You are on page 1of 14

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 28, NO.

2, APRIL 1998 205

Linguistic Labels for Expressing Fuzzy Preference


Relations in Fuzzy Group Decision Making
Marimin, Motohide Umano, Associate Member, IEEE, Itsuo Hatono, and Hiroyuki Tamura, Member, IEEE

Abstract—This paper proposes improvements to pairwise group represent the imprecise preference relations. The linguistic
decision making based on fuzzy preference relations in three labels are more appropriate for approximating preference val-
ways. First, it extends the fuzzy preference relation representation ues which are too complex to be represented using precise
using linguistic labels. Decision makers may express their pref-
erence relations in linguistic labels which are more practically numerical values [30]. We propose to revise the computation
implementable for solving group decision making problems. Sec- model accordingly, by using fuzzy sets representation and
ond, it modifies the computational procedures by using fuzzy sets computation, by eliminating the strict threshold value from
representation and computation, and by avoiding the use of strict the computational procedures and by using a neat ordered
threshold values. This allows natural representation, preserves weighted average (OWA) aggregation operator [27] for ag-
the preference accuracy, and produces more intuitively meaning-
ful solutions. Finally, it considers fuzzy criteria of the alternatives gregating the preferences. We also apply the operator for
explicitly. Solutions are first derived based on each criterion, and handling fuzzy linguistic quantifier (e.g., most) to all possible
then by using neat ordered weighted average (OWA) operator the computation procedures which were initially proposed in the
final solutions which accommodate all criteria are determined. group decision making area by Fedrizzi et al. [7] for a
The proposed method is verified for solving fuzzy group decision consensus degree derivation. Moreover fuzzy criteria have
making problems, i.e., advertising media selection cases.
been explicitly incorporated into the proposed model.

I. INTRODUCTION II. GROUP-DECISION ANALYSIS BASED


ON FUZZY PREFERENCE RELATIONS
I N a pairwise group decision making based on fuzzy pref-
erence relations, there are a set of
, and a set of
alternatives,
individual decision makers,
For individual decision maker , the fuzzy preference rela-
tion is given by its membership function
. Each decision maker provides his in [4], [9], and [13]
or her preference over , which is usually fuzzy in nature. A
solution is an alternative (or a set of alternatives) which is the
most acceptable to the group of decision makers as a whole. if is definitely preferred to
One of the possible methods for deriving the solution is if is slightly preferred to
proposed by Skala [23] and Bezdek et al. [4]. This model is if there is no preference (i.e., indifference)
rigid, moreover, each decision maker has to provide numerical if is slightly preferred to
values directly which are not convenient. This model has been if is definitely preferred to
extended into Nurmi’s -degree model [19] which is applied (1)
by Kacprzyk [9], Kacprzyk et al. [13], and Fedrizzi et al.
[6]–[8], for deriving solution of the pairwise fuzzy group is a matrix of preference relations for individual
decision making problem. However, in some extent some whose elements are given by , where
information is still lost in the computation processes. More- , (it could also be represented as
over, criteria for selecting the alternatives are not considered ‘ ’), for all . So, in that case, the preference degree
explicitly in the model. is represented by using numerical value in directly.
In this paper, first we briefly explain pairwise group de- There are indefinite values, i.e., the term “slightly” represents
cision analysis based on fuzzy preference relations, and then indefinite numbers of the numeric preference values which are
propose the use of linguistic labels for expressing the fuzzy within the range or . This representation is
preference relations. Because the decision makers usually tend impractical to decision makers especially who are not familiar
to represent their preferences in linguistic labels rather than with the fuzzy method.
in numerical values, this representation is more realistic to There are two approaches for solving the pairwise group
decision making problem; direct and indirect derivations of a
Manuscript received June 18, 1995; revised March 17, 1996 and November solution set [9], [13]. In the direct approach, a core concept
25, 1996.
Marimin is with the Department of Agro-industrial Technology, Faculty of is used to derive solutions by processing the individuals
Agricultural Technology, Bogor Agricultural University, Bogor, Indonesia. fuzzy preference relations directly. The core is defined as a
M. Umano, I. Hatono, and H. Tamura are with the Department of Systems set of undominated alternatives, i.e., those are not defeated
and Human Science, Graduate School of Engineering Science, Osaka Univer-
sity, Toyonaka, Osaka 560, Japan (e-mail: tamura@sys.es.osaka-u.ac.jp). in pairwise comparisons by a required majority
Publisher Item Identifier S 1083-4419(98)00218-0. such that for at
1083–4419/98$10.00  1998 IEEE
206 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 28, NO. 2, APRIL 1998

least individual decision makers . This concept is extended degree computation is then revised by using OWA aggregation
into fuzzy -core by using Nurmi’s -degree concept [19]: operator [7]. A cost function is introduced in order to allow
such that for at the decision makers to revise their preference relations during
least individual decision makers . a consensus reaching process [17].
Whether an alternative defeats or not can be identified The consensus degree is derived in three steps:
by using a strict threshold valued model: i) derive a degree of agreement from each pair of indi-
if viduals as to their preferences between all the pair of
(2) alternatives;
otherwise
ii) aggregate the degree of agreement of each pair of
Then, it can be relaxed to degree by replacing its condition individuals;
into instead of . The degree to which iii) aggregate the degree of agreement which results from
individual supports alternative is step ii) to obtain a degree of agreement of most pairs
of important individuals on most pairs of relevant alter-
(3) natives.
Again, the threshold valued model is used in the computational
procedures.
Based on the result of (3), by using some equations, a
solution set can be derived. The details of these equations can
be found in [9], [13]. A linguistic majority concept is used to III. LINGUISTIC LABELS FOR EXPRESSING
approximate the final solutions. For example most is defined FUZZY PREFERENCE RELATIONS
as follows: In Section II, we have summarized the pairwise group
for decision analysis based on fuzzy preference relations. Each
for (4) decision maker has to express his or her preference relations
for directly into numerical values, . This representation
is impractical to decision makers. They are used to represent
The final solution set is their evaluations linguistically.
We propose to extend the numerical representation by using
(5)
linguistic labels. Assume, is a variable which takes a value
that is a fuzzy set of alternatives which are not defeated from a base set , where is a linguistic
by (e.g., most) individuals. is the degree of support to label representing the preference relation. The decision makers
alternative by most decision makers. It can also be extended choose a value from the base set for representing their
into fuzzy / - core. preference evaluations of every pair of alternatives. Suppose
In the indirect approach, a solution set is derived in two for example, the base set consists of 13 linguistic labels,
stages: then the decision makers can express their fuzzy preference
i) derive , a group fuzzy preference relation, from indi- relations as follows:
viduals fuzzy preference relations;
ii) find the solution from .
if is preferred to in definite degree
One way to compute from is by using if is preferred to in very high degree
a threshold valued model [7], [11] as follows: if is preferred to in high degree
if if is preferred to in moderate degree
(6) if is preferred to in low degree
otherwise
if is preferred to in very low degree
where if is about the same as (indifferent to)
if if is preferred to in very low degree
(7) if is preferred to in low degree
otherwise
if is preferred to in moderate degree
Then, a consensus winner, or / winner can be used to derive if is preferred to in high degree
the solution set from . It also uses the threshold valued if is preferred to in very high degree
model. if is preferred to in definite degree
A degree of consensus computation model is proposed (8)
by Kacprzyk and Fedrizzi [11], and Kacprzyk [10], and
then advanced by Fedrizzi et al. [5], and Kacprzyk et al. For individual decision maker , the fuzzy preference relation
[12]. It is considered as a rigid measure. Full consensus is given by its linguistic label . The preference values
occurs only when all decision makers agree to all issues. are reflective-reciprocal with is the center. The evaluation
Kacprzyk et al. [13], improve it into soft-consensus by using values are actually consist of two groups: 1) prefer to
linguistic majority, in which full consensus can also be reached , and 2) prefer to , in which may belong to
when for example most decision makers agree on almost both groups. Since the two groups are similar and each
all relevant issues. The most function for the consensus group (plus ) consists of seven values, it is similar to a
MARIMIN et al.: LINGUISTIC LABELS FOR EXPRESSING FUZZY PREFERENCE RELATIONS 207

TABLE I
THE MAPPING OF THE FUZZY PREFERENCE RELATION VALUES

TABLE II
THE MAPPING OF THE LINGUISTIC LABELS INTO THE CORRESPONDING TFN
Fig. 1. Linguistic labels represented by using TFN.

as follows [14]:
i) Addition:
;
ii) Subtraction:
.
For multiplication and division operations, triplets cannot be
used directly. However the computation can be approximated
using the confidence interval at each level . Defining
the interval of confidence at level , TFN, say and can
be characterized as
;
.
seven points evaluation which is commonly used in the fuzzy
Then, the two values can be manipulated using interval ma-
decision making area [28]. The seven points evaluation is
nipulation formulas.
also suggested by Miller’s [18] observation that humans can
Suppose that there are two intervals and defined in
manage to keep in mind around seven items.
as and , respectively and there is a positive
The mapping of the preference relation into or vice
constant , then some operations are defined as follows [14]:
versa is shown in Table I. If alternative is preferred to
alternative , say, in very high degree , then alternative i) multiplication: ;
is dispreferred to alternative in very high degree ii) division: ;
by the decision maker . Accordingly, the decision maker only iii) scalar multiplication: ;
provides elements of the upper triangular part of the preference iv) power: , where in the right side
matrix. means power.
The linguistic label can be represented by three possible In the previous method, as summarized in Section II, the
ways: a fuzzy set, an interval, and a single value representa- fuzzy preferences are processed according to equations in
tions. The fuzzy sets representation is the most appropriate to which some of them are actually rigid, e.g., (2). When the
represent the linguistic labels [30]. There are several ways to values exceed a certain threshold, there is no difference
represent the fuzzy sets, one of them is by triangular fuzzy between a small difference and a big difference. Although it
numbers. In this analysis, we propose to use the triangular is relaxed to a degree , it is still quite rigid.
fuzzy number rather than using a trivial single value represen- In our method, the fuzzy preference relation matrices are
tation which is used in the previous model [4], [9], [13]. processed directly without converting into 0 or 1. This method
Each linguistic label is represented by a triangular fuzzy preserves the degree of preference accuracy during the compu-
number (TFN), say , which is defined as a triplets tation process. To aggregate the individual preference relations
, where . Without loss of generality, into a group preference relation, we use a neat OWA aggrega-
here, we deal with TFN represented in . Table II is one tion operator. It provides for aggregations lying between the
of possible transformation tables for converting the linguistic logical or and and.
labels into the corresponding TFN. Fig. 1 shows the visual An OWA operator [25], for with
form of the TFN. In that case are represented using associated weight , such that
special form of TFN in which . , is defined as a mapping: .
Suppose that there are two linguistic labels which are rep- And, ; where is the th largest of
resented by using the TFN defined on Table II as the . The values must be ordered to satisfy the symmetry
and , respectively, then some operations are defined (general commutativity) of the operator.
208 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 28, NO. 2, APRIL 1998

Since the linguistic labels are represented by using partially The benefits of using the operator for representing most is its
ordered TFN, the OWA operator cannot be used directly to intuitively meaningful solution [7], [11].
manipulate them. In order to handle this problem we use an For our case, the similar method can also be used to aggre-
extension of the OWA operator in which the aggregates do gate individuals support to alternative
not need to be ordered. The operator is called a neat OWA into a group support . Accordingly, we will use the OWA
operator. Moreover, by using the neat OWA operator, one may operator to represent most individuals (in the group)
be allowed to use weights that do not fixed, i.e., the weights support alternative as follows:
to be a function of the aggregates.
h (12)
A neat OWA operator [27], for with associated weight
, is defined as where is the neat OWA aggregation of
. The degree of satisfaction of the most
can be accommodated by the corresponding weight function,
(9) of (9).
The fuzzy -support to the alternatives is defined as a fuzzy
Thus, by using this kind of weight function, the weights are set
directly deduced from the values to be aggregated. When
we get a simple average operator. When approaches (13)
to infinity we get a maximum operator. This is a buoyancy or
that is a fuzzy set of alternatives which are supported by
orlike operator which can also be used for aggregating the
(e.g., most) individuals. These solutions on the TFN forms can
preference values [27], [29].
be converted back into the corresponding linguistic labels.
Using this operator, one can aggregate the preference re-
One of the possible way to transform the solutions back into
lation values which are more intuitively satisfied by the
the corresponding linguistic labels is a similarity method. The
decision makers. Moreover, this operator fulfills Arrow’s
corresponding linguistic label is the one with the most similar
conditions in aggregating preferences: unrestricted domain,
between the solution with each linguistic label which is
nonnegative response, independence of irrelevant alternatives,
defined in Table II. Similarity, can be represented as
citizen sovereign, and nondictatorship [1]. There has to be
the cardinality of the intersection between and divided by
no dictatorship among the decision makers and each decision
the cardinality of the label as follows:
maker is free to give his or her own preference relation
evaluations. Therefore, each decision maker has the same
(14)
participation in the group decision and also each alternative
has the same chance to be selected.
The corresponding linguistic label is the one with the highest
Computational models can be restated in the following
value. However, this equation cannot be used for
subsections.
label and , because these labels correspond to TFN
and , respectively, which have
A. The Direct Approach zero cardinality. In this case, when is or
A concept of fuzzy support, which is similar to that of the corresponding label is or respec-
the fuzzy core, is used to identify the alternatives which are tively.
supported by the decision makers. The fuzzy support value is
similar to the fuzzy preference value. B. The Indirect Approach
The degree to which individual supports alternative is A solution is derived in two stages:
i) a group fuzzy preference relation, , is derived from
(10) individuals fuzzy preference relations;
ii) the solution is derived from .
Unlike (3), which processes , the 0 or 1 converted values is derived from by using the following
from , (10) processes values directly. Therefore, it can equation:
preserve the preference accuracy and it also can differentiate if
between, say, a small difference and a big difference. (15)
otherwise
Then, the degree to which all the individuals support is
Unlike (6), which processes , the 0 or 1 converted values
(11) from , (15) processes values directly. Therefore, it can
preserve the preference accuracy.
Obviously, the aggregation process shown in (15), can also
Both (10) and (11) can also be applied for TFN. be done by using the neat OWA operator. Accordingly, (15)
Yager [25], [26] shows that the OWA operator can be used can be replaced by the following equation:
to represent aggregation guided by linguistic quantifier. This
method is also used to represent a linguistic quantifier most for if
(16)
consensus degree measure in a group decision making area [7]. otherwise
MARIMIN et al.: LINGUISTIC LABELS FOR EXPRESSING FUZZY PREFERENCE RELATIONS 209

where is the neat OWA aggregation of . Then, finally the consensus (the degree of agreement of
In that case, the mean operator is correspond to the neat OWA pairs of individuals of relevant pairs of alternatives) is
operator with the weight function’s parameter . (24)
The group fuzzy preference relation which is obtained either
from (15) or (16) can be processed further. where is the neat OWA aggregation of .
We calculate the average of the preferences to which alter-
native is preferred to all other alternatives as follows: D. Incorporating Fuzzy Criteria Into the Model
On each alternative there is a set of criteria
(17) to be considered. For example in an advertising media
selection, the alternatives, i.e., the advertising media, can be
television, radio, and newspapers. The corresponding criteria
and the degree of preferences to which alternative is can be the time and space availability of each medium, the
preferred to (e.g., most) other alternatives is targeted consumers, the area covered by each medium, and
(18) the type of programming in which the advertising is embed-
ded. The decision makers can be an advertising manager, a
where is the neat OWA aggregation of . product development manager, and a marketing manager in
Finally the fuzzy -support winner is defined as a fuzzy set a particular company. The criteria can be incorporated into a
of alternatives which are preferred to other alternatives (or three-dimensional evaluation matrix ; where is
which are supported by the decision makers) the number of alternatives, is the number of criteria, and
(19) is the number of decision makers. Each element of the
matrix represents the linguistic preference relation value of
By using the similarity method which is explained in the individual decision maker to each pair of alternatives with
Section III-A. the support value for each alternative can be respect to each criterion.
converted back into the corresponding linguistic label. Another method is that each decision maker may represent
his or her preference relation value based on each criterion
C. The Consensus Degree Measure into a matrix , which means that for each decision maker
provides matrices of dimension .
Based on the latest model on consensus degree measure
The criteria set can be classified into two cases: conflicting
[7], we propose to use the linguistic values directly. The
criteria exist or not exist in the system under consideration.
computations can be restated with the following equations.
For example, in industrial equipment selection, cost of pol-
The degree of agreement between individuals and as
lution device against environmental quality are the conflicting
to their preferences between alternatives and is
criteria. When the conflicting criteria exist, the evaluation may
if be based on a tradeoff function of the two conflicting criteria
(20)
otherwise rather than directly based on each criterion.
The matrix can be processed as follows.
Unlike in the original model, are taken from the linguis-
i) On each criterion, the preference relation matrices of
tic labels in preference relation matrices directly. Then, the
all decision makers are processed by using the same
next computational procedures are similar to that the original
algorithm as discussed in Section III-A–C. There will
model.
be a solution, say .
In order to follow Arrow’s conditions in aggregating the
ii) When the importance of the criteria are the same, a
preference relations, we consider the degree of importance of
solution set is then aggregated into
individuals is the same. We also consider that the relevance of
a total solution by using the neat OWA operator as
every pair of alternatives is the same. The degree of agreement
follows:
between individuals and as to their preferences between
all the relevant pairs of alternatives is (25)
where is the neat OWA aggregation of .
(21) In some cases, the decision makers may assign differing
degree of importance to each criterion. Assume for each
criterion has a degree of importance and its degree of
The degree of agreement between individuals and as
satisfaction is , then the values can be aggregated at
to their preferences between relevant pairs of alternatives is
by using one of the possible formula as follows [27]:
(22)
(26)
where is the neat OWA aggregation of .
The degree of agreement of all individuals is
where

(23)
and is reflective-reciprocal to .
210 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 28, NO. 2, APRIL 1998

TABLE III camera). For each product the decision makers evaluate the
THE MAPPING OF THE WEIGHTS OF THE CRITERIA INTO THE CORRESPONDING TFN alternatives based on each criterion.
For the first product, i.e., the telephone machines, a set of
fuzzy preference relation matrices of decision maker
based on criterion are as follows:

Fig. 2. Weights of the criteria represented by using TFN.

The other method is a fuzzy weighted average. It aggregates


the solutions on each criterion which may have difference
weights into a final solution as follows:

(27)

The weights of the criteria can be assigned by the group


of decision makers through a discussion based on the logical
nature of the criteria and consensus [20]. The weights are
stated linguistically by using seven-point values [18], [28].
Those are certainly not/none, very low, low, medium, high, very
high, and certain. Each weight value is represented by using
a TFN which is defined in Table III and visualized in Fig. 2.

IV. VERIFICATION
Suppose, in a case of fuzzy group decision making, there
are four alternatives, three criteria, and four decision makers.
In our case, i.e., an advertising media using television, the
alternatives are four television broadcasting companies which
offer advertising programs. In Osaka, Japan, the four broad-
casting companies are TV Osaka, Mainichi Television, Kansai
Television, and Asahi Television. The three criteria considered
are the time and space availability of each medium, the area
covered by each medium, and the type of programming in
which the advertising is embedded. The decision makers can
be an advertising manager, a product development manager, a
research and development manager, and a marketing manager
of a particular company.
The group of decision makers have to suggest the best tele-
vision media for advertising each of five electronic products Similarly the decision makers may assign the preference
(i.e., telephone machines, facsimile machines, mobile tele- evaluation matrices for the other four products. The data are
phone machines, television machines, and handy-cam video shown in the Appendix.
MARIMIN et al.: LINGUISTIC LABELS FOR EXPRESSING FUZZY PREFERENCE RELATIONS 211

The data are converted into the corresponding TFN (by


using Table II). The TFN data on each criterion are processed
to identify the alternative degree of preferences by using (10)
to (13) and (15) to (19) for the direct and indirect approaches,
respectively. By using (25) when the weights of criteria are the
same or by using (26) or (27) when the weight of criteria are
different, the solutions on each criterion are aggregated into
the final solutions. Then, by using (14), the final solutions are
converted back into the corresponding labels. The consensus
degree on each criterion is obtained by processing the original
data in labels form using (20) to (24). Similarly, the consensus
degree on all criteria is computed by using (25) or (26) or
(27). These computation models are mainly consist of matrix
of TFN arithmetic manipulations, aggregation functions by
using neat OWA operator and transformation function from
the TFN into the labels. These computations can be done in
polynomial time complexity which is still reasonably fast to
be implemented.
Because of a space limitation, the step by step computations
are not shown. The important results are shown on several
histograms and a table.
We use the neat OWA operator for aggregating the prefer-
ences in which they do not need to be ordered. The parameter
that corresponds to the weight function, i.e., is the decision
parameter. Figs. 3 and 4 show the solutions of product 1 by
using the direct and indirect approaches on several . As the
value is increased, the numbers of preferred alternatives is Fig. 3. Solution of the direct approach (for product 1) on several with the
also increase (see Fig. 4), the solutions’ ranks are comparable, same importance of the criteria.
however, some preference degrees of the alternatives are
different. We stop increasing the value when there is no
dispreferred alternative or when the solutions are the same
even if is increased by quite large value (the solution is
on a stable condition). In pairwise evaluations, when there
are preferred alternatives there should also be at least one
dispreferred alternative. In our case, when there
is no dispreferred alternative in the solution set, so that we
stop increasing the value. Based on these experiments, the
decision makers may select the intended value (in our case,
). The selection process may also be based on
the intended degree of orness of the operator [27] which is
different from a case to case. In our case, the decision makers
are satisfied when So, this value is used for the
following analysis.
Figs. 5 and 6 show the solutions of the direct and indirect
approaches for the five products in linguistic labels form. The
two approaches give slightly different rank of alternatives
and also slightly different degree of preferences. However,
the best alternative is always the same, i.e., alternative
4. Product 1 for instance, the direct approach produces
a solution set ,
while the indirect approach produces a solution set
.
For comparison purposes, the data are also processed by
using the previous single points method [9], [13]. For this
purpose, we ask the decision makers to provide preferences in
labels and also in numerical (single points) forms. The other
way is that the single point values can be approximated by the Fig. 4. Solution of the indirect approach (for product 1) on several with
midpoint of the corresponding TFN of the labels. Figs. 7 and the same importance of the criteria.
212 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 28, NO. 2, APRIL 1998

Fig. 5. Solution of the direct approach with the same importance of the
criteria.

Fig. 7. Solution of the direct approach by using single points method.

Fig. 6. Solution of the indirect approach with the same importance of the
criteria.

8 show the solutions of the direct and indirect approaches by


using the single points method. In this case, a mean operator
Fig. 8. Solution of the indirect approach by using single points method.
is used to aggregate solutions on each criterion into the final
solution based on all criteria.
Even if that corresponds to an average operator, single points method in term of the alternatives rank and also
solutions of the new method are slightly different to that the in their degree of preferences. In some cases, although the
MARIMIN et al.: LINGUISTIC LABELS FOR EXPRESSING FUZZY PREFERENCE RELATIONS 213

TABLE IV
THE DEGREE OF CONSENSUS DETERMINED BY
THE PREVIOUS AND THE PROPOSED METHOD

rank is the same the preferred alternatives are different, e.g.,


on the third product: by using the single points method with
indirect approach, alternative 2 is “slightly preferred” (with its
preference degree is 0.59), however by using the new method
alternative 2 degree of preference is about the same (AS).
In term of consensus degree, the new method relatively
gives lower value than the single points method (Table IV). It
is because in the new method, two preferences are considered
the same if they have the same values. However, for the single
points method, two preferences are considered the same if they
are both 0.5 or 0.5. So, the new method is more accurate
to represent the degree of consensus.
In some cases, the decision makers may assign differing
degree of importance (weight) to each criterion. The weights
of the criteria can be assigned by the group of decision makers
Fig. 9. Solution of the direct approach with different importance of the
through a discussion based on the logical nature of the criteria criteria.
and consensus [20].
In our case, the weight for the criteria 1, 2, and 3 are
medium, high, and very high, respectively. The results are
shown in Figs. 9 and 10. In term of the rank of alternatives,
Figs. 9 and 10 show that the results are comparable to that
the same weight of the criteria. However, in some alternatives
their degree of preferences are different.
Figs. 3–10 show that alternative 4 is always preferred and
its degree of preference is the highest, therefore it should
be selected as the best solution candidate. When the recom-
mended alternatives should have degree of preferences at least
(VLP) both by using the direct and indirect approaches, the
second solution candidate is alternative 2 for product 2, and
alternative 3 for product 1 and 5. There is no second alternative
recommended for product 3 and 4.

V. DISCUSSION
Fuzzy preference relations have been widely used in group
decision making [6], [7], [9], [13]. The preference relation
is represented using numerical value . The decision
makers have to provide certain values drawn from numerical
value in for representing their preference relations.
However, in real applications, it is difficult to assign the
preferences by using numerical (single point) values. The
decision makers are often notoriously unwilling to give precise
numerical estimates to the preference values. It becomes more
complicated when they are not familiar with fuzzy methods.
This seems to indicate that it is very difficult to achieve
any scheme which relies on the user providing consistent Fig. 10. Solution of the indirect approach with different importance of the
and precise numerical values [16]. They tend to express their criteria.
214 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 28, NO. 2, APRIL 1998

preferences qualitatively in the form of linguistic labels. The range is between 0.0 and 1.0. The decision makers may judge
linguistic labels representation method proposed in this article the results of the experiments and then decide the decision
will significantly improve simplicity and flexibility in the parameter accordingly based on their degree of satisfaction
applications of the group decision making based on fuzzy and the intended degree of orness of the operator. In our case,
preference relations. the decision makers are satisfied with . However, in
The previous model computation procedures use a strict the other case, it is possible that the second criteria is reached
threshold, for example when a preference relation value is first. When it is occurred, the suitable is the one which
less than or greater than the threshold value, a particular corresponds to the stable condition.
equation just simply produces 0 or 1. This rigid procedure In a particular case, an andlike operator may be consid-
has been relaxed to the -degree within a range around the ered. For instance, when it is required that there is exactly
threshold. This relaxed procedure is more flexible. However, one preferred alternative from a set of alternatives. For this
in some extent it is still rigid. In our method, the preference purpose, the neat OWA operator can also be used with different
relations, which are represented internally using TFN, are weight function which allows the operator behaves as an
processed directly without using the threshold. It preserves andlike operator. Some examples of the weight functions for
the preference accuracy during the computation processes and the andlike neat OWA operator are shown in [27]. In that case,
produces more intuitively meaningful solutions. Moreover, the when the neat OWA operator is similar to an average
core concept can be extended into a fuzzy support concept, operator, when approaches to infinity, it is similar to the
by allowing a fuzzy boundary between the solution (core) minimum operator. The value is increased from 0.0 until
and the possible alternatives. For example, a solution set there is no preferred alternatives or when the solutions are the
may consist of alternatives which are slightly dispreferred in same even if is increased by quite large value.
pairwise comparisons by a decision maker but they are highly Although the ordinary OWA operators treat the aggregates
preferred by the other decision makers as a whole. equally or give more weight to the highest opinion, in special
In some cases, the decision makers may have the same case the operator is also able to attach weight to a special
preferences to a particular alternative, e.g., highly preferred aggregates. In our case, we do not treat the decision mak-
but with different degrees. However, TFN treats the values ers opinions differently, each decision maker has the same
similarly. The representation can also be extended into fuzzy participation level. This situation is consistent to the Arrow’s
sets of Type 2 which allow to represent for example highly condition in aggregating the preference (undictatorship) [1].
preferred with different membership functions which lead However, we consider the solutions on each criterion differ-
to more complex analysis and computation which can be ently depending on the criterion’s degree of importance.
considered for the future work. In the previous model, the preference values given by the
For aggregating the preferences, the neat OWA aggregation decision makers are based on the overall criteria implicitly.
operator is used to replace the strict mean operator. The The decision makers have to consider all criteria thoroughly,
operator allows to aggregate preferences into a value lying then express their preference relations in a numerical value.
between and and or operators. As exposed by Yager [25], [27], Obviously, this evaluation procedure is difficult. In many
this operator is more appropriate, since the type of aggregation cases, the preferences can be different on each criterion.
implicitly desired by decision makers is neither the pure and Fuzzy criteria have been incorporated explicitly to the
nor or operator. new model. It allows the decision makers to express their
Unlike the ordinary OWA operator in which the values preference relations on each criterion. It gives results both
must be ordered, the neat OWA operator does not need to the solutions on each criterion and on all criteria. When
do so. Therefore, the neat OWA operator is also suitable there are conflicting criteria to be considered, the evaluation
for aggregating the linguistic labels represented by partially becomes more complex. In that case, the evaluation should be
ordered TFN. Accordingly, it is applied to the proposed model based on a tradeoff function between the conflicting criteria.
thoroughly. Determining the suitable tradeoff function is different from a
In this operator, one is allowed to assign weight which is case to case that could be considered for further research.
not fix, but it is as the function of the values to be aggregated. To aggregate solutions on each criterion into a final solution,
The parameter that corresponds to the weight function, i.e., there are two choices. First, the criteria have the same weights.
is a decision parameter. It can be determined based on In that case, the neat OWA operator is used. Second, the
experiments’ results and the intended degree of orness or criteria have different weights, then a special case of the neat
andness of the operator [27]. In our case, we use the weight OWA operator that is able to attach difference importance level
function that represents orlike operator, since in the pairwise to the values is used. The other choice is by using the fuzzy-
fuzzy group decision making the decision makers tend to give weighted average method. In our case, when with the
a ranking of the preferred alternatives rather than to give similar weights, the solutions of the neat OWA operator with
ranking of the dispreferred alternatives [13], [28]. different weights are comparable to that the fuzzy weighted
In the experiments, it can be tried from 0.0 (simple average method.
average) and it is increased by , say 0.1 until there is no A number of other methods for aggregating preferences
dispreferred alternative or when the solutions are the same in group decision making have been suggested, for example
even if the is increased by quite large value. In our case, methods proposed by Lehrer and Wagner [15], [24], and
when is the first criteria is reached, so that the by Saaty [21], [22]. In the Lehrer-Wagner method, each
MARIMIN et al.: LINGUISTIC LABELS FOR EXPRESSING FUZZY PREFERENCE RELATIONS 215

decision maker gives numerical preference to every alternative be extended into fuzzy sets of Type 2 which allow to rep-
and assigns numerical weights to the other decision makers resent for example two values: highly preferred with different
participated in the processes. The values are processed by membership functions.
using weighted average aggregation formulas. Saaty’s method Fuzzy criteria have been incorporated explicitly to the
of analytic hierarchy process (AHP) is capable of determining new model. It allows the decision makers to express their
the rank of alternatives in a hierarchical structure. However, preference relations based on each criterion in a single-level
the basic method is actually dealing with a single decision structure. It gives results for on each criterion and on all
maker not a group decision makers. It simply aggregates the criteria. The criteria may have the same or different weights.
individual data by using arithmetic or geometric mean, then For further work, the new method can be extended for solving
the results are processed further by using the same model as multilevel structure problems.
in the decision making by one decision maker. The proposed method has been applied to some sample data
Our proposed method uses pairwise fuzzy preference eval- sets of a fuzzy group decision making case. It allows for
uations and it is directly intended for solving group decision the use of fuzzy members. Its results are slightly different
making problem. The preferences as well as the weights to that the previous single points method in term of the
are expressed nonnumerically using linguistic labels. This rank of alternatives and also in the corresponding degree
expression is more practical, i.e., the decision maker does not of preferences. The new method preserves the preference
need to be as precise and can express his/her opinion in a more accuracy during computational processes and produces more
natural manner related to the alternative at hand. However it intuitively meaningful solutions. Applications to the other real
deals with single-level structure problem. For further work, fuzzy group decision making cases could be done in the future
our proposed method can be extended for solving multilevel works.
structure problems which is similar to that the AHP method.
Then, more in depth comparison between the two methods APPENDIX
can be done. THE DATA USED FOR VERIFICATION
Besides these models, for aggregating values with different
weights Baldwin’s approach [2], [3] can also be used with Case-2: facsimile products
some modifications to our model. The use of the Baldwin’s
model, its affects to the results and also comparison among
several models can be considered for the future work.
The proposed method has been applied to some sample
data sets of a fuzzy group decision making case, i.e., the
advertising media selection. It allows for the use of fuzzy
members. Its results are slightly different to that the previous
single points method in term of the rank of alternatives and
also in the corresponding degree of preferences. It preserves
the preference accuracy during computational processes and
produce more intuitively meaningful solutions. The proposed
method can be used as one of the possible basis for developing
group decision support systems. Applications of the model to
other real fuzzy group decision making cases could be done
in the future work.

VI. CONCLUDING REMARKS AND FUTURE WORKS


We have extended the group decision analysis based on
fuzzy preference relations by allowing the decision makers
to express their preference relations in linguistic labels rather
than in numerical values. In the real applications, this method
is more flexible and practical.
The strict threshold valued computational models have been
modified into fuzzy computational models by using TFN
representation and computation, and by avoiding the use of
strict threshold. The proposed model uses neat OWA operators
which allow to aggregate unordered preference relations, and
for manipulating the linguistic quantifier such as most.
In some cases, the decision makers may have the same
preferences to a particular alternative, e.g., highly preferred
but with different degrees. However, TFN treats the values
similarly. For the future work, the representation can also
216 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 28, NO. 2, APRIL 1998

Case-4: television products

Case-3: mobile telephone products


MARIMIN et al.: LINGUISTIC LABELS FOR EXPRESSING FUZZY PREFERENCE RELATIONS 217

REFERENCES
Case-5: handy-cam video camera products
[1] K. J. Arrow, Social Choice and Individual Values, 2nd ed. New York:
Wiley, 1963.
[2] J. F. Baldwin, “Evidential support logic, FRIL and case based reason-
ing,” Int. J. Intell. Syst., vol. 8, pp. 939–960, 1993.
[3] , “FRIL methods for soft computing, fuzzy control, and classifi-
cation,” in Proc. 4th IEEE Int. Conf. Fuzzy Systems, Yokohama, Japan,
1995, pp. 309–316.
[4] J. C. Bezdek, B. Spillman, and R. Spillman, “Fuzzy relation spaces
for group decision theory: An application,” Fuzzy Sets Syst., vol. 4, pp.
5–14, 1979.
[5] M. Fedrizzi and J. Kacprzyk, “On measuring consensus in the setting of
fuzzy preference relations,” in Non-Conventional Preference Relations
in Decision Making, K. Kacprzyk and M. Roubens, Eds. New York:
Springer-Verlag, 1988, pp. 129–141.
[6] M. Fedrizzi, J. Kacprzyk, and S. Zadrozny, “An interactive multi-user
decision support system for consensus reaching process using fuzzy
logic with linguistic quantifiers,” Decision Support Syst., vol. 4, pp.
129–141, 1988.
[7] M. Fedrizzi and J. Kacprzyk, “Consensus degrees under fuzzy majorities
and preferences using ordered weighted average operators,” in Proc. 5th
IFSA World Congr., Seoul, Korea, 1993, pp. 624–626.
[8] M. Fedrizzi and R. A. Pereira, “Consensual dynamics: An unsuper-
vised learning model in group decision making,” in Proc. Int. Conf.
Fuzzy Logic, Neural Nets Soft Computing, Iizuka, Japan, 1994, pp.
99–100.
[9] J. Kacprzyk, “Group decision making with a fuzzy linguistic majority,”
Fuzzy Sets Syst., vol. 11, pp. 105–118, 1986.
[10] J. Kacprzyk, “On some fuzzy core and ‘soft’ consensus measures in
group decision making,” in The Analysis of Fuzzy Information, J. C.
Bezdek, Ed. Boca Raton, FL: CRC, 1987, vol. 2, pp. 119–130.
[11] J. Kacprzyk and M. Fedrizzi, “Soft consensus measures for monitoring
real consensus reaching processes under fuzzy preferences,” Contr.
Cybern., vol. 15, pp. 309–323, 1986.
[12] Multiperson Decision Making Models Using Fuzzy Sets and Possibility
Theory, J. Kacprzyk and M. Fedrizzi, Eds. Dordrecht, The Nether-
lands: Kluwer, 1990.
[13] J. Kacprzyk, M. Fedrizzi, and H. Nurmi, “Group decision making and
consensus under fuzzy preferences and fuzzy majority,” Fuzzy Sets Syst.,
vol. 49, pp. 21–31, 1992.
[14] A. Kaufmann and M. M. Gupta, Fuzzy Mathematical Models in En-
gineering and Management Science. Amsterdam, The Netherlands:
Elsevier, 1988.
[15] K. Lehrer and C. Wagner, Rational Consensus in Science and Society.
Boston, MA: Reidel, 1981.
[16] M. Marimin, “Developing and evaluating a reasoner that reasons about
discrete stochastic simulation design parameters,” M.Sc. thesis, Univ.
Western Ontario, London, Ont., Canada, 1990.
[17] L. Mich, L. Gaio, and M. Fedrizzi, “On fuzzy logic-based consensus in
group decision,” in Proc. 5th IFSA World Congr., Seoul, Korea, 1993,
pp. 698–700.
[18] G. A. Miller, “The organization of lexical memory,” in The Pathology of
Memory, G. A. Talland and N. C. Waugh, Eds. New York: Academic,
1969.
[19] H. Nurmi, “Approaches to collective decision making with fuzzy
preference relations,” Fuzzy Sets Syst., vol. 6, pp. 249–259, 1981.
[20] K. J. Schmucker, Fuzzy Sets, Natural Language Computations, and Risk
Analysis, Rockville, MD: Computer Science, 1984.
[21] T. L. Saaty, The Analytic Hierarchy Process. New York: McGraw-Hill,
1980.
[22] , “Axiomatic foundation of the analytic hierarchy process,” Man-
age. Sci., vol. 32, pp. 841–855, 1986.
[23] H. J. Skala, “Arrow’s impossibility theorem: Some new aspects,” in
Decision Theory and Social Ethics, H. W. Gottinger and W. Leinfellner,
Eds. Dordrecht, The Netherlands: Reidel, 1978, pp. 215–225.
218 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART B: CYBERNETICS, VOL. 28, NO. 2, APRIL 1998

[24] C. Wagner, “On the formal properties of weighted averaging as a method Itsuo Hatono received the B.S. and M.S. degrees in
of aggregation,” Synthese, vol. 62, pp. 97–108, 1985. engineering from Osaka University, Osaka, Japan,
[25] R. R. Yager, “On ordered weighted averaging aggregation operators in in 1984 and 1986, respectively, and the Ph.D. degree
multi-criteria decision making,” IEEE Trans. Syst., Man, Cybern., vol. in engineering from Osaka University in 1994.
18, pp. 183–190, 1988. From 1986 to 1988, he was a Research Engineer
[26] , “Applications and extensions of OWA operators,” Int. J. Man- with C&C Systems Research Laboratories, NEC
Mach. Stud., vol. 37, pp. 103–132, 1992. Corp., Kawasaki, Japan. From 1988 to 1993, he
[27] , “Families of OWA operators,” Fuzzy Sets Syst., vol. 59, pp. was a Research Associate with the Department
125–148, 1993. of Precision Engineering, Faculty of Engineering,
[28] , “Non-numeric multi-criteria multi-person decision making,” Osaka University. From 1993 to 1996, he was a
Group Decision Negot., vol. 2, pp. 81–93, 1993. Research Associate with the Department of Systems
[29] R. R. Yager and D. P. Filev, “Parameterized andlike and orlike OWA Engineering, Faculty of Engineering Science, Osaka University. Since 1996,
operators,” Int. J. Gen. Syst., vol. 22, pp. 297–316, 1994. he has been an Assistant Professor with the Department of Systems Engineer-
[30] L. A. Zadeh, “The concept of a linguistic variable and its application to ing, Faculty of Engineering Science, Osaka University. His research interests
approximate reasoning,” Inf. Sci., vol. 8, pp. 301–357, 1975. include production scheduling for complex and large production systems,
and modeling of discrete event systems such as manufacturing systems and
computer networks.
Dr. Hatono is a member of Information Processing Society of Japan, and
Marimin was born in Sukoharjo, Central Java, the Institute of Systems, Control and Information Engineers of Japan.
Indonesia, in 1961. In August 1984, he received the
B.S. degree (Sarjana/Ir.) in agroindustrial technol-
ogy, from the Faculty of Agricultural Technology,
Bogor Agricultural University, Bogor, Indonesia.
In January 1990, he received the M.Sc. degree in Hiroyuki Tamura (M’69) was born in Nishi-
computer science from the University of Western nomiya, Hyogo, Japan, on March 9, 1940. He
Ontario, London, Ont., Canada. From April 1994 received the B.S. and M.S. degrees in engineering
to March 1995, he was a Research Student with from Osaka University, Osaka, Japan, in 1962 and
the Department of System Engineering, Faculty 1964, respectively, the M.S. degree in engineering
of Engineering Science, Osaka University, Osaka, economic systems from Stanford University,
Japan. He received the Ph.D. degree in engineering from Osaka University Stanford, CA, in 1968, and the Ph.D. degree in
in 1997; his dissertation was titled “Linguistic labels based methodology for engineering from Osaka University in 1971.
fuzzy group decision making.” From 1964 to 1971, he was a Research Engineer
From 1985 to 1987, he was an Assistant Researcher and Assistant Lecturer with the Central Research Laboratory, Mitsubishi
at Bogor Agricultural University. Since 1990, he has been a Lecturer and Electric Corp., Amagasaki, Japan. During this
Researcher with the Department of Agroindustrial Technology, Bogor Agri- period he received the Stanford Graduate Fellowship and spent 1966
cultural University. Where he has done research on applications of expert to 1968 at the Department of Engineering-Economic Systems, Stanford
systems in agroindustry, management technology, decision support systems, University. He held summer employment with Stanford Research Institute
and management information systems. His current research interests are fuzzy (now SRI International), Menlo Park, CA, in 1967. From 1971 to 1987,
systems and (fuzzy) expert systems applied in the (group) decision making he was an Associate Professor and from 1987 to 1993, he was a Professor
area. with the Department of Precision Engineering, Faculty of Engineering,
Marimin is a member of the Indonesian Engineer Association (PII), the Osaka University. During this period he spent 1972 to 1973 at the Control
Indonesian Industrial Engineering and Management Association, the Institute and Systems Group, Department of Engineering, Cambridge University,
of Systems, Control and Information Engineers of Japan (iSCIe), and the Cambridge, U.K., as a Visiting Researcher and received a British Council
Japan Society for Fuzzy Theory and Systems (SOFT). Scholarship. He was also a Visiting Researcher at the Politecnico di Milano,
Milano, Italy, and at the International Institute for Applied Systems Analysis
(IIASA), Laxenburg, Austria, from March to June 1976. He was invited to
give a series of lectures on Water Quality Systems Analysis, Control and
Management at the Northeastern Normal University, Changchun, and at the
Motohide Umano (A’94) received the B.S. degree Beijing Normal University, Beijing, China, in July 1981. Since 1993, he
in 1974, the M.S. degree in 1976, and the Dr.Eng. has been a Professor with the Department of Systems Engineering, Faculty
degree in 1979, all in information and computer of Engineering Science, Osaka University. His research interests center on
sciences, from Osaka University, Osaka, Japan. the systems methodology for large-scale systems such as modeling, control
He was an Assistant Professor of Applied Math- and decision making, and their applications to manufacturing systems and
ematics, Faculty of Science, Okayama University public systems. He has written more than 70 research papers in this field.
of Science, Okayama, Japan, from 1979 to 1985, He is a coauthor of the book Modeling and Control of River Quality (New
then moved to Osaka University, where he was a York: McGraw-Hill, 1979), and Large-Scale Systems Control and Decision
Research Associate, an Assistant Professor, and an Making (New York: Marcel Dekker, 1990).
Associate Professor of Computation Center from Dr. Tamura was awarded the Paper Prizes from the Society of the Instrument
1985 to 1990; an Associate Professor of Department and Control Engineers (SICE) of Japan in 1976, and from the Institute of
of Precision Engineering, Faculty of Engineering, from 1991 to 1993; and an Systems, Control and Information Engineers (ISCIE) of Japan in 1990. He
Associate Professor of Systems Engineering, Faculty of Engineering Science, is a member of ORSA, SRA, SICE of Japan, ISCIE of Japan, etc. He was a
from 1993 to 1996. He is currently Professor of Mathematics and Information member of the Board of Governers of ISCIE from 1988 to 1992, and was a
Sciences, College of Integrated Art and Sciences, Osaka Prefecture Univer- member of the Board of Governers of SICE from 1994 to 1996. He was a
sity. His current research interests includes fuzzy-set-oriented programming Vice-Chairman of the Program Committee of 1992 Japan–U.S.A. Symposium
languages, fuzzy databases, fuzzy expert systems, and fuzzy knowledge on Flexible Automation, San Francisco, July 1992. He was a General Vice-
information processing. co-chair of the 1st Asian Control Conference (ASCC), Tokyo, July 1994,
Dr. Umano is a member of the International Fuzzy Systems Association and is presently a Steering Committee Member of ASCC. He is presently
(IFSA), the Association for Computing Machinery (ACM), and several IPC Chairman of APMS’96 organized by IFIP WG 5.7. He is presently the
Japanese societies on computer science and systems science and engineering. Chairman of the Kansai Branch of the Operations Research Society of Japan.

You might also like