[ieee 2007 international symposium on computational intelligence and intelligent informatics -...
TRANSCRIPT
Janos FodorInstitute of Intelligent Engineering Systems
Budapest Tech
Becsi ut 96/b, H-1034 Budapest
Hungary
Email: [email protected]
Imre J. RudasInstitute of Intelligent Engineering Systems
Budapest Tech
Becsi ut 96/b, H-1034 Budapest
Hungary
Email: [email protected]
Abstract—We investigate the problem of aggregation of fuzzypreference structures in multicriteria decision making. Startingfrom monocriterion weak preferences, there are two ways todefine global strict preference, indifference and incomparability.We prove that, under very mild conditions, the only aggregationfor which those two ways result in the same global preferencestructure is dictatorial.
I. INTRODUCTION
In a multicriteria decision making (MCDM) problem, when
criteria are given in terms of (fuzzy) preference relations on the
set of alternatives, two major steps are distinguished: aggrega-
tion and exploitation. The aggregation phase defines a global
preference relation, and the exploitation phase transforms this
global relation into a ranking of the alternatives.
In the present paper we deal with the aggregation phase
when monocriterion preferences are binary fuzzy relations.
Thanks to keen interest in research in the last decade, theo-
retical aspects of fuzzy preferences have become rather well-
developed and sound. The situation most often considered is
the following. Suppose that A is a set of alternatives, and
R is fuzzy weak preference relation on A (that is, a function
R : A×A → [0, 1] such that R(a, a) = 1 for all a ∈ A), where
the value R(a, b) ∈ [0, 1] expresses the degree to which a is
not worse than b . Three further binary fuzzy relations can be
attached to such an R:
• strict preference: P (a, b) means the degree to which a is
better than b;
• indifference: I(a, b) gives the degree to which a and bare indifferent;
• incomparability: J(a, b) yields the degree to which a and
b cannot be compared.
There is an axiomatic framework of constructing P , I and
J from a given R (see [5]), which we will briefly recall in
the next section. It is easy to prove that the resulted triplet
(P, I, J) is a ϕ-fuzzy preference structure on A. This notion
was introduced in [8].
The natural area of using preference structures is multi-
criteria decision making. In that environment we start from
monocriterion preferences R1, . . . , Rn on A. Concerning any
of the previous three relations, we can define global strict
preference, global indifference and global incomparability in
two different ways. To illustrate this, consider e.g. the strict
preference.
One possibility is to aggregate first R1, . . . , Rn to obtain
global weak preference R := M(R1, . . . , Rn). Then we can
use the above axiomatic construction to define P , I and J .
The other possibility is to define first monocriterion strict
preference, indifference and incomparability, and then to ag-
gregate each type to obtain the related global relations.
Generally speaking, these two ways yield different global
relations. Thus, it is a natural need to search for appropriate
aggregation functions which give the same results in both of
the above approaches.
The main result of the paper says that the only type
of aggregation procedures satisfying these requirements is
dictatorial: it takes into account only one criterion in the global
relations. As a consequence, the global (P, I, J) is a ϕ-fuzzy
preference structure too.
The paper is organized as follows. In Section 2 we recall
some preliminary notions and results that are indispensable in
the sequel. In Section 3 we formulate the aggregation problem
in a mathematical way, by the help of functional equations.
Then we solve the problem for strict preference first. This
results is easily applied for indifference and incomparability.
In some extremely important particular cases the general
dictatorial aggregation is proved to be the identity function.
Finally we present some concluding remarks.
II. PRELIMINARIES
Let (T, S, N) be a continuous De Morgan triplet (i.e., Tis a continuous t-norm, N is a strong negation, and S is
the N -dual t-conorm of T ) representing logic operations in
the fuzzy framework. Consider a set of alternatives A, and
a reflexive binary fuzzy relation R on A. According to the
axiomatic approach in [5], strict preference P , indifference I ,
and incomparability J attached to R is constructed by the use
of some generator function for each of these relations:
P (a, b) := p(R(a, b), R(b, a)),I(a, b) := i(R(a, b), R(b, a)),J(a, b) := j(R(a, b), R(b, a)),
for any a, b ∈ A, where p, i, j are functions from [0, 1]2 to
[0, 1] with some prescribed properties. These functions can be
On Aggregation of Fuzzy Preference Structures
1-4244-1158-0/07/$25.00 © 2007 IEEE.85
obtained as solutions of a system of functional equations
S(p(x, y), i(x, y)) = x,
S(p(x, y), j(x, y)) = N(y),
where the t-conorm S and strong negation N are also un-
known. It was proved in [5] that these functional equations
imply the existence of an automorphism ϕ of the unit interval
such that (T, S, N) is of the following form
T (x, y) = ϕ−1(max{ϕ(x) + ϕ(y) − 1, 0}),S(x, y) = ϕ−1(min{ϕ(x) + ϕ(y), 1}),
N(x) = ϕ−1(1 − ϕ(x))
for all x, y ∈ [0, 1]. Then (T, S, N) is called a Łukasiewicz-triplet, and the above functional equations can be written as
ϕ(p(x, y)) + ϕ(i(x, y)) = ϕ(x), (1)
ϕ(p(x, y)) + ϕ(j(x, y)) = 1 − ϕ(y), (2)
and a solution is denoted by < p, i, j >ϕ.
There is a parametric family of t-norms which plays a key
role in several areas of fuzzy set theory, in particular in fuzzy
preference modelling. Let s ∈ ]0, 1[∪ ]1,∞[, then the Frank
t-norm with parameter s is defined by
T s(x, y) := logs
(1 +
(sx − 1)(sy − 1)s − 1
), x, y ∈ [0, 1].
The limit t-norms are given by the minimum (s → 0), the
algebraic product (s → 1) and the Łukasiewicz t-norm (s →∞). So, we employ the respective notations T 0, T 1 and T∞
for these limit cases. We also identify T1∞ and T 0, as well
as T10 and T∞. The family (T s)s∈[0,∞] is called the Frank
t-norm family [7]. The dual t-conorm is usually denoted by
Ss.
Now we can recall the following result from [5] which is
the most relevant one in studying the aggregation problem.
Theorem 1. Assume that p(x, y) = T1(x,Nϕ(y)) andi(x, y) = T2(x, y), where T1 and T2 are continuous t-norms.Then < p, i, j >ϕ satisfies (1) and (2) if and only if thereexists a number s ∈ [0,∞] such that
T1(x, y) = ϕ−1(T s(ϕ(x), ϕ(y))),T2(x, y) = ϕ−1(T 1/s(ϕ(x), ϕ(y))),
where T s and T 1/s belong to the Frank family.
The notion of fuzzy preference structures has been intro-
duced in [8]. We recall now a minimal definition after [2].
Consider an automorphism ϕ of [0, 1]. A triplet (P, I, J) of
binary fuzzy relations on A is called a ϕ-fuzzy preferencestructure (ϕ-FPS) on A if I is reflexive, I and J are sym-
metric, and we have the following equality for all a, b ∈ A:
ϕ(P (a, b)) + ϕ(P (b, a)) + ϕ(I(a, b)) + ϕ(J(a, b)) = 1. (3)
Because of (3), (P, I, J) can also be called a ϕ-additive FPS
on A.
It is immediate that any (P, I, J) generated by a solution
< p, i, j >ϕ of the functional equation system (1)-(2) is a ϕ-
FPS on A. In [6] the interested reader can find more details
where we presented an overview of fundamentals and recent
advances.
III. FORMULATING THE AGGREGATION PROBLEM
Let A be a set of alternatives and R1, . . . , Rm be reflexive
binary fuzzy relations on A (each weak preference Rk repre-
sents a criterion). Let a, b ∈ A, and denote xk := Rk(a, b),yk := Rk(b, a) for k ∈ {1, . . . ,m}. Then we obtain the global
weak preference as follows:
R(a, b) = MR(R1(a, b), . . . , Rm(a, b))= MR(x1, . . . , xm),
R(b, a) = MR(R1(b, a), . . . , Rm(b, a))= MR(y1, . . . , ym).
Using these global values, finally we can define P, I, J as
follows:
P (a, b) := p(MR(x1, . . . , xm),MR(y1, . . . , ym)),I(a, b) := i(MR(x1, . . . , xm),MR(y1, . . . , ym)),J(a, b) := j(MR(x1, . . . , xm),MR(y1, . . . , ym)).
On the other hand, monocriterion strict preferences, indif-
ferences, and incomparabilities are defined as Pk(a, b) :=p(xk, yk), Ik(a, b) := i(xk, yk), Jk(a, b) := j(xk, yk) for
k ∈ {1, . . . ,m}. Thus, we can introduce P ′, I ′, J ′ as follows:
P ′(a, b) := MP (p(x1, y1), . . . , p(xm, ym)),I ′(a, b) := MI(i(x1, y1), . . . , i(xm, ym)),J ′(a, b) := MJ(j(x1, y1), . . . , j(xm, ym)).
Therefore, equalities P = P ′, I = I ′, J = J ′ can be
formulated as the following functional equations:
p(MR(x1, . . . , xm),MR(y1, . . . , yn))= MP (p(x1, y1), . . . , p(xm, ym)), (4)
i(MR(x1, . . . , xm),MR(y1, . . . , ym))= MI(i(x1, y1), . . . , i(xm, ym)), (5)
j(MR(x1, . . . , xm),MR(y1, . . . , ym))= MJ(j(x1, y1), . . . , j(xm, ym)), (6)
where p, i, j : [0, 1]2 → [0, 1] are given strict preference, in-
difference, and incomparability generator functions (solutions
of (1) and (2)), and we are looking for the unknown functions
MR, MP , MI , and MJ which satisfy the equations (4)–(6)
for all xk, yk ∈ [0, 1] (k ∈ {1, . . . ,m}).
The generality of having possibly different aggregation
functions is apparent, as we can see by the following lemma.
3rd International Symposium on Computational Intelligence and Intelligent Informatics – ISCIII 2007 • Agadir, Morocco • March 28-30, 2007
86
Lemma 1. (1) If MR and MP are solutions of (4) withMR(0, . . . , 0) = 0 then MR = MP . Moreover, MR is self-dual with respect to N .
(2) If MR and MI are solutions of (5) with MR(1, . . . , 1) =1 then MR = MI .
(3) If MR and MJ are solutions of (6) with MR(0, . . . , 0) =0 then MR = MJ .
Thus, we have to deal only with one unknown aggregation
function, denoted by M(= MR = MP = MI = MJ), in
all cases. Moreover, M is self-dual with respect to the strong
negation N ; that is, we have for all x1, . . . , xm ∈ Rm:
M(x1, . . . , xn) = N(M(N(x1), . . . , N(xm))).
In addition, we require the following properties too (see
conditions in Lemma 1).
M(0, . . . , 0) = 0, (7)
M(1, . . . , 1) = 1, (8)
M is nondecreasing and continuous (9)
IV. ONLY DICTATORIAL AGGREGATION EXISTS
In this section we show a general result on the form of the
aggregation function M . First we investigate strict preferences,
because the corresponding functional equation (4) is very
restrictive in itself.
A. Aggregating Strict Preferences: Solution of (4)
The following theorem was already stated in [4].
Theorem 2. The general solution M of (4) is of the followingform:
M(x1, . . . , xm) = gk(xk), (10)
where gk : [0, 1] → [0, 1] is a nondecreasing function withgk(0) = 0, gk(1) = 1, satisfying
gk(p(x, y)) = p(gk(x), gk(y)) (11)
for all x, y ∈ [0, 1].
Therefore, the only possible aggregation takes into account
one criterion, and thus it is dictatorial, even if we consider only
strict preferences. The obtained class of aggregation functions
will be narrowed further if we consider indifferences and
incomparabilities.
B. Indifferences and Incomparabilities
Since MI = MJ = MP = MR = M , now we must use
the same formula (10) for defining global indifference and
incomparability. In addition to (11), the function gk has to
satisfy the following functional equations too:
gk(i(x, y)) = i(gk(x), gk(y)) (x, y ∈ [0, 1]), (12)
gk(j(x, y)) = j(gk(x), gk(y)) (x, y ∈ [0, 1]). (13)
These are simple consequences of the respective equations (5)
and (6).
C. The General Form of the Aggregation
We can summarize the previous facts in the following
statement.
Theorem 3. The general solution M of equations (4), (5), (6)is of the following form:
M(x1, . . . , xm) = gk(xk),
where gk : [0, 1] → [0, 1] is a nondecreasing function withgk(0) = 0, gk(1) = 1, satisfying the following three functionalequations:
gk(p(x, y)) = p(gk(x), gk(y)),gk(i(x, y)) = i(gk(x), gk(y)),gk(j(x, y)) = j(gk(x), gk(y))
for all x, y ∈ [0, 1]. In particular, gk must be self-dual withrespect to the strong negation N :
gk(N(x)) = N(gk(x)) (x ∈ [0, 1]).
We can be certain that there is at least one function gk which
satisfies all the previous conditions. Indeed, gk(x) := x for all
x ∈ [0, 1] is a continuous solution. Generally speaking, we do
not know if there are other continuous solutions. The study in
the next section might suggest a negative answer.
V. PARTICULAR CLASSES
Now we determine continuous solutions of equations (11),
(12), and (13) in the case indicated by Theorem 1. That
is, generator functions of strict preference, indifference and
incomparability are of the following form now:
ps,ϕ(x, y) := T sϕ(x, Nϕ(y)), (14)
is,ϕ(x, y) := T 1/sϕ (x, y), (15)
js,ϕ(x, y) := T 1/sϕ (Nϕ(x), Nϕ(y)) (16)
for all x, y ∈ [0, 1], where T sϕ is a ϕ-transform of the t-norm
T s belonging to the Frank family with parameter s ∈ [0,∞],and 1/∞ := 0, 1/0 := ∞. Fortunately, we are able to use
some of our previous results related to decomposable measures
[3].
A. Case 1: s = 0If s = 0 then i0,ϕ(x, y) = ϕ−1(max(ϕ(x) + ϕ(y) − 1, 0)),
the ϕ-transform of the Łukasiewicz t-norm. It is continuous,
Archimedean, with zero divisors (i.e., it is nilpotent).
In this case equation (12) is given as follows (x, y ∈ [0, 1]):
gk(ϕ−1(max(ϕ(x) + ϕ(y) − 1, 0)))= ϕ−1(max(ϕ(gk(x)) + ϕ(gk(y)) − 1, 0)). (17)
Lemma 2. The general continuous solution of (17) is of thefollowing form (x ∈ [0, 1]):
gk(x) = ϕ−1(max{1 − αk + αkϕ(x), 0}), (18)
where αk ≥ 1 is a real parameter.
J. Fodor, I. J. Rudas • On Aggregation of Fuzzy Preference Structures
87
Since gk must be self-dual with respect to the strong
negation N (see Lemma 1 (1)), it is immediate that this can
hold if and only if αk = 1, i.e., if and only if gk(x) = x for
all x ∈ [0, 1] .
B. Case 2: s = ∞If s = ∞ then i∞,ϕ(x, y) = min(x, y), the minimum
t-norm (independently of ϕ). It is continuous and non-
Archimedean. Moreover, p∞,ϕ(x, y) = ϕ−1(max(ϕ(x) −ϕ(y), 0)).
Therefore in the present particular case equation (12) is
automatically satisfied by any nondecreasing function gk.
However, equation (11) is as follows (x, y ∈ [0, 1]):
gk(ϕ−1(max(ϕ(x) − ϕ(y), 0)))= ϕ−1(max(ϕ(gk(x)) − ϕ(gk(y)), 0)). (19)
After a suitable substitution, this equation can be transformed
into (17), thus the general solution of (19) is given by (18),
and among those functions there is only one which is self-dual
with respect to N : this is again gk(x) := x for all x ∈ [0, 1].
C. Case 3: s ∈ ]0,∞[
If s ∈ ]0,∞[ then is,ϕ(x, y) = T1/sϕ (x, y) is a continuous,
Archimedean t-norm without zero divisors (i.e., it is strict).
In this case is,ϕ can be obtained by an appropriate ϕs-
transformation of the product t-norm:
T 1/sϕ (u, v) = ϕ−1
s (ϕs(u) · ϕs(v)) (u, v ∈ [0, 1]).
More explicitly, we have for all x ∈ [0, 1]:
ϕs(x) =
⎧⎪⎪⎨⎪⎪⎩
(1/s)ϕ(x) − 1(1/s) − 1
if s ∈]0,∞[, s �= 1,
ϕ(x) if s = 1
, (20)
and thus
ϕ−1s (x) =
⎧⎨⎩
ϕ−1(x) if s = 1
ϕ−1(log(1/s)(1 + ((1/s) − 1)x)) otherwise,.
From the solution of Cauchy’s famous functional equations
(see e.g. [1]) we can prove the following result (see [3]).
Lemma 3. For s ∈]0,∞[, the general continuous solution of(12) is of the following form:
gk(x) = ϕ−1s (ϕλk
s (x)) (x ∈ [0, 1]), (21)
where ϕs is the automorphism of the unit interval given in(20), and λk is a positive parameter.
As in the previous two particular cases, it is easy to see that
the only self-dual function (with respect to N ) among (21) is
gk(x) := x for all x ∈ [0, 1].We can summarize the results of this section in the following
theorem.
Theorem 4. Let (T, S, N) be a given Łukasiewicz triplet. Sup-pose generator functions of strict preference, indifference and
incomparability are given by (14), (15), and (16), respectively.Then the only continuous aggregation function which satisfiesalso equations (11), (12), (13) is given as follows:
M(x1, . . . , xm) = xk
for some k ∈ {1, . . . ,m} (x1, . . . , xm ∈ [0, 1]).
Therefore, if we have a profile of ϕ-fuzzy preference
structures
〈(P1, I1, J1), . . . , (Pk, Ik, Jk), . . . , (Pm, Im, Jm)〉then the aggregated triplet is (Pk, Ik, Jk). Consequently, the
aggregated triplet is a ϕ-FPS.
VI. CONCLUDING REMARKS
As we have shown, the only continuous aggregation of
ϕ-fuzzy preference structures is dictatorial. That is, it is
impossible to assign weights other than 0 or 1 to each criterion
in order to have the same global preferences in both types of
aggregation procedures mentioned in the Introduction.
Note, however, that there exist non-dictatorial continuous
aggregations of particular preference structures (for instance,
when monocriterion weak preferences are complete and nega-
tively transitive fuzzy relations). More results on such partic-
ular cases can be found in [5].
ACKNOWLEDGMENT
The authors would like to thank supports from OTKA
T046762 and from SZVT (Scientific Society for Management
and Organization, Hungary).
REFERENCES
[1] J. Aczel, Lectures on Functional Equations and their Applications, NewYork: Academic, 1966.
[2] B. De Baets, and B. Van de Walle, “Minimal definitions of classicaland fuzzy preference structures”, in: Proc. of the Annual Meeting ofNAFIPS, Syracuse, New York, pp. 299–304, 1997.
[3] D. Dubois, J. Fodor, H. Prade and M. Roubens, “Aggregation ofdecomposable measures with application to utility theory”, Theory andDecision, vol. 41, pp. 59–95, 1996.
[4] J. Fodor and M. Roubens, “Aggregation of strict preference relations inMCDM procedures”, in: V. Novak, J. Ramik, M. Mares, M. Cerny andJ. Nekola (eds.), Fuzzy Approach to Reasoning and Decision Making,Dordrecht: Kluwer, pp. 163–171, 1992.
[5] J. Fodor and M. Roubens, Fuzzy Preference Modelling and MulticriteriaDecision Support, Dordrecht: Kluwer, 1994.
[6] J. Fodor and B. De Baets, “Fuzzy Preference Modelling: Fundamentalsand Recent Advances,” in: H. Bustince, F. Herrera and J. Montero(eds.), Fuzzy Sets and Their Extensions: Representation, Aggregationand Models. Intelligent Systems from Decision Making to Data Mining,Web Intelligence and Computer Vision, New York: Springer, in press.
[7] J.M. Frank, “On the simultaneous associativity of F (x, y) and x+ y−F (x, y)”, Aequationes Mathematicae, vol. 19, pp. 194–226, 1979.
[8] B. Van de Walle, B. De Baets and E. Kerre, “Characterizable fuzzypreference structures”, The Annals of Operations Research, vol. 80, pp.105–136, 1998.
3rd International Symposium on Computational Intelligence and Intelligent Informatics – ISCIII 2007 • Agadir, Morocco • March 28-30, 2007
88