2

FOUNDATIONS OF
FUZZYAPPROACHES

2.1 FROM CRISP TO FUZZY SETS

The mathematical foundations of fuzzy logic rest in fuzzy set theory, which
can be thought of as a generalization of classical set theory. A familiarity with
the novel notions, notations, and operations of fuzzy sets is useful in studying
fuzzy logic principles and applications; acquiring it will be our main goal in
this chapter.
Fuzziness is a property of language. Its main source is the imprecision
involved in defining and using symbols. Consider, for example, the set of
chairs in a room. In set theory the set of chairs may be established by
pointing to every object in a room asking the question, Is it a chair? In
classical set theory we are allowed to use only two answers: Yes or No. Let us
code Yes as 1 and No as O.Thus, our answers will be in the pair {O,I}. If the
answer is 1, an element belongs to a set; if the answer is 0, it does not. In the
end we collect all the objects whose label is 1 and obtain the set of chairs in a
room. Suppose, however, that we ask the question, Which objects in a room
may function as a chair? Again we could point to every object and ask, Could
it function as a chair? The answer here too could artificially be restricted to
{O,1}. Yet, the set of objects in a room that may function as a chair may
include not only chairs but also desks, boxes, parts of the floor, and so on. It
is a set not uniquely defined. It all depends on what we mean by the word
function. Words like function have many shades of meaning and can be used
in many different ways. Their meaning and use may vary with different
persons, circumstances, and purposes; it depends on the specifics of a
situation. We say therefore that the set of objects that may function as a chair
is a fuzzy set, in the sense that we may not have crisply defined criteria for
13
14 FOUNDATlONS OF FUZZV APPROACHES FUZZV SETS 15

deciding membership into the set. Objects such as desks, boxes, and part of 2.2 FUZZY SErS
the fIoor may function as chairs, to a degree. It should be noted, however,
that there is nothing fuzzy about the material objects themselves: Chairs, As we saw in the previous section, in classical set theory there is a rather
boxes, and desks are what they are. Fuzziness is a feature of their representa- strict sense of membership to a set; that is, an element either belongs or does
tion i;Jl a milieu of symbols and is generally a property of models, computa- not belong to the set. In 1965 Lotfi A. Zadeh introduced fuzzy sets, where a
tional procedures, and language. more fIexible sense of membership is possible (Zadeh, 1965). In fuzzy sets
Let us now review some notions of classical set theory. Classical sets are
crisply defined collections of distinct elements (numbers, symbols, objects,
etc.), and for this reason we also call them crisp sets. The elements of all the
is indicated by a number between
[0, 1]. The point of departure
°
many degrees of membership are allowed. The degree of membership to a set
and 1-that is, a number in the interval
for fuzzy sets is simply generalizing the
sets under consideration in a given situation belong to an invariable, constant valuation set from the pair of numbers {O,I} to all numbers found in [0, 1]. By
set, called the universal set or universe or more often the universe of expanding the valuation set we alter the nature of the characteristic function,
discourse.: The fact that elements of a set A either belong or do not belong now called membership function and denoted by P-A(x)' We no longe r have
to a crisp set A can be formally indicated by the characteristic function of A, crisp sets but instead have fuzzy sets. Since the interval [0, 1] contains an
defined as infinity of numbers, infinite degrees of membership are possible. Thus, in

XA(X) == q iff
iff
x
x
E
(tE:
A
A
(2.1-1)
view of equation (2.1-2) we say that a membership function maps every element
of the universe of discourse X to the interval [0,1], and we formally write this
mapping as
where the symbols E and (tE: denote that x is and is not a member of A,
P-A(X): X ~ [0,1] (2.2-1)
respectively, and iff is shorthand for "if and only if." The pair of numbers
{O, I} is called the valuation set. Another way of writing equation (2.1-1) is
Equation (2.2-1) is a generalization ofthe mapping shown in equation (2.1-2).
Membership functions are a simple yet versatile mathematical tool for
XA(X): X ~ {O, I} (2.1-2)
indicating fIexible membership to a set and, as we shall see, for modeling and
The notation of equation (2.1-2) is read as follows: There exists a function quantifying the meaning of symbols. A question often asked by people
XA(X) mapping every element of the set X (our universe of discourse) to the set beginning the study of fuzzy sets is, How are membership functions found?
{0,1}. It emphasizes that the characteristic function is a mechanism for Membership functions may represent an individual's (subjectíve) notion of a
mapping the set X to the valuation set {0,1}. Important operations in crisp vague class-for example, objects in a room functioning as chairs, tall people,
sets such as union, intersection, and complementation are familiar to us from acceptable performance, small contribution to system stability, little improvement,
elementary mathematics. They are usually represented through Venn dia- big benefit, and so on. In designing and operating controllers or automatic
grams but may also be expressed in terms of the characteristic function. decision-making tools, for example, modeling such notions is a very impor-
Fundamentally, sets are categories. Defining suitable categories and using tant task. Membership functions may ais o be determined on the basis of
statistical data or through the aid of neural networks. In Part III of this book
operations for manipulating them is a major task of modeling and computa-
tion. From image recognition to measurement and control, the notion of we will look at the synergistic relation between neural networks and fuzzy
category, or set, is essential in the definition of system variables, parameters, logic toward this end (Kosko, 1992). At this point we can simply say that
their ranges, and their interactions. The constraint to have a dual degree of membership functions are primarily subjective in nature; this does not mean
membership to a set, an all-or-nothing, is a consequence of a desire to that they are assigned arbitrarily, but rather on the basis of application-specific
abstract a system description away from the multitude of intricacies and criteria (Kaufrnann, 1975; Dubois and Prade, 1980; Zimmermann, 1985).
complexities that exist in reality and focus on factors of primary infIuence. There are two commonly used ways of denoting fuzzy sets. If X is a
Nevertheless, given our modern-day computational technologies, it may be universe of discourse and x is a particular element of X, then a fuzzy set A
unduly restrictive. This is particularly the case when it is desired to develop defined on X may be written as a collection of ordered pairs
compute r models easily calibrated to the specifics of a system and endowed
with adaptive and self-organizing capabilities (Zadeh, 1973, 1988). A = {(x, p-Ax»)}, XEX (2.2-2)

IThe term universe of discourse is used in fuzzy logic; it comes from classical logic and describes where each pair (x, P-A(X)) is called a singleton and has x first, followed by its
the complete set of individual elements able to be referred to or quantified. membership in A, P-A(X). In crisp sets a singleton is simply the element x by
16 FOUNDATlONS OF FUZZV APPROACHES BASIC TERMS AND OPERATIONS 17

itself. In fuzzy sets a singleton is two things: x and pjx). For example, the J1,(x)
set of small integers, A, defined (subjectively) over the universe of discourse
of positive integers may be given by the collection of singletons
1 I....... :
I I: I
:
:
I
I

0.8 +·······················'\:····1························ "( j .......................•.•.••••• + .

A = {(1, 1.0), (2, 1.0), (3,0.75), (4,0.5), (5, O.i), (6,0.3), (7,0.1), (8, 0.1)}

Thus the fourth singleton from the left tells us that 4 belongs to A to a
0.6 :1,11-
j
: :j Ps(X) = --;-;-ri
l+(X): i
:

:::ll:'r:J
degree of 0.5. A singleton is also written as JLix)lx-that is, by putting
membership first, followed by the marker "I" separating it from x.2 Single-
tons whose membership to a fuzzy set is zero may be omitted. The support set
of a fuzzy set A is the set of its elements that have membership function
other than the trivial membership of zero.
An altemative notation, used more often than equation (2.2-2), explicit1y O • I I
I x
indicates a fuzzy as the union of alI JLix)lx singletons-that is, O 4 8 12 16 20
Figure 2.1 Zadeh diagram for the fuzzy set B = {small numbers}.

A = L: JLA(xJlx i (2.2-3)
XiEX
membership function JLix) given by

The summation sign in equation (2.2-3) indicates the union of alI singletons 1
(the union operation in set theory is like "addition". Equation (2.2-3) JLB(X) = (X)3 (2.2-5)
assumes that we have a discrete uniuerse of discourse. In this altemative 1+ -
notation the set of small integers above may be written as 5

Using the form of equation (2.2-4) the fuzzy set B is written as

A = JLA(1)/1 + JLA(2)/2 + JLA(3)/3 + JLA(4)/4 + JLA(5)/5
+ JLA(6)/6 + JLA(7)/7 + JLA(8)/8
= 1.0/1 + 1.0/2 + 0.7513 + 0.5/4 + 0.315 + 0.316 + 0.1/7 + 0.1/8
B = fX2:0
JLB(x)lx = f [ 1 + (xr1lX
X2:0 _
(2.2-6)

For a continuous universe of discourse, we write equation (2.2-3) as The membership function of fuzzy set B is shown in Figure 2.1.4 A graph like
this is called a Zadeh diagram.

A = fx JLA(x)lx (2.2-4)
2.3 BASIC TERMS ANO OPERATlONS

where the integral sign in equation (2.2-4) indicates the union of alI JLix)lx Many fuzzy set operations such as intersection and union are defined through
singletons." Consider, for example, the fuzzy set small numbers defined the min (1\) and max (v) operators. Min and max are analogous to product
(subjectively) over the set of non-negative real numbers through a continuous (-) and sum (+ ) in algebra (Dubois and Prade, 1980; Klir and Folger, 1988;
Terano et al., 1992). Let us take a look at how they are used.
2It should be noted that "/" does not indicate "division"; it is merely a marker.
3Note that the integral sign is not the same as the integral sign of differential and integral 4Fuzzy sets are sometimes called fuzzy subsets, reflecting the fact that they are subsets of a larger
calculus. It is used here in the sense that the integral sign is used in set theory-that is, to set-that is, the universe of discourse. A1though the term fuzzy subsets is factually correct, we will
indicate the sum or union of individual singletons. use the standard term fuzzy set for convenience.
18 FOUNOATIONS OF FUZZV APPROACHES
BASIC TERMS ANO OPERATIONS 19

First, min (1\) and max (v) may be used to select the minimum and min and max and the operations of muliiplication and addition. They both
maximum of two elements-for example, 2 1\ 3 = 2, or 2 V 3 = 3. We also have the same properties of associativity and distributivity, and thus in
write min(2,3) = 2, or max(2,3) = 3. Formally, the minimum of two ele- equations that involve min and max we may employ them in the same manner
ments J.LI and J.L2 denoted either as mine J.LI' J.L2)' 1\ (J.LI> J.L2)' or J.LI 1\ J.L2 is as multiplication (-) and addition ( +). We will see an interesting example of
defined as these properties in the composition of fuzzy relations (Chapter 3), where we
treat composition as matrix multiplication with (1\) and (v) in place of
iff J.LI s J.L2 product (-) and sum ( + ).
J.LI 1\ J.L2 = mine J.Lp J.L2) == (J.LI (2.3-1) Min (1\) and max (v) can ais o operate on a collection of sets as for
J.L2 iff J.LI > J.L2
example in
where, the " == "symbol means "by definition" and iff is shorthand for" if and
only if." Similarly the maximum of two elements J.LI and J.L2' denoted as A = I\(ApA2, ... ,Am) (2.3-8)
max( J.LI' J.L2) or J.LI V J.L2' is defined as
which can be succinctly written as
J.LI iff J.LI ~ J.L2
J.LI V J.L2 = max( J.LI' J.L2) == (2.3-2)
( J.L2 iff J.LI < J.L2 m
A = A (Ak) (2.3-9)
k~l
Second, min (1\) and max (v) may operate on an entire set, selecting the
least element (called infimum in mathematical analysis) or the greatest ele-
ment (called supremum) of the set. For example, 1\ (0.01,0.33,0.44,0.999) = Using primarily min (1\) and max (V), a number of useful notions and
0.01 and V (0.01,0.33,0.44,0.999) = 0.999. Formally we write this as operations involving fuzzy sets can be defined.?

J.L = I\A = inf A (2.3-3) Empty Fuzzy Set

and A fuzzy set A is called empty (denoted as A = 0) if its membership function
is zero everywhere in its universe of discourse X -that is,
J.L = V A = sup A (2.3-4 )

where J.L is an elementof A-that is, J.L E A. A==0 if J.LA(X) = O, Vx E X (2.3-10)

In addition, min (1\) and max (v) may be used as functions operating on
single elements or on entire sets, for example, to find the smallest element J.L where "Vx E X" is shorthand notation indicating "for any element x in X."
out of a list of elements (J.LI' J.L2"'" J.Lm)-that is,

J.L = 1\ ( J.Lp J.L2' ... , J.Lm) (2.3-5) Normal Fuzzy Set

A fuzzy set is called normal if there is at least one element Xo in the universe
which is the same as
of discourse where its membership function equals one-that is,
J.L = J.LI 1\ J.L2 1\ ... 1\ J.Lm (2.3-6)
J.LA( xo) = 1 (2.3-11)
We sometimes use a shorthand notation for equations (2.3-5) and (2.3-6) and
write them as
More than one element in the universe of discourse can satisfy equation
m (2.3-11).6
J.L = A (J.Lk) (2.3-7)
k~l 5These operations can also be defined in terms of T-nonns (see Appendix),
6It should be noted that the term normal does not refer to the area under the curve of the
This notation is analogous to finite product notation in algebra (or finite membership function. It simply means what the definition says: At least one point, maybe more,
summation when V is used). There is in fact a more general analogy between needs to have full membership value.
20 FOUNDATIONS OF FUZZV APPROACHES BASIC TERMS AND OPERATIONS 21

Equality 01 Fuzzy Sets the algebraic product of the membership functions of A and B,
Two fuzzy sets are said to be equal if their membership functions are equal J-LA·B(X) == J-LAx) . J-LB(X) (2.3-16)
everywhere in the universe of discourse-that is,
The product of two fuzzy sets can be generalized to any number of fuzzy sets
A==B if J-LA(X) = J-LB(X) (2.3-12)
on the same universe of discourse.

Union 01 Two Fuzzy Sets Multiplying a Fuzzy Set by a Crisp Number

The union of two fuzzy sets A and B defined over the same universe of We can multiply the membership function of a fuzzy set A by the crisp
discourse X is a new fuzzy set A u B also on X, with membership function number a to obtain a new fuzzy set called product a . A. Its membership
which is the maximum of the grades of membership of every x to A and function is
B-that is,
J-LaA(X) == a· J-LA(X) (2.3-17)
J-LAUB(X) == J-LA(X) V J-LB(X) (2.3-13)
The operations of multiplication and raising a fuzzy set to a power that we
The union of two fuzzy sets is related to the logical operation of disjunc- see next are useful for modifying the meaning of linguistic terms (Zadeh,
tion (OR) in fuzzy logic. Equation (2.3-13) can be generalized to any number 1975).
of fuzzy sets over the same universe of discourse.
Power 01 a Fuzzy Set
Intersection 01 Fuzzy Sets
We can raise fuzzy set A to a power a (positive real number) by raising its
The intersection of two fuzzy sets A and B is a new fuzzy set A íI B with membership function to a. The a power of A is a new fuzzy set, Aa, with
membership function which is the minimum of the grades of every x in X to membership function
the sets A and B, i.e.,
J-LAa(X) == [J-LA(X)r (2.3-18)
J-LAnB(X) == J-LA(X) 1\ J-LB(X) (2.3-14)
Raising a fuzzy set to the second power is usually taken to be equivalent to
The intersection of two fuzzy sets is related to conjunction (AND) in fuzzy linguistically changing it through the modifier VERY (Zadeh, 1983) (see
logic. The definition of intersection in (2.3-14) can be generalized to any Chapter 5). Thus the square of the membership function of B = {small
number of fuzzy sets over the same universe of discourse. numbers} in Figure 2.1 is taken to represent the fuzzy set B2 = {VERY small
numbers}.
Complement 01 a Fuzzy Set
Raising a fuzzy set to the second power is a particularly useful operation
and therefore has its own name. It is called concentration or CON. Taking
The complement of a fuzzy set A is a new fuzzy set, Ã, with membership the square root of a fuzzy set is called dilation or D/L (an operation useful
function for representing analytically the linguistic modifier MORE OR LESS).

J-L,.r(X) == 1 - J-LA(X) (2.3-15) Example 2.1 Union, Intersection, and Complement of Fuzzy Sets. Consider
the Zadeh diagram of fuzzy sets A and B shown in Figure 2.2a and defined
Fuzzy set complementation is equivalent to negation (NOT) in fuzzy logic. by membership functions

Product 01 Two Fuzzy Sets 1 1

J-LA(X) = 1 + O.3(x _ 8)2 and J-LB(X) = (3 (E2.1-1)
The product of two fuzzy sets A and B defined on the same universe of
1+ ~)
discourse X is a new fuzzy set, A . B, with membership function that equals
22 FOUNDATIONS OF FUZZV APPROACHES
BASIC TERMS AND OPERATIONS 23

1
discourse to either A or B in accordance with equation (2.3-13). Figure 2.2b
J.l(x) J-l•• (x) = 1+ 0.3 (x _ 8)2 shows the membership function of the union A u B. The interpretation of
A u B is "about 8 OR small number." Similarly the membership function of
the intersection of fuzzy sets A and B, shown in Figure 2.3a, represents the
new fuzzy set "about 8 AND small number." We observe that although the
0.84·····················'\,1.': union of A and B is a normal fuzzy set, the intersection shown in Figure
2.3a is not, because fuzzy set A n B has no point in the universe of

0.4
p(x)

". ::\
0.2
.: : \

O t=:= i
I
:
I
--I --. i
~x 0.8

o 4 8 12 16 20

0.6 •.• ············· .. ······

(a)
0.4
J.I.(x)

0.2

0.84····················:'\.· o: t::::= i
:
i
--. :
.
""..
.•.•.• ,......
i ·.h
L x

0.6
+\~,,~~.(~! o 4 8

(a)
12 16 20

J.I.(x)

o.•
0.2
4·.............::.~><J\r··················L . 0.8 J............ ..\
.
1 ~ ..•.. , .•.. j"

J....... ..
-» , ••~'

.. .. '. :

o
'
.
: : ~ .tu ..... 'i n-l- \.
I I i
x 0.6
......
........A·(~)'
o 4 8 12 16 20

(b)

Figure 2.2 Zadeh diagram for (o)

Exampie 2.1.
fuuy sets A and B and (b) their union in
:::-=~r:"><.,bcC.~~(~~mJm mmm.
Fuzzy set A may be thought of as defining the set of numbers "about 8," and i ; ,
x
fuzzy set B may be thought of as defining "small numbers." We take numbers o 4 8 12 16 20
between O and 20 to be the universe of discourse and, would like to find the
union and intersection of A and B and the complement of B. (b)
The membership function of the union of fuzzy sets A and B is the Figure 2.3 Zadeh diagram for (o) the intersection of fuuy sets A and B and (b) the
maximum grade of membership of each element x of the universe of Complement of B in Example 2.1.
24 FOUNDATIONS OF FUZZV APPROACHES BASIC TERMS AND OPERATIONS 25

discourse with grade of membership equal to 1. The complement of fuzzy set p(x)
B is a new fuzzy set with membership function given by equation (2.3-15).
Figure 2.3b shows the membership function of the complement B. The
complement B represents the logical negation (NOT) of B-that is, the set 0.8
"NOT small numbers." 07
0.6

Concentration
0.4
The concentration of a fuzzy set A defined over a universe of discourse, X,
is denoted as CON(A) and it is a new fuzzy set with membership function 0.2
given by
o-I
o
i =::t =r= : ~ x

I-LCON(A)(X) == (I-LA(X))2 (2.3-19) 4 8 12 16 20

Figure 2.4 The fuzzy sets VERY small numbers and MORE OR LESSsmall numbers
obtained by concentrating and di/ating the fuzzy set small numbers.
As we said in the previous paragraph, squaring or concentrating a fuzzy set is
equivalent to linguistically modifying it by the term VERY. Figure 2.4 shows
the concentration operation applied to the fuzzy set B = {small numbers}.
The membership function of the new fuzzy set CON(B) = B2 = {VERY small Contrast Intensification
numbers} is
In certain applications it is desirable to control the fuzziness of a fuzzy set A
by modifying the contrast between low and high grades of membership. For
1
2 instance, we may want to increase the membership function on that part of A
I'CON{B)(X) ~ (I'.(x)) ~ [1 + (~)T where membership values are higher than 0.5, and. decrease it for values
lower than 0.5. We define the contrast intensification of A as

I-LINT(A)(X) == 2[ I-LA(X)]2 for 0~I-LA(x)~0.5

Dilation (2.3-21)

The dilation of a fuzzy set A, denoted as DIL(A), produces a new fuzzy set I-LINT(A)(X) == 1 - 2[1 - I-LA(X)]2, for 0.5 ~ I-LA(x) s 1.0
in X, with membership function defined as the square root of the member-
ship function of A-that is,
Contrast intensification may be repeatedly applied to a fuzzy set. In the
extreme, when the maxirnum possible contrast is achieved we no longer have
I-LDIL(A)(X) == VI-LA(X) (2.3-20) a fuzzy set. We are back to a crisp set. The opposite effect-that is, going
from a crisp set to fuzzy set-may be achieved through fuzzification.
Dilation (DIL) and concentration (CON) are operations with opposing ef-
fects. Concentrating a fuzzy set reduces its fuzziness while dilating it in-
creases its fuzziness. The dilation operation corresponds to linguistically Fuzzyfication
modifying the meaning of a fuzzy set by the term "MORE OR LESS." Figure
2.4 shows the dilation of B = {small numbers}, resulting in a new fuzzy set Fuzzification is used to transforrn a crisp set into a fuzzy set or simply to
DIL(B) = Bl/2 = {MORE OR LESS small numbersi. increase the fuzziness of a fuzzy set. For fuzzification we use a fuzzyfier
function F that controls the fuzziness of a set. F may be one or more simple
7Here and throughout this book, the end of an example is indicated by the symbol "D.u parameters. For instance, consider the fuzzy set A that describes large
 26 FOUNDATlONS OF FUZZVAPPROACHES BASIC TERMSAND OPERATIONS 27

numbers. We define it (subjectively) through the membership function J.lA(X)

1
(2.3-22)
ILlarge Ilumbers( x) = - (X) -FI
0.8
1+ -
F2
0.7

where x is any positive real number. The membership function in equation 0.5

(2.3-22) has two fuzzifying parameters: an exponential fuzzyfier, Fi' and a

0.3
denominationalfuzzyfier, F2• Through them the fuzzy set A = {large numbers}
can be written as . 0.2

(*r}
O x
O 10 20 30 40 50 60 70 80 90 100
(2.3-23)
A ~ Ix [1 + Figure 2.6 The effect of varying the denominational fuuifier F2 while keeping the
exponential fuuifier F, constant in fuuifying the set A.

The membership function inside the brackets of equation (2.3-23) can be tial fuzzyfier at Fi = 4 is shown in Figure 2.6. Varying F2 results primarily in
adjusted when needed in order to better represent the meaning of the term translating the membership function left and right, and to a lesser extent it
large numbers. Consider the case when we fix the value of denominational affects the fuzziness of A. Such fuzzifiers are often used in fuzzy pattern
fuzzifier as F2 = 50 and vary the exponential fuzzyfier Fi' The result is a recognition and image analysis in defining, for instance, the meaning of the
family of fuzzy sets with decreasing fuzziness as Fi increases. Figure 2.5 words vertical, horizontal, and oblique lines (Pai and Majumder, 1986).
shows membership functions that result from such a variation. Note that Fuzzification may be used more systematically by associating a fuzzyfier F
when Fi becomes very large, the set A appears almost like a crisp set. The with another function, namely a fuzzy kemel, Kt;x), which is the fuzzy set that
effect of varying the denominational fuzzyfier F2 while keeping the exponen- results from the application of F to a singleton x. This is often done in
control applications where the input to an on-line control or diagnostic
system comes from sensors and is therefore crisp, usually a real number. In
J.1(X) order to use it in fuzzy algorithms (see Chapters 5 and 6), it is often necessary
to convert a crisp number to a fuzzy set, a step known as fuzzification. As a
I (Ff=4. Fr50)
result of the application of K to a fuzzy set A, we have
0.8

0.7

0.5
~=::l=-l~+FI-, .
•••••••••••••••••••• : ••••••••••••••••••••

!
t- ••••••••••••••••••••••••••••••••••••••••

! i:
:............ •• •••••••••••••••••••••••••••••••••••

!!
i 1
: •••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••

I !
(Ff=2.

(Ff=f.
Fr50)

Fr50)
F(A; K) = fx ILA(X) . ILK(xix)/x

where F(A; K) is a fuzzy set that results from changing the fuzziness of A in
accordance with K. The fuzzy kemel Ki x) is simply a fuzzy set imposed on a
singleton. It functions as a "mask" that covers the singleton to produce a
(2.3-24 )

t--1-1-1---t-
0.3 •• •••• ····, •• • •••• ·O: •••••••••• • •••••••• ·c·..·.. ..... ".,,, ... ,,...•.................... :....................•.... ,,.............••....................
fuzzy set. For example, suppose that we have the universe of discourse
X = {l,2, 3, 4, 5, 6, 7, 8, 9, iO} and a fuzzy kernel Kt.x) that centers a triangu-
0.2
i '- lar fuzzy set around 5 given by
III1 o x
o 10 20 30 40 50 60 70 80 90 /00 K(5) = 0.33/3 + 0.67/4 + 1.0/5 + 0.67/6 + 0.33/7 (2.3-25)

Figure 2.5 The effect of varying the exponential fuuifier F, while keeping the denomi- with all other elements of the universe of discourse having trivial (zero)
national fuuifier F2 constant in fuuifying the set A. membership. Now suppose that we have the value of 3, which may be a crisp
28 FOUNDATlONS OF FUZZV APPROACHES
PROPERTIES OF FUZZV SETS 29

measurement taken at a certain time. We write it as a singleton A given by In fuzzy sets all these properties can be expressed using the membership
function of the sets involved and the definitions of union, intersection, and
A = JLA(3)/3 = 1.0/3 (2.3-26)
complemento For example, consider the associa tive property given by equa-
We fuzzify A using equation (2.3-24) as follows: tions (2.4-4). ln terms of membership functions the associative property is
written as
F(A; K) = fx JLA(X) . JLK(x)(X)/x (JLA(X) V JLB(X)) V JLc(x) = JLA(X) V (JLB(X) V JLcCx))
= f [JLA(3)· J.LK(3)(X)]/x (J.LA(X) 1\ JLB(X)) 1\ JLc(x) = J.LA(X) 1\ (JLB(X) 1\ J.Lc(x))
x
= 0.33/1 + 0.67/2 + 1.0/3 + 0.67/4 + 0.33/5 (2.3-27) Similarly, the distributive property, equations (2.4-5), in terms of membership
functions is written as
which results in shifting the fuzzy kernel of (2.3-25) so that its peak is located
at the singleton '3'. ln other words, the effect of equations (2.3-27) is to mask
J.LA(X) V (JLB(X) 1\ JLc(x)) = (J.LA(X) V J.LB(X)) 1\ (J.LA(X) V J.Lc(x))
the crisp value '3' by the fuzzy set K(5), shifting its peak from '5' to '3'.
J.LA(X) 1\ (JLB(X) V JLc(x)) = (J.LA(X) 1\ J.LB(X)) V (J.LA(X) 1\ J.Lc(x))
2.4 PROPERTIES OF FUZZY SETS De Morgan's law, equation (2.4-7), is written as

Fuzzy set properties are useful in performing operations involving member- JLA(X) V JLB(X) = JLÃ(x) 1\ JL]j(x)
ship functions. The properties we list here are valid for crisp and fuzzy sets as
well, but some of them are specific to fuzzy sets only; more detailed where the bar over the membership functions indicates that we take the
treatment of properties may be found in Dubois and Prade (1980) and in Klir complement. De Morgan's law says that the intersection of the complement
and Folger (1988). Consider sets A, B, C defined over a common universe of of two fuzzy sets equals the complement of their union; in terms of member-
discourse X. We indicate the complement of a set by a bar over it. The ship functions, this is the same as saying that the minimum of two member-
following properties are true: ship functions equals the complement of their maximum. There are also
some properties generally not valid for fuzzy sets (although valid in crisp
Double Negation Law: (..:4) =A (2.4-1 )
sets), such as the law of contradiction,
A uA =A
Idempotency: (2.4-2) AnÃ,=0 (2.4-8)
AnA=A
AnB=BnA and the law of the excluded middle,
Commutativity: (2.4-3)
AuB=BuA AuÃ,=X (2.4-9)
(A U B) uC = A u (B u C)
Associative Property: (2.4-4) The law of the exc1uded middle in crisp sets states that the union of a set
(A n B) n C = A n (B n C) with its complement results in the universe of discourse. This is generally not
true in fuzzy sets. A property unique to fuzzy sets is
A U (B n C) = (A u B) n (A u C)
Distributive Property: (2.4-5)
A n (B u C) = (A n B) U (A n C) An0=0 (2.4-10)

An(AUB)=A Equation (2.4-10) says that the intersection of a fuzzy set with the empty set
Absorption: (2.4-6) -that is, a set with a membership function equal to zero everywhere on the
AU(AnB)=A
universe of discourse-is also the empty set. ln terms of membership
A UB=ÃnB
functions equation (2.4-10) is written as
De Morgan's Laws: (2.4-7)
AnB=ÃuB JLA(X) 1\ O = O
 lHE EXlENSION PRINCIPLE 31
30 FOUNDAllONS OF FUZZV APPROACHES

Also, the union of a fuzzy set A with the empty set, 0, is A itself; that is, where every single image of Xi under f-that is, Yi = f(x)-becomes fuzzy
,A U 0 = A or, equivalently, P-A(X) V O = P-A(X), The intersection of a fuzzy to a degree p-ix). Recalling that functions are generalIy many-to-one
set A with the universe of discourse is the fuzzy set A itself; that is, mappings, it is conceivable that several x's may map to the same y. Thus for
A n X = A or, equivalently, p-ix) Â 1 = p-ix). The union of a fuzzy set A a certain Yowe may have more than one x: Let us say that both x2 and x13
with the universe of discourse X is the universe of discourse; that is, in (2.5-1) are mapping to Yo' 'Hence, we have to decide which of the two
A U X = X, which, in terms of the membership function, is written as membership values, P-A(X2) or p-ix13), we should take as the membership
P-A(X) V 1 = 1. The universe of discourse may be viewed as a fuzzy set whose value of Yo' The extension principle says that the maximum of the member-
membership function equals 1 everywhere; that is, P-x(x) = 1 for ali x in X. ship values of these elements in the fuzzy set A ought to be chosen as the
grade of membership of Yo to the set B-that is,

2.5 THE EXTENSION PRINCIPLE P-B(YO) = P-A(X2) V P-A(X13) (2.5-4 )

While fuzzification operations such as the ones we saw in Section 2.3 are If, on the other hand, no element X in X is mapped to Yo-that is, no
useful for fuzzifying individual sets or singIetons, more general mathematical inverse image of Yo exists-then the membership value of the set B at Yo ís
expressions may also be fuzzified when the quantities they involve are zero. 'Having accounted for these two special cases (rnany x's mapping to the
fuzzyfied. For example, the output of arithmetic operations when their same Y and no inverse image for a certain y), we can compute the set
arguments are fuzzy sets becomes also a fuzzy quantity. The extension B-that is, the grades of membership of eIements y in Y produced by the
principie is a mathematical tool for extending crisp mathematical notions and mapping f(A)-using equation (2.5-3).
operations to the milieu of fuzziness. It provides the theoretical warranty that In a more general case where we have several variables, u, v, ... , w, from
fuzzifying the parameters ar arguments of a function results in computable different universes of discourse U, V, ... ,W and m different fuzzy sets
fuzzy sets. It is an important principle, and we will use it on several AI' A2, ... , Am defined on the product space U X V x ... x W, the multi-
occasions, particularly in conjunction with fuzzy relations (Chapter 3) and variable function, y = f(u, v, ... , w), may also be used to fuzzify the space Y
fuzzy arithmetic (Chapter 4). We give here an informal heuristic description through the extension principie. In this case, the grade of membership of any
of the extension principle; detailed formulations may be found in (Zadeh y equals the minimum of the membership values of u, v, ... ,w in
(1975), and in Dubois and Prade (1980). AI' A2, ... , Am, respectively. The membership function of B is given by
Suppose that we have a function f that maps elements xl> x2, ... ,XII of a
universe of discourse X to another universe of discourse Y -that is,
YI = f(XI)
P-B(y) = f
UxVx 000 xY
[P-A (U)
1
 P-A2(V) Â'"  P-A,.cw)]/f(u,v, ... ,w)

(2.5-5)
Y2 = f(x2)
(2.5-1)
where there is also a max (v) operation implicit in the union operatiofi [the
integral sign in equation (2.5-5) indicates a union (v) operation], The max
YII = f(xlI)
operation is performed over all u, v, ... , w such that y = f(u, v, ... , w). This
Now suppose that we have a fuzzy set A defined on Xl' X2' X3' ... ,XII (the is indicated by the union over the product space U X V x ... x W of all the
input to the function f). A is given by universes Ou which the m-tuples u, v, ... ,w are defined under the integral
signo If the inverse image does not exist, then the membership function is
A = P-A(XI)/XI + P-A(X2)/X2 + ... + P-A(XII)/XII (2.5-2)
simply zero.
We then ask the question, If the input to our function f becomes fuzzy-for In many engineering applications, the interpretation of numerical data
example, the set A of equation (2.5-2)-what happens to the output? Is the may not be precisely known. We consider this type of data to be fuzzy. Using
output also fuzzy? In other words, is there an output fuzzy set B that can be the extension principie, it is quite possible to adapt ordinary algorithms,
computed by inputting A to f. WelI, the extension principle tells us that which are used with precise data, to the case where the data are fuzzy.
there is indeed such an output fuzzy set B and that it is given by Example 2.2 is a mathematical illustration of the extension principle.
B = f(A) = P-A(xl)/f(xl) + P-A(x2)/f(x2) + ... +p-AxlI)/f(xlI)
Example 2.2 Using the Extension Principie. As an illustration of how the
(2.5-3) extension principle may be used, consider the function f that maps points
32 FOUNDATIONS OF FUZZV APPROACHES THE EXTENSION PRINCIPLE 33

from the x axis to Y axis in the Cartesian plane according to the equation y
b=l

Y = f(x) = V=>
1 - 4" (E2.2-1)

Figure 2.7a shows the function y of equation (E2.2-l). It is the upper half of x
an ellipse located on the center of the plane with major axis, a = 2, and
minor axis (height), b = 1. The general equation of the ellipse shown in
Figure 2.7a is
(a)
x2 Y2
(E2.2-2) J1(x)
a2 + b2 = 1
1
In our case with a = 2 and b = 1, equation (E2.2-2) becomes

2
x
_4
+ Y2 = 1 (E2.2-3)
-2
(b) 2 x
Equation (E2.2-l) is one of the two solutions of equation (E2.2-3).
Now suppose that we define a fuzzy set A on X as shown in Figure 2.7b: J1(y)
We fuzzify the x's of equation (E2.2-l) by specifying a grade of membership
ILix) for each x to fuzzy set A-that is, ILA(X) = ~Ixl and ................. ~1 _ .

A = f -2::s;x::S;2
Dlxl]/x (E2.2-4)

where Ixl is the absolute value of x, and we limit the support of A between -I 1 y
- 2 and + 2 as indicated by the limits under the integration sign iunion) of
equation (E2.2-4).
Having the x values fuzzyfied by the fuzzy set A, we want to know the
effect of fuzzification on y. The extension principie tells us that the fuzziness
of A will be extended to y as well. In other words, we will have a fuzzy set B
on Y derived by equations (2.5-3) or (2.5-5). To avoid the case where more
than one x will map to the same y, we consider first the function f in the
first quadrant of the plane (where both x and y are positive). Later we will 2
x
look at the entire function. The fuzzy set, B, defined on Y is

B = f(A) = 1 y
ILB(Y)/Y (E2.2-5)
(d)

We need to find ILB(Y) in equation (E2.2-5). In terms of the membership Figure 2.7 Graphs for Example 2.2. (o) The function y, which is the upper part of the
function of A and according to the extension princíple, equation (2.5-3), the ellipse shown. (o) The membership function of the set A (c) The membership function
of B. (d) Fuzzifying the interior of the ellipse.
 FOUNDATIONS OF FUZZVAPPROACHES ALPHA-CUTS 35
34

set B will be The o-cut of a fuzzy set A denoted as Aa is the crisp set comprised of all
the elements x of a universe of discourse X for which the membership
B = f(A) = i JLAx)/f(x)
y
(E2.2-6) function of A is greater than ar equal to a; that is,

Of course we want to transforrn the x variable to y in equation (E2.2-6) since Aa = {x EXI JLA(X) ~ a} (2.6-1 )
the union (integration) is formed with respect to Y, the universe of discourse
for B. We use equation (E2.2-1) to solve for x: where a is a parameter in the range O < a ~ 1; the vertical bar "I" in
equation (2.6-1) is shorthand for "such that."
x=2~ (E2.2-7)
Consider, for example, a fuzzy set A with trapezoidal membership func-
tion as shown in Figure 2.8. The O.5-cut of A is simply the part of its support
Then we substitute (E2.2-7) in (E2.2-6), noting that f(x) = y and that JLix) where its membership function is greater than 0.5. In Figure 2.8 we can see
is given by (E2.2-4). Thus we obtain the fuzzy set B: the 0.5-cut of A. Reflecting the fact that the o-cut is a crisp set, its
membership function appears like a characteristic function. As another
B= 1 ~/y
Osysl
(E2.2-8) example consider the set A of small integers given by

Now if we consider nega tive values for x as well, we would have to take the A = 1.0/1 + 1.0/2 + 0.75/3 + 0.5/4 + 0.3/5 + 0.3/6 + 0.1/7 + 0.1/8
maximum of the membership value of A at (x) and (-x) in accordance with
equation (2.5-5). Due to the symmetry of the problem these values are The O.5-cut of A is simply the crisp set AO.5 = {l, 2, 3, 4}.
actualIy the same and therefore B is still as derived in (E2.2-8). The In the next section we will see tha.t o-cuts provide a useful way both for
membership function of Bis, resolving a membership function in terms of constituent crisp sets as well as
for synthesizing a membership function out of crisp sets.
JLB(Y) (E2.2-9)
= ~ A fuzzy set can have an extensive support since its membership function
can be zero or near1y zero, or very smalI. In order to deai with situations
as shown in Figure 2.7c. Figure 2.7d shows the geometric interpretation of where smalI degrees of membership are not worthy of consideration, levei
fuzzyfying the interior of the ellipse in accordance with the fuzzy sets A and
B above. The result is a kind of fuzzy elliptic region, strongest near the x axis
and particular1y at its x = ± 2 sides and weakest near the origin and the
Y = ± 1 sides. O ,LIA O.S (x)
JL (x)

2.6 ALPHA-CUTS

With any fuzzy set A we can associa te a colIection of crisp sets known as
a-cuts (alpha-cuts) or levei sets of A. An o-cut is a crisp set consisting of
0.8
···········..·········r....······..····

· ·r·····, , ,
elements of A which belong to the fuzzy set at least to a degree a. As we
shalI see in the next section, o-cuts offer a method for resolving any fuzzy set 0.5

in terms of constituent crisp sets (something analogous to resolving a vector

into its components). In Chapter 4 we will see that a-cuts are indispensable
in performing arithmetic operations with fuzzy sets that represent various 0.3 .... , , ~
:
.:.
:
~
:
.:,
:
.
qualities of numerical data. It should be noted that o-cuts are crisp, not ~ ~ ~ ~
fuzzy, sets."
o
~ ~ 1 1
8 Formally,a distinction is made between two types of o-cuts, the strong and the weak o-cut 6 8 10
Figure 2.8 A fuzzy set A and its O.5-cuts. x
(Dubois and Prade, 1980). We use the weak o-cut, simply calling it o-cut,
36 FOUNDATIONS OF FUZZV APPROACHES THE RESOLUTION PRINCIPLE 37

}1(X) 2.7 THE RESOLUTlON PRINCIPLE

I There are several ways of representing fuzzy sets, and we have already seen a
0.8 ~ . few of them. They all involve two things: identifying a suitable universe of
discourse and defining membership functions. One way to represent a fuzzy
set would be to list all the elements of the universe of discourse together with
the grade of membership of each element (omitting the possibly infinite
elements that have zero membership), Alternatively, we can just provide an
analytical representation of the membership function. The resolution principIe
offers another way of representing membership to a fuzzy set, namely
0.2
through its o-cuts. It asserts that the membership function of a fuzzy set A
O
!""''''.''--
!
:
!
':
1
:
1
,;
i
-I
~
x
can be expressed in terms of its o-cuts as follows:
o 20 40 60 80 100
Figure 2.9 The O.2-levelfuzzy set of fuzzyset A.
J-LA(X)= V [a·J-LA.(x)] (2.7-1)
O<a,,;!

where the maximum is taken over all a's. Equation (2.7-1) indicates that the
fuzzy sets were introduced to exc1ude undesirable grades of membership membership function of A is the union (notíce the max operator) of all
(Radecki, 1977). We define the level fuzzy sets of a fuzzy set A as fuzzy sets e-cuts, after each one of them has been multiplied by a.
Ãa whose membership values are greater than a, where O < a < 1. Formally Consider, for example, the fuzzy set A with triangular membership
function shown in Figure 2.10. Several o-cuts of A, each multiplied by a, are
also shown. Knowing many a 's and the a-cuts of A, we can form their
Ãa = {(x, J-LA(x))lx EAa} (2.6-2) products and put them together (in the sense of taking their uníon) to
approximate the function. For example, we multiply the 0.25-cut by 0.25 to
where Aa is the o-cut of A. Equation (2.6-2) indicates that for a given a we get the 0.25-cut pushed down to 0.25, and similarly we multiply the O.5-cut by
have a level fuzzy set which is the part of A that has membership greater 0.5, the 0.75-cut by 0.75, and so on. When put together we have an
than a. Let us consider, for example, a fuzzy set A whose membership approximation of the membership function of A as shown in Figure 2.10.
function is
I-I(X)
1
(2.6-3) I i _........ - .. - I-IA(x)
J-LA(X) = 1 + O.OI(x _ 50)2 - .

0.75 r·············· _ f \
as shown in Figure 2.9 (dotted curve). Suppose that we are not interested in
the part of the support that has membership less than 0.2. We obtain the
0.5 ,·························1 I i \
0.2-level fuzzy set of A by chopping the part of the membership function
which is less than 0.2 as shown in the figure. Its membership function J-LÃ (x) 0.2
0.25 t············ .... ·····./ I I I I \ O.25·I-IM.25 (x)
is shown by the solid curve. It is the same as J-Lix) between x = 30 and
x = 70 and zero everywhere else. Level fuzzy sets should not be confused
with level sets, which is a synonym for o-cuts. Level fuzzy sets are indeed
o
fuzzy sets, whereas o-cuts are crisp sets. They provide a useful way of x
considering fuzzy sets in the significant part of their support, and hence they Figure 2.10 Putting many a-cuts of A multiplied by a together approximqtes the
save on computing time and storage requirements. membership function of A.
38 FOUNDATIONS OF FUZZV APPROACHES POSSIBILlTY THEORY AND FUZZV PROBABILlTIES 39

Thus, a large enough family of a-cuts provides another way of representing a . failure rate and error rate requires a large amount of data, which is often not
fuzzy ser. Although we often know the membership function exactly, in some practically possible to obtain. Is is obviously impractical to melt nuclear
applications only a-cuts are known and out of them we need to approximate reactors to collect failure rate data. Thus, in practice, the failure rate and
the membership function (see Chapter 4). error rate are estimated by experts based on their engineering judgment
(Onisawa, 1990); from this point of view, fuzzy possibilities and probabilities
(which we will examine momentarily) can be used to model such judgments in
2.8 POSSIBILlTY THEORV ANO FUZZV PROBABILlTIES a flexible and efficient way. Engineering judgment enters many are as of
systems and reliability analysis including estimating the effect of environmen-
In the late 1970s Zadeh advanced a theoretical framework for information tal factors, operator stress, dependence between functions or units, selection
and knowledge analysis, called possibility theory, emphasizing the quantifica- of sequence of events, expressing the degree of uncertainty involved in the
tion of the semantic, context-dependent nature of symbols-that is, meaning formulation of safety criteria, assuming parameter ranges, and so on
rather than measures of information. The theory of possibility is analogous, (Shinohara, 1976). Alternatives to failure and error rates have been devel-
and yet conceptually different from the theory of probability. Probability oped employing the notion of possibility measures, called failure and error
is fundamentally a measure of the frequency of occurrence of an event. possibilities, and have been applied to the reliability analysis of nuclear power
Although there are several interpretations of probability (subjectivistic, plants, structural damage assessments, and earthquake engineering. Failure
axiomatic, and frequentistic), probabilities generally have a physical event possibilities and error possibilities are essentially fuzzy sets on the interval
basis. They are tied to statistical experiments and are primarily useful for [O, 1] that employ the notions we examine in this section.
quantifying how frequently a sample occurs in a population. Possibility Over the years, two views, or schools of thought, of the definition of
theory, on the other hand, attempts to quantify how accurately a sample fuzziness have emerged. The first view, which we implicitly held in the
resembles an ideal element of a population. The ideal element is a prototypi- previous sections, has to do with categorizing or grouping the elements of a
cal class or a category of the population which we think of as a fuzzy set. In a universe of discourse into classes or sets whose boundaries of membership
sense, possibility theory may be viewed as a generalization of the theory of are fuzzy. Thus when we defined the set of small numbers in Example 2.1 we
probability with the consistency principIe, which we will see later on, providing identified a category of numbers within the universe of all numbers. Implic-
a heuristic connection between the two. Possibility theory focuses more on itly, what we dealt with in the example was the problem of imprecision. Our
the imprecision intrinsic in language, whereas probability theory focuses more main problem was to find the membership function that most appropriately
on events that are uncertain in the sense of being random in nature. In or accurately described the category of small numbers. The other view of
natural language processing, automatic speech recognition, knowledge-based fuzziness has to do with the problem of uncertainty. Here our main concern is
diagnosis, image analysis, robotics, analysis of rare events, information re- to quantify the certainty of an assertion such as "a number x is a small
trieval, and related areas, major problems are encountered on quantifying number," where x is an element of the universe of discourse X of numbers
the meaning of events-that is, the efficacious and accurate interpretation of (whose location on X is not known in advance) and is therefore called a
their significance and consequence and not the extent of their occurrence. nonlocated element. Possibility theory was advanced in order to address this
Let us illustrate with a simple example. type of problem. Possibility is more generally known as a fuzzy measure,
In the field of reliability analysis, probabilistic methods have been the which is a function assigning a value between O and 1 to each crisp set of the
basic instrument for quantifying equipment and human reliability as well. universe of discourse, signifying the degree of evidence or belief that a
Two very important concepts used are the failure rate and the error rate. particular element belongs to the set. Other types of fuzzy measure are belief
Knowing the failure rate of a component amounts to knowing the duration of measures, plausibility measures, necessity measures, and probability measures.
time that the component may be trusted to operate safely, and thus a The theory of fuzzy measures was advanced in 1974 by Sugeno as part of his
schedule for replacement and maintenance activities can be devised. It is not Ph.D. dissertation at Tokyo University. Fuzzy measures subsume probability
unusual, however, that after a component is fixed or replaced, the entire measures as well as belief and plausibility measures used in what is known as
system breaks down, a problem particularly acute with electronic compo- the Dempster-Shafer Theory of Evidence.
nents. Indeed, such general failures sometimes cause extremely negative Let us now take a closer look at possibility. Possibility is a fuzzy measure,
consequences, leading to catastrophic accidents. The problem here is that which means that possibility is a function with a value between O and 1,
failure rates are not sufficiently meaningful to account for the complex indicating the degree of evidence or belief that a certain element x belongs
interactions that a human being, such as a maintenance technician or an to a set (Zadeh, 1978; Dubois and Prade, 1988). A possibility of 0.3 for
operator, may have with a machine. In addition, the correct estimation of element x, for example, may indica te a 0.3 degree of evidence or belief that
 POSSIBILlTYTHEORY AND FUZZV PROBABILlTIES 41
40 FOUNDATIONS OF FUZZV APPROACHES
variable V, and 7TV(X) is the possibility distribution associated with V (or the
x belongs to a certain set. How this belief is distributed to elements other
possibility distribution function of Il). In equation (2.8-1), xix) is the charac-
that x is quantified through a possibility distributioã- In possibility theory, the
teristic function of A (see Section 2.1). Mathematically, Il is considered a
concept of possibility distribution is analogous to the notion of probability
measure which is a special function mapping the universe to the interval
distribution in probability theory. A possibility distribution is viewed as a fuzzy
[O, 1]. Knowing that the values that V may take are members of A is the
restriction acting as an elastic constraint on the values that may be assigned
same as knowing which values of the universe X are restricted to be values of
to a variable. What does this mean? Well, it is best to review the notion of a
V and which are restricted not to be values of V. We indicated this in
variable, first. Let A be a crisp set defined on a universe X and let V be a
equation (2.8-1) by using the characteristic function of the crisp set A. We
va:-jable taking values on some element x of X, a situation illustrated in
think of the crisp set A as a restriction on the values of the variable V, and in
Figure 2.11. The crisp set A is what in the parlance of probability we call an
view of the nonfuzzy nature of A this type of restriction is called a crisp
evento Events are comprised out of one or more basic events. Thus, the
restriction.
element x may be thought of as a basic event. If x is within A and x occurs,
Next, suppose that A is a fuzzy set and that its boundary no longer crisp
then we say that the event A has occurred as well. For example, in reliability
(i.e., does not sharply divide members from nonmembers) but is instead a
analysis, equipment failure and human errar are considered to be events
fuzzy boundary allowing an element x to be a member of A to some degree.
whose occurrence is based on the occurrence of basic events known as
As with any fuzzy set, A is uniquely identified by its membership function
initiating events. To say that V takes its values in A is to indicate that any
p.,A(X). In terms of events we think of A as a fuzzy event, and we can
element (basic event) of event A could possibly be a value of V and that any
associate with each basic event x a membership function indicating its
element outside of A, the complement of A, cannot be a value of V. Thus,
membership to A. Let us again consider a variable V whose arguments are
the statement V takes its value in A can be viewed as inducing a possibility, Il
elements of X9• Now suppose that V is constrained to take values on X. The
over X, associating with each value x the possibility that x is a value of V.
fuzzy set A aiso restricts the possible values that the variable V may take, but
This can be written as in a fuzzy manner-that is, to a degree. In such a case we consider the fuzzy

n(v:= x) = 7TV(X) = XA(X) = {6 if

if
x EA
x $A
(2.8-1)
set A to act as a fuzzy restriction on the possible values of V. Generalizing
equation (2.8-1) to the fuzzy case we say that the fuzzy set A induces a
possibility n. The associated possibility distribution 7TV(x) on the values that
where ":= " is an assignment symbol indicating that x is assigned to the V may assume is defined to be equal to the membership function of
A, p.,ix) and is written as

Il ( V := x) = 7T V( x) = p.,A( x) (2.8-2)

• ,///::
:::>4.::::::::: V
::~~~~~:\:~~~:'.
x

x
Thus, the possibility that Vis assigned x-that is, V:= x, which is some times
indicated as "V is x"-is postulated to be equal to the membership function
of A evaluated at x-that is, p.,ix). It is important to observe in equation
(2.8-2) that possibility distributions are fuzzy sets, while possibilities are just
numbers between O and 1. The possibility Il in (2.8.2) is a measure of the
compatibility of a given crisp value x that V may take with an a priori
defined set A. In this way, V becomes a variable associated with the
possibility distribution 7Tv(x) in much the same way as a random variable is
associated with the probability distribution.
What equation (2.8-2) indicates is that in certain situations, such as in the
definition of failure and error possibilities, it is of interest to interpret the
membership function p.,ix) of a fuzzy set as a possibility distribution of a
variable V. In this sense the fuzzy set A is viewed as the set of more or less
possible values for V.
Figure 2.11 Thestatement about a variable
V, "V takes its vaJues in A." has a different
'iJ:W,\~Jl\J:I) V
9In Chapter 5 the variable V will be generalized to a fuzzy variable, which is a variable that takes
meaning when the set A is crisp (top) than x fuzzy sets as values,
when the set Ais fuzzy (boffom),
 POSSIBILllY THEORY AND FUZZV PROBABILlTIES 43
42 FOUNDATIONS OF FUZZV APPROACHES

sures, and the interpretation of probability and possibility is quite different.

Given a possibility distribution 7Tv(x), the possibility that x may belong to
Probability is used to quantify the frequency of occurrence of an event, while
another crisp set B is defined as
possibility (along with fuzzy tools) is used to quantify the meaning of an
event. Consider the following example offered by Zadeh (1978). Suppose that
n(VcB) = V 7Tv(X) (2.8-3)
we have the proposition "Hans ate V eggs for breakfast," where V =
xEB
{I, 2, 3, ... }. A possibility distribution and a probability distribution may be
What equ-ltion (2.8-3) indicates is that the possibility of B is the possibility of associated with V, as shown in the following table:
the most possible elementary event x of B. Generalizing this relationship, it
can be shown (Dubois and Prade, 1988; Kandel, 1986) that the possibility x 1 2 3 4 5 6 7 8 9
measure of the union of two crisp sets B and C is the maximum of the 7Tv(x) 1 1 1 1 0.8 0.6 0.4 0.2 0.1
possibilities of B and C and can be written as
Pv(x) 0.1 0.8 0.1 O O O O O O
nr s U C) = n(B) V n(C) (2.8-4 )
The possibility distribution is interpreted as the degree of ease with which
Given a fuzzy set A and a possibility distribution function, 7Tv(x), the Hans can eat x eggs, while the probability distribution might have been
possibility of A, denoted as ll(A), is given by determined by observing Hans at breakfast for 100 days. Note that the
probability distribution function Pv(x) is given a frequentistic interpretation
n(A) = V [JLA(X) 1\ 7Tv(x)] (2.8-5) and that it sums to '1', while the possibility distribution function 7Tv(x) is
xEX imputed with a situation or context-dependent interpretation and does not
have to sum to '1'.
Consider two fuzzy events A and B defined over the universe of discourse Possibility is an upper bound for probability: A high degree of possibility
X. The possibility of A with respect to B is defined as does not imply a higher degree of probability. If, however, an event is not
possible, it is not als o probable. This is referred to as the probability /possibil-
n(AIB) = V [JLA(X) 1\ JLB(X)] (2.8-6) ity consistency principie (Zadeh, 1978). This heuristic principle is useful for
xEX drawing a distinction between the objectivistic use of probability measures
and the subjectivist use of possibility or fuzzy measures. When we attempt to
The possibility measure of A with respect to B reflects the extent to which use the two to describe a similar thing, we can use the possibility /probability
A and B coincide or overlap. Thus, possibility may be viewed as a measure of consistency principie as a guide. Possibility measures are more flexible mea-
comparison of fuzzy sets. sures useful for epistemic (i.e., cognitive) or context-dependent descriptions.
Conditional possibilities have been defined in analogy with conditional In general, according to Zadeh a variable may be associated with both a
probabilities; an entire body of theoretical results has been achieved, known possibility distribution and a probability distribution, with the weak connec-
generally as possibility theory. It is finding an increasing number of applica- tion between the two given by the consistency principle (Zadeh, 1978).
tions in the fields of knowledge representation and applied artificial intelli- In the language of probability theory the set A in Figure 2.11 may be
gence (Ragheb and Tsoukalas, 1988). A very comprehensive treatment of viewed as a fuzzy evento Such a fuzzy event induces a distribution on the
possibility may be found in the book entitled Possibility Theory by Dubois and values of a variable which we called the possibility distribution function and
Prade (1988). The theory of possibility has assumed particular significance in defined in equation (2.8-2). We can also define the probability of a fuzzy event
the field of natural language processing due to the inherent fuzziness of A. Suppose that a fuzzy event A is comprised of elementary events x, and
natural language. In the late 1970s Zadeh constructed a universal language with each x we associate a basic probability pi;x).
called PRUF, in which the translation of a proposition expressed in natural Zadeh defined the probability offuzzy eventA as the mathematical expecta-
language takes the form of a procedure for computing the possibility distribu- tion (the first moment) of its membership function, that is,
tion of a set of fuzzy relations in a database. The procedure, then, may be
interpreted as a semantic computation transforming the meaning of a propo-
sition to a computed possibility distribution quantifying the information P(A) = fXJLA(x)p(x) dx
conveyed by the proposition (Zadeh, 1983). (2.8-7)
There are certain differences between probability and possibility measures fxp(x) dx
worth pointing oiít. Possibility measures are "softer" than probability mea-
 FOUNDATIONS OF FUZZV APPROACHES
REFERENCES 45
44

where A is a fuzzy event on the universe X, x is an element of X, also called For another fuzzy set B = {integers that are not small} given by
an elementary event, and p(x) is a probability distribution (Zadeh, 1968).
When A is not a fuzzy event, equation (2.8-7) reduces back to the usual crisp B = 0.2/3 + 0.3/4 + 0.6/5 + 0.8/6 + 1.0/7
probability P(A). In equation (2.8-7) we assume that the probability measure
using equation (E2.3-3), we could obtain that the possibility of B is
on the entire universe of discourse must equal unity-that is, !xp(x) dx = l.
In addition, given equation (2.8-7) we can define a fuzzy mean as ll( B) = 0.2 V 0.3 V 0.3 V 0.3 V 0.1 = 0.3 (E2.3-5)

1 It should be noted in equations (E2.3-4) and (E2.3-5) that the possibility is

mA = P(A) JXXI-LA(X)P(X) dx (2.8-8)
simply a number between O and 1, whereas the possibility distribution is a
fuzzy set-for example, equation (E2.3-l).
and a fuzzy uariance as Let us now consider a simple instance of how to generate the possibility
distribution itself. Let C = 1/1 + 1/2 + 0.8/3 + 0.6/4 + 0.4/5 + 0.2/6 be

(TA2 = 1
P(A) fx(x-mA)2I-LA(X)P(x)dx (2.8-9)
a fuzzy set that represents small numbers. Then the proposition "V is a small
number" associates with V the possibility distribution, 1TV(X), taken in view of
equation (2.8-2) to be equal to the membership function of C-that is,
The probability of a fuzzy event as defined in equation (2.8-7) has been an 1TV(X) = 1/1 + 1/2 + 0.8/3 + 0.6/4 + 0.4/5 + 0.2/6 (E2.3-6)
extremely useful notion with wide application in the field of quantification
theory (Terano et al., 1992). Quantification methods are useful in analyzing In equation (E2.3-6) a singleton such as 0.6/4 indicates that the possibility
data involving human judgments which are not normally given numerical that x is 4, given that x is a small intege" is 0.6. D
expression, as well as in interpreting and understanding such data. -- (/r.•. u ),

Example 2.3 Possibility Measures and Distributions. Let us illustrate the REFERENCES
distinction between possibility measure or possibility and possibility distribu-
tion. We consider a possibility distribution induced by the proposition "V is a Dubois, D., and Prade, H., Fuzzy Sets and Systems: Theory and Applications, Academic
small integer" where the possibility distribution is (subjectívely) defined as Press, Boston, 1980.
Dubois, D., and Prade, H., Possibility Theory, Plenum Press, New York, 1988.
1TV(X) = 1.0/1 + 1.0/2 + 0.75/3 + 0.5/4 + 0.3/5 Kandel, A, Fuzzy Mathematical Techniques with Applications, Addison-Wesley,
Reading, MA, 1986.
+ 0.3/6 + 0.1/7 + 0.1/8 (E2.3-1)
Kaufmann, A, Introduction to the Theory of Fuzzy Subsets, Volume I, Academic Press,
New York, 1975.
We aiso consider the crisp set A = {3,4, 5} which we can write as
Klir, G. J., and Folger, T. A, Fuzzy Sets, Uncertainty, and Information, Prentice-Hall,
Englewood Cliffs, NJ, 1988.
A = I: I-LA(X)/X = 1/3 + 1/4 + 1/5 (E2.3-2)
Kosko, B., Neural Networks and Fuzzy Systems, Prentice-Hall, Englewood Cliffs, NJ,
xEX
1992.
What is the possibility of A? The possibility measure ll(A) is found using Onisawa, T., An Application of Fuzzy Concepts to Modelling of Reliability Analysis,
equation (2.8-5); that is, Fuzzy Sets and Systems, No. 37, pp. 267-286, North-Holland, Amsterdam, 1990.
Pai, S. K., and Majumder, D. K. D., Fuzzy Mathematical Approach to Pattem Recogni-
ll(A) = V [I-LA(X) 1\ 1Tv(X)] (E2.3-3) tion, John Wiley & Sons, New York, 1986.
XEX Radecki, T., Levei Fuzzy Sets, Joumal ofCybemetics, Vol. 7, pp. 189-198,1977.
Ragheb, M., and Tsoukalas, L. H., Monitoring Performance of Devices Using a
Using equations (E2.3-l) and (E2.3-2) in (E2.3-3), we can obtain the possibil- Coupled Probability-Possibility Method, Intemational Joumal of Expert Systems,
ity of A: Vol. 1, pp. 111-130,1988.
Shinohara, Y., Fuzzy Set Concepts for Risk Assessment, Intemational Institute for
ll(A) = 0.75 V 0.5 V 0.3 = 0.75 (E2.3-4) Applied Systems Analysis, Report WP-76-2, Laxenburg, Austria, January 1976.
46 FOUNDATIONS OF FUZZV APPROACHES PROBLEMS 47

Terano, T., Asai, K., and Sugeno, M., Fuzzy Systems Theory and Its Applications, (:) Given fuzzy set A, describing pressure p is higher than 15 mPa, through
Academic Press, Boston, 1992. the membership function:
Zadeh, L. A., Fuzzy Sets, Information and Control, Vol. 8, pp. 338-353, 1965.
Zadeh, L. A., Probability Measure of Fuzzy Events, Ioumal of Mathematical Analysis 1
andApplications, Vol. 23, pp. 421-427,1968. J.LA(X) = 1+ (x -15)-2 x> 15,
Zadeh, L. A., Outline of a New Approach to the Analysis of Complex Systems and
Decision Processes, IEEE Transactions on Systems, Man and Cybemetics, SMC-3, =0 x s 15,
pp. 28-44, 1973.
Zadeh, L. A., The Concept of a Linguistic Variable and its Application to Approxi- and fuzzy set B, describing pressure p is approximately equal to 17 mPa,
mate Reasoning, Infonnation Sciences, Vol. 8, pp. 199-249, 1975. with membership function:
Zadeh, L. A., Fuzzy Sets as a Basis for Theory of Possibility, Fuzzy Sets and Systems,
Vol. 1, pp. 3-28, 1978.
1
Zadeh, L. A., A Computational Approach to Fuzzy Quantifiers in Natural Languages,
J.LB(X) = 1 + (x - 17)4·
Computer and Mathematics, Vol. 9, pp. 149-184, 1983.
Zadeh, L. A., Fuzzy Logic, IEEE Compuser, pp. 83-93, Apri11988.
Zimmermann, H. J., Fuzzy Set Theory and its Applications, Kluwer-Nijhoff, Boston, Find the membership function of the fuzzy set C, describing pressure p is
1985. higher than 15 mPa and approximately equal to 17 mPa. Use at least four
different norms for interpreting AND (see Appendix) and draw all
membership functions.
PROBLEMS
(@ Using the data given in Problem 7, find the membership function of the
1. What happens to the curves in Figure 2.5 if we set F2 = 40 and vary FI fuzzy set D, describing pressure p is higher than 15 mPa or approximately
as in the figure? equal to 17 mPa. Use at least four different norms for interpreting OR
(see Appendix) and draw all membership functions.
2. In Figure 2.6, what is the significance of the intersection between the
J.LA = 0.5 line and the curves? r{j) Using the data given in Problem 7, find the membership function of the
fuzzy set E, describing pressure p is not higher than 15 mPa and approxi-
3; In Example 2.2, substitute y = sin x for equation (E2.2-l) and utilize the mately equal to 17 mPa. Use four different norms for interpreting AND
extension principb., in the same way as in the example. Choose an (see Appendix) and draw all membership functions.
appropriate range for x and assume any additional information needed
as in the example. 10. Determine all o-cuts for the following fuzzy sets, given that a = 0.0, 0.1,
0.2, ... 0.9, 1.0.
4. The fuzzy variable of Figure 2.9 is given by the equation J.Lix) =
I. A = 0.1/3 + 0.2/4 + 0.3/5 + 0.4/6 + 0.5/7 + 0.6/8 + 0.7/9 +
1/[1 + 0.3(x - 50)2]. Show that the 0.2level fuzzy set of fuzzy set A can
0.8/10 + 1.0/11 + 0.8/12
be represented by a-cuts using the resolution principlt
5. The fuzzy sets A and B are given by 11. B =
f-oo<x< +00 [11
1 + (x - 15)-
2 /x
A = 0.33/6 + 0.67/7 + 1.00/8 + 0.67/9 + 0.33/10
Write a MATLAB program that takes a number of o-cuts (minimum 10)
B = 0.20/3 + 0.60/4 + 1.00/5 + 0.60/6 + 0.20/7 and reconstructs the membership function.
(a) Write an expression for A V B. 11. Let X = N X N, and the fuzzy sets:
(b) Write an expression for A 1\ B.
1
6. Different fuzzy symbols are often used to mean similar things.
J.LA(X) = 1 + 10(x _ 2)2
(a) Write all symbols or terms that have the same general meaning as
max (v).
1
(b) Write all symbols or terms that have the same general meaning as J.LB(y) = 1 + 2y2
min (1\).
48 FOUNDATiONSOF FUZZVAPPROACHES

Let the mappings z = f(x, y), f: N x N ~ N be the following quadric

surfaces

2 y2
@ Z =
~
-
4 + -2' x EA , y EB .

x2 y2 Z2
(b) -+-~-=1
9 15 8
(c) 2y2 + 12z2 = x2
Sketch the surfaces and determine the image f(A X B) by the extension
principie, for each of the above.