You are on page 1of 9

Soft

n 0

Lomputing
;r’ 1
and Fuzzy
Logic
LOTFIA. ZADEH,University of California at Berkeley

0 Soft computing is a collection


of methodologies that aim to
exploit the tolerance for impreci-
sion and uncertainty to achieve
0 ne of the deepest time as a function of precision of solu-
traditions in science is that of accord- tion. As the data in Table 1 show, low-
ing respectability to what is quantita- ering the accuracy to 3.50 percent
tive, precise, rigorous, and categorical- reduces the computing time by an
ly true. It is a fact, however, that we order of magnitude for a ten-fold
live in a world that is pervasively increase in the number of cities.
imprecise, uncertain, and hard to be A more familiar example that illus-
wactability, robustness, and low categorical about. It is also a fact that trates the point is the problem of
solution cost. Its principal con- precision and certainty carry a cost. parking a car. W e find it relatively
Driven by our quest for respectability, easy to park a car because the final
stituents are fizzy logic, neuro- we tend to close our eyes to these facts position of the car is not specified pre-
and thereby lose sight of the steep cisely. If it were, the difficulty of park-
computing, and probabilistic rea- price we must pay for high precision ing would increase geometrically with
soning. Softcomputing is likely and low uncertainty. Another visible the increase in precision, and eventu-
concomitant of the quest for ally parking would become impossible.
to play an increasingly impor- respectability is that in much of the
tant role in many application scientific literature elegance takes Guiding principle. These and many
precedence over relevance. similar examples lead t o the basic
areas, including sof2ware engi- A case in point is the traveling premises and the guiding principle of
salesman problem, which is frequently soft computing.
neering. The role modelfor SOB used as a testbed for assessing the The basic premises of soft comput-
computing is the human mind. effectiveness of various methods of ing are
solution. What is striking about this + T h e real world is pervasively
problem is the steep rise in computing imprecise and uncertain.

48 07407459/94/$04 W 0 1994 IEEE NOVEMBER 1994


+ Precision and certainty carry a soft computing is aimed at accommo-
cost. dating the pervasive imprecision of the
The guiding principle of soft com- real world. Although soft computing
puting is has not as yet had a visible impact on Number Accuracy Computing
+ Exploit the tolerance for impre- software engineering, it is likely to do of cities
______~
time
cision, uncertainty, and partial truth to so in the years ahead. Among the areas
achieve tractability, robustness, and in which it is likely to be applied are
low solution cost. programming languages, computer
The label soft computing is grow- security, database management, user-
ing in use. What does it mean? Where 1 friendly interfaces, automated pro-

does it stand today and where is it gramming, fault diagnosis, and net-
~
Soztrce h$u. I‘07l Ti?iiu\ \1/19ib 12, 1991
headed? And what is the role of fuzzy working.
logic in soft computing? In this article, I will focus on some
i of the basic ideas that underlie soft In this article, I focus on fuzzy
Sample application. Some of the most 1 computing and relate them to its guid- logic.
striking examples of the application of ing principle.
~

the guiding principle of soft comput-


ing are the data-compression tech- FUZZY LOGIC CONCEPTS
niques that play a key role in high- SOFT COMPUTING AND FUZZY LOGIC
definition television and audio record- ~
As one of the principal constituents
ing and reproduction. ~ Basically, soft computing is not a of soft computing, fuzzy logic is play-
For example, in N H K s Muse sys- ’ homogeneous body of concepts and ing a key role in what might be called
tem, a motion-compensating tech- techniques. Rather, it is a partnership high MIQ (machine intelligence quo-
nique determines the outline, direc- of distinct methods that in one way or tient) systems.
tion, and speed of the moving body, another conform to its guiding princi- T w o concepts within fuzzy logic
then shifts the moving image without ple. L4tthis juncture, the dominant aim play a central role in its applications.
waiting to receive all the pixel data. of soft computing is to exploit the tol- + T h e first is a linguistic variable;
The resulting moving image does not erance for imprecision and uncertainty that is, a variable whose values are
have the resolution of the still picture. to achieve tractability, robusmess, and words or sentences in a natural or syn-
Muse exploits the fact that the human low solution cost. The principal con-
~ thetic language.?
eye cannot grasp the details of moving ’ stituents of soft computing are fuzzy + The other is afizzy $-then d e ,
objects with the same precision as still logic, neurocomputing, and proba- in which the antecedent and conse-
objects. Even more impressive is what bilistic reasoning, with the latter sub- quents are propositions containing lin-
is achieved in the recently developed suming genetic algorithms, belief net- guistic variables.!
digital HDTV systems. For example, works, chaotic systems, and parts of
~
The essential function of linguistic
the General Instrument system, learning theory. In the partnership of variables is that of granulation of vari-
instead of transmitting data for every fuzzy logic, neurocomputing, and ables and their dependencies. In effect,
color dot in a blue sky, sends an probabilistic reasoning, fuzzy logic is the use of linguistic variables and fuzzy
instruction to paint the sky. The com- mainly concerned with imprecision if-then rules results - through granu-
pression ratio this system achieves is and approximate reasoning; neuro- lation - in lossy data compression. In
on the order of 60 t o 1. In audio computing with learning and curve- this respect, fuzzy logic mimics the
recording and reproduction, similar fitting; and probabilistic reasoning remarkable ability of the human mind
ideas are embodied in Sony’s R/ID-l with uncertainty and belief propaga- to summarize data and focus on deci-
system and Philips’ DCC. tion. sion-relevant information.
In its current incarnation, the con- In large measure, fuzzy logic, neu- With regard to fuzzy logic, there is
cept of soft computing has links to rocomputing, and probabilistic rea- an issue of semantics that is in need of
many earlier influences, among them , soning are complementary, not com- clarification. Specifically, it is fre-
my 1965 paper on fuzzy sets;’ 1973 petitive. It is becoming increasingly
~
quently not recognized that the term
paper on the use of linguistic variables clear that in many cases it is advanta- fuzzy logic is actually used in two dif-
in the analysis and control of complex geous to combine them. A case in ferent senses. In a narrow sense, fuzzy
systems;‘ and 1979 report (1981 paper) point is the growing number of “neu- logic (FLn) is a logical system - an
on possibility theory and soft data rofuzzy” consumer products and sys- extension of multivalued logic that is
analysis.3 tems that use a combination of fuzzy intended to serve as a logic of approxi-
Unlike traditional hard computing, logic and neural-network techniques. mate reasoning. In a wider sense, fuzzy
~~~ . ~

IEEE SOFTWARE 49
from the knowledge of the member-
ship functions of the so-called przmary
terms -young and old, for example. A
primary term plays the role of a gener-
ator whose meaning (its membership
function) must be calibrated in context.
For example, the meaning of not very
young might be computed as
knot i.m"OUnE (4 = 1 - (Jlyoung(4)2
where very plays the role of an intensi-
fier and young is a primary term whose
membership function is specified in
context.
Most current applications of fuzzy
logic employ a simpler framework,
Figure 1. Interpretation of' middle-aged as a linguistic value. illustrated in Figure 2 . Specifically, the
membership functions are assumed to
be triangular or trapezoidal, and the
number of linguistic values is usually in
logic (FLm)is more o r less synony- robustness. and low solution cost. the range of three to seven.
mous with fuzzy set theory; that is, the The concept of a linguistic variable
theory of classes with unsharp bound- linguistic variables. A concept in fuzzy plays a central role in the applications
aries. In this perspective, FL = FLU, logic that plays a key role in exploiting of fuzzy logic because it goes to the
and FLn is merely a branch of FL. the tolerance for imprecision is the heart of the way in which humans per-
\%%at is important to recognize is that linguistic variable. A linguistic vari- ceive, reason, and communicate.
today the term fuzzy logic is used pre- able, as its name suggests, is a variable Quintessentially, the use of words
dominantly in its wider sense. It is in whose values are words or sentences in may be viewed as a form of data com-
this sense that any field X can be a natural or synthetic language. For pression that exploits the tolerance
"fuzzified" - and hence generalized example, age is a linguistic variable if for imprecision to achieve tractabili-
- by replacing the concept of a crisp its linguistic values are young, old, mid- ty, robustness, and economy of com-
set in X by a fuzzy set. In application to dle-aged, very old, not very young, and so munication. This fits almost precisely
basic fields such as set theory, arith- on. A linguistic variable is interpreted the guiding principle of soft comput-
metic, topology, graph theory, proba- as a label of a fuzzy set that is charac- ing.
bility theory, and logic, fuzzification terized by a membership function, as
leads to fuzzy set theory, fuzzy arith- illustrated in Figure 1. Thus, if u is a Granulation. In a related sense, the
metic, fuzzy topology, fuzzy graph the- numerical age, say 53, then F , ~ use~ of ~ words~ may
~ be viewed as a form
ory, and fuzzy logic in its narrow sense. ,A53) is the grade of membership of 53 of fuzzy quantization or more general-
Similarly, in application to applied in middle-aged. Subjectively, you ma) ly as granulation, as Figure 3 shows.
fields like neurocomputing, stability interpret F ~ ~ ~ ~as ~the~ degree . ~ ~ tc~ ~ ( Basically,
u ) granulation involves a
theory, pattern recognition and mathe- which u fits your perception of mid- replacement of a constraint of the form
matical programming, fuzzification dle-aged in a specified context. X=a
leads to fuzzy neurocomputing, fuzzy In a general setting, a linguistic
stability theory, fuzzy pattern recogni- variable, V , can be viewed as 2 with a constraint of the form
tion, and fuzzy mathematical program- microlanguage with context-free gram- Xis A
ming. What is gained through fuzzifi- mar and attributed-grammar seman-
cation is greater generality, higher tics. The context-free grammar define: where A is a fuzzy subset of U, the uni-
expressive power, an enhanced ability the legal values of V. For example, ir verse of X.For example,
to model real-world problems, and - the case of age, the legal values arc x=2
most important - a methodology for young, not young, not very young, quit(
exploiting the tolerance for impreci- old, middle-aged, and so on. The attrib. might be replaced with
sion, a methodology that fits the guid- uted-grammar semantics provides ; X is small
ing principle of soft computing and mechanism for computing the mem
thus serves to achieve tractability, bership function of any value of L In fuzzy logic, X is a is interpreted as a

50 N O V E M B E R 1994
characterization of the possible values
of X , with A representing a possibility
X 0
distribution. Thus, the possibility that X
X can take a value U is given by __ __-
Pon{X = U ] = M(U)

It is in this sense that X is ~ ( u ) with


,
possibility interpreted as ease of attain-
ment or assignment, may be interpret-
ed as an elastic constraint on X .

1 .o

0
lSJ

0
3oc\.
Figure 3. (A) Quantization versus (B) granulation @zzy quantization).

achieving a higher MIQ. Y,,...,Y,.and has the form:


Interestingly, the development of if X , is A , and ... X , is A,
fuzzy-set theory was motivated by the then Y,is B , and ... Y,is B,
first situation, but today most applica-
tions of fuzzy logic in the realm of where X = ( X , ,...,X,) and Y = ( Y ,
consumer products are motivated by ,...,YJ are linguistic variables and (A,
the second. ,...4,)and (B, ,...,B,) their respective
F D C L has many facets. Here, I linguistic values. For example:
shall sketch some of the basic ideas if Pressure is high and Tmperature
that underlie FDCL and the calculi of is high then Volume is small
fuzzy rules and fuzzy graphs.
Like any language, FDCL is char- For simplicity, I will discuss only rules
not, the calculi of fuzzy rules and fuzzy acterized by its syntax and semantics. in which m= n = 1.
graphs provide an alternative method- T h e syntax of F D C L is concerned A rule can have a rurface stmctzlre or
ology aimed at results that are in the with the form of admissible fuzzy a deep stmcture. The surface stmcture is
spirit of the guiding principle of soft rules; the semantics is concerned with the rule in its symbolic form:
computing. their meaning. It is important to note i f X i s A then Y i s B
that FDCL is not a “fuzzified”version
of a standard programming language, Such a rule is said to be uncalibrated,
as is true of Fuzzy Prolog.6 which means that the membership
functions of A and B are not specified.
Fuzzy rules. FDCL allows the use of T h e deep structure is the surface
a wide variety of fuzzy if-then rules, or structure together with a characteriza-
simply fuzzy rules. A typical fuzzy tion of the membership functions of
rule relates m antecedent variables linguistic values of variables. In this
X I ,...,X , to n consequent variables, case, the rule is said to be calibrated.
-
51
. .-

Low High

W
n
P
s
’ 30
Speed
130 180
[A)

i-
- -
Control rules
1. If (speed 1s low) ond (shift IS high) lhen (-3) 3 If (throt IS low) and (speed is high) then (t3) 5 If (throt is high) ond (speed is high) then (-1)
14 2 ll (speed 15 high) and (shift is low) then (4)4 If (throt IS low) and (speed is low) then (tl) 6 If (throt is high) and (speed IS low) then (-3)
- - ._~__ - ._ - - - __ ._ -- _____- -- - - _.
Figure 4. Fuzzy rules used in Honda ifizzy-logic transmission. Here, the meaning of the numeric values associated with the
d e s is not important; they only illustrate how d e s are calibrated

Figure 4 shows an example of the cali- ’ later, one of the central problems in the proposition ;l/lary is young might be
brated fuzzy rules used in Honda’s
fiiz7y-logic transmission. .As I explain
~

1
the applications of fuzzy logic is that of
deriving the deep structure of a set of 1 expressed as
,wdF3,isvozrng--$ is young

.______
-
71
fuzzy rules from U 0 data.
I where Age(Mary) is the focal variable
and young is a fuzzy constraint on
, DERIVING RULES AND GRAPHS Age(lMd?y).

’i
Applying this concept of meaning
...................
In the semantics of FDCL, the representation to the fuzzy rule
1
~

basic questions are, what is the mean- i f x i S A then y i s


ing of a single rule and what is the
meaning of a collection of rules? we can express the meaning of the rule

!I I in question as a fuzzy constraint on the


Deriving rules. Consider the simplest joint variable (X,Y).More specifically,
I ? type of rule:
1


U
i f X i s A then I’is B + (X,qis A x B ,
1,
A
n

I if X is A then Y is B I
where A x B is the Cartesian product
Fig41-e 5. .I x E ~ n t c i p ctrd
r 1ir n f i i z : ) where A and B are linguistlc values of of A and B The membership function
X and Y, respectlvely. The questlon is, of A x B is given by
what 1s the meaning of this rule P e n
the membership functlons of A and B? 1
p , x R (.,i) = p I (U) r\ p&)

1 In fuzzy logic, the meaning of a where A is the conjunction operator,


f‘

’ film
p+Z1sC
1
propositlon p is expressed as a ranonzcal usually defined as min. A x B may be
interpreted as afizzy point or a grunule,
as shown in Figure 5 .
I
I
I where + means “translates into,” Z is Deriving graphs. In the case of a col-
1
the constrained variable, and C is an lection of rules expressed as
fuzzy relation that plays the role of a i f X i s , 4 , t h e n I ’ i s B , , t = l , ...,n
fuzzy constraint on 2.W h a t this
implies is that the meaning of p is and the meaning of the collection is
expressed as a fuzzy - or, equivalent- defined as
Figzlre 6. InteVretation of a colkction ly, elastic - constraint on a designated IfXisA, then 1-1sB,,
of&zzy rules as a@zzy graph. variable. T o illustrate, the meaning of (2 = 1, ...,n) + (XY)1s
-

52 NOVEMBER 1994
(A,x B , + ... +A,>xBJ
where + is used in place of v to denote
the disjunction operator, which is usu-
ally defined as max. For simplicity, the
right-hand member of the collection 1 ~

may be written as
(X,I.? is (E, 4 x BJ
T h e expression 1,A, x B, may be Figure 7. Rcpm~titingn collection of fiizzy rr1le.r ns n fiizzy gmph, f: which
viewed as a superposition of fuzzy nppl~orlltI"rer to f.
points o r granules, as illustrated in
Figure 7 . In effect, it represents a
coarse - or, equivalently, compressed
- characterization of the dependency
and for this reason it is called a&zzy
graph.5 Thus, a collection of fuzzy
rules is represented as a fuzzy graph.
For example:
if X is small then Y is small
if X is medium then Y is large I

if X is large then Y is small

is a coarse characterization of the


dependency illustrated in Figure 6.

INTERPOLATION

If we interpret a collection of rules Figure 8. Interpolation of a@zzy gvaph. The value of Y may be interpreted as the
as a coarse representation of the func- projection of the intersection of thehzzy graph with the cylindrical extension ofA.
tional dependence of Yon X , the prob-
lem of interpolation may be defined as
that of computing the value of Y given (A', kJ is K
a value of X that may not be a perfect or, equivalently,
XisA
match with any of the antecedent vari- B = Z, p, A B ,
ables in the collection. More specifi- Y is R*A
cally, this problem can be expressed as in which the composition operation is in which
the inference schema defined by P2 = (F,4,(.) A 1 ,(U)
(X,r? is (1,
A , x B,) p R e.3 (v)= rLip&R(zi,z') Ap4('))
The sequence of computations that
XisA in which pR(u,u)and pz4(u)are, respec- leads to B is standard in most fuzzy
Y is?B tively, the membership functions of R logic applications and is usually imple-
in which ? B signifies that B is the and A. mented in software or hardware. In
object of computation. In graphical In the example considered earlier, R some implementations, called max-
terms, as shown i n Figure 8 , the is given by product implementations, the conjunc-
problem may be viewed as that of tion A is interpreted as the arithmetic
assigning a linguistic value to X and R = (E,A,x B,) product.
computing the corresponding linguis- Interpolation lies at the heart of the
tic value of Y. utility of fuzzy rule-based systems
In fuzzy logic, computation of B is because it makes it possible to employ
carried out through the basic rule of a relatively small number of fuzzy rules
inference, called the compositional rule to characterize a complex relationship
of inference.? T h e rule in question between two or more variables. In a
reads typical application in a consumer

IEEE SOFTWARE 53
INDUCING RULES FROM OBSERVATION

One of the central problems in the


applications of fuzzy logic relates to
the question, how can rules be inferred
from observations; that is, from the
knowledge of a collection of U 0 pairs?
[n the context of self-organizing sys-
tems, this problem was first formulat-
ed and analyzed by T.J. Procyk and
E.H. Mamdani. Later, a seminal paper
by T. Takagi and M. Sugeno made a

1 Figure 9. Interpolation of t w o mles and defizzq5cation.


major contribution.9
During the past several years,
researchers have made important
advances toward at least a partial solu-
tion to the problem by applying neur-
product or industrial-control system, al-network techniques or, more gener-
the number of rules is on the order of i f X i s A , then Yis B,
if X is A, t h e n Y is B, ally, dynamic and gradient program-
i o to 20.73 ming,7.10-J1 Other promising approach-
In applications in which B plays the and the input is es involve the use of genetic algo-
role of a control variable (an input to a Xis A rithmsl?$13and reinforcement learn-
motor, for example), B must be ing.'+ Figure 10 summarizes the ways
"defuzzified" - converted to a single- m l and m2 are, respectively, the to derive the deep structure of a set of
ton - before it is applied. In current degrees to which A matches A1 and rules from the surface structure.
practice, the center-of-gravity method A2. The expression for the output is A basic idea underlying these
is generally used to achieve defuzzifi- approaches involves representing a
P B ( U ) = 2, P/ A P d 4
cation. Figure 9 shows a simple exam- fuzzy rule-based system as a multilay-
ple of interpolation and defuzzifica- which upon center-of-gravity dehzzi- ered structure, such as that shown in
tion. fication leads to a numerical value of Figure 11. ' I In a simple version of this
In this figure, the rules are E. architecture11 that is rooted in the
Takagi-Sugeno-Kang approach,Y the
rules are assumed to be of the form
(ifX, isA,, and ... andX,isA,,
then Y =bJ, i = I,..,,n

where b, are constants (singleton con-


sequents). If the numerical values of
X I , ... X , are U,,, ... U,, , respectively,
and the grades of membership of u I ,,
... U,, , in A,,, ... A,, are p Ir(ul,),...
~ ~ ~ then ( u ~ the~ combined
) , degree to

41 1
Grodient pragromming which the input n-tuple X ( u I , , ... u,J
Genetic:olgorithms 1 matches the antecedents is taken to be
1 Reinforcementlearning f I I the product
m,= P I I ( U l i I,... Ilrn,(U,"J,

Then, defining the normalized weight


wras
w, = m,/m i + ... + mrn

1
l!
Figure 10. Summary of alternative methods t o deduce the deep stmctures of a set
oj-rmles.
the output is expressed as
Y = Z, m,b,

54 NOVEMBER 1994
Note that in this architecture there
is no defuzzifier because the inputs X,,
... X , are assumed to be singletons.
In the application of gradient pro-
gramming to this architecture, the
membership functions of A,, , ... A,,
are assumed to be triangular, trape-
zoidal, or Gaussian in form. Then,
using backward iteration, the values of
membership-function parameters are
computed from right to left.10-11 In
this way, from the knowledge of 110
pairs we can compute the values of
parameters and thereby induce the
rules from observations. Figure 11. Representing a fuzzy Jystem as a multilayered structure. Il and N
The approach sketched here is one denote multipliers and normalizers, respectively.
way the methodologies of fuzzy rule-
based systems and neural networks can
be combined, leading to “neurofuzzy”
systems. Such systems are growing in of equations governing system behav- algorithm), expressed in terms of
number and visibility and are illustra- ior - can be employed because we do uncalibrated fuzzy rules of the form
tive of the advantages derived from not know how to model a ball rolling if State is A then Action is B
combining soft computing’s con- or sliding on a rug-like surface. This
stituent methodologies. In this con- rules out the use of classical control 3. Test the system without trying
text, it is important to note that inter- theory as well as any approach that to solve the problem
polation and induction of rules from requires simulation. 4. Calibrate the fuzzy rules in step
observations are key issues in both The set-interval may be viewed as a 2 using metarules, rules that modify
fuzzy logic and neurocomputing. disjunctive goal. This feature makes it other rules
difficult to employ neural-network 5. Test the algorithm constructed
techniques. in step 2.
FUZZY BALL AND BEAM PROBLEM By contrast, the problem is easy to 6. Refine the calibrated fuzzy rules
solve with fuzzy logic because it is rel- derived in step 5 .
atively easy for a human. In fact, the 7 . Iterate steps 5 and 6 until the
presence of a fuzzy layer makes the ball stays in the set-interval.
ball-and-beam problem easy for
humans and difficult or impossible for
alternative methodologies. As in most
fuzzy-logic applications, the solution is
in effect a translation of a human solu-
tion into FDCL. A human solution
would normally involve seven steps:
1. Compile uncalibrated fuzzy or
crisp rules from knowledge of natural
laws, to govern the behavior of the
ball and beam. For example
if 0 isPegative
then Y is positive
if 0 is,positive
then Y is negative
the more negative 0,
the more positive Y
the more positive 0 ,
t h e more negative Y
approach that is model-dependent - Figure 12. Fuzzy ball and beam prob-
in the sense of requiring a formulation 2. Construct a plan of action (an lem.

IEEE SOFTWARE 55
Translating these steps into a col-
lection of fuzzy rules expressed in
FDCL is by no means a trivial prob-
more difficult for a human because it
involves a conjunction of two goals:
+ confine the motion of the ball to
A lthough soft computing is still in
its initial stages of evolution, i t is
lem. This is particularly true of the so- the prescribed set-interval [a,,a,], and rapidly growing in importance and
called gradual rulesls of the form + enter the set-interval a t a time t tisibility. In the years ahead, soft com-
the more Xis A the more Y is B which is constrained to lie in a pre- puting and its principal constituents
scribed temporal set-interval [t,,t J . - fuzzy logic, neurocomputing, and
because such rules describe the global In this case, formulating a human probabilistic reasoning - are likely to
behavior of a functional dependency solution and translating it into FDCL emerge as essential tools for the con-
rather than its local properties. is a real challenge. W e do not yet ception, analysis, and design of high
However, what is important is that, completely understand how to apply MIQ systems. In the final analysis, the
though it is not easy, it is feasible to fuzzy logic to problems like this. But it role model for soft computing is the
translate a human solution i n t o is evident that fuzzy logic - used human mind.
FDCL, whereas it is not feasible to alone or in combination with neuro-
translate it into anal~aicaltechnisues. computing and probabilistic reasoning
Now suppose you wanted the ball - is the methodology of choice when
to reach the set interval at some time t ~ analytic models are impossible or hard
and stay there. This is significantly to formulate.

ACKNOWLEDGMENTS
This research was supported in part by the BISC Prograin, N . A S h (;rant NCC 2-?ij, EPRI
Agreement RP 8010-34 and MICRO State Program KO.92-180.

REFERENCES
I. LA. Zadeh, “Fuzzy Sets,” Infirmation and Conti-ol,J u n c 196.i, pp. 3 38-15!.
2. L.A. Zadeh, “Outline of a New Approach to the Analysis of Complex Systems and Dccision
Processes,” IEEE Trans. Systems, Man and Cybmzetics, 1973, pp. 28-44,
3. L.A. Zadeh, “Possibility Theory and Soft Data Analysis,” .\.lathernotical Frontreis ofthe Social und
PoliT Sciences, L. Cobb and R. M. Thrall, eds., \Vesmew Press, Boulder, Colo., 1081, pp.69-129. ..
4. L.A. Zadeh, “Fuzzy Logic, Neural Network? and Soft Computing,” fConrrn. :IC.%f,Mar. 1994, tions to artificial intelligence, linguistics, logic,
pp. 77-84. decision analysis, expert systems, and neural net-
5. L.A. Zadeh, “On the Analysis of Large-Scalc Systems,” in $.itmu ,?pproache.r and Enmnmnrnzt work.
Problems, H . Gottinger, ed., L’andenhoeck and Ruprecht, <;orringen, lY74, pp. 23-3 7 , Zadeh is a graduate of the University of
6. M. Mukaidono, Z.L. Shen, and L. Ding, “Fundamentals of Fuzzy Prolog,” lnt’/y.:lpproxmrutr Teheran, M I T , and Columhia University. H e has
Reasoning, No. 3, 1989, pp. 179-193. received honorary doctorates from t h e Paul-
7. B. Kosko, Neural Networks and Fuzzy Systems: ‘-1 L)ynnrnical Sy.itenrs. Ippi-nachto .llai.hriw Sabatier University, France. the State Universih of
Intelligence, Prentice-Hall, Englewood Cliffs, NJ., 1091. Nen York at Ringhaniton, Dortmund University,
8. T. Terano, K. Asai, and M. Sugeno, “Fuzzy Systems T h e o n and lts Xpplications,” .Academic Germany; and the University of Granada and
Press, San Diego, Calif., 1992. Oviedo, Spain, in recognition of his development of
9. T. Takagi and M. Sugeno, “Fuzzy Identification of Systems and it5 .Applications to &lodeling the theory of fuzzy sets. He is a fellow of the IEEE,
and Contro1,”IEEE Trans. Systems, Man, and Cvhevwtirr l i . 1985, pp. 116-132. AA~AS,hCh1, and .AAAAI and a member of the
10. C:T. Lin and C.S. George Lee, “Neural-Network-Based F u z q Logic Control and Decision National -\cademy of Engineering and the Russian
System,”IEEE Trans. Computeir. Dec. 1991, pp. 1310-1776. Academy of Natui 11 Sciences.
11.J.-S.R. Jang, “Self-Learning Fuzzy Controller Based on Temporal BacL~Pr(,pagation,” IEEE
Trans. Neural,Venuorks, May 1992, pp. 714-723.
12. C.C. Lee, “Fuzzy Logic in Control Systems: Fuzz)- 1-ogic Controller, Part 1 and Part 11,” IEEE
Trans. Systems, LVIan,and Cybernetics 20, 1990, pp. 404-418.
1 3 . C. Karr, “Genetic Algorithms for Fuzzy Controllers.”.?I Exprit, h-ov. 1001, pp. 26-33.
14. M.A. Lee and H . Takagi, “Integrating Design Stages of 1;uzzy Systems Using Genetic
Algorithms,” Proc. Int‘l Con$ on Fuzzy System. IFEE Press, S e w York, 1903, pp. 612-617.
15. D. Dubois and H. Prade, “Gradual lnference Rules in Approximate Rcasoning,” Z ~ r f b ~ m o t i o a Address questions ahout this article to Zadeh at
Sciences, 1992, pp. 103-122. the CS Div., D e p t . of E E C S , University of
California, Berkeley, Berkeley, Calif. 94720;
zadeh~cs.berkeley.edu

56 NOVEMBER 1994

You might also like