×

Close

Touch here to read

Page-1

Topic:

11
Fuzzy Logic
11.1 Fuzzy sets and fuzzy logic
We showed in the last chapter that the learning problem is NP-complete
for a broad class of neural networks. Learning algorithms may require an
exponential number of iterations with respect to the number of weights until
a solution to a learning task is found. A second important point is that in
backpropagation networks, the individual units perform computations more
general than simple threshold logic. Since the output of the units is not limited
to the values 0 and 1, giving an interpretation of the computation performed
by the network is not so easy. The network acts like a black box by computing a
statistically sound approximation to a function known only from a training set.
In many applications an interpretation of the output is necessary or desirable.
In all such cases the methods of fuzzy logic can be used.
11.1.1 Imprecise data and imprecise rules
Fuzzy logic can be conceptualized as a generalization of classical logic. Modern fuzzy logic was developed by Lotﬁ Zadeh in the mid-1960s to model those
problems in which imprecise data must be used or in which the rules of inference are formulated in a very general way making use of diﬀuse categories
[170]. In fuzzy logic, which is also sometimes called diﬀuse logic, there are not
just two alternatives but a whole continuum of truth values for logical propositions. A proposition A can have the truth value 0.4 and its complement Ac
the truth value 0.5. According to the type of negation operator that is used,
the two truth values must not be necessarily add up to 1.
Fuzzy logic has a weak connection to probability theory. Probabilistic
methods that deal with imprecise knowledge are formulated in the Bayesian
framework [327], but fuzzy logic does not need to be justiﬁed using a probabilistic approach. The common route is to generalize the ﬁndings of multivalued logic in such a way as to preserve part of the algebraic structure [62]. In
R. Rojas: Neural Networks, Springer-Verlag, Berlin, 1996

290
11 Fuzzy Logic
this chapter we will show that there is a strong link between set theory, logic,
and geometry. A fuzzy set theory corresponds to fuzzy logic and the semantic
of fuzzy operators can be understood using a geometric model. The geometric
visualization of fuzzy logic will give us a hint as to the possible connection
with neural networks.
Fuzzy logic can be used as an interpretation model for the properties of
neural networks, as well as for giving a more precise description of their performance. We will show that fuzzy operators can be conceived as generalized
output functions of computing units. Fuzzy logic can also be used to specify
networks directly without having to apply a learning algorithm. An expert
in a certain ﬁeld can sometimes produce a simple set of control rules for a
dynamical system with less eﬀort than the work involved in training a neural network. A classical example proposed by Zadeh to the neural network
community is developing a system to park a car. It is straightforward to formulate a set of fuzzy rules for this task, but it is not immediately obvious
how to build a network to do the same nor how to train it. Fuzzy logic is now
being used in many products of industrial and consumer electronics for which
a good control system is suﬃcient and where the question of optimal control
does not necessarily arise.
11.1.2 The fuzzy set concept
The diﬀerence between crisp (i.e., classical) and fuzzy sets is established by
introducing a membership function. Consider a ﬁnite set X = {x1 , x2 , . . . , xn }
which will be considered the universal set in what follows. The subset A of
X consisting of the single element x1 can be described by the n-dimensional
membership vector Z(A) = (1, 0, 0, . . . , 0), where the convention has been
adopted that a 1 at the i-th position indicates that xi belongs to A. The set
B composed of the elements x1 and xn is described by the vector Z(B) =
(1, 0, 0, ..., 1). Any other crisp subset of X can be represented in the same way
by an n-dimensional binary vector. But what happens if we lift the restriction
to binary vectors? In that case we can deﬁne the fuzzy set C with the following
vector description:
Z(C) = (0.5, 0, 0, ..., 0)
In classical set theory such a set cannot be deﬁned. An element belongs to
a subset or it does not. In the theory of fuzzy sets we make a generalization
and allow descriptions of this type. In our example the element x1 belongs to
the set C only to some extent. The degree of membership is expressed by a
real number in the interval [0, 1], in this case 0.5. This interpretation of the
degree of membership is similar to the meaning we assign to statements such
as “person x1 is an adult”. Obviously, it is not possible to deﬁne a deﬁnite
age which represents the absolute threshold to enter into adulthood. The act
of becoming mature can be interpreted as a continuous process in which the
membership of a person to the set of adults goes slowly from 0 to 1.
R. Rojas: Neural Networks, Springer-Verlag, Berlin, 1996

11.1 Fuzzy sets and fuzzy logic
291
There are many other examples of such diﬀuse statements. The concepts
“old” and “young” or the adjectives “fast” and “slow” are imprecise but easy
to interpret in a given context. In some applications, such as expert systems,
for example, it is necessary to introduce formal methods capable of dealing
with such expressions so that a computer using rigid Boolean logic can still
process them. This is what the theory of fuzzy sets and fuzzy logic tries to
accomplish.
1
young
mature
0.8
old
degree of
membership
0.2
0
10
20
30
40
50
60
70
age
Fig. 11.1. Membership functions for the concepts young, mature and old
Figure 11.1 shows three examples of a membership function in the interval
0 to 70 years. The three functions deﬁne the degree of membership of any
given age in the sets of young, adult, and old ages. If someone is 20 years old,
for example, his degree of membership in the set of young persons is 1.0, in
the set of adults 0.35, and in the set of old persons 0.0. If someone is 50 years
old the degrees of membership are 0.0, 1.0, 0.3 in the respective sets.
Deﬁnition 11. Let X be a classical universal set. A real function μA : X →
[0, 1] is called the membership function of A and deﬁnes the fuzzy set A of X.
This is the set of all pairs (x, μA (x)) with x ∈ X.
A fuzzy set is completely determined by its membership function. Note
that the above deﬁnition also covers the case in which X is not a ﬁnite set.
The set of support of a fuzzy set A is the set of all elements x of X for
which (x, μA (x)) ∈ A and μA (x) > 0 holds. A fuzzy set A with the ﬁnite set
of support {a1 , a2 , . . . , am } can be described in the following way
A = μ1 /a1 + μ2 /a2 + · · · + μm /am ,
where μi = μA (ai ) for i = 1, . . . , m. The symbols “/” and “+” are used only
as syntactical constructors.
R. Rojas: Neural Networks, Springer-Verlag, Berlin, 1996

292
11 Fuzzy Logic
Crisp sets are a special case of fuzzy sets, since the range of the function
is restricted to the values 0 and 1. Operations deﬁned over crisp sets, such as
union or intersection, can be generalized to cover also fuzzy sets.
Assume as an example that X = {x1 , x2 , x3 }. The classical subsets A =
{x1 , x2 } and B = {x2 , x3 } can be represented as
A = 1/x1 + 1/x2 + 0/x3
B = 0/x1 + 1/x2 + 1/x3 .
The union of A and B is computed by taking for each element xi the maximum
of its membership in both sets, that is:
A ∪ B = 1/x1 + 1/x2 + 1/x3
The fuzzy union of two fuzzy sets can be computed in the same way. The
union of the two fuzzy sets
C = 0.5/x1 + 0.6/x2 + 0.3/x3
D = 0.7/x1 + 0.2/x2 + 0.8/x3
is given by
C ∪ D = 0.7/x1 + 0.6/x2 + 0.8/x3
The fuzzy intersection of two sets A and B can be deﬁned in a similar way, but
instead of taking the maximum we compute the minimum of the membership
of each element xi to A and B. The maximum or minimum of the membership
values are just one pair of possible deﬁnitions of the union and intersection
operations for fuzzy sets. As we show later on, there are other alternative
deﬁnitions.
11.1.3 Geometric representation of fuzzy sets
Bart Kosko introduced a very useful graphical representation of fuzzy sets
[259]. Figure 11.2 shows an example in which the universal set consists only
of the two elements x1 and x2 . Each point in the interior of the unit square
represents a subset of X. The convention is that the coordinates of the representation correspond to the membership values of the elements in the fuzzy
set. The point (1, 1), for example, represents the universal set X, with membership function μA (x1 ) = 1 and μA (x2 ) = 1. The point (1, 0) represents the
set {x1 } and the point (0, 1) the set {x2 }. The crisp subsets of X are located
at the vertices of the unit square. The geometric visualization can be extended
to an n-dimensional hypercube.
Kosko calls the inner region of a unit hypercube in an n-dimensional space
the fuzzy region. We ﬁnd here all combinations of membership values that a
fuzzy set could assume. The point M in Figure 11.2 corresponds to the fuzzy
set M = 0.5/x1 + 0.3/x2 . The center of the square represents the most diﬀuse
of all possible fuzzy sets of X, that is the set Y = 0.5/x1 + 0.5/x2 .
The degree of fuzziness of a fuzzy set can be measured by its entropy. In
the geometric visualization, this corresponds inversely to the distance between
R. Rojas: Neural Networks, Springer-Verlag, Berlin, 1996

## Leave your Comments