
be determined based on
A1.. ~ A. if e .. 1 and e .. < ~. (5.105)

J 1.J J1.
since e 12 < ~ 0.9 < e 21 1, we can conclude that U2 is better than
u1 ; while e 32 < ~ and e 23 1, we know U2 is better than u3 · However,

we can only conclude that U1 -'" U3 because e 13 = e 31 = 1. Thus, the
ranking order is A2 > Al Eif A3 ·
368
Remark
To show that Saaty's AHP approach is a good method (at least as
good as its fuzzy method peers), we shall modify the numerical example
presented in section 5.3.3 (Buckley [B39])and solve it using Saaty's

AHP approach. Table 5.12 presents the modified data.
Table 5.12 Reciprocal Matrices
xl A1 A2 A3 x2 A1 A2 A3
A1 1 1/3 1/2 Al 1 6 3
A2 3 1 2 A2 1/6 1 1/2
A3 2 1/2 1 A3 1/3 2 1
X3 Al A2 A3 w xl x2 x3
Al 1 2 8 Xl 1 1/6 1/2
A2 1/2 1 3 x2 6 1 3
A3 1/8 1/3 1 x3 2 1/3 1
The eigenvalues and eigenvectors are given as:
3.0093 3.0000 3.0091 3.0000 ]

max
Xl X2 X3
0 Al
[.53 96 .6667 .0964 ~
~ [.UU]
A2
A3
.1634
.2969
.1111
.2222
.7006
.2029
] .6667
.2222
The composite vector for the candidates is given by:
.5153
u o x W [ .2479
.2262 1
369
The differences between A1 and A2 and A1 and A3 are sufficiently
large. We conclude that A1 is the best, and A2 and A3 are more or
less equivalent, with A2 having a slight edge over A3 •
The results are compared with the solution derived by using
Buckley's approach. Let the crisp utilities derived from saaty's
approach be represented by
.5153, z2 .2479, and z3 .2262.
By inserting zi' i = 1,2,3, into the fuzzy utilities obtained in

Section 5.3.3, we find that
1, i 1,2,3.
This phenomenon seems to indicate that the crisp utility, which is
only part of the fuzzy utility, may be as good as the fuzzy utility in
terms of discriminating among alternatives. If so, Saaty's AHP
approach is better than its fuzzy method peers because it requires
less computations.
2. In 1983, Wagenknecht and Hartmann [W1] proposed a fuzzy weight
assessing method. The fuzzy weights are calculated using the
geometric mean. But the fuzzy ratio is represented by a special
type of membership function (instead of a triangular or a trapezoidal
fuzzy number)
xO: ij
k ..
1J
, x > 0
{x + d ij ) (3 1J
..
1
/.la .• (x) (5.106)
1J 0 , X :s 0
370
where 0 :s CI. •• :s f3 •. :s co, d .. ~ 0 for adjustment and k .. for

1) 1) 1) 1)
normalization. Since the fuzzy number used is much more complicated
than a trapezoidal fuzzy number, this approach ends up with a very
involved computation in order to get wj . We do not recommend this
approach for any size problem.
Since Wagenknecht and Hartmann's approach is even more difficult

to use than Laarhoven and Pedrycz's and Buckley's approaches, we will
not discuss it here. Interested readers should refer to the original
paper for details.

371
5.4 Fuzzy Conjunctive/Disjunctive Method

The classical conjunctive method is an intuitive approach used to
dichotomize alternatives into acceptable/not acceptable categories
(Hwang and Yoon [H13]). The OM sets up the minimum attribute values
(standard levels) s/he will accept for each of the attributes. Any
alternative which has an attribute value less than the standard level
will be rejected (not acceptable).
The classical disjunctive method, on the other hand, is one in
which an alternative is evaluated based on its greatest value of an
attribute. For example, an NFL team picks its kicker according to
this method. The player is selected simply because he kicks extremely
well, regardless of his passing and running skills.
Dubois et al. [037] pointed out that when data in a decision
problem are fuzzy, the match between standard levels provided by
the OM and attribute values becomes vague and, naturally, a matter of
degree. The degree of matching is computed using the possibility
measure and the necessity measure. The alternative that has the
highest degree of matching is considered the best.
5.4.1 Dubois, Prade, and Testemale's Approach

Dubois et al. [037] proposed the fuzzy version of the conjunctive
and disjunctive methods. They pointed out that when data in a
decision matrix and the OM's standard levels are fuzzy, the matching
between these two fuzzy data becomes vague and, naturally, a matter of
degree. The degree of matching is measured by the following member-
ship function:
a }, Va. (5.107)
where ITQ(X) represents the degree of possibility that x is the (unique)

372
value which describes an object modeled by Qi ~p(x) is the degree of
compatibility between the value x and the meaning of P. Thus, ~PIQ(a)
denotes the degree of compatibility of Q with respect to P. Eq.(5.107)
was first introduced by Zadeh [Z14] who interpreted ~PIQ(a) as the

fuzzy truth value of predicate P, given a referential predicate Q
describing a true state of facts.
The derivation of ~pIQ is best explained by the following

example in which P and Q are trapezoidal numbers (see Figs. 5.25a and
5.25b). set a = o. The x values that satisfy ~p(x) = 0.0 are 7 and
28 (see Fig. 5.25a). Consequently, ITQ (X=7) = .20 and ITQ (X=28) 0.0.
~PIQ(a=o.o) = sup [.20, 0.0] = .20 based on Eq. (5.107) (see Fig.
5.25b). By setting different a levels, we can obtain ~PIQ(a) as
denoted by the solid line and the point (1,1) on Fig. 5.25b. Note
that the dashed line from point (0,0) to (1,1) on Fig. 5.25b indicates
a perfect match between P and Q.
Although the interpretation is clear from a theoretical point of
view, Dubois et al. believed that Eq.(5.107) is not easily understood
by users, and difficult to manipulate at an operational level. As a
consequence, two scalar indices are used to approximate the ~PIQ(a)
",(x)
1.
.9
.8
.7
.6
.5
~
.4
. .'
~
~
.3 ~\
~\~
.2 ·-------I--
I-
.1 .' -
l:
o 0~-5~-1...0------~--- x
15 20 25 30
Fig. 5.25a Trapezoidal fuzzy numbers P and Q.

373
/L. «~)
Pia (11)
1. ---- ---- ------- --- ------------ --- --- -------------------;:,
.8 .... _______________________________________ ::;,6: _____
.9 _- ----- i.
.7 ;"i'
.6
.5
.4
.3
.2
.1
0
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1. a
Fig. S.2Sb The degree of matching ~PIQ(a).
measure so that compatibility between fuzzy sets can be estimated.
The two indices are (1) the possibility of matching IT(P;Q), and (2)
the necessity of matching N(P;Q).
Possibility and Necessity of Matching
The possibility of matching is defined as:
IT(P;Q) sup min (~p(x), ~Q(X» (S.108)

xeU
which estimates to what extent it is possible that P and Q refer to
the same x value. In other words, the possibility of matching is the
degree of overlapping of the fuzzy set of values compatible with P,
with the fuzzy set of possible values of Q.
The necessity of matching is defined as:
N(P;Q) inf max ( ~p(X), l-~Q(X» (S.109)

xeU
which estimates to what extent it is certain that the value to which Q
refers is among the ones compatible with P. In other words, the

374
necessity of matching is the degree of inclusion of the set of
possible values of Q into the set of values compatible with P.
The necessity of an event corresponds to the impossibility of the
opposite event, i.e.,
N(PjQ) 1 - II(PjQ) (5.110)
where ~p(X) = 1 - ~p(X) is the membership function of the complement

of the fuzzy set of values compatible with P. Clearly, we always have
II(PjQ) ~ N(PjQ) . (5.111)
Generally, if Q is a crisp number, then II(PjQ) = N(PjQ) = ~p(Q)
which is also a crisp number in [0,1]. When both P and Q are fuzzy,
then the following relation holds (given ~ is a modal where
N(PjQ) s ~ S II(PjQ). (5.112)
This relation is constructed based on the following equations:
II(PjQ) sup min (X,/.lPIQ(X» ~ min (~,/.lPIQ(~» ~,
x
(5.113)
and
N(PjQ) inf max (X,/.lPIQ(X» S max (~,1 - /.lPIQ(~» = ~.
x
(5.114 )
Hence [N(PjQ),II(PjQ)] is a bracketing of ~, which provides information
about the imprecision of /.lp(Q). We may conclude this section by
stating that II(PjQ) and N(PjQ) together are reasonable approximations
of /.lp(Q). Fig. 5.26 demonstrates the II and N indices.

375
,.,.(x)
1.
.9
.8 U(~!~L. __ _
.7
.6
.5
.4
.3
Fig. S.26 The IT(PjQ) and N(PjQ) indices.
Eqs.(S.107) and (S.108) can only be used in single-attribute

conditions. When there are multiple attributes involved, as in most
real world problems, Eqs.(S.107) and (S.108) can be modified using the
min operator:
min (S .11S)
j=l, ... ,n
min (S.116)
j=l, ... ,n
o 0 0 0 0
where A = (Xl'···' Xn ), Ai = (X i1 ,···, x in ), and Xj and Xij are
defined on the same domain U. The vector AO is the cutoff vector
specified by the DM, while Ai' i = 1, ••• , m, is the vector that
contains the performance scores of the ith alternative under all
attributes.
Eqs.(S.llS) and (S.116) suggest that the matching is done
attribute by attribute. These matching results are to be aggregated
using the min operator to preserve the respective semantics of
possibility and necessity of the indices. Eqs.(S.llS) and (S.116)
376
implicitly suggest that all attributes are of equal importance. If
unequal weights are used, one of the following formulas can be used.
S min max ( 1 - wj ' Sj) (for conjunctive case) (5.117)

j
or
S max min (wi' Sj) (for disjunctive case) (5.118)

j
where Wj denotes the relative importance of the attributes and
max Wj 1, (5.119)
j=l, ... ,n
i.e., the most important attributes are rated 1i and S expresses to
what extent we are certain that the fuzzy set of importance is includ-
ed in the fuzzy set of the requirements xj possibly (or necessarily)
satisfied by the performance score x ij defined by the equation
0
Sj II(Xj,X ij ), j 1, ... ,n, (5.120)
or
Sj
0
N(Xj,X ij ) , j 1, ... , n. (5.121)
Thus, for the conjunctive case, Eq.(5.117) may be rewritten (given
different Sj'S) as the aggregated II and N indices:
min max ( 1-W j ,II(Xj0 iXij» 0

II(A iAi)' (5.122)
j
0 0
min max ( 1-Wj,N(XjiXij» N(A iAi). (5.123)
j
For the disjunctive case, Eq. (5.118) may be rewritten as the
aggregated IT and N indices:

377
max min (5.124)

j
max min N(A o.,Ai ) (5.125)

j
The reason for constructing Eqs.(5.117) and (5.118) is as
follows. In the case of aggregation via the arithmetic mean, import-
ance can easily be accommodated in the aggregation (given P 1 ' ... , P n
fuzzy sets) through
n
L wJ' I-Lp, (x), "</ X e U. (5.126)
j=l J
Note that Eq. (5.126) does not have the intersection or union
operations. In order to get a weighted counterpart of the minimum and
maximum operation, we view Eq.(5.126) as the probability of a fuzzy
event ~x defined on the crisp set { 1, ... , n}. ~x is the fuzzy set
of Pj's containing x, and the wj's define a probability allocation on
the crisp set { 1, ... , n}. Hence
(5.127)
in the sense of Zadeh [Z2]. Changing probability into a possibility
or a necessity measure leads us to consider the following analogs of
Eq. (5.126), namely,
max min (5.128)

j
min max (5.129)

j
378
It is clear that Eq.(5.128) yields a weighted disjunction of the P.'s

l.
while Eq.(5.129) yields a weighted conjunction. Thus Eqs.(5.117) and

(5.118) are the direct applications of Eqs.(5.122) and (5.123), and
(5.124) and (5.125), respectively.
Generally, the ideal ordering is used to rank the alternatives
according to the following:
If Then
[II(Ao;A i ) > II(Ao;A j ) and N(Ao;A i ) ~ N(Ao;A.)] Ai > Aj (5.130)

]
[II(Ao;A i ) l!: II(A0 ;A j ) and N(Ao;A i ) > N(Ao;A j )] Ai > Aj (5.131)
[II(Ao;A i ) - N(Ao;A i ) < II(Ao;A j ) - N( Ao.,Aj ) Ai > Aj (5.132)
However, when the stated rule is not followed, we can't say that Ai >
Aj nor Aj > Ai. Also note that the necessity index, N, is more
important than the possibility index, II, because when the N index is
positive we can be certain that the alternative (more or less) matches
the requirements set by the OM.
Numerical Example--for conjunctive case (0[37])

This example deals with (more or less approximate) descriptions
of second-hand cars for sale in a garage. The different attributes
involved are the age of the car, the purchase price, the gas consump-
tion and the speed of the car. The available data is shown in Table
5.13. The linguistic terms are represented by L-R type trapezoidal
fuzzy numbers (a,b,ex,~) (see Fig. 5.27). For example, the term "new
age" is represented by the fuzzy number (0,1,0,1), where a = 0, b = 1,
ex = 0, and ~ = 1. ThUS, the fuzzy number labeled as "new age" in
Fig. 5.28a is easily understood. The fuzzy sets that represent the
lingustic terms in Table 5.13 are summarized below:
379
Table 5.13 Selection of Used Car
Age of the Purchase Consumption Maximum

Car Price of Gas Speed
A1 new expensive economical rather-fast
A2 less-than-3 around- rather- 180-200

yrs old 4500 economical
A3 very-recent between- heavy fast
50000 - 60000
A4 around-5 less-than- 8-9 180-200
yrs 20000
A5 5-10 around-10000 heavy rather fast
A6 old cheap economical not-very-fast
A7 new 32000-40000 very between-

economical 140 and 160
p.<x)
1.
Fig. 5.27 Trapezoidal fuzzy number (a,b,a,~).
1) Attribute: age-of-the-car (Fig. 5.28a); domain [0,20].
new less-than-3 very-recent around-5 old
(0,1,0,1) (0,3,0,0) (1,2, ,1,1) (5,5,1,1) (10,15,2,5)

380
.7
.6
.5
.4
.3
.2
.1
0
0 2 4 6 8 10 12 14 16 18 20 X
Fig. 5.2Sa The fuzzy set of the age of used car.
2) Attribute: selling-price (Fig. 5.2Sb); domain: [5000,100000].
cheap around-10000 around-45000

(5000,10000,0,5000) (9000,11000,1000,1000) (44000,46000,1000,1000)
between 50000-60000 moderate expensive

(50000,60000,5000,5000) (35000,60000,5000,5000) (SOOOO,100000,5000,0)
c~eap around 45000

P.(x) 1 around 10000 ~eW<f8~60000
1.: : ~
.9 : ~
: ~
.8
.7
:~
, \
,, \\
,l
.6 :~
.5 ":~
,I
.4 :~
.3
.2 f
l
expensiv
.1 §~
°0~~~2~~3~~4~~5--~6~~7~~8--~9--+10X
Fig. 5.2Sb The fuzzy set of used car price.

381
3) Attribute: fuel consumption (Fig.5. 28c); domain: [5,15].
very economy economy rather-economy heavy
(5,6,0,0.5) (6,7,0.5,1) (7,S,1,1) (9,15,0.5,0)
very economy rather economy
~.(X) \~~:.~~_ _ _--,

.9 I \ s; heavy
.8 ~ : ~
~: ~
.7 ,.
~,
, I
~ : f
.6 ~ :
~ .
.5
~ ~
.4
.3 ~
.2 ~
~
.1 ~
o ~~--~~~~~~~--~~--~~x
~
o 2 4 6 8 10 12 14 16 18 20
Fig. 5.2Sc The fuzzy set of used car fuel consumption.
4) Attribute: maximum speed (mph) (Fig. 5.2Sd); domain [100,250].
not-very-fast between-140-and-160 rather-fast fast
(120,140,10,10) (140,160,20,20) (150,1S0,20,20) (1S0,200,20,20)
between 140-160
/L(x)not very fast ~rather fast
1. \ .. '" 'I""""" fast
.9 : l"
. \ ~~ /
8 ! ~ : ~
• : ~ ". ~$
.7 : \ :, ~
.6 ! ': ~
" • $
.5 ,: ~
.4 ." ~\
.3
,2 \ ~
.1 \
~
Ol--...........L...,i-i-..,....I.....I-...;....-+--I---.--....--..... x
o 100 120140 160180200220240260
Fig. 5.2Sd The fuzzy set of used car maximum speed (mph).
382
Let's assume that the cutoff vector is given as:
o 0 0 0
«xl' 0.8), (X 2 ' 0.5), (x 3 , 1), (x 4 , 0.2»
where the numbers 0.8, 0.5, 1, and 0.2 are weights associated with
each attribute, and the xj, j =1,2,3,4, are summarized as:
new age moderate price very-economy fast
For the first alternative, we find that the possibility of

matching between x~ and x 11 is (see Fig. 5.28a):
rr(x~;x11) = sup min (~xo (u), ~x (u» 1

ue[0,20] 1 11
sup min (~xo(u), ~x (u» o

ue[5000,100000] 2 12
sup min (~ o(u), ~x (u» 1

ue[5,15] x3 13
sup min (~xo(u), ~x (u» 1

ue[100,250] 4 14
o
The necessity of matching between Xl and x 11 is (see Fig. 5.29):
inf max (~xo(u), 1-~ (u» .50

ue[0,20] 1 x11
inf max (~XO(u), 1-~ (u» 0,

ue[5000,100000] 2 x 12
383
P.(x)
1.
.9
.8
.7
.6
.4 (
(
.3 ~
.2 ~~
.1 ~
o O~~-2~~4--~6--~8---1~O--1~2--1~4--~16--~18--~20X
Fig. 5.29 The derivation of N(X~;X11)'
inf max (~XO(u), 1-~ (u» 0,

ue[5,15] 3 x 13
inf max (~xo(u), 1-~ (u» o.

ue[100,250] 4 x 14
Applying Eqs. (5.122) and (5.123), we compute
IT(A O; A1 ) ~_min max (l-W j , IT(xj,X ij »)

J-1, ... ,4
= min [«1-0.8) V 1), «1-0.5) V 0), «1-1) V 1),

(1-0.2) V 1)]
min [1" 0. 5, 1, 1]
0.5,
and
min max (l-W j , N(X oj , Xij »
j=l, ... ,4
= min [«1-0.8) V 0.5), «1-0.5) V 0), «1-1) V 0),
«1-0.2) V 0)]
min [0. 5, O. 5 , 0, O. 8 ]
o.
384
Similarly, we can obtain the possibility measure and necessity

measure for other alternatives. They are summarized as:
i 1 2 3 4 5 6 7
Il(A O ; Ai) .5 .3 o o o .2 .8
N(Ao ; A.) o o o o o o .5
Observing all the possibility and necessity indices values, we find

that A7 , whose II and N values are the maximum among all possibility
measures and necessity measures, respectively, is the one that matches
AO the best. But A1 , A2 , A6 are more or less acceptable (given the
OM's standard, AO ).
Note
This method utilizes the min operator to combine the
single-attribute index values for the conjunctive case (Eq.(S.117»
and utilizes the 'max' operator to combine the single-attribute index
values for the disjunctive case (Eq.(S.118». In both cases, only a
small portion of the available information is used in deriving the
final conclusions. That is, the matching of one alternative over
another with a cutoff vector is determined solely on one attribute.
Thus, even if an alternative matches fairly well on all but one
attribute, another alternative that matches about average on all
attributes would be chosen over it (for the conjunctive case).
Similar remarks apply to the disjunctive case.
The applicability of this method is rather limited. In general,
it would only be reasonable to use this method if the OM is assumed to
have a pessimistic nature in the decision process (for Eq. (5.117»,
while Eq. (5.118) should only be used when the OM is assumed to have an
optimistic stance toward the decision situation.
385
5.5 Heuristic MAUF Approach
Efstathiou and Rajkovic [E3] and Efstathiou [E2,E4] argued that
the Multiple Attribute utility Function (MAUF) cannot be practically
obtained by the combination of single attribute utility functions
because of the dependency among attributes. Therefore, a heuristic
approach is needed to define the MAUF. Since decision data may be
numerically and/or linguistically expressed, fuzzy set theory must

be incorporated in this heuristic approach. The utility function is
represented in the "IF .•• THEN ... " decision rule format.
Algorithm
The algorithm may be summarized by the following steps.
Step 1. Identify interest groups of people involved in the
decision environment.
Step 2. Identify attributes (X j ) and establish a universe of
discourse for each attribute.
Step 3. Interact with the OMs to identify their heuristics and
identify the ideal solutions for each interest group.
Step 4. Construct heuristic decision rules for each group. The
rules are constructed according to the heuristics obtained in Step 3.
Step 5. List all alternatives and measure along the attributes.
step 6. Calculate the utilities of alternatives for each group.
Step 7. Rank alternatives on the basis of calculated utilities.
The OMs are actively involved in the decision process in this
method. For example, in Step 3 heuristic information is obtained
through a question-and-answer system, where dialogue between the DMs
and analysts takes place; and in Step 4, where the decision rules are
constructed through serious discussions between the OMs and analysts.

386
Numerical Example (modified from Efstathiou [E41)

Four typewriters, Al , A2 , A3 , and A4 , are to be evaluated with
respect to three attributes: price (Xl)' downtime (X 2 ), and special
features (X 3 ). Let's assume there is only one interest group
involved--office typists). since the interest group and attributes
are all identified, we shall begin the decision process with step 2.
step 2. Establish the universes of discourse for attributes and

utility. Through careful discussion, the universes of discourse in
use are:
Special
Price Downtime Features UtilitX
intolerable
::s $400 low poor poor
sa! $600 moderate medium
$800 high good good
'"
very good very good
ideal
step 3. Identify the heuristics and ideal solutions for the

interest group. Assume the OMs reveal their heuristics as follows.
a. If the down time is "high" and the special features are at
least "good," the price is a very important factor in determining
utilities.
b. If both the features and downtime are good, the price is not
an important factor.
c. If the special features are poor, the price is over $600,
and the downtime is high, the machine is not acceptable.
d. If the machine is over $800 and has moderate downtime, it is
not acceptable.
e. The machine should be fairly cheap and have a low downtime.
In addition, the OMs' ideal machine is described as having "low price,
387
low downtime, and very good special features."

step 4. Obtain the decision rules
The rules may take the form of:
IF THEN
Price Downtime special Features utility
:s $400 moderate good good
This is interpreted as: if the price is :s $400, the downtime is
moderate, and the special features are good, then the utility is
"good." After serious discussions with the DMs, we construct the
decision rules summarized in Table 5.14.
Step 5. List all the alternatives. The alternatives in the
decision matrix can be written as:
Xl X2 X3
D A1 $520 moderate poor
[
A2 $455 moderate-high very good
A3 $430 low good
A4 $875 moderate good 1
where the term "moderate-high" may be interpreted as a fuzzy set
{ ( . 5/moderate), (. 5/high) } .
step 6. Obtain the utility for each alternative. Normally, we
can simply match the performance data with the decision rules in
Table 5.14 and determine the fuzzy utility of each alternative. For
example, because one of the rules in Table 5.14 is matched, the fuzzy
utility of A4 is "very poor". The matched rule is:
IF THEN
Price Downtime Special Features Utility
~ $800 moderate good very poor
388
Table 5.14. The Heuristic Decision Rules
Price Downtime SEecial Features utility

:S $400 low poor good
good good
very good ideal
moderate poor medium
good good
very good very good
high poor poor
good good
very good good
- $600 low poor medium
good medium-good
very good good
moderate poor medium-poor
good medium
very good medium
high poor intolerable
good poor
very good poor
2:: $800 low poor poor
good medium
very good medium-good
moderate poor very poor
good very poor
very good poor
high poor intolerable
good intolerable
very good poor
However, for Ai' i = 1,2,3, there is no appropriate match between
the alternative performance data and the decision rule. In order to
obtain fuzzy utilities, Ui ' i = 1,2,3, some additional rules are
needed:
389
a. When the price, p (crisp number), is between $400 and $600,

p can be rewritten as the fuzzy set
{« 600 - p )/s $400),

200
« p - 400 )/~ $600)}.
200
b. When the price, p (crisp number), is between $600 and $800,
p can be rewritten as the fuzzy set
{ « p - 600 ) /~ $600), « 800 - P ) /~ $800)}.

200 200
with the additional rules, the fuzzy utility of Al can be
determined as follows. Since the price of Al is $520 e [$400,$600],

it can be represented as the fuzzy set
or
{(.4/S $400), (.6)/~ $600)}.
That is, Al can now be represented as
Al {.4/[S $400, moderate, poor], .6/[~ $600, moderate, poor]}.
By matching the elements [S $400, moderate, poor] and [~ $600,
moderate, poor] with the decision rules in Table 5.14, we obtain fuzzy
utilities "medium" and "medium-poor", respectively. By combining the
two fuzzy utilities, we obtain the fuzzy utility of AI:
U1 {(.4/medium), (.6/medium-poor)}.
The determination of U2 is even more invloved. The price ($455)
and the downtime (moderate-high) of A2 are interpreted as the fuzzy

390
sets {(.725/~ $400), (.275/f!!! $600)} and {(.5/moderate), (.5/high)},
respectively. Note that the membership values .725 and 275 are
obtained as we did in deriving the membership values of the price of
A1 • The numbers .5 and .5 are assigned arbitrarily. A2 can then be
represented as a cartesian product:
A2 {(.725/~$400),(.275/f!!!$600)} x {(.5/moderate), (.5/high)} x
{(l/very good)}
We then apply the min-operator to determine the composite membership

values and obtain the fuzzy utility:
U2 = {(.5/[~$400,moderate,very good]), (.5/[~$400,high,very good),
(.275/[f!!!$600,moderate,very good]), (.275/[f!!!$600,high,very good)}
= {(.5/very good), (.5/good), (.275/medium), (.275/poor)}.
For example, the membership value of the element [~$400,moderate,very
good] is
min [.725,.5,1] .5.
U3 can be obtained in the same manner as u 1 . The four utilities
are summarized as:
1 {(.4/medium),(.6/medium-poor)}
2 {(.5/very good), (.5/good),(.275/medium), (.275/poor)}
3 {(.85/good),(.15/medium-good)}
4 {(l/very poor)}
391
Note that these fuzzy utilities are fuzzy sets which contain
linguistic terms and the real numbers e [0,1] associated with each
linguistic term may be treated as the membership value. The member-
ship values of a fuzzy set (e.g., fuzzy utility) need not sum to one.
step 7. Rank the fuzzy utilities. Let's assume the meaning of
the atomic linguistic terms {good, medium good, medium, medium poor,
poor} can be shown in Fig. 5.30 or expressed in discrete membership
functions as:
x 0 .1 .2 .25 .3 .4 .5 .6 .7 .75 .8 .9 1
good 0 .16 .3 .7 1
med.goOd 0 .3 .7 1 .7 .3 0
medium 0 .5 1 .5 0
med.poor 0 .3 .7 1 .7 .3 0
poor 1 .7 .3 .16 0
,.,.(x)
med.poor medium med.good
1.
. 9-
.8
.7
.6-
.5
.4-
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 5.30 The meaning of linguistic terms -- continuous case.
The term "very good" which appears in U2 can be determined as follows

(Zadeh [Z5,Z6]):
392
2
Ilvery good(X) {Ilgood (x)} •
This results in a fuzzy set
.7 .75 .8 .9 1
very
Xgoodl 0 .0256 .09 .49 1
If these discrete membership functions are used, the fuzzy

utilities Ui' i = 1,2,3,4, may be approximated as shown in Fig. 5.31.
For example, to graph U1 , the following process is used. By using
the min operator, we can obtain
x o .1 .2 .25 .3 .4 .5 .6 .7
(I) 0.4/medium o .4 .4 .4 o
(II) 0.6/med. poor o .3 .6 .6 .6 .3 o
For example, when x .4, we can obtain
.4/medium min [.4, Ilmed ium(.4)] min [.4,.5] .4
.6/med.poor min [.6, Ilmed.poor(.4)] min [ .6, .3] .3.
Using the max operator to combine (I) and (II), we can obtain U1 as:
o .1 .2 .25 .3 .4 .5 .6 .7
o .3 .6 .6 .6 .4 .4 .4 o
1
which can be approximated as a continuous function in Fig. 5.31.

utilities Ui' i =2,3,4, can be obtained in a similar manner (see
Fig. 5.31).
393
x 0 .1 .2 .25 .3 .4 .5 .6 .7 .75 .8 .9 1
/.lU (x) .275 .275 .275 .16 0 .275 .275 .275 0 .16 .3 .5 .5
2
/.lU (x) 0 .15 .15 .16 .3 .7 .85
3
/.lU (x) 1 .49 .09 .0256 0
4
JL(X)
1.
.9
.8 ~,4
.7 "
.6
.5
.4
.3
.2
.1
0 ~-T--~~~-T--~--r-~~~--r-~x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Observing Fig. 5.31, it seems that U2 and U3 are better than U1
and U4 . But it is not so clear if U2 is better than U3 or the other
way around. More criteria may be needed to distinguish between U2 and
U3 . The solution to this numerical example is: {A 2 ,A 3 } > A1 > A4 .
Note
1. In step 6, when an alternative takes some composite linguistic
terms such as "moderate-high," "between $400 and $600," etc. as its
performance data, this alternative must be represented as the
cartesian product of fuzzy sets on Xj , vj. Assume that A. has some

1
data items which are represented by composite linguistic terms.
Alternative Ai can be represented as:
(5.133)
394
where F(X j ) indicates the performance score of Ai with respect to Xj .
For example, A2 of the numerical example in this section is
A2 [$455, moderate-high, very good].
Its cartesian product is
A2 {(.725/~$400), (.275/~$600)} x {(.5/moderate), (.5/high)} x
{(l/very good)}.
The utility of alternative Ai' Ui , is also a fuzzy set whose

membership function is defined as:
JiA (x..
(n) ) min (Ji A (x .. » (5.134)
i 1) j=l, ... ,n i 1)
where xij(n) denotes an n-dimensional vector. For example, U2 is
{(.5/very good), (.5/good), (.275/medium), (.275/poor)} where the real
numbers (such as .5, .275) are obtained using Eq. (5.134).
2. The major disadvantages of this approach are: (a) It is
extremely time consuming to construct the decision rules, even with
the help of the OM's heuristics. For a small problem like the
presented numerical example, it may be okay to use this method.
However, when the problem size increases to 10 attributes, each having
four values in its universe of discourse, there will be a total of 4 10
= 1,048,576 decision rules, and even with the help of the OM's
heuristics, the number of rules is still high. That makes the
algorithm impractical for application to large real-world problems.
(b) The involvement of the OM is tremendous. The discussion between
the OMs and analysts is time consuming. When the OMs are not
available for consultation, this approach is not applicable.

395
5.6 Negi's Approach
Negi [N7] believes that the MADM problem containing fuzzy data
can be solved using classical MADM algorithm in which fuzzy data are
assumed to be trapezoidal numbers and are handled by fuzzy
mathematics. The method requires no new algorithm since numerous
classical MADM algorithms are readily available (Hwang and ¥oon
[H13]). In addition, the proposed approach is very general and can be
applied to any type of MADM problem involving fuzzy data. Negi left
it to the system analysts to decide which classical MADM algorithm
should be used.
We shall use one of the well known classical MADM methods, TOPSIS
(Hwang and ¥oon [H13]), to illustrate the proposed approach. It is
used to solve a problem described as follows. Given a decision
matrix, D, of m x n dimension:
Xl X. X
) n
D A1 x 11 x 1j x 1n
A. Xu x .. X.
1 1) 1n
Am x x x
m1 mj mn
where x ij ' Vi, j may be crisp or fuzzy. If x ij is fuzzy, it is
represented by a trapezoidal number as x ij (a ij , b ij , c ij ' d ij )·

The fuzzy weights shall be described by Wj (a j , ~j' 7j' OJ) (see
Fig. 5.32).
Algorithm
The problem is solved using the following steps.
step 1. Normalize the Decision Matrix. The decision matrix must
first be normalized so that the elements

396
P.(x) xij Wj
1.
.9
.8
.7
.6-
.5
.4
.3
.2
.1
0 x
Fig. 5.32 Trapezoidal fuzzy numbers used in Negi's method.
are unit-free. To avoid the complicated normalization formula used in
classical TOPSIS, we use linear scale transformation as follows:
r x / x * , vj, X.J is a benefit attribute

ij ij j
{ r ij -
Xj / x ij , vj, X. is a cost attribute
J
(5.135)
By applying Eq. (5.135) , we can rewrite the decision matrix as:
Xl Xj X
n
D' A1 r 11 r 1j r 1n
Ai r i1 r ij rin
A r m1 r rmn
m mj
When x ij is crisp, its corresponding r ij must be crisp; while when x ij
is fuzzy, its corresponding r ij must be fuzzy. Eq.{5.135) is then
replaced by the following fuzzy operations: Let Xij = (a ij , b ij , c ij '
d ij ) and Xj* = (a *j , b *
j , c*
j , d*
j ), we have
397
a .. b ij c ij d ..
r ij Xij (+) Xj* (21. , -¥) (5.136a)
d*
j ' c *j b*
j ' aj
- -
- a.
(~, c
bj , c. d:
r ij Xj (+) x ij
~J
ij b~j' ....l.-)
a ij (5.136b)
step 2. Obtain the Weighted Normalized Decision Matrix. This
matrix is obtained using
(5.137)
When both r ij and w. are crisp, v ij is crisp; while when either r ij or

J
wj , (or both) are fuzzy, Eq. (5.137) may be replaced by the following
fuzzy operations:
a .. b ij c ij d ..
v ij r ij ( . ) Wj (-¥ aj , ~j' T.,-¥ .s.) (5.138)
dj c *j b*
j J a. J
J
-
a. - d:
V ij r ij (.) Wj (....l.- a j ,
d ij
bj
c ij ~j' b ij
l T.,.:.:.L
J a
.s.)
J
(5.139)
ij
Eq.(5.138) is used when the jth attribute is a benefit attribute.
Eq.(5.139) is used when the jth attribute is a cost attribute. The
result of Eqs(5.138) and (5.139) can be summarized as:
v
398
step 3. Obtain the positive Ideal Solution (PIS), A* , and the

Negative Ideal Solution (HIS), A-. PIS and HIS are defined as:
A * *
[VI' ... , Vn* ) ,
A - [V~, ... , Vn ) ,
where Vj* = max v ij and Vj- min v ij •

i i
For crisp data, Vj* and Vj are obtained straight forward. For
fuzzy data, Vj* and Vj- may be obtained through some ranking procedures
as those described in Chapter 4. Negi used Lee and Li's [L3) ranking
method (see section 4.6) for comparison of fuzzy numbers. The Vj* and
Vj- are the fuzzy numbers with the largest generalized mean and the
smallest generalized mean, respectively. The generalized mean for
fuzzy number v ij ' Vi,j, is defined as:
(5.140)
For each column j, we find a v ij which has the greatest mean as the Vj*
and a v ij which has the lowest mean as the v j .
Step 4. Obtain the Separation Measures Si* and si. In the

classical case, separation measures are defined as:
n
Si* r * i 1, ..., m• (5.141)
j=1 °ij'
and
n
si r °ij' i 1, ... , m• (5.142)
j=1
399
* and Dij are given as:

For crisp data, the difference measures Dij
* v .. - *
Dij 1.) Vj
Di j v .. - Vj
1.)
The computation is straight forward. For fuzzy data, Negi defined the
difference between two fuzzy numbers g (x) and g *(x) (based on Zadeh
v ij Vj
[Zl]) as:
1 - { sup [gv .. (x) 1\ g *(x) ]} 1 - L ij , Vi,j, (5.143)

x 1.) V j
where Lij is the highest degree of similarity of v ij and v *

j. The
value of Lij is best depicted in Fig. 5.33.
Similarly, the difference between g (x) and g -(x) is defined

v ij Vj
as:
1 - {sup [gv .. (x) 1\ gv:(x)]}, Vi,j, (5.144)

x 1.) )
* and Dij are crisp numbers.

Note that both Dij
Step 5. Compute the Relative Closeness to the Ideals. This

* and Si- indices calculated in
index is used to combine the Si
Step 4. Since S: and si are crisp numbers, they can be combined:
(5.145)
The alternatives are ranked in descending order of the Ci index.

400
p.(x)
1.
Fig. 5.33 The derivation of L ij .
Numerical Example (Hwang and Yoon [H13])

A country wants to buy fighter jets. six attributes must be
considered: maximum speed (Xl)' ferry range (x 2 ), maximum payload
(X 3 ), price (x 4 ), reliability (x 5 ), and maneuverability (x 6 ). Four

types of fighter jets are to be evaluated with respect to these
attributes. The decision matrix is given as:
Xl X2 X3 X4 X5 X6
(Mach) (NM) (lbs) ($M)
D A1 1500 20000 5.5 avg. v. high
[ 2.0
A2 2.5 2700 18000 6.5 low avg.
1.8 2000 21000 4.5 high high
1
A3
A4 2.2 1800 20000 5.0 avg. avg.
[ mol indif. indif. indif. important very ]

important important
The decision matrix contains both crisp and fuzzy data. It is
assumed that the meanings of the linguistic terms in the decision
problem are known and can be represented by trapezoidal fuzzy numbers.
The decision matrix and the weight set are then rewritten as:
401
Xl X2 X3 X4 X5
D A1 2.0 1500 20000 5.5 (.3, .5, .5, .7)
[
A2 2.5 2700 18000 6.5 ( .1, .2, .2, .3)
A3 1.8 2000 21000 4.5 (.7, .8, .8, .9)
A4 2.2 1800 20000 5.0 (.3, .5, .5, .7)
X6
(.9, .95, .95, 1. 0) A1
(.3, .5, .5, .7) A2
( .7, .8, .8, .9) A3
( .3, .5, .5, .7) 1 A4
! [(.6,.675,.675,.75), (.4,.5,.5,.6), (.4,.5,.5,.6), (.4,.5,.5,.6),

( • 75, .825, .825, .90) , ( .9, .95, .95,1. 0) ] •
step 1. Obtain the Normalized Decision Matrix. Since all the
fuzzy numbers e [0,1], the normalization is carried out only on Xl'
X2 , X3 , and X4 using
r ij x ij I Xj* , for benefit attributes Xl' X2 , and X3
r ij -
Xj I x ij , for cost attribute X4 •
The results are given as:
Xl X2 X3 X4 X5
R A1 .80 .55 .95 .82 ( . 3, .5, .5, .7)
[
A2 1.00 1. 00 .86 .69 ( .1, .2, .2, .3)
A3 .72 .74 1.00 1.00 (.7, .8, .8, .9)
A4 .88 .67 .95 .90 ( . 3, .5, .5, .7)
X6
( .9, .95, .95, 1. 0) A1
( .3, .5, .5, .7) A2
( .7, .8, .8, .9) A3
( .3, .5, .5, .7) 1 A4
402
step 2. Obtain the Weighted Normalized Decision Matrix. Since
the weights are fuzzy data, we shall apply fuzzy multiplication on Xl'
X2 , X3 , X5 and X6 using Eq. (5.138) while using Eq. (5.139) on X4 (cost
attribute). Note the crisp numbers may be represented in trapezoidal
form. For example, given x 11 = 0.8 = (.8,.8,.8,.8) and w1 = (.6,.675,
.675,.75), we have
(.8,.8,.8,.8) (0) (.6,.675,.675,.75) ( .48, .54, .54, .6) .
By the same token, we can obtain all the vij's and present them as:
Xl X2
v (.48,.54,.54,.60) (.22, .275, .275, .33)
(.6,.675,.675.75) (.4,.5,.5,.6)
(.432, .486, .486, .54) (.296,.37, .37.444)
(.528, .594, .594, .66) (.268, .335, .335, .402)
X3 X4
(.38, .475, .475, .57) (.328, .41, .41, .492)
(.336, .43, .43, .516) (.276, .345, .345, .414)
(.4,.5,.5,.6) (.4, .5,.5,.6)
(.38, .475, .475, .57) (.36,.45,.45,.54)
X5 X6
(.225, .412, .412, .63) (.81,.903,.903,1) A1
(.075, .165, .165, .27) (.27, .475, .475, .70) A2
( .428, .66, .66, .81) (.63, .76, .76, .90)
1
A3
(.225, .412, .412, .63) (.27, .475, .475,.70) A4
step 3. Obtain the Positive and Negative Ideal Solutions.
First, M(V ij ), Vi,j, is computed using Eq.(5.140). They are
summarized as:
403
Xl X2 X3 X4 X5 X6
M A1 .54 .275 .475 .41 .422
.904 :]
[
A2 .675 * .500 * .43 .345 .170 .481
A3 .486 .37 .50 * .50 * .633 * .763
A4 .594 .335 .475 .450 .422 .482
For example, given v 11 (.48,.54,.54,.60), the M(v 11 ) is computed as:
[-(.48)2_(.54)2+(.54)2+(.60)2_(.48) (.54)+(.54) (.60)]

.54.
[3 (-.48 -.54 + .54 + .60)]
Once the v ij ' vi,j, are obtained, Vj* and Vj- are easily identified
as:
A* (V 1* , V2* ' V3* ' V4* ' V5* ' V6* ) (V 21 , V22 ' V33 ' V34 ' V35 ' V16 )
A - - -
(V~, V2 ' v 3 ' V4 ' V5 ' V~)
- - (V 31 , V12 ' V23 ' V24 ' V25 ' V26 )
step 4. Obtain the separation Measures S.* and S-:-.

1 1
Since all the data are fuzzy, the difference measures between v ij
and its corresponding v.* and v j , respectively, are: -

)
D*.. 1 - {sup /.I V (x) A /.I * (x) ]}, Vi, j

1)
x ij v.
)
1 - {sup A ]}, vi, j

Dij /.IV .. (x) /.I _(x)
x 1) V.
)
The results are two matrices shown as:
Xl X2 X3 X4 X5 X6
D* A1 1 1 .128 .495 .551 0
[ ]
A 0 0 .376 .917 1 1
A3 0 .747 0 0 0 .571
A4 .575 .988 .128 .263 .551 1
404
Xl X2 X3 X4 X5 X6
-
D A1 .474 0 .249 .430 .846 1
[
A 1 1 0 0 0 0
1
A3 0 .736 .376 .917 1 .803
A4 .900 .492 .249 .660 .846 0
For example, given v 11 , v *

1 , and v 1 , we have
-
* 1 - {sup (x) 1\ (x) ]} *

1 - Lll 1
v*
D11 Jl v Jl
x 11
1
D~l 1 - {sup
x
Jl v
11
(x) 1\ Jl
v1 - (x) ] } 1 - L~l .474
For alternative A1 , we compute its separation measures to PIS and
NIS, respectively, as:
6
*
Sl *
L D1j 1 + 1 + .128 + .495 + .551 + 0 3.174,
j=l
6
S~ L D~j .474 + 0 + .249 + .430 + .846 + 1 2.999.
j=l
S: and si, i = 2,3,4, can be obtained in a similar manner. The

separation measures are summarzed as:
i 1 2 3 4
3.174 3.293 1.318 3.505
2.999 2.000 3.832 3.147
Step 5. Obtain the Relative Closeness to the Ideals.
We shall use Eq.(5.145) to compute the relative closeness index

as:
S~ / (S~+ S~) .486.

405
We can obtain Ci , i 2,3,4, in a similar manner. Ci , vi, are

summarized as:
i 1 2 3 4
.486 .378 .744 .473
The rank ordering of the the four alternatives according to Ci is:
A3 > A1 > A4 > A2 •
Note
1. The algorithm of TOPSIS is not the only classical MADM method
available. There are other approaches such as Simple Additive Weight-
ing method, Linear Assignment Method, ELECTRE, and so on. Note that
the original algorithms must be modified slightly so as to incorporate
fuzzy mathematics or to simplify the computation. For example, in
classical TOPSIS normalization of the decision matrix is done using
m 2 1/2
x .. /
~J
[r
i=1
x .. ]
~J
(5.146)
which is not the same as that of the fuzzified TOPSIS proposed by

Negi. He argues that such SUbstitution is made simply for ease of
computation, because Eq.(5.146) implies immediate involvement of
complicated use of the extension principle.
2. Recall that in Negi's approach, PIS and NIS are obtained
using Lee and Li's ranking method (see Section 3.6). It is worth
noting, however, that other ranking methods such as Dubois and Prade
[029] (see Section 3.5) or Murakami et. al [M23] (see section 3.9) can
all be applied. Lee and Li's ranking method is chosen over other
ranking methods because it is easy to apply. Caution must be taken,
though, because different ranking methods may result in different PIS
and NIS which in turn may alter the ranking order of the alternatives.
3. Negi's approach suffers a great drawback which is shared by
all existing fuzzy MADM methods, i.e., it can't solve problems of
large dimensions, say m ~ 10, n ~ 10, with reasonable effort. Recall
that the numerical example stated earlier is only of 4 by 6 and the
amount of calculation is enormous. One can imagine that a problem
with 100 alternatives and 50 attributes is almost impossible to solve.
This certainly limits the applicability of this approach to large
real-world problems.
407
5.7 Fuzzy outranking Methods
The outranking relation of alternative ~ ~ Al says that even
though two alternatives k and 1 do not dominate each other
mathematically, the OM accepts the risk of regarding Ak as almost
surely better than Al (Roy [R9], Hwang and Yoon [H13]). Through the
successive assessments of the outranking relations of the other
alternatives, the dominated alternatives defined by the outranking
relation can be eliminated. To derive outranking relation between
pairs of alternatives is the key issue in classical outranking method.
Since it requires lengthy discussion, we will not describe it here.
The use of fuzzy outranking relations in MADM is first seen in
Roy [Rll,R13]. It is considered a fuzzy edition of the classical
ELECTRE method (see Roy [R9,R10], Hwang and Yoon [H13]). Since then
Takeda [T2], Siskos et al. [S20], Brans et al. [B35], and Martel et al.
[Mll] have developed various procedures in deriving fuzzy outranking

relations. The ranking of alternatives is conducted using the fuzzy
outranking relations.
Roy [R13] uses the degree of concordance and discordance to
derive fuzzy outranking relations; siskos et al. [S20] follows Roy's
approach but use different formulas and threshold values in deriving
the degree of concordance and discordance. Takeda [T2] proposed an

interactive approach for building fuzzy outranking relations from
which the OM's preference structure could be extracted as a fuzzy
multilevel graph. Brans et al. [B35] assist the analyst with six
different formulas for computing the degree of outranking. Each
formula is to be used under a unique circumstance. The analyst needs
to discuss with the OM and carefully select the appropriate formula
for each pair of alternatives. The Brans et al. approach is probably
the simplest one (in terms of computational requirement) in this
category. Martel el al.'s approach takes possibility distributions as

408
the decision data. A confidence index and a doubt index are computed
first which in turn are used to calculate fuzzy outranking relations.
We shall present Roy's approach for its originality, siskos et
aI's approach and Brans et aI's methods for their simplicity. We will
only present the basic concepts and the algorithm of Takeda's approach
because it is too cumbersome to use. Martel et al. approach is more
difficult to use than Takeda's approach and will not be presented
here. Interested readers should refer to the original paper [Mll].
Definitions
outranking Relation (Roy [Rll])

Given two alternatives, ~ and AI' "~ outranks AI" signifies
that the analyst has enough reasons to admit that in the eyes of the
OM, ~ is at least as good as AI. Hence, ~ is indifferent from or
preferred to AI.
The statement "Ak does not outrank AI" implies that the arguments
in favor of the proposition "~ is at least as good as AI" are judged
insufficient and that there exist arguments in favor of "AI is at
least as good as Ak ". Hence, Al is preferred or incomparable to ~.
Fuzzy outranking Relations (Roy [Rll], Zadeh et al. [Z5])

A fuzzy outranking relation Sd(k,l) can be characterized by a
membership function ~(k,l) which indicates the degree of outranking
associated with each pair of alternatives (~,AI). Note that ~(k,l)
does not indicate the intensity of Ak over Al rather it shows the

credibility of a preference existing of Ak over AI. ~(k,l) must
possess the following properties:
409
1. ~(k,l) increases with the reliability of the outranking of ~
over AI; thus, in particular, ~(l,k) is a nondecreasing function of
r lj , Yj and a non increasing function of r kj , Yj. The terms r lj and
r kj represent the performance of Ak and Al with respect to the jth
attribute, respectively.
2. ~(k,l) = 1 implies a certain outranking of ~ over AI' while
~(k,l) = 0 indicates either a certain nonoutranking of Al by Ak or
the total absence of arguments in favor of an outranking.
Consequently, ~(k,l) e [0,1].
A fuzzy outranking relation Sd(k,l) should be:
1. reflexive: Sd(k,k) = 1, Yk
2. nonsymmetrical: Sd(k,l) does not necessarily equal Sd(l,k).
3. transitive: Sd(k,l) ~ max min [S d (k,h),S d (h,l)], Yk,l.

h=l, ... ,m
h *" k,l
Based on these properties, we can induce some preference relations
such as (siskos et al. [S20]):
d d
Ak is preferred to Al iff S (k,l) > S (l,k);
Ak is indifferent from Al iff Sd(k,l) = Sd(l,k) > 0;
Ak is incomparable to Al iff Sd(k,l) = Sd(l,k) = o.
5.7.1 Roy's Approach
Roy [R11] proposed the use of the degree of concordance and the
degree of discordance to construct fuzzy outranking relations. There
are three thresholds must be specified: (1) t i (indifference
threshold), (2) t P (preference threshold), and (3)t V (veto threshold).
The interpretation for these thresholds under the jth attribute, Xj ,
are given as the following:
r kj ~ r lj + t; indicates Ak is at least as good as AI;

410
r kj ~ r lj + t3 indicates ~ is strictly preferred to AI; and
r kj ~ r lj + t j indicates ~ is largely better than AI.
In relation to the outranking of Al by ~, an attribute u appears more
or less discordant when the unfavorably different (r lu - r ku ) becomes
significant, (i.e., r lu - r ku > t~), and even if the other attributes
confirm the superiority of ~ over AI. When (r lu - r ku ) is too large,

attribute u plays the role of "dictator" when alternatives k an I are
compared and produces phenomena of incomparability. In order to
resolve the incomparability problem, the veto threshold is used to
denote the case where ~ can never outrank Al if r lu - r ku > t~. It
follows that t3 $ tj. For any attribute j, 0 s t~ s t3 $ tj.
Degree of Concordance and Degree of Discordance
The degree of concordance, Cj , which expresses the credibility of
proposition "Ak over AI" with respect to the jth attribute, is defined
as (see Zimmermann [Z31] p.184):
i, ~ r
1 ,r
kj + t J lj (5.147)
r (r + tl?
lj - kj J ti,
-..::::...<'---:.---....:..:;~----"''---, rkJ' + J $ r l J'
t~ - tl?
J J
Similarly, the degree of discordance, d j , which expresses the
credibility of the proposition "~ is not at least as good as AI", is
defined as (see Zimmermann [Z31], p. 184):

411
dj(k,l) (5.148)
r + t P,
t~ - tl? kj J
J J
Fig.5.34 presents the concordance and the discordance functions.
The concordance and the discordance relations have been defined
for each pair of alternatives and with respect to a single attribute.
The next step is to aggregate all the single-attribute concordance
relations, Cj , Vj, into a unified relation, C, as:
n
C(k,l) L wJ' CJ,(k,l), Vk,l, (5.149)
j=l
where wj , vj, are the weights assigned by the OM. The concordance and
the discordance relations must be aggregated to arrive at the
outranking relation. To this end a discordance relation, 0, is
defined in order to reduce the concordance by the discordance. Roy
defines the discordance relation, 0, as:
1 n
O(k,l) n L f(dJ,(k,l),C(k,l» (5.150)
j=l
where
1, d, (.) < C (. ) (5.151)

J
f(dj(k,l), C(k,l»
{ 1 - d, (,)
J
1 - C(')
, d j (, ) > C (,) , C (,) ~ 1.
The concordance relation, C, and the discordance relation, 0, are then

412
elk,1)
1"1---------,.
Fig. 5.34a A fuzzy concordance function.
Fig. 5.34b A fuzzy discordance function.
combined to give the fuzzy outranking relation Sd(k,l) as:
C(k,l)"D(k,l), Vk,l. (5.152)
Algorithm
Roy's approach will be presented as a series of successive steps
as follows.
step 1. Initialization. The decision maker must set the
413
i
thresholds: (1) indifference threshold, t j , (2) preference threshold,
t pj , and (3) veto threshold, t vj , where j = 1, •.• ,n.

step 2. Construct the concordance relation Cj for the jth
attribute, X., using Eg.(5.147).

J
step 3. Construct the discordance relation d j for the jth
attribute, X., using Eg. «5.148).

J
step 4.1. Given the single-attribute concordance relations, Cj , Vj,
an aggregated concordance relation C can be obtained using Eg. (5.149).
step 4.2. Given the single-attribute discordance relations, d j ,
Vj, and the aggregated concordance relation C, an aggregated
discordance relation D can be obtained using Egs.(5.150) and (5.151).
step 5. By using Eg. (5.152) to combine the aggregated concordance
relation C and the aggregated discordance relation D, we can obtain
the fuzzy outranking relation sd.
step 6. Given the fuzzy outranking relation sd, three possible
ranking schemes--dichotomy, trichotomy, and rank ordering--can be used
to compare alternatives. Dichotomy method is to select the best
alternative. Trichotomy method is bascially for screening purpose.
Rank ordering method is to prioritize the alternatives in descending
order.
Numerical Example (siskos et al. [S20])
Four candidates are evaluated for a managerial position. The
attributes used are experience (Xl)' communication skill (X 2 ), and
maturity (X 3 ). The decision matrix is given as:
Xl X2 x3
A1 5.2 5.6 2.6
[
A2 7.1 4.5 1.8
3.8 8.0 6.5 (5.153)
1
A3
A4 6.4 4.2 9.0
414
The weight is given by the OM as:
W (.25, .40, .35).
The data are subjective judgments of the OM based on a scale of 0 to

10, with 0 being the worst performance and 10 being the best perfor-
mance. The problem is solved by the following steps.
step 1. Initialization. The OM sets the thresholds as t~ .5,

J
j 1,2,3, t3 = 2, j = 1,2,3, and tj = 5, j 1,2,3.
step 2. Construct the single-attribute concordance relation, Cl ,

using Eq.(4.147) as:
Al A2 A3 A4
A1 1 .067 1 .53
C1 A2
A3
A4
[ .40
1
1
1
0
.867
1
1
1
1
0
1 1
For example, we can obtain C1 (1,2) using Eq.(5.l47) as:
7.1 - 7.2
C1 (1,2) i p
.067,
.5 - 2
b ecause r 11 + t 1i s r 21 s r 1l + ti is found. Similarly, we determine
because the condition r 21 + ti ~ r 1l is met.

415
similarly, we can obtain C2 and C3 as:
A1 A2 A3 A4
A1 1 1 0 1
[
A2 .60 1 0 1
C2
1 1 1 1
1
A3
A4 .40 1 0 1
A1 A2 A3 A4
A1 1 1 0 0
[
A2 .80 1 0 0
C3
1 1 1 0
1
A3
A4 1 1 1 1
step 3. Construct the single-attribute discordance relation, d 1 ,
using Eq. (4.148) as:
A1 A2 A3
0 0 0
.~o 1
A1
[
A2 0 0 0
d1
A3 0 .433 0
A4 0 0 0
For example, by using Eq.(5.148) we can obtain
d 1 (3,2)
7.1 - (3.8+2)
.433,
5 - 2
because r 31 + ti ~ r 21 ~ r 31 + t~ is found.
Similarly, we can obtain d 2 and d 3 as:

416
A1 A2 A3 A4
A1 0 0 .133 0
[
A2 0 0 .500 0
d2
1
A3 0 0 0 0
A4 0 0 .600 0
A1 A2 A3 A4
A1 0 0 .633 1
[
d3 A2 0 0 .90 1
A3 0 0 0 .167
A4 0 0 0 0 1
step 4.1 Construct the aggregated concordance relation, C, using
Eq. (4.149) as:
A1 A2 A3 A4
A1 1 .767 .25 .533
[
A2 .77 1 .25 .65
C
1
A3 .85 .750 1 .40
A4 .76 .967 .60 1
For example, we can obtain C(1,2) using Eq.(5.149) as:
C(1,2) .767.
step 4.2. Construct the aggregated discordance relation, D,

using Eqs.(4.150) and (4.151) as:
A1 A2 A3 A4
A1 1 1 .830 .667
[
A2 1 1 .600 .667
D
A3 1 1 1 1
A4 1 1 1 1 1
417
For example, we can obtain 0(1,2) using Eqs.(5.150) and (5.151) as:
[f(d 1 (1,2),C(1,2»+f(d 2 (1,2),C(1,2»+f(d 3 (1,2) ,C(1,2»]

0(1,2)
3
1,
where f(d 1 (1,2),C1,2» = 1 because C(1,2) = .767 > d 1 (1,2) = 0; and
f(d 2 (1,2) ,C(1,2» = f(d 3 (1,2),C(1,2) = 1 because C(1,2) > d 2 (1,2) and
C(1,2) > d 3 (1,2), respectively.
step 5. To combine concordance and discordance relations into
the outranking relation, sd, we apply Eq.(4.152) as:
A1 A2 A3 A4
A1 1 .767 .208 .356
[
sd A2 .77 1 .15 .434
.85 .750 1 .40
1
A3
A4 .76 .967 .60 1
For example, Sd(2,3) = C(2,3)oO(2,3) = .15.
Step 6. Given the outranking relation sd, three possible ranking
schemes--dichotomy, trichotomy, and rank orderings--can be used.
(I) Dichotomy method: Selecting the best alternatives. First of
all, we can derive a crisp outranking relation sd by setting an a
level. For example, by setting a = 1, we get a crisp outranking
relation sd as:
A1 A2 A3
A1 1 0 0
[
sd A2 0 1 0
1 0
A3 0 1
A4 0 0 0 1
where none of the alternatives outrank the others.
418
In general, the chance of getting the best alternative, A* , is

not too high. The decision analyst may have to determine Sa* for a ~ 1
such that a is considered high enough to assert dominance of alter-
native A* over the others. If A* is still not obtainable, we can at
least reduce the number of alternatives from m to a smaller number.
For example, at a = 0.85 we get the nondominated alternatives {A 3 ,A4 }
because Sd(3,1) 0.85 ~ a and Sd(4,2) = 0.967 ~ a.
This method is particularly useful when selecting applicants for
a well-defined post, or new investment plans for a corporation.
(II) Trichotomy method: This method is suitable for problems
such as admitting students to graduate school or increasing the salary
of those who work hard.
The OM is to set three thresholds, A1, A2, and A3 with which the
alternatives are classified into "acceptable", "nonacceptable", and
"nondecidable". All A'S are confined in [0,1]. All alternatives
which have sd ~ A1 will be put in the acceptable class. The rest of
the alternatives are to be evaluated against A2 and A3. Fig. 5.35
presents a decision tree which defines the segmentation of
alternatives.
Fig. 5.35 Roy's decision tree.

419
(III) Rank ordering the alternatives: The decision maker wants

to prioritize the alternative set A. We shall start with setting
crisp relation S~.8 which produces a subset B = {A 3 ,A 4 } of two
alternatives for which none is outranked in alternative set A with a
credibility greater than 0.8. We may consider A3 and A4 as candidates
for first place in a classification P' that is being sought. When ~ =
.6, the crsip relation S~6 gives B' = {A4 }. Thus A4 is ranked first.
It follows that A3 ranked sencond. By carrying out similar procedure
on alternative set {A1 ,A 2 }, we find that we cannot tell the preference
relation between Al and A2 . The ranking order--a weak order--is given
as:
Before proposing P' to the OM, we need to construct another weak

order P" following the same reasoning but proceeding in the opposite
direction. When P' and P" are too far different, Roy [Rll] suggests
reconstructing the problem from the beginning, i.e., reevaluating
problem data. If they are similar, the principle discussed in ELECTRE
II (Roy and Bertier [R12]) can be used to resolve minor differences.
The rank ordering, P, is then obtained.
The analyst is to discuss with the OM which ranking methodes) to
use. If the problem contains too many alternatives, one may use
trichotomy method first to screen out those dominated alternatives,
and then apply the rank ordering method to get the linear order.
5.7.2 siskos et al.'s Approach

siskos et al. [S20] present a fuzzy outranking method that is
similar to Roy's algorithm (Roy [R9,R13] Zimmerman [Z31]). The are
two major differences between these algorithms. The formulas used in
deriving concordance relations, Cj , and discordance relations, d j ,
420
are different. And after the outranking relation is constructed,
sisko et al. build a fuzzy dominance relation and subsequently a fuzzy
nondominance relation. The alternative with the highest degree of
nondominance is said to be the best.
Fuzzy Partial Concordance Relation Cj(k,l}
The concordance index is defined in such a way that Cj(k,l} gives
the degree of outranking of alternative Ak over alternative Al under

'
f UZZlness regar d'lng th e J,th a tt rl'bu t e a I one. The uncertainty occurs
when the Ir kj - rIjl differences are very small and can be considered
nonsignificant.
To make the concept operational, a notion of the maximum
nonsignificant threshold, t j , is introduced. The difference between
Ak and Al under jth attribute is significant if Ir kj - rIjl ~ t j . And

it is clear that as long as r Ij stays within the range of [r kj , r kj +
tj] alternative Ak continues to be "at least as good as" alternative
Al with decreasing credibility (see Fig. 5.36a for details). The
threshold tj allows the OM to express his/her uncertainty concerning
the fuzziness inherent in comparing alternatives with respect to a

single attribute.
with the nonsignificant threshold tj as the indicator that
represent the maximum error of imprecision in the evaluation, we can
define the fuzzy partial concordance relation Cj(k,l} as:
1, i f r Ij - r kj :s 0,
Cj(k,l}
{ 0, i f r Ij - r kj tj , (5.154)
~
r Ij - r kj
[1 - ]L, L > 0, otherwise
t,
J
Note that the decrease of Cj(k,l} can be determined by linear

421
p,(x)
C.(k,1)
1.1----+-'---..
Fig. 5.36a Fuzzy partial concordance relation at L 1.
interpolation (when L = 1) or any other form (when L * 1). The
nonsignificant threshold, t j , may be different from attribute to
attribute.
Fuzzy Concordance Relation, C(k,l)
Having obtained n partial fuzzy outranking relations, n being the
number of attributes, we can aggregate these partial outranking
relations into a single one that takes into account the weight of each
attribute. The aggregation formula is a simple weighted sum that
defines a new fuzzy relation called fuzzy concordance relation,
n
C(k,l) L WJ. C J' (k,l), (5.155)
j=l
where C(k,l) indicates the aggregated degree of dominance of Ak over
Al.
Fuzzy Partial Discordance Index dj(k,l)
In relation to the outranking of Ak over Ai' an attribute j
appears to be more or less discordant when the unfavorable [r lj - r kj ]

422
on that attribute becomes significant, i.e., [r Ij - r kj ] > t j , and even

if the other attributes confirm the superiority of ~ over AI. When
the difference is too unfavorable, attribute j plays the role of
"dictator" when two alternatives are compared and produces phenomena
of incomparabililty (see Fig. 5.36b). To account for the
incomparabillity between alternatives, a veto threshold, v j , is
introduced. That is when [r Ij - r kj ] > v j , alternative Ak can never
outrank AI. Therefore, our fuzzy discordance index can be defined
as:
I
1, i f r Ij - r kj ~
rj
dj(k,l) 0, i f r Ij - r kj ~
tj (5.156)
r Ij - r kj
]L, L >0, otherwise
v. - t.
J J
Note that the increase of the discordance of (~,AI) can be determined
by linear interpolation (L = 1) or any other form (L * 1).
p.(x)
1.
Fig. 5.36b Fuzzy partial discordance relation at L 1.

423
sometimes it is difficult for the OM to precisely determine all
the Vj values. Indeed, in cases where such a value is fuzzy and
varies within a certain interval, it is wise to give Vj the maximum
value of the interval so as to avoid uncertain and tricky
incomparability.
Fuzzy outranking Relation Sd(k,l)
Fuzzy outranking relation Sd(k,l) is obtained through linking
fuzzy concordance matrix and fuzzy discordance matrices. It is
important to note that
If Then
C(k,l) 1, Vk,l Cj(k,l) 1
and
dj(k,l) 0
C(k,l) < 1, Vk,l dj(k,l) > 0
Given the global character of C(k,l) as a binary preference indicator,
we propose that outranking relation Sd(k,l) = C(k,l) when discordance

attribute effect are not too great going beyond C(k,l), i.e.,
Hence, the outranking relation is defined as:
C(k,l), if C(k,l) ~ dj(k,l), Vj
C(k,l) (5.157)
{IT*[l - d *(k,l)]} 1-C(k,1) ,
j j
for { jl C(k,l) < dj*(k,l)}.
424
The second part of this definition allows us to take into account only
the most significant discordances in the calculation of Sd (k,l). In
the case of dj*(k,l) =1 for at least one attribute j*, the attribute
Xj * plays the role of "dictator" to the outranking of Ak over AI.
The properties of Sd(k,l) are summarized as:
(i) reflexive: Sd(k,k) = 1, Vk.
(ii) nonsymmetric: Sd(k,l) does not necessarily equal Sd(l,k).

(iii) transitivity: Sd(k,l) ~ max min [Sd (k,q),S d (q,I)],Vk,l.
q=I, .•• ,m
q .. k,l
Based on the properties, we can induce some better known preference

relations as:
~ is preferred to Al iff Sd(k,l) > Sd(l,k)
~ is indifferent to Al iff Sd(k,l) = Sd(l,k) > 0
~ is incomparable to Al iff Sd(k,l) = Sd(l,k) = 0
Dominance and Nondominance Relations

A fuzzy dominance relation is defined by a membership function
~D(k,l) that measures the outranking intensity. It is computed as:
{ 0, otherwise. (5.158)
since ~D(k,l) is the fuzzy set of alternatives AI' VI, that are
dominated by ~, it is natural to define the nondominance relation
as the complement of ~D(k,l). THat is a nondominance relation can be
defined as: (see Section 3.2.2.1 Complement of a Fuzzy Set)
1 - ~D(k,l). (5.159)
425
similarly, we know ~ND(I,k) is the fuzzy set of alternatives ~,
Vk, that are not dominated by AI. The intersection of all ~ND(I,k)
for all alternatives will result in the fuzzy set of alternatives that
are not dominated by any alternatives. Thus, we get a fuzzy set of
nondominated alternative, ~*(Ak)' which is defined by
min ~ND(I,k) (5.160)

1=1, ••• ,m
l .. k
min [1 - ~D(I,k)]
1 - max ~D(I,k)
1 - max [Sd(l,k) - Sd(k,I)]
The best alternative is the one that satisfies
max ~ * (Ak ) (5.161)

k=l, ... ,m
1 - min { max [Sd(l,k) Sd(k,I)] }
k=l, ... ,m 1=1, ••• ,m
l .. k
Algorithm
The algorithm is described by the following steps.
step 1. Initialization: The DM is asked to give his/her maximum
nonsignificant threshold, t j , veto threshold, v j , and weight, wj , for

each attribute.
step 2. Construct partial fuzzy concordance matrices.
step 3. Construct fuzzy concordance matrix by aggregating
partial fuzzy concordance matrices.
step 4. Construct fuzzy discordance matrices.
step 5. Construct fuzzy outranking relation matrix using fuzzy
concordance matrix and fuzzy discordance matrices.

426
Step 6. Construct fuzzy dominance relation using fuzzy
outranking relation.
stee 7. Construct nondominance relation, /.LNO(k,l) .
stee 8. * (A )
Obtain a set of nondominated alternatives using, Ilk .
stee 9. Identify the best alternative, A* , with
Il * (A * ) max Il * (~).
k=l, ... ,m
Numerical Example (siskos et ale [S20])
A company has to choose among four candidates, A1 , A2 , A3 , and
A4 , for a managerial position. There are three attributes,
Xl (experience) , X2 (communication skill), and X3 (maturity) to be
considered. The rating for each alternative under each attribute is
given based on a scale of ° to 10 (10 being the maximal evaluation).

The decision matrix, G, is given as
Xl X2 X3
A1 5.2 5.6 2.6
G A2 7.1 4.5 1.8
A3 3.8 8.0 6.5
A4 6.4 4.2 9.0
The problem is solved by following steps.
step 1. Initialization.
Assume that the maximal nonsignificant threshold vector ~, the
veto threshold vector ~, and the weight set ware given by the OM as:
(2,2,2) (5,5,5) (0.25,0.40,0.35)

427
step 2. Construct the partial fuzzy concordance matrices.
For the first attribute, Xl' we construct the partial fuzzy
outranking matrix as:
A1 A2 A3 A4
A1 1 .05 1 .40
A2 1 1 1 1
A3 .30 0 1 0
A4 1 .65 1 1
where C1 (1,2) is calculated using Eq. (5.154) as follows. Since
(r 21 - r 11 ) = 7.1 - 5.2 = 1.9 E [0,2), C1 (1,2) is computed as:
(7.1 - 5.2)
C1 (1,2) = 1 - 2
0.05.
Similarly, we get C2 and C3 as:
A1 A2 A3 A4
A1 1 1 0 1
C2 A2 .45 1 0 1
A3 1 1 1 1
A4 .30 .85 0 1
A1 A2 A3 A4
A1 1 1 0 0
C3 A2 .60 1 1 0
A3 1 1 1 0
A4 1 1 1 1
428
step 3. Construct the fuzzy concordance matrix.
We aggregate all three partial concordance matrices and form a
fuzzy concordance matrix using Eq.(5.155) as:
A1 A2 A3 A4
A1 1 .76 .25 .50
C A2 .64 1 .25 .65
A3 .83 .75 1 .40
A4 .72 .85 .60 1
For example C(l,2) is computed as:
C(l,2) .76.
step 4. The first fuzzy discordance matrix is computed as:
A1 A2 A3 A4
A1 0 0 0 0
d1 A2 0 0 0 0
A3 0 .43 0 .20
A4 0 0 0 0
The element d 1 (3,2) is calculated using Eq.(5.156) as follows. Since
r 21 - r 31 = 7.1 - 3.8 = 3.3 e [2,5], we compute d 1 (3,2) as: (L=1)
d 1 (3,2) 0.43.
Similarly, we get martrices d 2 and d 3 as:

429
A1 A2 A3 A4
0 0 .13 0
0 0 .50 0
0 0 0 0
0 0 .60 0
A1 A2 A3 A4
0 0 .63 1
0 0 .90 1
0 0 0 .17
0 0 0 0
step 5. Construct the fuzzy outranking matrix using concordance
and discordance matrices.

The fuzzy outranking matrix may be obtained using Eq. (5.157) as:
A1 A2 A3 A4
A1 1 .76 .12 0
sd A2 .64 1 .02 0
A3 .83 .75 1 .40
A4 .72 .85 .60 1
where Sd(1,3) is calculated as follows: Since C(1,3) = .25 is
greater than both d 1 (1,3) = 0 and d 2 (1,3) = .13 but is smaller than
d 3 (1,3) = .63, the value of Sd(1,3) is computed as:
C(1,3) .12.
1-C(1,3)
step 6. Compute the dominance relation.
The dominance relation is calculated using Eq.(5.158) as:

430
A1 A2 A3 A4
A1 0 .12 0 0
/.1 0 A2 0 0 0 0
A3 .71 .73 0 0
A4 .72 .85 .20 0
For example, since C(4,3) .60 > C(3,4) 0.40, the value of /.10(4,3)
is
• 60 - .40 .20 •
step 7. Compute the nondominance relation.
The nondominance relation, /.INO' is computed using Eq. (5.159) as:
A1 A2 A3
A1 1 .88 1
/.I NO A2 1 1 1 1
A3 .29 .27 1 1
A4 .28 .15 .80 1
where /.INO(l,2) 1 - /.10(1,2) 1 - .12 .88 based on Eq.(5.159).
step 8. Compute the degree of dominance.
For alternative A1 , the degree of nondominance of A1 by other
alternatives is computed as:
min [1,.29,.28]
.28
We may obtain /.I * (Ak ), k 2,3,4, in a similar manner. They are
summarized as:
431
k 1 2 3 4
j.1 * (~) .28 .15 .80 1
step 9. Identify the best alternative.
Since the alternative that has the highest j.1* value is considered
the best one, we select A4 because j.1 * (A4) = 1.
Note
The same numerical example in this section is solved again using
the classical SAW method. Assume that the weight ~ = (.25,.40,.35)
are used. The normalized weighted final ratings are:
.564 .545 .786 .785
For example,
3
E wJorlJo (.25) (.73)+(.40) (.70)+(.35) (.29) .564.
j=1
Notice that r 1j , vj, have been normalized using
rO1J°
where x ij is the performance score of Ai with respect to Xj and Xj* is
the best value among x ij . The final ranking order is A3 > A4 > Al >
A2 •
It is worth emphasizing that both the classical SAW method and
siskos et al.'s fuzzy approach result in a solution. But the two
ranking orders do not always agree with each other. It is difficult
to judge which method gives consistent, reliable solution. It is
clear, however, that the classical SAW method is much easier to apply,
while fuzzy outranking method is very cumbersome to use. This seems

432
to suggest that while fuzzy outranking methods may be theoretically

sound but it is too costly to apply to solve MADM problems regardless
the problem size.
5.7.3 Brans et al.'s Approach

Brans, Mareschal, and Vincke [B35] proposeed a family of
outranking methods called Preference Ranking Organization METHods for
Enrichment Evaluations (PROMETHEE). The families of PROMETHEE are
Promethee I, Promethee II, Promethee III, and Promethee IV. A partial
order can be obtained using Promethee Ii a complete order can be
obtained using Promethee IIi and an interval order is provided by
Promethee III. As for Promethee IV, it deals with multiple objective
decision making problems and will not be discussed here.
In general, these methods include the following three phases.
1. For each attribute, Xj , a generalized criterion {Xj,H(d)} is
constructed by considering a preference function, P.
2. A multi-attribute preference index is defined in order to
obtain an outranking relation representing the preferences of the DM.
3. The evaluation of the outranking relation is obtained by
considering for each alternative a leaving and an entering flow.
Generalized criterion
Let Xj be an benefit attribute, i.e, bigger better. When two
alternatives, Ak and AI' are compared with respect to Xj ' the result
of the comparison must be expressed in terms of preferences.
Therefore, a preference function P(k,l) which gives the intensity of
the preference of ~ over Al is defined. The preference structures
pertaining to P(k,l) and their meaning are summarized below.
There are four possible preference relations between alternatives
Ak and AI· They are:
433
( i) P(k,l) = 0: No preference of ~ over AI.

(ii) P(k,l) close to 0: Ak is weakly preferred to AI.
(iii) P(k,l) close to 1: Ak is strongly preferred to AI.
(iv) P(k,l) = 1: ~ is strictly preferred to AI.
It is realistic to consider each preference function P(k,l) as a

nondecreasing function of the difference between ~ and AI. Let
the difference of Ak and Al with respect to the jth attribute be
The preference function--with respect to the jth attribute--P(k,l)

may be characterized by the function P(d j ) shown in Fig. 5.37.
Given the preference function of ~ over Al (when d > 0) and
the preference function of Al over ~ (when d ~ 0), a function R(d)
can be defined as (see Fig. 5.38):
P(k,l) , if d ~ 0,
R(d)
{ P(l ,k) , if d ~ o.
............. _;;,;;_._---
Fig. 5.37 A preference function P(d).

434
H(d)
PREFERENCE 0
AtOVER"
~------~~~------~d
Fig. 5.38 A B(d) function.
The function B(d) is used to characterize the OM's preference
structure of ~ over AI' since the OM's preference structure varies
from attribute to attribute, the function B(d) must be different from
one attribute to another. To indicate that a particular B(d) function
represents the OM's preference structure with respect to the jth
attribute, Martel et al. create the term {Xj,B(d)}. This term is
called "generalized criterion."
For each attribute Xj , a generalized criterion must be defined.
This can be done through interactive discussion between the OM and
system analyst. In order to facilitate the identification of the
preference functions, Brans et al. have listed six possible types of
generalized criteria. They are:
0, d = 0,
B(d) { 1, Idl > 0,
(see Fig. 5.39) (5.162)
0, Idl :s u,
B(d) { 1, otherwise,
(see Fig. 5.40) (5.163)
I v , Idl :s v,
B(d) { Idl1, otherwise,
(see Fig. 5.41) (5.164)
435
0, Idl :S u,
R(d) { 0.5, u :S Idl :S v, (see Fig. 5.42) (5.165)
1, otherwl.se,
0, Idl :s u,
R(d) { (Idl - u) I(v - u),u :S Idl :s v, (5.166)
1, otherwise,
( see Fig. 5 . 43 )
d 2
R(d) 1 - exp{- -----}. (see Fig. 5.44) (5.167)
2 0- 2
H(d)
~-----------±-------------d
Fig. 5.39 General criterion type 1: usual criterion.
H(d)
~------_~q----~I~--~q------~d
Fig. 5.40 General criterion type 2: quasi criterion.

436
H(d)
~----=-----~----~----~d
Fiq. 5.41 General criterion Type 3: linear preference.
H(d)
~------~--~------------~d
-p -q q p
Fiq. 5.42 General criterion Type 4: level criterion.
H(d)
~---7---_~q--~--~q~--~p----d
Fiq. 5.43 General criterion Type 5: criterion with linear

preference and difference area.
437
H(d)
-- ----------- -- ---- --- --1 --- ---------- ----- ----
~~------~~~----~--~d
a
Fig. 5.44 General criterion Type 6: Gaussian criterion.
Each of the shapes can be very easily defined because only one or two
parameters are to be fixed. The possible parameters are:
1. u j is an indifference threshold--the largest d j value below
which the DM considers there is no difference.
2. v. is a strict preference threshold--the lowest value of d.
J J
above which the DM considers there is strict preference.
3. u j is the standard deviation of a normal distribution.
These formulas do not represent an exhaustive list. other shapes of
the H(d) function could be considered, too.
If criterion Xg is to be minimized, then the formula for
calculating the difference between ~ and Al must be defined as:
(5.168)
The generalized criterion associated to Xg can then be defined as

{Xg,H(dg )}, where H(dg ) can take the same shape as those in H(d j ).
The parameters for H(d g ) are: u g ' v g ' and u g ' which have the same
meanings as those in Xj .
438
Algorithm
The following steps are used in sequence to solve a MADM problem.
step 1. Initialization.
Through iterative discussion with the DM, the type of H(d)
function with respect to each attribute and its parameters u, v, or ~
are determined.
step 2. Calculate the difference between pairs of alternatives.
Calculate the difference between the alternatives Ak and Al with
respect to cost attribute Xg or benefit attribute Xj using
'Vi (5.169a)
'Vi (5.169b)
step 3. Construct outranking relation matrix.
The elements in the outranking relation matrix are defined as:
~ {[[ H(d.)] + [[ H(d g )]}, d g , d j ~ 0, 'Vg,j, (5.170)

j J g
where n is the number of attributes, and Sd(k,l) E [0,1].
Note that when Sd(k,l) approaches 0, Ak is slightly better than
AI' while when Sd(k,l) approaches 1, Ak is strongly preferred to AI'
The outranking relation matrix is defined as:
A1 A2 Am
I
sd A1 Sd(1,2) sd(1,m) (5.171)
A2 sd (2,1) sd(2,m)
Am sd(m,1) sd(m,2)
439
Notice that Sd(k,l) does not necessarily equal Sd(l,k)

step 4. Determine the degree of optimality.
In order to evaluate the alternatives, the following relations
are considered.
1. The degree of outranking
F+(k) =E Sd(k,I), Vk; (5.172)

1=1, •.. ,n
l~i
2. The degree of dominance
F-(k) =E Sd (l,k), Vk,• (5.173)

1=1, .•. ,n
1 ~ k
3. The degree of optimality
F(k) (5.174)
Physically, F+(k) gives a measure of the degree of outranking of

~ over other alternatives. F-(k), on the other hand, measures the
degree of dominance of ~ by other alternatives. F(k) measures the
overall performance of ~ against other alternatives.
step 5. Obtain the ranking order.

The ranking order can be obtained using one of the following
methods.
1. Promethee I: The following decision rules are used to
determine the relation of alternatives ~ and AI.
440
If Then
F+(k) > F+(l) p+

~ Al
F+(k) = F+(l) 1+
~ Al
F-(k) < F-(l) -

~ p Al
F-(k) = F-(l)
~ I
- Al
p+
AI] and [~ P
-
[~ Al ]
~ outranks Al
p+ A ] and I-
[~ l [~ Al ] ~ outranks Al
1+ A ] and [~ P-
[~ l Al ]
~ outranks Al
[Ak 1+ A l ] and [~ I- Al ] Ak is indifferent to
none of the previous comditions Ak and Al are

incomparable
2. Promethee II: In this method, two general rules are applied
to determine the relationship of ~ to Al •
If Then
F(k) > F(l) ~ outranks Al
F(k) F(l) ~ is indifferent to Al
3. Promethee III: For~, an interval [Zk'Y k ] is defined as:
(5.175)
(5.176)
where a < 0 controls the number of strict outranking,

441
F(k) ! F(k), (5.177)
(5.178)
Note that the center of [Zk'Y k ] is the mean of F(k) and the
length of [Zk'Y k ] is proportional to the standard error of the

distribution of [d(k,l) - Sd(k,I)]. The choice of the a value depends
on the application. However, in order to avoid too many indifferences
between pairs of alternatives, it is better to have the mean length of
all the intervals be less than the mean distance between two
successive centers of intervals. This leads in general to a value of
about 0.15 for a.
with the intervals, the Ak and Al relation is determined as:
If Then
~ is indifferent to Al
Numerical Example (Brans et al. [B35])
Six criteria are considered by the OM to rank five power plant
proposals (AI' A2 , A3 , A4 , AS). The criteria are: manpower needed
(Xl)' electricity generated (X 2 ), construction cost (X 3 ), maintenance
cost (X 4 ), number of villages to evacuate (X 5 ), and safety impact
(X 6 ). Of the six attributes, X2 and X6 are to be maximized while the
others are to be minimized. The problem is then put in a decision
matrix, o.
442
(min) (max) (min) (min) (min) (max)

Xl X2 X3 X4 X5 X6
D A1 80 90 6 5.4 8 5
A2 65 58 2 9.7 1 1
.,
A3 83 60 4 7.2 4 7
A4 40 80 10 7.5 7 10
A5 52 72 6 2.0 3 8
The problem is solved by the following steps.
Step 1. Initialization.
After serious discussion with the DM, the analyst concludes that
the H(d) functions for attribute, Xj , j = 1, ... , 6, are summarized
in Table 5.15.
Table 5.15 H(d) Functions Used

attribute H(d) type parameters
Xl II u 10
X2 III v 30
X3 V u 0.5; v 4.5
X4 IV u 1; v 5
X5 I
X6 VI (j = 5
Step 2. Calculate the difference between pairs of alternatives
with respect to each attribute.
For cost attributes, i.e., smaller better, Eq.(5.169a) is used to
calculate the difference between alternative Ak , k = 1,2,3,4,5, and
alternative AI' 1 = 1,2,3,4,5, where k ~ 1. For benefit attributes,
i.e. bigger better, Eq. (5.169b) is used to calculate the difference
between pairs of alternatives with respect to each attribute. The
results are summarized in Table 5.16.

443
Table 5.16 Summary of d Values
type of H(d) II III V IV I II
d j or d g Xl X2 X3 X4 X5 X6
(A 1 , A2 ) -15 32 -4 4.3 -7 4
(A 1 ' A3 ) 3 30 -2 1.8 -4 -2
(A 1 , A4) -40 10 4 2.1 -1 -5
(A 1 , A5) -28 18 0 -3.4 -5 -3
(A 2 , A1 ) 15 -32 4 -4.3 7 -4
(A 2 , A3 ) 18 -2 2 -2.5 3 -6
(A 2 , A4) -25 -22 8 -2.2 6 -9
(A 2 , A5) -13 -14 4 -7.7 2 -7
(A 3 , A1 ) -3 -30 2 -1.8 4 2
(A 3 , A2 ) -18 2 -2 2.5 -3 6
. (A 3 ,A4) -43 -20 6 0.3 3 -3
(A 3 , A5) -31 -12 2 -5.2 -1 -1
(A 4 , A1 ) 40 -10 -4 -2.1 1 -5
(A 4 , A2 ) 25 22 -8 2.2 -6 9
(A 4 , A3 ) 43 20 -6 -0.3 -3 3
(A 4 , A5) 12 8 -4 -5.5 -4 2
(A 5 , A1 ) 28 -18 0 3.4 5 3
(A 5 , A2 ) 13 14 -4 7.7 -2 7
(A 5 , A3 ) 31 12 -2 5.2 1 1
(A 5 , A4) -12 -8 4 5.5 4 -2
For example, the difference between A1 and A2 with respect to Xl
(cost attribute) is calculated using Eq. (5.169a) as:
65 - 80 -15.
On the other hand, since X2 is to be maximized, the difference between
A1 and A2 with respect to X2 is computed using Eq. (5.169b) as:
90 - 58 32.
444
step 3. Calculate the outranking relation, Sd(k,l).
For each pair of alternatives (Ak,A1 ), we calculate Sd(k,l) using
the H(d} functions defined in step 1. From Table 5.15, we know that
Eq. (5.163) is used for H(d 1 }; Eq.(5.164} is used for H(d 2 }; Eq. (5.166)
is used for H(d 3 }; Eq. (5.165) is used for H(d 4 ); Eq. (5.162) is used
for H(d 5 }; and Eq. (5.167) is used for H(d 6 ).
Therefore, Sd(1,2) can be obtained as:
1
Sd(1,2}
n {[1:H(d.}] + [1: H(d g } ]}
j J g
1 .296
"6 [H(d 4 ) + H(d 2 } + H(d 6 }]
where 0, and g = 2,6, j 4

dj,d g
'"
H(d 4 } 1 because d 2 32 > v 30
H(d 4 } 0.5 because u 2 1 :s d 4 4.3 :s v 5
d6
- 1 - exp (- 42
H(d 6 ) 1 - exp( .274
20"2 2 ( . 5) 2
Similarly, we can get preferences for other pairs of alternatives.
The results may be summarized as an outranking matrix:
A1 A2 A3 A4 A5
Sd A1 .296 .250 .285 .100
A2 .479 .396 .333 .313
A3 .242 .180 .333 .063
A4 .333 .506 .305 .224
A5 .444 .515 .570 .479
445
step 4. Determine the degree of optimality.
In order to calculate the degree of optimality for alternative
AI' the degree to which Al outranks others and the degree of dominance
of Al by others are computed first as:
F+(l) L Sd(l,l) .931

1=2,3,4,5
F-(l) L Sd(l,l) 1. 498.

1=2,3,4,5
The degree of optimality of Al is then calculated using Eq.(5.174) as:
F (1) -.567.
similarly, we can calculate all the F+, F-, and F for all the
other alternatives as summarized below.
F+ (1) -
F (1) F (1)
Al .931 1.498 -.567

A2 1. 521 1. 497 .024
A3 .818 1. 521 -.703
A4 1. 368 1.430 -.062
A5 2.008 .700 1. 308
step 5. Obtain the ranking orders.
For demonstration purpose, we shall derive the ranking order
using Promethee I, II, and III, respectively.
1. Promethee I. If the Promethee I method is used, the follow-
ing conclusion can be made. Clearly, A5 outranks other alternatives
because its F+ is the highest but its F- is the smallest. That is A5
has the highest degree of domination over other alternatives but has
the lowest degree of dominance by other alternatives. It is also
observed that A2 and A4 are incomparable. A2 is very cheap, generates
a very small amount of electricity, and has a low degree of safety,
while A4 generates a large amount of electricity but costs more to
construct and has a high degree of safety. The ranking order is a
partial order because one cannot tell whether A2 or A4 is better. It
is given as:
2. Promethee II. By comparing F index alone, we get a linear

ranking order as:
It is desirable to have a linear ranking order. However, the case of

incomparability is ignored. Brans et al. pointed out that the
information provided by Promethee I is more realistic than that of
Promethee II because incomparability is also a very useful information
for decision making.
3. Promethee III. The interval of alternative Al is computed as

follows. First of all, the center of the interval is defined as:
F(l) ! F(l) 51 (-.567) -.113
The variance of Al is computed using

447
.0IS4.
The standard error can be obtained by taking the square root of u~ as
.124, i.e., u 1 = .124.
Assume that a = O.IS, the lower and upper bounds of the interval
are computed as:
(-.113) - (.IS)(.124) -.132,
(-.113) + (.IS)(.124) -.094.
Similarly, we calculate the intervals for all the other
alternatives. The results are summarized as:
Al A2 A3 A4 AS
Z -.132 -.021 -.169 -.033 .244

Y -.094 .031 -.112 .008 .279
The promethee III gives a ranking order as shown in the following
graph where dotted lines denote indifference.
..-----~ A2 Al
AS ---+
A4 A3
Note
1. The a level in Promethee III can be varied. For example, we
can set a = O.OS level and get new intervals for Al and A3 as [-.119,
448
-.106], and [-.150, -.131], respectively. Since Zl > Y3' we obtain a

preference of A1 over A3 . In other words, indifference among the
alternatives disappears rather quickly when the a level decreases.
2. The numerical example in this section is solved again using
the classical SAW method and TOPSIS, respectively. Equal weights
among attributes are assumed for the following computation.
(i) SAW method: The normalized final ratings for the
alternatives are:
.4715 .5943 .4794 .5831 .6726
(ii) TOPSIS: The distance to positive ideal solution, A*

(.2708, .5513, .1443, .1304, .0848, .6468), for each alternative can
be summarized as:
.8142 .8111 .5972 .8512 .3842
The distance to negative ideal solution, A- = (.5619, .3553, .7217,

.6323, .6786, .0647), for each alternative can be summarized as:
S~ S~ S~
.5173 .8371 .6928 .6853 .8784
The relative closeness to positive ideal for each alternative is then

computed as:
.3917 .4899 .5232 .4441 .6863
The ranking order is given as: A5 > A3 > A2 > A4 > A1 •

449
The same numerical example can be solved by both classical MADM

methods and the Promethee method. The results are different. It is
inconclusive that fuzzy method will qive better solution than the
classical methods will and vice versa. The obvious fact is that the
classical MADM methods are easier to apply then the Promethee method.
Aqain, the Promethee method may be theoretically sound but is too
costly to apply to MADM problems of any size.·
400
5.7.4 Takeda's Approach

Takeda [T2] proposed an interactive procedure for building fuzzy
outranking relations from which the decision maker's preference
structure could be extracted as a fuzzy multilevel graph, of which a
vertex corresponds to an alternative.
Similar to Roy's [Rll,R13] approach, Takeda used concordance and
discordance relations to obtain fuzzy outranking relation. The
difference is that Roy [RIll assumes that in concordance analysis
certain a priori weights about attributes are available. This
assumption is not valid for many cases because the OM may not be
certain about the weights. This makes Roy's approach hard to apply in
some decision making situations. To resolve the difficulty, Takeda
proposed an interactive procedure to obtain the weights of attribute.
The weights are used to obtain the concordance relation. combining
the concordance relation with the discordance relation, a fuzzy
outranking relation is determined.
Algorithm
For a MAOM problem, the OM must initially specify a set of
weights pertaining to the attributes. The discordance relation is
constructed first. The set of weights is modified iteration by
.iteration until the maximum (minimum) degree of credibility ~(~,AI)
of the outranking of ~ over Al is greater (less) than a specified

threshold value a, which should be between 0.5 and one. This will
make the concordance relation readily available. The resulting fuzzy
outranking relation is considered consistent with the OM's preference
structure. A multilevel graph can be drawn as the solution of the
MAOM problem.
The basic steps are summarized as follows.
451
step 1. Define the initial weight set, ~1. Arrange the
attributes according to the importance for the DM such that if h < j
Xj is at least as important as Xh . The weights must sum to one. Thus,
the initial weight set is defined as:
n
{Wlw1~· .. ~w.~ ... ~w ~O, L wk 1}.
J n k=l
step 2. Construct the discordance relation. The element of the
discordance relation is denoted by d kl , which can be computed as
= max
reD kl
where x kr and x lr are the performance data of alternatives Ak and Ai'
respectively, with respect to attribute Xr ; Dkl = {rlxkr<x lr }; and

d max = max IXk - xl I which guarantees that d kl e [0,1]. The
r lsk,lsm r r
number m is the number of alternatives in question.
A critical threshold value ~ must be set by the DM. the
discordance relation is then modified using
This equation indicates that if d kl > ~ then Ak can never outrank Ai.
Iteration No. L. set L 1.
step 3. Obtain a set of new weight sets, ~

LP , P = l, ... ,r, ... ,n,
where the jth element of vector wLr can be computed as:
l/r, if lsjsr,
{
0, if j>r.
452
step 4. Obtain the fuzzy outranking relation. For the weight
set ~Lr, a fuzzy outranking relation MLr can be constructed. The
element in MLr, ~r(~,Al)' is computed as
,
if dkl=O,
,
if dkl>O,
where Ckl {tIXkt~Xlt} is the concordance relation. Let
max
l:sr:sn
be the maximum and the minimum degrees of outranking of Ak over AI'
respectively. If there are pairs of alternatives (Ak,A l ) such that
mkl ~ a and ~kl < a,
where a is a predetermined threshold e [0.5,1], select any of them
(say (~,Al» and go to step 5. Otherwise, go to step 6.
step 5. Ask the OM: "Is Ak at least as good as AI? respond yes
or no." If yes, constraints
n
1: w.
j=l ]
are used to compute a new weight set wL . If no, the constraints
n
1: w. 1 and 1: Wt:s a-c (c > 0 and small)
t=l ] eC kl
453
are used to obtain a new weight set wL .

Set L = 2 and return to Step 3.
Step 6. Construct a fuzzy multilevel graph. Consider the
ordinary graph associated with the fuzzy outranking relation defined
in Step 4. Remove all arcs implied by transitivity and construct a
fuzzy multilevel graph in which the numbers associated with arc (k,l)
represents the maximum and minimum degrees of credibility of the
outranking of Ak over AI' i.e., mkl and ~kl.
Note
Takeda's method shares the same disadvantage found in the other
approaches in the fuzzy outranking methods category. That is, they
all require involved and complex computations. This pitfall will
definitely limit the applicability of Takeda's as well as other fuzzy
outranking approaches.
Due to the complex computation requirement of Takeda's approach,
we will not present a numerical example here. Interested readers

should refer to the original paper [T2] for detailed discussion on
Takeda's procedure.
454
5.8 Maximin Methods

The classical maximin method is used to select an alternative A*
such that
A * max min xi]' }, j l, . . . ,n; i 1, ... ,m.

i j
where xij's are in a common scale.

The term "maximin" signals the selection of the maximum (across
alternatives) of the minimum (across attributes) values. In this
situation, where the overall performance of an alternative is
determined by the weakest or poorest attribute, a OM would examine the
attribute values for each alternative, note the lowest value for each
alternative, and then select the alternative with the most acceptable
value in its lowest attribute. In general, this method would be
reasonable only if the OM is assumed to have a pessimistic nature in
the decision making situation (Hwang and Yoon [H13]).
The decision matrix for the maximin method is given as:
Xl Xj x
n
A1 IJ. 1 (X 1 ) IJ. 1 (X j ) IJ. 1 (X n )
A, IJ.i(X 1 ) IJ.i(X j ) IJ.i(X n )

1
Am IJ.m(Xl) IJ.m(X j ) IJ.m(Xn)
where IJ.i(X j ) e [0,1] is interpreted as how well Ai satisfies attribute

Xj . It represents a subjective judgement of the OM, and hence, is
fuzzy. The best alternative A* is determined as:
A * max min IJ.i(x],)}, j l, . . . ,n; i l, . . . ,m.

i j
455
In a classical MADM problem, values of different attributes may
be measured in different units. The values must be normalized such
that interattribute values are comparable. However, in a fuzzy case,
the values in the decision matrix are all given as degrees of "how one
alternative satisfies a certain attribute." There is no need for
normalization when the decision data are fuzzy. The decision data
~. (x.) in the decision matrix is referred to as the fuzzy singleton

1 J
[Z5].
The concept of maximin applied in a fuzzy environment was first
seen in Bellman and Zadeh [BS]. Although its original intention was
for general fuzzy decision making, this concept is readily applicable
to fuzzy multiple attribute decision making problems. Yager [Y4]
utilizes this concept and develops an algorithm for fuzzy MADM
problems with unequal weights.
5.S.1 Bellman and Zadeh's Approach
Bellman and Zadeh [BS] asserted that in the conventional approach
to decision making, the principal ingredients of the decision process
are (1) a set of alternatives; (2) a set of constraints on the choice
between different alternatives; and (3) a performance function which
associates with each alternative the gain or loss resulting from the
choice of that alternative.
In a fuzzy environment, the performance function may be replaced
by the concept "fuzzy goal". A fuzzy goal, G, may be represented by a
fuzzy set {(x'~G(X)) I xeU} where U is the universe of the fuzzy set G.
~G(x) is the membership function of the fuzzy goal and takes its
values in [0,1]. For example, the fuzzy goal expressed in words such
as "x should be in the vicinity of 15" may be represented by the
membership function (Bellman and Zadeh [BS]):

456
4 -1
(1 + (X - 15» E [0,1].
The x value that makes the highest ~G(x) value is the preferred one.
Clearly, the membership function serves the same purpose as a conven-
tional performance function. Furthermore, ~G(x) may be considered a

normalized performance function. such normalization provides a common
denominator for the various fuzzy goals and fuzzy constraints and
thereby makes it possible to treat them alike. This line of reasoning
explains why it is perfectly appropriate to regard the concept of the
"fuzzy goal"-rather than the performance function-as one of the major
components for decision analysis in a fuzzy environment.
Similar remarks can be made to fuzzy constraints. For example,
the constraints "x should be approximately between 2 and 10," may be
represented by (Bellman and Zadeh [B8]):
m -1
(1 + a(x - 6» E [0,1],
where a is a positive number and m is a positive even integer chosen
in such a way to reflect the sense in which the approximation to the
interval [2,10] is to be understood. If we set m = 4 and a = 5- 4 ,
then at x=2 and x=10 we have approximately ~C(X) = 0.71; while at x=l
and x=ll, ~C(x) = 0.50; and at x=o and x=12, ~C(x) is about 0.32.
The above definitions of goals and constraints in a fuzzy
environment (i.e., when both sets are fuzzy) make it appropriate to
treat the fuzzy goals and fuzzy constraints identically in the
formulation of a decision. By contrast, in the conventional approach
to decision-making, the use of Lagrangian multipliers and penalty
functions makes it apparent that there is an intrinsic similarity
between performance functions and constraints. This similarity is
made explicit in the formulation of fuzzy decision analysis.

457
Thus, a fuzzy decision may be stated as the fuzzy set D resulting
from the intersection of the goals and constraints. That is, given
the fuzzy goals, Gi , i = 1, ••. , m, and the constraints, Cj , j = 1,
... , n, we can determine the solution D, using
D (5.187)
Its membership function is defined as:
~G (x)n .•. n~G (x)n~c (x)n •.. n~c (x). (5.188)

1 m 1 n
The selection of the most appropriate x value for fuzzy set D is then
given as:
~ *(x) max ~D(x), x eKe U, (5.189)

D
where K is the set of points in U on which ~D(x) attains its maximum,
if it exists. Note that ~ *(x) is the optimal decision and any x in

D
the support of D* will be referred to as a minimizing decision.
For example, given the universe U {1, 2, .•• , 10} and the fuzzy
goals G1 and G2 , and fuzzy constraints C1 and C2 :
Table 5.17 Fuzz~ Goals and Constraints
x 1 2 3 4 5 6 7 8 9 10
~G (x) 0 .1 .4 .8 1.0 .7 .4 .2 0 0
1
~G (x) .1 .6 1.0 .9 .8 .6 .5 .3 0 0
2
~C (x) .3 .6 .9 1.0 .8 .7 .5 .3 .2 .1
1
~C (x) .2 .4 .6 .7 .9 1.0 .8 .6 .4 .2
2
458
The decision D {(X'~D(X»} may be obtained using Eq.(5.188) as:
1 2 3 4 5 6 7 8 9 10
° .1 .4 .7 .8 .6 .4 .2
° °
By Eq.(5.189), we obtain
~D*(X) = max [0,.1,.4,.7,.8,.6,.4,.2,0,0] .8
which indicates that when x = 5, the best decision has been reached.
We can conclude that the value 5 is the best selection among all
possible candidates, i. e., {1, 2, ... , 10}.
The computational procedure of Eq.(5.188) can be demonstrated by
calculating ~D(2) as:
~D(2) ~G (2) A ~G (2) A ~c (2) A ~c (2) .1.

1 2 1 2
Note that none of the ~D(X), Vx, have full membership, i.e., all ~D(X)
are less than one. This indicates that a certain degree of conflict
exists among goals and constraints and none of the x values fully
satisfy all of them.
Numerical Example (modified from Yager [Y4])

Three candidates are being considered for a job opening. The
criteria considered are: young age (Xl)' experience (X 2 ), communication

ability (X 3 ), and maturity (X 4 ). The decision matrix is:
Xl X2 X3 X4
[
.7 .3 .3 .5
1
.5 .8 .3 .1
.4 .6 .8 .2
where ~A (X 2 ) = .30 indicates that candidate Al satisfies the
1
experience requirement at a degree of 0.30; while ~A (X 3 ) = 0.8 shows
3
that candidate A3 'S communication ability is subjectively rated at .8.
The intersection of all attribues' values for each candidate is:
x
.3 .1 .2
For example, ~O(Al) = ~x (A l ) A ~x (A l ) A ~x (A l ) A ~X (A l ) = .3.

1 2 3 4
Since ~ .(x) =.3, the best choice is Al based on Eq.(5.l89).
o
Note
1. We include Bellman and Zadeh's approach in our study because
it is the pioneering work on fuzzy decision analysis. In their
approach, although the data are expressed by crisp, real numbers,
their intrinsic characteristics are fuzzy. The decision variable x in
this algorithm is treated as an alternative set. The values that x
may take are the alternatives to be evaluated. The decision problem
modeled by Bellman and Zadeh can be treated as a classical MADM
problem in that all decision data are crisp and the decision is to
select one out of several possible alternatives.
2. It has long been recognized that the applicability of the
maximin method in MADM is very limited (Hwang and Yoon [H13). The
maximin method utilizes only a small portion of the available
information in making a choice. It may happen that if an alternative
is clearly superior in all but one attribute which is below average,
another alternative which is only average on all attributes will be
chosen over it. We do not recommend this approach in any circumstance
except when the OM takes a defensive strategy, i.e., a pessimistic
attitude toward risk.
460
3. Hannan's [H4] criticism is that the values given for each
alternative-attribute combination are considered fuzzy membership
values, but since only single values are provided, they can also be
solved by classical MADM methods, such as SAW, TOPSIS, etc. In other
words, the information provided is no fuzzier than in the classical
MADM problem.
Yager [Y4) proposes a method based on the idea of assigning to
each of the attribute in a MADM problem a number indicating its
importance to the decision maker. The weights are calculated using
saaty's method [Sl). When the fuzzy decision takes place, the
performance data under all attributes for each alternative are raised
to their appropriate power and the alternative that satisfies
max min (5.190)

i i
is preferred.
Algorithm
A fuzzy MADM problem is given as:
Xl x. Xn
J
A1 ~1(X1) ~l(Xj) ~l(xn)
Ai ~i(X1) ~i(Xj) ~i(Xn)
A ~m (Xl) ~m(Xj) ~m (X n )
m
where ~. (x.) E [0,1] indicates how well alternative A1. satisfies

1 J
criteria X .. ~i(Xj) is a measure of subjective judgment. The
J
selection of the best alternative is done using the following steps.
461
step 1. Compute the relative importance for each criteria.

saaty's method is used to calculate weight, wj , Vj, (See section 5.2.2
for details).
Step 2. Obtain the weighted decision matrix. The weights are
used to modify the decision matrix. A weighted decision matrix is:
Xl Xj Xn
(W 1 ) (W j ) (Wn )
Al ~11 ~1j ~ln
{wI) (W j ) (wn )
Ai ~i1 ~ij ~in (5.191)
{wI) (W j ) (wn )
Am ~m1 ~mj ~~
where ~ij = ~i(Xj) and Wj is the weight obtained in Step 1.
step J. Select a compromise alternative. Ideally, our goal is

to select the alternative which has the highest membership values
with respect to all the criteria, Xj , Vj. However, this rarely
happens, because one alternative that has the highest membership value
with respect to Xl does not necessarily have the highest membership
value under other criteria. Since an alternative must be chosen, some
forms of compromise are to be made. In this case, Yager proposed the
use of the max and the min operators to select the best alternative.
The selected alternative is said to maximize the minimum membership
values over all the criteria, i.e.,
Wj
max [min ~ij ]. (5.192)
i j
462
Numerical Example (Yager [Y4])
Three candidates are being considered for a job opening. The
criteria considered are: young age (Xl)' experience (X 2 ), communica-
tion ability (X 3 ), and maturity (X 4 ). The decision matrix is:
Xl X2 X3 X4
[
.7 .3 .3 .5
1
.5 .8 .3 .1
.4 .6 .8 .2
The problem is solved by the following steps.
step 1. Obtain Weights. The reciprocal comparison matrix of the
attributes obtained after serious discussion with the OM is:
Xl X2 X3 X4
Xl 1 3 7 9
1
X2 3" 1 6 7
1 1
X3 1 3
"1 6
1 1 1
X4 1
9 "1 3"
The unit weight vector is obtained using the Eigenvector method:
~I = (.773,.400,.106,.053)
By multiplying ~ by 3, we obtain the weight vector:
~ (2.32,1.20,.32,.16)
step 2. Calculate the weighted decision matrix. By applying
the weights to raise the membership values to their appropriate power,
we can get a weighted decision matrix:

463
Xl X2 X3 X4
[
A1 .44 .24 .60 .90
.20 .76 .68 .69
1
A2
A .12 .54 .93 .77
For example, the element ~1(x2) of this matrix is obtained using
( ~12 ) 1.20 24
..
where 1.20 is the relative importance of attribute X2 .
step 3. Select an alternative. The minimum membership values for
each alternative are given as:
[.24, .20, .12], i 1, 2, 3.
The first candidate, A1 , has the highest ~min value and thus will be
selected.
Note
1. The effect of raising membership values to the w > 1 level is
to reduce the membership values of all the alternatives in a manner
that those which have higher membership values are reduced much less
than those which have smaller membership values. The effect of raising
membership values to the w < 1 level is to decrease the stringency of
the requirement of the fuzzy set, i.e., attributes that are less
stringent are considered less important.
The incorporation of weights in the algorithm makes linguistic
hedge interpretation much more meaningful. For example, a fuzzy set

F = "close to 1" may be denoted as:
x .1 .3 .5 .7 1.0
.1 .4 .6 .8 1.0
464
It can be raised to the power of 2 as:
.1 .3 .5 .7 1.0
~:2(X) I .01 .16 .36 .64 1.0
which may be treated as a fuzzy set "very close to 1". On the other
hand, F may also be raised to the power of 1/2 to represent another
fuzzy set "sort of close to 1" as:
x .1 .3 .5 .7 1.0
~ 5 (x) .32 .63 .76 .89 1.0
F·
2. Yager's approach suffers the same drawbacks as the classical

maximin method in that only a small portion of information is used to
determine the rank ordering of alternatives. Also worth emphasizing
is that even though the data are conceptually fuzzy they may be
regarded as normalized performance scores. Any classical MADM methods
(see Chapter 2) may give a better result than Yager's approach can.
4~
5.9 A New Approach to Fuzzy MADM Problems

After a systematic and critical study of the existing fuzzy MACM
approaches, we frequently encountered the following difficulties.
First of all, we found that the majority of the approaches, such as
Baas and Kwakernaak [B1], Buckley [B36,B39], Negi [N7], Roy [R11,R13],
etc., require cumbersome computations. As a result, none of them is
suitable for solving problems with more than ten alternatives
associated with more than ten attributes. That drawback certainly
limits their applicability to real world problems. Second, most
approaches, such as Baas and Kwakernaak [B1], Dubois and Prade [D28],
Bonissone [B26,B27], Laarhoven and Pedrycz [L1], etc., require that
the elements in the decision matrix be presented in a fuzzy format,
even though they are crisp in nature. Such an assumption violates the
original intent of fuzzy set theory--to cope with human subjective
judgment. If the data is precisely known, there is no subjectivity
involved in the decision problem. Such data should never be
represented in any fuzzy format. The conversion of crisp data into
fuzzy format will increase the computational requirements. This in
turn makes thses fuzzy methods cumbersome to use and incapable of
solving problems which contain more than ten alternatives and ten
attributes. Finally, some approaches, such as Yager [Y19], etc.,
assume that fuzzy datum can be represented by a fuzzy singleton (Zadeh
[Z5]), which is a fuzzy set of only one element with its membership
value e [0,1]. Such an assumption is not practical. As Hannan [H3]
pointed out if a fuzzy singleton is only a real number in [0,1] then
its fuzzy matrix is no fuzzier than a normalized decision matrix in
the classical MADM study domain. Yager's method is not fuzzy at all.
The proposed new approach is designed to avoid the aforementioned
difficulties so that MADM problems can be meaningfully and efficiently
solved in a fuzzy environment. The basic assumption of the proposed
466
approach is that the MACH problem may contain fuzzy and crisp data.
Furthermore, fuzzy data may be expressed in linguistic terms or in
fuzzy numbers.
The proposed approach is composed of two major phases. The first
phase converts fuzzy data into crisp scores. The fuzzy data can be
linguistic terms, fuzzy sets, or fuzzy numbers. If the fuzzy data
are linguistic terms, they are transformed into fuzzy numbers first.
Then all the fuzzy numbers (or fuzzy sets) are assigned crisp scores.
The result of the first phase is a decision matrix which contains only
crisp data. In the second phase, classical MACH methods (see Chapter
2) can be utilized to determine the ranking order of alternatives. In
general, mathematical computations are reduced to a minimum. A SO-by-
SO decision matrix can be solved as easily as a S-by-S one. The
easy-to-use and easy-to-understand characteristics of this new
approach make it valuable to management and system analysts.
S.9.1 converting Linguistic Terms to Fuzzy Numbers

A numerical approximation system is proposed to systematically
convert linguistic terms to their corresponding fuzzy numbers. It
contains eight conversion scales (see Figs. S.4Sa through S.4Sh). The
conversion scales are proposed by synthesizing and modifying the work
of Baas and Kwakernaak [B1], Bonissone [B27], Chen [C14], Efstathiou
and Rajkovic [E3], Efstathiou and Tong [E6], Kerre [K17], and Wenstop
[W6]. The linguistic terms used in our conversion scales are
summarized in Table S.18. Note that even when the number of terms
allowed is the same, the actual verbal terms may be slightly
different. It is also worth noting that even when the same term such
as "high" is used, the fuzzy numbers graphed are quite different from
figure to figure. This reflects the fact that the same linguistic
term may possess different meanings for different occasions.
467
",(x)
medium high
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0
0 .1 .2 .3 .4 .5 .6 .7 .8 .9
Fig. 5.45a Scale 1.
",(x)
low medium high
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 5.45b Scale 2.
",(x) very
1. low low medium
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 5.45c Scale 3.

468
/L(x) medium medium

low medium high
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 5.45d Scale 4.
/L(x) fairly fairly

low low high
1.
.9
.8
.7
.6
.5
.4
.3
x
.1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 5.45e Scale 5.
/L(x) very mol mol very

low low low medium high high
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 5.45f Scale 6.

469
medium high to
P.(x~Ow to medium
high v.high
1. ,V.lOW low low medium
.9 \
\ $~
~\ ~
.8 \
\
I~
I ,
\
.7 v. ~l lv. ~
.6 lo~\ lhi~h
.5
\ t ~
~
~
.4 ~
~
.3 ~
~
.2 ~
.1 ~
l
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 5.45g Scale 7.
P.(x)
low 10
v.low
medium
low
1
medium
medium
high
high 10
v.high
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
Fig. 5.45h Scale 8.

470
Table 5.18 Summary of Verbal Terms Used in the System
Scale 1 2 3 4 5 6 7 8
No. of
terms used two three five five six seven nine eleven
none yes
v. low yes yes yes yes yes
low-v. low yes yes
low yes yes yes yes yes yes yes
fairly low yes yes yes yes

mol low yes yes
medium yes yes yes yes yes yes yes
mol high yes yes
fairly high yes yes yes yes

high yes yes yes yes yes yes yes yes
high-v. high yes yes
v.high yes yes yes yes yes
excellent yes
v. : very med: medium mol: more or less
The principle of this system is simply to pick a figure that

contains all the verbal terms given by the OM and use the fuzzy
numbers in that figure to represent the meaning of the verbal terms.
As an example, assume the OM gives terms (medium, very high). Fig.
5.45c (Scale 3) should be used because Fig. 5.45c contains the terms
medium and very high. Or assume the terms (medium, high) are used by
a OM. Although all the scales contain these two terms, we choose the
simplest scale--Fig. 5.45a (Scale l)--to be our conversion scale. If
the terms (medium, high, excellent) are used, we find that Fig. 5.45h
(Scale 8) is the only figure which matches all the terms given by the
OM and should be used as the conversion scale.
471
The verbal terms used in our scales are in the universe U =

{excellent, very high, high to very high, high, fairly high, medium,
fairly low, low, low to very low, very low, none}. This universe of
verbal terms may be appropriate to describe the reliability of an
airplane but certainly is not suitable for describing the distance of
two places or two objects. Fortunately, our system does not confine
itself to that universe. Rather, the universe can be adjusted to fit
the nature of attributes used in a decision problem. For example, if
price is one of the attributes, the possible universe will be
{extremely expensive, very expensive, .•. , fair price, fairly cheap,
•.. , extremely cheap}. Or if size is one of the attributes, the
possible universe will be {extremely small, very small, •.. , medium,
medium large, ... , extremely large}. For any type of attributes, we
can always find a pair of words that represents extreme meanings, such
as high vs. low, good vs. poor, small vs. large, and so on. A set of
76 pairs of opposite words may be found in Osgood, suci, and
Tannenbaum [04]. For example, eight pairs of opposite words are shown
in Table 5.19. Ultimately, the proposed standard scales system is
capable of converting linguistic terms into fuzzy numbers in a
systematic manner. Such characteristics guarantee the consistency of
translating linguistic terms to fuzzy numbers.
The determination of the number of conversion scales in our
system is rather intuitive. Too few conversion scales provide no more
help than previous research results; while too many conversion scales
may make the system too complex to be practical. Miller [M16] pointed
out that "seven plus or minus two" represents the greatest amount
of information an observer can give us about the objects on the basis
of an absolute judgment. In our case, the number of scales used is
about right considering Miller's theory. One may argue that if "seven
plus or minus two" is valid, then Scale 8 which has 11 verbal terms
472
Table 5.19 Ei2ht Linguistic Universes

dis- wei- Hazar- tech- exper-
General price size tance 2ht dous nige ences
high expen- large far heavy danger ad- good
sive vanced
fair fairly
priced local
low cheap small local light safe basic poor
obviously violates that rule or is at the extreme of the rule. We

feel, however, that a detailed conversion scale is very important when
the OM is familiar with the decision problem context. In general,
there is no strong theorem(s) to support the argument that a system
with eight scales is better than systems with seven, nine, or any
other number of scales. We can only conclude that our scale system is
simple enough to be understood by the OM, easy to use by system
analysts, and yet thorough enough for real-world applications.
Example
The concept behind our system is to match the linguistic terms
under each attribute with one of the conversion scales, assuming that
the OM is not available for consulting (a non interactive situation).
Once a scale is found which matches all the linguistic terms used, it
will be employed to convert linguistic terms to their corresponding
fuzzy numbers. If more than one scale is found to match all the
linguistic terms used, the scale with the least number of terms is
used for conversions. The following example will illustrate our
concept and procedure.
Four commercial nonlinear programming software packages are
evaluated with respect to their ease of use, cost, editing facilities,
problem storage and retrieval, and supporting documents. The decision
matrix (modified from Golden and Wasil [GG]) is given as:
473
Table 5.20 Decision Matrix for Software Evaluation
Ease of Cost Editing Storage & Documen-

use retrieval tat ion
--
Al v. good 395 mol good v. fast clear
A2 fair 250 fair fast clear
A3 excellent 400 v. good medium very clea r
A4 excellent 195 poor slow vague
The proposed system is to convert all the linguistic terms into

fuzzy numbers column by column.
The first column contains three terms (very good,fair,excellent).
These terms are equivalent to the terms (very high,medium,excellent).
The only scale that contains such terms is Scale 8.
The third column contains four terms (more or less good, fair,
very good, poor). These terms are equivalent to "more or less
high,""medium," "very high," and "low," respectively. By observation,
we find that Scale 6 is an appropriate conversion scale.
The fourth column contains four terms (very fast, fast, medium,
slow). These four terms are equivalent to "very high," "high,"
"medium," and "low," respectively. By observation, we select Scale 3
as the conversion scale.
The last column contains three terms (clear, very clear, vague).
This three-term set is equivalent to (high, very high, low). By
observation, we select Scale 3 as the conversion scale. After
identifying the appropriate scale, we can simply convert the linguistic
terms into their corresponding fuzzy numbers column by column. Notice
that since the second column contains crisp data, the conversion system
will not be applied to it.
474
5.9.2 converting Fuzzy Numbers to crisp Scores
A fuzzy scoring method which converts fuzzy numbers to crisp
scores is proposed. The proposed scoring method is a modification of
Jain's [J2,J5] and Chen's [C12] fuzzy ranking approaches. The crisp
score of a fuzzy number M is obtained as follows (see section 4.8.3).

Given a maximizing set and a minimizing set as:
X,
{ (5.193)
0, otherwise
1 - x, 0 :s X :s 1
{ (5.194)
0, otherwise
The right score of M can be determined using:
(5.195)
The left score of M can be determined using:
(5.196)
Given the left and right scores of M, we can compute the total score
of Musing:
(5.197)
Example: Given two fuzzy numbers M1 and M2 as shown in Fig. 5.46,
their crisp scores are computed as follows.

475
P.(x)
Ml M2
1.
.9
.8
.5
.4
.3
.2
.1 .' .
0
0 .1 .2 .3 .4 .5 .6 .7 .8 .9
Fig. 5.46 The left and right scores of MI and M2 .
The maximizing set and minimizing set are defined as:
{
x, o " x " 1
/.lmax(X)
0, otherwise
and
{ o " x "
I - x, 1
/.lmin(X)
0, otherwise
Also known are the membership functions of MI and M2 :
x - .4
, 0.4 " x < 0.6
.2
1
/.lM (x)
1
.8 - x , 0.6 x 0.8
.2 " "
x - .6
.2
, 0.6
" x < 0.8
1
/.lM (x)
2 I - X
, 0.8 x 1.0
.2 " "
The right, left and total scores are:

476
1 .667 .500 .584

2 .833 .333 .750
For example, we compute the total score of M1 as:
~R(M1) sup ~max(X) A ~M (x)] .667,

x 1
~L(M1) sup [ ~min(x) A ~M (x)] .500,

x 1
[~R(M1) + 1 - ~L(M1)]
~T(M1) .584.
2
5.9.3 The Algorithm
The proposed algorithm will be presented as a series of steps.
step 1. Transform the linguistic expressions into fuzzy numbers
attribute by attribute. We will match all linguistic terms used
under an attribute with one of the conversion scales. If more than
one scale is identified, the one with the fewest linguistic terms will
be adopted. The process continues until all linguistic terms under
every attribute have been converted to fuzzy numbers.
step 2. Assign crisp scores to fuzzy numbers. This is done by
applying Eqs.(5.195), (5.196) and (5.197) on every fuzzy number in
question (see section 4.8.3).
step 3. Up to this point, we have transformed a decision matrix
with fuzzy elements into one with real numbers. It can be solved
using any classical MADM method (see Chapter 2). We have chosen to
use the TOPSIS method for the following numerical examples because of
its general and broad acceptability in many problem domains. TOPSIS
(Hwang and Yoon [H13]) will give cardinal order of the alternatives.
477
Numerical Example (Hwang [H15])

The Tisza River basin, having a total area of 130,000 sq. km.,
is shared by five countries. The region considered in Hungary is
surrounded by mountains, and covers about 30,000 sq. km. The main
river flowing through this study area is the Tisza River.
The development of the water resources system was started in the
middle of the last century by flood control and river regulation

works. Needs under present consideration include the further
development of the water supply for agricultural purposes, industrial
and domestic supplies, navigation, and riparian recreation. Careful
management of both the quantity and the quality of natural supplies
has been introduced in recent years. The decision problem is:
Table 5.21 The Tisza River Basin Problem

Attribute Alternative Systems
Xl Total cost 99.6 85.7 101.1 95.1 101.8

(in millions)
Probability of
water shortage 4 19 50 50 50
water quality v.good good poor v.good fair
Energy (reuse 0.7 0.5 0.01 0.1 0.01
factor)
Recreation v.good good fair poor poor

Flood good exl. fair exl. poor
protection (%)
Land and forest 90 80 80 60 70
use (1000 hal
Manpower impact v.good v.good good fair fair
X Environmental v.good good poor good fair

9
architecture
International v.easy easy ditf. ditf. ditf.
cooperation
Development v.good good fair poor fair
possibility
478
where v.good = very good, exl. excellent, v.easy very easy, and
diff. = difficult.
Assume the weights for all attributes are equal. The problem is
solved by the following steps.
step 1. We have identified that attributes X3 ' Xs ' X6 ' X8 ' X9 '
X10 ' and X11 contain linguistic terms. These will be converted to
fuzzy numbers.
The translation of linguistic terms to fuzzy numbers is easily
done using the numerical approximation system. We know attributes X3 '

X5 , X9 , and X11 can be translated using standard Scale 3. For X6 , we
match the linguistic terms with Scale 8. For X8 ' Scale 3 is used
again, while Scale 5 is used for the 10th attribute. The results are
shown in Figs. S.47, 5.48, 5.49, and 5.S0.
Step 2. To assign scores to fuzzy numbers, the fuzzy scoring
method presented in Section 4.8.3 is used. The results are:
x+
1
x+
2 X3
• X4 Xs
• •
X6 X7 •
X8
A1 99.6 4 .909 .7 .909 .667 90 .909
A2 85.7 19 .717 .5 .717 .954 80 .909
A3 101.1 SO .283 .01 .500 .500 80 .717
A4 9S.1 50 .909 .1 .283 .954 60 .500
AS 101.8 SO .SOO .01 .283 .333 70 .500
X9* X10 • X*
ll
.909 .917 .909 A1
.717 .7S0 .717 A2
.283 .416 .500 A3
.717 .2S0 .283 A4
.500 .416 .500 AS
where attributes with a "+" sign are cost attributes (the smaller the
better), while attributes with a II." sign are attributes being
assigned crisp scores.

479
,.,.(x) very
poor fair good good
1. ,
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
JlT(poor) .283 JlT(fair) = .500

JlT(good) .717 JlT(very good) = .909
Fig. 5,47 Linguistic ~ fuzzy set ~ crisp score for attributes

X3 , X5 , X9 , and X11 using Scale 3.
,.,.(x)
poor fair good exel
1.
,
.9
.8
.7
.6
.5
.4
.3
.2
.1
0
0 .1 .2 .3 .4 .5 .6 .7 .8 .9
JlT(poor) .333 JlT(fair) .500

JlT(good) .667 JlT(exel) .954
Fig. 5.48 Linguistic ~ fuzzy set ~ crisp score for attributes

X6 using Scale 8.
480
p,(x) very
fair good good
1.
.9
.8
.7
.6
.5
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
J..I T (fair) = .500 J..I T (very good) = .909

Xs using Scale 3.
p,lx) fairly very

difficult difficult easy easy
1. - .
.9
.8
.7
.6
.5-
.4
.3
.2
.1
0 x
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1.
J..I T (difficult) .250 J..I T (fairly difficult) .416
J..I T (easy) = .750 J..I T (very easy) = .917

X10 using Scale 5.
481
SteE 3. The TOPSIS method (see Chapter 2 for the algorithm) is
applied to the decision matrix shown in step 2. (Recall that the
weights are assumed equal for all attributes.)
step 3.1 Construct the normalized decision matrix as:
x+
1
x+
2 X3* X4 X5* X6*
A1 0.4599 0.0451 0.5753 0.8082 0.6870 0.4116
A2 0.3958 0.2141 0.4538 0.5773 0.5419 0.5887
A3 0.4669 0.5634 0.1791 0.0115 0.3779 0.3085
A4 0.4392 0.5634 0.5753 0.1155 0.2139 0.5887
A5 0.4701 0.5634 0.3164 0.0115 0.2139 0.2055
X7 x8* X9* *
X10 *
X11
0.5249 0.5566 0.6150 0.6812 0.6559 A1
0.4666 0.5566 0.4851 0.5572 0.5174 A2
0.4666 0.4391 0.1915 0.3090 0.3608 A3
1
0.3499 0.3062 0.4851 0.1857 0.2042 A4
0.4082 0.3062 0.3383 0.3090 0.3608 A5
step 3.2 Determine the positive and negative ideal solutions.
The positive ideal solution A* is
A* [0.3958, 0.0451, 0.5753, 0.8082, 0.6870, 0.5887, 0.5249,
0.5566, 0.6150, 0.6812, 0.6559].
The negative ideal solution is
A- [0.4701, 0.5634, 0.1791, 0.0115, 0.2139, 0.2055, 0.3499,

0.3062, 0.1915, 0.1857, 0.2042].
step 3.3 Compute the separation measures. Based on the
formula
482
11 2
{ ~ ( ) }1/2, ~ 1,2,3,4,5,
LVi]' - v]' ...
j=1
we can obtain the separation measure of each alternative to A* :
i 1 2 345
0.1884 0.4154 1.2890 1.2386 1.3134
Based on the formula
11 _ 2
s, { ~
LVi]'
(
- v]'
) } 1/2,
~-
j=1
we can.obtain the separation measure of each alternative to A-:
i 1 2 345
s, 1. 4314 1. 0845 0.3294 0.6339 0.2890
~-
step 3.4 Compute the relative closeness to the ideal
solutions, Ci , ~i. Using the formula
1,2,3,4,5,
We can obtain
i 1 2 345
0.8837 0.7230 0.2035 0.3385 0.1804
step 3.5 Rank the alternatives. According to the descending
order of Ci , the rank ordering of alternatives is:

Note
1. The existing fuzzy MADM methods share one or more of the
following pitfalls.
1. Size of Problem.
The first important issue is about the size of problems fuzzy
methods can handle. Most real world problems have a number of
alternatives ranging from 5, 10, 50, or 100 to over 1,000. The
attributes are most likely presented in hierarchical structure. For
instance, we may have a problem where alternatives are to be ranked
according to five major criteria. Each major criterion may contain up
to seven subcriteria. Each subcriterion may have up to four sub-
subcriteria. Thus, the number of attributes can easily go beyond 100.

The fuzzy MADM methods we reviewed so far consider less than 10
attributes. And the alternatives evaluated are less than 10. While
the algorithms of these methods may be valid, real world MADM problems
seldom come in such small dimensions. As a result, existing methods
cannot effectively (from labor and cost aspects) solve real world
problems.
The first example is the selection of the most promising
laboratory models of grain cleaning and separating equipment. The

attributes are detailed in Table 5.22. There are four main
attributes. Each has a different number of subattributes. Together
they form a MADM problem of 20 attributes. This example illustrates
the complexity of attribute structure which cannot be handled easily
by any existing fuzzy MADM methods. The next example is a college
admission problem summarized in Table 5.23. Notice that there are only
eight attributes, but the number of alternatives (applicants) can go
up to several thousand.
484
Table 5.22 criteria and Weighting Factors for selecting the

Most Promising Pieces of Laboratory Model Grain
cleaning and separating Equipment (Hwang[H15])
Main Criteria Sub-Criteria
I. Design (1) First Sieve

(2) Second Sieve
(3) Third Sieve
(4) Fourth Sieve
(5) Aspiration
(6) Sieve Cleaning
(7) Sieve Shaking
(8) Cylindrical Sieve
(9) Indented Cylinder
(10)Scalping unit
II.Performance ( 1) Efficiency (I)
(2) Efficiency (II)
(3) Operation
III.Applicability ( 1) Testing Time
(2) Test Sample
(3) Applicability
(4) Mobility
IV. Cost ( 1) Purchase Price
(2) Maintenance Cost
(3) operating Cost
2. Fuzzy vs. Crisp Data.
The majority of the fuzzy MADM methods assume the decision data
is fuzzy. The decision data regardless of whether they are fuzzy or
crisp must be denoted in fuzzy format. This practice is against the
original intention of fuzzy set theory which is to incorporate human
imperfect perceptions, subjective judgement, and imperfect information
into a decision model. When decision data can be precisely captured,
such data should not be forced into a fuzzy format because it will
complicate the already complex decision problem.

485
Table 5.23 Admission to a Private College (Klahr [K19a])
Attributes
Alumni interview average highest high high

Campus Interview 5 5 1 1
College Board Scores SOO SOO SOO SOO
Activ.:Leader/Member none/ none/ none/ none/
sevrl. some some none
High school Grades A A C C
High school Recommend good excellent excellent average
IQ 150 110 150 110
Rank in Senior Class top 5% top 5% top 5% top 5%
Attribute A6 A7 AS
Alumni Interview above average average highest high

Campus Interview 3 9 1 9
College Board Scores SOO SOO 400 400
Activ.:Leader/Member none/ some/ none/ some/
sevrl. sevrl. sevrl. sevrl.
High School Grades C C A A
High school Recommend average excellent average good
IQ 150 110 150 110
Rank in Senior Class top 33% top 33% top 5% top 5%
Attribute
Alumni Interview very high above average average
Campus Interview 9 1 9
College Board Scores 400 400 400
Activ.:Leader/Member none/ none/none none/sevrl.
sevrl.
High School Grades A A C
High School Recommend good superior good
IQ 150 110 110
Rank in Senior Class top 33% top 33% top 33%
486
Most real world problems contains a mixture of fuzzy and crisp

data. Example in Table 5.23 is a typical real world problem in which
both fuzzy and crisp data coexist. When precise information can not
be obtained or is too costly to assess, the imperfect information must
be noted in fuzzy format; however, when information is easily
measurable or accessible, the information must be coded in crisp
notation. We can easily find decision problems in which data are
either half crisp and half fuzzy, a majority crisp and a small amount
fuzzy, or a large portion fuzzy and a small amount crisp. Seldom do
we see a problem presented in pure crisp or pure fuzzy notation.
3. Fuzzy singleton.
A fuzzy singleton is a fuzzy set which contains only one element.
For example, the fuzzy set M = {(x,0.7)} is a fuzzy singleton because
it has only one element x with a membership value of 0.7. In MADM
decision analysis, we do not see how a fuzzy singleton differentiates
itself in a decision matrix from a real number in [0,1].
For example, Yager [Y3] uses an example to show a fuzzy singleton
application. The example involves the selection of a faculty member.
Xl X2 X3
A1 .5 .5 .2
.7 .4 .01
[
A2
A3 .3 .8 .6
A4 .6 .4 .9
1
In this problem, Xl = candidate should be young, X2 = should be
experienced, X3 = be able to communicate. The datum may be explained
as: "Candidate Ai satisfies attributes Xj at degree of a e [0,1]".
Theoretically, the data is fuzzy. But from an operational aspect, the
decision matrix is no more than a classical MADM model which can be
easily solved by an appropriate classical MACM method.
VI. CONCLUDING REMARKS
This study is a sequel to our previous works on "Multiple

Objective Decision Making--Methods and Applications" [H12], "Multiple
Attribute Decision Making--Methods and Applications" [H13], and "Group
Decision Making Under Multiple criteria--Methods and Applications"
[H14]. It gives a state-of-the-art survey of the existing methods
which solve fuzzy MADM problems and their applications. It also
provides readers with a capsule look into the existing methods, their
characteristics, and their applicability to the analysis of fuzzy MADM
problems. Many diversified methods are reviewed thoroughly and
critically, and classified systematically. We also present a new and
practical fuzzy MADM approach.
6.1 MADM Problems and Fuzzy Sets

As we presented in Chapter I (INTRODUCTION), a MADM problem can
be concisely expressed in matrix format as:
Xl X2
A1 x 11 x 12
x 21 x 22 (6.1)
[
A2
D
~ xm1 xm2 x~ 1
where Ai represents an alternative, Xj denotes an attribute, and
x ij is the performance rating of alternative Ai with respect to
attribute Xj • x ij is also referred to as the decision data.
In most of the real-world problems, some of the decision data x ij
can be precisely assessed while others cannot. We use real (crisp)
numbers to represent data which can be precisely measured. For those
data which cannot be precisely assessed, we use Zadeh's [Zl] fuzzy
sets (numbers) to denote them. The use of fuzzy set theory allows us
to incorporate unquantifiable information, incomplete information,
nonobtainable information, and partially ignorant facts into the
decision model.
6.2 On Existing MACM Solution Methods

Basically, a MACM problem can be tackled in two phases.
Phase I: The aggregation of the performance ratings with respect to
all attributes for each alternative. Phase II: The rank ordering of
alternatives according to the aggregated scores.
6.2.1 Classical Methods for MACM Problems

Classical (crisp) MADM methods solve problems in which all
decision data are assumed to be known and must be represented by crisp
numbers. The methods are to effectively aggregate performance scores.
Once the aggregated scores are determined, the ranking order of
alternatives can be automatically decided. A thorough and systematic
survey of the classical MACM methods is presented in Hwang and Yoon
[H13] (see Fig. 1.2).
6.2.2 Fuzzy Methods for MADM Problems

Fuzzy methods have been proposed specifically for MADM problems
which contain fuzzy decision data. The introduction of fuzzy data
into a MADM problem complicates the decision analysis in the following
ways. First of all, fuzzy data are operationally difficult to
manipulate; hence, they greatly increase the computational require-
ments. Secondly, whereas we can clearly say that a real number is
better than, worse than, or equal to another real number, we cannot
crisply distinguish which fuzzy numbers are better (or worse).
Therefore, we have difficulty in judging the preferred alternatives
because all aggregated scores are fuzzy data.
489
Researchers in this field have developed methods to tackle these

complexities. These methods focus on either phase I problems, phase
II problems, or both phases of MADM problems. In this monograph,
methods for solving phase II problems are referred to as "fuzzy ranking
methods," and methods for solving phase I problems and/or solving both
phases of MADM problems are referred to as "fuzzy MADM methods."
6.2.2.1 Fuzzy Ranking Methods

About two dozen fuzzy ranking methods have been developed in the
past. They have been reviewed in the works of Bortoland and Degani
[B34], Freeling [F1], Li and Lee [L7], Tseng et ale [T13,T14], and
Zimmermann [Z31]. By combining the many fine classification ideas
from the previous works and our intensive study, we have come up with
a system of classifying the existing fuzzy ranking methods. The
advantages and disadvantages of each method are discussed. Simple
numerical examples are used to facilitate the explanation of basic
concepts and computational procedures of these methods. A taxonomy of
fuzzy ranking methods is presented in Fig. 1.3.
6.2.2.2 Fuzzy MADM Methods

A system of classifying over a dozen fuzzy MADM methods is
presented in this study (see Fig. 1.4). These methods have been
developed by various researchers and presented in diverse publications.
The concepts, the computational procedures, and the characteristics of
each method are discussed and critically reviewed. The computational
procedure of each method is illustrated by solving a simple numerical
example.
6.3 critiques of the Existing Fuzzy Methods

The study of fuzzy methods for MADM problems is still in its
infancy and still has a lot of room for improvement. We have assessed
400
the disadvantages of the existing methods from a pratical point of

view as follows:
6.3.1 Size of Problem

The first important issue deals with the size of problems that
fuzzy methods can handle. The size of a MACM problem is measured by
m (the number of alternatives) and n (the number of attributes) which
are shown in Eq.(6.1).
Most of the real-world problems have a number of alternatives
ranging from 5, 10, 50, 100, 1,000, or over thousands. The
attributes may well be represented in a hierarchical structure (as
shown in Fig. 2.1a), in which we may have seven major criteria. Each
major criterion may contain up to seven subcriteria. Each sub-
criterion may have up to another seven sub-subcriteria. In this
case, we will have a total of about 350 attributes.
Almost all the existing fuzzy MACM methods reviewed are cumbersome
to use, even for a problem which has less than 10 alternatives (m < 10)
and less than 10 attributes (n < 10). In reality, real-world problems
rarely come in such small dimension. While these methods may be
theoretically sound, they cannot effectively solve most real-world
problems.
6.3.2 Fuzzy vs. Crisp Data

Most real-world problems contain a mixture of fuzzy and crisp
data. However, some of the fuzzy MACM methods assume the decision
data are fuzzy or at least have to be represented in fuzzy format.
That is, crisp data must be fuzzified in order to apply these methods.
This requirement not only complicates the computational procedure but
also violates the original intention of fuzzy set theory which is to
incorporate unquantifiable information, imperfect information,
491
nonobtainable information, and partially ignorant facts into the
decision model. When decision data are precisely known, they should
not be forced into a fuzzy format in the decision analysis.
6.4 A New Approach to Fuzzy HADM Problem Solving

since the existing fuzzy MADM methods are cumbersome to use and
difficult to understand, a new and practical method for fuzzy HADM
problem solving has been proposed in this study. It is easy to
understand and is capable of solving large size real-world problems
which contain a mixture of fuzzy and crisp data.
The proposed new approach tackles a fuzzy HADM problem in two

steps. The first step deals with converting the fuzzy decision data
into crisp scores. The fuzzy data may be linguistic terms or fuzzy
numbers. If the fuzzy data take linguistic terms as their values,
they are transformed into fuzzy numbers first. Then by using fuzzy
ranking methods, the fuzzy numbers of each column (attribute) in the
decision matrix are assigned crisp scores. The result of the first
step is a decision matrix which contains all crisp real numbers. In
the second step, an appropriate classical MADM method (see Chapter 2)
can be applied to determine the rank ordering of alternatives.
6.4.1 Semantic Modeling of Linguistic Terms
The vague, fuzzy information may be frequently expressed in a
linguistic expression (term). Linguistic terms are not mathematically
operable. To cope with that difficulty, each linguistic term is
associated with a fuzzy set or a composition of fuzzy sets which
represents the meaning of that linguistic term. Since the meaning of
each linguistic term varies from circumstance(s) to circumstance(s),
to assign a fuzzy set(s) to a linguistic term is a constant challenge.
The existing works on this topic are few and seem quite arbitrary. We
4~
have proposed a standard conversion system, obtained through
synthesizing and modifying the existing works, which may be considered
as the first step to a systematic and rational approach to associate
fuzzy sets with linguistic terms.
The system contains eight scales. There are generic verbal terms
(ranging from 2 to 11) in the system where Scale 1 contains only two
verbal terms and Scale 8 contains 11 verbal terms. The meaning of
each generic verbal term is represented by a fuzzy set (number). The
principle of this system is to pick a scale that matches all the
linguistic terms in a column (attribute) of the decision matrix and

use the fuzzy sets on that scale to represent the meaning of these
linguistic terms. The system is used on all columns which contain
linguistic terms, one by one.
6.4.2 Fuzzy scoring System
After all the linguistic terms have been converted to fuzzy sets
(numbers), a decision matrix which contains either fuzzy numbers, real
numbers, or both is obtained. The fuzzy numbers are to be converted
to numbers E [0,1) which are called the fuzzy scores. This is done
through the fuzzy scoring method.
The fuzzy scoring method is meant to compare fuzzy numbers.
First, a total score is determined for each fuzzy number. The fuzzy
number with a larger total score is considered better. In our fuzzy
scoring system, the total scores are not compared, rather, they
replace the fuzzy numbers in the decision matrix.
6.4.3 The Solution
At this stage, the decision matrix contains only crisp data.
That is, we have transformed a fuzzy MADM problem into a classical
one. Any appropriate classical MADM method can be applied to
determine the ranking order of the alternatives.

493
6.4.4 The Advantages of the New Approach

The advantages of this new approach over the existing ones are:
1. Classical MACM methods have been used in solving nonfuzzy MADM
problems and the majority of them are capable of handling large size
MADM problems. Our new approach has successfully extended that
ability to the fuzzy problem domain.
2. Our approach allows MADM problems to take data in the forms of
linguistic terms, fuzzy numbers, and/or crisp numbers. This allows

more realistic decision models to be created than the existing methods
do.
3. The existing methods are very cumbersome to use and are not
capable of solving many large size real-world problems. Our new
approach eliminates that difficulty.
6.5 Other Multiple criteria Decision Making Methods

Other multiple criteria decision making methods include Multiple
Objective Decision Making methods and methods of Group Decision Making
under multiple criteria. They will be briefly discussed in this
section.
6.5.1 Multiple Objective Decision Making Methods

Throughout this monograph, we assume that there are predetermined
alternatives to be evaluated/ranked/prioritized with respect to
multiple, usually conflicting attributes. In the study of decision
making in a complex environment, the term "Multiple Objective Decision
Making" is also frequently cited. It refers to the decision situation
in which an infinite number of alternatives which are implicitly
defined by constraints are evaluated. MODM methods thus address design
problems rather than choice problems. The terms "Multiple Attribute
Decision Making" and "Multiple Objective Decision Making" may be
494
designated by the term "Multiple Criteria Decision Making" (MCDM).

Literature on MODM methods and applications for single decision
maker problems has been studied in [H12] and has been systematically
classified as shown in Fig. 6.1.
6.5.2 Methods of Group Decision Making under Multiple criteria

Moving from a single decision maker to a multiple decision maker
setting introduces a great deal of complexity into the analysis. The
problem is no longer the selection of the most preferred alternative
among the nondominated solutions according to one individual's
preference structure. The analysis must be extended to account for
the conflicts among different interest groups who have different
objectivs, goals, criteria, and so on.
Due to the broad applicability of group decision making under
multiple criteria, group decision making methods have evolved from
research fields such as utility theory, social choice theory, theory
of voting, general game theory, expert evaluation analysis, economic
equilibrium theory, etc. A taxonomy of these methods has been
developed as shown in Fig. 6.2 (Hwang and Lin [H14]).
6.5.2.1 Social Choice Theory

voting is a group decision making method in a democratic society,
an expression of the will of the majority. It is a multiple criteria
decision making process whenever a voter casts a vote to select a
candidate or alternative policy. The candidate's qualifications may
be judged by multiple criteria such as trustworthiness and/or honesty,
capability, general political stance --conservative, moderate, or
liberal--and position on specific issues. These criteria are
summarized, in a voter's mind, to be a value function (utility
function), but in the counting of votes cast, the multiple criteria do
not appear explicitly.
I. Stage at Which II. Type of III Major Classes
Information is needed Information of Methods
1.1.1 Global Criterion Method
1. No Articulation 1.1.2 TOPSIS for MODM
of Preference
Information
2.1 Cardinal 2.1.1 Utility Function
Information 2.1.2 Bounded Objective Method
2. A Prior
Articulation of 2.2 Ord inal and 2.2.1 Lexicog raph ic Method
I Preference Cardinal 1 - 2.2.2 Goal Programming
Information r---.. Information 2.2.3 Goal Attainment Method
3.1.1 Method of Geoffrion

and Interactive
Multiple Goal Programming
Objective A3.1 Explicit 3.1.2 Surrogate Worth
Trade-off Trade-off Method
Decision r- 3.1.3 Method of Satisfactory
Making Goals
3.1.4 Method of Zionts-Wellenius
~
3.2.1STEM and Related Methods
3. Progressive 3.2.2SEMOPS and SIGMOP Methods
, Articulation of 3.2.3 Method of Displaced Ideal
Preference 3.2.4GPSTEM Method
Information N 3.2 Implicit 3.2.5Method of Steuer
(Interactive Methods) Trade-off J- (Interactive MOLP Method)
3.2.6 ISGP (Interactive
Sequential Goal Programming
3.2.7 MOST (Multiple Objective
Superior Technique)
\ 4. Posterior
Articulation 4.1.1 Parametric Method
Preference Information ~l4.1 Implicit 4.1.2 -Constraint Method
(Nondominated Solutions Trade-off }-- 4.1.3 MOLP Methods
Generation Method) 4.1.4 Adaptive Search Method
Fig. 6.1 A taxonomy of methods for classical MODM problems (Hwang and Masud, H12J.
Types of fields Major class of methods and forms
Social choice theory

Social choice function
Social welfare function
Creative confrontation
and stimulating ideas
Expert judgement!
group participation Systematic structuring
~
Simulation
Implementation and
controlling
The normal form payoff function

Game theory
The characteristic function form
Fig. 6.2 A taxonomy of methods for group decision making

[Hwang and Lin, H14J.
497
However, voting and counting in an electoral system are two
different processes, as they are not performed by the same people.
The voting process is carried out by all the voters who may be in a
country or on a committee. The counting process, on the other hand,
is carried out by a small group of selected workers, under expert
direction and subject from start to finish to the strictest possible
supervision and checking. Obviously the voting process should be kept
reasonably simple and straightforward, so as to cause no difficulty to
the general run of electors. On the other hand, the primary concern
of the counting process is not simplicity but accuracy and
effectiveness.
Let us use an example to illustrate the voting and counting
processes. Suppose an electoral body of 60 individuals voted for an
office holder from a field of three candidates a, b, and c in the
following manner:
23 have given the order a pcp b

19 have given the order b pcp a
16 have given the order c p b P a
2 have given the order c p a p b
The results depend on the methods of voting being employed. Any of
the three candidates could be elected: candidate a by the plurality
method, candidate b by the second ballot of the majority representa-
tion system, and candidate c by the Condorcet principle. This is a
clearly undesirable situation since each individual would prefer the
method of voting to elect his candidate. In another example, the
simple majorities could be intransitive in the situation when x beats
y, y beats z, and z beats x. This outcome is called the paradox of
voting. The paradox was known and developed by the Marquis de

498
Condorcet in the eighteenth century, and is referred to as the
Condorcet effect.
How do we solve this Condorect effect problem? It is a vital
question when dealing with methods of election, particularly the
social choice theory. The question is: What kind of decisions are
necessary and sufficient in light of the real world to enable us to
derive social orderings of the relevant candidates? The study of the
problem, so called the counting process, has been classified in two
ways: (1) the social choice function, and (2) the social welfare
function. Fig. 6.3 presents approaches in social choice theory,
which includes voting, social choice function and social welfare
functions.
6.5.2.2 Experts Judgement/Group Participation

The problem of group decision making can be broadly classified
into two categories in this field: experts judgment, and group
participation. The experts judgment process entails making a decision
by inventing a new alternative. specifically, it is concerned with
forecasting, and involves constructing supplemental objects which may
be new designs or new technical solutions. On the other hand, the
group participation process entails groups which have common
interests, such as a community or an organization, making a decision.
Fig. 6.4 illustrates phases, activities, and methods of experts
judgment/group participation.
6.5.2.3 Game Theory
Game theory is a mathematical technique used in analyzing
conflict-of-interest situations, and may be classified into two
different forms--normal form and ch~racteristic function form. The
normal form of the game includes a specified number of players, the

SOCIAL CHOICE THEORY
1. VOTING 2. SOCIAL CHOICE FUNCTION 3. SOCIAL WELFARE FUNCTION
1.1 Nonranked Voting System Condorcet Principle 3.1 Arrow's Conditions for Social Welfare Function
1.1.1 One Member Elected From Two Candidates 2.1 Condorcet's Function 3.2 Arrow's Possibility Theorem for Two Alternatives
1.1.2 One member Elected From Multi-Candidates .
(a) The First Past the Post System 2.2 Borda's Function 3.3 Arrow's General Possibility Theorem
(b) Majority Representation System
Repeated Ballots 2.3 Copeland's Function 3.4 Black's Single-Peaked Preferences
The Second Ballods
2.4 Nanson's Function 3.5 Bowman and Colantoni's Approach
1.1.3 Election of Two or More Members
1.1.3.1 The Single Non-Transferable Vote 2.5 Dodgson's Function 3.6 Goodman and Markowitz's Approach
1.1.3.2 Multiple Vote
1.1.3.3 Limited Vote 2.6 Kemeny's Function 3.7 Cardinal Social Welfare Function
1.1.3.4 Cumulative Vote
1.1.3.5 List Systems 2.7 Cook and Seiford's Function 3.7.1 Value Function for Certainty Case
(a) Highest Average $
(b) Greatest Remainder 2.8 Fishburn's Function 3.7.2 Utility Function for Uncertainty Case
1.1.3.6 Approval Vote
2.9 Eigenvector Function 3.7.2.1 Additive Group Utility Function
1.2 Preferential Voting System
2.10 Bernardo's Assignment Approach 3.7.2.2 Multiplicative Group Utility Function
1.2.1 Simple Majority Decision Rule
1.2.1.1 Two-Alternative Case 2.11 Cook and Seiford's Ordinal 3.7.3.1 Gymnastics Competitions
1.2.1.2 More Than Two Alternatives Case Intersection Method
(a) Paradox of Voting 3.7.3.2 Extended Contribution Rule (ECR) Method
(b) The Condorcet Effect
1.3 Nonminority Rule
Fig. 6.3 Approaches in Social Choice Theory [Hwang and Lin, H14].
Phases Extraction Exploration Selection Execution
/",/,,~t~t
Idea Issue Problem Problem Planning
Stimulation Clarification Structuring Solving Solution
Activities
t :> <: t ::::>c< t>< t ~Ietentlng
Creative Polling of Systematic Simulation and
Confrontation Experts/participant Structuring A Controlling
_- Brainstorming Surveys - Morphological - KSIM - Gantt chart

1
-- Trigger Method Delphi method analysis - aSIM - PERT

-- SIL Method Conferences - ISM - XIMP - CPM
-- Gordon's Approach SPAN technique - DEMATEL - SPIN - PPBS
-- Bralnwriting - Cognitive map - DYNAMO - Delta chart
-- Method 6-3-5
Methods -- Pin-card method
,-- Gallery method
-- BBB method
8
-- Collective notebook
-- Nominal Group
technique (NGT)
- Synectics
Fig. 6.4 Phases, activities, and methods of experts judgement and

group participation (Hwang and Lin, H14),
501
number of alternative strategies available to each player, and the
payoff function of the game. The characteristic function involves
coalitions. In normal form, however, coalitions are nonexistent. The
three basic abstract forms and approaches of game theory are shown in
Fig. 6.5.
Extensive form of game - Game tree
Parato optimal set

Game Theory Normal form of
g~i Nash-Harsanyi arbitration
solution
Compromise solution
Core concept
Characteristic
Function form of
game
-E Shapley value
Parametrical approach
Fig. 6.5 The three basic abstract forms and approaches of

game theory [Hwang and Lin, H14].
6.6 On Future Studies
We have identified some research areas that require special
attention in the future. These research areas include semantics of
linguistic term, fuzzy ranking methods, fuzzy MADM methods, and MADM
expert decision support system.
6.6.1 Semantics of Linguistic Term
The meaning of a linguistic term can be expressed by a fuzzy
set. Currently, the assignment of a fuzzy set to a linguistic term
seems arbitrary. The conversion system we proposed in Section 5.9.2
is an attempt to assign fuzzy sets to linguistic terms in a systematic
manner. This system has not been tested vigorously, nor theoretically
proved. Further study of this topic is desirable.

502
6.6.2 Fuzzy Ranking Methods
Many aspects of fuzzy set theory applications require the
comparison of fuzzy numbers (sets). The comparison of fuzzy utilities
to determine the preferred solution is only one of them. The

importance of the fuzzy ranking methods cannot be overemphsized.
A fuzzy number (or a fuzzy set) may be characterized by the
following factors: (1) the shape, (2) the height, (3) relative
location on x-axis and y-axis, (4) the spread, (5) the centroid point,
(6) fuzzy mean, (7) proportion to fuzzy ideals, and so on. As we
mentioned earlier, none of the existing methods are perfect.
Specifically, they may result in counter-intuitive solutions. This
may be attributed to the fact that they use only one or two factors to
compare fuzzy numbers.

It is worth emphasizing that even though almost all the existing
ranking methods are not perfect, they have shown the process of human
efforts to find ways to solve problems. There are always some good
points produced by each method. We wish to recognize and continue the
effort in and the process of improving these methods. Flawless
ranking methods may possibly be obtained by combining some of the good
points of each of these methods into one algorithm.
Another approach to improve the performance of the existing
ranking methods is to treat the problem of ranking fuzzy numbers as a
MADM problem. The many factors that characterize fuzzy numbers may be
considered as the attributes. The fuzzy numbers in comparison are

seen as the alternatives. The decision data are real numbers. How to
solve this type of MADM problem is not difficult. But to identify the
proper attributes and to effectively determine the decision data is
not straightforward. The need for further study in this direction
is obvious.
503
6.6.3 Fuzzy MADM Methods
We will state once more that a MADM problem can be concisely
expressed in a matrix format of Eq. (6.1), where there are m
alternatives and n attributes. In most of the real-world problems,
the number m can be 5, 10, 50, 100, 1000, or over thousands, and the
number n can be up to around 350 (see the hierachical tree presented
in Fig. 2.1a). The decision data x ij can take fuzzy or crisp data as
its value. A fuzzy data can be expressed in linguistic term, fuzzy

number, or fuzzy set.
The existing fuzzy MADM methods are complex and difficult to
apply to most large size real-world problems. A good and simple
method which is conceptually easy to understand and practically
capable of solving real-world problems is desirable. The proposed new
method to fuzzy MADM problem solving is an attempt toward that goal.
It is hoped that researchers in this field will make further
investigations to meet that goal.
6.6.4 MADM Expert Decision Support Systems
Currently, there is no single method which is good for solving
all the different types of decision problems. In this monograph, we
have frequently suggested that any "appropriate" method can be applied
to solve a MADM problem. In reality, the so-called "appropriate"
method cannot be identified with ease. Most of the time, people
simply apply the method that they are familiar with (or willing to
use). This practice will often result in an ad hoc decision.
To resolve this problem, we must carefully choose the decision
making technique which is appropriate for the particular problem on
hand. This argument is supported by the work of H.A. Simon [S17]
which states that we must let the problems which we are trying to
solve determine the methods we need to use rather than having the
5~
methods (or tools) determine the problems we are willing to tackle.
Some preliminary studies on picking appropriate classical MADM methods
have been conducted during the past decade. For example, Hwang and
Yoon [H13] have proposed a general guideline on when to use which
classical MADM method; and Hwang [H15] has conducted an extensive
study on classifying the types of MADM problems and identifying the

appropriate classical MADM method(s} for each problem type.
This study, combined with our previous books, is a first step
toward this goal. We have thoroughly and systematically surveyed an
entire armory of both the classical and fuzzy MADM methods. The next
step is to develop an Expert Decision Support System that can assist
the system analyst to pick an appropriate method for his/her decision
problem. This will require a collection of a variety of problems,
which are then systematically classified into different types.
Inference rules must be built to match the problem types and the
decision methods.
VII. BIBLIOGRAPHY
A1. Adamo, J.M., Fuzzy decision trees, Fuzzy Sets and Systems,
Vol.4, No.3, 1980, pp. 207-220.
A2. Aguilar-Martin J., and R.L. Mantaras, The process of

classification and learning the meaning of linguistic
descriptors of concepts, In: Approximate Reasoning in
Decision Analysis, Gupta and Sanchez (eds.),
North-Holland, 1982, pp. 165-175.
A3. Alexeyev, A.V., A.N. Borisov, V.I. Glushkov, O.A. Krumberg,

G.V. Merkuryeva, V.A. Popo~, and N.N. Slyadz, A linguistic
approach to decision-making problems, Fuzzy Sets and Systems,
Vol. 22, 1987, pp.25-42.
A4. Alley, H.A., C.P. Bacinello, and K.W. Hipel, Fuzzy set
approaches to planning in the grand river basin,
Advances in water Resources, Vol. 2, 1979, pp. 3-12.
A5. Alsina, C., E. Trillas, and L. Valverde, On some logical

connectives for fuzzy set theory, Journal of Mathematical
Analysis and Applications, Vol. 93, 1983, pp.15 26.
A6. Alsina, C. On a family of connectives for fuzzy sets,

Fuzzy Sets and Systems, Vol. 16, 1985, pp. 231-235.
A7. Asai, K., and H. Tanaka, Applications of fuzzy sets theory
decision-making and control, J. JAACE, Vol. 19, 1975,
pp.235-242.
A8. Asai K., H. Tanaka, T. Okuda, Decision-making and its

goal in a fuzzy environment. In: Fuzzy Sets and Their
Applications to cognitive and Dec1sion Processes,
L.A. Zadeh, K.S. Fu, K. Tanaka, and M. Shimura (eds.),
Academic Press, New York, 1975, pp. 257-277.
A9. Asai, K., H. Tanaka, and T. Okuda, On discrimination of

fuzzy states in probability space, Kybernetes, Vol. 6, 1977,
pp.185-192.
A10. Asse, A., P. Mangin, and D. Willaeys, Assisted diagnosis

using fuzzy information: Method of inverting equations of
fuzzy relations with ~ fuzzy sets, In: Analysis of Fuzzy
Information, Vol. II - Artificial Intelligence and Decision
Systems, J.C. Bezdek (ed.), CRC Press, Boca Raton, Flor1da,
1987, pp. 153-162.
B1. Baas, S.M., and H. Kwakernaak, Rating and ranking of

multiple aspect alternative using fuzzy sets, Automatica,
Vol. 13, 1977, pp. 47-58.
B2. Baldwin, J.F. and N.C. Guild, A model for multi-criterial

decision-making using fuzzy logic, In: Proc. Workshop of Fuzzy
Reasoning, Queen Mary College, University of London, London,
England, 1978.
B3. Baldwin, J.F., and N.C. Guild, Comparison of fuzzy sets
on the same decision space, Fuzzy sets and Systems, Vol. 2,
No.2, 1979, pp. 213-231.
B4. Baldwin, J.F., Fuzzy logic and fuzzy reasoning, In:

Fuzzy Reasoning and its Applications, B.R. Gaines (ed.),
B5. Baptistella, L.F.B., and A. Ollero, Fuzzy methodologies for

interactive multicriteria optimization, IEEE Trans. On
System, Man, and Cybernetics, Vol. SMC-10, No.7, 1980,
pp.355-365.
B6. Basu, K., Fuzzy revealed preference theory, Journal of

Economic Theory, Vol. 32, 1984, pp. 212-227.
B7. Bedrosian, S.D., A role for fuzzy concept in interactive

decision-making, In: Management Decision Support Systems
using Fuzzy sets and Posslbility Theory, J. Kacprzyk (eds) ,
Verlag, 1985, pp. 38-47.
B8. Bellman, R., and L.A. Zadeh, Decision making in a fuzzy
environment, Management Science, Vol. 17B, No.4, 1970,
pp. 141-164.
B9. Bellman, R., and M. Giertz., On the analytic formalism of the

theory of fuzzy sets, Information Sciences, Vol.5, 1973,
pp. 149-156.
B9a. Benayoun, R., B. Roy, and N. Sussman, Manual de reference du

programme Electre, Note de Synthese et Formation,
Direction scientifique SEMA, No. 25, Paris, 1966.
B9b. Bernardo, J.J. and J.M. Blin, A programming model of consumer

choice among multi-attributed brands, Journal of Consumer
Research, Vol. 4, No.2, 1977, pp.111-118.
B10. Bezdek, J.C., B. Spillman, and R.Spillman, Fuzzy measure

of preference and consensus in group decision-making,
Proceedings of 1977 IEEE Conference on Decision and Control,
1977, pp. 1303-1309.
B11. Bezdek, J.C., B.Spillman, and R. spillman, Fuzzy relation

spaces for group decision theory: An application, Fuzzy Sets
and Systems, Vol. 2, No.1, 1979, pp. 5-14
B12. Bezdek, J.C. (ed.), Analysis of Fuzzy Information,

Vol. I - Mathematics and Logic, CRC Press, Boca Raton,
Florlda, 1987.
Vol. II Artificial Intelligence and Decision Systems,
CRC Press, Boca Raton, Florida, 1987.

Vol. III - Applications in Engineering and science,
CRC Press, Boca Raton, Florida, 1987.
B15. Blin, J.M., Fuzzy relations in group decision theory,

Journal of Cybernetics, Vol. 4, 1974, pp. 17-22.
W7
B16. Blin, J.M., and A.B. Whinston, Fuzzy sets and social
choice, Journal of Cybernetics, Vol. 3, 1974, pp. 28-36.
B17. Blin, J.M., K.S. FU, A.B. Whinston. and K.B. Moberg,
Pattern recognition in micro-economics, Journal of
Cybernetics, Vol. 3, 1974, pp. 17-27.
B18. Blin, J.M., Fuzzy sets in multiple criteria decision making,

In: TIMS/studies in the Management Sciences, Vol. 6,
North-Holland, Amsterdam, 1977, pp. 129-146.
B19. Blin, J.M., Fuzzy sets in multi-criteria decision-making,

In: Multiple-Criteria Decision-Making, M.K. Slarr and
M. Zeleny (eds.), Academic Press, New York, 1977.
B20. Blishun, A.F., Fuzzy adaptive learning model of decision-

making process, Fuzzy Sets and Systems, Vol. 18, No.3,
1986, pp. 273-282.
B21. Blishun, A.F., Fuzzy learning models in expert systems,

Fuzzy Sets and Systems, Vol. 22, No.1, 1987, pp. 57-70.
B22. Bolanos, M., M.T. Lamata, and S. Moral, Decision making problems
under a Shaffer's evidence, First International Fuzzy System
Association (IFSA) Congress, July 1985.
B23. Bolanos, M.J., M.T. Lamata, and S. Moral, Decision making

problems in a general environment, Fuzzy Sets and Systems,
Vol. 25, No.2, 1988, pp. 135-144.
B24. Bonissone, P.P., The Problem of Linguistic Approximation in

System Analysis, Ph.D. Dissertation, Department of Electrical
Engineering and Computer Sciences, University of California,
Berkeley, 1979.
B25. Bonissone, P.P., A pattern recognition approach to the problem

of linguistic approximation in system analysis, IEEE 1979
International Conference on Cybernetics and society, 1979,
1979, pp. 793-798.
B26. Bonissone, P.P., A fuzzy set based linguistic approach:

Theory and applications, Proceedings of the 1980 Winter
Simulation Conference, Orlando, Florida, 1980, pp 99-111.
B27. Bonissone, P.P., A fuzzy sets based linguistic approach:

Theory and applications, In: Approximate Reasoning In
Decision Analysis, M.M. Gupta And E. Sanchez (eds.),
B28. Borisov A., o. Krumberg, A theory of possibility for decision-

making, Fuzzy Sets and Systems, Vol. 9, No.1, 1983, pp.13-23.
B29. Borisov, A.N., and G.V. Merkuryeva, Methods of utility

evaluation in decision-making problems under fuzziness and
randomness, Proc. IFAC Symposium on Fuzzy Information,
Knowledge Representation, and Decision Analysis,
Marseille, Pergamon Press, New York, 1983, pp. 307-312.
508
B30. Borisov, A.N., Y.Y. Luns., V.A. Popov, Fuzzy decision analysis
in the tasks of electric network development, Proc. IFAC
symposium on Fuzzy Information, Knowledge Representat~on and
Decision Analys~s, Marse~lle, Pergamon Press, New York, 1983,
pp. 289 294.
B31. Borisov, A., and L. Naglis, Multi-criteria choice of

alternatives in an expert system for computer-aided design of
industrial robot installation, Fuzzy Sets and Systems, Vol.
16, No.2, 1985, pp.93-101.
B32. Borisov, A., and V. Glushkov, The choice of alternatives

described by fuzzy evidences, In: Management Decision
Support Systems using Fuzzy Sets and Poss~bility
Theory, J. Kacprzyk (eds.), Verlag, 1985, pp. 85-100.
B33. Bortolan, G., and R.T. Degani, Ranking of fuzzy

alternatives in electrocardiography, IFAC Symposium on
Fuzzy Information, Knowledge Representation and
Decision Analysis, 1983, pp. 409 414.
B34. Bortolan, G. and R. Degani, A review of some methods

for ranking fuzzy subsets, Fuzzy Sets and Systems,
Vol. 15, No.1, 1985, pp. 1-19.
B35. Brans, J.P., B. Mareshal, and Ph. Vincke, Promethee: A

new family of outranking methods in multicriteria
analysis. In: J.P.Brans (ed.) operational Research
'84. North-Holland, Amsterdam. (proceedings of the
Tenth IFORS International Conference on Operational
Research, Washington, D.C.) 1984, pp. 477-490.
B36. Buckley, J.J., The multiple-judge, multiple-criteria

ranking problem: A fuzzy-set approach, Fuzzy Sets and
Systems, Vol. 13, No.1, 1984, pp. 25-38.
B37. Buckley, J.J., Fuzzy decision-making with data: Applicat-

tions to statistics, Fuzzy Sets and Systems, Vol. 16,
No.2, 1985, pp. 139-147.
B38. Buckley, J.J., Ranking alternatives using fuzzy numbers,

B39. Buckley, J.J., Fuzzy hierarchical analysis, Fuzzy Sets

and Systems, Vol. 17, No.3, 1985, pp. 233-247.
B40. Buckley, J.J., Fuzzy programming and the multicriteria

decision problem, In: optimization Models using Fuzzy
Sets and Possibility Theory, J. Kacprzyk(ed.),
D. Re~del Publishing Co., Boston, 1987, pp. 226-244.
B41. Buckley, J.J., Generalized and extended fuzzy sets with

applications, Fuzzy Sets and Systems, Vol. 25, No.2,
1988, pp. 159-174.
B42. Buckley, J.J., and S. Chanas, A fast method of ranking

alternatives using fuzzy numbers (Short communications),
Fuzzy Sets and Systems, Vol.30, No.3, 1989, pp. 337-339.
W9
C1. Capocelli, R.M., and A. DeLuca, Fuzzy sets and decision

theory, Information and Control, Vol. 23, 1973, pp. 446-473.
C2. Carlsson, C., An approach to handling fuzzy problem structures,

cybernetics and Systems, Vol. 14, No.1, 1983, pp.33-54.
C3. Carlsson, C., On the relevance of fuzzy sets in management

science methodology, In: TIMS/Studies in the Management
Sciences, Vol. 20, H.J. Zimmermann et al. (eds.), Elsevier
Science Publishers, 1984, pp. 11-28.
C4. Carlsson, C., Fuzzy multiple criteria for decision

support systems, In: Management Decision Support
Systems using Fuzzy Sets and Poss1bility Theory,
J. Kacprzyk (eds.), Verlag, Berlin, 1985, pp. 48-61.
C5. Chameau, J.L., M. Gunaratne, and A.G. Altschaeffl, An

application of type 2 fuzzy sets to decision-making in
engineering, In: Analysis of Fuzzy Information, Vol. II -
Artificial Intelligence and Decision Systems, J.C. Bezdek
(ed.), CRC Press, Boca Raton, Florida, 1987, pp. 145-151.
C6. Chang, R.L.P., and T. Pavlidis, Fuzzy decision trees,

IEEE International Conference on Cybernetics and society,
1976, pp. 564-567.
C7. Chang, S.S.L., Fuzzy mathematics, man, and his environment,

IEEE Trans. On Systems, Man, and Cybernetics, Vol. SMC-2,
1972, pp. 93-97,
C8. Chang, S.S.L., Application of fuzzy-set theory to economics,

Kybernetes, Vol. 6, 1977, pp. 203-207.
C9. Chang, S.Y., E.D. Brill, and L.D. Hopkins, Modeling to

generate alternatives - A fuzzy approach, Fuzzy Sets
C10. Chang, W., Ranking of fuzzy utilities with triangular membership

functions, Proceeding of International Conference on Policy
Analysis and Information Systems, 1981, pp. 263-272.
C11. Chang, W.K. A Study on the Ranking of Fuzzy Alternatives and its
Apelication to Dec1sion Making, Ph.D. dissertation, Tamkang
un1versity, Ta~an, 1982.
C12. Chen, S.H., Ranking fuzzy numbers with maximizing set

and minimizing set, Fuzzy Sets and Systems, Vol. 17,
No.2, 1985, pp. 113 129.
C13. Chen, S.J., and C.L. Hwang, Fuzzy Scoring of fuzzy

number - A direct comparison index, unpublished paper,
1989.
C14. Chen, S.M., A new approach to handling fuzzy decision-making

problems, Proceedings of the 18th International symposium
on Multiple-Valued Logic, Spain, 1988, pp. 72-76.
C15. Chen, Y.C., Fuzzy Linear Programming and Fuzzy Multiple

Objective Linear Programming, Master Thesis, Department of
Industrial Engineering, Kansas State university, 1984.
510
C16. Cheng, Y.M., A Computer Program for Multi-Attribute Multiple

Alternative Decision Making Problem using Fuzzy Sets, MS thesis,
University of Houston, Houston, Texas, 1978.
C17. Cheng, Y.M., and B. McInnis, An algorithm for multiple

attribute, multiple alternative decision problem based
on fuzzy sets with application to medical diagnosis,
IEEE Trans. On System, Man, and Cybernetics, Vol. SMC-10,
1980, pp. 645-650.
C18. Choate, S.W., and A. Kandel, Fuzzy resource management analysis,

Fuzzy Sets and Systems, Vol. 20, No.3, 1986, pp.291-316.
C19. Cholewa, W., Aggregation of fuzzy opinions - An axiomatic

approach, Fuzzy Sets and Systems, Vol. 17, No.3, 1985,
pp. 249-258.
C20. Chow, L.R., and W. Chang, A new ranking technique of fuzzy

alternatives and its applications to decision making,
Policy and Information, Vol. 7, 1983, pp. 31-48.
C21. Chu, A.T.W., R.E. Kalaba, and K. Spingarn, A comparison

of two methods for determining the weights of belonging
to fuzzy sets, Journal of optimization Theory and Appli-
cations, Vol. 27, 1979, pp. 531-538.
C21a. Chu, F., Quantitative evaluation of university teaching

quality - An application of fuzzy set and approximate
reasoning, Fuzzy Sets and Systems, Vol. 37, 1990, pp.1-11.
C22. Churchman, C.W., and R.L. Ackoff, An approximate measure of

value, Journal of the operations Research Society of
America, Vol. 2, 1954, pp.172-187.
C23. Civanlar, M.R., and H.J. Trussell, constructing membership

functions using statistical data, Fuzzy sets and Systems,
Vol. 18, No.1, 1986, pp. 1-13.
C24. Czogala, E., On distribution function description of

probabilistic sets and its application in decision making,
Fuzzy sets and Systems, Vol. 10, No.1, 1983, pp. 21-29.
C25. Czogala, E., and H.J. Zimmermann, The aggregation operations

for decision making in probabilistic fuzzy environment,
Fuzzy sets and Systems, Vol. 13, No.3, 1984, pp.223-239.
C26. Czogala, E., and P. Zysno, A contribution to the construc-

tion of the r-operator, cybernetics and Systems Research 2,
R. Trappl (ed.), Elsevier Science Publishlng, North Holland,
1984, pp. 531-534.
C27. Czogala, E., A dominance of alternatives for decision

making in probabilistic fuzzy environment, Proceedings
of the Eighth European Meeting on Cybernetics and Systems
Research, R. Trappl (ed.), Reidel Publishing Company,
1986, pp. 591-598.
C28. Czogala, E., Multi-criteria decision making by means of fuzzy

and probabilistic sets, Fuzzy Sets and Systems, Vol. 36, 1990,
pp. 235-244.
511
D1a. Dawes, R.M., Social selection based on multidimensional

criteria, Journal of Abnormal and Social Psychology, Vol.68,
No.1, 1964, pp.104-109.
01. Degani, R., and G. Pacini, Linguistic pattern recognition

algorithms for computer analysis, Proceedings of BIOSIGMA,
1978, pp. 18-26.
02. Degani, R., and G. Bortolan, Fuzzy numbers in computerized

electrocardiography, Fuzzy Sets and Systems, Vol. 24, No.3,
1987, pp. 345-362.
03. Degani, R., and G. Bortolan, The problem of linguistic

approximation in clinical decision making, International
Journal of Approximate Reasoning, Vol. 2, 1988, pp. 143-162.
04. Delgado, M., J.L. Verdegay, and M.A. Villa, A procedure for
ranking fuzzy numbers using fuzzy relations, Fuzzy Sets and
Systems, Vol. 26, No.1, 1988.
05. Delgado, M., J.L. Verdegay, and M.A. vila, Ranking

linguistic outcomes under fuzziness and randomness,
Proceedings of the 18th International symposium on
Multiple-Valued Logic, spain, 1988, pp. 352-357.
06. DeLuca, A. and S. Termini, Algebraic properties of fuzzy

sets, Journal of Mathematical Analysis and Applications,
Vol. 40, 1972, pp. 373-386.
07. Deutsch, S.J. and C.J. Malborg, A fuzzy set approach to data
set evaluation for decision support, IEEE Trans On Systems,
Man, and Cybernetics, Vol. SMC-15, No.6, 1985, pp. 777-783.
08. Dhar, S.B., Power-System long-range decision analysis

under fuzzy environment, IEEE Trans. on Power Apparatus
and Systems, Vol. PAS-89, 1979, pp. 585-596.
09. Dijkman, J., H. Van Haeringen, and S.J. DeLange, Fuzzy

numbers, Journal of Mathematical Analysis and Applications,
Vol. 92, No.2, 1983, pp. 302-341.
010. Dimitrov, V., Learning decision-making with fuzzy automata,

In: computer-oriented Learning Processes, J.C. Simon (ed.),
Noordhoff, 1976, pp. 149-154.
011. Dimitrov, V., Social choice and self-organization under

fuzzy management, Kybernetes, Vol. 6, 1977, pp. 153.
012. Dimitrov, V., Group choice under fuzzy information,

D12a. Dockery, J., and E. Murray, A fuzzy approach to aggregating

military assessments, International JoUrnal of Approximate
Reasoning, 1987, Vol. 1, pp.251-271.
013. Dinola, A., and A.G.S. Ventre, The Mathematics of Fuzzy

Systems, Verlag TUV Rheinland, 1986.
512
014. Dombi, J., A general class of fuzzy operators, the De Morgan

class of fuzzy operators, and fuzziness measures induced by
fuzzy operators, Fuzzy Sets and Systems, Vol. 8, 1982,
pp. 149-164.
015. Dombi, J., Basic concepts for a theory of evaluation:

The aggregative operator, European Journal of Opertional
Research, Vol. 10, 1982, pp. 282-293.
016. Dombi, J., and P. Zysno, Comments on the 7-model, cybernetics

and Systems Research, R. Trappl (ed.), North-Holland, 1982,
pp. 711-714.
017. Dombi, J., and Z. Vas, Basic theoretical treatment of fuzzy

connectives, Acta Cybernetica, Vol. 6, 1983, pp. 191-201.
018. Dompere, K.K., On epistemology and decision-choice rationality,

Cybernetics and Systems Research, R. Trappl (ed.),
019. Driankov, D., and I. Stantchev, An approach to multiobjectives

decision making with ordinal information, Cybernetics and
Systems Research, R. Trappl (ed.), North-Holland, 1982,
pp.253 260.
020. Dubois, D., A teaching system using fuzzy subsets and
multi-criteria analysis, Int. J. Math, Ed. Sci. Tech.,
Vol. 8, No.2, 1977, pp. 203 217.
D21. Dubois, D. and H. Prade, Comment on tolerance analysis using

fuzzy sets and a procedure for multiple-aspect decision-making,
International Journal of System Science, Vol. 9, 1978,
pp. 357-360.
D22. Dubois, D., and H. Prade, operations on fuzzy numbers,

International Journal of Systems Science, Vol. 9, No.6,
1978, pp. 613-626.
D23. Dubois, D., and H. Prade, Decision-making under fuzziness,

In: Advances in Fuzzy Set Theory and Applications, M.M. Gupta
(eds.), North-Holland, New York, 1979, pp.279-303.
D24. Dubois, D., and H. Prade, Outline of fuzzy set theory: An

introduction, In: Advances in Fuzzy Set Theory and Applications,
M.M. Gupta (ed.), North-Holland, New York, 1979, pp.27-48.
D25. Dubois, D., and H. Prade, Systems of linear fuzzy constraints,

026. Dubois, D., and H. Prade, Fuzzy Sets and Systems: Theory
and Applications, Academic Press, New York, 1980.
D27. Dubois, D., and H. Prade, A unifying view of comparison

indices in a fuzzy set-theoretic framework, In: Fuzzy Set
and Possibility Theory-Recent Development, R.R. Yager (ed.),
Pergamon Press, New York, 1982, pp. 3 13.
028. Dubois, D., and H. Prade, The use of fuzzy numbers in decision
analysis, In: Fuzzy Information and Decision Processes, M.M.
Gupta and E. Sanchez (eds.), North Holland, 1982, pp.309-321.
513
D29. Dubois, D., and H. Prade, Ranking of fuzzy numbers in

the setting of possibility theory, Information Sciences,
Vol. 30, 1983, pp. 183-224.
D30. Dubois, D., and H. Prade, criteria aggregation and ranking

of alternatives in the framework of fuzzy set theory,
In: TIMS/Studies in the Management Science, Vol. 20,
H.J. Zimmermann (ed.), Elsevier Science Publishers,
D31. Dubois, D., and H. Prade, Recent models of uncertainty

and imprecision as a basis for decision theory: Toward
less normative frameworks, In: Intelligent Decision
Support in Process Environments, E. Hollnagel, G. Mancini,
and D. Woods (eds.), Sprlnger-Verlag, New York, 1985,
pp. 3-24.
D32. Dubois, D., and H. Prade, A review of fuzzy set aggregation

connectives, Information Sciences, Vol. 36, 1985, pp.85-121.
D33. Dubois, D., and H. Prade, A tentative comparison of numerical

approximation reasoning methodologies, International Journal
of Man-Machine Studies, Vol. 27, 1987, pp.717-728.
D34. Dubois, D., and H. Prade, Recent literature - collected by

Didier Dubois and Henri Prade, Fuzzy Sets and Systems, Vol.
24, 1987, pp.385-390.
D35. Dubois, D., and H. Prade, Recent literature - collected by

Didier Dubois and Henri Prade, Fuzzy Sets and Systems, Vol.
26, 1988, pp. 225-242.
D36. Dubois, D., and H. Prade, Possibility Theory: An
Approach to computerized Processing of uncertainty,
Plenum Press, 1988.
D37. Dubois, D., H. Prade, and C. Testemale, Weighted fuzzy

pattern matching, Fuzzy Sets and Systems, Vol. 28, No. 3
1988, pp. 313-331.
D38. Dyckhoff, H., and W. Pedrycz, Generalized means as a model

of compensative connectives, Fuzzy Sets and Systems,
Vol. 14, 1984, pp. 143-154.
D39. Dziech, A., and M.B. Gorzalczany, Decision making in signal
transmission problems with interval-valued fuzzy sets,
E1. Easton, A., One-of-a-kind decisions involving weighted

multiple objectives and disparate alternatives, J.L.
Cochrane and M. Zeleny (eds.), In: Multiple criteria
Decision Making, university of South Carolina Press,
columbia, South Carolina, 1973, pp. 657-667.
E2. Efstathiou, J., A Practical Development of Multi-Attribute

Decision Making using Fuzzy Set Theory, Ph.D. Dissertation,
Department of Comput., university of Durham, England, 1979.
514
E3. Efstathiou, J., and V. Rajkovic, Multiattribute decision-

making using a fuzzy heuristic approach, IEEE Trans. On
Systems, Man, cybernetics, Vol. SMC-9, 1979, pp. 326-333.
E4. Efstathiou, J., The incorporation of objective and subjec-

tive aspects into the assessment of information systems,
In: The Information Systems Environment, Lucas, Land,
Lincoln, and Supper(eds.), North-Holland, 1980, pp.187-197.
E5. Efstathiou, J., and R. Tong, Ranking fuzzy sets using

linguistic preference relations, Proceedings of the
10th International s~osium on Multiple-Valued Logic,
Northwestern Univers~ty, Evanston, 1980, pp. 137-142.
E6. Efstathiou, J., and R. Tong, Ranking fuzzy sets: A decision

theoretic approach, IEEE Trans., Systems, Man, and cybernetics,
Vol. SMC-12, 1982, pp.655-659.
E7. Efstathiou, J., Practical multi-attribute decision-making

and fuzzy set theory, In: TIMS/Studies in the Management
Sciences, Vol. 20, H.J. Zimmermann (eds.), Elsevier Science
Publishers, North-Holland, 1984, pp. 307-322.
E8. Enta, Y., Fuzzy decision theory, In: Fuzzy Set and Possibily
Theory - Recent Developments, R.R. Yager (ed.), Pergamon
Press, New York, 1982, pp. 439-449.
E9. Eshragh, F., and E.H. Mamdani, A general approach to

linguistic approximation, International Journal on
Man-Machine Studies, Vol. 11, 1979, pp. 501-519.
E10. Eshragh, F., Conversational Programs for Decision Making using

Fuzzy Set Theory, Ph.D. Thesis, Dept. of Elec. and Electron.
Eng., Queen Mary College, Univ. of London, London, England,
1979.
Ell. Eshragh, F., Subjective multi-criteria decision making,

International Journal on Man-Machine Studies, Vol. 13,
1980, pp. 117-141.
E12. Eshragh, F., N.J. Mandic, and E.H. Mamdani, Multi-criteria

decision making using fuzzy sets, Progress in Cybernetics
and Systems Research, Vol. 8, 1980, pp. 483-489.
E13. Esogbue, A.O., and R.C. Elder, Fuzzy sets and the modeling
of physician decision process, part II: Fuzzy diagnosis
decision models, Fuzzy Sets and Systems, Vol, 3, No.1,
1980, pp. 1-9.
E14. Esogbue, A.O., A fuzzy sets approach to public participa-

tion effectiveness measurement in water quality planning,
Proceedings of the International Congress On Applied
Systems Research and Cybernetics, Acapulco, Mexico, 1980,
pp. 3076-3081.
E15. Esogbue, A.O., and Z.M. Ahipo, Fuzzy sets and water
resources planning, In: Fuzzy Set and Possibility Theory
Recent Development, R.R. Yager(ed.), Pergamon Press, New
York, 1982, pp. 450-465.
515
Fl. Freeling, A.N.S., Decision analysis and fuzzy sets, Masters

Thesis, Cambridge university, England, 1979.
F2. Freeling, A.N.S., Fuzzy sets and decision analysis,

IEEE Trans. On Systems, Man, and Cybernetics, Vol. SMC-10,
1980, pp. 341-354.
F3. Freksa, C., Linguistic description of human judgments

in expert systems and in the "soft" sciences, In:
Approximate Reasoning in Decision Analysis, M.M. Gupta
and E. Sanchez (eds.), 1982, pp. 297-306.
F4. French, S., Fuzzy decision analysis: Some criticisms,

In: TIMS/studies in the Management Sciences, Vol. 20,
H.J. Zimmermann (ed.), Elsevier Science Publishers,
F5. Freire, E., and A. Ollero, A method of multicriteria

analysis, In: Fuzzy Information and Decision
Processes, M.M. Gupta and E. Sanchez (eds.), North-Holland,
1982, pp. 289-300.
F6. Fung, L.W., and K.S. Fu, An axiomatic approach to rational

decision-making in a fuzzy environment, In: Fuzzy Sets and
Their Applications to Cogn1tive and Decision Processes,
G1. Gaglio, S., R. Minciardi, and P.P. puliafito, Multiperson

decision aspects in the construction of expert systems, IEEE
Trans. On Systems, Man, and Cybernetics, Vol. SMC-15, NO~
1985, pp. 536-539.
G2. Gaines, B.R., and L.J. Kohout, The fuzzy decade - A bibliography
of fuzzy systems and closely related topics, International
Journal on Man-Machine Studies, Vol. 9, 1977, pp. 1-68.
G3. Gaines, B.R., Fundamentals of decisions: Probabilistic

possibilistic and other forms of uncertainty in
decision analysis, In: TIMS/Studies in the Management
Sciences, Vol. 20, H.-J. Zimmermann(eds.}, Elsevier
Sciences Publishers, North-Holland, 1984, pp. 47-65.
G4. Gaines, B.R., L.A. Zadeh, and H.-J. zimmermann, Fuzzy

sets and decision analysis - A perspective, In:
TIMS/studies in the Management sciences, Vol.20, H.-J.
zimmermann(eds.}, Elsevier Science Publishers,
G5. Glushkov, V.I., and A.N. Borisov, Analysis of fuzzy

evidence in decision making models, In: Optimization
Models using Fuzzy Sets and Possibility Theory, J.
Kacprzyk(ed.}, D. Reidel Publishing Co., Boston, 1987,
pp. 141-153.
G6. Golden, B.L., and E.A. Wasil, Nonlinear programming on a

microcomputer, Computer and Operations Research, Vol.13,
1986, pp.149-156.
516
G7. Godo, L., and C. Sierra, A new approach to connective

generation in the framework of expert systems using
fuzzy logic, Proceedings of the 18th International
symposium on Multiple-Valued Logic, spain, 1988, pp.
157-162.
G7a. Gonzalez, A., A study of the ranking function approach

through mean values, Fuzzy Sets and Systems, Vol. 35, 1990,
pp.29-41.
G8. Grobelny, J., The fuzzy approach to facilities layout problems,

G9. Gupta, M.M., and E.H. Mamdani, Second IFAC round table on fuzzy
automata and decision processes, Automatica, Vol. 12, 1976,
pp.291-296.
G10. Gupta, M.M., G.N. Saridies, and B.R. Gaines (eds.), Fuzzy
Automata and Decision Processes, North-Holland, Amsterdam, 1977.
G11. Gupta, M.M., R.K. Ragade, and R.R. Yager (eds.), Advances in
Fuzzy Set Theory and Applications, North-Holland, New York,
1979.
G12. Gupta, M.M., and E. Sanchez(eds.), Approximate Reasoning

in Decision Analysis, North-Holland, Amsterdam, 1982.
G13. Gupta, M.M., and E. Sanchez(ed.), Fuzzy Information

and Decision Processes, North-Holland, Amsterdam, 1982.
G14. Gupta, M.M., Fuzzy information modeling in decision making,

International Conference on Advances in Information Sciences
and Technology, Golden Jubllee Conference at the Indian
Conference at the Indian statistical Institute, Calcutta,
India, 1982.
G15. Gupta, M.M., Fuzzy interval and its applications to

decision making processes, IFAC Large Scale Systems,
Warsaw, Poland, 1983, pp. 235-240.
G16. Gusev, L.A., and I.M. Smirnova, Fuzzy sets - Theory and
applications (A survey), Automation and Remote Control, No.5,
May 1973, pp. 66-85.
H1. Hagg, C., Possibility and cost in decision analysis,

Fuzzy Sets and Systems, Vol. 1, No.1, pp. 81-86.
H2. Hammerbach,I.M., and R.R. Yager, The personalization of

security selection - An application of fuzzy-set theory,
H3. Hannan, E.L., On the efficiency of the product operator

in fuzzy programming with multiple objectives,
H4. Hannan, E.L., Fuzzy decision making with multiple objectives

and discrete membership functions, International Journal of
Man-Machine Studies, Vol. 18, 1983, pp.49-54.
517
H5. Hersh, H.M., and A. Caramazza, A fuzzy-set approach to

modifiers and vagueness in natural language, Journal
of Experimental Psychology, Vol. 105, 1976, pp. 254-276.
H6. Hersh, H.M., A, Caramazza, and H.H. Brownell, Effects of

context on fuzzy membership functions, In:
Advances in Fuzzy set Theory and Applications, M.M. Gupta
(eds.), North-Holland, New York, 1979, pp. 389-408.
H7. Hesketh, T., R. Pryor, and B. Hesketh, An application of

computerized fuzzy graphic rating scale to the psychological
measurement of individual differences, International Journal
of Man-Machine Studies, Vol. 29, 1988, pp.21-36.
H8. Hipel, K.W., Fuzzy set methodologies in multicriteria

modeling, In: Fuzzy Information and Decision Processes,
M.M. Gupta and E. Sanchez(eds.), North-Holland, Amsterdam,
1982, pp. 279-287.
H9. Hisdal, E., Decision based on statements in natural

language, In: TIMS/studies in the Management Sciences,
Vol. 20, H.J. Zimmermann(eds.), Elsevier Science
HID. Hisdal, E., Infinite-valued logic based on two-valued

logic and probability Part 1.1. Difficulties with
present-day fuzzy-set theory and their resolution in
the TEE model, International Journal of Man-Machine
Studies, Vol. 25, 1986, pp. 89 Ill.
H11. Hisdal, E, Infinite-Valued logic based on two-valued

logic and probability Part 1.2 Different sources of
fuzziness, International Journal of Man-Machine Studies,
Vol. 25, 1986, pp. 113 138.
H12. Hwang, C.L., and A.S.M. Masud, Multiobjective Decision

Making - Methods and Applications, A State-of-the-Art
Survey, springer-Verlag, New York, 1979.
H13. Hwang, C.L., and K. Yoon, Multiple Attribute Decision
Making - Methods and Applications, A State-of-the-Art
Survey, Springer-Verlag, New York, 1981.
H14. Hwang, C.L., and M.J. Lin, Group Decision Making Under
Multiple Criteria, springer-Verlag, New York, 1987.
H15. Hwang, F., An Expert Decision Making Support System

For Multiple Attribute Decision Making, Ph.D. Thesis,
Dept. of Industrlal Engg, Kansas State university, 1987.
J1. Jacquet-Lagreze, E., Modeling preference among distribution

using fuzzy relation, 5th Res. Conference on subjective
Probability, utility, and Decision Making, Darmstadt,
West German, pp.1-4, 1975.
J2. Jain, R., Decision making in the presence of fuzzy

variables, IEEE Trans. On Systems, Man, and Cybernetics,
Vol. SMC-6, 1976, pp. 698-703.
518
J3. Jain, R., outline of an approach for the analysis of fuzzy

systems, International Journal of Control, Vol. 23, 1976,
pp. 627-640.
J4. Jain, R., Decision making in the presence of fuzziness

and uncertainty, Proceedings of the 1977 IEEE Conference,
on Decision and Control, 1977, pp. 1318- 1323.
J5. Jain, R. A procedure for multi-aspect decision making

using fuzzy sets, International Journal of System
science, Vol. 8, 1977, pp. 1-7.
J6. Jain, R., Fuzzyism and real world problems, In: Fuzzy
Sets - Theory and Applications to Policy Analysis and
Information Systems, P.P. Wang and S.K. Chang(eds.),
Plenum Press, New York, 1980, pp. 129-132.
J7. Johnston, D.M., Multiobjective Decision Making Under

Uncertainty: An Application of Fuzzy Set Theory, Ph.D. Thesis,
University of washington, U.S.A., 1986.
J8. Jones, A., A. Kaufmann, and H.J. Zimmermann, Fuzzy Set

Theory and Applications, D. Reidel, Dordrecht, 1985.
K1. Kacprzyk, J., Decision-making in a fuzzy environment

with fuzzy termination time, Fuzzy Sets and Systems,
Vol. 1, No.3, 1978, pp. 169-179.
K2. Kacprzyk, J., staniewski, P., Long-term inventory policy-

making through decision-making models, Fuzzy Sets and
Systems, Vol. 8, No.2, 1982, pp. 117-132.
K3. Kacprzyk, J., and R.R. Yager, "Softer" optimization and

control models via fuzzy linguistic quantifiers,
Information sciences, Vol. 34, 1984, pp. 157-178.
K4. Kacprzyk, J. Some "commonsense" solution concepts in group

decision making via fuzzy linguistic quantifiers,
In: Management Decision Support Systems Using Fuzzy Sets
and Possibility Theory, J. Kacprzyk (ed.), Verlag, Berlin,
1985, pp. 125-135.
K5. Kacprzyk, J., and R.R. Yager (eds)., Management Decision

support Systems Using Fuzzy Sets and posslbility Theory,
Verlag, Berlin, 1985.
K6. Kacprzyk, J., Group decision making with a fuzzy linguistic

majority, Fuzzy Sets and Systems, Vol. 18, No.2, 1986,
pp. 105-118.
519
K7. Kacprzyk, J., On some fuzzy cores and "soft" consensus

measures in group decision making, In: Analysis of
Fuzzy Information, Vol II - Artificial Intelligence
and Decis~on Systems, J.C. Bezdek (ed.), CRC Press,
Boca Raton, Florida, 1987, pp. 119-130.
K8. Kacprzyk, J., and S.A. Orlovski (eds.), optimization
Models Using Fuzzy Sets and Possibility Theory, Reidel,
Boston, 1987.
K9. Kahne, S., A procedure for optimizing development decision,

Automatica, Vol. 11, 1975, pp. 261-269.
K10. Kahne, S. A contribution to decision making in environmental

design, Proceedings of IEEE 63, 1975, pp. 518-528.
K11. Kandel, A., and R.R. Yager, A 1979 bibliography on fuzzy

sets, their applications, and related topics, In: Advances
in Fuzzy Set Theory and Applications, M.M. Gupta (eds.), 1979.
K12. Kandel, A., Fuzzy Mathematical Techniques with Application,

1986.
K13. Kaufmann, A., Theory of Fuzzy Subsets, Vol. 1, Academic
Press, New York, 1975.
K14. Kaufmann, A. Bibliography on fuzzy sets and their applications,

Busefal, No. 1-3, LSI Lab, University of Paul Sabatier,
Toulouse, France, 1980.
K15. Kaufmann, A., and M.M. Gupta, Introduction to Fuzzy

Arithmetic, Van Nostrand, New York, 1985.
K16. Kaufmann, A., and M.M. Gupta(eds.), Fuzzy Mathematical

Models in Engineering and Management Science, North-Holland,
Amsterdam, 1988.
K17. Kerre, E.E., The use of fuzzy set theory in electrocardio-

logical diagnostics, In: Approximate Reasoning in Decision
Analysis, M.M. Gupta and E. Sanchez (eds.), 1982,
pp. 277-282.
K18. Kickert, W.J.M., Fuzzy Theory on Decision-Making, A Critical

Review, Martinus Nijhoff Soc~al Sciences Division, Leiden,
1978.
K19. Kickert, W.J.M., Towards an analysis of linguistic modelling,

K19a. Kim, K., and K.S. Park, Ranking fuzzy numbers with index of
optimism, Fuzzy Sets and Systems, Vol. 35, 1990, pp. 143-150.
K19b. Klahr, D., Decision making in a complex environment: The use

of similarity judgements to predict preference, Management
Science, Vol. 15, NO. 11, 1969, pp.595-617.
K20. Klement, E.P., Operations on fuzzy sets and fuzzy numbers

related to triangular norms. Proceedings of the 11th
International Symposium On Multi-valued Logic, Oklahoma,
1981, pp. 218 225.
520
K21. Klement, E.P., operations on fuzzy sets: An axiomatic

approach, Information Sciences, Vol. 27, 1984, pp. 221-232.
K22. Klin, A.J., The role of decision models in the evaluation

of competing environment health alternatives, Management
Science, Vol. 18, pp. B52-B67, 1971.
K23. Klir, G.J., and T.A. Folger, Fuzzy Sets, Uncertainly, and
Information, Prentice Hall, Englewood cliffs, New Jersey, 1988.
K24. Kokawa, M., K. Nakamura, and M. Oda, Fuzzy-theoretic and

concept-formational approaches to inference and hint-effect
experiments in human decision processes, Proceedings of the 1977
IEEE Conference on Decision and Control, 1977, pp. 1330-1335.
K25. Korhonen, P.J., A hierarchical interactive method for ranking
alternatives with multiple qualitative criteria, European
Journal of operations Research, Vol. 24, 1986, pp. 265-276.
K26. Kovalerchuk, B. Ya., On foundation of fuzzy decision making,

IFAC Large Scale Systems, Warsaw, Poland, 1983, pp. 241-245.
K27. Kolodziejczyk, W., Orlovsky's concept of decision-making

with fuzzy preference relation-further results, Fuzzy Sets
K28. Kuzmin, V.B., and S.V. Ovchinnikov, Design of group
decisions II. In: Spaces of partial-order fuzzy relations,
K29. Kuzmin, V.B., and S.V. Ovchinnikov, Group decisions in

arbitrary spaces of fuzzy binary relations, Fuzzy Sets
K30. Kuzmin, V.B., A reference approach to obtaining fuzzy

preference relations and problem of choice, In: Fuzzy
Set and possibility-Recent Developments, R.R. Yager (ed.)
Pergamon Press, New York, 1982, pp. 107-118.
K31. Kuzmin, V.B., and S.I. Travkin, Fuzzy choice, In: optimization
Models Using Fuzzy Sets and Possibility Theory, J. Kacprzyk
(eds.), D. Reidel Publishing Co., Boston, 1987, pp. 99-121.
K32. Kwakernaak, H., An algorithm for rating multiple-aspect

alternatives using fuzzy sets, Automatica, Vol. 15, 1979,
pp. 615-616.
L1. Laarhoven, P.J.M., and W. Pedrycz, A fuzzy extension of

Saaty's priority theory, Fuzzy Sets and Systems, Vol. 11,
No.3, 1983, pp. 229-241.
L2. Lakov, D.R., and N. Naplatanoff, Decision-making in vague

conditions, Kybernetes, Vol. 6, 1977, pp. 91-93.
L3. Lee, E.S., and R.L. Li, Comparison of fuzzy numbers based on
the probability measure of fuzzy events, Computer and
Mathematics with Applications, Vol. 15, 1988, pp. 887-896.
521
L4. Lemmer, J.F., and L.N. Kanal (eds.), Uncertainty In

Artificial Intelligence 2, North-Holland, Amsterdam, 1988.
L5. Leung, Y., A concept of a fuzzy ideal for multicriteria

conflict resolution, In: Advances In Fuzzy Sets, Possibility
Theory, and Application, P.P. Wang (ed.), Plenum Press,
New York, 1983, pp. 387-403.
L6. Leung, Y., Hierarchical programming with fuzzy objective
constraints, In: Optimization Models using Fuzzy Sets and
Possibility Theory, J. Kacprzyk (eds.), D. Reidel, Boston,
1987, pp. 245-257.
L7. Li, R.J., and E.S. Lee, Ranking fuzzy numbers - A comparison,
Proceedings of the North American Fuzzy Information Processing
society Workshop, J.L. Chameau and James T.P. Yao (eds.),
NAFIPS 1987, Purdue university 1987, pp. 169-204.
L8. Liu, W.J., Fuzzy invariant subgroups and fuzzy ideals,

L9. Luce, R.D., Semiorder and a theory of utility discrimination,

Econometrica, Vol. 24, No.2, 1956, pp.178-191.
LI0. Luhandjula, M.K., Compensatory operators in fuzzy linear

programming with multiple objectives, Fuzzy Sets and
Systems, Vol. 8, No.3, 1982, pp. 245-252.
Ll1. Lukasik, S., Hierarchical multicriteria systems - A game

approach, IFAC Large Scale System, Warsaw, Poland, 1983,
pp. 297-302.
L12. Lusk, E.J., and J. Neves, The analysis of the consistency
measure of the eigenvalue priority model: A possibility
perspective, In: Management Decision Support Systems Using
Fuzzy Sets and Poss1b111ty Theory, J. Kacprzyk (ed.),
Verlag, 1985, pp. 101-124.
M1. Mabuchi, S., An approach to the comparison of fuzzy subsets

with an a-cut dependent index, IEEE Trans. On Systems, Man,
and cybernetics, Vol. SMC-18, No.2, 1988, pp. 264-272.
M2. MacCrimmon, K.R., Decision making among multiple attribute

alternatives: A survey and consolidated approach, RAND
Memorandum, RM-4823-ARPA, 1968.
M3. Maeda, M., S. Murakami, and o. Toshikazu, Fuzzy mUltiattri-

bute decision analysis on the choice of clinical tests
and/or treatments for anovulatory women, In: Analysis
of Fuzzy Information, Vol. II - Artificial Intelligence and
Decision Systems, J.C. Bezdek (ed.), CRC Press, Boca Raton,
Florida, 1987, pp. 163-173.
M4. Maiers, J., and Y.S. Sherif, Applications of fuzzy set

theory, IEEE Trans. On Systems, Man, and cybernetics,
Vol. SMC-15, No.1, 1985, pp. 175-189.
522
M5. Mamdani, E.H., and B.R. Gaines (eds.), Fuzzy Reasoning and
its Applications, Academic Press, London, 1981.
M6. Mandic, N.J., F. Eshragh, and E.H. Mamdani, Multiple-criteria

decision-making using fuzzy sets, In: Proc. 5th European Meeting
on cybernetics and System Research, vienna, 1980.
M7. Mandic, N.J., A Multi-Attribute Decision Making Scheme With

Fuzzy Rule-Based Modification of pr10rity, Ph.D. Thesis,
Univers1ty of London, 1983.
M8. Mandic, N.J., and E.H. Mamdani, A multi-attribute

decision-making model with fuzzy rule-based modification
of priorities, ·In: TIMS/studies in the Management
sciences, Vol. 20, H.J. zimmermann(ed.), Elsevier Science
M9. Mantara, R.L., P. Meseguer, F. Sanz, C. Sierra, and A.

Verdaguer, A fuzzy logic approach to the management of
linguistically expressed uncertainty, IEEE 1988 International
Symposium on Multiple-Valued Logic, 1988, pp. 144-151.
M10. Martel, J.M., and G.R. D'avignon, Project ordering with

multicriteria analysis, European Journal of Operational
Research, Vol. 10, 1982, pp. 56-69.
M11. Martel, J.M., and G.R. D'avignon, A fuzzy outranking relation

in multicriteria decision making, European Journal of
Operational Research, 1986, pp. 258-271.
M12. Martin, J.A., and R.L. De Mantaras, The process of

classification and learning the meaning of linguistic
descriptors of concepts, In: Approximate Reasoning in
Decision Analysis, M.M. Gupta and E. Sanchez(eds.),
North-Holland, Amsterdam,1982, pp. 165-175.
M13. Mathieu-Nicot, B., Fuzzy expected utility, Fuzzy Sets

M14. Mattila, J.K., On some logical points of fuzzy conditional

decision making, Fuzzy Sets and Systems, Vol. 20, No.2,
1986, pp. 137-145.
M15. Merkuryeva, G.V., and A.N. Borisov, Decomposition of

multiattribute fuzzy utility functions, Fuzzy Sets and
Systems, Vol. 24, No.1, 1987, pp. 35-50.
M16. Miller, G.A., The magic number seven, plus or minus seven,
Psychological Review, Vol.63, 1965, pp 81-97.
M17. Mizumoto, M., and K. Tanaka, Algebraic properties of

fuzzy numbers, IEEE International Conference on cybernetics
and Society, 1976, pp. 559-563.
M18. Mizumoto, M., and K. Tanaka, Algebraic properties of fuzzy

numbers, In: Advances in Fuzzy Set Theory and Applications,
M.M. Gupta (eds.), North-Holland, New York, 1979, pp.153-164.
M19. Mizumoto, M., Fuzzy sets and their operations, Part I,

Information and Control, Vol. 48, 1981, pp. 30-48.
523
M20. Mizumoto, M., Fuzzy sets and their operations, Part II,
Information and Control, Vol. 50, 1981, pp. 160-174.
M21. Montero, de Juan F.J., and J. Tejada, Some problems on

the definition of fuzzy preference relations, Fuzzy
Sets and Systems, Vol. 20, No.1, 1986, pp. 45-53.
M22. Montero, De Juan, F.J., Aggregation of fuzzy opinions

in a nonhomogeneous group, Fuzzy Sets and Systems,
Vol. 25, No.1, 1988, pp. 15-20.
M23. Murakami, S., S. Maeda, and S. Imamura, Fuzzy decision analysis

on the development of centralized regional energy control
system, IFAC Symposium on Fuzzy Information, Knowledge
Representatlon and Decision Analysis, 1983, pp.363-368
Mc1. McCahone, C., Fuzzy Set Theor¥ Applied to Production and

Inventory Control, Ph.D. Thesls, Department of Industrial
Engineering, Kansas State university, 1987.
N1. Nagashima, K., Inference System for Selection of an

Appropriate Multiple Attribute Decision Making Method,
MS. Thesis, Dept. of Industrial Engineering, Kansas State
University, 1986.
N2. Nakamura, K., Preference relation on a set of fuzzy

utilities as a basis for decision making, Fuzzy Sets
N3. Naplatanoff, N., and D. Lakov, Decision-Making in vague

conditions, Kybernetes, Vol. 6, 1977, pp. 91-93.
N4. Natvig, B., Possibility versus probability, Fuzzy Sets

N5. Nauta Lemke, H.R. van, T.G. Dijkman, H. van Haeringen, and M.
Pleeging, A characteristic optimism factor in fuzzy decision-
making, Proceedings of IFAC Symposium on Fuzzy Information,
Knowledge Representation and Decislon Analysls, Pergamon Press,
New York, 1983, pp. 283-288.
N6. Navarrete, N.R. Jr., Studies on Robust Methods for Multicriteria

Decisionmaking with Nonquantitative Information, Doctoral
Thesis, Department of Applled Mathematics and Physics, Kyoto
University, Japan, 1980.
N7. Negi, D.S., Fuzzy Analysis and optimization, Ph.D. Thesis,

Department of Industrial Engineering, Kansas State University,
1989.
N8. Negoita, C.V., and D.A. Ralescu, Application of Fuzzy Sets

to Systems Analysis, Birkhavuser, Basel and Stuttgart, 1975.
N9. Negoita, C.V. and D.A. Ralescu, Applications of Fuzzy Sets to

System Analysis, Wiley, New York, 1975.
524
N10. Negoita, C.V., Review of fuzzy sets and their application

to cognition and decision processes, IEEE Trans. On Systems,
Man, and Cybernetics, Vol. SMC-7, No.2, 1977.
N11. Negoita, C.V., P. Flondor, and M. Sularia, On fuzzy

environments in optimization problems, Econ. Comput.
Econ. Cybernet, Stud. Res., 1977, pp. 13~24.
N12. Negoita, C.V., Management Applications of System Theory,

Birkhauser Verlag, Vasel and Stuttgart, 1979.
N13. Negoita, C.V., Fuzzy Systems, Abacus Press, Tunbridge, Wells,

1981.
N14. Negoita, C.V., Fuzzy sets in decision-support systems,
Human System Management, Vol. 4, 1983, pp. 27-33.
N15. Negoita, C.V., Expert Systems and Fuzzy Systems, Benjaminl

Cummings Publisher, Menlo Park, CA, 1985.
N16. Nojiri, H., A model of fuzzy team decision, Fuzzy Sets
N17. Nojiri, H., On the fuzzy team decision in a changing

environment, Fuzzy Sets and Systems, Vol. 3, No.2,
1980, pp. 137-150.
N18. Nojiri, H., A model of the executives' decision processes

in new - product development. Fuzzy Sets and Systems,
Vol. 7, No.3, 1982, pp. 227-242.
N19. Norwich, A.M., and I.B. Turksen, The construction of
membership functions, In: Fuzzy Set and Possibility
Theory-Recent Developments, R.R. Yager (ed.), Pergamon
N20. Nowakowska, M., New ideas in the decision theory,

International Journal on Man-Machine Studies, Vol. 11,
1979, pp. 213-234.
N21. Nurmi, H., Approaches to collective decision making with

fuzzy preference relations, Fuzzy Sets and Systems,
Vol. 6, No.3, 1981, pp. 249-259.
N22. Nurminen, M.I., About the fuzziness in the analysis of

information systems, Progress In Cybernetics and
Systems Research, Vol. 3, 1976, pp. 337-345.
N23. Nurminen, M.I., and A. Paasio, Some remarks on the fuzzy

approach to multigoal decision making, The Finnish Journal
of Business Economics, Vol. 3, 1976, pp. 291-302.
01. Okuda, T., H. Tanaka, and K. Asai, A formulation of

fuzzy decision problems with fuzzy information using
probability measures of fuzzy event, Information and
Control, Vol. 38, 1978, pp. 135-147.
525
02. Orlovsky, S.A., Decision-making with a fuzzy preference

relation, Fuzzy sets and Systems, Vol. 1, 1978, pp. 155-167.
03 . . Orlovski, S.A., Mathematical programming problems with

fuzzy parameters, In: Management Decision Support
Systems using Fuzzy Sets and possibility Theory, J.
Kacprzyk (ed.), Verlag, 1985, pp. 136-145.
04. Osgood, C.E., G.J. Suci, and P.H. Tannenbaum, The Measurement
of Meaning (9th ed.), University of Illinois Press, Urbana,
1975.
05. Ovchinnikov, S.V., General negation in fuzzy set theory,

Journal of Mathematical Analysis and Applications, Vol. 92,
1983, pp. 234-239.
06. Ovchinnikov, S.V., Fuzzy choice functions, IFAC Symposium

on Fuzzy Information, Knowledge Representation and Deci-
sion Analys1s, 1983, pp. 369 374.
07. Ovchinnikov, S.V., and M. Migdal, On ranking fuzzy sets

(short communication), Fuzzy Sets and Systems, Vol. 24,
No.1, 1987, pp. 113-116.
08. Ovchinnikov, S.A., Preference and choice in a fuzzy

environment, In: optimization Models Using Fuzzy Sets
and Possibility Theory, J. Kacprzyk (ed.), D. Reidel
Publishing Co., Boston, 1987, pp. 91-98.
Pl. Pappis, C.P., On a Fuzzy Set Theoretic Approach to Aspects

of Decision-Making in Ill-Def1ned Systems, Ph.D. Thesis,
Department of Electrical and Electronic Engineering,
Queen Mary College, London, 1976.
P2. Pedrycz, W., An approach to the analysis of fuzzy systems,

International Journal of Control, Vol. 34, 1981, pp.403-421.
P3. Pedrycz, W., Some aspects of fuzzy decision-making,
Kybernetes, Vol. 11, 1982, pp. 297-301.
P4. Pedrycz, W., Fuzzy relational equations with generalized

connectives and their applications, Fuzzy Sets and Systems,
Vol. 10, No.2, 1983, pp. 185-202.
P5. Pedrycz, W., Ranking multiple aspect alternatives - Fuzzy

relational equation approach, Automatica, Vol. 22, No.2,
1986, pp. 251-253.
P6. Piera, N., J. Aguilar-Martin, and M. Sanchez, Mixed

connectives between min and max, Proceedings of the
18th International Symposium on Multiple-Valued Logic,
spain, 1988, pp. 244-247.
P7. Pollatschek, M.A., Hierarchical Systems and fuzzy-set

theory, Kybernetes, Vol. 6, 1977, pp. 147-151.
526
P8. Ponsard, C., Fuzzy sets in economics - Foundations of soft

decision theory, In: Mana~ement Decision Support Systems
Using Fuzzy Sets and Poss1bility Theory, J. Kacprzyk and
R.R. Yager (eds.), Verlag, 1985, pp. 25-37.
P9. Ponsard, C., spatial fuzzy consumer's decision making:

A multicriteria analysis, European Journal of operational
Research, Vol. 25, 1986, pp. 235-246.
P10. Prade, H., Operations research with fuzzy data, In:

Fuzzy set Theory and Applications to Policy Analysis
and Information System, P.P. Wang and S.K. Chang (eds.),
Plenum Press, New York, 1980.
P11. Procyk, T.J., Linguistic Representation of Fuzzy Variables,

Fuzzy Logic Working Group, Queen Mary College, London, England,
1976.
R1. Ragade, R.K., Profile transformation algebra and group

consensus formation through fuzzy sets, In: Fuzzy Automata
and Decision Processes, M.M. Gupta, G.N. Saridies, and
B.R. Gaines (eds.), North-Holland, Amsterdam, 1979, pp. 19-25.
R2. Rajkovic, V., and M. Bohanec, A cybernetic model of the

computer aided decision making process, Proceedings of the
9th International Congress on Cybernetics, Namur, 1980,
pp. 185-199.
R3. Ralescu, D., optimization in fuzzy environment, Proceedings
IFAC symposium on Fuzzy Information, Knowledge Representa
tion and Decision Analysis, Marseille, Pergamon Press,
New York, 1983, pp. 301-305.
R4. Ramik, J., and J. Rimanek, Inequality relation between

fuzzy numbers and its use in fuzzy optimization, Fuzzy
Sets and Systems, Vol. 16, No.2, 1985, pp. 123-1~
R5. Ramik, J., Extension principle in fuzzy optimization,

R6. Rose, J., Current topics in cybernetics and Systems,
section 8, Springer-Verlag, New York, 1978.
R7. Roubens, M., and P. Vincke, Fuzzy preference in an optimiza-

tion perspective, In: optimization Models using Fuzzy Sets
and Possibility Theory, J. Kacprzyk (ed.), D. Reidel, Boston
1987, pp. 77-90.
R8. Rouse, W.B., Fuzzy models of human problem solving,

In: Advances In Fuzzy Sets, Possibility Theory, and
Application, P.P. Wang (ed.), Plenum, New York, 1983,
pp. 377-386.
R9. Roy, B., How outranking relation helps multiple criteria

decision making. In: Multiple Criteria Decision Makin~,
Cochrane and Zeleny (eds.), university of South Caral1na
Press, 1973, pp.179-201.
527
R10. Roy, B., and P. Bertier, La methode Electre II: Une

application au media-planning, In: Vlleme Conference
Internationale de Recherche operationnelle, Dublin, M.
Ross(ed.), OR72 , North-Holland, Amsterdam, pp.291-302,
1973.
R11. Roy, B., Partial preference analysis and decision-aid:

The fuzzy outranking relation concept, In: Conflicting
Objectives in Decisions, D.E. Bell, R.L. Keeney, and
H. Raiffa (eds.), Wiley, New York, 1977, pp. 40-75.
R12. Roy, B., Electre III: Un algorithme de classement

fonde sur une representation floue des preferences en
presence de criteres multiples, Cahiers du Centre
d'Etude de Recherche operationnelle, Vol. 20, 1978, pp.3-24.
R13. Roy, B., Selektieren, sortieren und orden mit hilfe von
Pravalenzrelationen, zeitschr. F. Betriebsw. Forschung,
Vol.32, 1980, pp. 465-496.
Sl. Saaty, T.L., A scaling method for priorities in hierarchical

structures, Journal of Mathematical Psychology, Vol. 15,
1977, pp.234-281.
S2. Saaty, T.L., Exploring the interface between hierarchies,
multiple objectives, and fuzzy sets, Fuzzy Sets and Systems,
Vol. 1, No.1, 1978, pp. 57-68.
S3. Sagaama, S., sbjective probabilities, fuzzy sets and

decision making, Progress in cybernetics and Systems
Research, Vol. 3, 1976, pp. 289-296.
S4. Sakawa, M., Interactive fuzzy decision making for multi-

objective linear programming problem and its application,
Proceedings IFAC Symposium on Fuzzy Information, Know-
ledge Representation and Decision Analysis, Marseille,
Pergamon Press, New York, 1983, pp. 295-300.
S5. Sakawa, M., Interactive multiobjective decision-making

by the fuzzy sequential proxy optimization technique -
FSPOT, In: TIMS/Studies in the Management sciences,
Vol. 20, H.J. Zimmermann (eds.), Elsevier Science
Publishers , North-Holland, 1984, pp. 241-260.
S6. Sanchez, E. (ed.), Fuzzy Information, Knowledge Representa-

tion and Decision Analysis (Proceedings of the IFAC
symposium), Marse11le, France, July 1983, Pergamon Press,
New York, 1983.
S7. Sanchez, E., Solution of fuzzy equations with extended

operations, Fuzzy Sets and Systems, Vol. 12, No.3,
1984, pp. 237-248.
S8. Sanchez, E., and L.A. Zadeh (eds.), Approximate Reasoning in

Intelligent Systems, Decision and Control, Pergamon Press,
New York, 1987.
528
S9. Scheffe, P., On foundations of reasoning with uncertain

facts and vague concepts, In: Fuzzy Reasoning and its
Applications, B.R. Gaines (ed.), Academic Press, New
York, 1981, pp. 189-216.
S10. Seo, F., and M. Sakawa, Fuzzy extension of multiattribute
utility analysis for collective choice, Cfbernetics and
System Research 2, R. Trappl (ed.), Elsev~er Science
S11. Seo, F., and M. Sakawa, Fuzzy assessment of multiattribute

utility functions, In: Proceedings of an Interactive Decision
Analysis and Interpretative computer Intelligence, A. Wierzbicki
and M. Grauer (eds.), Verlag, New York, 1984, pp. 97-104.
S12. Seo, F., and M. Sakawa, Fuzzy multiattribute utility

analysis for collective choice, IEEE Trans. on Systems,
Man, and Cybernetics, Vol. SMC-15, No.1, 1985, pp. 45-53.
S13. Shaw, M.L.G., and B.R. Gaines, Deriving the constructs

underlying decision, In: TIMS/studies in the Management
sciences, Vol. 20, H.J. Zimmermann (ed.), Elsevier Science
Publishers , North-Holland, 1984, pp. 335-355.
S14. Shimura, M., Fuzzy sets concept in rank-ordering objects,

Journal of Mathematical Analysis and Applications, Vol. 43,
1973, pp. 717-733.
S15. Shimizu, K., Group decision problem under competitive

situation and its interactive solution by optimization-
satisfaction approach, IFAC Large Scale System, Warsaw,
Poland, 1983, pp. 259-262.
S16. Silvert, W., symmetric summation: A class of operations

on fuzzy sets, IEEE Trans. on Systems, Man, and Cybernetics,
Vol. SMC-9, 1979, pp. 657-659.
S17. Simon, H.A., Two heads are better than one: The collaboration
between AI and OR, Interfaces, Vol. 17, No.4, 1987, pp. 8-15.
S18. Siskos, J., A way to deal with fuzzy preferences in

multicriteria decision problem, European Journal of
Operation Research, Vol. 10, 1982, pp. 314-324.
S19. siskos, J., and Ph. Hubert, A survey and a new comparative
approach, European Journal of operation Research, Vol. 13,
1983, pp. 278-299.
S20. siskos, J.L., J. Lochard, and J. Lombard, A multicriteria
decision making methodology under fuzziness: Application to the
evaluation of radiological protection in nuclear power plants,
In: TIMS/studies in the Management sciences, Vol. 20,
H.J. Zimmermann (eds.), Elsevier Science Publishers,
S21. Schweizer, B. and A. Sklar, Probabilistic Metric Spaces,

Masson, Paris, 1985.
529
S22. Slyadz, N., and A. Borisov, Anaysis of fuzzy initial

information in decision-making models, In: Cybernetics
and Systems Research, Vol. 1, R. Trappl (ed.), North
Holland, Amsterdam, 1982, pp. 739-742.
S23. Smets, Ph., Information content of an evidence,

International Journal of Man-Machine Studies, Vol. 19,
1983, pp. 33-43.
S24. Sommer, G., Fuzzy Bayes-Decision making, Progress in

cybernetics and Systems Research, Vol. 8, 1980, pp. 499-507.
S25. Spillman, B., J. Bezdek, and R. spillman, Coalition analysis

with fuzzy sets, Kybernetes, Vol. 8, No.3, 1979, pp. 203-211.
S26. Spillman, B., R. Spillman, and J. Bezdek, A fuzzy

analysis of consensus in small group, In: Fuzzy Set
Theory and Applications to Policy Analysis and Informa-
tion System, P.P. Wang, and S.K. Chang (eds.), Plenum
S27. spillman, R., and B. Spillman, A survey of some

contributions of fuzzy sets to decision theory, In:
Analysis of Fuzzy Information, Vol. II - Artificial
Intel11gence and Dec1sion Systems, J.C. Bezdek (ed.),
CRC Press, Boca Raton, Florida, 1987, pp. 109-118.
S28. Sugeno, M., Fuzzy decision-making problems, Transactions

of society of Instrument and Control Engineers, Vol. 11,
1975, pp. 709 714.
S29. Sugeno, M., and G.T. Kang, Structure identification of

fuzzy model, Fuzzy Sets and Systems, Vol. 28, No.1, 1988.
T1. Takeda, E., and T. Nishida, Multiple criteria decision

problems with fuzzy domination structures, Fuzzy Sets
T2. Takeda, E., Interactive identification of fuzzy outranking

relations in a multicriteria decision problem, In: Fuzzy
Information and Decision Processes, M.M. Gupta and -----
E. Sanchez(eds.), North Holland, 1982, pp. 301-307.
T3. Tamura, H., and K. Yurimuka, Identification of group

utility functions for group decision making based on
convex dependence, IFAC Large Scale System, Warsaw,
Poland, 1983, pp. 263-268.
T4. Tanaka, H.T., T. Okuda, and K. Asai, Decision making and its
goal in a fuzzy environment, Proceedings of US-Japan Seminar On
Fuzzy Sets and their Applicat10ns, Berkeley, CA, July 1974.
T5. Tanaka, H., T. Okuda, and K. Asai, A formulation of fuzzy

decision problems and its application to an investment
problem, Kybernetes, Vol. 5, 1976, pp. 25-30.
530
T6. Tanaka, H., T. Okuda , and K. Asai, Fuzzy information

and decision in statistical model, In: Advances in
Fuzzy set Theory and Applications, M.M. Gupta (ed.),
North-Holland, New York, 1979, pp. 303-320.
T7. Tanaka, H., H. Ichihashi, and K. Asai, Fuzzy decision

in linear programming problems with trapezoid fuzzy
parameters, In: Management Decision Support Systems
using Fuzzy Sets and Possibility Theory, J. Kacprzyk (ed.),
Verlag, 1985, pp. 146-154.
T8. Tanino, T., Fuzzy preference orderings in group decision
making, Fuzzy Sets and Systems, Vol. 12, No.2, 1984,
pp. 117-131.
T9. Thole, U., H.J. Zimmermann, and P. Zysno, On the suitability
of minimum and product operators for the intersection of fuzzy
sets, Fuzzy Sets and Systems, Vol. 2, No.2, 1979, pp.167-180.
T10. Tong, R.M., and J. Efstathiou, A critical assessment of

truth functional modification and its use in approximate
reasoning (short communication), Fuzzy Sets and Systems,
Vol. 7, No.1, 1982, pp. 103-108.
T11. Tong, R.M., and P.P. Bonissone, Linguistic solutions

to fuzzy decision problems, In: TIMS/studies in the
Management Science, Vol. 20, H.J. Zlmmermann (ed.),
Elsevier Science Publishers B.V., North-Holland, 1984,
pp. 323-334.
T12. Toth, H., From fuzzy-set theory to fuzzy set-theory:

Some critical remarks on existing concepts, Fuzzy Sets
T13. Tseng, T.Y., C.M. Klein, and M.S. Leonard, A formalism

for comparing ranking procedures, Proceedings of the
7th Annual Meeting of the North American Fuzzy
Information processing society, NAFIPS 1988, pp. 231-235.
T14. Tseng, T.Y., and C.M. Klein, A comparison of ranking procedures
for fuzzy sets, Working Paper #8812101, Department of Industrial
Engineering, University of Missouri-Columbia, Columbia,
Missouri, 1988.
T15. Tsukamoto, Y., P.N. Nikiforuk, and M.M.Gupta, On the

comparison of fuzzy sets using fuzzy chopping, In:
Control Science and Technology for Progress of Society,
H. Akashi (ed.), Pergamon Press, New York, 1983, pp. 46-51.
T16. Tversky, A., Choice by elimination, Journal of Mathematical

Psychology, Vol. 9, No.4, 1972, pp. 341-367.
U1. Umano, M., M. Mizumoto, and K. Takeda, FSTDS system: A

fuzzy-set manipulation system, Information sciences,
Vol. 14, 1978, pp. 115-159.
531
U2. Urban, B., and V. Hansel, A fuzzy concept in the theory of

strategic decisions where several objects exist, In:
Proceedings IFAC symposium on Fuzzy Information, Knowledge
Representation and Decision Analysis, Pergamon Press, New
York, 1983, pp. 313-320.
V1. Van Gigch, J.P., and L.L. Pipino, From absolute to probable
and fuzzy in decision-making, Kybernetes, Vol. 9, No.1, 1980.
V2. Viliums, E.R., and L.Y. Suker, Practical aspects of

alternatives evaluating and decision making under
uncertainty and multiple objectives, In: Cybernetics
and Systems Research 2, R. Trappl (ed.), Elsevier science
W1. Wagenknecht, M., and K. Hartmann, On fuzzy rank-ordering

in polyoptimization, Fuzzy Sets and Systems, Vol. 11,
No.3, 1983, pp. 253 264.
W1a. Wang, P.P., and S.K. Chang, Fuzzy Sets - Theory and Applications
to Policy Analysis and Information Systems, Plenum Press,
New York, 1980.
W2. Watada, J., and H. Tanaka, The perspective of possibilis-
tic models in decision making, Proceedings of the 7th
International Conference on Multiple criteria Decision
Making, Kyoto, Japan, 1986, pp. 328-337.
W3. Watson, S.R., J.J. Weiss, and M.L. Donnell, Fuzzy decision
analysis, IEEE Trans. On Systems, Man, and cybernetics,
SMC-9, 1979, pp. 1-9.
W4. Weber, S., A general concept of fuzzy connetives, negations

and implications based on t-norms and t-conorms, Fuzzy Sets
and Systems, Vol. 11, 1983, pp. 115-134.
W5. Wenstop, F., Deductive verbal models of organizations,

International Journal of Man-Machine Studies, Vol. 8,
1976, pp. 293-311.
W6. Wenstop, F., Fuzzy set simUlation models in a systems

dynamic perspective, Kybernetes, Vol. 6, 1976, pp. 209-218.
W7. Wenstop, F., Quantitative analysis with linguistic values,

W8. Whalen, T., Decision-making under uncertainty with various

assumptions about available information, IEEE Trans. On
Systems, Man, and Cybernetics, Vol. SMC-14, 1984,
pp. 888-900.
W9. Whalen, T., and B. Schott, SAPIR: A new approach to

linguistic approximation, IEEE 1985 International
Conference on Cybernetics and Systems, 1985, pp. 341-345.
532
W10. Whalen, T., Introduction to decision making under

various kinds of uncertainty, In: optimization Models
Using Fuzzy Sets and Possibility Theory, J. Kacprzyk (ed.),
D. Reidel, Boston, 1987, pp. 27-49.
W11. Whalen, T., and B. Schott, Interactive task assignment

with composite fuzzy preference relations: an application
to faculty scheduling, In: Analysis of Fuzzy Information
Vol II - Artificial Intelligence and Decision Systems,
J.C. Bezdek(ed.), CRC Press, Boca Raton, Florida, 1987,
pp. 131-143.
W12. Wierzchon, S.T., Applications of fuzzy decision-making
theory to coping with ill-defined problems, Fuzzy Sets
Y1. Yager, R.R., and D. Basson, Decision-making with fuzzy

sets, Decision Sciences, Vol. 6, 1975, pp. 590-600.
Y2. Yager, R.R., Decisions under uncertainty using fuzzy

sets, Proceedings of American Institute for Decision
Sciences Northeast Regional Conference, 1977, pp. 17-20.
Y3. Yager, R.R., Multiple objective decision-making using

fuzzy sets, International Journal on Man-Machine Studies,
Vol. 9, 1977, pp. 375 382.
Y4. Yager, R.R., Fuzzy decision-making including unequal

objectives, Fuzzy Sets and Systems, Vol. 1, No.2, 1978,
pp. 87-95.
Y5. Yager, R.R., On a general class of fuzzy connectives,
In: Fourth European Meeting on cybernetics and Systems
Research, Amsterdam, 1978, (see also Fuzzy Sets and
Systems, Vol. 4, 1980, pp. 235-242.)
Y6. Yager, R.R., Ranking fuzzy subsets over the unit interval,
Proc. 1978 CDC, 1978, pp.1435-1437.
Y7. Yager, R.R., A measurement - informational discussion

of fuzzy union and intersection, International Journal
of Man-Machine Studies, Vol. 11, No.2, 1979, pp. 189-200.
Y8. Yager, R.R., Possibility decision making, IEEE Trans.

On Systems, Man, and cybernetics, Vol. SMC-9, No.9,
1979, pp. 388-392.
Y9. Yager, R.R., Fuzzy sets, probability, and decision, Journal

of Cybernetics, Vol. 10, 1980, pp. 1-18.
Y10. Yager, R.R., Decisions with ordinal preferences and

importance, Kybernetes, No.9, 1980, pp.109-114.
Y11. Yager, R.R., On choosing between fuzzy subsets, Kybernetes,

Vol. 9, 1980, pp. 151-154.
533
Y12. Yager, R.R., competitiveness and compensation in decision

making: A fuzzy set based interpretation, Computers and
operations Research, Vol. 7, 1980, pp. 285-300.
Y13. Yager, R.R., Finite linearly ordered fuzzy sets with

application to decisions, International Journal of Man
Machine studies, Vol. 12, 1980, pp. 299-322.
Y14. Yager, R.R., A foundation for a theory of possibility,

Journal of cybernetics, Vol. 10, 1980, pp. 177-204.
Y15. Yager, R.R., Satisfaction and fuzzy decision functions,

In: Fuzzy Set Theory and Applications to Policy Analysis
and Information System, P.P. Wang and S.K. Chang (eds.),
Plenum Press, New York, 1980, pp. 171-194.
Y16. Yager, R.R., On a general class of fuzzy connectives,

Y17. Yager, R.R., A linguistic variable for importance, Journal of

Cybernetics, Vol. 10, 1980, pp. 249-260.
Y18. Yager, R.R., A procedure for ordering fuzzy subsets of the unit
interval, Information Sciences, Vol. 24, 1981, pp. 143-161.
Y19. Yager, R.R., A new methodology for ordinal multiple

aspect decisions based upon fuzzy sets, Decision Science,
Vol. 12, 1981, pp. 589-600.
Y20. Yager, R.R., Applications of information granularity to

political and other decisions, Policy and Information,
Vol. 5, No.1, 1981, pp. 15-32.
Y21. Yager, R.R., Some procedures for selecting fuzzy set-

theoretic operators, International Journal of General
System, Vol. 8, 1982, pp. 115-124.
Y22. Yager, R.R. (ed.), Fuzzy Set and Possibility Theory -

Recent Developments, Pergamon Press, New York, 1982.
Y23. Yager, R.R., Level sets for membership evaluation of

fuzzy subsets, In: Fuzzy Set and Possibility Theory -
Recent Developments, R.R. Yager (ed.), Pergamon Press,
New York, 1982, pp. 90-97.
Y24. Yager, R.R., Linguistic hedges: Their relation to context

and their experimental realization, Cybernetics and Systems,
Vol. 13, 1982, pp. 357-374.
Y25. Yager, R.R., Membership in a compound fuzzy subset,

cybernetics and Systems, Vol. 14, 1983, pp. 173-184.
Y26. Yager, R.R., Quantifiers in the formulation of multiple

objectives decision functions, Information Sciences,
Vol. 31, 1983, pp. 107-139.
Y27. Yager, R.R., and J. Kacprzyk, Multicriteria and multistage

decision-making via fuzzy linguistic quantifiers,
Technique Report MII-306/301, Machine Intelligence Institute
Iona College, New Rochelle, New York, 1983.
5~
Y28. Yager, R.R., General multiple objective decision making

and linguistically quantified statements, International
Journal of Man-Machine studies, Vol. 21, 1984, pp. 389-400.
Y29. Yager, R.R., Aggregating evidence using quantified

statements, Information sciences, Vol. 36, 1985,
pp. 179-206.
Y30. Yager, R.R., Optimal alternative selection in the face

of evidential knowledge, In: Optimization Models Using
Fuzzy Sets and Possibility theory, J. Kacprzyk (eds.),
D. Reidel Publishing co., Boston, 1987, pp. 123-140.
Y31. Yager, R.R., On ordered weighted averaging aggregation

operators in multicriteria decisionmaking, IEEE Trans.
On Systems, Man, and Cybernetics, Vol. SMC-18, 1988,
pp. 183-190.
Y31a. Ying, M., The alternative measures of fuzzy sets, Fuzzy Sets
and Systems, Vol. 37, 1990, pp. 105-110.
Y32. Yoon, K., System Selection by Multiple Attribute Decision
Making, Ph.D. Thesis, Dept. of Industrial Engineering,
Kansas State University, 1980.
Y33. Yoon, K., A reconciliation among discrete compromise solutions,
Journal of Operational Research Society, Vol. 38, 1987,
pp.277-286.
Y34. Yoon, K. The propagation of errors in multiple attribute

decision analysis: A practical approach, Journal of
Operational Research Society, Vol. 40, 1989, pp. 681-686.
Y35. YU, P.L., Dissolution of fuzziness for better decision-

perspective and techniques, In: TIMS/studies in the
Management Sciences, Vol. 20, H.J. Zimmermann (ed.),
Elsevier Science Publishers, North-Holland, 1984, pp.171-207.
Y36. YU, P.L., Multiple-criteria Decision Making - Concepts,

Techniques, and Extensions, Plenum Press, New York, 1985.
Zl. Zadeh, L.A., Fuzzy sets, Information and Control, Vol.8,

1965, pp. 338-353.
Z2. Zadeh, L.A., Probability measure of fuzzy events, Journal

of Mathematical Analysis and Applications, Vol. 23, 1968,
pp. 421-427.
Z3. Zadeh, L.A., similarity relations and fuzzy orderings,
Information Science, Vol.3, 1971, pp. 177-200.
Z4. Zadeh, L.A., A fuzzy-set-theoretic interpretation of

linguistic hedges, Journal of cybernetics, Vol. 2, No.3,
1972, pp. 4-34.
Z5. Zadeh, L.A., Outline of a new approach to the analysis of

complex system and decision processes, IEEE Trans. On Systems,
Man, and Cybernetics, Vol. SMC-2, 1973, pp. 28-44.
535
Z6. Zadeh, L.A., The concept of a linguistic variable and

its application to approximate reasoning-I, Information
Sciences, Vol. 8, 1975, pp. 199-249.
Z7. Zadeh, L.A., The concept of a linguistic variable and

its application to approximate reasoning-II, Information
Sciences, Vol. 8, 1975, pp. 301-357.
Z8. Zadeh, L.A., A bibliography on fuzzy sets and their

application to decision processes, In: Fuzzy Sets and
Their Applications to Cognitive and Decision Processes,
Z9. Zadeh, L.A., K.S. Fu, K. Tanaka, and M. Shimura (eds.),

Fuzzy sets and their Applications to Cognitive and Decision
Processes, Academic Press, New York, 1975.
Z10. Zadeh, L.A. The linguistic approach and its application to

decision analysis, In: Directions in Large-Scale Systems, Y.C.
Ho (ed.), Plenum, New York, 1976, pp.339-370.
Zll. Zadeh, L.A., Fuzzy sets as a basis for a theory of possibility,

Z12. Zadeh, L.A., Fuzzy sets and information granularity

In: Advances in Fuzzy Set Theory and Applications,
M.M. Gupta (eds.), North Holland, New York, 1979, pp.3-19.
Z13. Zadeh, L.A., Fuzzy logic and its applications to decision

and control analysis, Proceedings 1978 IEEE Conference on
Decision and Control, 1979.
Z14. Zadeh, L.A., PRUF-a meaning representation language for natural

language, In: Fuzzy Reasoning and its Applications, B.R. Gaines
(ed.), Academic Press, New York, 1981, pp. 1-58.
Z15. Zadeh, L.A., A computational approach to fuzzy quantifiers,

In: Natural Languages Compo and Maths. with Appls., Vol. 9,
Vol. 9, 1983, pp. 149-184.
Z16. Zahariev, S., An approach to group choice with fuzzy

preference relations, Fuzzy Sets and Systems, Vol. 22,
No.3, 1987, pp. 203-213.
Z17. Zeleny, M., Membership functions and their assessment,

In: Current Topics in Cybernetics and Systems, J. Rose
(ed.), Verlag, Berlin, 1978, pp. 391-392.
Z18. Zeleny, M., On the (ir)relevancy of fuzzy set theory,

Human System Management, Vol. 4, 1984, pp. 301-306.
Z19. Zhukovin, V.E., The multicriteria decision making with

vector fuzzy preference relation, In: Cybernetics and
Systems Research, Trappl, R. (ed.), 1984, pp. 179-181.
Z20. Zhukovin, V.E., F.V. Burshtein, And E.S. Korelov, A decision

making model with vector fuzzy preference relation, Fuzzy Sets
and Systems, Vol. 22, No. 1/2, 1987, pp.71-79.
536
Z21. Zhukovin, V.E., A fuzzy multicriteria decision model,

In: optimization Models Using Fuzzy sets and Possibility
Theory, J. Kacprzyk (eds.), D. Reidel Publish1ng Company,
Boston, 1987, pp. 203-215.
Z22. Zimmermann, H.J., Description and optimization of fuzzy

systems, International Journal of General System,
Vol. 2, No.4, 1975, pp. 209-215.
Z23. Zimmermann, H.J., Results of empirical studies in fuzzy set

theory, In: A~plied General Systems Research, G.J. Klir (ed.),
Plenum Publish1ng Company, 1978, pp. 303-312.
Z24. Zimmermann, H.J., and P. Zysno, Latent connectives in human

decision-making, Fuzzy Sets and Systems, Vol. 4, No.1, 1980,
1980, pp. 37-52.
Z2S. Zimmermann, H.J., and P. Zysno, Decisions and evaluations
by hierarchical aggregation of information, Fuzzy Sets and
Systems, Vol. 10, No.3, 1983, pp. 243-260.
Z26. Zimmermann, H.J., Using fuzzy sets in operational research,

European Journal of Operational Research, Vol. 13, 1983,
pp. 201-216.
Z27. Zimmermann, H.J., Fuzzy programming and linear programming

with several objective functions, In: TIMS/studies in
the Management sciences, Vol. 20, H.J. Zimmermann (ed.),
Elsevier Science Publishers, North-Holland, 1984, pp. 109-121.
Z28. Zimmermann, H.J., L.A. Zadeh, And B.R. Gaines (eds.),

Fuzzy Sets and Decision Analysis, TIMS/studies in the
Management Sciences, Vol. 20, North-Holland, Elsevier
Science Publisher, 1984.
Z29. Zimmermann, H.J., and P. Zysno, Quantifying vagueness

in decision model, European Journal of operational
Research, Vol. 22, 1985, pp. 148-158.
Z30. Zimmermann, H.J., Fuzzy Set Theory and Its Applications,

Kluwer, Nijhoff Publishing, Boston, 1985.
Z31. Zimmermann, H. J., Fuzzy Set, Decision Making, and Expert
System, Kluwer, Boston, 1987.
Z32. Zimmermann, H.J., Modeling and solving ill-structured

problems in operation research, In: Analysis of Fuzzy
Information, Vol. III, Applications in Engineering and
science, J.C. Bezdek (ed.), CRC Press, Boca Raton,
Florida, 1987, pp. 217-240.
Z33. Zysno, P. One class of operators for the aggregation of fuzzy

sets, EURO III Congress, Armsterdam, 1979.
Z34. Zysno, P., The integration of concepts within judgmental

and evaluative processes, Progress in cybernetics and
Systems Research, Vol. 8, 1980, pp. 509-516.
Lecture Notes in Economics
and Mathematical Systems
For infonnation about Vols. 1-210
please contact your bookseller or Springer-Verlag
Vol. 211: P. van den Heuve1, 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. Nono, 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
we1tschutz 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. Miiller, 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. 151 pages. 1983. Vol. 241: U. Willmann, 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 Trans10g 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 Miirk- 1985.
ten mit unvollstiindiger Nachfragerinformaton. IX, 125 Seiten.
Vol. 246: P.W. Oller, Dynamic Feature Space Modelling,
1984.
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. Bally VI, 371 pages. 1985.
Bamberg and K. Spremann VII, 306 pages. 1984.
Vol. 248: Pural Rationality and Interactive Decision Processes.
Vol. 228: Nonlinear Models ofFluclUating Growth. Proceedings, Proceedings, 1984. Edited by M. Grauer, M. Thompson and A.P.
1983. Edited by R.M. Goodwin, M. Kriiger 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. Miiller, 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. Liitkepohl. 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 1. 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,128 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. Schneeweiss 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. Biittler, 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: J.K. Ho, R.P. Sundarraj, DECOMP: an Implementation
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: J. 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.J.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, D.J. 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. 351: 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. Riccil (Ed.), Declslon Processes In Economics.
Trade. V. 132 pages. 1989. Proceedings. 1989. III. 209 pages 1991.
Vol. 328: W. GUth. 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 Lotsizlng 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 Crltena Decision Supporl. Proceedings, 1989. XII, 393
Vol. 331: A. Lewandowski, A.P. Wierzbicki (Eds.), Aspiration pages. 1991.
Based Decision Support Systems. X, 400 pages. 1989. Vol. 358: P. Knottnerus, Linear Models with Correlaled Distur-
Vol. 332: T.R. Gulledge, Jr., L.A. Litteral (Eds.), Cost Analysis bances. VIII, 196 pages. 1991.
Applications of Economics and Operations Research. Vol. 359: E. de Jong, Exchange Rate Determination and Opti-
Proceedings. VII, 422 pages. 1989. mal Economic Policy Under Various Exchange Rate Regimes.
VII, 270 pages. 1991.
Vol. 360: P. Stalder, Regime Translations, Spillovers and Buffer
Stocks. VI, 193 pages. 1991.
Vol. 361: C. F. Daganzo, Logistics Systems Analysis. X, 321
pages. 1991.
Vol. 362: F. Gehreis, Essays In Macroeconomics of an Open
Economy. VII, 183 pages. 1991.
Vol. 363: C. Puppe, Distorted Probabilities and Choice under
Risk. VIII, 100 pages. 1991
Vol. 364: B. Horvath, Are Policy Variables Exogenous? XII,
162 pages. 1991.
Vol. 365: G. A Heuer, U. Leopold-Wildburger. Balanced
Silverman Games on General Discrete Sets. V, 140 pages. 1991.
Vol. 366: J. Gruber (Ed.), Econometric Decision Models. Pro-
ceedings, 1989. VIII, 636 pages. 1991.
Vol. 367: M. Grauer, D. B. Pressmar (Eds.), Parallel Computing
and Mathematical Optimization. Proceedings. V, 208 pages.
1991.
Vol. 368: M. Fedrizzi, J. Kacprzyk, M. Roubens (Eds.),
Interactive Fuzzy Optimization. VII, 216 pages. 1991.
Vol. 369: R. Koblo, The Visible Hand. VIII, 131 pages. 1991.
Vol. 370: M. J. Beckmann, M. N. Gopalan, R. Subramanian
(Eds.), Stochastic Processes and their Applications. Proceedings,
1990. XLI, 292 pages. 1991.
Vol. 371: A. Schmutzler, Flexibility and Adjustment to Infor-
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 Equilibra. 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.

1992 Fuzzy Multiple Attribute Decision Making Methods and Applications Compress

Uploaded by

Copyright:

Available Formats

1992 Fuzzy Multiple Attribute Decision Making Methods and Applications Compress

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

1992 Fuzzy Multiple Attribute Decision Making Methods and Applications Compress

Uploaded by

Copyright:

Available Formats

Lecture Notes in Economics

and Mathematical Systems 375

Fuzzy Multiple Attribute

ISBN-13: 978-3-540-54998-7 e-ISBN-13: 978-3-642-46768-4

Typesetting: Camera ready by author

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

This monograph is intended for an advanced undergraduate or

graduate course as well as for researchers, who want a compilation of

developments in this rapidly growing field of operations research.

This is a sequel to our previous works: "Multiple Objective Decision

Making--Methods and Applications: A state-of-the-Art Survey" (No.164

of the Lecture Notes); "Multiple Attribute Decision Making--Methods

and Applications: A State-of-the-Art Survey" (No.186 of the Lecture

Notes); and "Group Decision Making under Multiple Criteria--Methods

and Applications" (No.281 of the Lecture Notes).

In this monograph, the literature on methods of fuzzy Multiple

Attribute Decision Making (MADM) has been reviewed thoroughly and

critically, and classified systematically. This study provides

readers with a capsule look into the existing methods, their

characteristics, and applicability to the analysis of fuzzy MADM

an overview of the classical MADM in Chapter II.

Chapter III presents the basic concepts and mathematical

operations of fuzzy set theory with simple numerical examples in

a easy-to-read and easy-to-follow manner.

Fuzzy MADM methods basically consist of two phases: (1) the

aggregation of the performance scores with respect to all the

alternatives according to the aggregated scores.

Chapter IV deals with the fuzzy ranking methods. Fuzzy ranking

to all aspects of fuzzy applications. A systematic classification of

critically reviewed. The basic concepts, the computational procedures,

and the characteristics of each method are concisely discussed.

Numerical examples are used to illustrate the computational procedures.

A system for classifying over one dozen fuzzy MADM methods is

presented in Chapter V. The basic concept, the algorithm, and the

characteristics of each method are discussed. The computational

procedure of each method is illustrated by solving a simple numerical

hundreds of attributes. The existing fuzzy MADM methods are cumbersome

most real-world problems of more than ten alternatives and ten

attributes. We present a new and practical approach which can solve

understand and operationally easy to use.

An up-to-date bibliographical listing of more than 400 references

We wish to acknowledge the pioneering and outstanding work on

Professors D. Dubois, E.S. Lee, H. Prade, B. Roy, T.L. Saaty,

H. Tanaka, K. Yoon, P.L. Yu, and H.J. Zimmermann. We thank Mei-Hua

Chen for typing and Trenetta Jones for editing.

Shu-Jen Chen C.L. Hwang

II. MULTIPLE ATTRIBUTE DECISION MAKING -- AN OVERVIEW 16

III. FUZZY SETS AND THEIR OPERATIONS 42

3.2.2 Basic Concepts of Fuzzy Sets 45

3.2.2.1 Complement of a Fuzzy Set 45

3.2.2.2 Support of a Fuzzy Set 46

3.2.2.3 a-cut of a Fuzzy Set 46

3.2.2.4 convexity of a Fuzzy Set 47

3.2.2.5 Normality of a Fuzzy Set 48

3.2.2.6 Cardinality of a Fuzzy Set 48

3.2.2.7 The mth Power of a Fuzzy Set 49

3.3 Set-Theoretic operations with Fuzzy Sets 49

3.3.1 No compensation operators 52

3.3.2.3 Hamacher's Min Operator 53

3.3.2.4 Yager's Min Operator 54