Professional Documents
Culture Documents
R
1
,
R
2
, ,
R
n
).
Before the description of our method, Hsus aggregation method (SAM) is briey described.
In Hsu and Chens article, they calculated the average agreement degree of each expert
E
i
(i = 1, 2, , n) by averaging the degrees of similarity with respect to other experts:
A(E
i
) =
1
n 1
n
j=1,j=i
S(
R
i
,
R
j
), (1)
where
S(
R
i
,
R
j
) =
_
x
(min{
Ri
(x),
Rj
(x)})dx
_
x
(max{
Ri
(x),
Rj
(x)})dx
. (2)
It is a similarity measure function by Zwick et al.
[10]
, which means the proportion of the
consistent area (
_
x
(min{
Ri
(x),
Rj
(x)})dx) to the total area (
_
x
(max{
Ri
(x),
Rj
(x)})dx).
AGGREGATION OF FUZZY OPINIONS UNDER GROUP DECISION-MAKING 65
Then the aggregation weight of the ith expert is given by
RAD
i
=
A(E
i
)
n
j=1
A(E
j
)
. (3)
Without considering the importance of the ith expert, the aggregation result is therefore dened
as
R = F(
R
1
,
R
2
, ,
R
n
) =
n
i=1
RAD
i
R
i
, (4)
where
A,
B) = S(
A,
C) =
3
4
. However, one can see from Fig. 1 that these two similarities
should not be equal because the supports of the intersection of
A and
B are larger than the
supports of the intersection of
A and
C. To avoid this situation we need to consider the weight
66 CHENGGUO LU JIBIN LAN ZHONGXING WANG
of each element x R so that we have the following weighted similarity measures. For fuzzy
numbers
R
i
and
R
j
, assume the weight of each x R is w(x), 0 w(x) 1. Then the weighted
similarity between
R
i
and
R
j
is dened as
S
w
(
R
i
,
R
j
) =
_
x
(w
min
(x) min{
Ri
(x),
Rj
(x)})dx
_
x
(w
max
(x) max{
Ri
(x),
Rj
(x)})dx
, (5)
where w
min
(x) and w
max
(x) denote the weight functions of the consistent area and the total
area, respectively. In general, let w
min
(x) = min{
Ri
(x),
Rj
(x)}, w
max
(x) = max{
Ri
(x),
Rj
(x)} because min{
Ri
(x),
Rj
(x)} and max{
Ri
(x),
Rj
(x)} are membership functions of
the consistent area and total area, respectively. Thus the improved similarity is given as
S
w
(
R
i
,
R
j
) =
_
x
(min{
Ri
(x),
Rj
(x)})
2
dx
_
x
(max{
Ri
(x),
Rj
(x)})
2
dx
. (6)
It is easy to prove that S
w
(
R
i
,
R
j
) satises the following properties:
(SP1) 0 S
w
(
R
i
,
R
j
) 1;
(SP2) S
w
(
R
i
,
R
j
) = 1 if and only if
R
i
=
R
j
;
(SP3) S
w
(
R
i
,
R
j
) = S
w
(
R
j
,
R
i
);
(SP4) S
w
(
R
i
,
R
j
) S
w
(
R
i
,
R
k
) and S
w
(
R
j
,
R
k
) S
w
(
R
i
,
R
k
) if
R
i
R
j
R
k
.
Let us again consider Example 3.1. Now using our similarity measure Eq. (6) we obtain
S
w
(
A,
B) =
4
5
S
w
(
A,
C) =
3
5
.
The result is fair and reasonable.
3.2 Distance Between Fuzzy Numbers
Distance is an important concept in fuzzy set theory and is also a signicant index of the
comparison of fuzzy numbers. Many distance measure methods have been proposed up to now,
a survey of such distances can be found in papers of Dubois et al.
[12]
, and Heilpern
[13]
. We
synthesize the main characteristics of these methods and classify them as two categories as:
The rst category takes no consideration of the membership function of fuzzy numbers,
such as the distance used by Tong
[14]
and Hausedor metric. The advantages of these methods
are the convenience of operating and ease of understanding. However, it will lead to the loss of
information for the negligence of the supports of fuzzy numbers.
The second category is characteristic of taking the membership function into account. In
these methods, the Hamming metric and the Euclidean metric are most often used. For any
two fuzzy numbers
A and
B with membership functions
A
and
B
, respectively, we have (see
[15]) the Hamming distance d
H
(
A,
B)
d
H
(
A,
B) =
_
x
|
A
(x)
B
(x)|dx (7)
and the Euclidean distance d
E
(
A,
B)
d
E
(
A,
B) =
_
x
(
A
(x)
B
(x))
2
dx. (8)
These formulas are straightforward generalizations of distances with the membership functions.
However, these methods do not always operate. In the case that the intersection between two
fuzzy numbers is empty, no conclusion will be drawn. Let us see the following example.
AGGREGATION OF FUZZY OPINIONS UNDER GROUP DECISION-MAKING 67
Example 3.2 There are three fuzzy numbers
A = (1, 2.5, 3.5, 4),
B = (5, 6, 7, 8),
C =
(9, 10, 11, 12). Using the Hamming distance Eq. (7), the distance between
A and
B and the
distance between
A and
C are calculated as follows:
d
H
(
A,
B) = S
A
+S
B
, d
H
(
A,
C) = S
A
+S
C
,
where S
A
, S
B
, and S
C
imply the areas of
A,
B, and
C, respectively. Since S
B
= S
C
, it follows:
d
H
(
A,
B) = d
H
(
A,
C).
By analogy, we have d
E
(
A,
B) = d
E
(
A,
C).
But it is obvious that the distance between
A and
B should be smaller than the distance
between
A and
C. To overcome this disadvantage, we introduce a distance between two sets A
and B
d
inf
(A, B) = inf{d(a, b), a A, b B}, (9)
where d is the usual metric. For trapezoidal fuzzy numbers
A = (a
1
, a
2
, a
3
, a
4
) and
B =
(b
1
, b
2
, b
3
, b
4
),
d
inf
(
A,
B) = inf{d(a, b), a [a
1
, a
4
], b [b
1
, b
4
]}. (10)
Considering the Hamming distance is a linear distance and is easy to operate, now we dene
a new distance measure between fuzzy numbers
A and
B on the basis of the Hamming distance
d
H
(
A,
B) and the distance d
inf
(
A,
B) as follows:
D(
A,
B) =
1
2
_
d
H
(
A,
B) +d
inf
(
A,
B)
_
=
1
2
__
x
|
A
(x)
B
(x)|dx +d
inf
(
A,
B)
_
. (11)
Note that this distance can be regarded as the Hamming distance when d
inf
(
A,
B) = 0, but
it does overcome the shortcoming of the Hamming distance. To illustrate the eciency of this
distance method, now consider Example 3.2 again. By our distance measure Eq. (11), we have
D(
A,
B) =
1
2
(S
A
+S
B
+ 1), D(
A,
C) =
1
2
(S
A
+ S
C
+ 5) > D(
A,
B).
This coincides with our intuition.
For the fuzzy numbers
R
i
=(a
i
, b
i
, c
i
, d
i
)(i = 1, 2, , n) of each experts opinion, we cal-
culate the distance d(
R
i
,
R
j
) between each pair of
R
i
and
R
j
, then select the largest distance
D
R
p
,
R
q
) = max
i,j
D(
R
i
,
R
j
). We divide each distance by D
R
p
,
R
q
) so that we obtain the
following normalized distance:
d(
R
i
,
R
j
) =
D(
R
i
,
R
j
)
D
R
p
,
R
q
)
. (12)
This distance measure fullls the following properties:
(DP1) 0 d(
R
i
,
R
j
) 1;
(DP2) d(
R
i
,
R
j
) = 0 if and only if
R
i
=
R
j
;
(DP3) d(
R
i
,
R
j
) = d(
R
j
,
R
i
);
(DP4) d
R
i
,
R
j
) d(
R
i
,
R
k
) and d(
R
j
,
R
k
) d(
R
i
,
R
k
) if
R
i
R
j
R
k
.
3.3 Aggregation Method
Let
R
1
,
R
2
, ,
R
n
be nfuzzy numbers representing each experts opinion for an alternative
under a given criterion. Since we have obtained the similarity S
w
(
R
i
,
R
j
) and the distance
68 CHENGGUO LU JIBIN LAN ZHONGXING WANG
d(
R
i
,
R
j
) for i, j = 1, 2, , n, now we dene a new consistency measure between fuzzy opinions
R
i
and
R
j
:
r(
R
i
,
R
j
) = S
w
(
R
i
,
R
j
) + (1 )(1 d(
R
i
,
R
j
)), (13)
where [0, 1] is the weight of S
w
(
R
i
,
R
j
), which reects the relative importance degree
between the similarity and the distance with respect to the decision maker. 1 is the weight
of d(
R
i
,
R
j
). The basic idea of this aggregation method is that the larger S
w
(
R
i
,
R
j
) and the
smaller d(
R
i
,
R
j
), the larger consistency degree r(
R
i
,
R
j
) between fuzzy opinions
R
i
and
R
j
.
Since r(
R
i
,
R
j
) is the linear combination of S
w
(
R
i
,
R
j
) and d(
R
i
,
R
j
), it follows
(RP1) 0 r(
R
i
,
R
j
) 1;
(RP2) r(
R
i
,
R
j
) = 1 if and only if
R
i
=
R
j
;
(RP3) r(
R
i
,
R
j
) = r(
R
j
,
R
i
);
(RP4) r
R
i
,
R
j
) r(
R
i
,
R
k
) and r(
R
j
,
R
k
) r(
R
i
,
R
k
) if
R
i
R
j
R
k
.
In practice, group decision-making is highly inuenced by the degrees of the importance
of participants. For example, there are some experts such as the managers of a company
with authority and some experts who are more experienced than the others. So an eective
aggregation method should consider the relative importance weight of each expert. For no loss
of generality, let the degree of importance of ith expert be e
i
(0 e
i
1), and
n
i=1
e
i
= 1. (14)
Then the weighted consistency degree of each expert E
i
is given as
C(E
i
) =
n
j=1
r(
R
i
,
R
j
)e
j
. (15)
Thus the aggregation weight of each expert E
i
is calculated by
w(E
i
) =
C(E
i
)
n
j=1
C(E
j
)
. (16)
The aggregation result is therefore dened as
R = F(
R
1
,
R
2
, ,
R
n
) =
n
i=1
w(E
i
)
R
i
, (17)
where
R
i
,
R
j
) between each pair of experts by Eq. (6).
Step 3: Calculate the distance d(
R
i
,
R
j
) between each pair of experts by Eq. (12).
Step 4: Given [0, 1], calculate the consistency degree r(
R
i
,
R
j
) between each pair of experts
by Eq. (13).
Step 5: Select the degree of importance e
i
of each expert E
i
, then calculate the weighted con-
sistency degree C(E
i
) of each expert E
i
by Eq. (15).
AGGREGATION OF FUZZY OPINIONS UNDER GROUP DECISION-MAKING 69
Step 6: Calculate the aggregation weight w(E
i
) of expert by Eq. (16).
Step 7: Aggregate each fuzzy opinion into a group fuzzy opinion by Eq. (17).
Our consistency aggregation method (CAM) also preserves some important properties as
follows.
Property 3.1 (Agreement preservation
[1]
) If
R
i
=
R
j
for all i, j, then
R =
R
j
. That is,
if all opinions of experts are identical the aggregation result is the common opinion.
Property 3.2 (Order independence
[1]
) The result of CAM would not be inuenced by the
dierent order with which individual opinions are combined. That is, if {(1), (2), , (n)} is a
permutation of {1, 2, , n}, then
R = F(
R
1
,
R
2
, ,
R
n
) = F(
R
(1)
,
R
(2)
, ,
R
(n)
).
Property 3.1 and Property 3.2 are consistency equipments.
Property 3.3 Let the uncertainty measure H(
R
i
) of individual opinion
R
i
be dened as
the area under its membership function
[1]
H(
R
i
) =
_
+
Ri
(x)dx, (18)
then the uncertainty measure dened in Eq. (18) satises the following equation:
H(
R) =
n
i=1
w(E
i
) H(
R
i
). (19)
This means that the uncertainty of aggregation result is between the uncertainties of all
experts, i.e., min
i
H(
R
i
) H(
R) max
i
H(
R
i
). Referring to Example 4.1, H(
R
1
) = 2, H(
R
2
) =
2.25, H(
R
3
) = 2.75, so H(
R) = 2.322 is between H(
R
1
) and H(
R
3
).
Property 3.4
[8]
The common intersection of all experts opinions is concluded in the nal
aggregation result. That is,
n
i=1
R
i
R.
Proof Let cut of
R
i
be
R
i
= [a
i
, b
i
], then
n
i=1
i
= [a
, b
], where
a
= max
i
a
i
, b
= min
i
b
i
.
Since
R =
n
i=1
w(E
i
)
R
i
,
it follows
=
n
i=1
w(E
i
)
i
=
_
n
i=1
(w(E
i
) a
i
),
n
i=1
(w(E
i
) b
i
)
_
.
Notice
n
i=1
(w(E
i
) a
i
) a
= max
i
a
i
,
n
i=1
(w(E
i
) b
i
) b
= min
i
b
i
.
We have proved this property.
Property 3.5 If
n
i=1
R
i
= , the consistent group opinion
R also can be derived.
When the opinions are disjoint, we can measure the distance d(
R
i
,
R
j
) (for all i, j) among
them, then calculate the consistency degree r(
R
i
,
R
j
) between each pair of experts by Eq. (13)
so that aggregation process can be continued.
70 CHENGGUO LU JIBIN LAN ZHONGXING WANG
4 Numerical Example
Example 4.1
[8]
Consider a group decision-making problem with three experts. The
opinions of each expert are given as three positive trapezoidal fuzzy numbers (see Fig. 2):
R
1
= (1, 2, 3, 4),
R
2
= (1.5, 2.5, 3.5, 5),
R
3
= (2, 2.5, 4, 6).
Figure 2 Aggregation result of fuzzy opinions
R1,
R2, and
R3
Now use our consistency aggregation method (CAM) to study this problem.
Step 1: It is given above.
Step 2: Calculate the similarity S
w
(
R
i
,
R
j
) for i, j = 1, 2, 3 as follows:
S
w
(
R
1
,
R
1
) = 1, S
w
(
R
1
,
R
2
) =
1
2
, S
w
(
R
1
,
R
3
) =
1
3
;
S
w
(
R
2
,
R
1
) =
1
2
, S
w
(
R
2
,
R
2
) = 1, S
w
(
R
2
,
R
3
) =
2
3
;
S
w
(
R
3
,
R
1
) =
1
3
, S
w
(
R
3
,
R
2
) =
2
3
, S
w
(
R
3
,
R
3
) = 1.
Step 3: Calculate the distance d(
R
i
,
R
j
) for i, j = 1, 2, 3 as follows:
d(
R
1
,
R
1
) = 0, d(
R
1
,
R
2
) =
5
9
, d(
R
1
,
R
3
) = 1;
d(
R
2
,
R
1
) =
5
9
, d(
R
2
,
R
2
) = 0, d(
R
2
,
R
3
) =
4
9
;
d(
R
3
,
R
1
) = 1, d(
R
3
,
R
2
) =
4
9
, d(
R
3
,
R
3
) = 0.
Step 4: Select =
1
2
, calculate the consistency degree r(
R
i
,
R
j
) for i, j = 1, 2, 3 as follows:
r(
R
1
,
R
2
) = 1, r(
R
1
,
R
2
) = 0.472, r(
R
1
,
R
3
) = 0.167;
r(
R
2
,
R
1
) = 0.472, r(
R
2
,
R
2
) = 1, r(
R
2
,
R
3
) = 0.611;
r(
R
3
,
R
1
) = 0.167, r(
R
3
,
R
2
) = 0.611, r(
R
3
,
R
3
) = 1.
Step 5: Let the degrees of importance of three experts be e
1
= 0.42, e
2
= 0.25, e
3
= 0.33,
respectively, and we obtain: C(E
1
) = 0.593, C(E
2
) = 0.649, C(E
3
) = 0.553.
Step 6: Calculate the aggregation weight w(E
i
) of expert E
i
as follows: w(E
1
) = 0.330, w(E
2
) =
0.361, w(E
3
) = 0.309.
Step 7: Aggregate each fuzzy opinion into a group fuzzy opinion as
R =
3
i=1
w(E
i
)
R
i
= (1.489, 2.334, 3.489, 4.979).
Referring to Fig. 2, we can see the second opinion
R
2
is close to the other opinions, so the
aggregation weight of
R
2
is the largest, and the consensus opinion
R is therefore close to
R
2
. The consistency degree of the opinion
R
3
is the smallest of three ones so that
R
3
is less
important, which is fair and reasonable. However, in Hsu and Chens result, the consensus
degree coecient of
R
2
(CDC
2
= 0.33) is smaller than CDC
1
= 0.34 although
R
2
is closer to
the consensus opinion
R than
R
1
. This does not coincide with peoples intuition. Furthermore,
AGGREGATION OF FUZZY OPINIONS UNDER GROUP DECISION-MAKING 71
the uncertainty of our aggregation result (H(
R) = 3.495).
5 Conclusions
In this article, we dene a similarity measure, and a distance measure is also given to study
the consistency of one expert to the other experts. To deal with the situation when opinions are
disjoint, we introduce a consistency index of each individual opinion on the basis of similarity
and distance. The basic idea of our aggregation method is that the aggregation weight of the
opinion should be larger if the degrees of similarity between the experts opinion and the other
opinions are larger and the degrees of distance between his opinion and the other opinions are
smaller. The importance of each expert is also considered in the procedure of our method.
Finally, a numerical example shows that our method is rather ecient.
References
[1] A. Bardossy, L. Duckstein and I. Bogaradi, Combination of fuzzy numbers representing expert
opinions, Fuzzy Sets and Systems, 1993, 57: 173181.
[2] A. Ishikawa, M. Amagasa, T. Shiga, G. Tomizawa, R. Tatsuta and H. Mieno, The max-min Delphi
method and fuzzy Delphi method via fuzzy integration, Fuzzy Sets and Systems, 1993, 55: 241253.
[3] H. Nurmi, Approaches to collective decision making with fuzzy preference relations, Fuzzy Sets and
Systems, 1981, 6: 249259.
[4] J. Kacprzyk and M. Fedrizzi, A soft measure of consensus in the setting of partial (fuzzy) prefer-
ences, European Journal of Operational Research, 1988, 34: 315325.
[5] J. Kacprzyk, M. Fedrizzi and H. Nurmi, Group decision making and consensus under fuzzy pref-
erences and fuzzy majority, Fuzzy Sets and Systems, 1992, 49: 2131.
[6] M. Fedrizzi and J. Kacprzyk, On measuring consensus in the setting of fuzzy preference relations, in
Non-conventional Preference Relations in Decision Making (ed. by J. Kacprzyk and M. Roubens),
Springer, Berlin, 1988, 129141.
[7] R. N. Xu and X. Y. Zhai, Extensions of the analytic hierarchy process in fuzzy environment, Fuzzy
Sets and Systems, 1992, 52: 251257.
[8] H. M. Hsu and C. T. Chen, Aggregation of fuzzy opinions under group decision making, Fuzzy Sets
and Systems, 1996, 79: 279285.
[9] H. S. Lee, Optimal consensus of fuzzy opinions under group decision making environment, Fuzzy
Sets and Systems, 2002, 132: 303315.
[10] R. Zwick, E. Carlstein and D. V. Budescu, Measures of similarity among fuzzy concepts: A com-
parative analysis, Internat, J. Approximate Reasoning, 1987, 1: 221242.
[11] A. Kauman and M. M. Gupta, Introduction to Fuzzy Arithmetic: Theory and Applications, Van
Nostrand Reinhold, New York, 1985.
[12] D. Dubois, E. Kerre and R. Mesiar, Fuzzy interval analysis, in Fundamentals of Fuzzy Sets (ed.
by D. Dubois and H. Prade), The Handbooks of Fuzzy Sets Series, Kluwer Academic Publishers,
Boston, 2000, 483581.
[13] S. Heilpern, Representation and application of fuzzy numbers, Fuzzy Sets and Systems, 1997, 91:
259268.
[14] R. M. Tong and P. P. Bonissone, A linguistic approach to decision making with fuzzy sets, IEEE
Trans. Systems Man Cybernet, 1980, 10: 716723.
[15] J. Kacprzyk, Multistage Fuzzy Control, Wiley Chichester, 1997.