T h e role o f fuzzy logic in modeling, identification and control

LOTFI A. ZADEHt

Keywords: Fuzzy logic: fuzzy control: fuzzy modeling.

Fuzzy Sets, Fuzzy Logic, and Fuzzy Systems Downloaded from www.worldscientific.com

In the nearly four decades which have passed since the launching of the Sputnik,
by NANYANG TECHNOLOGICAL UNIVERSITY on 11/10/15. For personal use only.

great progress has been achieved in our understanding of how to model, identify and
control complex systems. However, to be able to design systems having high MIQ
(Machine Intelligence Quotient), a profound change in the orientation of control
theory may be required. More specifically, what may be needed is the employment
of soft computing—rather than hard computing—in systems analysis and design.
Soft computing-unlike hard computing—is tolerant of imprecision, uncertainty
and partial truth.
At this juncture, the principal constituents of soft computing are fuzzy logic,
neurocomputing and probabilistic reasoning. In this paper, the focus is on the role
of fuzzy logic. The basic ideas underlying fuzzy logic and its applications to
modeling, identification and control are described and illustrated by examples. The
role model for fuzzy logic is the human mind.

1. Introduction—A personal perspective

The origins of modern control theory go back to the days of World War II when the
analysis and design of servomechanisms played a central role in the conception and
construction of electromechanical systems which could transform low-power inputs
into high-power outputs.
As a student at MIT—and later a professor at Columbia University—I witnessed
the evolution of control theory from a simplistic, frequency-based theory of linear
control systems into a mathematically sophisticated theory of linear/nonlinear systems
based on the theory of differential equations and dynamical systems.
It was, above all, the advent of the space age in 1957 that changed the orientation
and spirit of control theory. The competition in space—and the challenges posed by the
complexity of control problems related to the guidance of space vehicles—attracted to
control theory a number of prominent mathematicians, most notably L. S. Pontryagin
in the USSR and R. E. Bellman in the United Stales. As a result, the level of
mathematical sophistication of the theory began to grow very rapidly, swinging the
pendulum all the way from the low-brow approaches of the forties and fifties to the
high-brow mathematical formalism of the seventies, eighties and nineties.
In the fifties and early sixties, my interests centred not on control but on a more
general theory—system theory—a theory which was aimed at the development of a

Received 31 March 1994.

tComputer Science Division and the Electronics Research Laboratory, Department of
EECS, University of California, Berkeley, CA 94720; Tel: 510-642-4959; Fax: 510-642-8271,510-
642-5775; Email: zadeh(ei'cs.berkeley.edu.
Research supported in part by the BISC Program, NASA Grant NCC 2-275, EPRI
Agreement RP 8010-34 and MICRO State Program No. 92-180.

783
 784 Selected Papers by LotS A. Zadeh

192 L. A. Zadeh

better understanding of how abstract systems behave and how they could be designed
to perform in a specified way.
As a mathematically oriented system theorist, I had been conditioned to believe
that the analytical tools based on set theory and two-valued logic were all that was
needed to build a framework for a precise, rigorous and effective body of concepts and
techniques for the analysis of almost any kind of man-made or natural system,
including control systems. Then, in 1961-1963, in the course of writing a book on
system theory with C. A. Desoer, I began to feel that highly complex systems—typified
by economic and biological systems cannot be dealt with effectively by the use of
conventional approaches. My feeling derived, in the main, from a realization that
Fuzzy Sets, Fuzzy Logic, and Fuzzy Systems Downloaded from www.worldscientific.com
by NANYANG TECHNOLOGICAL UNIVERSITY on 11/10/15. For personal use only.

system description languages based on classical mathematics are not sufficiently

expressive to serve as a means of characterization of complex input output relations in
an environment of imprecision and uncertainty.
The culprit as 1 saw it was the universally made assumption that classes have
sharply defined boundaries. They do in classical mathematics, but in the real world that
we live in the opposite is the case, that is, almost all classes are fuzzy in the sense that the
transition from membership to nonmembcrship in such classes is gradual rather than
abrupt. Accepting this fact, the obvious thing to do is to assume that membership in a
class is a matter of degree. This assumption is the genesis of the theory of fuzzy sets.
After I wrote my first paper on fuzzy sets in 1965, my aims as a system theorist
underwent a marked shift. I came to the conclusion that not just control theory and
systems analysis, but, more generally, most or all scientific methodologies will have to
undergo a critical reexamination and move toward the replacement of their crisp
foundations with foundations based on fuzzy set theory and fuzzy logic (Zadeh 1971,
1972), aiming at greater generality and better approximation to reality. In the realm of
control, in particular, the replacement of crisp logic with fuzzy logic could make it
possible to design systems with a much higher MIQ (Machine Intelligence Quotient)
compared to those which can be designed by traditional methods. In what follows, I
will attempt to describe very briefly the basis for my conclusion.

2. Fuzzy logic, linguistic variables and granulation

Fuzzy logic has been and still is somewhat controversial. In part, the controversy
relates to the fact that the term fuzzy logic is used in two different senses. In a narrow
sense, fuzzy logic (FLn) is a logical system which aims at a formalization of
approximate reasoning. As such, it is rooted in multivalued logic but its agenda is quite
different from that of traditional multivalued logical systems, e.g., Lukasiewicz's logic.
In this connection, it should be noted that many of the concepts which account for the
effectiveness of fuzzy logic as a logic of approximate reasoning are not a part of
traditional multivalued logical systems. Among these are the concepts of a linguistic
variable, canonical form, fuzzy rule, fuzzy graph, fuzzy quantifiers and such modes of
reasoning as interpolative reasoning, syllogistic reasoning and dispositional reasoning.
In a broad sense, fuzzy logic (FLw) is almost synonymous with fuzzy set theory.
Fuzzy set theory is much broader than fuzzy logic in its narrow sense and contains the
latter as one of its branches. Among the other branches of fuzzy set theory are fuzzy
arithmetic, fuzzy mathematical programming, fuzzy topology, fuzzy graph theory,
possibility theory, etc. What is important to recognize is that any crisp theory can be
fuzzified by generalizing the concept of a set within that theory to the concept of a fuzzy
set. Indeed, it is very likely that eventually most theories will be fuzzified in this way. As
 The Role of Fuzzy Logic in Modeling ... [1994] 785

The role of fuzzy logic 193

was alluded to earlier, the impetus for transition from a crisp theory to a fuzzy one
derives from the fact that both the generality of a theory and its applicability to real
world problems arc substantially enhanced by replacing the concept of a set with that of
a fuzzy set.
Today, the growing tendency is to use the term fuzzy logic (FL) in its broad sense. In
part this reflects the fact that fuzzy logic sounds more euphonious than fuzzy set theory.
A concept which plays a cenfral role in the applications of fuzzy logic is that of a
linguistic variable (Zadeh 1973, 1975). A linguistic variable, as its name suggests, is a
variable whose values are words rather than numbers. For example, age is a linguistic
variable if its values are: young, old, not very old, etc. Each linguistic value is interpreted
Fuzzy Sets, Fuzzy Logic, and Fuzzy Systems Downloaded from www.worldscientific.com
by NANYANG TECHNOLOGICAL UNIVERSITY on 11/10/15. For personal use only.

as a label of a fuzzy set in its universe of discourse and each such set, e.g., old, is defined
by its membership function, nol<i(u), which associates with each numerical value, u, of
age the degree to which n fits one's subjective, context-dependent perception of the
meaning of old. The grade of membership is assumed to take values in the interval
[0,1].
In current practice, in most applications to control, the membership functions of
linguistic values are assumed to be trapezoidal or triangular in shape (Fig. 1). The
number of linguistic values is usually in the range of three to seven, and the values are
labeled positive small (PS), negative large (NL), zero (Z), etc.
There is a basic aspect of the concept of a linguistic variable which is at the heart of
its utility. Specifically, consider a linguistic variable such as age whose linguistic values
are young, middle-aged, and old, with young defined by a membership function such as
shown in Fig. 2.
Clearly, a numerical value such as 25 is simpler than the function young. But young
represents a choice of one out of three possible values whereas 25 is a choice of one out
of, say, 100 values. The point of this simple example is that the use of linguistic values
may be viewed as a form of data compression. It is suggestive to refer to this form of
data compression as fuzzy granulation or simply granulation.

NL NM NS .i Z PS PM PL
\ \ \ / / / /

0 X

Figure 1. Triangular linguistic values.

LINGUISTIC FUZZY
VARIABLE GRANULATION

DATA
COMPRESSION

20 40 Age 25 Age
choice of one out of choice of one out of
3 labels 100 values

Figure 2. Granulation and data compression.

 786 Selected Papers by Lotfi A. Zadeh

194 L. A. Zadeh

^ k small medium targe ^ A smalt medium large

s s s / s s
, \XX\,
Figure 3. Quantization vs. granulation.

The same effect can be achieved, of course, by conventional quantization. But in the
case of quantization, the values are crisp intervals whereas in the case of granulation the
values are overlapping fuzzy intervals (Fig. 3), resulting in a gradual rather than abrupt
transition from one value to another.
Fuzzy Sets, Fuzzy Logic, and Fuzzy Systems Downloaded from www.worldscientific.com
by NANYANG TECHNOLOGICAL UNIVERSITY on 11/10/15. For personal use only.

The graduality of transition mimics the way in which humans perceive linguistic
values. More importantly—as will be seen later—it provides a basis for interpolative
reasoning. Such reasoning plays a central role in learning from experience.

3. Fuzzy rules
In fuzzy logic, the concept of a linguistic variable serves as a point of departure for
the construction of a language, referred to as FDCL (Fuzzy Dependency and
Command Language), which provides a framework for the description of imprecise
dependencies and commands through the use of fuzzy if-then rules or simply fuzzy
rules. In what follows, we shall sketch some of the basic ideas underlying FDCL.
Like any language, FDCL is characterized by its syntax and semantics. The syntax
and semantics of FDCL define, respectively, the form of rules and their meaning.
Simple examples of the rules in FDCL are the following.
(a) if pressure is high then volume is small
(b) if pressure is high and temperature is low then volume is very small
(c) if pressure is high then lower temperature slightly
(d) if pressure is high then volume is small unless temperature is high
(e) if pressure is high then usually volume is small.
In these examples, pressure, volume and temperature are linguistic variables and
small, low and high are their linguistic values. All of the rules except (c) represent
dependencies, with (c) representing a command. All of the rules except (d) and (e) are
categorical. Rules (d) and (e) are qualified, with (d) qualified through an exception and
(e) through usuality. (e) exemplifies what is referred to as a dispositional rule. All of the
rules except for (b) involve a single variable in the antecedent. A real-world example of
rules used in Honda's fuzzy logic transmission is shown in Fig. 4.
Assuming that small, low, high, usually and whatever other linguistic values may be
involved are defined by their membership functions, as in Fig. 4, a question of semantics
is: What is the meaning of a rule? And, more generally: What is the meaning of a
collection of rules? We shall examine these questions very briefly in the context of basic
generic rules of the form
if A" is A then Y is B, (1)
where A and B are linguistic values of X and Y, respectively.
In fuzzy logic, the answers to these questions are not unique. In what has become a
standard interpretation, (1) is interpreted as a fuzzy (elastic) constraint on the variables
X and y. More specifically, with -* representing translation, we have
if X is A then Y is B^(X, Y) is A x B, (2)
 The Role of Fuzzy Logic in Modeling ... [1994] 787

The role of fuzzy logic 195

CONTROL RULES FUZZY SET

1 1F (SPEED IS LOW) AND (SHIFT IS HIGH) NOT VERY
LOW
THEN ( - 3 )
2. IF (SPEED IS HIGH) AND (SHIFT IS LOW)
THEN ( + 3 )
v ' - » - MOT
LOW
3. IF (THROT IS LOW) AND (SPEED IS U G H )
THEN ( + 3 )
4.1F fTHROT IS LOW) AND (SPEED IS LOW)
THEN(+t)
5. * (TOROT IS HIGH) AND (SPEED IS HIGH)
THEN(-T)
6. IF (THROT IS HK3H) AND (SPEED IS LOW)
THEN ( - 3 )
7 * (RESIST IS POSITIVE BIG) AND
Fuzzy Sets, Fuzzy Logic, and Fuzzy Systems Downloaded from www.worldscientific.com

(THROT IS LOW) AND

by NANYANG TECHNOLOGICAL UNIVERSITY on 11/10/15. For personal use only.

(SHIFT IS NOT VERY LOW)

-so o • a
THEN ( - 1 5 ) THROTTLE <d«gl ROAD RESISTANCE (Xfl)
8. * (REStST IS NEGATIVE BG) AND
(THROT IS CLOSE) AND NEGATIVE BIG LONG
1 1
(SHIFT IS NOT LOW)
THEN ( - 2 )
9. f (BRAKE TIME IS LONG) AND
( A U IS NEGATIVE BIG) AND
o
(SHIFT IS NOT LOW) D
-I LI 0 Q.1 0
/
7 S 1
THEN ( - 2 ) BAAKE TIME (MC)
ACCELERATION <G)

Figure 4. Fuzzy rules employed in Honda's fuzzy logic transmission.

where A x B is the cartesian product of A and B. A x B is defined by

HA*B(U, V) = HA(U)AH^V), ueU, veV,

where nA and nB are the membership functions of A and B, respectively; U and V are the
universes of discourse of X and Y; and A is the operation of conjunction, which is
usually defined as min but, more generally, can be any t-norm (Zimmermann 1990).
More generally, if the dependence of y on X is characterized by a collection of n
rules of the form (1), that is:
MX is Ai then Vis Bh i = l , . . . , « , (3)
then the meaning of the collection in question is defined by
if X is Ai then Y is B;, i=l,...,n->(X,Y) is Al x B, +... + An x B„, (4)
where + denotes disjunction. Usually the disjunction is defined as max but, more
generally, could be any t-conorm. For convenience, the right-hand member of (4) can be
written as
(X, Y) is 1,-/1,. x B,,
If / is a crisp function of X, Y=f(X), the expression
f* = I.iAixBi (5)
constitutes the fuzzy graph o f / (Zadeh 1971, 1973/74a, 1974b/76). More generally, if
Y = / ( X , , . . . , X „ ) is defined via n fuzzy rules of the form
Y is B, i f * ! is A} and ... and Xt is A^ i = l , (6)
then the fuzzy graph, / * , is expressed as
f* = T.iBixAl x ... xAl i = 1,...,«.
As shown in Fig. 5, / * may be interpreted as a coarse relational approximation to / .
 788 Selected Papers by LotG A. Zadeh

196 L. A. Zadeh

Figure 5. Fuzzy graph approximation to f.

r f = small x small *
Fuzzy Sets, Fuzzy Logic, and Fuzzy Systems Downloaded from www.worldscientific.com

medium x large *
by NANYANG TECHNOLOGICAL UNIVERSITY on 11/10/15. For personal use only.

large x small
granulation

X X

Figure 6. Granulation and data compression.

What is important to note is that / * serves to represent the dependence of Y on X in

a data-compressed form. For example, in the case of the function shown in Fig. 6, the
fuzzy graph
/ * = small x small + medium x large + large x small

may be interpreted as a compressed representation of /

Manipulation of compressed representations expressed as fuzzy graphs is the
province of the calculus of fuzzy graphs (CFG) (Zadeh 1994). Most of the applications
of fuzzy logic in the realms of systems analysis and control are in reality applications of
CFG.
The importance of C F G derives from the fact that it provides a simple and yet
effective way of representing imprecise dependencies. What should be stressed is that
C F G does not replace the traditional calculi of differential equations, difference
equations, matrices, etc. Rather, it adds to the collection of tools for systems analysis a
body of concepts and techniques which are effective in dealing with dependencies which
are imprecise to begin with or can be treated imprecisely to achieve tractability,
robustness or low solution cost.

4. Fuzzy modeling and control

In an idealized sense, a system S has a model M(S) if M(S) makes it possible to
predict the response of S to any given input. In reality, the response of M(S) is usually
an approximation to the response of S—an approximation which holds for a restricted
set of inputs.
In traditional approaches to systems analysis, M(S) is generally represented as a
collection of differential or difference equations. In fuzzy-logic-based approaches, M(S)
is represented as a collection of fuzzy if-then rules or, more particularly, as a collection
of fuzzy rules of the form (6). In this sense, M(S) is a fuzzy model if through the use of the
calculi of fuzzy rules or fuzzy graphs the knowledge of M(S) makes it possible to predict
the response of S to any given fuzzy input.
 The Role of Fuzzy Logic in Modeling ... [1994] 789

The role of fuzzy logic

R fr—| - * H )
1-*-.

S "

Figure 7. Example of a fuzzy model.

Fuzzy Sets, Fuzzy Logic, and Fuzzy Systems Downloaded from www.worldscientific.com
by NANYANG TECHNOLOGICAL UNIVERSITY on 11/10/15. For personal use only.

Figure 8. The meaning of the degree of match between A and A-,.

An example of a fuzzy model is shown in Fig. 7. In this model, R and S are assumed
to be characterized by the fuzzy rules
Z is large if X is small and Y is medium
R- Z is small if X is medium and Y is not very small

| Y is small if Z is large
S< Y is medium if Z is not very large

The calculus of fuzzy graphs provides a computationally simple method of

computing the response of S to any given fuzzy input through the use of fuzzy
interpolation. The basis for the computation in question is the interpolation rule
(X, Y) is S,vl, x B,.
X is A
(7)
Y is r,m,. A Bt
where mi represents the degree to which the value of A", A, matches the antecedent of /th
rule. More specifically
mi = supu(fiA(u)AHAi(u)), (8)

where the supremum is taken over all u e U (Fig. 8).

The interpolation rule (7) is a derivative of the compositional rule of inference
(Zadeh 1973); it was described in Zadeh (1974a, 1974/76b) and was employed
independently by Mamdani and Assilian in their seminal work on fuzzy control
(Mamdani 1975). Today, the interpolation rule and its variants play a central role in
most of the applications of fuzzy logic to control and systems analysis.
In the foregoing discussion, we have tacitly restricted our attention to static fuzzy
systems. However, as shown in some ofour early work (Zadeh 1971, 1974), the method
in question can readily be extended to dynamic fuzzy systems.
 790 Selected Papers by LotR A. Zadeh

198 L. A. Zadeh

More specifically, let X„ Y, and S, denote the fuzzy input, fuzzy output and fuzzy
state at time r, t = 0,1,2,..., of a finite-state fuzzy system S. Then, the state equations of
S may be expressed as
S, + s=f(S„XXt = 0, 1,2,
y,=<?(s„x,). (9)

Expressed in FDCL, the state equations can be represented as a collection of fuzzy

rules of the form
if X, is Aj and S, is Bi then S, + , is C,, i = 1, n
(10)
Fuzzy Sets, Fuzzy Logic, and Fuzzy Systems Downloaded from www.worldscientific.com
by NANYANG TECHNOLOGICAL UNIVERSITY on 11/10/15. For personal use only.

if X, is /4, and S, is B, then Y, is £>,,

where /1„ B, and C, are linguistic values of A",, S, and Y„ respectively.
Equivalently, the state equations can be represented as fuzzy graphs
/ * = 1,-/1,. x Bi x C,

a* = I 1 / f i x B ; x D , (11
As a simple illustration, assume that the linguistic values of X„ S, and Y, are S
(small), M (medium) and L (large), respectively, with the understanding that the
membership functions of S, M and L are not necessarily the same for Xr S, and Y,.
Expressed in tabular form, the state equation may have, as an example, the form shown
in Table 1.
In cases where the spaces of states, inputs and outputs are continua rather than
finite sets it may be convenient to represent St+l in incremental form, that is,

S,+ l=S, + dS,.

As an illustration, in the case of an inverted pendulum, S, = (0„0,), where 0, is the

vertical inclination and 0, is the time-derivative of 6„ with both 6, and 0, interpreted as
linguistic variables with linguistic values Z (zero), PS (positive small), NS (negative
small), etc. In this case, an entry in the state transition table may be expressed as a fuzzy
rule exemplified by
if S, is (Z, PS) and X, is PS then SS, is (NS, Z).
Knowledge of a fuzzy model of S in a direct or incremental form provides a basis for
the formulation and solution of various types of generic control problems centring on
S. One such problem is the following.

Table 1. State equations of afinite-statefuzzy system.

transition table
S M L current state
current input M i> L M next state
L L MS
output table
S M L current state

current input
sML 5 current output
MS L M
L L MS
 The Role of Fuzzy Logic in Modeling ... [1994] 791

The role of fuzzy logic 199

Given an initial state S, and a goal state SG in the state space of S find a policy
function n,

which would take S from the initial state at time t = 0 to a slate at time t = N which is
nearest to SG.
Problems of this type can be solved through the use of fuzzy dynamic programming
(Bellman and Zadch 1970). More specifically, assume that the error at time N is defined
as the distance, J, between SN and SG, i.e.,
CN(SN) = d(SN, SG).
Fuzzy Sets, Fuzzy Logic, and Fuzzy Systems Downloaded from www.worldscientific.com
by NANYANG TECHNOLOGICAL UNIVERSITY on 11/10/15. For personal use only.

Using backward iteration, we can write

^ ^ ( S ^ , ) = minA.N ,U/(/(S N _„ * „ _ , ) , SG)) (12)
= minimum error at / = N
achievable from SN _ ,
eN_2(SN^2) = minXN ,(<?N- , ( / ( S N _ 2 , XN.2)))
= minimum error at t = N
achievable from S N _ 2

e 0 (S 0 ) = min X 0 ( e ,(/(S 0 ,X 0 ))).

Then, proceeding in the forward direction we can determine the policy function by
successively minimizing the error at times f = 0 , . . . , N.
At this juncture, fuzzy-logic-based control is used extensively in a wide variety of
consumer products and industrial systems ranging from cameras and washing
machines to cement kilns and subways. In such applications, a fuzzy model of the
system is not known. Instead, the point of departure is the designer's or operator's
perception of the fuzzy rules which should be used to achieve a desired objective.
Viewed in this perspective, in most of its current applications fuzzy control involves a
representation of what is basically a human solution in the form of fuzzy if-then rules
expressed in FDCL.
It should be realized, however, that a human solution is implicitly rooted in a fuzzy
model of the system under control. Thus, the main difference between the prescriptive
fuzzy-model-based approach described earlier and the descriptive 'model-free'
approach which is employed in current practice reduces to the question of whether the
fuzzy model is explicit or implicit. At this juncture, our understanding of how fuzzy
models can be constructed and dealt with is still rather rudimentary. In coming years,
as our understanding of fuzzy models improves, prescriptive approaches to fuzzy
control are likely to gain in importance and visibility.

5. Fuzzy system identification and induction of rules from observations

In dealing with fuzzy rules, it is important to differentiate between the surface
structure and the deep structure of a fuzzy rule.
The surface structure of a rule is the rule expressed in a symbolic form, e.g.,
if X is A then Y is B. (13)
 792 Selected Papers by Lotfi A. Zadeh

200 L. A. Zadeh

The deep structure of a rule is its surface structure together with the definitions of
linguistic values which appear in its antecedent and consequent. In the case of the rule
represented by (13), the deep structure is the rule itself together with a specification of
the membership functions of A and B.
The concepts of surface and deep structure provide a convenient point of departure
for the formulation of the identification problem in the context of fuzzy rule-based
systems.
The basic identification problem has two variants. In the first variant, the starting
point is the knowledge of the surface structure of a system, S, which is the object of
identification. The problem is to approximate to the deep structure of S based on given
Fuzzy Sets, Fuzzy Logic, and Fuzzy Systems Downloaded from www.worldscientific.com

input-output data, that is, a collection of input-output pairs.

by NANYANG TECHNOLOGICAL UNIVERSITY on 11/10/15. For personal use only.

The second variant is considerably more difficult than the first. In this variant, the
problem is the same but the surface structure is not known initially. A seminal
contribution to the solution of this problem was made by Takagi and Sugeno (1985).
Today, there is a fairly extensive literature on the problem in question but a definitive
solution is not yet in hand. In what follows, we shall discuss very briefly the simpler (first
variant) of the identification problem.
As is typical of standard approaches to the identification problem, the point of
departure is the postulated structure of a fuzzy system modulo the parameters of
linguistic variables in the fuzzy if-then-rules.
Many of the approaches described in the literature are based on variants of the
Takagi-Sugeno-Kang (TSK) architecture, which is illustrated in Fig. 9. In this
architecture, the input X is assumed to have k components Xi,...,Xk and the
consequent of rule R' is assumed to be a linear combination of the constituents of X.
The output, V, is a convex combination of the consequents, with the weight, w,, being
the conjunction of the grades of membership of A 1 ,,..., Xk in the antecedents A\,...,A'k,
respectively. In a frequently employed variant of this architecture, A is replaced with *
(product) and the membership functions are assumed to be triangular or trapezoidal. In

xo—»—(..~o. *—oy

// X1 is A\ and ... andXk Is Ak

then Y'= p'0 + p',*, + ... + p'kXk

Y= I,P,>"
P,.= w,./ I,.w,
w,= A'j(X,) A ... /\Ak{Xk)

Figure 9. Takagi-Sugeno-Kang architecture.

 The Role of Fuzzy Logic in Modeling ... [1994] 793

The role offuzzy logic 201

Fuzzy Sets, Fuzzy Logic, and Fuzzy Systems Downloaded from www.worldscientific.com
by NANYANG TECHNOLOGICAL UNIVERSITY on 11/10/15. For personal use only.

Figure 10. The Jang architecture. In the illustrated system, there are two inputs, x and y, and
three rules.

other variants, the membership functions are assumed to be gaussian. This assumption
leads to methods which are similar to the radial basis function approaches used in
neurocomputing (Warwick et al. 1992).
A basic idea underlying the approaches described in the literature is that a fuzzy
system based on the TSK architecture can be represented as a multilayer feedforward
network. In the Jang approach (Jang 1992), for example, the system is represented as
shown in Fig. 10. The circles with n and N represent multipliers and normalizers,
respectively.
Representing a fuzzy system as a multilayer structure lays the groundwork for the
application of neurocomputing techniques and, in particular, the backpropagation
algorithm (Hertz et al. 1991) to the determination of parameters of membership
functions. In a more basic way, the techniques in question are rooted in dynamic
programming (Werbos 1974) and, more particularly, in gradient programming, which
through chain differentiation leads to the backpropagation algorithm (Zadeh 1994).
More detailed expositions of these methods may be found in (Jang 1992, Lin and Lee
1991, Wang 1993 and Takagi 1991).
In addition to the use of techniques borrowed from neurocomputing, there are
many promising approaches based on reinforcement learning (Berenji 1993) and
genetic algorithms (Lee and Takagi 1994). Important progress had been made in our
understanding of the identification problem in the context of fuzzy systems, but many
difficult problems remain to be addressed.

6. Concluding remarks
In coming years, a significant paradigm shift in systems analysis and control is likely
to take place. The shift in question involves a move toward the replacement of crisp
foundations of modern control theory and systems analysis by a foundation based on
fuzzy logic and soft computing—a collection of computing methodologies which are
tolerant of imprecision, uncertainty and partial truth.
 794 Selected Papers by LotR A. Zadeh

202 L. A. Zadeh

In my perception, the shift in question will h a v e a n i n v i g o r a t i n g effect o n c o n t r o l

theory a n d systems analysis. It will o p e n the d o o r to the d e v e l o p m e n t of systems which
will have a m u c h higher M I Q ( M a c h i n e Intelligence Q u o t i e n t ) t h a n t h o s e which could
be conceived a n d designed t h r o u g h the use of t r a d i t i o n a l m e t h o d s .

REFERENCES A N D R E L A T E D PAPERS

BELLMAN, R. E.,and ZADEH, L. A. (1970). Decision-making in a fuzzy environment, Management

Science, 17, B141-B164.
BERENJI, H. R. (1991). Fuzzy logic controllers, In An Introduction to Fuzzy Logic Applications in
Intelligent Systems, edited by R. R. Yager and L. A. Zadeh (Kluwer Academic Publishers),
pp. 69-96.
Fuzzy Sets, Fuzzy Logic, and Fuzzy Systems Downloaded from www.worldscientific.com
by NANYANG TECHNOLOGICAL UNIVERSITY on 11/10/15. For personal use only.

CHANG, S. S. L., and ZADEH, L. A. (1972). On fuzzy mapping and control, IEEE Trans. Systems,
Man, and Cybernetics, 2, 30-34.
DRIANKOV, D., HELLENDOORN, H., and REINFRANK, M. (1993). An Introduction to Fuzzy Control
(Springer-Verlag, Berlin).
DUBOIS, D., PRADE, H., and YAGER, R. (Eds.) (1993). Readings in Fuzzy Sets for Intelligent Systems
(Morgan Kaufmann, San Mateo).
HERTZ, J., KROGH, A., and PALMER, R. (1991). Introduction to the theory of neural computation
(Addison Wesley).
JAMSHIDI, M., VADIEE, N. and Ross, T. (Eds.) (1993). Fuzzy Logic and Control (Prentice Hall,
Englewood Cliffs, NJ).
JANG, J.-S. R. (1992). Self-learning fuzzy controller based on temporal back-propagation. IEEE
Trans, on Neural Networks, 3, 714-723.
KANDEL, A., and LANGHOLZ, G. (Eds.) (1994). Fuzzy Control Systems (CRC Press, Boca Raton).
KARR, C. (1991). Genetic algorithms for fuzzy controllers. AI Expert, 6, 26-33.
KOSKO, B. (1991). Neural Networks and Fuzzy Systems: A Dynamical Systems Approach to
Machine Intelligence (Englewood Cliffs, Prentice-Hall).
LEE, C. C. (1990). Fuzzy logic in control systems: Fuzzy logic controller, Part I and Part II, IEEE
Trans. Systems, Man and Cybernetics, 20, 404-418.
LEE, M. A. and TAKAGI, H. (1993). Integrating design stages of fuzzy systems using genetic
algorithms. 2nd International Conference on Fuzzy Systems (FUZZ-IEEE '95), pp. 612-
617, March 28-April 1.
LIN, C.-T. and GEORGE LEE, C. S. (1991). Neural-network-based fuzzy logic control and decision
system. IEEE Transactions on Computers, 40, 1320-1336.
MAMDANI, E. H. and ASSILIAN, S. (1975). An experiment in linguistic synthesis with a fuzzy logic
controller. Int. J. Man-Machine Studies, 7, 1-13.
PEDRYCZ, W. (1989). Fuzzy Control and Fuzzy Systems (John Wiley, New York).
SUGENO, M. and KANG, G. T. (1988). Structure identification of fuzzy model. Fuzzy Sets and
Systems, 28, 15-33.
TERANO, T., ASAI, K. and SUGENO, M. (1992). Fuzzy Systems Theory and its Applications
(Academic Press).
TAKAGI, T. and SUGENO, M. (1985). Fuzzy identification of systems and its applications to
modeling and control, IEEE Trans. Systems, Man and Cybernetics, 15, 116-132.
WANG, L.-X. (1994). Adaptive Fuzzy Systems and Control: Design and Stability Analysis (PTR
Prentice Hall, Englewood Cliffs, New Jersey).
WARWICK, K., IRWIN, G. W. and HUNT, K. J. (1992). Neural networks for control and systems
(London, Peter Peregrinus Ltd.).
WERBOS, P. J. (1974). Beyond Regression: New Tools for Prediction and Analysis in the Behaviour
Sciences, Ph.D. Thesis, Harvard University, Committee on Applied Mathematics.
ZADEH, L. A. (1971). Toward a theory of fuzzy systems, In Aspects of Network and System Theory,
R. E. Kalman and N. DeClaris (eds.) (New York: Rinehart & Winston), 469^190.
ZADEH, L. A. (1972). A rationale for fuzzy control, J. Dynamic Systems, Measurement and Control
94, Series G, 3-4.
ZADEH, L. A. (1973). Outline of a new approach to the analysis of complex systems and decision
processes, IEEE Trans. Systems, Man and Cybernetics, 3, 28-44.
ZADEH, L. A. (1974a). On the analysis of large scale systems, In Systems Approaches and
Environment Problems, H. Gottinger (ed.), 23-27 (Gottingen, Vandenhoeck and Ruprecht).
 The Role of Fuzzy Logic in Modeling ... [1994] 795

The role of fuzzy logic 203

ZADHH, L. A. (1974b). A fuzzy-algorithmic approach to the definition of complex or imprecise

concepts, Electronics Research Laboratory Report ERL-M474, University of California,
Berkeley, (also published in Inl. Man-Machine Studies, 8, 249-291, 1976).
ZADEH, L. A. (1975). The concept of a linguistic variable and its applications to approximate
reasoning, I, Information Sciences, 8, 199-249.
ZADEH, L. A. (1994). Fuzzy Logic, Neural Networks and Soft Computing, Comm. ACM, 37, 7 7 -
84.
ZIMMERMAN, H. J. (1990). Fuzzy Set Theory and Its Applications, second edition (Kluwer-Nijhoff).
Fuzzy Sets, Fuzzy Logic, and Fuzzy Systems Downloaded from www.worldscientific.com
by NANYANG TECHNOLOGICAL UNIVERSITY on 11/10/15. For personal use only.