Professional Documents
Culture Documents
(Applied Optimization 44) Evangelos Triantaphyllou (Auth.) - Multi-Criteria Decision Making Methods - A Comparative Study-Springer US (2000)
(Applied Optimization 44) Evangelos Triantaphyllou (Auth.) - Multi-Criteria Decision Making Methods - A Comparative Study-Springer US (2000)
A Comparative Study
Applied Optimization
Volume 44
Series Editors:
Panos M. Pardalos
University of Florida, US.A.
Donald Hearn
University of Florida, US.A.
The titles published in this series are listed at the end of this volume.
Multi-Criteria Decision
Making Methods:
A Comparative Study
by
Evangelos Triantaphyllou
Department of Industrial and Manufacturing Systems Engineering,
College of Engineering,
Louisiana State University,
Baton Rouge, Louisiana, US.A.
AU Rights Reserved
© 2000 Springer Science+Business Media Dordrecht
OriginaIly published by Kluwer Academic Publishers in 2000
Softcover reprint of the hardcover 1st edition 2000
No part ot the matenal protecteo by thlS copynght not1ce may be reproouceO or
utilized in any fonn or by any means, electronic or mechanical,
including photocopying, recording or by any infonnation storage and
retrieval system, without written pennission from the copyright owner
This book is gratefully dedicated to all my students;
of the past, the present, and the future.
TABLE OF CONTENTS
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
qualitative data, deriving weights from ratio and difference comparisons, and
especially sensitivity analysis of MCDM methods. This to me seems as
valuable or even more so than suggesting a new method which may solve
another variant of the MCDM problem. At the end of the book Fuzzy MCDM
methods are described and evaluated.
What makes this book so valuable and different from other MCDM
books is, that even though the analyses are very rigorous, the results are
described very clearly and are understandable even to the non-specialist. Also,
very extensive numerical studies and comparisons are presented, which are
hard to find in any other text that I know. This book, in fact, provides a
unique perspective into the core of MCDM methods and practice. The
presented theoretical and empirical analyses are complementary to each other,
thus allowing the reader to gain a deep theoretical and practical insight into
the topics covered in this book. In addition to this, the author offers at the end
of each chapter and at the end of the book suggestions for further research
and I can only hope, that his suggestions will be accepted by many scientists.
Dr. Triantaphyllou has been involved in MCDM for almost two
decades. He has become internationally known as one of the leading experts
in the field and he is, therefore, qualified as hardly anybody else to write this
book. I can only congratulate him on his achievement and hope that many
practitioners will benefit from this excellent book and that scientists will
accept his suggestions for further research as fascinating challenges.
Hans-Jtirgen Zimmermann
PREFACE
The research and the writing of this book would never had been
accomplished without the decisive help and inspiration from a number of
people to which the author is deeply indebted. His most special thanks go to
his first M.S. Advisor and Mentor, Professor Stuart H. Mann, currently the
Dean of the W.F. Harrah College of Hotel Administration at the University
of Nevada. The author would also like to thank his other M.S. Advisor
Professor Panos M. Pardalos currently at the University of Florida and his
Ph.D. Advisor Professor Allen L. Soyster, currently the Dean of Engineering
at the Northeastern University. He would also like to thank Professors F.A.
Lootsma, S. Konz, J. Elzinga, T.L. Saaty, L.G. Vargas, E.H. Forman, Dr.
P. Murphy; the CEO of InfoHarvest, Inc., and the late Drs. I.H. LaValle,
Dr. H.K. Eldin, and C.L. Hwang. The author is also highly appreciative for
the excellent comments made by Professor H.-J. Zimmermann.
Many thanks go to his colleagues at LSU; especially to Dr. E.
McLaughlin; Dean Emeritus of the College of Engineering at LSU and to Dr.
L.W. Jelinski; Vice-Chancellor for Graduate Studies and Research at LSU for
her support and inspiration. The author is also very indebted to his
distinguished colleague Dr. Tryfon T. Charalampopoulos at LSU. Many
special thanks are also given to the Editor Mr. John Martindale at Kluwer
Academic Publishers for his encouragement and incredible patience and to his
graduate student Vetle I. Torvik for his always thoughtful comments. The
author would also like to recognize here the significant contributions to his
published papers made by the numerous and anonymous referees and editors
of the journals in which his papers have been published.
The author would also like to acknowledge his most sincere gratitude
to his graduate and undergraduate students, which have always provided him
with unlimited inspiration, motivation and joy.
April 2000
E-mail: trianta@fsu.edu
Personal Web page: http://www.imse. fsu.edu/vangeiis/
Chapter 1
INTRODUCTION TO MULTI-CRITERIA
DECISION MAKING
Alternatives:
Usually alternatives represent the different choices of action available to the
decision maker. In this book the set of alternatives is assumed to be finite,
Multiple Attributes:
Each MCDM problem is associated with multiple attributes. Attributes are
also referred to as "goals" or .,decision criteria". Attributes represent the
different dimensions from which the alternatives can be viewed.
In cases in which the number of criteria is large (e. g., more than a
dozen), criteria may be arranged in a hierarchical manner. That is, some
criteria may be major ones. Each major criterion may be associated with
several sub-criteria. Similarly, each sub-criterion may be associated with
several sub-sub-criteria and so on. Although some MCDM methods may
explicitly consider a hierarchical structure in the criteria of a decision
problem, most of them assume a single level of criteria (e.g., no hierarchies).
Incommensurable Units:
Different criteria may be associated with different units of measure. For
instance, in the case of buying a used car, the criteria "cost" and "mileage"
may be measure~ in terms of dollars and thousands of miles, respectively.
It is this nature of having to consider different units which makes MCDM
problems intrinsically hard to solve.
Decision Weights:
Most of the MCDM methods require that the criteria be assigned weights of
importance. Usually, these weights are normalized to add up to one. How
these weights can be determined is described in Chapters 4 and 5.
Decision Matrix:
An MCDM problem can be easily expressed in a matrix format. A decision
matrix A is an (m X n) matrix in which element au indicates the performance
of alternative Ai when it is evaluated in terms of decision criterion C; (for i =
1,2,3, ... , m, andj = 1,2,3, ... , n). It is also assumed that the decision
maker has determined the weights of relative performance of the decision
criteria (denoted as wj ' for j = 1, 2, 3, ... , n). This information is best
summarized in Definition 1-1 (according to [Zimmermann, 1996]) and Figure
Chapter 1: Introduction to MCDM 3
1-1. Please note that in this definition the term "goals" is used instead of the
usual term "criteria".
Definition 1-1:
Let A = {Ai' for i = 1, 2, 3, ... , n} be a (finite) set of decision alternatives
and G = {gj' for j = 1, 2, 3, ... , m} a (finite) set of goals according to
which the desirability of an action is judged. Determine the optimal
alternative A· with the highest degree of desirability with respect to all
relevant goals gj'
Criteria
CI C2 C3 Cn
Alts. ( WI w2 W3 wn )
Al all a l2 aJ3 a ln
A2 a21 a22 a23 a2n
Dominance
No Information r---------t~ Maximin
7'-------.J Maximax
Elimination by Aspect
Information Lexicographic Semi order
on the l...-_---'r---+i Lexicographic Method
Attributes
Figure ]-2: A Taxonomy of MCDM Methods (according to Chen and Hwang [1991]).
Chapter 2
This section is only concerned with the way the WSM, WPM, AHP,
revised AHP, ELECTRE, and TOPSIS methods process the numerical data
in step 3. The central decision problem examined in this book is described
as follows. Given is a set of m alternatives denoted as AI' A2, A3, ••• , Am and
a set of n decision criteria denoted as Cl , C2 , C3, ... , Cn. It is assumed that
the decision maker has determined (the absolute or relative) performance
value aij (for i = 1, 2, 3, ... , m andj = 1, 2, 3, ... , n) of each alternative
in terms of each criterion. That is, he/she has determined the matrix A with
the aij values, along with the criteria weights Wj (i.e., the decision matrix as
it was defined in Section 1.2). In the next chapter a number of procedures for
determining these data is discussed.
Given the aij and Wj values, then the problem examined in this chapter
is how one can rank the alternatives when all the decision criteria are
considered simultaneously. In the following sections, the criteria are assumed
to represent some kind of profit. That is, the higher the value, the better it
is. Next, a number of MCDM models for solving the above problem (i.e.,
step 3 above) is presented.
The weighted sum model (WSM) is probably the most commonly used
approach, especially in single dimensional problems. If there are m
alternatives and n criteria then, the best alternative is the one that satisfies (in
the maximization case) the following expression [Fishburn, 1967]:
n
A~M-scor" = ~ L aijwj , for i =1, 2, 3, ... , m. (2-1)
I j = 1
where: A*WSM-score is the WSM score of the best alternative, n is the number
of decision criteria, (lij is the actual value of the i-th alternative in terms of the
j-th criterion, and Wj is the weight of importance of the j-th criterion.
The assumption that governs this model is the additive utility
assumption. That is, the total value of each alternative is equal to the sum
of the products given as (2-1). In single-dimensional cases, where all the
units are the same (e.g., dollars, feet, seconds), the WSM can be used
without difficulty. Difficulty with this method emerges when it is applied to
Chapter 2: MCDM Methods 7
Example 2-1:
Suppose that an MCDM problem involves four criteria, which are
expressed in exactly the same unit, and three alternatives. The relative
weights of the four criteria were determined to be: wJ = 0.20, w2 = 0.15,
W 3 = 0.40, and W 4 = 0.25. Also, the performance values of the three
alternatives in terms of the four decision criteria are assumed to be as follows:
25 20 15 30
A = 10 30 20 30
30 10 30 10
Therefore, the data for this MCDM problem are summarized in the following
decision matrix:
Criteria
C1 C2 C3 C4
Alts. ( 0.20 0.15 0.40 0.25)
A1 25 20 15 30
A2 10 30 20 30
A3 30 10 30 10
When formula (2-1) is applied on the previous data the scores of the three
alternatives are:
A J• WSM-score = 25 xO.20 + 20xO.15 + 15 x 0.40 + 30xO.25 =
= 21.50.
Similarly, we get:
A 2• WSM-score = 22.00,
and A 3. WSM-score = 20.00.
Therefore, the best alternative (in the maximization case) is alternative
A2 (because it has the highest WSM score; 22.00). Moreover, the following
ranking is derived: A2 > A J > A3 (where the symbol" > " stands for "better
than") . •
8 MCDM Methods: A Comparative Study, by E. Triantaphyllou
where n is the number of criteria, aij is the actual value of the i-th alternative
in terms of the j-th criterion, and Wj is the weight of importance of the j-th
criterion.
If the term R(AK / AJ is greater than or equal to one, then it indicates
that alternative AK is more desirable than alternative AL (in the maximization
case). The best alternative is the one that is better than or at least equal to all
other alternatives.
The WPM is sometimes called dimensionless analysis because its
structure eliminates any units of measure. Thus, the WPM can be used in
single- and multi-dimensional MCDM. An advantage of the method is that
instead of the actual values it can use relative ones. This is true because:
11
aKjI E
i = 1
aKj
;:
I
aKj
(2-3)
11
aIJI E
i = 1
ali
A relative value d Kj is calculated using the formula:
where the aKj's are the actual values.
Example 2-2:
Consider the problem presented in the previous Example 2-1.
However, now the restriction to express all criteria in terms of the same unit
is not needed. When the WPM is applied, then the following values are
derived:
R(A/A:zJ = (25/10)°·20 x (20/30)°·15 X (15/20)°·40 X (30/30)°·25 =
= 1.007 > 1.
Similarly, we also get:
R(A/A~ = 1.067 > 1,
Chapter 2: MCDM Methods 9
In the previous expression the term P(A k ) denotes the performance value (not
a relative one) of alternative Ak when all the criteria are considered under the
WPM model. Then, when the previous data are used, exactly the same
ranking is derived. Some interesting properties of this method are discussed
in Chapter 11, Sections 11.6 and 11.7.
The similarity between the WSM and the AHP is clear. The AHP
uses relative values instead of actual ones. Thus, it can be used in single- or
multi-dimensional decision making problems.
Example 2-3:
As before, we consider the data used in the previous two examples
(note that as in the WPM case the restriction to express all criteria in terms
of the same unit is not needed). The AHP uses a series of pairwise
comparisons (more on this can be found in Chapter 3) to determine the
relative performance of each alternative in terms of each of the decision
criteria. In other words, instead of the absolute data, the AHP would use the
following relative data:
Criteria
CI C2 C3 C4
Alts. ( 0.20 0.15 OAO 0.25)
That is, the columns in the decision matrix have been normalized to add up
to one. When formula (2-5) is applied on the previous data, the following
scores are derived:
A 1, AHP.score = (25/65)xO.20 + (20/55)xO.15 +
+ (15/65)xOAO + (30/65)xO.25 =
Chapter 2: MCDM Methods 11
0.34.
Similarly, we get:
A 2• AHP-score = 0.35,
and A 3. AHP.score 0.31.
Therefore, the best alternative (in the maximization case) is alternative A2
(because it has the highest AHP score; 0.35). Moreover, the following
ranking is derived: A2 > Al > A3 • •
Criteria
C1 C2 C3
Alts. (113 113 113)
1 9 8
9 1 9
1 1 1
In real life problems the decision maker may never know the previous
real data. Instead, he/she can use the method of pairwise comparisons (as
described in Chapters 3 and 4) to derive the relative data. When the AHP is
applied on the previous data, the following decision matrix with the relative
data is derived:
12 MCDM Methods: A Comparative Study, by E. Triantaphyllou
Criteria
CI C2 C3
Alts. (113 113 113)
Therefore, it can be easily shown that the vector with the final AHP scores
is: (0.45, 0.47, 0.08). That is, the three alternatives are ranked as follows:
A2 > Al > A3 •
Next, we introduce a new alternative, say A 4 , which is an identical
Q!ID.'. of the existing alternative A2 (i.e., A2 == A4)' Furthermore, it is also
assumed that the relative weights of importance of the three criteria remain
the same (i.e., 113, 113, 113). When the new alternative A4 is considered,
it can be easily verified that the new decision matrix is as follows:
Criteria
CI Cz C3
Alts. ( 113 1/3 113)
Similarly as above, it can be verified that the vector with the final AHP scores
is: (0.37, 0.29, 0.06, 0.29). That is, the four alternatives are ranked as
follows: Al > A2 == A4 > A 3 • Belton and Gear claimed that this result is in
logical contradiction with the previous result (in which A2 > Aj).
When the revised AHP is applied on the last data (that is, when the
data are normalized by dividing by the largest entry in each column), the
following decision matrix is derived:
Criteria
CI C2 C3
Alts. ( 113 113 113)
Al 119 1 8/9
Az 1 119 1
A3 119 119 119
A4 1 119 1
Chapter 2: MCDM Methods 13
The vector with the final scores is: (2/3, 19/27, 1/9, 19/27). That is, the
four alternatives are ranked as follows: A2 == A4 > Al > A3 • The last
ranking is, obviously, the desired one. •
(2-6)
~i>~
k=l
Therefore, the normalized matrix X is defined as follows:
Y = X W, or:
where:
WI 0 0 0
0 w2 0 0
n
W and LW i l.
i= I
0 0 0 ... Wn
d m1 d m2 dm3 ...
as with the C matrix, the entries of matrix D are not defined when k = I.
Finally, it should be noted here that the previous two m x m matrices are
Chapter 2: MCDM Methods 17
not symmetric.
1 m m
" = m(m - 1)
E E
k=l 1= 1
C/d'
(2-8)
tIIt4 k., tIIt4 I.k
have an element equal to one. Then, the best alternative is the one which
dominates all other alternatives in this manner.
where xij denotes the performance measure of the i-th alternative in terms
of the J-th criterion. Next, the steps of the TOPSIS method are presented.
(2-11)
~i;x~
&=1
maker wants to have a minimum value among the alternatives. From the
previous definitions it follows that alternative A * indicates the most preferable
alternative or the ideal solution. Similarly, alternative A- indicates the least
preferable alternative or the negative-ideal solution.
n
L (vij - Vj J2, for i = 1,2,3, ... , m, (2-14)
j = 1
where Si* is the distance (in the Euclidean sense) of each alternative from the
ideal solution.
Similarly, for the distances from the negative-ideal solution we have:
n
S.1 - L (vij - Vj _)2, for 1, 2, 3, ... , m, (2-15)
j = 1
where Si_ is the distance (in the Euclidean sense) of each alternative from the
negative-ideal solution.
c. =
S.'
1*
Si. + 1-
decision maker has to choose his/her answer among 10-17 discrete choices.
Each choice is a linguistic phrase. Some examples of such linguistic phrases
are: "A is more important than B, " or "A is of the same importance as B, "
or "A is a little more important than B, " and so on. The main focus of this
chapter is not the wording of these linguistic statements, but, instead, the
numerical values which should be associated with such statements.
The main challenge with the use of pairwise comparisons is how to
quantify the linguistic choices selected by the decision maker during the
evaluation of the pairwise comparisons. All the methods which use the
pairwise comparisons approach eventually express the qualitative answers
(i.e., the linguistic statements) of a decision maker into some numbers. The
present chapter examines the issue of quantifying pairwise comparisons. Since
pairwise comparisons are the keystone in solving many MCDM problems,
correctly quantifying them is a very crucial step in MCDM methods which
use qualitative or fuzzy data.
Pairwise comparisons are quantified by using a scale. Such a scale
is nothing but an one-to-one mapping between the set of discrete linguistic
choices available to the decision maker and a discrete set of numbers which
represent the importance, or weight, of the previous linguistic choices. There
are two major approaches in developing such scales. The first approach is
based on the linear scale proposed by Saaty [1980] as part of the analytic
hierarchy process (AHP). The second approach was proposed by Lootsma
[1988], [1990], and [1991] and determines exponential scales. Both
approaches depart from some psychological theories and develop the numbers
to be used based on these psychological theories.
This chapter is organized as follows. The second section illustrates
the principals of the two classes of scales. The second section also presents
some ways for generating even more scales based on Saaty's linear scale and
on the exponential scales proposed by Lootsma. The third section discusses
some ways for evaluating the performance of various scales. This is achieved
in terms of two evaluative criteria. The next section describes the problem
of selecting the appropriate scale (or scales) as an MCDM problem. Some
computational results presented in the fifth section reveal that under different
conditions some scales are more efficient than others. These findings are
presented in depth in the final section which is the conclusion section of this
chapter.
Chapter 3: Quantification of Qualitative Data in MCDM Problems 25
Intensity of
Importance Definition Explanation
Besides the second scale, many other scales can be generated. One
way to generate new scales is to consider weighted versions between the
previous two scales. That is, within the interval [1, 1/9] the values can be
calculated using the formula:
NewValue = Value(Scalel) +
+ [Value(Scale2) - Value(Scale1)] x (a/IOO),
where a can range from a to 100. Then, the values in the interval [9, 1] are
the reciprocals of the above values. Observe that for a = a Scale1 is
derived, while for a = 100 Scale2 is derived.
28 MCDM Method: A Comparative Study, by E. Triantaphyllou
Intensity of
Importance Defmition
With these turning points interpolated between the major ones, we find a
geometric sequence of echelons with the progression factor equal to 1.8.
c) Size of nations. The above categorization is not only found on the time
axis, but also in spatial dimensions when we categorize the nations on the
basis of the size of their population. Omitting the very small nations with
less than one million inhabitants, we have:
Because the respective nations fall typically between the major echelons, the
refined sequence of echelons has the progression factor equal to 2.0.
Under the precaution that the precision should not be taken too
seriously because we have at each of these major echelons a mixture of
wave lengths, we can observe that the subjective light intensities also
constitute a geometric sequence with the progression factor equal to 4.0.
In the previous paragraphs we have used five examples to
demonstrate that exponential scales are common in human comparative
judgment when dealing with historical periods, planning horizons, size of
nations, sound intensities, and perception of light. Therefore, these
examples make exponential scales only plausible. Lootsma [1991] has
32 MCDM Method: A Comparative Study, by E. Triantaphyllou
studied the scale sensitivity of the resulting scores when exponential scales
are used. He observed that the rank order of the scores is not affected by
variations of the scale parameter. That is, the numerical values of the
calculated scores are weakly dependent on that parameter.
For a more detailed documentation on psychophysics we refer the
interested reader to [Marks, 1974], [Michon, et al., 1976), [Roberts,
1979], [Zwicker, 1982], and [Stevens and Hallowell Davis, 1983]. The
reader will find that the sensory systems for the perception of tastes, smells,
and touches follow the power law with exponents in the vicinity of one.
However, any norm which attempts to approximate the real (and thus
unknown) ratios with ratios taken from a finite and discrete set of values,
will always allow for the possibility that some real ratios (which are close
enough to each other) will be mapped to the same discrete value from the
current scale. The last statement indicates that Theorem 3-1 (stated later in
Section 3.3.2) will still be valid if alternative norms are considered
(however, its present proof assumes that the first norm is used).
34 MCDM Method: A Comparative Study. by E. Triantaphyllou
The matrix with the aij entries that we assume the decision maker is
able to construct as described above has entries from the discrete and finite
set e. We call this matrix the Closest Discrete Pairwise matrix or the CDP
matrix. The CDP matrix may not be perfectly consistent. That is, the
consistency index (CI) values (see the next section for an exact definition of
the CI index) of CDP matrices are not necessarily equal to zero. More on
this inconsistency issue will be discussed in the following section. It is
important to observe here that the CDP matrices are the reciprocal matrices
with pairwise comparisons that a decision maker will construct if we assume
that each of his/her pairwise comparisons is the closest possible to its actual
real value.
The decision maker is limited by the discrete numerical values (i.e.,
the values from the set e provided to him/her by a scale). He/She may
never know the actual values of his/her pairwise comparisons. He/She
simply attempts to approximate them. In other words, we assume here that
these approximations are the closest possible. Clearly, this is a highly
favorable assumption when one attempts to investigate the effectiveness of
various scales. The following example illustrates further the concepts of the
RCP and CDP matrices.
Example 3-1:
Suppose that the real (and hence unknown) relative performance
values (or weights of importance), after normalization, of a set with three
entities (i.e., three alternatives or three criteria) are WI = 0.77348, W 2 =
0.23804, and W3 = 0.23848. Then, the RCP matrix with the real values of
the pairwise comparisons is as follows:
1 3.24938 3.24342
0.30832 1.00183 1
1 3 3
-
1 1 1
CDP = 3
1
1 1
3
To see this consider the (1, 2) entry of the previous RCP matrix.
For this entry we have a l2 = 3.24938. Therefore, when the values in Table
3-1 are to be used in order to quantify the (1, 2) pairwise comparison then,
the a l2 entry is approximated by the value 3. The value 3 is the closest one
to the value 3.24938 when the values in Table 3-1 are used. Clearly, this
is an assumption which is made here in order to study different scales. A
similar explanation holds for the rest of the entries in the previous CDP
matrix . •
A - n
CI = _max_-::--
n - 1 '
where Amax is the maximum eigenvalue of the matrix with the pairwise
comparisons and n is the order of that matrix.
In the following paragraphs we will show that CDP matrices can be
inconsistent regardless of the scale used to quantify the actual pairwise
comparisons. This is stated in terms of the following theorem:
Theorem 3-1:
Regardless of the scale that is used to quantify the pairwise comparisons of
n (where n ~ 3) entities, the corresponding CDP matrices may be
inconsistent.
Proof:
Without loss of generality, suppose that AI> A 2 , and A3 are three items
(alternatives or criteria) of a collection of n (where n ~ 3) items that we
need to compare in terms of some common characteristic. Let the current
scale be defined on the following (2k + 1) discrete values (where k ~ 1):
[Vk' 1/Vk_l , Vk_2 , ••• , V2 , VI> 1, 1/VI , 1/V2 , ... , 1/Vk_2 , 1/Vk_l , 1/Vk],
where V; > 0 for any i = 1, 2, 3, ... , k, and VI > 1.
In this proof it will be shown that, when the previous scale is used,
then it is possible for the three comparisons a 12 , a 13 , a32 made by the
decision maker not to satisfy the consistency requirement:
a l2 = a13 X Cl:32'
Suppose that the actual values of the pairwise comparisons that
correspond to the previous three items AI' A 2 , and A3 are as follows:
Al 3Vl + 1
= (X 12 =
Az 4
Al VI + 3
= (X 13 =
A3 4
A3 (A l/A 2) (X 12 3Vl + 1
and = (X 32
A2 (Al/A3) (Xu VI + 3
Using the above relations it can be easily verified (since VI > 1) that the
following conditions (i) are true (see also Figure 3-1):
Chapter 3: Quantification of Qualitative Data in MCDM Problems 37
V1 + 1
V1 > cx 12 > 2
V1 + 1
2 ~ cx 13 >1,
V1 + 1
2 > CX 32 > 1.
1/V1 1.00
! ! MI
I
Note: MI = (VI + 1)/2; is the middle point of the interval [VI, 1].
Lemma 3-1:
Let a scale for quantifying pairwise comparisons be defined on the following
(2k+ 1) discrete values:
fVk' Vk _I, Vk _Z' ••• , Vz, VI' 1, 1lVI' IIVz, ... , IlVk _z, IIVk _I, IIVJ,
where V; > 0, for any i = 1, 2, 3, ... , k, and VI > 1.
Then, the maximum error, 0max' of the pairwise comparisons in a CDP
matrix is given by the formula:
V-V.
} }- I }
max { for j=1,2,3, ... ,k, and Vo 1.
°max
lJ + lJ-I '
Proof:
Suppose that a pairwise comparison has actual (and hence unknown) value
equal to ex, where: 1Il-j;;::: ex ;;::: 1Il-j_1' for some k ;;::: j ;;::: 1. Let M be the
middle point of the interval [1Il-j, lI"1-a. That is:
M =
lIlJ_I - lIV
--'---:::------..:..} +
1 V} + V}- I
2 lJ 2 V} V'} - I
Then, the largest 0 value for this particular pairwise comparison occurs
when the value of ex coincides with the middle point M. This is true
because in this case the closest value from the values permitted by the
current scale has the largest distance from ex. That is, under the assumption
that the decision maker will choose the closest value (i.e., he/she will be as
much accurate as possible), the value of this pairwise comparison will
become equal either to 1Il-j or to 11"1-1' In the first case the corresponding
ovalue, which we will call 0], becomes:
Chapter 3: Quantification of Qualitative Data in MCDM Problems 39
° = __
1
lIV.
M
J -1
VJ
V.J + V.J- I -1
VJ-I - VJ
VJ-I + VJ
2 V.J V·J- I
Similarly, in the second case the value of 0, which we will call ~, becomes:
v. - V. I
02 = J J-, that is: 1011 = 1021.
V.J- I + V.J
Since, in general, it is assumed that: lIVk ::;; O! ::;; Vk , it follows that the
maximum value of 0, denoted as 0max, is given by the following formula:
°mu
= max {
V.-V.
J J-
V.J + V.J- I
I }
'
Theorem 3-2:
Let a scale for quantifying pairwise comparisons be defined on the following
(2k+ 1) discrete values:
[Vk' Vk _]' Vk-2' ••• , V2, V]' 1, 1IV], IIV2, •• ·, IlVk_2, IlVk_], IIVJ,
where V; > 0 for any i = 1, 2, 3, ... , k.
Then an upper bound of the maximum consistency index, denoted as Clmox,
of the resulting CDP matrices is given by the following relation:
fl
CI ~~,
max 2
V.-V. I }
where' 5 = max { j j-
'max V V'
j + j-I
for } = 1, 2, 3, ... , k, and Vo = 1.
Proof:
The proof of this theorem is based on Theorem 7-16, stated in [Saaty,
1980]. According to that theorem the following relation is always true:
~x - n ~ (n - 1)/2 02max, (3-1)
where 5max is defined as:
Omax = max {eij - I}, and eij = ay(w/w i ),
for any i,} = 1, 2, 3, ... , n.
The aij's are the entries of the pairwise (judgment) matrix and Wi' Wj are the
real weights of entities i and}, respectively. From relation (3-1), above, we
get:
Amax -n 02max
~ or:
n- 1 T'
(3-2)
o~x
C1max ~
T'
For the case of CDP matrices the value of the maximum 0, denoted as 5max ,
can be determined as follows (see also Lemma 3-1):
5 = max { V.-V.
j }-
I }
(3-3)
max V . + V. I '
} }-
the relation:
O~ax
C1rnax ::::; 2'
where omax is given by the previous relation (3-3). •
Corollary 3-1:
When the original Saaty scale is used, an upper bound of the maximum
consistency index, Clmax , of the corresponding CDP matrices is:
(113)2 1
C1max ::::; 2 18·
0.030
0.028
0.026
0.024
0.022
0.020
0.018
-
rI.I 0.016
Average
=
~
;;;. 0.014
CIS
1-1
U 0.012 Minimum
0.010
0.008
0.006
0.004
0.002
O.OOO~.~-~~~~~~~~~r-~r-~r-~r-~r-r-o
o 10 20 30 40 50 60 70 80 90 100
Order of Set
Figure 3-2: Maximum, Average, and Minimum CI Values of Random CDP Matrices
When the Original Saaty Scale is Used.
The results in this section reveal that CDP matrices (which are
assumed to be the result of a highly effective elicitation of the pertinent
pairwise comparisons) are very unlikely to be perfectly consistent. That
is, some small inconsistency may be better than no inconsistency at all!
(since no CDP matrix with CI is equal to zero was found when sets with
more than five elements were considered). This counter-intuitive situation
forms another paradoxical phenomenon which, however, can be explained
theoretically by the lemmas and theorems in this section.
Chapter 3: Quantification of Qualitative Data in MCDM Problems 43
30 elements were considered. For each such set 21 scales defined on the
interval [9, 1/9] (which correspond to ex values equal to 0, 5, 10, 15, ... ,
90, 95, 100) and 57 exponential scales which correspond to 'Y values equal
to 0.02, 0.04, 0.06, ... , 1.10, 1.12, 1.14 were generated. The previous 78
(Le., 21 + 57) scales will also be indexed as scale #1, scale #2, scale #3,
... , scale #78.
In Figures 3-3 and 3-4 the results of the evaluations of scales #1,
#2, #3, ... , #21 (also called Class 1 scales) in terms of the first and second
evaluative criterion, respectively, are presented. Similarly, in Figures 3-5
and 3-6 the results of the evaluations of scales #22, #23, #24, ... , #78 (also
called Class 2 scales) in terms of the first and second evaluative criterion,
respectively, are presented. It should be noted here that only 57 exponential
scales were generated because in this way values of 'Y from zero to around
to 1.00 can be considered. In the original Lootsma scales the value of 'Y
was 0.50 and 1.00. In this investigation all the scales with 'Y = 0.02, 0.04,
0.06, ... , 0.50, ... , 1.00, ... , 1.14 are considered. For each case of a value
of n and one of the 78 scales, 1,000 random test problems were generated
and tested according to the procedure described in the previous paragraphs.
The computational results of this investigation are depicted in Figures 3-5
and 3-6.
At this point it should be emphasized that the present simulation
results are contingent on how the random weight values were generated.
Other possibilities, such as assigning weight values from a nonuniform
distribution (such as the normal distribution), would probably favor other
scales. However, the uniform distribution from the interval [1, 0] was
chosen in this study (despite the inherited restrictions of this choice) because
it is the simplest and most widely statistical distribution used in similar
simulation investigations.
46 MCDM Method: A Comparative Study. by E. TriantaphylLou
Inversion
100 Rate (%) Scale 1
~-s~~::: Scale 2
90
80
70
Scale 20
60
Scale 21
50
40
30
20
10
Size of Set
Figure 3-3: fuversion Rates for Different Scales and Size of Set
(Class 1 Scales).
Chapter 3: Quantification of Qualitative Data in MCDM Problems 47
Indiscrimination
100 Rate (%)
90
80
70
60
50
40
30
20
10
Size of Set
Figure 3-4: Indiscrimination Rates for Different Scales and Size of Set
(Class 1 Scales).
48 MCDM Method: A Comparative Study. by E. Triantaphyllou
Inversion
Rate (%)
:l~~I~~~~~~ Scale 7778
100
"~
90
80
70
60
50
40
30
20
Scale 23
10 Scale 22
5 10 15 20 25 30
Size of Set
Figure 3-5: Inversion Rates for Different Scales and Size of Set
(Class 2 Scales).
Chapter 3: Quantification of Qualitative Data in MCDM Problems 49
Indiscrimination
100 Rate (%)
90
Scale 22
80
Scale 23
70
60
50
40
30
20
10
10 15 20 25 30
Size of Set
Figure 3-6: Indiscrimination Rates for Different Scales and Size of Set
(Class 2 Scales).
50 MCDM Method: A Comparative Study, by E. Triantaphyllou
Figures 3-3, 3-4, 3-5, and 3-6 depict how the previous 78 different
scales perform in terms of the two evaluative criteria given in Section 3.3.3.
Figures 3-3 and 3-4 depict the inversion and indiscrimination rates (as
derived after applying the two evaluative criteria) for Class 1 scales. That
is, for the scales defined in the interval [9, 1/9]. Similarly, Figures 3-5 and
3-6 depict the inversion and indiscrimination rates for the exponential scales
(or Class 2 scales). It is also interesting to observe here that when scales
from both classes of scales are evaluated in terms of the second evaluative
criterion (see also the indiscrimination rates in Figures 3-4 and 3-6), then
they perform worse when the size of the set with the entities to be compared
is between 8 and 12.
Clearly, there is no single scale which outperforms all the other
scales for any size of set. Therefore, there is no single scale, or a group of
scales, which is better than the rest of the scales in terms of both evaluative
criteria. However, the main problem is to determine which scale or scales
are more efficient.
Since there are 78 different scales for which there are relative
performance data in terms of two evaluative criteria, it can be concluded
that this is a classical multi-criteria decision making problem. That is, the
78 scales can be treated as the alternatives in this decision making problem.
The only difficulty in this consideration is how to assess the weights of
importance of the two evaluative criteria. Which evaluative criterion is the
most important one? Which is the least important? Apparently, these type
of questions cannot be answered in a universal manner.
The weights for these criteria depend on the specific application
under consideration. For instance, if ranking indiscrimination of the entities
to be compared is not of main concern to the decision maker, then the
weight of the ranking reversals should assume its maximum value (i.e.,
becomes equal to 1). However, one may argue that, in general, ranking
indiscrimination is less severe than ranking reversal. Depending on how
more critical ranking reversals are, one may want to assign a higher weight
to the ranking reversal criterion. If both ranking reversal and ranking
indiscrimination are equally severe, then the weights of the two criteria are
equal (i.e., they are set equal to 0.50).
For the above reasons, the previous decision making problem was
solved for all possible weights of the two evaluative criteria. Criterion 1
was assigned weight WI while the second evaluative criterion (or Criterion
2) was assigned weight w2 = 1 - WI (where 1 ;:::: WI ;:::: 0). In this way, a
total of 100 different combinations of weights were examined.
Chapter 3: Quantification of Qualitative Data in MCDM Problems 51
~
19 Class 1
with a = 90 II 69 Class 2
with y = 0 .96
21 Class 1 77 Class 2
Weight Wi
~ with a = 100 with y = 1.12
22 Class 2
with y = 0.02 II 78 Class 2
with y = 1.14
24 Class 2
with y = 0.06 D other scales
Size of Set
Scale No Description
II 78 Class 2
with y = 1.14 Scale No Description
Weight WI
77 Class 2
with y '" 1.12 lID 22 Class 2
with y = 0.02
23 Class 2
~
75 Class 2
with y = 1.08 with y = 0.04
I@I 24 Class 2
with y = 0.06 D other scales
1.00
0.80
0.60
0.40
o.zo
0.00
o 10 15 20 25 30
Size of Set
3.6 CONCLUSIONS
Both the theoretical and empirical analyses described in this chapter
revealed that the scale issue is a complex one. The various theoretical and
empirical results demonstrated that there is no single scale which can always
be classified as the best scale or as the worst scale for all cases. The
present investigation is based on the assumption that there exists a
real-valued rating of the comparison between two entities, that ideally
represents the individual preference. However, the decision maker cannot
express it, hence he/she has to use a scale with finite and discrete options.
In order to study the effectiveness of various scales, we furthermore
54 MCDM Method: A Comparative Study, by E. Triantaphyllou
assumed the scenario in which the decision maker is able to express his/her
judgments as accurately as possible. Under this scenario, it is assumed that
the decision maker is able to construct CDP matrices with pairwise
comparisons instead of the unknown RCP matrices. Based on this setting,
a number of computational experiments was performed to study how the
ranking derived by using CDP matrices differs from the real (and hence
unknown) ranking implied by the RCP matrices. The computational results
revealed that there is no single scale which is best in all cases. It should be
emphasized here that given an RCP matrix (and a scale with numerical
values), then there is one and only one CDP matrix which best
approximates it. Moreover, this CDP matrix mayor may not yield a
ranking different than the ranking implied by the RCP matrix.
An alternative assumption to the current one, which accepts that
there exists a real-valued rating of the comparison between entities, is to
consider the premise that maybe the real entity is the CDP matrix as given
by the decision maker. In this case the RCP matrix is maybe just an
illusion. In the later case the preference reversal leads to a very different
conclusion: if the CDP is the only "real thing", then it means that the
individual should point at the interval [1/~, l/Vi_l ] or [~-l' ~] rather than
to the values V;. That is, the preference reversal effects indicate that two
entities will be indifferent (since their ranking changes in the interval).
To determine the appropriate scale in a given situation certain
factors will have to be analyzed. First the number n, of the entities to be
compared, has to be known. Secondly, the relative importance of the two
evaluative criteria has to be assessed. These evaluative criteria deal with
possible ranking inversions and ranking indiscriminations that may result
when a scale is used. When these factors have been assessed, Figure 3-7
depicts the best scale for each case. Similarly, Figure 3-8 depicts the worst
scale for each case.
For instance, suppose that one has to evaluate the weight values of
a set with 15 members. Furthermore, suppose that ranking reversal is
considered, in a particular application, far more severe than ranking
indiscrimination. In other words, the weight of the first evaluative criterion
is considered to be higher than the weight of the second criterion. Using
this information, we can see that Figure 3-7 suggests to use scale #22 from
Class 2 (i.e., an exponential scale with parameter 'Y = 0.02). Moreover,
Figure 3-8 suggests that the worst scale for this case is scale #77 from Class
2 (i.e., an exponential scale with parameter 'Y = 1.12).
The same figures also indicate that the choice of the best or worst
scale is not clear under certain conditions. For instance, when the number
of members is greater than 15 and the two evaluative criteria are of almost
equal importance. In cases like this, it is recommended to experiment with
Chapter 3: Quantification of Qualitative Data in MCDM Problems 55
different scales in order to increase the insight into the problem, before
deciding on what is the best scale for a given application.
The computational experiments in this chapter indicate (as shown in
Figure 3-7) that exponential scales are numerically more efficient (i.e., more
stable) than the original Saaty scale (i.e., Scale 1). Only two Saaty-based
scales (i.e., scales #19 and #21) are present in Figure 3-7. In a matter of
fact, for sets with up to 10 entities scale #21 was best over a wide range of
weights. It is also worth noting that all the worst scales in Figure 3-8 came
from the exponential class (i.e., Class 2).
However, as the various examples in Section 3.3 suggest, human
beings seem to use exponential scales in many diverse situations. Therefore,
exponential scales appear to be the most reasonable way for quantifying
pairwise comparisons. The computational results in this chapter provide a
guide for selecting the most appropriate exponential scale for quantifying a
given set of pairwise comparisons.
Finally, it needs to be emphasized here that the scale problem is a
very crucial issue when the data are qualitative and thus they need to be
determined by using pairwise comparisons. Such qualitative data can
capture the pertinent information for many real life decision making
problems. An alternative point of view of the study described in this
chapter would be to perform in the future a similar investigation with
methods which do not use pairwise comparisons and thus are counterparts
of the pairwise comparison methodologies. However, since pairwise
comparisons provide a flexible and also realistic way for estimating these
type of qualitative data, it follows that an in depth understanding of all the
aspects of the scale problem is required for a successful solution to an
MCDM problem.
Chapter 4
remaining eigenvalues are close to zero. Thus, in order to find the weights
in the non-consistent case, one should find an eigenvector that corresponds to
the maximum right eigenvalue Amax. That is, to find the principal right
eigenvector W that satisfies:
A W = Amax W, where Amax ~ n.
Saaty proposed to estimate the reciprocal right eigenvector W by
multiplying the entries in each row of matrix A together and taking the n-th
root (n being the number of the elements to be compared). Since we desire
to have values that add up to 1, one needs to normalize the previously found
vector by the sum of the above values. If one wants to have the element with
the highest value to have a weight value equal to 1, then he/she needs to
divide the previously found vector by the highest value.
Under the assumption of total consistency, if the judgments are
gamma distributed (something that Saaty claims to be the case), the principal
right eigenvector of the resultant reciprocal matrix A is Dirichlet distributed.
If the assumption of total consistency is relaxed, then Vargas [1982] proved
that the hypothesis that the principal right eigenvector follows a Dirichlet
distribution is accepted if the consistency ratio is equal to or less than 10%.
The consistency ratio (CR) is obtained by first estimating Amax. Saaty
proposed to estimate the value of >--max by adding the columns of matrix A and
then multiplying the resulting vector with the vector W. Then, he uses what
he calls the consistency index (Cn of the matrix A. He defined CI as
follows:
CI = (>--max - n)/(n -1).
Then, the consistency ratio CR is obtained by dividing the CI by the
Random Consistency Index (RCI) as given in Table 4-1. Each RCI is an
average random consistency index derived from a sample of size 500 of
randomly generated reciprocal matrices with entries from the set {9, 8, 7, ... ,
2, 1, 1/2, ... , 117, 1/8, 1/9} to see if its CI is 10% or less. The number of
500 replications was used due to the limitations of the early computers when
the above concepts were established by Saaty. If the previous approach yields
a CR greater than 10 %, then a re-examination of the pairwise judgments is
recommended until a CR less than or equal to 10% is achieved.
n 1 2 3 4 5 6 7 8 9
min S = L L (aijwj -
i=j j=j
wji,
n
S.t. LWj = 1,
1=1
t ij = dij - 1,
or after using expression (4-10), above:
wll _ 1/wll •
Therefore, the corresponding n(n - 1)/2 errors are (after using
relations (4-11) and (4-12»:
(4-13)
62 MCDM Methods: A Comparative Study, by E. Triantaphyllou
Relations (4-13) and (4-14), when the data are consistent (i. e., all the errors
are equal to zero), can be written as follows:
B W = b. (4-15)
The vector b has zero entries everywhere except the last one that is equal to
1, and the matrix B (on the next page) has the following form (please notice
that blank entries represent zeros. Also, the first row and last column
indicate the cell numbers of the rows and columns, respectively, and are not
part of the matrix):
Chapter 4: Deriving Relative Weights from Ratio Comparisons 63
1 2 3 4 5 6 7 n-1 n
-1 a 1,2 1
-1 a 1,3 2
-1 a 1,4 3
-1
1
B 2
3
a2,n-l
-1 a 2 ,n n-2
-1 an-1,n 1
1 1 1 1 1 1 1 . . . 1 1
(4-16)
which, apparently, expresses a typical linear least squares problem.
If we use the notation described previously, then the quantity S in
expression (4-6) which is minimized in [Chu, et al., 1979] becomes:
n n n n
(4-17)
However, as the second example illustrates, the performance of the Federov,
et al., method is greatly dependent on the selection of the WI or w2 functions.
Now, however, these functions are further modified by (4-17).
Chapter 4: Deriving Relative Weights from Ratio Comparisons 65
Amax 4.226,
Cl = (4.226-4)/(4-1) = 0.053,
CR Cl/0.90 = 0.0837 < 0.10.
The formulation (4-15) that corresponds to this example is as follows:
-1 2/1 0.0 0
-1 0.0 1/5 0 0
Vl 0
-1 0.0 0 1/9
V2 0
0.0 -1 1/8 0 x = 0
V3 0
0.0 -1 0 1/9 V4 0
0.0 0.0 -1 1/4 1
1 1 1 1
The vector V that solves the above least squares problem can be shown to be
as follows:
V = (0.065841, 0.039398, 0.186926, 0.704808).
Hence, the sum of squares of the residual vector components is 0.003030.
The average squared residual for this problem is 0.003030/«4 x(4 - 1)/2)+ 1)
The second extensive numerical example uses the same data used
originally in the paper by Saaty [1977], and later in the two papers by Chu,
et al. [1979] and Federov, et al. [1982]. These data are next presented in
Table 4-2.
Table 4-3 presents a summary of the results (as found in the
corresponding references) when the methods described in Sections 4.2, 4.3,
and 4.4 are used. The power method for deriving the eigenvector was applied
as presented by Kalaba, et al. [1979]. In the last row of Table 4-3 are the
results obtained by using the least squares method under the human rationality
assumption (HR).
As it is shown in the last column of Table 4-3, the performance of
each method is very different as far the mean residual is concerned. The
same results also illustrate how critical is the role of the functions it,(X, Z)
and itiX, Z) in Federov's, et al. [1982] method. The mean residual obtained
by using the least squares method under the human rationality assumption is,
as expected, the smallest one (actually by 16%).
(1) 1 4 9 6 6 5 5
(2) 114 1 7 5 5 3 5
(3) 119 117 1 115 115 117 115
Saaty's eigenvector 0.429 0.231 0.021 0.053 0.053 0.119 0.095 0.134
method
Power method's 0.427 0.230 0.021 0.052 0.052 0.123 0.094 0.135
eigenvector
Chu's, et aI., method 0.487 0.175 0.030 0.059 0.059 0.104 0.085 0.097
Federov's, et al., Model 0.422 0.232 0.021 0.052 0.052 0.127 0.094 0.138
1 with 'lr, = 1
Federov's, et al., Model 0.386 0.287 0.042 0.061 0.061 0.088 0.075 0.161
2 with 'lr2 = 1
Federov's, et al., Model 0.383 0.262 0.032 0.059 0.059 0.122 0.083 0.152
2 with 'lr2 = I w;- Wj I
Federov's, et al., Model 0.447 0.229 0.021 0.051 0.051 0.120 0.081 0.130
2 with 'lr2 = (w; / w)
Least squares method 0.408 0.147 0.037 0.054 0.054 0.080 0.066 0.082
under the HR assumption
4-2 the vertical axis represents the average residual while the horizontal one,
the consistency index (CI). The CI index was selected such that the CR index
would be equal to 0.02, 0.03, 0.04, ... , 0.10. Due to the difficulty in
obtaining large random matrices with the CI (or the corresponding CR) value
being very small, the Supercomputer facilities at the Cornell University were
used (those experiments took place in 1988). Figure 4-1 depicts average
residuals when the human rationality assumption is used, while Figure 4-2
depicts residuals when the eigenvector method is used.
As the plots in Figures 4-1 and 4-2 illustrate, the average residual is
a function of both the consistency index (CI) of the data as well as the order
of the input matrix (Le., the number of elements to be compared in the
pairwise manner). Regression analyses suggested that for sets of a given
order the average residual is linearly related to the consistency index (CI).
Also, as the CI index of the data decreases so does the average residual. This
is expected because as the CI reaches zero the input data tend to be perfectly
consistent. From these results it can be seen that the average residuals are
significantly smaller when the method that is based on the human rationality
assumption is used.
The same plots also illustrate the importance of the number of the
elements to be compared. For small sets the CI index has to be small in
order to keep the average error at a low level. However, as the size of the
sets increases the CI index can be relatively larger and still achieve small
average errors per comparison. That is, the input data in the case of large
sets can be more inconsistent than in the case of smaller sets, as far the error
per comparison issue is concerned.
Chapter 4: Deriving Relative Weights from Ratio Comparisons 69
Table 4-4: Average Residual and CI Versus Order of Set and CR When the Human
Rationality (HR) Assumption and the Eigenvalue Method (EM) are Used.
The Results Correspond to 100 Random Observations.
Corresponding CR Index
Order
of
Method 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.10
Set
Used
(*) Values were not found for these cases due to the excessive requirement for CPU time.
70 MCDM Methods: A Comparative Study, by E. Triantaphyllou
0.8
n=4
0.7
0.6
0.5
Average 0.4
Residual
0.3
0.2
0.1
CI Values
Figure 4-1: Average Residual and CI Versus Order of Set When the Human Rationality
Assumption is Used (the Results Correspond to 100 Random Observations).
Chapter 4: Deriving Relative Weights from Ratio Comparisons 71
1.4
1.2
1.0
n=7
0.8
Average
Residual
0.6
0.4
0.2
CI Values
Figure 4-2: Average Residual and CI Versus Order of Set When the Eigenvalue
Method is Used (the Results Correspond to 100 Random Observations).
72 MCDM Methods: A Comparative Study, by E. Triantaphyllou
4.8 CONCLUSIONS
The eigenvalue method, as proposed by Saaty, for evaluating the
relative weights of entities that share some mutual characteristic of interest has
captured the interest of many researchers. However, the findings of the
present chapter demonstrate that there is a new dimension to the weight
estimation problem. Although the CR value can be kept less than 10% (and
hence ensure satisfaction of the Dirichlet distribution criterion as described
previously) the mean residual can vary significantly.
The results in this chapter demonstrate that even with data that yield
CR values less than 10%, the CR (or the corresponding CI) value has to be
kept at low levels for small comparison sets and at somewhat higher levels for
large sets, even when the method which is based on the human rationality
assumption is used. However, it should also be stated here that Saaty's
method is the easiest one from the computational point of view.
As it was demonstrated in other chapters of this book, small changes
in the weight values can mean the difference between selecting one alternative
instead of another in many MCDM problems. Accurately estimating relative
weights of key decision components (alternatives and criteria) is crucial in
many real life MCDM problems. Thus, a deeper understanding of this
problem is very critical in MCDM theory and practice.
Chapter 5
(Please notice that the left-hand-side part of (ii) refers to (Ai' Ai) and not to
(Aj , Ai ).)
On the other hand, a dissimilarity relation d: U x U - R+ satisfies the
following properties:
(i) d(A i , A) d(Aj , A;), for all Ai' Aj E U,
(ii) d(Ai' A) ~ 0, for all Ai' Aj E U,
(iii) d(Ai' A;) 0, for all Ai E U.
is called a distance. This is the case presented in this chapter. Through out
this chapter the similarities among the entities are measured by a dissimilarity
(not similarity) relation, particularly by the distance named in the literature
as the "city-block-metric" [Mardia, 1979].
From the previous expressions it follows that the function d is a
symmetric one. Furthermore, in this chapter it is assumed that entities can
be compared because their similarities can be measured by means of
difference of degrees of a common characteristic or feature. We denote by
WA and WA , the degree that this common feature is present in the two entities
A and A I, respectively. The values WA and WA , can also be viewed as
membership values of the members of a fuzzy set defined in terms of a
common feature. More on fuzzy sets can be found in Chapters 4 and 12 of
this book or in [Dubois and Prade, 1980] and [Kaufmann, 1975]. Usually,
these membership values take on values in the interval [1, 0]. A value of 1
indicates that the feature is fully present, while a value of 0 indicates that the
feature is completely absent.
Using the definition of the two values ~ and WA , it follows that the
closer the two values ~ and WA , are with each other, the more similar the
two entities A and A I should be. Therefore, a dissimilarity function can also
be defined in terms of the degrees two entities share a common feature as
Chapter 5: Deriving Relative Weights from Difference Comparisons 75
follows:
d(A, A 1 = I WA - WA , I'
for any entities A, A lEU.
The main problem that this chapter examines is how to estimate the
similarity relation among any pair of n entities, concepts, or objects. It is
assumed that these n entities share some common feature and that these
similarity relations are viewed in terms of that common feature.
As input data for determining the previous similarity relations we will
use a set of pairwise comparisons. Since there are n(n - 1)/2 possible pairs
of entities, the maximum number of pairwise comparisons is also equal to n(n
- 1)12.
For each comparison the decision maker is asked to do his/her best
in estimating the similarity only between two entities A and A I at a time.
The answer of the decision maker is (as was the case earlier in Chapter 3) a
phrase from a finite set of linguistic phrases. Each such linguistic phrase is
pre-assigned to a numerical value which attempts to capture the numerical
value of the difference I WA - WA , I (where WA and WA , are as defined
above). This is done by using a scale (see also Chapter 3).
In the previous chapter pairwise comparisons have been used to
estimate the relative importance of n entities when they are examined in
terms of a common feature. In that context pairwise comparisons express
ratios of relative importance. Usually, the values of these comparisons are
not in the interval [1, 0]. In the present treatment, however, pairwise
comparisons express the relative similarity of a pair of entities. Furthermore,
they express differences instead of ratios and take values in the interval [1,
0]. The issue which is raised in the light of these pairwise comparisons of
relative similarity is how to combine all of them together and estimate the
actual similarity relations among any pair of the n entities.
The present chapter is organized as follows. The next section
introduces the concept of pairwise comparisons of relative similarity.
Moreover, the same section illustrates how a scale of discrete choices can be
used to quantify this type of comparisons. The third section describes the
main contribution of this chapter. It presents the methodology for processing
pairwise comparisons of relative similarity. The comparisons lead to the
formulation of a quadratic programming problem which minimizes the
errors associated with these comparisons. In this way, the actual similarity
relations among a number of entities can be estimated effectively. Finally,
the last section presents a summary of the contributions described in this
chapter and some possible extensions for more research in this area.
76 MCDM Methods: A Comparative Study, by E. Triantaphyllou
0.50 Similar
1 Completely dissimilar
Suppose that the real (and hence unknown to the decision maker)
value of the (i, j) pairwise comparison is equal to (Xi} (where (Xi} ~ 0). This
value (Xi} is equal to the absolute value of the difference (W; - llj), where Wi
and llj is the degree (similarity value) a certain feature is present in the i-th
andj-th entities, respectively. That is, the following is true:
(Xi} = (Xji = I W; -llj I . (5-1)
Since the decision maker (in his/her assessment of the value of the
78 MCDM Methods: A Comparative Study, by E. Triantaphyllou
n-l n
E E (1 - Xi)2, or equivalently:
i = 1 j = i+l
(5-5)
n-l n n-l n
L L
i = 1 j = i+ 1
Xi~ - 2x L L
i = 1 j = i+l
Xij'
In the above matrix the entry (1,2) is equal to 0.18 because 0.92 - 0.74 =
0.18. A similar explanation holds for the remaining entries in matrix A.
80 MCDM Methods: A Comparative Study, by E. Triantaphyllou
The decision maker cannot determine the exact values of the previous
comparisons. However, he/she can use the scale depicted in Table 5-1 to
quantify his/her judgments. If we asswne that the decision maker is always
able to make that selection from the scale which has a numerical value closest
to the corresponding actual value in matrix A then, the following matrix B
represents the pairwise comparisons which we assume that the decision maker
can derive for this example:
subject to:
Xkj akj + ~i aji - Xile aile = 0 (i)
(for any n ~ i > j > k ~ 1)
and all the Xg's ~ O.
4
- L A;xgj(XI2' X13 , Xl4' ... , X34 )·
;=1
- - - - ----
Then, any point (X12' X 13• X I4> .00' X34• AI' A2, A3• A4) which satisfies the
following relations:
6L 6L 6L 6L (5-6)
= - - = - - = .•. = - - = 0,
6 X34 6 Al 6 A4
is also an optimal solution to the previous quadratic programming problem.
For the present extensive numerical example expressions (5-6) indicate
that the quadratic programming system (i), above, is equivalent to the
following system of linear equations (given in matrix form):
82 MCDM Methods: A Comparative Study. by E. Triantaphyllou
1 0 0 0 0 0 -a 12 -a 12 0 0 X 12 1
0 1 0 0 0 0 a 13 0 -a 13 0 X13 1
0 0 1 0 0 0 0 a 14 a 14 0 X 14 1
0 0 0 1 0 0 -a23 0 0 -a23 X23 1
0 0 0 0 1 0 0 -a24 0 a 24 X24 1
x
0 0 0 0 0 1 0 0 -a34 -a34 X34 1
a 12 -a13 0 a 23 0 0 0 0 0 0 1..1 0
a 12 0 -a14 0 a 24 0 0 0 0 0 1..2 0
0 a 13 -a14 0 0 a 34 0 0 0 0 1..3 0
0 0 0 a 23 -a24 a 34 0 0 0 0 1.. 4 0
[ 1 , -A T
m
A, 0"
xl X I
A
= [! l.
0
(ii)
=
=
~}
_
0
or:
(iii)
Chapter 5: Deriving Relative Weights from Difference Comparisons 83
222 2 2 2
-(a I2 + a l3 + a23 ) -a l2 +a13 -az3
-(a I2 + a l4 + a2J
2 222 2 2
-a 12 -a 14 +~
2 2 222 2
+a13 -a 14 -(a I3 + a l4 + a3J -a34
2 2 222
-az3 +a24 -a34 -(~+~+a3J
The above matrix is of rank N-1. This is true because any column (or
row) is linearly dependent on the remaining columns (or rows) and the
remaining columns (rows) are linearly independent. Therefore, anyone of
the \ variables can be set to an arbitrary value and then solve for the
remaining N-1 variables. Then the Xu variables can be determined from the
first relation in system (iii).
For instance, suppose that in the current example we set A4 = O.
When the numerical data (that is, matrix B) of this example are use..Q., the
linear syste!!J. defined by the ~econd relation in (iii) yields the solution Al = -
0.087015, A2 = -0.066934.z..A3 = -0.147256 (and A4 = 0).
From the previous Ai values and the first relation in (iii) the following
optimal solution to the original quadratic programming problem is derived:
-
X 12
0.96921
X13 0.97590
-
X 14 1.12851
=
X23 0.98260
x24 0.96653
X34
0.95582
From the structure of the AAT matrix (as it was depicted earlier) it follows
that expression (5-7) is true if and only if the difference (AI - A") is equal to
the following vector:
+1
-1
+1 xt,
-1
where t is any real number.
Given the previous observation on the difference (AI - A"), the
structure of the matrix A\ and relations (iii), it follows that the following
relations are also true:
Xl - X" = AT (AI - A") = 0, or: Xl = X",
t
where Xl = i + AT AI, and X" = + AT A". In other words, altho.!!gh
system (iii) may have infinitely many Ai solutions, the optimal solution X is
unique.
As it was stated earlier, this is also the optimal solution to the
original quadratic programming problem. In general, if there are n entities
to be compared, then the resulting system of linear equations has C3n - 1 =
34
pairwise comparisons may yield the wrong results. This is the case when the
triangular property ai j :::;; ai k + ak j does not hold for all possible
combinations of the i, k, and j indices. From the definition of the pairwise
comparisons of relative similarity, however, it was assumed that the triangular
inequality should always be satisfied. This is introduced in order to capture
the intuitive feeling by many decision makers that there is a close relation
between the notion of similarity and that of distance (see, for example,
[Ruspini, 1991]). Therefore, if the decision maker reaches a situation in
which the triangular property does not hold for all possible combinations, then
some or all of his/her comparative judgments need to be revised until the
triangular property holds true for all combinations of the indices.
Another interesting issue here is to observe that the proposed
approach always reaches a feasible solution. This is indeed the case
because from the transformation of problem (i) into its equivalent form (iii)
it follows that the variables A.; (for i = 1, 2,_3, ... , N) can always be
calculated. Furthermore, the solution vector X cannot have· a negative
element (and hence be infeasible).
To see this cOI.!.sider the two relations in system (iii). If some element
in the solution vector X is negative, then the system (iii), and the fact that the
matrix A is formed from the coefficients of the constraints expressed as (5-4),
implies that the expression (that is, an element of the vector Ai): alrj + aij -
aile is negative for some n 2:: i > j > k 2:: 1.
However, the above situation can never occur because from the
triangular property aile 2:: aij + ajk (for any n 2:: i, j, k 2:: 1) and the fact that
aij = aji (for any n 2:: i, j 2:: 1) it follows that the expression akj + aij - aile
can never be negative. Therefore, the proposed approach always reaches a
feasible (and consecutively an optimal) solution.
5.3 CONCLUSIONS
This chapter presented an approach for estimating similarity relations
among n entities. Pairwise comparisons have been used intensively as the
means for extracting the pertinent data for many decision making problems
(as discussed in the previous chapters). In this way, imprecise judgments of
an expert can be processed and accurate estimates of the unknown parameters
of a problem can be derived.
In the past, pairwise comparisons have been used in estimating the
relative importance among the members of a set (of alternatives or criteria).
In that context pairwise comparisons estimate the ratio of the relative
importance of two entities when they are considered in terms of a property
86 MCDM Methods: A Comparative Study, by E. Triantaphyllou
In this chapter it is always assumed that the pairwise comparisons for the
same pair of entities is the same regardless of the matrix. This is the reason
why the comparison between the elements A3 and A4 is the same (i.e., equal
to 0.333) in both the Ml and Ml matrices.
It can be verified that both of the previous matrices are satisfactorily
consistent (i.e., their CI values are less than 0.10) and thus they can be used
to derive the relative priorities of the two groups of elements. In [Saaty,
1980], or in Section 4.2, an effective way for estimating the relative weights
of importance from reciprocal matrices is described. According to that
procedure, one has first to calculate the geometric means of each row and
then normalize these means so they add up to 1. When this procedure is
applied to the previous two Ml and M2 matrices, then the following two
vectors PI and P2 with the relative priorities are derived, respectively.
0.1855
0.9710 0.1634 ]
, and [ 0.5396 .
0.1855
0.2970
0.5318
It can be observed from the first vector that the ratio of the relative
priorities of the elements A3 and A4 is equal to 0.3488 (= 0.1855/0.5318),
while from the second vector the same ratio is equal to 0.3028 (= 0.1634/
0.5396). Therefore, the question we seek to answer in this example is what
are the relative priorities when all the five elements are considered together?
One can view the previous two matrices as parts of a larger matrix
which is defined on the entire five elements. When matrices Ml and Ml are
combined, then the following 5 x 5 matrix M is derived (where "*" indicates
an undetermined comparison):
90 MCDM Methods: A Comparative Study, by E. Triantaphyllou
..
Reciprocals of
missing com parisons
~ ". .
"
.. •
..
II,
II
In the next section two procedures are developed for estimating the
missing comparisons. The first procedure is a simple and straightforward
one, while the second procedure attempts an error minimization strategy and
it is based on a linear programming (LP) formulation.
L
k=n-"2+1 ajk
(n 1 + n2 - n)'
for i = 1,2,3, ... , n-~,
andj = (n\+1), (n\+2), (n\+3), ... , n. (6-3)
After these averages are calculated the missing entries of the entire
matrix have been estimated. Next, the eigenvector approach (or any other
pertinent approach like the ones discussed in Chapter 4) can be applied on the
complete matrix and thus the final weights of the n entities can be estimated.
Although the above averages can be calculated in a straightforward
manner, the above approach fails to capture the requirement for the following
relationship (6-4):
92 MCDM Methods: A Comparative Study. by E, Triantaphyllou
It should be noticed that relationship (6-4) follows directly from the fact that
au = lIaji , for any i,} = 1,2,3, .. " n,
"1
L
k = 11-"2+1 ajk
(n 1 + n2 - n)'
for i = 1, 2, 3, .. " n-~,
and} = (nl + 1), (nl +2), .. " n, (6-5)
LNl
ail;
k=N-N2+1 ajk
(n1 + n2 - n)
for i = 1, 2, 3, .. " n-~,
and} = (n l +l), (n l +2), .. ,' n, (6-6)
problem of estimating the missing entries (and thus determining the final
relative weights) is to attempt to minimize the sum of all the previous error
terms, This is in accordance with the implicit assumption that the decision
maker attempts to be as consistent in his/her judgements as possible, Since
the errors may be positive or negative, we would like to minimize the sum of
their absolute values, Thus, this consideration leads to the following LP
formulation:
n-~ n
Minimize f = L L I eij I +
i=1 j=n 1 +1
n n
I eiji' j' I
n-~ n-~
+ L L L L
i=1 j=n t +l i'=1 j'=n t +l
subject to:
nt
L a ik
k=n-~+1 ajk
(n 1 + n2 - n)
for i = 1,2,3, .. " n-~,
and} = (n 1 +1), (n 1 +1), ,." n,
(a i i' x ajj') x X;, j' + e/ j',
for i, il = 1,2,3, .. " n-~,
and},/ = (n 1 +1), (n 1 +2), (n 1 +3), .. ,' n,
where all Xy's ~ 0, and all eij and e/j' variables are unrestricted,
The above nij (unknown) variables need not to be confused with the sizes
94 MCDM Methods: A Comparative Study, by E. Triantaphyllou
(i.e., the cardinality values), denoted as n1 and ~, of the two previous input
sets.
Similarly to the above constraint transformations, the previous
objective function! can be modified as follows:
11-"2 II
! = L L [Pij + nij] +
i=1 j=1I1 +1
11-"2 II 11-"2 II
+ L L L i' L
i = 1 j =111 + 1 j' = 1 =111 + 1
This is true because if, say the actual term eij' has to be negative then
it can be easily seen (from the linear dependencies in the columns of the new
constraints) that the value of the variable Pij will be equal to zero (as a
nonbasic variable) while the variable nij will be greater than zero. In other
words:
Pij + nij,
and P iji'j' + Ni'j'
ij .
11-"2 II
Min ! = L L [Pij + nij] +
i= 1 j=1I1 + 1
11-"2 II 11-"2 II
+
+ LLLL
i'
i = 1 j = III + 1 j' = 1 = III + 1
[pf
y
j' N yi ' j']
subject to:
III
L a ik
k=II-"2+1 ajk
(n 1 + n2 - n)
for i = 1, 2, 3, .,., n-~,
andj = (n 1+1), (n 1+2), (n 1+3), ... , n.
Xij = (a ii , x ajj') x Xi'j' + P/j' - N/j',
for i, il = 1,2,3, ... , n-~,
Chapter 6: A Decomposition Approach for Deriving Weights from Comparisons 95
where all the Xii' Pii' nii' and P/' j', N/, j' variables are ~ O.
a 13 a 14 1.000 0.333
- + +
a53 a54 2.000 0.500
0.583.
2 2
Similarly, we get:
a23 a24
+
a53 a54
= 0.325.
2
and (all x a55 ) = 1.00, (a'2 X a55 ) = 2.00.
Therefore, the constraints of the LP formulation for this numerical example
are as follows:
X'5 0.583 + P'5 - n'5
X25 = 0.325 + P25 - n25
X'5 2.000 X25 + P'5 25 - N'5 25
X25 = 0.500 X'5 + P25 '5 - N25'5
X'5 1.000 X'5 + P 15 '5 - N'5'5
X25 = 1.000 X25 + P25 25 - N25 25
96 MCDM Methods: A Comparative Study, by E. Triantaphyllou
It can be observed at this point that the last two constraints are
redundant, since we can directly set P I5 15 = N I5 15 = P25 15 = N2525 = O. In
general, we can always set P;/ = N;/ = O. This observation suggests that the
LP formulation, defined in the previous section, can have fewer variables and
constraints than the ones used here.
From the previous discussions it follows that the LP formulation for
this example is as follows:
subject to:
X I5 P15 + n15 = 0.583
X 25 P25 + n 25 = 0.325
X15 2.000 P1525
X 25 - + N
15
25 = 0
X 25 0.500 X15 - P2515 + N25 15 = 0
and P 25 P 15 N 25 N 15 ~ O.
X 15, X 25• PI5' P25' n 15• n 25• 15' 25' 15' 25
0.1392
0.0722
p' 0.1350
0.4137
0.2398
Chapter 6: A Decomposition Approach for Deriving Weights from Comparisons 97
If the non-LP approach is used, then it can be easily verified that X/5
0.583, and X25 = 0.325. Therefore, M" and p" the matrix with the
pairwise comparisons and the vector with the relative priorities, respectively,
are as follows:
0.1478
0.0783
and P" = 0.1433
0.4390
0.1916
set: {9, 8, 7, ... , 2, 1, 112, ... , 1/7, 118, 1I9} (if the original Saaty scale is
to be used). That is, instead of the real (and thus unknown) value ex one is
able to determine ali such that:
the difference I ex - ali I is minimum,
and ali E {9, 8, 7, ... , 2, 1, 112, ... , 117, 118, 1I9}.
In other words, it is assumed here that one's judgments about the value of the
pairwise comparison of the i-th element when it is compared with the j-th one,
is so accurate that in real life it is the closest (in absolute value terms) of the
values one is supposed to choose from.
The matrix with the ali entries that we assume the decision maker is
able to construct has entries from the discrete and finite set: {9, 8, ... , 2, 1,
1/2, ... , 1/8, 1I9}. This second matrix is called the Closest Discrete
Pairwise (CDP) matrix. More on some intriguing properties of RCP and
CDP matrices can be found in Section 3.3.
For illustrative purposes suppose that the actual relative weights of a
set of five elements, denoted as {AI' A 2, A 3, A 4 , and A5}, are as follows:
0.1328
0.0745
n 0.1542
0.3888
0.2498
By using the previous data, the corresponding RCP matrix can be found to
be as follows:
This is true because, say entry (1,2) is equal to 1.783 (= Wl/W2 = 0.1328 /
0.0745). A similar interpretation holds for the remaining entries in this RCP
matrix. Given the previous RCP matrix it can be easily verified that the
corresponding CDP matrix is as follows:
Chapter 6: A Decomposition Approach for Deriving Weights from Comparisons 99
This is true because, say entry (1, 2), is equal to 2.000, since the value 2.000
(taken from the current scale on use) is the closest value to the corresponding
entry in the RCP matrix (i.e., 1.783). A similar interpretation holds for the
remaining entries in this CDP matrix. Given the previous CDP matrix the
relative weights (priorities) derived by using the eigenvector approximation
approach (see also Section 4.2) are:
0.1352
0.0743
P 0.1352
0.4143
0.2410
0.1392 0.1478
0.0722 0.0783
0.1350 and P2 0.1433
0.4137 0.4390
0.2398 0.1916
From the previous two sets of relative weights we can observe that the
weights derived by using the LP approach are closer to the weights when the
CDP matrix is considered (i.e., with the ones implied by vector P). However,
the ranking of the five elements, when the LP and non-LP approaches are
used, is different than the ranking implied by the weights in vector P in this
illustrative example.
100 MCDM Methods: A Comparative Study, by E. Triantaphyllou
3 3 4 20 83.33 2 2
3 3 5 0 60 15 5
3 3 6 0 40 82 100
3 4 5 12.5 80 7 5
3 4 6 0 60 21 11
3 4 7 0 42.86 88 100
3 5 6 8.33 80 8 10
3 5 7 0 61.9 27 21
3 5 8 0 46.43 94 100
3 6 7 5.88 80.95 5 8
3 6 8 0 64.29 38 27
3 6 9 0 50 92 100
3 7 8 4.35 82.14 21 19
3 7 9 0 66.67 40 25
3 7 10 0 53.33 98 100
3 8 9 3.33 83.33 15 16
3 8 10 0 68.89 48 41
3 8 11 0 56.36 75 45
3 9 10 2.63 84.44 23 18
3 9 11 0 70.91 57 32
3 9 12 0 59.09 99 100
3 10 11 2.13 85.45 34 27
3 10 12 0 72.73 53 42
3 10 13 0 61.54 99 100
3 11 12 1.75 86.36 33 30
3 11 13 0 74.36 74 48
3 12 13 1.47 87.18 33 35
4 4 5 33.33 90 0 0
4 4 6 9.09 73.33 12 9
4 4 7 0 57.14 21 9
4 4 8 0 42.86 93 100
4 5 6 23.08 86.67 0 2
4 5 7 6.67 71.43 14 15
4 5 8 0 57.14 33 18
4 5 9 0 44.44 94 100
4 6 7 16.67 85.71 3 10
4 6 8 5 71.43 28 18
4 6 9 0 58.33 47 33
4 6 10 0 46.67 100 100
4 7 8 12.5 85.71 7 5
4 7 9 3.85 72.22 32 21
4 7 10 0 60 58 40
4 7 11 0 49.09 98 100
4 8 9 9.68 86.11 13 16
4 8 10 3.03 73.33 23 25
102 MCDM Methods: A Comparative Study, by E. Triantaphyllou
4 8 11 0 61.82 72 72
4 8 12 0 51.52 100 100
4 9 10 7.69 86.67 5 13
4 9 11 2.44 74.55 47 35
4 9 12 0 63.64 76 49
4 9 13 0 53.85 100 100
4 10 11 6.25 87.27 18 23
4 10 12 2 75.76 51 42
4 10 13 0 65.38 88 61
4 11 12 5.17 87.88 24 26
4 11 13 1.67 76.92 55 44
4 12 13 4.35 88.46 11 19
5 5 6 42.86 93.33 0 1
5 5 7 17.65 80.95 9 11
5 5 8 5.26 67.86 24 16
5 5 9 0 55.56 56 28
5 5 10 0 44.44 98 100
5 6 7 31.58 90.48 4 2
5 6 8 13.64 78.57 12 18
5 6 9 4.17 66.67 38 28
5 6 10 0 55.56 59 35
5 6 11 0 45.45 100 100
5 7 8 24 89.29 4 4
5 7 9 10.71 77.78 18 16
5 7 10 3.33 66.67 35 32
5 7 11 0 56.36 98 100
5 7 12 0 46.97 89 79
5 8 9 18.75 88.89 6 8
5 8 10 8.57 77.78 26 36
5 8 11 2.7 67.27 53 45
5 8 12 0 57.58 84 62
5 8 13 0 48.72 90 75
5 9 10 15 88.89 11 14
5 9 11 6.98 78.18 24 29
5 9 12 2.22 68.18 57 45
5 9 13 0 58.97 95 66
5 10 11 12.24 89.09 11 24
5 10 12 5.77 78.79 43 27
5 10 13 1.85 69.23 55 49
5 11 12 10.17 89.39 13 18
5 11 13 4.84 79.49 43 50
5 12 13 8.57 89.74 18 21
6 6 7 50 95.24 0 0
6 6 8 25 85.71 3 11
6 6 9 11.11 75 19 30
6 6 10 3.45 64.44 44 34
6 6 11 0 54.55 79 51
6 6 12 0 45.45 99 100
6 7 8 38.46 92.86 2 2
6 7 9 20 83.33 10 15
Chapter 6: A Decomposition Approach for Deriving Weights from Comparisons 103
6 7 10 9.09 73.33 39 31
6 7 11 2.86 63.64 57 42
6 7 12 0 54.55 83 52
6 7 13 0 46.15 98 100
6 8 9 30.3 91.67 1 6
6 8 10 16.22 82.22 20 21
6 8 11 7.5 72.73 32 38
6 8 12 2.38 63.64 65 50
6 8 13 0 55.13 97 66
6 9 10 24.39 91.11 5 10
6 9 11 13.33 81.82 20 30
6 9 12 6.25 72.73 45 43
6 9 13 2 64.1 76 51
6 9 11 20 90.91 10 13
6 10 12 11.11 81.82 20 20
6 10 13 5.26 73.08 47 54
6 11 12 16.67 90.91 6 8
6 11 13 9.38 82.05 29 35
6 12 13 14.08 91.03 14 24
7 7 8 55.56 96.43 0 1
7 7 9 31.25 88.89 6 8
7 7 10 16.67 80 16 22
7 7 11 7.69 70.91 45 36
7 7 12 2.44 62.12 68 56
7 7 13 0 53.85 93 66
7 8 9 44.12 94.44 2 3
7 8 10 25.64 86.67 14 20
7 8 11 13.95 78.18 26 23
7 8 12 6.52 69.7 46 40
7 8 13 2.08 61.54 77 55
7 9 10 35.71 93.33 3 7
7 9 11 21.28 85.45 14 20
7 9 12 11.76 77.27 30 32
7 9 13 5.56 69.23 66 52
7 10 11 29.41 92.73 8 12
7 10 12 17.86 84.85 16 23
7 10 13 10 76.92 41 48
7 11 12 24.59 92.42 7 19
7 11 13 15.15 84.62 20 24
7 12 13 20.83 92.31 9 9
8 8 9 60 97.22 0 0
8 8 10 36.59 91.11 4 6
8 8 11 21.74 83.64 13 22
8 8 12 12 75.76 38 40
8 8 13 5.66 67.95 51 47
8 9 10 48.84 95.56 4 4
8 9 11 30.61 89.09 12 14
8 9 12 18.52 81.82 19 25
8 9 13 10.34 74.36 46 38
8 10 11 40.38 94.55 3 7
104 MCDM Methods: A Comparative Study, by E. Triantaphyllou
8 10 12 25.86 87.88 14 25
8 10 13 15.87 80.77 23 31
8 11 12 33.87 93.94 7 10
8 11 13 22.06 87.18 10 26
8 12 13 28.77 93.59 8 10
9 9 10 63.64 97.78 1 1
9 9 11 41.18 92.73 7 13
9 9 12 26.32 86.36 17 21
9 9 13 16.13 79.49 24 38
9 10 11 52.83 96.36 1 8
9 10 12 35 90.91 14 23
9 10 13 22.73 84.62 17 26
9 11 12 44.44 95.45 4 15
9 11 13 30 89.74 13 22
9 12 13 37.84 94.87 3 6
10 10 11 66.67 98.18 0 0
10 10 12 45.16 93.94 4 10
10 10 13 30.43 88.46 23 26
10 11 12 56.25 96.97 4 3
10 11 13 38.89 92.31 8 16
10 12 13 48 96.15 3 9
11 11 12 69.23 98.48 1 0
11 11 13 48.65 94.87 6 12
11 12 13 59.21 97.44 3 4
12 12 13 7l.43 98.72 1 0
Chapter 6: A Decomposition Approach for Deriving Weights from Comparisons 105
For instance, consider the seventh row in Table 6-1 (part A). This
row has the numbers: [3, 5, 6, 8.33, 80, 8, 10]. In this case the size of the
original set is equal to 6. Also, the same data indicate that the first sub-set
is comprised of the first 3 elements, while the second one is comprised of the
last 5 elements. From the values of n, nj, and n2 we can verify that the
number of the available comparisons is equal to 12, the number of common
comparisons is 1, and the number of all possible comparisons is equal to 15.
Therefore, from the available comparisons, 8.33 % are common, and the
available comparisons represent 80.00% of all possible comparisons. When
the LP approach was applied on 100 randomly generated test problems with
the previous characteristics, in 8 cases (i.e., 8 % of the cases) the derived
ranking was different than the one implied by the corresponding CDP
matrices. For the case of using the non-LP approach, however, the same rate
was equal to 10% (or 10 cases out of the 100 random replications).
These results are also plotted in Figures 6-2 through 6-7. Figures 6-2
and 6-3 depict the error rates, for sets of different size, when the LP and non-
LP approaches are used, respectively. In these figures the error rates are
presented as a function of the size of the set (i.e., the n value) and the percent
of available comparisons (i.e., the fifth column in Table 6-2, all parts).
Figures 6-4 and 6-5 also depict the error rates for different sets, but now the
percent of common comparisons (i.e., the fourth column in Table 6-2) are
used instead. Figures 6-6 and 6-7 present the error rates of the two
approaches when the results are in terms of the averages of all set sizes. This
is the reason why there are only two curves in Figures 6-6 and 6-7; one
represents the performance of the LP approach, while the other one the
performance of the non-LP approach. Figure 6-6 refers to the available
comparisons, while Figure 6-7 refers to the common comparisons.
From the previous results a number of interesting conclusions can be
derived. First of all, as it is natural the percent of common comparisons
increases with the number of available comparisons. From Figures 6-2 to 6-5
we can observe that for larger set sizes the error rates are lower.
Furthermore, the LP approach yields, on the average, smaller error rates.
The effectiveness of the two approaches becomes more transparent in Figures
6-6 and 6-7. We can observe that when the percentages of available or
common comparisons are low, then both methods are unreliable, since they
result in too high error rates. For instance, if the percent of available
comparisons is less than 50 %, then both methods yield error rates higher than
50% (see also Figure 6-6).
106 MCDM Methods: A Comparative Study. by E. Trianraphyllou
o? 0 Q
lifJ ::
::
r
rfJ
(J
q
II
.,0
fIJ
c,. 1J
~
~
,..
Li
y
!/J
fJ
It
~
10
. 9'
Figure 6-2: Error Rates Under the LP Approach for Sets of Different Size as
a Function of the Available Comparisons.
Chapter 6: A Decomposition Approach for Deriving Weights from Comparisons 107
o
Q
i
?,.
Q? o
r,.!. :.,: : f
,rfl
Irtl
I?
i)
II II I
rjJ ~ i ::
~ ::
to ~j
~
c.- 1J
~
.-
~
a fJ
!/
I/J
II
fJ
10
o·
Figure 6-3: Error Rates Under the Non-LP Approach for Sets of Different
Size as a Function of the Available Comparisons.
108 MCDM Methods: A Comparative Study . by E. Triantaphyllou
\rfJ
rjJ
~
to
UJ
c... 1J
!l
...
J/J
t
6y 1} ~
~
fJ
0
10 0
Figure 6-4: Error Rates Under the LP Approach for Sets of Different Size as
a Function of the Common Comparisons.
Chapter 6: A Decomposition Approach for Deriving Weights from Comparisons 109
9
9
9 o
\rfJ
gO
tJ
l'
ffJ
~
... IjJ ~
..
!1
IP ff
t ?,
6
!/ fJ I ?
fJ ?
10 0 0 o o
Figure 6-5: Error Rates Under the Non-LP Approach for Sets of Different
Size as a Function of the Common Comparisons.
110 MCDM Methods: A Comparative Study. by E. Triantaphyllou
100
90 o ,.''-''-''-.-''~·---"-:··" ·" -, l, -, -" · ·, ·" , , , ·, -l·!, , ·, " ' ' ' ' -' ' ' '-:r-' ' '- -'-' ' ' ' '
Non-LP apprdach
-""-""'-""''' ':'''f--:;~-'''' . -.. "-1":7-":'"' ·rp"~pp-roiiCh·--. -t. -"·, , · " ."""'"' ' r' ' ' "" "--.._,,
". : 0
" 0 i i i i
80
! I 0 ! l
-------t·-.. . .·. . . .---·-t·--,---· . -_ ..--·-i----·-------
0 !
70 .------.--. --t......-:.. - i - _.. -
! A , A . l 0 .1 : I
r-.
; ' .to I 0 I l !
~ 60
(/)
1!0
.... 50
....
0
....
....
w 40
000
30
20
J
!
I
t
o !
. !
0
-------·--··'t. -------------:--r---·--_. . ,· . -.. .'1'-:'·'··. . ··'o-'--: ..
! o~.. ! ..
: OJ
..
,
~···,t,-: ·
tl·
..
- -0-
'"
~.. -a'.
~ao· ! :
f
I
- 6··-..·-···--..--·
· •
10 . - . .- . "". . ."L"""""-,,--I--,,-.-.
! . I . """.,,.-I.!. ~.-~"".-""-.,,
" ,o""l.:~_".:. . ;-<>~":'F}o
~ :. _..~"""~ . . . .". 0
! i ! . o..·!: 0 0
0
40 50 60 70 80 90 100
Figure 6-6: Error Rates for the two Approaches as a Function of the
Available Comparisons.
Chapter 6: A Decomposition Approach for Deriving Weights from Comparisons 111
100
90
80
70
.,.....
~
'-"
60
III
!!0
..... 50
.....
0
l.
.....
w 40
30
20
10
0
a 10 20 30 40 50 60 70 80
Figure 6-7: Error Rates for the Two Approaches as a Function of the
Common Comparisons.
112 MCDM Methods: A Comparative Study, by E. Triantaphyllou
However, for error rates at lower levels, say less than 30 %, the LP
approach is always the best method. This is the case when the percent of the
available comparisons is 80 % or higher. This is also the case when we
consider the percent of the common comparisons. Figure 6-7 suggests that
when the percent of common comparisons is equal to 10% or higher, then the
two methods have error rates less than 20 %. Moreover, the LP approach
consistently outperforms the non-LP approach. As it is anticipated, the
performance of the methods converges to being perfect (Le, the error rates
vanish) when the percentages of available or common comparisons approach
the upper limit of 100 %.
The fact that in these results the LP approach performs better than the
non-LP approach as the number of common comparisons increases, is in
direct agreement with the way the two approaches were designed. The main
difference of the two approaches is that the LP approach utilizes the presence
of common comparisons in minimizing the sum of the absolute errors. In the
non-LP approach, the issue of the common comparisons is totally ignored.
However, it is interesting to observe here that when the percent of common
comparisons is very high (e.g., more than 60%), then the non-LP approach
is as good as the LP approach. Moreover, the non-LP approach is extremely
simpler (one just needs to calculate the means in relations (6-3» compared to
the more CPU time consuming LP formulation.
6.7 CONCLUSIONS
This chapter examined the problem of decomposing a large set of
pairwise comparisons into two sub-sets. These two sub-sets may have some
common comparisons. These sub-sets may be defined by groups of elements
which are more similar within these two sub-sets, than when they are
considered all together. Two approaches were developed and tested. The first
is a simple approach, while the second one is based on an LP formulation.
Both approaches first attempt to estimate the missing comparisons and then
to derive the relative priorities and rankings by using all the pairwise
comparisons .
The LP formulation attempts to minimize the sum of the absolute
errors in the missing comparisons. The simulation results suggest that the LP
approach is consistently better than the non-LP approach. Furthermore, the
effectiveness of this approach improves with the amount of the common
comparisons. It should be also stated here that an interesting issue for future
research in this area might be how to decompose a large set of elements to be
compared into more than two sub-sets.
Finally, it is important to notice that the computational results in this
Chapter 6: A Decomposition Approach for Deriving Weights from Comparisons 113
chapter are contingent on the way the random data were generated. If the
random data are generated from other distributions, then it is possible for the
computational results to be different. However, it can still be highly
anticipated that the LP approach will perform better than the non-LP
approach, and moreover, its performance to improve with the amount of the
input data (since it processes the input data in a more comprehensive manner).
Since deriving relative weights from pairwise comparisons is a popular
approach in MCDM problem solving, finding ways of reducing the amount
of the required data, is of critical importance.
Chapter 7
evaluated and vice-versa. In other words, since criteria and alternatives are
the key entities in a given MCDM problem, one has to simultaneously keep
them in focus all the time during the decision making process.
For instance, if one considers the two criteria "cost" and
"functionality " in purchasing a product, then these two criteria have different
relative importance if the problem is to purchase a TV set or a new house.
In the first case a 20% price change may not be so critical, while in the case
of purchasing a new house a 20 % price change may be more detrimental.
Therefore, it is also natural to accept the premise that perfect independence
of the criteria and the alternatives seldomly exists. This is a highly debatable
issue in the decision analysis community and different authors have expressed
different positions (see, for instance, [Saaty, 1994, Chapter 8]). Thus, the
use of the proposed comparison question format is not totally new.
When the decision maker elicits all possible comparisons for a given
alternative, then the weight vector of this judgment matrix corresponds to a
normalized row (as opposed to a column with the traditional approach) of the
decision matrix. Therefore, by using the previous type of dual comparisons
the decision maker can determine all the rows of the decision matrix in a
normalized manner (i.e., the elements in each row sum up to one or are
divided by the maximum entry).
Let by (for i = 1, 2, 3, ... , m andj = 1, 2, 3, ... , n) denote the (i,
j) entry in the decision matrix normalized in terms of each row via a sequence
of m judgment matrices which are defined with dual comparisons. It can be
noticed that now it is required to form m judgment matrices of size n xn,
while in the traditional (primal) approach there are n such matrices of size
mXm each. Please recall that a typical MCDM problem is assumed to have
m alternatives and n decision criteria. Moreover, suppose that the decision
maker also forms the pairwise comparisons of a single judgment matrix in the
traditional fashion. That is, the decision maker forms a single mxm
judgment matrix. To help fix ideas, suppose that this is the judgment matrix
which examines all the m alternatives in terms of the first decision criterion.
Let ail (for i = 1, 2, 3, ... , m) be the elements of the corresponding weight
vector derived from the last judgment matrix.
Given the values of the m rows (e.g., the by values) normalized as
above and the values of the single normalized column (e.g., the ail values),
it is now straightforward to derive the elements of any column in the decision
matrix. It can be easily verified that this can be achieved by employing the
following formula (7-1):
(7-1)
118 MCDM Methods: A Comparative Study, by E. Triantaphyllou
It is important to note here that the sum of all the ayl elements in a given
column (except the first one) is not necessarily equal to one.
In general (if the k-th column is selected for normalization via the
extra judgment matrix), then the previous expression becomes:
(7-2)
Next the previously derived elements are normalized by dividing each one of
them by the sum of the entries of its column. Thus, the last expression
yields:
(7-3)
Theorem 7-1:
The percent (%) of change of the number of comparisons between the prime
and the dual problem is given by the following formula:
m-'(n'---_1~)..,.(m_-_n_-_l~)
_ (7-4)
x 100.
n(n - 1 + m(m - 1»
Proof:
As it was described in the previous paragraphs, a typical MCDM problem
with m alternatives and n decision criteria, under the traditional (prime)
approach requires the formation of a single n x n judgment matrix to derive
the criteria weights. It also requires the formation of n judgement matrices
of size m x m each to derive the relative weights of the m alternatives in terms
of each one of the n decision criteria. Thus, according to the traditional
Chapter 7: Reduction of Pairwise Comparisons Via a Duality Approach 119
1 (7-7)
-m(n - 1)(m - n - 1).
2
Thus, the percent (%) change of the number of comparisons between the
prime and the dual problem is given as expression (7-4). •
Corollary 7-1:
The dual problem requires less pairwise comparisons if the number of
alternatives in the problem is greater than the number of decision criteria plus
one.
Proof:
The above corollary follows directly from the fact that expression (7-7) must
be greater than zero. Thus, m - n - 1 > 0, or m > n + 1. This is true
because the value of n is always greater than 1 (and also m > 0). •
criteria (which describe their job performance). The rate of reduction on the
number of required comparisons is given by expression (7-4) of Theorem 7-1.
The next corollary describes a theoretical statement and it states that these
reduction rates converge to a fixed quantity when the number of criteria is
kept constant and the number of alternatives approaches to infinity
[Triantaphyllou, 1999].
Corollary 7-2:
The percent (%) of change of the number of comparisons between the prime
and the dual problem, for a given number of criteria N, approaches the value
(N - J)/N when the number of alternatives approaches infinity.
Proof:
This follows directly from expression (7-4) in Theorem 7-1, if one sets n =
N and then takes the limit when m approaches to infinity. Also, the function
in (7-4) is continuous and increases monotonically. Therefore, this limit can
also serve as an ugper bound on the reduction rate which can be achieved
by using the proposed duality approach. •
Criteria
CI C2 C3
Alts. (5/8 1/8 2/8)
Al 5 3 4
A2 2 5 3
A3 4 2 5
A4 2 4 1
As 3 3 2
The previous criteria weights can be viewed as the weight vector of the
following 3 x 3 judgement matrix:
1 5/1 5/2
1/5 1 1/2
2/5 2/1 1
Criteria
CI C2 C3
Alts. (5/8 118 2/8)
In standard AHP practice (see also Section 2.2.3), the decision maker
derives the normalized columns from a sequence of three judgment matrices
of size 5 x 5 each. For instance, the first column is derived from the
following primal judgment matrix (please note that all the comparisons are
perfectly consistent for easy demonstration of the main concepts in this
chapter):
122 MCDM Methods: A Comparative Study, by E. Triantaphyllou
Criteria
CI C2 C3
Alts. (5/8 118 2/8)
1 5/3 5/4
3/5 1 3/4
4/5 4/3 1
vector (besides the three normalized rows) of the decision matrix. In other
words, to form a single 5x5 judgment matrix. This extra matrix represents
the pairwise comparisons derived when the five alternatives are compared in
terms of anyone of the three decision criteria. For easy demonstration,
suppose that the decision maker chooses to compare the five alternatives in
terms of the first decision criterion (the case of using the second or the third
decision criterion can be developed in an identical manner). The
corresponding judgment matrix was provided in the previous sub-section.
Therefore, the normalized column is: (5/16, 2/16, 4/16, 2/16, 3/16?
The next step is to use the previous three normalized rows and the
normalized column to derive the normalized columns of the decision matrix.
It can be easily observed that only the last two columns of the decision matrix
need to be calculated in this example. When formula (7-3) is applied to
calculate the value of al2 , it turns out that its value is equal to 3/17. This is
true because from relation (7-3) the value of al2 should be equal to:
5/16 x3/12
5/12
( 5/16 X3/12) +(2/16 X5/10) +(4/16 X2/11) +( 2/16 X4f7) +( 3/16 X3/8)
5/12 2/10 4/11 2f7 3/8
3/16 = 3/17.
3/16 + 5/16 + 2/16 + 4/16 + 3/16
Criteria
C1 C2 C3
Alts. (5/8 1/8 2/8)
Obviously, the above matrix is identical to the matrix derived with the
primal approach. However, in the dual approach the decision maker needed
to form 5 (i.e., 1 + 3 + 1) judgment matrices which required a total of 28
(= 3(3-1)/2 + 5[3(3-1)/2] + 5(5-1)/2) pairwise comparisons. This represents
a reduction of 15.14 % from the total number of pairwise comparisons
required under the primal approach. Although this reduction may not seem
to be very significant, when the number of alternatives is much higher than
the number of decision criteria, then the benefits of using the dual approach
increase dramatically. This is further investigated in the numerical results
presented in the next section.
3000
. . . 1 l .
2500 . .
I -
·······················i·········N~~b·;~· ~f···c~~;~; ·i·~~~~···f; ···p;·i ·~;· *~b·i·;~·····"T········
...,
1 : :
-2~-
! ~~::1=I -s=~~=::--~-
500
Figure 7-1: Total Number of Comparisons and Reduction Achieved When the
Dual Approach is Used. The Number of Criteria n = 5.
6000
5000
Numb~r of Comparisons for Prlme !Problem l. . .·. . ·. ·. . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .1"0 . . . . . . . . . . . . . . . . . . . . . . . , . . . . . . . . . . . . . . . . . . . . . . . .
: ;
on • • • • • ~. . . . . . . . . . . . . . .
'"
.t:
o
JOOO
Q.
E
8
'0
~ 2000
.I:J
E
:;)
z
-1000
o 5 10 15 20 25 30 J5
Number of Alternatives in Problem
Figure 7-2: Total Number of Comparisons and Reduction Achieved When the
Dual Approach is Used. The Number of Criteria n = 10.
126 MCDM Methods: A Comparative Study. by E. Triantaphyllou
8000
6000
: :
o ........................,..............
~-~ ...- .
-2000 ~~~~--~~~--~~~~~~--~----~--~~
o 5 10 15 20 25 30 35
Figure 7-3: Total Number of Comparisons and Reduction Achieved When the
Dual Approach is Used. The Number of Criteria n = 15.
12000
.............. -- .._._ .. :..... _- ........ ---_.:.__ .............. "' .... ............ ........ ......--
8000
'"
c:
'0
; 4000
.&J
E
:>
z 2000
- 2000
0 5 10 15 20 25 30 35
Number of Alternatives in Problem
Figure 7-4: Total Number of Comparisons and Reduction Achieved When the
Dual Approach is Used. The number of Criteria n = 20.
Chapter 7: Reduction of Pairwise Comparisons Via a Duality Approach 127
4000 ~~~-'~~~~~~~-r~~~'-~~~,-~~~~~--~
8
'0 1000 """""""".. ,""."."."" ....
"-
OJ
.<:>
E
z'" o
-1000
o 5 10 15 20 25 30 35
Figure 7-5: Net Reduction on the Number of Comparisons When the Dual
Approach is Used. Results for Problems of Various Sizes.
100 ~~~~r-~~~~~~~~~--~~~~~~--~~~ __ ~~
Number bf Criteriia n =4
50 .. ~~~~~
'"
<::
o ........ , ........................ ...................... ,.
o .;..
.~
<>
0..
E
o -50 ............................................
<..>
'0
~ - 100
(.J
".,
=>
a:: -150
N
"",j" '''·''''''''·.. ".... ''t''·'' '·' '·. ·. ···,',·.·.-_i"",: .....,.. ,.. ,
-200 ......... i .......................... ;.. ........ " ....... ,._ .......... ,....... ................
-250
o 5 10 15 20 25 30 35
Figure 7-6: Percent (%) Reduction on the Number of Comparisons When the
Dual Approach is Used. Results for Problems of Various Sizes.
128 MCDM Methods: A Comparative Study, by E. Triantaphyllou
From the previous figures and analyses it follows that if the number
of alternatives is more than 35, then the number of comparisons under the
dual approach will assume a more quadratic rate of increase. However, one
may reasonably argue that most of the real life problems involve less than 35
alternatives (which is the upper limit in the previous plots).
On the other hand, the number of comparisons required under the
traditional (prime) approach noticeably increases quadratically in these
figures. The previous observations are a compelling reason why the duality
approach can be significantly beneficial to most real life decision problems
which use pairwise comparisons to elicit qualitative data.
Figure 7-5 depicts the net reduction on the number of comparisons
(given as expression (7-4)) achieved when the dual approach is applied, for
different size problems. Figure 7-6 presents the percent (%) reductions
achieved when the dual approach is used on problems of various sizes. As
before, these increases follow quadratic patterns on the size of the decision
problems. It is noticeable that in Figure 7-6 the reduction rates seem to
converge to a constant value when the number of alternatives increases.
Obviously, this is in direct agreement with Corollary 7-2.
7.5 CONCLUSIONS
criterion) .
The sensitivity analysis approach proposed in [Triantaphyllou and
Sanchez, 1997], which is also described in detail in this chapter, is
complementary to the one developed by Masuda. Thus, these two approaches
can be used together (since the proposed approach can focus on individual
judgments while the Masuda approach considers a single vector at a time).
Also, Armacost and Hosseini in [1994] presented a procedure for determining
the most critical criterion for a single level hierarchy AHP problem. Their
work does not explicitly determine what should be the smallest change in the
current weight of a criterion, such that the existing ranking of the alternatives
will change.
As a related comment it should also be stated here that the Expert
Choice (http://expertchoice.com/) software package which is based on
the AHP, performs a type of elementary sensitivity analysis. The user has the
option to graphically alter the weights of the decision criteria and see on the
screen how the rankings of the alternatives will change. However, the issue
of criteria sensitivity analysis is not studied systematically. Moreover, Expert
Choice does not offer any means for studying the effects of changes in the
measures of performance of the alternatives (which is part of the methodology
proposed in this chapter).
In decision making the weights assigned to the decision criteria
attempt to represent the genuine importance of the criteria. When criteria
cannot be expressed in quantitative terms (such as cost, weight, volume, etc.),
then it is difficult to represent accurately the importance of these criteria. In
a situation like the above one, the decision making process could be improved
considerably by identifying the critical criteria (the formal definition is given
later) and then re-evaluate more accurately the weights of these criteria. The
intuitive belief is that the criterion with the highest weight is the most critical
one [Winston, 1991; p. 754]. This may not always be true and in some
instances the criterion with the lowest weight may be the most critical one.
The decision maker can make better decisions if he/she can determine
how critical each criterion is. In other words, how sensitive the actual
ranking of the alternatives is to changes in the current weights of the decision
criteria. Thus, in this chapter we examine two closely related sensitivity
analysis problems. In the first problem we determine how critical each
criterion is, by performing a sensitivity analysis on the weights of the
criteria. This sensitivity analysis approach determines what is the smallest
change in the current weights of the criteria, which can alter the existing
ranking of the alternatives. In the second problem, we use a similar concept
to determine how critical the various performance measures of the alternatives
(in terms of a single decision criterion at a time) are in the ranking of the
alternatives.
Chapter 8: A Sensitivity Analysis Approach for MCDM Methods 133
weight Wi and performance measure aij is. These two problems are examined
in more detail in the subsequent sections.
8.3 PROBLEM 1:
DETERMINING THE MOST CRITICAL CRITERION
Definition 8.1:
Let Ok,jJ (for 1 ~ i < j ~ m and 1 ~ k ~ n) denote the minimum change
in the current weight W k of criterion Ck such that the ranking of alternatives
Aj and Aj will be reversed.
That is, the parameter Ofk,jJ expresses changes in relative terms. As it will be
shown later (in Theorem 8-1), it is possible for a given pair of alternatives
and a decision criterion the value of the critical change to be infeasible.
The most critical criterion is next defined in two possible ways (please
recall that from relations (8-1) alternative A I is always assumed to be the best
alternative). The first of these two definitions (Le., Definition 8-2) applies
when one is interested only in changes in the best alternative, while the
second definition (Le., Definition 8-3) applies when one is interested in
changes in the ranking of any alternative. Recall that Is I stands for the
absolute value function (e.g., I-51 = +5).
Definition 8-2:
The Percent-Top (or PT) critical criterion is the criterion which corresponds
to the smallest Iffk'lJI (for 1 ~ j ~ m and 1 ~ k ~ n) value.
Definition 8-3:
The Percent-Any (or PAl critical criterion is the criterion which corresponds
to the smallest Iffk,j,jl (for 1 ~ i < j ~ m and 1 ~ k ~ n) value.
136 MCDM Methods: A Comparative Study, by E. Triantaphyllou
Definition 8-4:
The criticality degree of criterion Cle• denoted as D I", is the smallest percent
amount by which the current value of w" must change, such that the existing
ranking of the alternatives will change. That is, the following relation is true:
Di = min
1:0: i<j :o:m
{I aLJ I}, for all n ~ k ~ l.
Definition 8-5:
The sensitivitv coefficient of criterion Cle• denoted as sens(Ck ), is the
reciprocal of its criticality degree. That is, the following relation is true:
1
sens(Ck ) = -, for any n ~ k ~ 1.
Di
lfthe criticality degree is infeasible (i.e., impossible to change any alternative
rank with any weight change), then the sensitivity coefficient is set to be equal
to zero.
The previous two Definitions 8-4 and 8-5 are based on changes in the
ranking of any alternative. However, one may be interested only in changes
in the ranking only of the best (top) alternative. For instance, in a problem
involving the purchase of a house, the focus is on the best house and the
ranking of all alternative houses may be of secondary interest. In cases like
the above one, one may want to use modifications of the criticality degree and
sensitivity coefficient concepts in which changes are only defined on the
ranking of the best alternative. It can also be observed from the above
definitions that since D Ik is always less than or equal to one, then it follows
that the value of sens(Ck) is always greater than or equal to one (except when
it is set to be equal to zero as described above).
Chapter 8: A Sensitivity Analysis Approach for MCDM Methods 137
For this case it is assumed that a decision maker uses the WSM or the
AHP method and he/she wishes to alter the existing ranking of the two
alternatives A] and A2 by modifying only the current weight w] of criterion C].
Currently, the following relation is true (as it was assumed in (8-1»: Pi ;;:::
P2 • In the Appendix of this chapter it is shown that the minimum quantity
0U2' needed to reverse the current ranking of the two alternatives A] and A2 ,
should satisfy the following relation:
(P2 - PI)
al ,l,2 <
(a21 - au)
, or :
(8-3a)
(P2 - PI)
al ,l,2 > -~--=---,
(a21 - au)
Furthermore, the following condition should also be satisfied for the new
weight w*] = w] - OJ,J,2 to be feasible:
o or:
o or: (8-3b)
a1,1,2 ;s; WI'
In these developments it is not required to have w*i ::;; 1 because these weights
are re-normalized to add up to one.
From the last two expressions (8-3a) and (8-3b), it can be seen that
sometimes the quantity OJ,J,2 may not have a feasible value. In other words,
it may be impossible to reverse the existing ranking of the two alternative A]
and A2 by making changes in the current weight w] of criterion C]. This
situation occurs when the value of the ratio:
( P2 - PI )
(a 21 - au)
is greater than w].
The previous considerations can be easily generalized and thus lead
to the proof of the following theorem which covers the general case (recall
that currently the following relation is assumed to be true from (8-1): Pi;;:::
Pj , for all 1 ::;; i ::;; j ::;; m).
138 MCDM Methods: A Comparative Study, by E. Triantaphyllou
Theorem 8-1:
When the WSM, or the AHP method is used, the quantity O/k. ;. j (jor 1 :::; i
< j :::; m and 1 :::; k :::; m), by which the current weight W k of criterion Ck
needs to be modified (after normalization) so that the ranking of the
alternatives A; and Aj will be reversed, is given as follows:
Furthermore, the following condition should also be satisfied for the value of
o/k,;,j to be feasible:
(Pj - Pi)
(8-4b)
(a jk - aik)
Table 8-1: Decision Matrix for the Numerical Example on the WSM.
Criteria
CI C2 Cj C4
Alts. (0.3277 0.3058 0.2876 0.0790)
Suppose that we want to apply the WSM (the case of the AHP is
identical since the data are already normalized). Then, by using formula (2-
1), in Section 2.2.1, the final preferences and the corresponding ranking of
the four alternatives are as shown in Table 8-2.
I Alternative
I Preference (P)
I Ranking
I
Al 0.3162 1*
A2 0.2768 2
Aj 0.2621 3
A4 0.1449 4
(0.2621 - 0.3162)
l>113 < , or
,, (0.4509 - 0.3088)
Table 8-3: All possible fJk.i,j Values (Absolute Change in Criteria Weights).
Pair of Criterion
Alternatives C1 C2 CJ C4
A 1 - A2 N/F -0.7023 0.1866 -0.0902
Note: N/F stands for Non-Feasible. That is, the corresponding fJ value does not
satisfy condition (8-4b).
coincidence that both definitions of the most critical criterion indicate the
same criterion (i.e., criterion C3) in this numerical example.
At this point it should be stated that if a decision maker wishes to
define the most critical criterion in absolute changes, then the previous two
definitions of Percent-Top (PT) and Percent-Any (PA) critical criterion
correspond to the Absolute-Top (AT) and the Absolute-Any (AA) critical
criterion, respectively. From Table 8-3 it can be easily verified that the AT
criterion is C4 and also, by coincidence, the AA criterion is C4 (the
corresponding minimum changes are boldfaced). Later, in Section 8.3.3,
some computational results indicate how frequently various alternative
definitions of the most critical criterion may point out to the same criterion.
When Definition 8-4 is used, then from Table 8-4 it follows that the
criticality degrees of the four criteria are: VI = 1-19.13341 = 19.1334, V z
= 48.7901, V3 = 9.1099, and V 4 = 32.5317. Therefore, the sensitivity
coefficients of the four decision criteria (according to Definition 8-5) are:
sens(CI ) = 0.0523, sens(Cz) = 0.0205, sens(C3) = 0.1098, and sens(C4) =
0.0307. That is, the most sensitive decision criterion is C3 , followed by the
criteria C I , C4 , and Cz in that order.
Table 8-4: All possible eik "ij Values (Percent Change in Criteria Weights).
Pair of Criterion
Alternatives CI C2 C3 C4
Al - A2 N/F -229.7 64.8818 -114.1772
Note: N/F stands for Non-Feasible. That is, the corresponding ()' value does
not satisfy condition (8-4b).
142 MCDM Methods: A Comparative Study, by E. Triantaphyllou
R A2
IT
(A1] -_ i=l (aa1i ]Wl .
2i
(8-5)
log ( IT (a 1y ]W' 1
Y =1 ~Y
5 1,1,2 > (8-6)
Cl all]
loaf
a21
The right-hand-side of the last relationship gives the minimum change
value needed to modify the current weight WI of criterion CI such that
alternative A2 will become more preferred than alternative Al (in the
maximization case). Similarly as in the previous section, this quantity needs
to satisfy condition (8-4b). The previous considerations can be easily
generalized and thus lead to the proof of the following theorem:
Theorem 8-2:
When the WPM method is used, the critical quantity 0 'k,iJ (jor 1 :::;; i < j :::;;
m and 1 :::;; k :::;; n), by which the current weight W k of criterion Ck needs to
be modified (after normalization) so that the ranking of the alternatives Ai and
Aj will be reversed, is given as follows:
log :iI
y = 1
a iy
a jy
w" 1 100
K = (8-7a)
loafail:) x Wl; •
"ajl:
Furthermore, the following constraint should also be satisfied:
Consider a decision making problem with the four alternatives AI' A2,
A3 , and A4 and the four decision criteria el , e2, e3 , and e4 • Please note that
this numerical example is different than the first one (and the two examples
which follow), in order to provide a wider exposure of various numerical
scenarios. Also, the decision matrices in these extensive numerical examples
are square (Le., m = n) of mere coincidence. The proposed procedures can
be applied to any size of decision matrix without any modification at all.
Next, suppose that Table 8-5 depicts the decision matrix for this numerical
example.
Table 8-5: Decision Matrix for the Numerical Example on the WPM.
Criteria
CI C2 C3 C4
Alts. (0.4504 0.1231 0.0848 0.3417)
Then, by applying the WPM approach (that is, by using formula (2-2)
in Section 2.2.2) the ranking of the four alternatives is as shown in Table 8-6.
144 MCDM Methods: A Comparative Study, by E. Triantaphyllou
The product expressed by (2-2) for alternative A] is greater than one for all
possible combinations which include A], thus, the most preferred alternative
is A]. Also, according to the weights of the four criteria, criterion C]
appears to be the most important one, because this is the criterion with the
highest weight.
Next, we consider the minimum quantity needed to alter the current
weight w4 , so that the current ranking of the two alternatives A1 and A2 will
be reversed. This quantity, denoted as K (and expressed as %), can be found
by using relation (8-7a) of Theorem 8-2 as follows:
Al - A2 1.0192 Al -+ I-st *
AJ - Aj 4.6082 A2 -+ 2-nd
Al - A4 5.3062 Aj -+ 3-rd
A2 - Aj 4.5216 A4 -+ 4-th
A2 - A4 5.2065
A3 - A4 1.1515
Note: * indicates the most preferred alternative (in the maximization case).
Pair of Criterion
Alternatives C1 C] Cj C4
Al - A] 21.21 -48.44 13.50 -11.04
Note: N/F stands for non-feasible. That is, the corresponding K value does not satisfy
condition (8-7b).
generated completely randomly. For the WPM and WSM cases the data were
generated randomly from the uniform distribution in the interval [9, 1].
According to the~~test problem generation approach described in
et
[Triantaphyllou, Lootsma, al., 1994], or in Chapter 3 of this book, the data
were generated as follows. First a random weight vector W was generated
such that the ratio of the largest to the smallest element was less than 9 (in
order to comply with the values in the Saaty scale which was used here).
From these weights the entries of the matrix with the actual pairwise
comparisons were determined by using the relationship aij = w/Wj. It is
assumed that the decision maker does not know these values. This matrix is
called in [Triantaphyllou, Lootsma, et al., 1994] and also in Chapter 3 the
Real and Continuous Pairwise (RCP) matrix. However, it was again
assumed here that the decision maker is capable of estimating the entries of
the RCP matrix by forming a matrix in which each entry in the RCP matrix
is replaced by a number which is as close as possible to the values allowed
in the traditional Saaty scale (i.e., the numbers from the set {9, ... , 2, 1,
1/2, ... , 1/9}). This is the Closest and Discrete Pairwise (CDP) matrix (see
also Chapter 3). Next, the eigenvector of the COP matrix is estimated as
before and the corresponding vector of the decisioll: matrix is formed.
For instance, if the real (and hence unknowri1 performance values of
three alternatives in terms of a single criterion are: (0.77348, 0.23804,
0.23848), then, the (1,3) entry of the corresponding RCP matrix is equal to
3.24342 (= 0.77348/0.23848). Thus, the corresponding COP element will
be equal to 3 (because this value is the closest one from the Saaty scale
values: {9, ... , 2, 1, 1/2, ... , 1/9}). More on this approach and some
interesting properties of the COP matrices can be found in [Triantaphyllou,
Lootsma, et al., 1994] and also in Chapter 3.
Two parameters were considered in these test problems. The first
parameter was the number of decision criteria. The second parameter was the
number of alternatives. The number of decision criteria was equal to 3, 5,
7, ... , 21. Similarly, the number of alternatives was equal to 3, 5, 7, ... ,
21. In this way we formed 100 (= 10 x 10) different combinations of
numbers of criteria and alternatives and 1,000 random test MCOM problems
were generated for each such combination. This simulation program was
written in Fortran, using the IMSL library of subroutines for generating
random numbers. These results are depicted in Figures 8-1 to 8-12.
For each test problem we examined whether the PA or the PT critical
criterion was the criterion with the highest or the criterion with the lowest
weight. The results of the computational experiments, when the relative
(percent) changes are considered, are depicted in Figures 8-1 to 8-4. Figures
8-5 to 8-8 illustrate the same concepts but when changes are expressed in
absolute terms. Finally, Figure 8-9 depicts some specific results when the
Chapter 8: A Sensitivity Analysis Approach for MCDM Methods 147
WPM is used. In the present study, we solved each problem using the WSM,
the WPM, the AHP and the ideal mode (revised) AHP method.
The four curves in each figure represent the results from each one of
the three different MCDM methods used plus one curve for the ideal mode
AHP. The most profound observation is that all MCDM methods generated
almost identical results. This is indicated by the fact that their curves in
Figures 8-1 to 8-8 are very close to each other.
Figures 8-1 to 8-8 indicate that it makes a significant difference
whether critical changes are expressed as percentages (Le., in relative terms)
or in absolute terms. When changes are expressed as percentages, then more
frequently the criterion with the highest weight is the most critical criterion.
This is true both when the concept of the critical criterion is defined in terms
of changes in the ranking of the top alternative or in terms of changes in the
ranking of any alternative. This is evident when one compares Figure 8-1
with Figure 8-2 and Figure 8-3 with Figure 8-4. The reverse situation occurs
when one defines change in absolute changes. That is, now more frequently
the most critical criterion is the criterion with the lowest weight. Figures 8-5
to 8-8 depict the corresponding results.
As it was anticipated, the sensitivity importance of any weight
(including the highest or the lowest) reduces gradually as the number of
decision criteria in a problem increases. In a matter of fact, when changes
are measured in relative terms (Le., as a percentage), then the lowest weight
is hardly ever sensitive in problems with more than 10 criteria (see also
Figures 8-2 and 8-4). On the other hand, the number of alternatives has only
a minor practical influence. This is indicated in Figures 8-11 and 8-12 in
which the bottom curve corresponds to problems with 3 alternatives and the
top curve to problems with 21 alternatives.
The question which is raised at this point is what kind of changes a
decision maker should consider: The ones defined as percentages or the
ones defined in absolute tenus? One may argue here that percentage
changes are the most meaningful ones. After all, a change, say of 0.03, does
not mean much unless one also considers what was the initial value (for
instance, was the initial value equal to 0.95 or to 0.05?).
Figure 8-9 depicts how frequently the AT and PT definitions pointed
out to the same criterion. Please recall that this situation also occurred in
some of the illustrative examples analyzed earlier.
Similarly, Figure 8-10 depicts how frequently the AA and PA
definitions pointed out to the same criterion. As expected, the frequency of
matching the top rank is always higher than matching all ranks. Moreover,
this distinction fades away as the number of criteria in a problem increases.
Finally, Figure 8-11 depicts how frequently all alternative definitions (i.e.,
AT, PT, AA, and PA) pointed out to the same criterion when the WSM
148 MCDM Methods: A Comparative Study, by E. Triantaphyllou
55
50
45
40 ~
~
>- 30
~
..
______ MCDM methods
~
o
c
~ 25
~
0-
m
20
~~
~ ~
15 ---- ~
10 --------------
5
o
3 5 7 9 11 13 15 17 19 21
Number of Criteria
Figure 8-1: Frequency of the time that the PT Critical Criterion is the Criterion with
the Highest Weight.
55
50
45
40
~35
"--'
~
15 ~
10 ~ L MCDM methods
~,
5
o ~-
3 5 7 9 11 13 15 17 19 21
Number of Criteria
Figure 8-2: Frequency of the time that the PT Critical Criterion is the Criterion with
the Lowest Weight.
150 MCDM Methods: A Comparative Study, by E. Triantaphyllou
55
50
0,
45
40 ~
~
>- 30
~ ~ MCDM methods
~
o
c
~ 25
0-
OJ
~ 20
~
15
---- ~
10
5
o
3 5 7 9 11 13 15 17 19 21
Number of Criteria
Figure 8-3: Frequency of the time that the PA Critical Criterion is the Criterion with
the Highest Weight.
55
50
45
40
"
/ MCDM methods
15
/
10
5
\,
o ~
3 5 7 9 11 13 15 17 19 21
Number of Criteria
Figure 8-4: Frequency of the time that the PA Critical Criterion is the Criterion with
the Lowest Weight.
Chapter 8: A Sensitivity Analysis Approach for MCDM Methods 151
55
50
45
40
>- 30
u ./ MCDM methods
c
~ 25
0-
Q) ~~
.: 20
15
~
10
~
~
5
o
3 5 7 9 11 13 15 17 19 21
Number of Criteria
Figure 8-5: Frequency of the time that the AT Critical Criterion is the Criterion with
the Highest Weight.
55
50
45
40 \\
\\
\\
\\ MCDM methods
,~ /
15 ~'Y
10
~
~~
5 =
o
3 5 7 9 11 13 15 17 19 21
Number of Criteria
Figure 8-6: Frequency of the time that the AT Critical Criterion is the Criterion with
the Lowest Weight.
152 MCDM Methods: A Comparative Study, by E. Triantaphyllou
55
50
45
40
~35
.........
>- 30
()
c
~ 25
(J
OJ
~ 20
/ MCDM methods
15
10
~
5
0
3 5 7 9 11 13 15 17 19 21
Number of Criteria
Figure 8-7: Frequency of the time that the AA Critical Criterion is the Criterion with
the Highest Weight.
55
I
50
45 \
40
\
\
\
~35
.......
>-30
()
c \. / MCDM methods
~ 25
(J
OJ
~ 20
,,/
15 ~
10
~
~
5
0
3 5 7 9 11 13 15 17 19 21
Number of Criteria
Figure 8-8: Frequency of the time that the AA Critical Criterion is the Criterion with
the Lowest Weight.
Chapter 8: A Sensitivity Analysis Approach for MCDM Methods 153
100
90
80
70
r--.
~
"-" 60 ~ / MCDM methods
>-
u
c 50 ~
-----
OJ ~
::J
cr
OJ
L.
40 --
L.-
30
20
10
0
3 5 7 9 11 13 15 17 19 21
Number of Criteria
Figure 8-9: Frequency of the time that the AT and PT Definitions point to the Same
Criterion.
100
90
80
70
r--.
MCDM methods
~ 60 /
>-
u
c
(l)
50 ~ ,/
-
::J
cr
-===;
40
::::----::::
Q)
....
LL
30
20
10
0
3 5 7 9 11 13 15 17 19 21
Number of Criteria
Figure 8-10: Frequency of the time that the AA and PA Defmitions point to the Same
Criterion.
154 MCDM Methods: A Comparative Study, by E. Triantaphyllou
55
50
45
40
~35
'--'
No of alternatives
>- 30
u
c
~ 25
0-
OJ
~ 20
15
10
0
3 5 7 9 11 13 15 17 19 21
Number of Criteria
Figure 8-11: Frequency of the time that the AT, PT, AA, and PA Definitions point to
the Same Criterion Under the WSM Method.
55
50
45
40
~35
'--'
No of alternatives
>- 30
u
c
~ 25
0-
OJ
~ 20
15
10
0
3 5 7 9 11 13 15 17 19 21
Number of Criteria
Figure 8-12: Rate the AT Critical Criterion is the one with the Lowest Weight for
Different Size of Problems Under the WPM Method.
Chapter 8: A Sensitivity Analysis Approach for MCDM Methods 155
8.4 PROBLEM 2:
DETERMINING THE MOST CRITICAL au
MEASURE OF PERFORMANCE
Definition 8-6:
Let Tij,k (jor 1 : : :; i < k : : :; m and 1 : : :; j : : :; n) denote the threshold value of
!!g, which is the minimum change which has to occur in the current value of
aij such that the current ranking between alternatives A; and Ak will change.
Definition 8-7:
The criticality degree of alternative Ai in terms of criterion~, denoted as
~ lij' is the smallest amount (%) by which the current value of aij must change,
such that the existing ranldng of alternative Ai will change. That is, the
following relation is true:
Definition 8-8:
Alternative AL is the most critical alternative if it is associated with the
smallest criticality degree. That is, if and only if the following relation is
true:
Definition 8-9:
The sensitivity coefficient of alternative Ai in terms of criterion C, denoted J
as sens(aij)' is the reciprocal of its criticality degree. That is, the following
condition is true:
-,'
1
dij (8-11)
for any m ~ i ~ 1, and n ~ j ~ 1.
If the criticality degree is infeasible, then the sensitivity coefficient is set to be
equal to zero.
From Definition 8-7 it follows that the smaller the criticality degree
~ lij
is, the easier is for the ranking of alternative Ai to change. Alternatively,
Definition 8-9 indicates that ranking changes are easier, as the sensitivity
coefficients sens(aij) are higher. Finally, when Definition 8-8 is combined
with Definitions 8-7 and 8-9, it indicates that the most sensitive alternative is
the one with the highest sensitivity coefficient. The next section shows how
to calculate these new important terms.
Chapter 8: A Sensitivity Analysis Approach for MCDM Methods 157
8.4.2.1 Case (I): When Using the WSM or the AHP Method
The Appendix of this chapter presents the highlights for a proof for
Theorem 8-3. This theorem provides the main formula used to calculate the
threshold 7 / iJ,k values and it is stated next.
Theorem 8-3:
When the WSM method is used, the threshold value 7 /iJ'k (in %) by which the
pelj'ormance measure of alternative Ai in terms of criterion ~, denoted as ai}'
needs to be modified so that the ranking of the alternatives Ai and Ak will be
reversed, is given as follows:
't.' .
1.1. k
< R, when i < k, or :
(8-lOa)
't .' .
1.1. k > R, when i > k.
where R is defined as follows:
100
R = x --.
Wj aij
Furthermore, the following condition should also be satisfied for the threshold
value to be feasible:
From the requirement to have: 0 ~ ail - 7 i,j,k and relation (8-8), the
new condition 0 ~ ail - ail X 7 /iJ,,JlOO is derived. Next, this leads to condition
(8-lOb). For the case of the AHP method it can be easily shown (see also the
Appendix of this chapter) that the corresponding theorem is as follows:
Theorem 8-4:
When the AHP method is used, the threshold value 7 /iJ,k (in %) by which the
measure of pelj'ormance of alternative Ai in terms of criterion ~ needs to be
modified so that the ranking of alternatives Ai and Ak will change, is given as
follows:
158 MCDM Methods: A Comparative Study, by E. Triantaphyllou
(Pi - P k ) 100
• .1. ----------------------x---. (8-12a)
'." k [Pi - P k + wAakj - aij + 1)] aij
Furthermore, the following condition should also be satisfied for the threshold
value to be feasible:
.~.
',J. k
:s; 100. (8-12b)
The sensitivity analysis of the aij values, when the WSM or the AHP method
is used, is next demonstrated in terms of an extensive numerical example.
Consider the decision matrix depicted in Table 8-8 (along with the
corresponding final preferences P;) of an application of the WSM model.
This illustrative problem has five alternatives and five decision criteria). The
AHP case can be developed in an analogous fashion.
Table 8-8: Decision Matrix and Initial Preferences for the Example.
Criteria
C, C, C3 C4 Cs Preference Values
Alts. (0.4146 0.0129 0.2958 0.0604 0.2164) Pi
When Theorem 8-4 is used on the above data, the corresponding 7 / i.j,k
threshold values are as in Table 8-9. The boldfaced entries in Table 8-9
correspond to the criticality degrees !l. lij (i.e., the smallest entry per column
in each row section, as given in Definition 8-7). The criticality degrees are
next summarized in Table 8-10.
To help interpret the entries in Table 8-10, consider anyone of them,
say entry (3,1) (i.e., 89.3). This entry indicates that 7 11,1,4 = 89.3%. That
is, the measure of performance a11 must be decreased by 89.3 % from its
current value (i.e., 0.3576) to (1 - 0.893)x0.3576, in order for aiternativeA4
Chapter 8: A Sensitivity Analysis Approach for MCDM Methods 159
(which is shown on the right column in this table) to become more preferred
than alternative Al (note that currently Al is more preferred than A4)' A
similar interpretation holds for the rest of the entries. Note that some of the
entries in the previous table are marked as NIP, because they correspond to
infeasible values (Le., condition (8-12b) in Theorem 8-4 is violated).
It can be noticed that the entries in Table 8-9 are greater than 100
only when the sign is negative. Recall that negative changes in reality mean
increases. If a rating becomes greater than 100, that is all right. In the case
of the criteria weights the numbers will be re-normalized to add up to one. In
the case of the au performance measures, the numbers may be re-normalized
(for instance, in the AHP model) or they may become greater than one (for
instance, in the WPM or WSM models). As before, the boldfaced numbers
indicate minimum values.
Table 8-9: Threshold values T 'jJ.k (%) in Relative Terms for the WSM/AHP Example.
Criterion 0
Alt.
Alt. (AJ C} Cz CJ C4 Cs At
A} 74.1 N/F N/F N/F N/F Az
A} N/F N/F N/F N/F N/F AJ
Al 89.3 N/F N/F N/F N/F A4
Al 96.1 N/F N/F N/F N/F As
------- --------------------------------------- --------
Az -157.9 N/F -1,677.6 N/F N/F Al
Az 5.4 N/F 35.9 79.4 N/F A3
Az 5.3 N/F 41.9 N/F N/F A4
Az 8.9 N/F 78.4 N/F N/F As
------- -------------------------------------- --------
AJ -1,535.7 N/F N/F N/F N/F Al
A3 -39.3 -355.2 -8.7 -52.5 -13.2 Az
AJ 8.0 38.9 1.1 8.3 2.2 A4
AJ 41.1 N/F 6.8 29.9 7.3 As
------- --------------------------------------- --------
A4 -286.1 N/F N/F N/F N/F Al
A4 -8.2 N/F -10.2 -163.0 -98.9 Az
A4 -1.7 -41.4 -1.1 -19.7 -11.5 A3
A4 5.9 N/F 5.3 65.9 40.4 As
------- ----------------------------------- --------
As -460.8 N/F -970.6 N/F N/F Al
As -20.7 N/F -44.4 -87.4 -21.2 Az
As -12.8 -544.7 -15.9 -29.7 -6.7 AJ
As 10.0 N/F 18.7 14.9 3.5 A4
160 MCDM Methods: A Comparative Study, by E. Triantaphyllou
Table 8-9 was next used to derive Table 8-10. From Table 8-10 it
follows that the most critical alternatives (according to Definition 8-8) are
alternatives A3 and A4 • This is true because these alternatives correspond to
the minimum criticality degrees (which happened to be identical and equal to
1.1) among all the values in Table 8-10. It can be noticed that the
corresponding alternatives from the right column in Table 8-9 are now within
the parentheses in the entries in Table 8-10. As before, boldfaced numbers
represent the corresponding minimum values. Finally, Table 8-11 presents
the various sensitivity coefficients (as given in Definition 8-9). It can be
noticed that if any entry in Table 8-10 was infeasible (Le., it was denoted by
the "_" symbol), then the corresponding sensitivity coefficient in Table 8-
11 is defined to be equal to zero.
Table 8-10: Criticality Degrees A lij (%) for each aij Performance Measure.
Criterion Cj
CI C1 C3 C4 Cs
AIt.(AJ
AI 74. 1(A 2 ) - - - -
Al 5.3(A.) - 35.9(A3) 79.4(A3)
-
A3 8.0(A4) 38.9(A4) 1. 1(A4) 8.3(A4) 2.2(A.)
A4 1.7(A3) 41.4(A 3) -1.1 (A 3) 19.7(A3) 11.5(A3)
Table 8-11: Sensitivity Coefficients sens(aij) for each aij Performance Measure.
Criterion Cj
CI C2 C3 C4 Cs
Alt.(AJ
AI O.014(A 2 ) 0 0 0 0
A2 O.189(A.) 0 O.028(A 3) O.013(A) 0
A3 O.125(A.) O.026(A 4) O.909(A.) O.121(A4) O.455(A.)
The Appendix of this chapter also presents the highlights for a proof
for Theorem 8-5. This theorem provides the main formula for calculating the
threshold values r/i,j,k when the WPM method is used and it is stated next.
Theorem 8-5:
When the WPM model is used, then the threshold value r/i,j,k (in %) by which
the performance measure of alternative Ai in terms of criterion ~, denoted as
aij' needs to be modified so that the ranking of the alternatives Ai and Ak will
be reversed, is given as follows:
I
t: ij,k > Q, when i > k, or:
(8-14a)
I
t: iJ,k < Q, when i < k.
where Q is defined as follows:
Furthermore, the following condition should also be satisfied for the value to
be feasible:
(8-14b)
Criteria
CI C2 C3 C4 Cs
Alts. (0.2363 0.1998 0.0491 0.2695 0.2453)
Please recall that in the WPM method normalization of the aij values is not
required. Then, by applying formula (2-2) (in Section 2.2.2), the current
ranking of the alternatives is as shown in Table 8-13.
Pair of
Alternatives Ratio
Ap - Aq R(ApIAq) Ranking
AI - A2 1.580 A J .... I-st
AI - Aj 2.415 A2 .... 2-nd
AJ - A4 2.692 Aj .... 3-rd
AI - As 6.152 A4 .... 4-th
A2 - A3 1.529 A5 .... 5-th
A2 - A4 1.704
A2 - As 3.893
Aj - A4 1.115
A3 - As 2.547
A4 - As 2.285
For instance, the entry (8,1), (Le., -2E +05) actually represents the value: -
2.0x 105 = -200,000. It can be argued here that very significant changes
(such as the ones represented in exponential format or those which measure
in terms of thousands of % change) are not realistic and thus practically they
can also be classified as "NIF" (e.g., Non-Feasible) cases. Finally, it can be
observed that the highlighted entries in Table 8-14 correspond to the criticality
degrees Ll lij (as given by Definition 8-7). The criticality degrees are best
summarized in Table 8-15.
Table 8-14: Threshold Values 7 /;jk (%) in Relative Terms for the WPM Example.
"
Criterion ~
Alt.
AIt(A,) CI C2 C3 C4 Cs At
Al 85.6 89.9 N/F 81.7 84.5 A2
Al 97.6 98.8 N/F 96.2 97.3 A3
Al 98.5 99.3 N/F 97.5 98.2 A4
Al N/F N/F N/F As N/F N/F
------- ----------------------------------- ------
A2 -593 -887 -IE+06 -446 -546 Al
A2 83.4 88.0 N/F 79.3 82.3 A3
A2 89.5 93.0 N/F 86.2 88.6 A4
A2 99.7 99.9 N/F 99.4 99.6 As
------- ----------------------------------- ------
A3 -4,072 -8156 -6E+09 -2,538 -3,540 Al
A3 -502 -736 -6E+05 -383 -464 A2
A3 36.8 41.9 89.0 33.1 35.7 A4
A3 98.1 99.1 N/F 96.9 97.8 As
------- ----------------------------------- ------
A4 -6,501 -14,105 -6E+1O -3,844 -5,562 Al
A4 -853 -1,339 -5E+06 -622 -777 A2
A4 -58 -72 -811 -50 -56 A3
A4 97.0 98.4 N/F 95.3 96.6 As
------- ----------------------------------- ------
As -2E+05 -9E+05 -IE+ 18 -8E+04 -2E+05 Al
As -3E+04 -9E+04 -IE+ 14 -2E+04 -3E+04 A2
As -5,124 -10,672 -2E+1O -3,113 -4,420 A3
As -3,202 -6,161 -2E+09 -2,049 -2,805 A4
164 MCDM Methods: A Comparative Study, by E. Triantaphyllou
From Table 8-15 it follows that the most critical alternative (according
to Definition 8-8) is alternative A 3 . This is true because this alternative
corresponds to the minimum criticality degree (e.g., 33.1) among all the
values in Table 8-15. Table 8-16 presents the various sensitivity coefficients
(as defined by Definition 8-9). Note that if there was an infeasible entry
(denoted by the "_" symbol) in Table 8-15, then the corresponding
sensitivity coefficient in Table 8-16 is defined to be equal to zero. A
comparison of Tables 8-10 and 8-15 (or Tables 8-11 and 8-16) indicates that
the current WPM example is much more robust than the WSMI AHP
example. This is true because the sensitivity coefficients in the WPM
example are much smaller. This is a consequence of the specific data used
in these two examples and the different nature of the AHP and WPM
procedures.
Table 8-15: Criticality Degrees d lij (in %) for each aij Measure of Performance.
Criterion 0
AIt.(A)
C1 C2 C3 C4 Cs
A1 85.6(A2) 89.9(A2)
- 81.7 (A 2) 84.5(A 2)
A2 83.4(A 3) 88.0(A 3) 79.3(A J) 82.3(A J)
-
A3 36.8(A4) 41.9(A4) 89.0(A 4) 33.1(A 4) 35.7(A4)
A4 58.0(A3) n.0(A3) - 50.0(AJ) 56.0(A 3)
As - - - - -
Table 8-16: Sensitivity Coefficients sens(aij) for each aij Measure of Performance.
Criterion 0
C] C2 CJ C4 Cs
AIt.(AJ
A1 0.012(A 2) O.oII(A 2) 0 0.012(A 2) 0.012(A2)
A2 0.012(AJ) O.oII(A 3) 0 0.0 13 (A J) 0.012(AJ)
A3 0.027(A 4) 0.024(A 4) 0.011 (A4) 0.030(A 4) 0.028(A 4)
A4 0.0 17(AJ) 0.014(A 3) 0 0.020(A 3) 0.018(A 3)
As 0 0 0 0 0
Chapter 8: A Sensitivity Analysis Approach for MCDM Methods 165
8.5 CONCLUSIONS
By knowing which data are more critical, the decision maker can more
effectively focus his/her attention on the most critical parts of a given MCDM
problem.
Another area of application is during the phase of gathering the data
for a MCDM problem, given a limited budget. Often, in real life applications
of MCDM, data are changeable and cannot be precisely determined. In such
cases it makes more sense to determine with higher accuracy the weights of
the criteria (as well as the aij measures of performance) which are more
critical and with less accuracy the less critical weights. A sensitivity analysis,
contacted at an early stage, may reveal which Wj and aij values have a
tendency to be more critical to the final (ranking) decisions. Therefore, these
data can be determined with higher accuracy at a second stage. Next, a new
sensitivity analysis cycle can be initiated again. This process can be repeated,
in this stepwise manner, for a number of times until the entire budget is used
or the decision maker is satisfied with the robustness of the final results.
The three (and a variant of one) MCDM methods examined in this
chapter have been fuzzified by Triantaphyllou and Lin [1996] and are also
presented in Chapter 13. Thus, a natural extension of this research is to
develop a sensitivity analysis approach for cases in which the data are
described by fuzzy numbers. An additional area of possible extension is to
extend these results to AHP problems with multiple hierarchies.
Chapter 8: A Sensitivity Analysis Approach for MCDM Methods 167
APPENDIX TO CHAPTER 8
W; = WI - 01.1.2' (AI)
As it was stated in the definitions given in Section 1.1, usually the
criteria weights are normalized to add up to one. Therefore, the new
normalized weights, denoted as W Ii' will be as follows:
WI*
wII
WI* + w2 + .•. + wn
Wz
wI2
WI* + W2 + + Wn (A2)
WIn =
W; + W2 + ... + wn
Given the new weights W Ii (for i = 1, 2, ... , n) it is necessary to
express the conditions for the new ranking. Let p I1 and p I2 denote the new
final preference values for the two alternatives A1 and Ab respectively. Since
it is desired that the new ranking of the previous two alternatives to be
reversed, the following relation should be satisfied:
(A3)
168 MCDM Methods: A Comparative Study, by E. Triantaphyllou
Since the WSM (or AHP) model is used, the previous two preference
values are given by the formulas:
/I /I
-
Lwja Ij Lwj a2j
WI* all j=2 WI* a21 j=2
+ < +
/I /I /I /I
WI* + LWj
•
WI + LWj WI* + LWj WI* + LWj
j=2 j=2 j=2 j=2
/I /I
-
(P2 - PI)
() 1,1,2 < if (a21 > all)' or :
(a21 - all)'
(A5)
(P2 - PI)
()
1,1,2
> if (a21 < all)·
(a21 - all)'
The above derivations can easily be expanded and are generalized in Theorem
8-1.
Chapter 8: A Sensitivity Analysis Approach for MCDM Methods 169
R(Al)
A2
= IT
j=l
(a 1j
a2j
)W1• (A6)
RI[~) -
RI[ ~:) where: (A7)
rl = (all),
a
r2 = (aa I2 ), ... , and rn
l2 22
... ,
w1
where:
=
w··w" w2
= w;
=
+
w··
w wn .
w··
2 + ••• +
-
.
WI W2
- Wft
1og(r WI r r
a > 1 Z
W2 ••• Wft)
n
1,1,2 logr1
When relation (A7) is used on the previous expression, the following relation
is obtained after some simple algebraic manipulations:
log ( II
n a )W1
--.!l 1
y=1 a Zy
(A8)
109( au]
a Zl
The above derivations can easily be expanded and are generalized in Theorem
8-2.
o
o
~
I
a34 ,
From the previous relations, it can be seen that the 7 3,4,2 value must be less
than or equal to a34 in order to have a feasible meaning.
Next, let us assume that the measure of performance that we wish to
alter is again a34 , but now we want to reverse the ranking between alternatives
A3 and A5 (observe that now: P3 ~ P5 from (8-1». Then, by following an
approach similar to the previous one, one can derive that in order to reverse
the current ranking of the two alternatives A3 and A5 the threshold value 73,4,5
should satisfy the following relation:
Moreover, the following condition should be satisfied for the new a /34 value
to have a feasible meaning:
a34 •
t 3,4,5 ~
The above derivations can easily be expanded and are generalized in Theorem
8-3.
current value of the a y measure of performance, such that the ranking between
the two alternatives Ai and Ak will be reversed, Let a'i} denote the modified
a y measure of performance, That is:
a'y = ay - 7 i ,i,k' (AlO)
For an easy demonstration of the main ideas, suppose that we are
interested in reversing the current ranking between alternatives A2 and A 3,
(where P2 ~ P 3 from (8-1» by altering the a34 value (only), Therefore, it is
necessary to determine the threshold value 7 3,4,2 required to get the following
relation (where P /2 and P /3 denote the new final preference values of
alternatives A2 and A3 , respectively):
172 MCDM Methods: A Comparative Study, by E. Triantaphyllou
a14
a14 =
a14 + ~ + *
~4 + ... + am4
a24
~4 =
a14 + a24 + ~4* + + am4
* (A12)
~4
a34 =
a14 + a24 + ~4* + ... + am4
am4 =
a14 + ~ + ~4
* + + am4
Furthermore, the denominator of relations (A12) can be simplified by using
relation (AlO) as follows:
1 - '1: 3,4,2'
(A13)
Therefore, relations (A12) can be expressed as follows:
----,
1 - 't 3,4,2
I am4
am4 = ----=-'"'---
1 - 't 3,4,2
-
(A15)
P3 - Ps
'1: 3•4,5 >
174 MCDM Methods: A Comparative Study, by E. Triantaphyllou
The following condition should also be satisfied for the new d 34 value to have
a feasible meaning:
't" 3,4,5 ~ a 34 ·
The above derivations can easily be expanded and are generalized in Theorem
8-4.
or:
(
a34 )W4 xR (A2)
- < 1.
a 34 - 't" 3,4,2 A3
Furthermore, the following condition should also be satisfied for the new
a 134 value to be feasible:
Chapter 8: A Sensitivity Analysis Approach for MCDM Methods 175
o
o
-
=>
"3,4,2 ~ a34 ,
Now, assume that the measure of performance to alter is again a34 ,
but now we want to reverse the ranking between the two alternatives A3 and
A5 (now: P3 2!: P5 from (8-1», Then, by following an approach similar to the
previous one, one can derive that in order to reverse the current ranking of
the two alternatives A3 and A5 , the threshold value 7345 should satisfy the
following relation:
The MCDM methods presented in the second chapter are among the
most widely used ones. As it will be seen in the following sections, however,
the part of the methods that processes a decision matrix (i.e., step 3 in Section
2.2) may give different answers to the same problem. Because only the
WPM, the AHP, the revised AHP, and the TOPSIS method are applicable
both in single- and multi- dimensional decision making, these are the methods
that are examined in this chapter (i.e., the WSM will not be tested). The
ELECTRE method (as described in Chapter 2) was not tested because it is
already known that this method presents a different ranking philosophy and
it does not assume that a unique ranking always exists in practice. The main
part of this chapter is based on the research results presented in
[Triantaphyllou and Mann, 1989] plus some additional investigations.
Since the truly best alternative is the same regardless of the method
chosen, an estimation of the accuracy of each method is highly desirable. The
most difficult problem that arises here is how one can evaluate a multi-
dimensional decision making method when the true best alternative is not
known. Two evaluative criteria are introduced for the above purpose in the
following section.
Example 9-1: Testing the method used by the Analytic Hierarchy Process by
using the first evaluative criterion
Suppose that the matrix below depicts the actual values, measured in
the same units.
Criteria
C1 · C2 C3
Alts. (8/13 2/13 3/13)
Al 1 9 9
A2 5 2 2
A3 1 5 9
Criterion C1 Criterion C2
1
5/1
1/5
1
1/1
5/1
1 9/2
9/5
2/9 1 2/5
1
1/1 1/5 1 5/9 5/2 1
Criterion C3
1 9/2 9/9
2/9 1 2/9
9/9 9/2 1
180 MCDM Methods: A Comparative Study, by E. Triantaphyllou
Therefore, the mXn (Le., 3x3) decision matrix with the relative
importances of the alternatives in terms of each criterion, along with the
criteria weights, that is used in the final step of the AHP is:
Criteria
C1 Cz C3
Alts. (8/13 2113 3/13)
Applying the last step of the AHP it turns out that alternative A2 is the best
one (A 2• AHP-score = A*AHP-score = 0.48). Obviously, this is in contradiction with
the conclusion derived using the WSM. •
Table 9-1: Contradiction Rate (%) Between the WSM and the AHP.
Table 9-2: Contradiction Rate (%) Between the WSM and the Revised AHP.
Table 9-3: Contradiction Rate (%) Between the WSM and the WPM.
Figure 9-1: Contradiction Rate (%) Between the WSM and the AHP.
Chapter 9: Evaluation of Methods for Processing a Decision Matrix 185
20
lS~
16
-
~
~
14 c:
C
12
ti
:a
10 ~
C
U
Iv,
IIlJJber of C .
ljterJa
Figure 9-2: Contradiction Rate (%) Between the WSM and the RevisedAHP.
Figure 9-3: Contradiction Rate (%) Between the WSM and the WPM .
186 MCDM Methods: A Comparative Study. by E. Triantaphyllou
Example 9-2: Testing the method used by the Analytic Hierarchy Process by
using the second evaluative criterion
Suppose that the following is a matrix of eigenvectors produced by the
AHP process. That is, the matrix contains the relative values of the
importance of the alternatives instead of the actual values. Also, assume that
the criteria have weights WI = 217, W2 = 217, and W3 = 317.
Criteria
CI C2 C3
Alts. (217 217 317)
Criteria
CI Cz C3
Alts. (217 217 317)
Criterion C1 Criterion C2
Criterion C3
1 2/4 2/1
4/2 1 4/1
1/2 1/4 1
Table 9-4: Rate of Change (%) of the Indication of the Optimum Alternative When a
Non-Optimum Alternative is Replaced by a Worse One. The AHP Case.
Table 9-5: Rate of Change (%) of the Indication of the Optimum Alternative When a
Non-Optimum Alternative is Replaced by a Worse One. The Revised AHP
Case.
Cases with
Nwnber of alts. 3 3 3 21 3 3 3 21
Nwnber of criteria 3 5 7 21 3 3 7 21
Method:
WPM 14.7 12.7 12.1 ... 25.7 0 0 0 ... 0
AHP 8.2 10.5 12.3 ... 10.9 0.16 0.26 0.26 0.02
RevisedAHP 7.4 8.2 8.8 ... 3.0 0.50 0.50 0.50 ... 0.60
Chapter 9: Evaluation of Methods for Processing a Decision Matrix 191
o. Z4
~
'-"
0.11 ~
O.lS g
O.ll
~
:a
0.12 f!
0.09
-=o
U
06
Figure 9-4: Rate of Change (%) of the Indication of the Optimum Alternative
When a Non-Optimum Alternative is Replaced by a Worse One.
The AHP Case.
Figure 9-5: Rate of Change (%) of the Indication of the Optimum Alternative
When a Non-Optimum Alternative is Replaced by a Worse One.
The Revised AHP Case.
192 MCDM Methods: A Comparative Study, by E. Triantaphyllou
case), does not depend on the number of criteria as seen in Figure 9-4. This
occurs because, for every criterion, the same constant value (Le., -1) was
used to alter the nonoptimal alternative. It is expected (although not tested)
that if the aij were to be changed by a random number and not a constant
value (Le., -1), the contradiction rate would be dependent on the number of
criteria. Because the constant value (Le., -1) was used the bias, if any, would
be in favor of the original AHP over the revised AHP.
However, the role of the alternatives is a critical one. When m, the
number of alternatives, increases the role of the multipliers is more critical
than the impact of replacing a nonoptimal alternative by a worse one. Since
the number of alternatives overrides the role of the previous replacements the
rates in Figure 9-4 decrease as the number of alternatives increases.
Finally, Figure 9-5, for the revised AHP, illustrates that the rate of
change of the indication of the optimal alternative when a nonoptimal
alternative is replaced by a worse one decreases as the number of alternatives
increases. This occurs because as the number of alternatives increases the
multipliers (lImax{aij' for i = 1,2,3, ... , m}) becomes closer to 9 (Le., the
maximum) and thus the altered alternative still remains nonoptimal. Similarly
with the previous paragraph, as the number of criteria increases there is
greater chance that the multipliers will differ substantially and alternative
changes will have greater impact.
Because in the revised AHP the multipliers (Le., the quantities
(lImax{aij, for i = 1,2,3, ... , m}) are more stable than the corresponding
ones of the original AHP the revised AHP performs better in terms of the
first evaluative criterion. This is not the case when the second evaluative
criterion is used. However, now the difference in the rates does not seem to
be so dramatic.
194 MCDM Methods: A Comparative Study, by E. Triantaphyllou
Table 9-7: Contradiction Rate Between the WSM and the TOPSIS Method.
60
SO
()
g 40
~ cr ~
:30
~ ~
(l ZO
/0 Q".
~
~
'l,.."
"q,
,,"I. ~"i:>
,,~ :..~
,,~ -<;;:-r,,:
,," '<.,.<z--<'
q, ~
~ '" ::<:;:,<z--<' c3-.
~v~
Figure 9-6: Contradiction Rate (%) Between the WSM and the TOPSIS
Method.
Figure 9-7: Rate of Change (%) of the Indication of the Optimum Alternative
When a Non-Optimum Alternative is Replaced by a Worse One.
The TOPSIS Case.
Chapter 9: Evaluation of Methods for Processing a Decision Matrix 197
9.7 CONCLUSIONS
The previous analyses make it clear that none of the MCDM methods
reviewed is perfectly effective in terms of both evaluative criteria. The
results that were obtained by testing the methods using the two evaluative
criteria described in Section 9.2 indicate that the MCDM methods yield
different rates of contradiction.
The findings of the analyses are summarized in Table 9-8. This table
presents the structure of the typical decision-making problem with which the
typical MCDM method is concerned with and it also illustrates a decision
making paradox. In this table, MCDM methods are treated as alternatives
and the two evaluative criteria as the decision criteria of the problem. The
two evaluative criteria were considered in terms of cases with 3 alternatives
and 3 criteria, 3 alternatives and 5 criteria, 3 alternatives and 7 criteria, and
so forth. Thus, 100 (= 10 X 10) subcriteria were generated per criterion.
The numbers WI and w2 = 100 - WI> where WI = 1,2,3, ... , 100%, were
assumed to be the relative weights of the two evaluative criteria. The relative
weights of the first 100 sub criteria were assumed to be equal to WI / 100, and
for the last 100 subcriteria equal to W 2 / 100.
Since it is not clear which method is the best, the problem of selecting
the best decision making method was solved in [Triantaphyllou and Mann,
1989] by using successively the WSM, the AHP, and the revised AHP
method. The occurrence of zeros in Table 9-6 made the applicability of the
WPM infeasible (because of the divisions by zeros). Thus, the WPM was not
considered for solving the previous decision problem. Figure 9-8 presents the
results for different values for the weights WI and w2 • Since in Table 9-6 all
values are presented in the same unit (recall that these numbers represent
percentages), the WSM is the method that would yield the most reliable
answer to the problem of choosing the most effective decision making
method.
The first column in the chart of Figure 9-8 represents the evaluation
derived by using the WSM. That chart suggests that the revised AHP appears
to be the best method over a wide range of different combinations of weights
WI and w2 • Only for WI less than 10 per cent the revised AHP is not the best
method. For WI less than 10 and above 3 per cent the AHP appears to be
superior than the others, while for WI less than 3 per cent the WPM becomes
the best method.
From Figure 9-8 one can generally conclude that as WI increases, then
WPM is replaced by the AHP and later by the revised AHP as the best
method. This probably happens because both methods behave in the same
way for problems with 100 criteria and 3 alternatives (as is the case in Table
9-6).
198 MCDM Methods: A Comparative Study, by E. Triantaphyllou
100
80
60
KEY:
40
20
o
WSM AHP Revised
AHP
However, both methods recommend the WPM to be the best and not
themselves! They also recommend the revised AHP to be the second best
approach, while the original AHP is the best one only for a very small portion
of combinations of weights WI and W 2 .
The previous study, with the findings summarized in Figure 9-8,
demonstrates that it is impossible to determine precisely the best MCDM
method, for to do so one needs to use the best MCDM method! This
problem of finding the best MCDM method always reaches a
decision-making paradox which makes any attempt in solving this problem
to be of limited success. However, the results in this chapter recommend that
for most of the cases of different weights of the two evaluative criteria the
revised AHP appears to be the best decision making method of the four
examined in terms of these two evaluative criteria, while the original AHP,
appears to be the most inaccurate one when the previous two evaluative
criteria are considered.
Chapter 10
some entities (Le., decision criteria or alternatives) to have the same rank
while in reality they are distinct. However, the empirical analyses of this
chapter demonstrate that it is possible for alternatives which in reality are less
important than others to appear to be more important after the AHP or the
revised AHP are used. Furthermore, this problem of rank reversals is more
severe in the present investigation. Also, the present findings reveal that the
rate of the rank reversals increases dramatically as the number of
alternatives increases.
This chapter uses extensively the concepts of the RCP (Real and
Continuous Pairwise) and CDP (Closest Discrete Pairwise) matrices (which
are defined and analyzed in Section 3.3). The problem which is investigated
in this chapter is best described by the numerical example presented in the
next section. In that numerical example it is assumed that the actual relative
values in a decision making problem, which involves three alternatives and
four criteria, are known. The AHP and the revised AHP are applied to these
actual data and the relative weight and ranking of each alternative is derived
according to each one of these two MCDM methods.
Next, we will use the concepts of the RCP and CDP matrices to
derive the discrete data required in the test problems. These discrete data are
the data which a decision maker could have if one were able to make the
closest approximations to the real (and thus unknown) pairwise comparisons.
Again, this is a very favorable assumption since there is no certainty that in
reality one can achieve these closest approximations for all the pairwise
comparisons. The following numerical example reveals that it is possible for
the AHP and the revised AHP method to fail to preserve the ranking of the
alternatives when the continuity assumption is utilized and even if the decision
maker is as accurate as possible.
Criteria
Ct C2 C3 C4
Alts. (0.1325 0.0890 0.5251 0.2533)
That is, the actual relative weights of the four criteria are: (0.1325
0.0890 0.5251 0.2533). The relative importances of the three alternatives
in terms of the first criterion are: (0.5008 0.3785 0.1207), and so on.
The above values can also be viewed as the principal right
eigenvectors of the following 1 + 4 = 5 perfectly consistent Rep matrices.
For the case of the four criteria we have:
For instance, in the previous matrix the (1, 2) entry is equal to 1.4890
because the corresponding pairwise comparison is: 0.1325 / 0.0890 (=
1.4890). Similar interpretations hold for the remaining entries.
For the three alternatives we have in terms of each one of the four
decision criteria:
1 1.3230 4.1502
0.7559 1 3.1371
0.2410 0.3188 1
1 2.2552 1.5682
0.4440 1 0.6963
0.6377 1.4362 1
1 0.1917 0.1803
5.2156 1 0.9403
5.5468 1.0635 1
204 MCDM Methods: A Comparative Study, by E. Triantaphyllou
1 0.5561 0.5167
1.7982 1 0.9291
1.9355 1.0763 1
By applying the AHP and the revised AHP (see Sections 2.2.3 and 2.2.4,
respectively, for the appropriate formulas) it can be shown that the following
results are obtained:
In reality, however, the decision maker does not know the previous
real values. Instead, one has to use matrices with entries from the set: {9, 8,
... , 2, 1, 1/2, ... , 118, 1/9} (that is, to use the scale presented in Table 3-1
in Section 3.2.1). The COP matrices which best approximate the previous
five RCP matrices are shown next.
For the case of the four decision criteria the corresponding COP
matrix and its corresponding relative weights (approximated eigenvector) is:
RCPo -+ CDPo =
For instance, the (1, 2) entry in the previous matrix is equal to 1.5000
because the corresponding entry in the RCP matrix is 1.4890 and the value
1.5000 is the closest value from the permissible values (when the scale
depicted in Table 3-1 in Section 3.2.1 is used). Similar interpretations hold
for the remaining entries.
Similarly with above, for the three alternatives in terms of each one
of the four decision criteria the COP matrices and their corresponding relative
weights are as follows:
Chapter 10: A Computational Evaluation of the Original and Revised AHP 205
RCP l -;0.
CDP l 0.7500 1 3.0000 -;0. 0.3750
That is, the data which are assumed to be available in reality to the
decision maker for the present problem are:
Criteria
C1 C2 C3 C4
Alts. (0.1304 0.0870 0.5217 0.2609)
Similarly, as in the first part of this example, the AHP and the revised AHP
yield the following results:
206 MCDM Methods: A Comparative Study, by E. Triantaphyllou
These findings clearly contradict the results found in the first part
of this analysis. In this numerical example the weights of the two alternatives
A z and A3 are almost identical when the AHP is applied (namely, 0.3982 and
0.3945, respectively) on the actual data. However, after the original Saaty
scale is used, and the assumption that the decision maker is as accurate as
possible is made, then the AHP yields the weights 0.3875 and 0.4074 for
alternatives A2 andA 3 , respectively. In other words, the AHP causes the two
alternatives to look very distinct, while in reality they are almost identical.
If for illustrative purposes one considers that the alternatives represent
projects to be funded, then the AHP would suggest that project A3 to be
funded 5.14% (i.e., (0.4074 - 0.3875) x 100 I 0.3875) more than project
A 2 , while in reality project A3 should be funded by almost 1.0% less than
project A z. That is, after the AHP or the revised AHP are applied, then it is
likely that alternatives which in reality are almost identical, to appear to be
rather distinct. The reverse situation is also possible to occur.
In the previous numerical example both the AHP and the revised AHP
yield the same alternative ranking. However, in general this is not always the
case. It is possible for the AHP to yield the correct alternative ranking and
the revised AHP an incorrect one, or vice-versa. In order to study the
performance of the two methods in depth, a simulation approach was
followed. Random problems with different number of decision criteria and
alternatives were generated and then treated as in the previous extensive
numerical example.
For each such random problem actual relative weights were assumed
as in the previous numerical example. However, because the Saaty matrices
use values from the set: {9, 8, ... , 2, 1, 112, ... , 2/8, 2/9} only the random
problems which are associated with Rep matrices with entries within the
continuous interval [9, 119] were considered. Each test problem was treated
as in the previous example. If the AHP or the revised AHP yielded a ranking
different than the one derived when actual data were used, then the case was
recorded as a failure. Tables 10-1 and 10-2 present the results of this
Chapter 10: A Computational Evaluation of the Original and Revised AHP 207
Table 10-1: The Failure Rates are Based on 1,000 Randomly Generated Problems.
The AHP Case.
Nwnber of Alternatives
Nwnber of
Criteria 3 5 7 9 11
Nwnber of Alternatives
Nwnber of
Criteria 13 15 17 19 21
Nwnber of Alternatives
Nwnber of
Criteria 3 5 7 9 11
Nwnber of Alternatives
Nwnber of
Criteria 13 15 17 19 21
Number of Alternatives
100 .....-' 21
90 ·~·=0·~ ·~·~·i·~··~· ~·~ ·~·~·~ +.:.;.~.~.~ .~.f·~·~;'·~·~·~·~·f ·~·~·~·~·~·~·f·~·=:·= ·~·~ .:. . 7
20
10
o
:3 6 9 12 15 18 21
Number of Decision Criteria
Figure 10-1,' The Failure Rates are Based on 1,000 Randomly Generated Problems.
The AHP Case.
Chapter 10: A Computational Evaluation of the Original and Revised AHP 211
Number of Alternatives
..-.
~
.,'"
60 ···················_······1······· ··················+········· · :·· =·· J·· =~··=···=···t··=···=···= ·):·· =~:::·· .:;;;. .-
- .+ . .
5
"0
a:: 50 ~. ::: ·~ ~ · =··. =r . ·~·· . · · . · · . . ·. ·I· .. ·..··· .. · ..···..·......·+··. ·. ··· . · . ··· . ·..··· . t·········· .. ········· ..·· .. ..
~
.2
'0
40 .............................~ .. , ..........-....- ........+. . . . . . . . . . . . . . . ~. . . . . . . . . . . . . .····t···. ·······..·············.. ·~·; ..·........·
:
.....
30 ......... ...................:.............................. ~ .... ·· .. ··· .. ··· .... ·..·.. +·························.. ··t...... ·..··· .... ·......······t···· .. ··· ..·.. ·· .... ~ 3
20 ~.. ........................ ,..... ·· ....·· .... ·.. t············ .. ·.. ··· .. ··· .. ··;· ..· ......................................................
10 ..............................:.. ... ---.....................; ...................... ,...·,··.;"···"···"··"··,····,····"·7"···'······.. ···.. ··.. ·······t .. ·,·· ..·
a L-~~~~~ __~-L__~~-J____~__~~__~-L~~~-J
:> 6 9 12 15 18 21
Figure 10-2: The Failure Rates are Based on 1,000 Randomly Generated Problems .
The Revised AHP Case.
212 MCDM Methods: A Comparative Study, by E. Triantaphyllou
10.4 CONCLUSIONS
the AHP and now it is also called the ideal mode AHP.
The issue of ranking reversals continues to be a hot and controversial
one. This subject has also been studied by Dyer and Ravinder [1983] and
Dyer and Wendell [1985]. Saaty in [1983] and [1987] responded to these
criticisms with some axioms and guidelines on how close a near copy can be
to an original alternative without causing a rank reversal. He asserted that the
decision maker has to eliminate alternatives from consideration that score
within 10% of another alternative. This recommendation was later sharply
criticized by Dyer [1990a].
Dyer in [1990a; and 1990b] claimed that the original AHP may result
in arbitrary rankings, even when no identical alternatives are considered. He
also asserted that in some cases the AHP may yield results which may be
highly correlated with the actual performances of a consistent decision maker,
and even provide an accurate ranking of alternatives in other special cases.
However, the same author claimed that, in general, this cannot be guaranteed
a priori.
Dyer attributed this problem to the principle of hierarchic
composition, which is one of the main assumptions in the AHP. This
principle assumes that the weights of the criteria do not depend on the
alternatives under consideration. As an alternative approach, Dyer suggested
that the AHP be combined with utility theory. That paper by Dyer stimulated
some critical responses by Saaty [1990] and Harker and Vargas [1990].
As Winkler in [1990] remarked in a related technical note, multi-
attribute utility theory (MAUT) is not a panacea. He cautions the reader to
be careful in accepting alternative theories. Clearly, these problems are still
controversial in decision theory. Some additional discussions beyond the ones
given in Chapters 9 and 10, on these truly important issues are due to Bell
and Farquhar [1986], Hogarth and Reder [1986], Weber and Camerer [1987],
Munier [1988], Fishburn and LaValle [1989], and Sarin [1989]. In contrast
with all previous analyses, the new ranking abnormalities described in this
chapter do not involve the introduction of any new alternatives.
The new type of ranking abnormalities are based on decomposing an
MCDM problem into a set of smaller problems. Each such problem is
defined on a pair of the alternatives. It is important to state here that no
change of the numerical data is assumed (except any normalizations which do
not alter the relative preferences). Then, the rankings derived from the
smaller problems are compared with the rankings derived when all the
alternatives are considered simultaneously. The above are best described in
terms of the following illustrative examples.
Chapter 11: New Ranking Abnormalities with Some MCDM Methods 215
Criteria
CI Cz Cj
Alts. (2/7 2/7 3/7) Combined
Priorities:
Al 9/19 2/12 2/7 0.305
Az 5/19 1112 4/7 0.344
Aj 5/19 9/12 117 0.351
Note that the above data are assumed to have been derived from
perfectly consistent judgment matrices with pairwise comparisons. This was
assumed in order to block out any effects due to inconsistent pairwise
comparisons among the alternatives in terms of each one of the decision
criteria. When the original AHP is applied on the above decision matrix, then
it can be easily seen that the three alternatives have the priority values shown
under the "Combined Priorities" vector. For instance, the combined priority
for alternative Al is 0.305, because: (9/19) X (2/7) + (2/12) X (2/7) +
(2/7)x(37) = 0.305. Therefore, in the maximization case, the above
combined priorities indicate that the three alternatives are ranked as follows:
Aj > A z > Al (where the symbol "> " as before stands for "is better than ").
In the following paragraphs the three alternatives are compared in a
pairwise fashion. That is, the original problem is broken into three smaller
problems, each one having two alternatives. The first such problem (which
considers the pair of alternatives A2 and A 3) is described by the following
decision matrix:
216 MCDM Methods: A Comparative Study, by E. Triantaphyllou
Criteria
CI C2 C3
Alts. (2/7 2/7 3/7) Combined
Priorities: Derived Ranking:
5/10 1110 4/5 0.5143 A2 > AJ
5/10 9/10 1/5 0.4857
The columns in the above matrix have been normalized again, in order to be
consistent with the basic requirement of the original AHP (which requires the
vectors to add up to one).
Similarly, the second such problem (which considers the pair of
alternatives A] and A 2) is described by the following decision matrix:
Criteria
CI C2 C3
Alts. (2/7 2/7 3/7) Combined
Priorities: Derived Ranking:
9/14 2/3 2/6 0.5170 A] > A2
5/14 1/3 4/6 0.4830
Finally, the third and last problem (which considers the pair of alternatives
A] and A J) is described by the following decision matrix:
Criteria
CI C2 C3
Alts. (2/7 2/7 3/7) Combined
Priorities: Derived Ranking:
9/14 2111 2/3 0.5213 A] > A3
5/14 9/11 1/3 0.4787
The previous results, when are taken together, indicate that the
ranking of the three alternatives must be: Al > A2 > A 3. That is, this
ranking is different of the one derived when all the alternatives were
considered simultaneously (namely: A3 > A2 > AI). In matter of fact, the
second ranking is the reverse of the first one! Therefore, the questions which
naturally are raised at this point are: "Which one is the best ranking?" and
"What are the correct combined priorities of the three alternatives?" Before
we proceed to answering these questions, we present some further
elaborations on this illustrative example. Suppose that the combined priorities
of the three alternatives are as usually denoted as PI' P2 , and P3 ,
respectively. In the light of this notation, the previous three smaller problems
imply that:
Chapter 11: New Ranking Abnormalities with Some MCDM Methods 217
Alts. Al Az A3 Combined
Priorities
The above combined priorities were derived from the previous pairwise
comparisons by applying the Saaty approximation of estimating the right
principal eigenvector of the previous reciprocal matrix. This approximation
is done by multiplying the elements in each row and then by taking the n-th
root (i.e., the third root in this example) [Saaty, 1980] (or Section 4.2). The
values obtained that way are then normalized and the "Combined Priorities"
vector is obtained.
It can be noticed now that in this matrix the alternatives are compared
in terms of all the three criteria combined. This is a novel step since
traditionally in the AHP context alternatives are compared in terms of a single
decision criterion at a time. At this point it becomes apparent that the ranking
of the three alternatives implied by the last combined priorities is: Al > A2
> A j • Obviously, this ranking is identical to the ranking obtained from the
combination of the solutions of the earlier three sub-problems. Again, this
ranking is different (actually reverse) than the ranking obtained when all the
alternatives are considered simultaneously . •
Similarly with the previous example, the case of the ideal mode AHP
is best described in terms of an example. Suppose that the following is the
decision matrix of a simple problem with three criteria and the three
alternatives AI' A 2, and A):
Criteria
CI C2 C)
Alts. (4122 9/22 9/22) Combined
Priorities:
Al 9/9 5/8 2/8 0.5398
A2 119 8/8 5/8 0.6850
A) 8/9 2/8 8/8 0.6730
As earlier, the above data are also assumed to have been derived from
perfectly consistent judgment matrices with pairwise comparisons. When the
ideal mode AHP is applied on the above decision matrix, then it can be easily
seen that the three alternatives have the priority values shown under the
"Combined Priorities" vector. Therefore, in the maximization case, the above
combined priorities indicate that the three alternatives are ranked as follows:
A2 > A3 > AI'
Similarly with the discussion in Example 11-1, next the three
alternatives are compared in a pairwise fashion. That is, the original problem
is broken into three smaller problems, each one having two alternatives. The
first such problem (which considers the pair of alternatives A2 and A) is
described by the following decision matrix:
Criteria
CI C2 C)
Alts. (4122 9/22 9122) Combined
Priorities: Derived Ranking:
118 8/8 5/8 0.687500 AJ > A2
8/8 2/8 8/8 0.693182
Now the columns in the above matrix have been normalized in the way
prescribed by the ideal mode AHP (that is, the elements in each vector are
divided by the largest value in that vector).
Similarly, the second such problem (which considers the pair of the
alternatives Al and A2) is described by the following decision matrix:
Chapter 11: New Ranking Abnormalities with Some MCDM Methods 219
Criteria
CI C2 C3
Alts. (4/22 9/22 9/22) Combined
Priorities: Derived Ranking:
9/9 5/8 2/5 0.601136 A2 > Al
119 8/8 5/5 0.838384
Finally, the third and last problem (which considers the pair of
alternatives Al and A3) is described by the following decision matrix:
Criteria
CI C2 C3
Alts. (4122 9/22 9122) Combined
Priorities: Derived Ranking:
9/9 5/5 2/8 0.693181 A3 > Al
8/9 2/5 8/8 0.734343
The previous results, when are taken together, indicate that the ranking of the
three alternatives must be: A3 > A2 > AI' Similarly with the first example,
this ranking is different of the one derived when all the alternatives were
considered simultaneously (namely: A2 > A3 > AI)'
As with the previous illustrative example, it can be observed that the
following relations must be true:
PiP3
0.687500/0.693182 = 0.991803,
~
P/P2
0.601136/0.838384 = 0.717018,
~
and ~ 0.693182/0.734343 = 0.943948.
P/P3
These pairwise comparisons lead to the introduction of the following pairwise
comparison matrix and then to the attached combined priorities:
Alts. Al A2 AJ Combined
Priorities
a time, then the derived ranking is different. Namely, this ranking is: A3 >
A2 > AI. This ranking is (by definition) identical to the ranking derived from
solving the three smaller sub-problems and then combining the partial
solutions. That is, in this case the ranking derived from the combined
pairwise matrix is different than the ranking derived when the three smaller
problems are considered. Clearly, this was not the case in Example 11-1.
Besides the previous categories of ranking abnormalities, a different type of
ranking abnormalities may occur and it is discussed in the next section. •
The previous discussions clearly raise the issue that when the original
or ideal mode AHP is used, then sometimes it may not be very obvious what
is the correct ranking. Should one accept the ranking derived when all the
alternatives and decision criteria are used? Or should one accept the ranking
derived from combining the solutions of the smaller problems which consider
two alternatives (and all the criteria together)? It may also be the case when
the matrix with the combined pairwise comparisons is considered, then one
may derive a third ranking which is different than the previous two rankings.
The above scenarios are possible when the original or ideal mode AHP is
used.
One may argue here that when the decision maker encounters a
problematic situation as in the previous scenarios, then he/she must be more
careful in accepting the final results. A re-evaluation of the characteristics of
the problem (especially if new decision criteria can be introduced) may result
in more robust data which may lead to less ambiguous conclusions.
However, as the two numerical examples discussed in the previous sections
demonstrate, it is possible to reach inconclusive results even when all the
pairwise comparisons (i.e., the ones which consider pairs of alternatives in
terms of a single criterion) are perfectly consistent. Recall that this was
actually the case with the data in the decision matrices in these examples.
Therefore, the first question to be answered is which is the right ranking of
the alternatives.
Deciding on which one is the right ranking can be an open-ended
question. One may argue that the correct ranking is the one derived when all
the alternatives and decision criteria are considered simultaneously. However,
one may also claim that the solutions of the smaller problems are more
reliable because these involve simpler (only two alternatives are considered
Chapter 11: New Ranking Abnormalities with Some MCDM Methods 221
Criteria
C1 C2 C3 C4
Alts. (0.27 0.41 0.05 0.27)
Suppose that one uses the ideal mode AHP. When alternatives A I and
A z are compared as in the previous two examples, then it is derived that: Az
> AI. Similarly, when alternatives AI and AJ are compared, then it is derived
that: AI > AJ. That is, these relations suggest that the ranking of the three
alternatives must be as follows: A z > Al > A3. However, when the two
alternatives A3 and Az are compared, then the derived ranking is: A3 > A z.
That is, a logical contradiction (Le., a case of logical inconsistency) is
reached. A similar phenomenon can be observed with test problems which
use the original AHP method. •
Criteria
CI Cz C3 C4
Alts. (0.43 0.12 0.06 0.39)
Criteria
CI Cz C3 C4
Alts. (0.43 0.12 0.06 0.39)
Al 1.00 1.00 0 0
Az 0 0 1.00 0.22
A3 0.50 0.44 0.81 1.00
Chapter 11: New Ranking Abnormalities with Some MCDM Methods 223
show similar contradiction rates for the two versions of the AHP.
On the other hand, Figures 11-3 and 11-4 depict contradictions in the
ranking of any alternative. Now the number of alternatives plays a decisive
role, while the number of decision criteria is not as important. Moreover, the
contradiction rates are significantly more dramatic. For instance, for
problems with five alternatives, the contradiction rates are almost 50%. As
before, there is no much difference between the results obtained when the
original or the ideal mode AHP was used. As it was expected, the
contradiction rates in Figures 11-3 and 114 are much higher than those in
Figures 11-1 and 11-2. This was expected because the cases of contradiction
in the first two figures are naturally included in the results of Figures 11-3
and 11-4.
Figures 11-5 and 11-6 present the contradiction rates when the
alternatives are compared two at a time and cases of logical inconsistencies
were found (as in Example 11-3). The roles of the number of alternatives and
number of decision criteria are similar as before. However, now the ideal
mode AHP performs significantly worse than the original AHP.
In all these results problems with less alternatives yielded fewer
ranking contradictions than problems with more alternatives. This was
expected because the number of alternative pairs to be considered for a given
case is directly related to the number of alternatives in the problem. In matter
of fact, if the number of alternatives is equal to m, then the number of pairs
to be considered is equal to m(m-l)/2. Thus, the chances of finding a logical
inconsistency increase accordingly. As it can be seen from the illustrative
examples, the number of criteria did not playa prime role. This is also
evident in the computational results by the almost horizontal curves.
Chapter 11: New Ranking Abnormalities with Some MCDM Methods 225
100 ;
90 ---. ~- !i .
_______ .1 __ • ______L _ _ _ _ _ • !o------.~. .---.-.----.;.. . -.--
! ;
~! .:=t=+=+
.-...
~ 70
~
0::
0 60
c::: !! 1 ~
~
0
:;J 50
.~
'0
g 40
c:::
~+
0
u 30
20
10
. : '{..Probler6s w ti h 31 Alternot~es
0
4 6 8 10 12 14 16 18 20
Number of Criterio
Figure 11-1: Contradiction Rates on the Indication of the Best Alternative When
Alternatives are Considered Together and in Pairs. The Original AHP
Case.
100
I
~++=r++~=:::=:::::::::r
90
80
; I i i I f I
.-... 70
~
4>
'0 60
0::
c:::
0 50
~
'0
~ 40
C
0
u 30
i i i i I Prob l~ms with : 21 Alter?otives i
~~~;~r4i.:t3~~,~~::,:i~
20
10
' . ' i ' Problems with ! 3A1etre ~ otives
0
4 6 8 10 12 14 16 18 20
Number of Criteria
Figure 11-2: Contradiction Rates on the Indication of the Best Alternative When
Alternatives are Considered Together and in Pairs. The Ideal Mode
(Revised) AHP Case.
226 MCDM Methods: A Comparative Study. by E. Triantaphyllou
tEL=3 ~F~
80
~
~
70
'"
'0
a:::
60
! ; . , I /Prpblems 1ith 5 Alt1rnotives !
c:
50
·r~tt31~
:e'6
0
40
~
c:
0
! ! ! ! !
0
4 6 8 10 12 14 16 18 20
Number of Criteria
Figure 11-3: Contradiction Rates on the Indication of Any Alternative When Alternatives
are Considered Together and in Pairs. The Original AHP Case.
100
o
I I i i
4 6 8 10 12 14 16 18 20
Number of Criteria
Figure 11-4: Contradiction Rates on the Indication of Any Alternative When Alternatives
are Considered Together and in Pairs. The Ideal Mode (Revised) AHP Case.
Chapter 11: New Ranking Abnormalities with Some MCDM Methods 227
100
90
80
g 70
"
OJ
60
cr
c:
.2 50
-0
v 40
~
C
a 30
U
20
10
0
4 6 8 10 12 14 16 18 20
Problems with 3 A tlernatives
Number of Criteria
Figure 11-5: Contradiction Rates on the Indication of Any Alternative When Alternatives
are Considered in Pairs. The Original AHP Case.
i _ .-+--* .- - ; - - - - ! - -*·-··-i-
100
90
1"" ~/ ! i i i ' Problems with 15i Alternatives
80
g 70
OJ
"
cr
c:
.2
60
50
-0
'<;
e 40
ca ./1 .... i : i j l.Problems with 9: Alternatiites
u 30
20
10
0
4 6 8 10 12 14 16 18 20
Numbe r of Criteria
Figure 11-6: Contradiction Rates on the Indication of Any Alternative When Alternatives
are Considered in Pairs. The Ideal Mode (Revised) AHP Case.
228 MCDM Methods: A Comparative Study. by E. Triantaphyllou
iI (ali)w/
au1=1
> 1, or:
(11-1)
II (alit' II (a r'·
/I /I
> 2I
1= 1 i = 1
It can be noticed here that the last two products express the priority
values of the corresponding two alternatives under the WPM method. These
priority values' can be normalized and/or ranked and thus derive the ranking
of a set of alternatives. In [Triantaphyllou, 2000] it is proved that this
multiplicative variant of the AHP possesses certain interesting properties. In
particular, it is proven theoretically that the ranking abnormalities studied in
Chapter 11: New Ranking Abnormalities with Some MCDM Methods 229
Criteria
CI C2 C3 C4
Alts. (0.27 0.41 0.05 0.27)
When the ideal mode AHP was applied, the ranking derived when two
alternatives were considered at a time, had some internal contradictions.
Next, the multiplicative formula (11-1) is applied on these data. The first pair
to consider is alternatives Al and A2. The corresponding calculations are:
J«NAJ -_ (3.12)0.27
- (4.31)0.41 x (8.57)0.05
x - -
(7.11)0.27
x - = 0.932 < 1.
7.70 4.TI 7.45 3.29
Similarly with above, these alternatives are ranked as follows: A3 >
A 2 • From the last two rankings the following ranking is derived for all three
alternatives: A3 > A2 > AI' Observe, that when the last pair Al and A3 is
considered, then the derived ranking is always in perfect agreement with the
previous global ranking of the three alternatives. The corresponding
calculations for the pair Al and A3 are presented next:
The above result could also had been obtained by observing that:
When formula (11-1) is applied on the previous decision matrix, the derived
priority values (before normalization) are: PI = 0.735, Pz = 0.742, and P3
= 0.754. Therefore, the ranking of the three alternatives is: A3 > Az > AI'
which is identical (as it should be) with the one found earlier. •
11.6 COMPARATlVERESULTSFROMTWOREALLlFE
CASE STUDIES
Criteria
C1 C2 CJ C4 Cs C6 C7
Alts. (0.1760 0.2360 0.0700 0.0790 0.2190 0.0300 0.1900)
When the original AHP, ideal mode AHP, and the multiplicative AHP are
applied on the previous data, the priorities and rankings depicted in Table Il-
l are derived.
It can be observed from this table that the original and ideal mode
AHP derived identical rankings, while the ranking derived by the
multiplicative AHP differs on the third and second position (marked with "*"
on this table).
Table 11-1: Priorities and Rankings of the Alternatives in the "Bridge Evaluation" Case
Study [Saaty, 1994].
that under the AHP and the ideal mode AHP the derived ranking is: A2 < A 3 .
That is, a case of ranking reversal occurs. All the other pairs yield rankings
consistent with the original ranking. This is true when the original or the
ideal mode AHP method is used. It should also be stated here that no logical
inconsistencies as the ones found in Example 11-3 were detected with these
data.
Criteria
CI Cz C3 C4 Cs C6 C7
Alts. (0.1760 0.2360 0.0700 0.0790 0.2190 0.0300 0.1900)
It can be easily shown that when all the alternatives are considered
simultaneously and the AHP is applied, then the derived ranking is: A1 > A2
> A 3 > A4 > A5 > A6 > A 7• However, when the alternatives A 3 and A2 are
compared as a pair and the AHP method is used, then their ranking becomes:
A3 > A2 . All the other pairs of alternatives yield rankings consistent with the
original one. The same phenomenon also occurs when the ideal mode AHP
is applied on these data. Finally, it should be stated here that many more
such case studies examinations can be found in [Triantaphyllou, 2000].
Chapter 11: New Ranking Abnormalities with Some MCDM Methods 233
11. 7 CONCLUSIONS
The analyses on ranking abnormalities presented in this chapter, along
with other studies on ranking abnormalities, re-enforce a growing belief
among many decision analysts that ranking abnormalities are unavoidable
when the AHP or some of its additive variants are used. This phenomenon
seems to be an inherited difficulty when one deals with criteria which are
defined on different units. This chapter has demonstrated, via some
numerical examples, that ranking abnormalities are possible even when the
data are perfectly known (and thus perfectly consistent). Therefore, it can
very well be the case that the decision maker may never know the exact
ranking of the alternatives in a given MCDM problem when the original
AHP, or its current additive variants, are used. The computational study,
along with some additional analyses reported in [Triantaphyllou, 2000]
indicate that such ranking abnormalities may occur dramatically too often.
A multiplicative variant of the AHP, as described in [Barzilai and
Lootsma, 1994] and in [Lootsma, 1999] is also advocated in this chapter.
This multiplicative variant of the AHP is a direct adaptation of the well
known weighted product model (WPM) [Miller and Starr, 1969] and [Chen
and Hwang, 1992] in the last step of the AHP, in which the decision matrix
is processed and the final ranking of the alternatives is determined (see also
Section 2.2.2). In this way, no more ranking abnormalities of the type
described in this chapter are possible.
It should be emphasized at this point that the fact that the proposed
multiplicative AHP does not possess any of the ranking abnormalities studied
in this chapter does not necessarily guarantee that this method is perfect (i.e.,
the rankings derived are always the true ones). However, it is obvious that
the reverse argument should always be true. That is, a perfect MCDM
method should never possess any of the ranking abnormalities studied in this
chapter.
It is also important to state here that a fuzzy version of the WPM
performed poorly in the empirical study described in Chapter 13 (see
especially Section 13.7.1). Please also recall the contradiction rates between
the WSM and the WPM methods (in crisp environments) as studied in
Chapter 9. Ranking abnormalities in MCDM methods is a very important,
fascinating, and also very controversial issue in decision analysis, and more
research is required.
Chapter 12
12.1 BACKGROUND
- "
INFORMATION
For a long time it has been recognized that an exact description of
many real life physical situations may be virtually impossible. This is due to
the high degree of imprecision involved in real world situations. Zadeh, in
his seminal papers [Zadeh, 1965; and 1968], proposed fuzzy set theory as the
means for quantifying the inherent fuzziness that is present in ill-posed
problems (which by many accounts are the majority of the real life problems
in decision making). Fuzziness is a type of imprecision which may be
associated with sets in which there is no sharp transition from membership to
nonmembership [Bellman and Zadeh, 1970]. Examples of fuzzy sets are
classes of objects (entities) characterized by such adjectives as large, small,
serious, simple, approximate, etc. [Bellman and Zadeh, 1970].
A comprehensive description of the importance of fuzzy set theory in
engineering and scientific problems is best illustrated in the more than 1,800
references given in [Chang, 1971], [Dubois and Prade, 1980], [Gupta, Ragade
and Yager, 1979], [Xie and Berdosian, 1983], [Zadeh, Fu, Tanaka, and
Shimura, 1975], and [Zadeh, 1976; 1978; and 1979].
Currently, an increasingly large number of researchers has been faced
with the problem that either their data or their background knowledge is
fuzzy. This is particularly critical to people which build expert systems and
decision support systems, for the knowledge they are dealing with is almost
always riddled with vague concepts and judgmental rules (e.g., [Lee, 1971],
[Lee, Grize, and Dehnad, 1987], [Prade and Negoita, 1986], [Ramsay, 1988],
and [Zadeh, 1983]. The most critical step in any application of fuzzy set
theory is to effectively estimate the pertinent data (i.e., the membership
values). Although this is a fundamental problem, there is not a unique way
of determining membership values in a fuzzy set. This is mainly due to the
way different researchers perceive this problem.
The above problems and challenges are particularly pervasive in many
MCDM problems. Very often in MCDM problems data are imprecise and
fuzzy. For instance, what is the value of the j-th alternative in terms of an
environmental impact criterion? A decision maker may encounter difficulty
in quantifying such linguistic statements so they can be used in deterministic
decision making. Related are the developments presented in Chapter 3.
Therefore, it is also desirable to develop and compare decision making
methods which use fuzzy data in making a decision. It is also crucial to
E. Triantaphyllou, Multi-criteria Decision Making Methods: A Comparative Study
© Springer Science+Business Media Dordrecht 2000
236 MCDM Methods: A Comparative Study. by E. Triantaphyllou
1 I
--x - --, if x E [I, m]
m - I m - I
Ilm(X) = 1 I (12-1)
x - , if x E [m, u]
m-u m-u
0, otherwise.
In the above relation (12-1) I :::;; m :::;; u, and I and u stand for the
lower and upper value of the support of fuzzy number M, respectively, and
m for the modal value. A fuzzy triangular number, as expressed by equation
(12-1), will be denoted as (I, m, u).
Laarhoven and Pedrycz [1983], Buckley [1985] and Boender, et al.,
[1989] introduced fuzzy number operations in Saaty's AHP method by
substituting crisp numbers with triangular fuzzy numbers. The distinction of
Buckley's method from Boender's is that the fuzzy solution of a decision
making problem does not need to be approximated by fuzzy triangular
numbers. However, the triangular approximation of fuzzy operations is
plausible to fuzzify the implicit solution of a decision making problem and
provides fuzzy solutions with much smaller spread than Buckley's method
[Boender, et al., 1989]. Also, Boender, et al., proposed the use of a
geometric ratio scale as opposed to the original Saaty equi-distant scale in
quantifying the gradations of a human's comparative judgements. The basic
operations of fuzzy triangular numbers which were developed and used in
[Laarhoven and Pedrycz, 1983] are defined as follows:
where "=: " denotes approximation, and fl.] = (nu, n]m> n]u) and fl.2 = (n2f ,
n2m> n2u) represent two fuzzy triangular numbers with lower, modal, and
upper values. For the special case of raising a fuzzy triangular number to the
power of another fuzzy triangular number, the following approximation was
used:
- i!2 _ I n2f n2m n2u )
n] = \nlf ' n]m ,n]u
Please note that this formula was used only in the development of the fuzzy
WPM (as explained in the next chapter).
Then, it; dominates (or outranks) itj , written as it; > itj , if and only if eij = 1
and ej ; < Q, where Q is some fixed positive fraction less than 1. Values such
as 0.70, 0.80, or 0.90 might be appropriate for Q and the value of Q should
be set by the analyst and possibly be varied for a sensitivity analysis. In the
computational experiments reported in the next chapter the value of Q was set
equal to 0.90. The above concepts are best explained in the following
illustrative example.
Example 12-1:
Suppose that the importances of two fuzzy alternatives A] and A2 are
represented by the two fuzzy triangular numbers it] = (0.20, 0.40, 0.60) and
it2 = (0.40,0.70,0.90), respectively. Next, it can be observed from Figure
12-1 that e2] = 1 and el2 = 0.40 < Q = 0.90. Therefore, A] > A2 and thus
the best (fuzzy) alternative is A]. •
Figure 12-1: Membership Functions for the Two Fuzzy Alternatives Al and A2•
Chapter 13
using the Saaty scale, and the Lootsma scale in order to study the impact of
different scales in the final results. Determining the best fuzzy MCDM
method is the second major objective of this chapter.
In the context of this chapter the values which the decision maker
assigns to the alternatives in terms of the decision criteria in a fuzzy MCDM
problem are assumed to be fuzzy triangular numbers instead of crisp
numbers. ThuS,-'these fuzzy triangular-numbers will be denoted as fl = (I, m,
u), where I !5; m !5; u, and I, m, U are again the lower, modal and upper
value of fl, respectively. The parameter m (modal value) should not be
confused here with the number of alternatives which was also denoted as m
earlier in this book.
In the following sections the procedures applied by Laarhoven, et al. ,
[1983] and Boender, et al., [1989] will be used on the crisp MCDM methods
which have been introduced in the second chapter. As usually, some
numerical examples will also be analyzed in order to achieve a better
illustration of the key concepts in this chapter.
Example 13-1:
Let a decision problem be defined on the four fuzzy decision criteria
CI, C2, C3, C4, and the three fuzzy alternatives AI' A2, and A3 . Next, let us
Chapter 13: Fuzzy Multi-Criteria Decision Making 243
assume that the data for this problem are given in the following fuzzy
decision matrix.
Criteria
C1 C1 C3 C 4
AIts. (0.13,0.20,0.31) (0.08,0.15,0.25) (0.29,0.40,0.56) (0.17,0.25,0.38)
Therefore, when the fuzzy WSM (i.e., the FWSM) approach is used, then the
fInal priority scores (denoted as PI P2 • and P3 • respectively) of the
three alternatives are: •
PI = (0.13, 0.20, 0.31) x (3.00,4.00, 5.00) +
(0.08,0.15,0.25) x (5.00,6.00, 7.00) +
(0.29, 0.40, 0.56) x (5.00, 6.00, 7.00) +
(0.17,0.25,0.38) x (2.00,3.00,4.00) =
=
(2.583, 4.850, 8.750).
In a similar manner, the following results are also derived:
P2 = (1.979, 3.950, 7.625),
and P3 = (3.792,6.550, 11.188).
Q • 8.98
x
Figure 13-1: Membership Preference Functions of the Fuzzy Alternatives AI' A2• and AJ
of Example 13-1 According to the Fuzzy WSM Method.
244 MCDM Methods: A Comparative Study. by E. Triantaphyllou
The best alternative with the crisp version of this model is the one
which satisfies equation (2-2). For the fuzzy version of this model the
corresponding formula becomes:
R[~Kl
AL
= iI (~Kj)-WJ.
J=l alJ
(13-2)
where o'Kj' o'lj and Wj are Afuzzy triangular numbers. Si~ilarly to the crisp
setting, fuzzy alternative Ax dominates fuzzy alternative AL if and only if the
value of the numerator in equation (13-2) is greater than its denominator.
The application of formula (13-2) is also illustrated in the following example:
x
Figure 13-2: Membership Functions of the Fuzzy Alternatives A/. A2 • and AJ of Example
13-2 According to the Fuzzy WPM Method.
Chapter 13: Fuzzy Multi-Criteria Decision Making 245
Example 13-2:
The data used in Example 13-1 are also used here. Therefore, when
relation (13-2) is applied, then the following ratios are obtained:
[ (3.00,4.00,5.00)(·13,.20,.31) X (5.00,6.00,7.00)(·08,.15,.25) X
X (5.00,6.00,7.00)(·29,.40,.56) X (2.00,3.00,4.00)(·17,.25,.38) ] /
[ (6.00,7.00,8.00)(·13,.20,.31) x (5.00,6.00,7.00)(·08,.15,.25) X
X (0.50,1.00,2.00)(·29,.40,.56) X (4.00,5.00,6.00)(·17,.25,.38)] _
(2.355, 4.652, 13.516) / (1.473, 2.887, 9.008).
In the previous calculations the approximation symbol " == " was used,
because the corresponding formula in Section 12.2 uses an approximation as
well. Next, in a similar manner, we get:
Example 13-3:
Consider a decision problem with four fuzzy decision criteria and
three fuzzy alternatives. Now pairwise comparisons are represented as ratios
of fuzzy triangular numbers. Next suppose that when the decision maker is
asked to compare the three fuzzy alternatives in terms of the first fuzzy
criterion by using fuzzy pairwise comparisons, then the following reciprocal
judgment matrix was derived:
246 MCDM Methods: A Comparative Study, by E. Triantaphyllou
Criteria
C1 C2 C 3 C4
Alts. (0.08,0.18,0.46) (0.08,0.16,0.39) (0.17,0.40,0.86) (0.11,0.26,0.61)
~
Then, according to equation (13-1) the final fuzzy priority scores (denoted as
PI' P2 • and P3 ) of the three fuzzy alternatives are as follows:
(0.02, 0.14, 0.99) x (0.08, 0.18, 0.46) +
(0.18, 0.44, 0.95) x (0.08, 0.16, 0.39) +
(0.22,0.37,0.64) x (0.17,0.40,0.86) +
(0.12, 0.23, 0.55) x (0.11, 0.26, 0.61)
= (0.068, 0.474, 1.887).
Similarly, the following two fuzzy preference scores are derived:
= (0.772, 0.208, 1.908),
and (0.897, 0.480, 2.696).
When the procedure described in Section 12.3 for ranking triangular fuzzy
numbers is applied to PI' P2 • and P3 • it can be easily shown that fuzzy
alternative A3 is the best one. •
Example 13-4:
In this example the vectors in the fuzzy decision matrix of Example
13-1 are divided by the largest entry in that vector. In this way the following
fuzzy decision matrix is derived:
248 MCDM Methods: A Comparative Study, by E. Triantaphyllou
Criteria
C1 C2 C3 C4
A1ts. (0.08,0.18,0.46) (0.08,0.16,0.39) (0.17,0.40,0.86) (0.11,0.26,0.61)
In a manner similar to the one used in the original AHP, the final fuzzy
preference scores of the three fuzzy alternatives are calculated as follows:
= (0.08, 0.18, 0.46) x (0.01, 0.21, 9.90) +
(0.08, 0.16, 0.39) x (0.44, 1.00, 2.29) +
(0.17,0.40,0.86) x (0.41,0.69, 1.26) +
(0.11, 0.26, 0.61) x (0.26, 0.50, 1.26) =
= (0.130, 0.605, 2.923).
Example 13-5:
In this example we follow the fuzzy versions of the steps for the crisp
TOPSIS method as described in Section 2.2.6.
Criteria
C1 C2 C3 C4
AIts. (0.13,0.20,0.31) (0.08,0.15,0.25) (0.29, 0.40, 0.56) (0.17,0.25,0.38)
"
Al (0.08, 0.25, 0.94) (0.25, 0.93, 2.96) (0.34, 0.70, 1.71) (0.12, 0.24, 0.92)
..4.2 (0.23, 1.00, 3.10) (0.13,0.60,2.24) (0.03, 0.05, 0.09) (0.12, 0.40, 1.48)
..4.3 (0.15, 0.40, 1.48) (0.13, 0.20, 0.88) (0.62, 1.48, 3.41) (0.24, 1.00,3.03)
Criteria
Alts. C1 C2 C3 C4
"
Al (0.01,0.05,0.29) (0.02,0.14,0.74) (0.10,0.28,0.96) (0.02 ,0.06 ,0.35)
..4.2 (0.03, 0.20, 0.96) (0.01,0.09,0.56) (0.01,0.02,0.05) (0.02, 0.10, 0.55)
..4.3 (0.02, 0.08, 0.46) (0.01,0,03,0.22) (0.18, 0.59, 1.91) (0.04, 0.25, 1.15)
When the ranking procedure described in Section 12.3 is applied here, it can
be easily shown that the previous closeness measures are ranked as follows:
e 3* > e J* > e 2*. Therefore, the preference order of the three fuzzy
alternatives is: 13 > 1J > 1 2 , That is, the best alternative is 1 3, •
Example 13-6:
Let a decision problem involve four decision criteria and three
alternatives. Let all the data be expressed in terms of the same unit of
measurement (e.g. dollars, or time, or weight, etc.). Suppose that the
following decision matrix depicts the actual (and thus unknown to the
decision maker) crisp data for this problem.
Criteria
CI C2 C3 C4
Alts. (6.015 5.526 8.349) 5.721
The decision maker is assumed not to know these data. Next, he/she
is asked to use triangular fuzzy numbers in which the lower, modal, and
upper values are members of the set: {9, 8, 7, ... 2,1,112, ... 117, 118, 1I9}.
These are the values recommended by the Saaty scale [1994] when one wishes
to quantify pairwise comparisons (see also Chapter 3).
It is also assumed here that the decision maker is as accurate as
possible. For instance, when the decision maker attempts to estimate the
performance of the first alternative in terms of the first criterion, the fuzzy
number (3, 4, 5) is used. It can be observed now that the modal value of this
number is the closest number from the above set of numbers to the actual
value of 4.454. Also, the lower and upper values of that fuzzy number are
assumed to be one unit apart. In a similar manner the rest of the entries are
estimated and thus the following fuzzy decision matrix is assumed to had been
obtained by the decision maker.
Criteria
(;1 (;2 (;3 (;4
Alts. (5, 6, 7) (5, 6, 7) (7, 8, 9) (5, 6, 7)
'" (3, 4, 5)
Al (2, 3,4) (3, 4, 5) (5, 6, 7)
12 (3, 4, 5) (7, 8, 9)
(6, 7, 8)
(0.5, 1, 2)
(1, 2, 3)
(1, 2, 3)
(4, 5, 6)
A3 (5, 6, 7)
Therefore, the final fuzzy priority scores of the fuzzy alternatives are
calculated as follows:
= (5,6,7) x (3,4,5) + (5,6,7) x (2,3,4) +
(7,8,9) x (3,4,5) + (5,6,7) x (5,6,7)
= (71, 105, 157).
RCP matrix):
For instance, the value {332 = (1, 2, 3) is derived from a32 (= 1.52)
in which 1.52 has the absolute minimal difference with the value 2 from
Saaty's original scale. Therefore, the modal value of {332 is set to be equal to
2 and the upper and lower values of {332 are equal to 1 and 3, respectively,
which are one unit apart from the modal value. Similarly, the fuzzy pairwise
comparisons (i.e., the corresponding fuzzy CDP matrices) of the three
alternatives in terms of each decision criterion are derived in accordance with
the Saaty scale and are as follows:
- -
4.454 1.22 0.71 (1,1,1) (0.50,1,2) (0.30,0.50,1)
3.647 0.82 1 0.59 (0.50,1,2) (1,1,1) (0.30,0.50,1)
6.232 1.41 1.69 1 (1,2,3) (2,3,4) (1,1,1)
- -
3.253 0.38 0.45 (1,1,1) (0.25,0.30,0.50) (0.30,0.50, 1)
8.450 2.63 1 1.16 (2,3,4) (1,1,1) (0.50,1,2)
7.273 2.22 0.86 1 (1,2,3) (0.50,1,2) (1,1,1)
- -
3.987 1 2.76 (1,1,1) (2,3,4) (1,2,3)
1.447 0.36 1
!.OO] (0.25,0.30,0.50) (1,1,1) (0.30,0.50,1)
0.:8
2.496 0.63 1.72 (0.30,0.50,1) (1,2,3) (1,1,1)
254 MCDM Methods: A Comparative Study, by E. Triantaphyllou
The final fuzzy priority scores PI , p2' and p3' of the three fuzzy
alternatives are derived in a similar manner as stated earlier and are as
follows:
= (0.09,0.26,0.72) x (0.12,0.28,0.66) +
(0.08, 0.22, 0.60) x (0.09, 0.16, 0.34) +
(0.11, 0.30, 0.72) x (0.28, 0.54, 0.97) +
(0.08, 0.22, 0.60) x (0.22, 0.45, 0.90) =
= (0.070, 0.384, 1.946).
Similarly, the following two fuzzy priorities are derived:
= (0.045, 0.245, 1.376),
and = (0.063, 0.371, 1.914).
When these priority scores are ranked as before, then the three fuzzy
alternatives are ranked as follows: "C
> A3 > A2. That is, fuzzy alternative
A1 turns out now to be the best one. Obviously, this is in contradiction with
the results derived when the fuzzy WSM was applied at the beginning of this
illustrative example.
At the same time, it can also be observed that the entire ranking of
the fuzzy alternatives as derived by the fuzzy WSM and the fuzzy AHP has
also changed (that is, from: A3 > A1 > A2 it has changed to: A1 > A3 >
A2). Therefore, a contradiction occurs between the fuzzy WSM and the fuzzy
AHP when one compares the entire ranking orders of the three alternatives
Chapter 13: Fuzzy Multi-Criteria Decision Making 255
It is possible that the best alternative derived from the fuzzy WSM
and the other fuzzy methods be identical but the remaining alternatives change
their orders. The revised (i.e., ideal mode) AHP, the WPM and the TOPSIS
methods can be examined as above and it can be similarly demonstrated that
they also yield contradictions when a single-dimensional environment is
assumed and the fuzzy WSM is used as the norm.
Example 13-7:
As with the previous numerical example suppose that a decision
problem with four criteria and three alternatives has the following decision
matrix. Again, it is assumed that these values are unknown to the decision
maker.
Criteria
C1 C2 C3 C4
Alts. (2.885 3.987 2.434 4.894)
Criteria
C1 C2 C3 C4
Alts. (O.OS, O.lS, 0.47) (0.11,0.30,0.74) (O.OS, O.lS, 0.47) (0.13,O.35,O.S2)
41
~
Also, the final priority scores PI' P2 • and P3 • of the three alternatives
can be shown to be as follows:
PI = (0.073, 0.361, 1.889),
P2 = (0.073, 0.355, 1.824),
and P3 = (0.061, 0.2~4, 1.388).
Apparently, fuzzy alternative A 1 is the best one.
Next, the crisp alternative A3 (which is not the best one) in the
original matrix is replaced by A/ which is worse than the original A3 • The
performance values of A/ are the same as in the original alternative A3 except
that the third value in terms of criterion C3 of 8.501 is substituted by the least
one 1.064. Thus, the original matrix of crisp numbers is modified as follows:
Criteria
CJ C2 C3 C4
Alts. (2.885 3.987 2.434 4.894)
Criteria
C J C2 C3 C4
Alts. (0.08, 0.18, 0.47) (0.11,0.30,0.74) (0.08,0.18,0.47) (0.13, 0.35, 0.82)
Al (0.35, 0.59, 0.91) (0.22, 0.43, 0.83) (0.10, 0.17, 0.29) (0.13, 0.33, 0.84)
A2
A /
(0.21,0.32,0.57) (0.22, 0.43, 0.83) (0.45,0.67,0.97) (0.13, 0.33, 0.84)
A3 (0.06,0.09,0.13) (0.09,0.14,0.26) (0.10,0.17,0.29) (0.13, 0.33, 0.84)
Working as before, it can be seen that the final fuzzy priority scores now
become:
PI = (0.077, 0.377, 1.872),
P2 = (0.092, 0.418, 2.030),
and P!.3 = (0.040, 0.205, 1.082).
Chapter 13: Fuzzy Multi-Criteria Decision Making 257
From the above fuzzy scores it is obvious that now the best fuzzy
alternative is A2 . This result is in contradiction with the earlier result, namely
that the best fuzzy alternative is AI. This analysis indicates that a
contradiction may occur when the fuzzy AHP (original version) is used and
a nonoptimal alternative is replaced by a worse one. In a similar manner it
can be shown that the ideal mode (revised) AHP, WPM and TOPSIS methods
may also fail when they are tested in a similar manner. •
Rll: Is the rate that the fuzzy WSM and another fuzzy method
disagree in the indication of the best alternative.
R12: Is the rate that the fuzzy WSM and another fuzzy method
disagree on the entire ranking of the alternatives.
R21: Is the rate that a method changes the indication of the best
alternative when a nonoptimal alternative is replaced by a
worse alternative.
values according to the original Saaty scale). The fuzzy AHP (original and
revised), the WPM, and the TOPSIS methods were then examined in terms
of the two fuzzy evaluative criteria in a manner similar to the procedures
described in Examples 13-6 and 13-7. The computational results are depicted
in Figures 13-4 to 13-8 and are also discussed in the following two sub-
sections.
.35
.3
.25
20
0::
c .2
0
~
'0
.1 5 ,
+---i---r-+-b.:-;j2-: t-+-r-
~c ~-r=':~~-~~---;
0
<->
.1
0
4 6 8 10 12 14 16 18 20
Number of Criteria
Figure 13-3: Contradiction Rate R11 When the Number of Fuzzy Alternatives
is Equal to 3 .
.55
: :
.5
.45
.4
20 .35 ~- - - .. _..... .
0::
c
0 .3
:.-
0
'0 .25
~c
~4~g
0 .2
<->
.15
.1
.05
0
4 6 8 10 12 14 16 18 20
Number of Criteria
Figure 13-4: Contradiction Rate R11 When the Number of Fuzzy Alternatives
is Equal to 21.
260 MCDM Methods: A Comparative Study. by E. Triantaphyl/Ou
.1
r ,
i F-IjpM
\" , ,
.06 ···"..:··,~,,-·····---······--·1····-···-·--·----··!--·
G>
'0 \L ~
c::
c
.2
t) .04
"6
~
c:0
u .02
o
4 6 8 10 12 14 16 18 20
Number of Criteria
Figure 13-5: Contradiction Rate R21 When the Number of Fuzzy Alternatives
is Equal to 3 .
.2
.15
',I
20 l i ! I !
i, ! i, ! j
c::
c:
!"" !
i i I
l
!
i
:8u
"6
E
.1
iii:
i .--...-.. ---.--..i-.-.---..-.'.'.--~~ ..'-.....--.-....1·-----·---··-1..---.................L..............--•.•.. -·l--··-··-·····-~----------;-
! l .
t1t
c:
0
(.)
.05
o ~-L~__~~____~~__J-~__-L__~~~__~~~
...•~._
.. -_ •.~
.-~ ...._...~~~
4 6 8 10 12 14 16 18 20
Number of Criteria
Figure 13-6: Contradiction Rate R21 When the Number of Fuzzy Alternatives
is Equal to 21.
Chapter 13: Fuzzy Multi-Criteria Decision Making 261
.5
.45
.4
a
II>
.35
cr
.2 .3
t>
'g... .25
C
8 .2
.15
.05 ·.·---·.i---·-----.-----i·---··-·---··--t··-·------·.·-l--------·-f-·----·-·---·-r--···--·-·-···t---··~-··--·---·t-----·····--·--·····[·---····
! ! . 1 j ! ~
4 6 8 10 12 14 16 18 20
Number af Criteria
Figure 13-7: Contradiction Rate R12 When the Number of Fuzzy Alternatives
is Equal to 3.
parameter was equal to 0.50 and 1.00. However, the derived contradiction
rates were significantly higher than the ones derived when the Saaty scale was
used, and thus these results were not plotted. In summary, the findings of
this study reveal that some fuzzy MCDM methods are better than others in
some cases even though none of the methods is perfectly accurate.
13.9 CONCLUSIONS
The analyses presented in this chapter reveal that none of the four
fuzzy MCDM methods is completely perfect in terms of both fuzzy evaluative
criteria. Different contradiction rates are yielded when these fuzzy MCDM
methods are tested according to the two evaluative criteria. The fuzzy WSM
could be the simplest method to solve single-dimensional decision making
problems. However, the other more systematic approaches; the fuzzy AHP,
the fuzzy revised RAHP, the fuzzy WPM, and the fuzzy TOPSIS, are more
capable of capturing a human's appraisal of ambiguity when complex MCDM
problems are considered. This is true because pairwise comparisons provide
a flexible and realistic way to accommodate real life data. The experimental
results reveal that the fuzzy revised (ideal mode) AHP is better than the other
methods in terms of the previous two fuzzy evaluative criteria.
It needs to be emphasized here that these fuzzy MCDM methods are
best to be used as decision tools. Individual decision makers may reach their
own solution after applying anyone of them. This study provided only a
general view of different methods under certain situations. A broader
understanding of the characteristics of the methods, and fuzzy evaluative
criteria is required for a successful accomplishment in solving real life fuzzy
MCDM problems.
Chapter 14
It is widely accepted today that people do not always behave the way
the well studied normative theories say they ought to behave (see, for
instance, [Allais and Hagen, 1979], [Bell, et al., 1988], [Ellsberg, 1961], and
Raiffa [1984]. Many decision theories (especially game theories) assume that
the decision makers are always perfectly rational. Too often, however,
important decisions are based on non-scientific clues (see, for instance,
[Kadane and Larkey, 1982a; 1982b; and 1983], [McMillan, 1992], [Neale
and Bazerman, 1991], [Raiffa, 1982], and [Sebenius, 1992].
Even well defined classical statistic inference procedures can be of
limited benefit because the input probabilities are hard to be estimated or the
underlying assumptions are not valid at a given situation or they cannot be
validated properly (see, for instance, [French, 1986], [Pratt, et al., 1994],
[Raiffa, 1968], and [Raiffa and Schlaifer, 1968]). The above issues constitute
the three major concerns on the future of research and practice as was
expressed recently by Raiffa [1994].
As an antidote to the above problems, Raiffa in [1994] proposed a
new orientation to the study of decision making which he calls the prescriptive
orientation. According to this orientation, people should not be considered
as super rational individuals. Instead new methods should seek to establish
a systematic and wise approach in making decisions. Raiffa also proposed
that the study of decision making should combine normative theories with a
deep understanding of the cognitive and behavioral aspects involved in real
life decision making.
Ackoff, R.L., (1978), "The Art of Problem Solving, " John Wiley & Sons, Inc., New York,
NY, USA.
Allais, M., and O. Hagen (Eds.), (1979), "Expected Utility Hypothesis and the Allais
Paradox, " D. Reidel, Dordrecht, The Netherlands.
Anderberg, M.R., (1973};, "Probability and Mathematical Statistics, " Cluster Analysis for
Applications. Academic Press, New York, NY, USA.
Arbel, A., and A. Seidmann, (1984), "Selecting a Microcomputer for Process Control and
Data Acquisition, " IEEE Transactions, Vol. 16, No. I, pp. 16-25.
Armacost, R.L., and J.C. Hosseini, (1994), "Identification of Determinant Attributes Using
the Analytic Hierarchy Process, "Journal of the Academy of Marketing Science, Vol.
22, No.4, pp. 383-392.
Baas S.J., and H. Kwakernaak, (1977), "Rating and Ranking of MUlti-Aspect Alternatives
Using Fuzzy Sets, "Automatica, Vol. 13, pp. 47-58.
Barzilai, J., and F.A. Lootsma, (1994). "Power Relations and Group Aggregation in the
Multiplicative AHP and SMART, " Proceedings of the Third International Symposium
on the AHP, George Washington University, Washington, DC, USA, 157-168.
Bell, R.E., and P.H. Farquhar, (1986), "Perspectives on Utility Theory," Operations
Research, Vol. 24, pp. 179-183.
Bell, D.E., H. Raiffa, and A. Tversky, (1988), Decision Making: Descriptive, Normative,
and Prescriptive Interactions, Cambridge University Press, Cambridge, England.
Bellman, R.E., and L.A. Zadeh, (1970), "Decision-Making in a Fuzzy Environment,"
Management Science, Vol. 17B, No.4, pp. 141-164.
Belton, V., and T. Gear, (1983), "On a Short-Coming of Saaty's Method of Analytic
Hierarchies, " Omega, pp. 228-230.
Ben-Arieh, D., and E. Triantaphyllou, (1992), "Quantifying Datafor Group Technology with
Weighted Fuzzy Features, " International Journal of Production Research, Vol. 30,
No.6, pp. 1285-1299.
Benayoun, R., B. Roy, and N. Sussman, (1966), "Manual de Reference du Programme
Electre, " Note de Synthese et Formaton, No. 2S, Direction Scientifque SEMA, Paris.
France.
Boender, C.G.E., J.G. de Graan, and F.A. Lootsma, (1989), "Multi-Criteria Decision
Analysis with Fuzzy Pairwise Comparisons, " Fuzzy Sets and Systems, Vol. 29, pp.
133-143.
Boucher, T.O., and E.L. McStravic, (1991), "Multi-Attribute Evaluation within a Present
Value Framework and Its Relation to the Analytic Hierarchy Process," The
Engineering Economist, Vol. 37, No. I, pp. 55-71.
Bridgman, P.W., (1922), "Dimensional Analysis," Yale University Press, New Haven, CT,
USA.
Buckley, J.J., (1985), "Ranking Alternatives Using Fuzzy Numbers," Fuzzy Sets and Systems,
Vol. 15, pp. 21-31.
Buckley, J.J., (1985), "Fuzzy Hierarchical Analysis," Fuzzy Sets and Systems. Vol. 17, pp.
233-247.
Cambron, K.E., and G.W. Evans, (1991), "Layout Design Using the Analytic Hierarchy
Process, " Computers and Industrial Engineering, Vol. 20, No.2, pp. 221-229.
Chang, S.K., (1971), "Proceedings of the Brooklyn Poly technical Institute Symposium on
Computers and Automata," Fuzzy Programs, Brooklyn, New York, NY, USA. Vol.
268 MCDM Methods: A Comparative Study, by E. Triantaphyllou
111-115.
Khurgin, J. I., and V.V. Polyakov, (1986), "Fuzzy Approach to the Analysis of Expert
Data," Fuzzy Sets Applications. Methodological Approaches and Results,
Akademie-Verlag, Berlin, Germany, pp. 116-124.
Kuhn, H.W., and A.W. Tucker, (1951), "Nonlinear Programming," in Proceedings of the
Second Berkeley Symposium on Mathematical Statistics and Probability. J. Neyman
(Ed.), University of California at Berkeley Press, Berkeley, Calif., USA, pp. 481-
492.
Laarhoven, P.J.M., and W. Pedrycz, (1983), "A Fu¥y Extension of Saaty's Priority
Theory, " Fuzzy Sets and Systems. Vol. 11, pp. 229-241.
Lakoff, G., (1973), "Hedges: A Study in Meaning Criteria and the Logic of Fuzzy Concepts, "
J. Philos. Logic. pp. 12-22.
Lee, C.T.R., (1971), "Fuzzy Logic and the Resolution Principle," Second International Joint
Conference on Artificial Intelligence. London, England, pp. 560-567.
Lee, S.N., Y.L. Grlze, and K. Dehnad, (1987), "Quantitative Modelsfor Reasoning Under
Uncertainty in Knowledge Based Expert Systems", International Journal of Intelligent
Systems. Vol. 2, pp. 15-38.
Legrady, K., F.A. Lootsma, J. Meisner and F. ScheUemans, (1984), "Multi-Criteria
Decision Analysis to Aid Budget Allocation,· in M. Grauer en A.P. Wierzbicki,
(Editors), Interactive Decision Analysis, Springer-Verlag, Berlin-New York, NY,
USA, pp. 164-174.
Lootsma, F .A., (1999), "Multi-Criteria Decision Analysis via Ratio and Difference Judgment. "
Kluwer Academic Publishers. Applied Optimization Series, Vol. 29, Dordrecht, The
Netherlands.
Lootsma, F .A., (1997), "Fuzzy Logic for Planning and Decision Making." Kluwer Academic
Publishers. Applied Optimization series, Vol. 8, Dordrecht, The Netherlands.
Lootsma, F.A., (1990), "The French and The American School in Multi-Criteria Decision
Analysis,· Recherche Operationnele / Operations Research, Vol. 24, No.3, pp.
263-285.
Lootsma, F .A., (1980), "Saaty's Priority Theory and the Nomination of a Senior Professor in
Operations Research, • European Journal of Operational Research, Vol. 4, pp. 380-
388.
Lootsma, F.A., T.C.A. Mensch, and F.A. Vos, (1990), "Multi-Criteria Analysis and Budget
Reallocation in Long-Term Research Planning, " European Journal of Operational
Research, Vol. 47, pp. 293-305.
Lootsma, F.A., (1991), "Scale Sensitivity and Rank Preservation in a Multiplicative Variant
of the AHP and SMART," Technical Report 91-67, Faculty of Technical
Mathematics and Informatics, Delft University of Technology, Delft, The
Netherlands.
Lootsma, F.A., (1988), "Numerical Scaling of Human Judgment In Pairwise-Comparison
Methods For Fuzzy Multi-Criteria Decision Analysis," Mathematical Models for
Decision Support, NATO ASI Series F, Computer and System Sciences, Springer-
Verlag, Berlin, Germany, Vol. 48, pp. 57-88.
Ma, D., and X. Zheng, (1979), "9/9-9/1 Scale Method of the AHP," Proceedings of the 2nd
International Symposium on the AHP, Vol. I, Pittsburgh, PA, USA, pp. 197-202.
Mardia, K.V., J.T. Kent, and J.M. Bibby, (1979), "Multivariate Analysis." Academic
Press, New York, NY, USA.
Marks, L.E., (1974), "Sensory Processes. The New Psychophysics. " Academic Press, New
York, NY, USA.
References 271
Masuda, T., (1990), "Hierarchical Sensitivity Analysis of the Priorities Used in the Analytic
Hierarchy Process," Inter'l Journal of Systems Science, Vol. 21, No.2, pp.
415-427.
McCahon, C.S., and E.S. Lee, (1988), "Comparing Fuzzy Numbers: The Proportion of the
Optimum Method," Inter'l Journal of Approx. Reasoning, Vol. 4, pp. 159-163.
McMillan, J., (1992), Games, Strategies, and Managers, Oxford University Press, Oxford,
England.
Michon, J.A., Eijkman, E.G.J., and L.F.W. de Klerk, (Editors), (1976), "Handboek der
Psychonomie, " (in Dutch), Van Loghum Slaterus, Deventer, The Netherlands.
Miller, D.W., and M.K. Starr, (1969), "Executive Decisions and Operations Research, "
Prentice-Hall, Inc., Englewood Cliffs, NJ, USA.
Miller, C.A., (1956), "The Magic Number Seven Plus or Minus Two: Some Limits on Our
Capacity for Processing Information, " Psychological Review, Vol. 13, pp. 81-97.
Munier, B.R., (Editor), (1988), "Risk, Decision and Rationality," D. Reidel, Dordrecht,
The Netherlands.
Neale, M.S., and M.H. Bazerman, (1991), Cognition and Rationality in Negotiation, The
Free Press, Amsterdam, The Netherlands.
Pratt, J.W., H. Raiffa, and R.O. Schlaifer, (1994), Introduction to Statistical Decision
Theory, M.I.T. Press, Boston, MA, USA.
Prade, H., and C.V. Negoita, (Editors), (1986), "Fuzzy Logic in Knowledge Engineering,"
Verlag TUV Rheinland, Berlin, Germany.
Putrus, R., (1990), "Accounting for Intangibles in Integrated ManUfacturing (nonfinancial
justification based on the Analytical Hierarchy Process, "Information Strategy, Vol.
6, pp. 25-30.
Raiffa, H., (1994), "The Prescriptive Orientation of Decision Making: A Synthesis of
Decision Analysis, Behavioral Decision Making, and Game Theory, " in Decision
Theory and Decision Analysis: Trends and Challenges, Sixto Rios, (Ed.), Kluwer
Academic Publishers, Boston, Mass., USA.
Raiffa, H., (1984), Back from Prospect Theory to Utility Theory, Plural Rationality and
Interactive Decision Processes, LNEMS, Springer-Verlag, New York, NY, USA,
pp. 100-113.
Raiffa, H., (1982), The Art and Science of Negotiations, Harvard University Press. Boston,
Mass., USA.
Raiffa, H., (1968), Decision Analysis, Addison Wesley, Boston, Mass., USA.
Raiffa, H., and P.J.H. Schlaifer, (1968), Applied Statistical Decision Theory, M.I.T. Press,
Boston, Mass., USA.
Ramsay, A., (1988), "Formal Methods in Artificial Intelligence, " Cambridge University
Press, Cambridge, England.
Ray, T., and E. Triantaphyllou, (1999), "Procedures For the Evaluation of Conflicts in
Rankings of Alternatives, " Computers and Industrial Engineering, Vol. 36, No.1,
pp. 35-44.
Ray, T., and E. TriantaphyUou, (1998), "Evaluation of Rankings With Regard to the
Possible Number of Agreements and Conflicts, " European Journal of Operational
Research, Vol. 106, No.1, pp. 129-136.
Roberts, F.S., (1979), "Measurement Theory, " Addison-Wesley, Reading, Mass., USA.
Roper-Lowe, G.C., and J.A. Sharp, (1990), "The Analytic Hierarchy Process and its
Application to an Information Technology Decision, " Operational Research Society
Ltd., Vol. 41, No.1, pp. 49-59.
Ruspini, E,H., (1991), "On the Semantics of Fuzzy Logic," Inter'l Journal of Approximate
272 MCDM Methods: A Comparative Study, by E. TriantaphyUou
A
absolute-any (AA) critical criterion 134, 141, 152
absolute terms 133
absolute-top (AT) critical criterion 134, 141, 151
actual relative weights 98
additive function 228
additive utility assumption 6-7
AHP axioms 215
AHP variant by Dyer 222-223
all possible comparisons 105-112
alternatives 1
analytic hierarchy process (AHP) 9-11, 116, 137-141, 157-160,
165-166, 179-182, 186-188, 197-199,
201-212, 215-217, 225-227, 241
attributes 2
available comparisons 105-112
average consistency index of CDP matrices 41-42
average squared residual 65
B
best scales 51,53
best MCDM method 199
best fuzzy MCDM method 261-262
bound of maximum consistency 40
"Bridge Evaluation" problem 230-231
C
case studies 230-232
common characteristic 74
common comparisons 105-112
complete pairwise comparisons 88
conflict among criteria 1
concordance index 14, see also TOPSIS (fuzzy and crisp)
concordance matrix 16, see also TOPSIS (fuzzy and crisp)
concordance set 16,
see also under TOPSIS (fuzzy and crisp)
consistency see perfect consistency
276 MCDM Methods: A Comparative Study, by E. Triantaphyllou
D
decision criteria 1
decision making paradox see paradox
decision matrix 2-3
decision space 1
decision support tool 207
decision weights 2
decomposition of judgment matrices 88-113
definiteness property 74
descriptive theories 265
deterministic MCDM 2
difference comparisons 74-86
difference judgments 76-86
dimensionless analysis 8
discordance index 14, see also TOPSIS (fuzzy and crisp)
discordance matrix 16-17, see also TOPSIS (fuzzy and crisp)
discordance set 16, see also TOPSIS (fuzzy and crisp)
discrete decision spaces 1
dissimilarity relation 74
distance (of similarity) 74
dominance matrices 17-18, see also TOPSIS (fuzzy and crisp)
duality approach 115-129
dual problem see duality approach
Subject Index 277
E
eigenvalue approach 44, 58-59, 217
eigenvector see eigenvalue approach
eigenvector approximation 58-59
see eigenvalue approach
eigenvector method see eigenvalue approach
ELECTRE method 13-18, 241
error terms 93
evaluative criteria (crisp) 43, 177-199
evaluative criteria (fuzzy) 250-262
Euclidean distance 19-20
examples of exponential scales 29-32
Expert Choice (computer software) 132, 201, 212
exponential scales 24,28-32
H
hierarchic composition 214
hierarchies (multiple) 131
hierarchical structure 1
human rationality assumption 61-71
F
feasible solution 85
feasible value 137, 163
''flat maxima principle" 131
fuzzy AHP 245-247, 262
fuzzy alternatives 236, 239, 242-262
fuzzy decision criteria 236, 242-262
fuzzy CDP matrix 251-257
fuzzy data 235
fuzzy databases 87
fuzzy decision matrix 243, 254-256
fuzzy evaluative criteria see evaluative criteria
fuzzy operations 236-329
fuzzy MCDM 2, 241-262
fuzzy numbers 166, 236-237, 239, 241
fuzzy RCP matrix 251-257
fuzzy reciprocal (judgment) matrix 253-260
fuzzy revised AHP 247-248, 258, 262
278 MCDM Methods: A Comparative Study, by E. Triantaphyllou
G
goals 1-2
group decision making 2
guided (pairwise comparisons) 88
Q
quadratic problem see quadratic programming
quadratic programming 75,79-85
I
ideal mode AHP see, revised AHP
ideal solution 20
identical alternatives 12, 214
incommensurable units 2
inconsistent CDP matrix 37
InfoHarvest, Inc. 201
L
Langrangian multipliers 81
Langrangian (the) 81
large size decision problems 128
law of stimulus of measurable magnitude 26
law of stimulus perception 28
least squares 63-67
linguistic choices 24,28
linear equation, system of 79,82
linear programming 92-97, 112
linear scale 24, see also Saaty scale
Lootsma scales see exponential scales
logical contradiction 222
Subject Index 279
M
matrix partitioning see decomposition
matrix transpose 82
maximum eigenvalue see eigenvalue approach
maximum similarity 73-74
maximum dissimilarity 73-74
maximum consistency of CDM matrices 38-42
maximum consistency index 40-42
membership value 57, also see fuzzy numbers
modal value 237
most critical criterion 133-135, 138-139, 141
most important criterion 144
most sensitive alternative 155-156
missing pairwise comparisons 86,91
Multi-Attribute Decision Making (MADM) 1
multi-attribute utility theory (MAUT) 214
Multi-Criteria Decision Making (MCDM) (definitions) 1-22
multi-dimensional MCDM 8
Multi-Objective Decision Making (MODM) 1
multiple attributes 1
multiple hierarchies 131-132
multiple objective functions 1
multiplicative AHP 228-233, see also WPM
N
negative-ideal solution 20
normalized decision matrix 14-16, 19
normalized columns 117-118
normalized rows 117-118
normative theories 265
number of alternatives (role of) 207
o
optimization approaches 60-67
optimal solution 83-85
outranking relations 13-14
280 MCDM Methods: A Comparative Study, by E. Triantaphyllou
p
pairwise comparisons 23, 25-32
paradox 2, 42, 145, 197-199, 265
partial ranking 221
partitioning of pairwise comparisons 90
see also decomposition of judgment matrices
percent-any (PA) critical criterion 134-135, 140-141, 145, 150
percent-top (PT) critical criterion 134-135, 140-141, 145, 149
perfect consistency 95, 215, 218
performance values see decision matrix
power law 32
power method 67
prescriptive theories 265
prime approach see duality approach
prime problem see duality approach
primal approach see duality approach
R
random consistency index (RCI) 59, see also consistency index
ranking abnormalities 213-233
ranking of fuzzy numbers 238-239
ranking reversal 43
ranking indiscrimination 43
ratio comparisons 57-72
Real Continuous Pairwise (RCP) matrix 32-34, 80, 97, 146,
202-205, 251
real life case studies 230-232
reciprocal comparisons see pairwise comparisons
reciprocal matrices 88-89, see also pairwise comparisons
redundant constraints 96
relative closeness 20
relative importance 75,85
relative priorities 89, see also relative importance
relative magnitudes 25
relative ranking 80
relative similarity 75-85, also, see similarity
relative terms 132
relative weights 57-72, 73-86, 87-113
revised AHP 11-13, 116, 189, 191-193, 197-199,
201-212, 214, 218-222, 225-227
Subject Index 281
S
Saaty scale 26-27
scale 24-55,75
scale generation 27
scale evaluation 32-55
separation measure 20
SENSA ITO library 131
sensitivity analysis 131-175
sensitivity coefficient (of a criterion) 136
sensitivity coefficient (of an alternative) 156
sensitivity coefficient (of measures of performance) 160, 164
similarity function 73
similarity scale 77
similarity measure 74
single-dimensional MCDM 8
"Site Selection" problem 232
stepwise approach 166
stochastic MCDM 2
stimulus of measurable magnitude 26
stimulus perception 28
symmetric comparisons see difference judgments
symmetric comparisons 76
system of linear equations see quadratic programming
T
taxonomy of MCDM methods 4
threshold value(s) 155, 158-159, 161, 163, 175
TOPSIS method 18-21, 194-196, 241
triangular property 74,85
V
"vector-maximum" problem 1
U
union operation 87
282 MCDM Methods: A Comparative Study, by E. Triantaphyllou
units of measure 8
upper bound of reduction rate 120
unrestricted variables 93
utility theory 214
W
weighted product model (WPM) 8-9, 142-145, 154, 161-164, 165,
169-170, 174-175, 180, 183, 185,
192, 197-199, 228-233, 241
weighted sum model (WSM) 6-7, 137-141, 157-160, 165, 167-168,
170-171, 179-185, 241, 197-199
worst scales 52-54
AUTHOR INDEX
A
Ackoff, R. L. 23,25
Allais, M. 265
Anderberg, M.R. 74
Arbel, A. 201
Armacost, R. 313
Amoff 25
B
Baas, S.J. 238
Barzilai, J. 228,233
Bazerman, M.H. 265
Bell, D.E. 265
Bellman, R.E. 235
Belton, V. 5, 11, 12,53, 116201,212,
213, 236, 247
Benayoun, R. 13, 14
Benhajla, S. 201
Berdosian, S.D. 235
Boender, e.G.E. 236, 237, 242, 245
Bonissone, P.P. 238
Boucher, T.O. 201
Bridgman, P. W. 8
Buckley, J.J. 236, 237, 238, 241
C
Cambron, K.E. 201
Camerer, C. 214
Chang, S.K. 235
Chen, S.J. 1, 3, 4, 233
Chu, A.T.W. 32,57,60,64
Churchman 25
D
Dantzig, G. 165
Dehnad, K. 235
284 MCDM Methods: A Comparative Study, by E. Triantaphyllou
Dodd, F.J. 41
Donegan, H.A. 41
Dubois, D. 57, 74, 235, 236, 237
Dyer, J.S. 212, 214, 222
E
Ellsberg, D. 265
Evans, G.W. 201
F
Falkner, C. 201
Federov, V.V. 32, 57, 60, 64, 66
Finnie, J. 201
Fishburn, P. C. 6,214
French, S. 131, 265
Fu, K.S. 235
G
Gear, T. 5, 11, 12, 53, 116, 201,
212, 213, 247
Grize, Y.L. 235
Gupta, M.M. 57,235
Gutsche 264
H
Haerer, W. 201
Hagen, O. 265
Hallowell Davis, M.D. 32
Harker, P.T. 88,214
Hedge, G.G. 232
Hihn, J.M. 32
Hogarth, R.M. 214
Hosseini, J.C. 132
Hwang, F. 3
Hwang, C.-L. 1, 3, 4, 18, 233, 236
Author Index 285
J
Johnson, C.R 32
Insua, RD. 131
K
Kadane, J.B. 265
Kalaba, R.E. 66
Kaufmann, A. 74
Khurgin, J.1. 32
Kuhn H.W. 1
Kwakernaak, H. 238
L
Laarhoven, PJ.M. 236, 237, 239, 241, 242, 245
Lakoff, G. 57
Larkey, P.D. 265
LaValle, I.H. 214
Lee, C.T.R. 235
Lee, E.S. 238
Lee, S.N. 235
Lootsma, F.A. 14, 23, 24, 25, 28, 29,
32, 45, 146, 228, 233, 236, 239, 241, 261
Lin, C.-T. 166,241,267,261
M
Ma, D. 26
Mann, S.H. 13, 43, 44, 53, 57, 60, 145,
177, 194,201,207,258
Mardia, K. V. 74
Masuda, T. 131
Marks, L.E. 32
McCahon, C. S. 238
McMillan, J. 265
McStravic, E.L. 201
Michon, J .A. 32
Miller, C.A. 26,76
Miller, D.W. 8,233
Munier, B.R 214
286 MCDM Methods: A Comparative Study, by E. Triantaphyllou
N
Neale, M.A. 265
Negoita C.V. 235
o
O'Keefe, W.D. 201
p
Pardalos, P.M. 57, 61, 145
Pedrycz, W. 236, 237, 245
Polyakov, V. V. 32
Prade, H. 57, 74, 235, 236, 237
Pratt, J.W. 265
Putrus, R 201
R
Raiffa, H. 265
Ramsay, A. 235
Ragade, RK. 235
Ravinder, H. V. 212, 214
Raz, T. 201
Reder, M.W. 214
Roberts, F.S. 28,32
Roper-Lowe, G.C. 201
Roy, B. 13
Ruspini, E.H. 73, 77, 85
S
Saaty, T.L. 5, 9, 10, 23, 24, 25, 26, 27, 32,
40, 57, 58, 66, 89, 117, 201, 214, 217, 231
Sanchez, A. 131, 132, 133, 165
Sarin, RK. 214
Schlaifer, R. O. 265
Sebenius, J.K. 265
Seidmann, A. 201
Sharp, J.A. 201
Shimura, M. 235
Author Index 287
Sokal, RR. 74
Sneath, P.R.A. 74
Starr, M.K. 8,233
Stevens, S.S. 32
Stewart, S.M. 63
Swann, K. 201
T
Tadikamalla, P.R. 232
Tate, M.D. 61
Tong, RM. 238
Triantaphyllou, E. 13, 23, 32, 33, 41, 44, 53,
57,61,73,77, 115, 116, 118, 120, 131,
132, 133, 145, 146, 165, 166, 177, 184,201,
207, 213, 228, 232, 233, 241, 257, 258, 261
Tucker, A.W. 1
V
Vargas, L.G. 32,59,214
X
Xie, L.A. 235
y
Yager, R.Y. 235
Yoon, K. 18, 236
W
Wabalickis, RN. 201
Wang. L. 201
Weber, A. 26,76
Weber, M. 214
Wendell, W.E. 131, 212, 214
Winkler, RL. 212, 214
Winston, W. L. 81, 131
288 MCDM Methods: A Comparative Study, by E. Triantaphyllou
Write, C. 61
Z
Zadeh, L.A. 235
Zheng, X. 26
Zhu, Q. 238
Zimmermann, H.-J. Foreword, 1,2,264
Zwicker, E. 32
ABOUT THE AUTHOR