Professional Documents
Culture Documents
Aggregation Operator
Ronald R. Yager*
Machine Intelligence Institute, Iona College, New Rochelle, NY 10801
We discuss the fuzzy integral. The centrality of the ordering operation, based upon the arguments
to be aggregated, is pointed out. We then extend the fuzzy integral aggregation operator by
allowing the ordering operation to be based upon values other then those being aggregated. This
leads to the induced fuzzy integral aggregation operator. We look at this new operator and study
its properties. We show its relationship to a formulation called limited fuzzy integral aggregation.
It is shown how this new induced fuzzy integral operator provides a natural framework for the
implementation of nearest neighbor rules. Throughout this work, use is made only of the ordinal
aspects of the information used. © 2002 Wiley Periodicals, Inc.
1. INTRODUCTION
The Sugeno or Fuzzy Integral1,2 provides an aggregation operator which can
be seen as providing a generalization and extension of the logical anding and oring
operations. Some applications in which the Fuzzy Integral (FI) can be used are in
the aggregation of the truths of propositions, the construction of multicriteria
decision functions from individual criteria, and the determination of the satisfac-
tion of a hypothesis to a collection of evidence. It can be used to provide operators
for the aggregation of fuzzy subsets. In these applications the arguments being
aggregated can be seen as some kinds of truth or belief values. While applications
of this operator have typically used values drawn from the unit interval I ⫽ [0, 1],
an important feature of this operator is that it only requires that the information
used be drawn from an ordinal scale. This situation opens its usefulness to
applications in many environments in which we are trying to model aspects of
human intelligence. Particularly notable are the possibilities for aggregating infor-
mation expressed in linguistic terms as these terms can often only be ordered. This
situation makes this aggregation operator useful for the development of Zadeh’s
agenda of computing with words.3,4
Central to the implementation of the FI aggregation operator is an ordering
operation based upon the arguments to be aggregated. This ordering operation, in
*e-mail: yager@panix.com.
where a-index( j) is the index of the jth largest of the a i and H j ⫽ {A a-index(r) 兩r ⫽
1 to j}. Letting b j ⫽ a a-index( j) , the jth largest of the a i , and letting m j ⫽ (H j ),
we can express S (a 1 , a 2 , . . . , a n )) ⫽ Maxj [m j ∧ b j ].
For a particular aggregation, S (a 1 , . . . , a n ), we let B denote a vector having
b j as its jth element, we call this the ordered argument vector. We also let M denote
a vector, called the weighting vector, whose jth component m j is (H j ). Even
though the weighting vector M is dependent upon the arguments, it must always
have m j ⱖ m i if j ⬎ i and m n ⫽ t m . We shall say a vector having these two
properties is proper. It is well known that the FI aggregation operator it is bounded,
Mini [a 1 ] ⱕ S (a 1 , . . . , a n ) ⱕ Maxi [a i ] and monotonic, if â i ⱖ a i for all i then
S (â 1 , . . . , â n ) ⱖ S (a 1 , . . . , a n ).
A prototypical application of the FI aggregation operator is in the problem of
hypothesis verification. Let the Ꮽ ⫽ {A 1 , . . . , A n } be a collection of evidence
and let a i be the degree of compatibility of a proposed hypothesis with the evidence
A i . Here we let the fuzzy measure be defined such that (E) indicates the
credibility associated with a hypothesis that is compatible with all the pieces of
evidence in the set E. In this case the FI aggregation can be used to provide a
measure of the validation of the hypothesis.
By appropriate choice of the underlying fuzzy measure the FI operator leads
to a large class of different aggregation operators. We refer the interested reader to
Refs. 1, 12, and 13 for more details.
†
Essentially a-index is a mapping a-index: N 3 N, where N ⫽ {1, 2, . . . , n}.
1052 YAGER
where d j ⫽ ∧ j
a v-index(k) and w j ⫽ (G j ), where G j ⫽ {A v-index(k) 兩k ⫽ 1 to
k⫽1
j}, the subset of elements with the j highest values for the inducing variable. We
further note that d j is the minimum of the arguments of the objects with the j
highest-order inducing values. At times we shall find it convenient to denote a
vector W consisting of the w j as the weighting vector W and a vector D consisting
of the d j as the ordered argument vector.
A simple example will illustrate the I-FI aggregation.
This leads us to the observation that if the object with the largest order inducing
value has the smallest argument value, then its argument value is the valuation of
S (具v i , a i 典).
We can also show that there exists an even tighter boundary on the I-FI
aggregation operator then the Max and Min of the argument values. We first see
that d 1 ⫽ a v-index(1), the argument value of the object with the largest order
inducing value. Since d j ⱕ d 1 for all j, then:
Max 关共Gj 兲∧dj 兴 ⱕ d1 ⱕ av-index共1兲
j
Specifically we observe that the upper boundary of the I-FI operator is the
argument value of the object with the largest order inducing value.
4. EXAMPLES OF I-FI
Let us look at this induced FI operator for some special cases of fuzzy
measures. First we shall consider the case when is a cardinality based measure.
For these measures the weighting vector W, whose components are the (G j ), just
depends on the cardinality. Thus w j does not depend on the indexing function,
v-index, it is constant in this situation. In this case S (具v j , a j 典) ⫽ Maxj [w j ∧ d j ]
with w j independent of the argument and d j ⫽ Mink⫽1 to j [a v-index(k) ].
Consider first the special case of * , w n ⫽ t m and w j ⫽ t 0 for all j ⫽ n, here:
S 共具v j, a j典兲 ⫽ Max 关wj ∧dj 兴 ⫽ wn ∧dn ⫽ dn ⫽ Min 关ai 兴
j i
We note that both * and are special cases of this. If ␣ ⫽ t m , we get * and,
*
if ␣ ⫽ t 0 , we get .
*
Another special case of cardinality based measures are what we shall call the
Kth component measure. Here (E) ⫽ t 0 if Card(E) ⬍ K and (E) ⫽ t m if
Card(E) ⱖ K. From the nature of , w j ⫽ t 0 if j ⬍ K and w j ⫽ t m if j ⱖ K. Thus
in this case:
K
S 共具v j, a j典兲 ⫽ Max 关wj ∧dj 兴 ⫽ ∧ av-index共i兲
j i⫽1
It is the minimal argument value among the K objects with the largest order
inducing values. A special case of this is a kind of median. Here K is equal to n⫹1 2
if the number of arguments is odd or K ⫽ n/ 2 if the number of arguments is even.
Here we essentially take half the objects with the largest inducing value and then
use as our aggregation the minimal argument value of these. We note that if the
argument and order inducing variable are the same this reduces the median, “we
take the smallest argument value among the half the objects with the largest
argument value.”
Let us now consider the possibility measure in this induced environment. In
this case (E) ⫽ Maxi僆E [ ␣ i ] where ␣ i ⫽ ({A i }) and at least one ␣ i ⫽ t m . With
G j ⫽ {A v-index(k) 兩k ⫽ 1 to j} we see that (G j ) ⫽ ∨ k⫽1 j
␣ v-index(k) , it is the
highest possibility among the j elements with largest inducing values. In this
situation S (具v i , a i 典) ⫽ Maxj [∨ k⫽1 j
␣ v-index(k) ∧ ∧ k⫽1
j
a v-index(k) ].
We note that if v-index(1) has the largest possibility, ␣v-index(1) ⫽ t m , then
S (具v i , a i 典) ⫽ a v-index(1). Thus, if the object with the largest order inducing value
has the highest possibility, then its argument value is the aggregated value. More
generally, we see that if the object with the qth largest order inducing value has
possibility t m , ␣ v-index(q) ⫽ t m , then S (具v i , a i 典) ⱖ ∧ k⫽1
q
a v-index(k) . It is at least
as big as the smallest argument among the elements with the q largest order
inducing values.
Let us consider the case of the necessity measure. We recall for this measure
if we denote  i ⫽ (Ꮽ ⫺ {A i }), then (E) ⫽ ∧ iⰻE  i . It is also assumed that
at least one  i ⫽ t 0 . In this case with G j ⫽ {A v-index(k) 兩k ⫽ 1–j}, we have
(G j ) ⫽ ∧ k⫽j⫹1n
 v-index(k) . Using this, we have:
One issue that must be addressed when using this induced type aggregation is
related to the situation when there exists ties among the arguments with regard to
the order inducing variable.
1056 YAGER
A 1 ⫽ 具r 4, t 4典
A 2 ⫽ 具r 6, t 3典
A 3 ⫽ 具r 3, t 2典
A 4 ⫽ 具r 4, t 6典
A 5 ⫽ 具r 4, t 3典
Ordering these by their induced values, we get:
A2
A1 ⫺ A4 ⫺ A5
A3
INDUCED FUZZY INTEGRAL AGGREGATION OPERATOR 1057
We see that A 1 , A 4 , and A 5 are tied with respect to their order inducing value.
Adjudicating the tied elements in descending order of their argument value, we get:
A2
A4
A1
A5
A3
This adjudication gives us an order inducing function v-index( j) such that
v-index共1兲 ⫽ 2, v-index共2兲 ⫽ 4, v-index共3兲 ⫽ 1, v-index共4兲 ⫽ 5, v-index共5兲 ⫽ 3
In the following section we shall provide some additional ideas which enable
us to justify this method of dealing with ties.
Let us denote MinA i僆E [a i ] as Q(E). Q(E) indicates the degree to which all the
objects in E satisfy our condition.
As an example that brings some useful intuition to this aggregation we
consider a multicriteria aggregation problem, here we are trying to determine the
appropriateness of an alternative solution to our requirements. Here Ꮽ are a set of
criteria of interest, and (E) indicates the degree that we are satisfied with a
solution that satisfies all the criteria in E. We shall more simply denote (E) as the
degree to which E provides a credible collection of criteria. In this framework a i
indicates the degree that an alternative satisfies criteria A i . Under this semantics
Q(E) indicates the degree to which the alternative being considered satisfies all the
criteria in E. Here, then, our evaluation of an alternative is obtained as
Agg ( A 1 , . . . , a n ).
The semantics of this aggregation is the following: We note that the term
(E) ∧ Q(E) indicates the degree to which the subset E provides an acceptable
collection of criteria and the degree to which E is satisfied by the alternative being
evaluated. With this in mind, our operator Agg can be viewed as determining the
satisfaction using the following imperative:
1058 YAGER
11
What is of interest to us is that it can be shown that Agg (a 1 , . . . , a n ) is
the FI aggregation operator, Agg (a 1 , . . . , a n ) ⫽ S (a 1 , . . . , a n ).
First we note that (E) ⫽ t m only for sets G j ⫽ {A s-index(k) 兩k ⫽ 1 to j}; otherwise
it is t 0 . Hence, in this case, Agg 兩 (a 1 , . . . , a n ) ⫽ ∧ j⫽1
n
[ (G j ) ∧ MaxA i僆G j[a i ]].
However, from the definition of G j we see that MaxA i僆G j[a i ] ⫽ ∧ k⫽1 j
a s-index(k) .
Thus Agg 兩 (a 1 , . . . , a n ) ⫽ ∧ j⫽1 [ (G j ) ∧ ∧ k⫽1 a s-index(k) ], which is essentially
n j
‡
A more formal expression of this concept can be made if we represent P as a binary
relation R p such that R p ( A i , A j ) ⫽ 1 if A i ⱖ P A j and denote TOT( A i ) ⫽ ¥ j R p ( A i , A j ). A
subset E is called a rooted subordering of P if E contains all A j such that TOT( A j ) ⬎ TOT( A i* ),
where TOT( A i* ) ⫽ MinA k僆E TOT( A k ).
INDUCED FUZZY INTEGRAL AGGREGATION OPERATOR 1061
ordering is the same as the result obtained using a limited FI aggregation in which
the limited function is induced by a linear ordering, which is the same except the
tied elements are lexicographically ordered by their argument value.
Before proceeding we shall find the following terminology useful. Let P be a
weak ordering on a space Ꮽ. We shall say that P̂ is a linear ordering adjudicated
from P if the ordering between the untied elements are the same and the tied
elements are linear ordered lexicographically by some adjudication procedure. We
shall say that is a limiting function induced from a weak ordering.§ P if (E) ⫽
t m for all subsets that are rooted suborderings of P and (E) ⫽ t 0 if E is not a
rooted subordering.
§
Here we note that a linear ordering is a special case of weak ordering and hence the same
can be done with linear ordering.
**For simplicity we assume only one group of tied elements the following analysis can
easily be extended to multiple groups of tied elements.
1062 YAGER
冉 冊
m
where Ᏺ j is the collection of subsets of the tied elements that has A t j as its object
with the highest index. What is clear is that any M 僆 Ᏺ j also has M 債 T j . From
this it follows that:
∨ 共共E兲∧Q共E兲兲 ⫽ ∨ 共共E兲∧Q共E兲兲
ˆ2
E僆C E僆C 2
冉 冊
n n j
We see this is the induced FI aggregation operator. Thus the nearest neighbor
method can be seen as a type of I-FI aggregation.
We now look at some specific cases of nearest neighbor rule. These special
cases are determined by the structure of . First we shall consider the class of
cardinality based measures. For this class we let (G j ) ⫽ w j , where w j just
depends on j and w j ⱖ w i if j ⬎ i. If we consider the specific cardinality based
measure where w j ⫽ t m for all j, then:
冉 冊
n j
It is the degree to which the nearest neighboring prototype has the notable property.
We see this as corresponding to the simple strict nearest neighbor rule.
If we consider the case where w j ⫽ t 0 for j ⫽ 1 to p ⫺ 1 and w j ⫽ t m for
j ⫽ p to n, then:
冉 冊
n j p
It is the minimal degree to which any of the p nearest neighbor prototypes have the
property. This can be seen as a kind of pth nearest neighbor rule.
Another class of fuzzy measures occurs when we associate with each proto-
type A i a degree of credibility ␣ i and assume the (E) ⫽ MaxA i僆E [ ␣ i ], is a
possibility measure. In this case:
冉冉 冊冉 冊冊
n n n
8. CONCLUSION
We discussed the fuzzy measure and the related idea of fuzzy integral. We
pointed out the centrality within the implementation of fuzzy integral aggregation
of the ordering operation based upon the arguments to be aggregated. We then
extended the fuzzy integral aggregation operator by allowing the ordering opera-
tion to be based upon values other then those being aggregated. This lead us to the
induced fuzzy integral aggregation operator. We looked at this operator and studied
its properties. We showed its relationship to a formulation called limited fuzzy
integral aggregation. It was shown how this new induced fuzzy integral operator
provides a natural framework for the implementation of nearest neighbor rules. We
introduced the idea of prototype based reasoning. Throughout this work the ordinal
nature of fuzzy integral was exploited to enable us only to use data drawn from a
ordinal scale; this situation makes the techniques developed here particularly
suitable for application within the scope of Zadeh’s approach to computing
with words.
References
1. Sugeno M. Fuzzy measures and fuzzy integrals: A survey. In: Gupta MM, Saridis GN,
Gaines BR, editors. Fuzzy automata and decision process. Amsterdam: North-Holland;
1977. pp 89 –102.
2. Murofushi T, Sugeno M. Fuzzy measures and fuzzy integrals. In: Grabisch M, Murofushi
T, Sugeno M, editors. Fuzzy measures and integrals. Heidelberg: Physica-Verlag; 1999.
pp. 3– 41.
3. Zadeh LA. Fuzzy logic ⫽ computing with words. IEEE Trans Fuzzy Syst 1996;4:103–111.
4. Zadeh LA. From computing with numbers to computing with words—from manipu-
INDUCED FUZZY INTEGRAL AGGREGATION OPERATOR 1065
lation of measurements to manipulations of perceptions. IEEE Trans Circuits Syst
1999;45:105–119.
5. Yager RR. A new methodology for ordinal multiple aspect decisions based on fuzzy sets.
Decision Sci 1981;12:589 – 600.
6. Grabisch M, Murofushi T, Sugeno M. Fuzzy measures and integrals. Heidelberg: Physica-
Verlag; 1999.
7. Yager RR. A general approach to criteria aggregation using fuzzy measures. Int J Man-
Machine Stud 1993;38:187–213.
8. Zadeh LA. Fuzzy sets as a basis for a theory of possibility. Fuzzy Sets Syst 1978;1:3–28.
9. Dubois D, Prade H. Possibility theory: An approach to computerized processing of
uncertainty. New York: Plenum Press; 1988.
10. Zadeh LA. Fuzzy sets and information granularity. In: Gupta MM, Ragade RK, Yager RR,
editors. Advances in fuzzy set theory and applications. Amsterdam: North-Holland; 1979.
pp 3–18.
11. Sugeno M. Theory of fuzzy integrals and its application. Doctoral Thesis, Tokyo Institute
of Technology, 1974.
12. Grabisch M. Fuzzy integral in multicriteria decision making. Fuzzy Sets Syst 1995;69:
279 –298.
13. Marichal JL. On Sugeno integral as an aggregation function. Fuzzy Sets Syst 2000;114:
347–365.
14. Yager RR, Filev DP. Induced ordered weighted averaging operators. IEEE Trans Syst Man
Cybernet 1999;29:141–150.
15. Yager RR. Choquet aggregation using order inducing variables. Technical Report# MII-
2109 Machine Intelligence Institute, Iona College, New Rochelle, NY, 2001.
16. Yager RR. Quantifiers in the formation of multiple objective decision functions. Inf Sci
1983;31:107–139.
17. Yager RR. General multiple objective decision making and linguistically quantified state-
ments. Int J Man-Mach Stud 1984;21:389 – 400.
18. Sugeno M. Theory of fuzzy integrals and its application. Doctoral Thesis, Tokyo Institute
of Technology, 1974.