2% [IEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL, SMC, NO, asuany 1973) Outline of a New Approach to the Analysis of Complex Systems and Decision Processes LOTFI A. ZADEH Abstract—The approach described inthis paper represents substan- tive departure from the conventional quantitative techniques of system analysis. It has three main distinguishing featres: 1) wse of so-called “linguistic” variables in place of or in addition to numerical variables; 2) characterization of simple relations between variables by fuzzy ‘conditional statements; and 3) characterization of complex relations by ‘uray algorithms. ‘A linguistic variable defined a8 «variable whose values are sentences ‘n'a natural or artificial language. Ths, if tall, not tall, ery tall, very ery tal, ec. are values of height, then height is a lingulstie variable. Fuzzy conditional statements are expressions of the form 1 A THi2s By where A and J have fuzzy meaning, 1 is small rie y is large, ‘where small and large are viewed as labels of fuzy sets. A fuzzy algoritin {san ordered sequence of instructions which may contain fuzzy assignment and conditional statements, e.., = very small 1+ is small THEN ys {arge. The execution of such instructions is governed bythe compositional ‘ule of inference aud the rule ofthe preponderant alternative. ‘By relying on the use of linguistic variables and fuzzy algorithms, the approach provides an approximate and yet efetive means of describing the behavior of systems which are 0 complex or {00 ill-defined to admit of precise mathematical analysis. Its main applications le in economics, management science, artificial intelligence, psychology, linguistics, information retrieval, medicine, biology, and other fields In which the ‘dominant roles played by the animate rather than animate behavior of ‘system constituents. I. IntRopuction HE ADVENT of the computer age has stimulated a rapid expansion in the use of quantitative techniques for the analysis of economic, urban, social, biological, and other types of systems in which it is the animate rather than inanimate behavior of system constituents that plays a dominant role. At present, most of the techniques em- ployed for the analysis of humanistic, i., human-centered, systems are adaptations of the methods that have been developed over a long period of time for dealing with ‘mechanistic systems, ie., physical systems governed in the main by the laws of mechanics, electromagnetism, and thermodynamics. The remarkable successes ofthese methods in unraveling the secrets of nature and enabling us to build better and better machines have inspired a widely held belief that the same or similar techniques ean be applied with comparable effectiveness to the analysis of humani systems, As a case in point, the successes of modern control Manuscript ried August 1, 1972; vied August 13, 1972 ‘Ti work was sopprted by the Navy Escrone Stems Command ander Contact NOOOS.71.0555 the Army Research Ofice Dut ham, NC, under Grant DAVARO-DLTSUALGHE, ond NASA “be autor with the Depron of Era Enginerng and Suit with the Deparent of lea Engisring an Computer Senos and Electonics Research Laboratom, Unley oF Cioran Boke, Ca SP theory in the design of highly accurate space navigation systems have stimulated its use in the theoretical analyses of economic and biological systems. Similarly, the effective- ness of computer simulation techniques in the macroscopic analyses of physical systems has brought into vogue the use of computer-based econometric models for purposes of forecasting, economic planning, arid management. Given the deeply entrenched tradition of scientific think- ing which equates the understanding of a phenomenon with the ability to analyze it in quantitative terms, one is certain to strike a dissonant note by questioning the growing tendency to analyze the behavior of humanistic systems as if they were mechanistic systems governed by difference, differential, or integral equations. Such a note is struck in the present paper. Essentially, our contention is that the conventional ‘quantitative techniques of system analysis are intrinsically unsuited for dealing with humanistic systems or, for that matter, any system whose complexity is comparable to that cof humanistic systems. The basis for this contention rests con what might be called the principle of incompatibility Stated informally, the essence of this principle is that as the complexity of a system increases, our ability to make precise and yet significant statements about its behavior diminishes until a. threshold is reached beyond which precision and significance (or relevance) become almost mutually exclusive characteristics." It is in this sense that precise quantitative analyses of the behavior of humanistic systems are not likely to have much relevance to the real- world societal, political, economic, and other types of problems which involve humans either as individuals or in seroups. An alternative approach outlined in this paper is based ‘on the premise that the key elements in human thinking are not numbers, but labels of fuzzy sets, that is, classes of objects in which the transition from membership to non- membership is gradual rather than abrupt. Indeed, the pervasiveness of fuzziness in human thought processes suggests that miuch of the logic behind human reasoning is not the traditional two-valued or even multivalued logi but a logic with fuzzy truths, fuzzy connectives, and fuzzy rules of inference. In our view, itis this fuzzy, and as yet ‘not well-understood, logic that plays a basic role in what may well be one of the most important facets of human thinking, namely, the ability to summarize information—to extract from the collections of masses of data impinging + A corollary principle may be stated succinctly as, “The closer one Jooks ata reak-world problem, the Tuzziet becomes ie solution.”upon the human brain those and only those subcollections Which are relevant to the performance of the task at hand. By its nature, a summary is an approximation to what it summarizes. For many purposes, avery approximate characterization of a collection of data is sufficient because most of the basic tasks performed by humans do not require a high degree of precision in their execution. The human brain takes advantage of this tolerance for im- precision by encoding the “task-relevant” (or relevant”) information into labels of fuzzy sets wi ‘an approximate relation to the primary data. In this way, the stream of information reaching the brain via the visual, auditory, tactile, and other senses is eventually reduced to the trickle that is needed to perform a specified task with inimal degree of precision. Thus, the ability to manip- ulate fuzzy sets and the consequent summarizing capability constitute one of the most important assets of the human mind as well as a fundamental characteristic that dis- tinguishes human intelligence from the type of machine intelligence that is embodied in present-day digital com- puters, Viewed in this perspective, the traditional techniques of system analysis are not well suited for dealing with human- istic systems because they fail to come to grips with the reality of the fuzziness of human thinking and behavior. Thus, to deal with such systems realistically, we need ap- proaches which do not make a fetish of precision, rigor, ‘and mathematical formalism, and which employ instead a methodological framework which is tolerant of imprecision ‘and partial truths. The approach described in the sequel is a step—but not necessarily a definitive step—in this direction, ‘The approach in question has three main distinguishing features: 1) use of so-called “linguistic” variables in place of or in addition to numerical variables; 2) characterization of simple relations between variables by conditional fuzzy statements; and 3) characterization of complex relations by fuzzy algorithms. Before proceeding to a detailed discussion of our approach, it will be helpful to sketch the principal ideas behind these features. We begin with a brief explana- tion of the notion of a linguistic variable. 1) Linguistic and Fuzzy Variables: As already pointed ‘out, the ability to summarize information plays an essential role in the characterization of complex phenomena. In the case of humans, the ability to summarize information finds its most pronounced manifestation in the use of natural languages. Thus, each word x in a natural language L may be viewed as a summarized description of a fuzzy subset M(x) of a universe of discourse U, with M(x) representing the meaning of x. In this sense, the language as a whole may be regarded as a system for assigning atomic and ‘composite labels (ie., words, phrases, and sentences) to the fuzzy subsets of U. (This point of view is discussed in greater detail in [4] and [5].) For example, if the meaning of the noun flower is a fuzzy subset M(fiower), and the meaning of the adjective red is @ fuzzy subset M(red), then the meaning of the noun phrase red flower is given by the intersection of M(red) and M(flower), bear 2» If we regard the color of an object as a variable, then its values, red, blue, yellow, green, etc., may be interpreted as labels of fuzzy subsets of a universe of objects. In this sense, the attribute color isa fuzzy variable, that i a variable whose values are labels of fuzzy sets. It is important to note that the characterization of a value of the vatiable color by ‘4 natural label such as red is much less precise than the numerical value of the wavelength of a particular color. In the preceding example, the values of the variable color are atomic terms like red, blue, yellow, etc. More generally, the values may be sentences in a specified language, in which case we say that the variable is linguistic. To illustrate, the values of the fuzzy variable height might be expressible as tall, not tall, somewhat tall, very tall, not very tal, very very tal, tall but not very tll, quite tall, more ‘or less tall, Thus, the values in question are sentences formed from the label rall, the negation not, the connectives and and but, and the hedges very, somewhat, quite, and more or les. In this sense, the variable height as defined above is a linguistic variable. ‘As will be seen in Section III, the main function of linguistic variables is to provide a systematic means for an approximate characterization of complex or ill-defined phenomena. In essence, by moving away from the use of quantified variables and toward the use of the type of linguistic descriptions employed by humans, we acquire a capability to deal with systems which are much too complex to be susceptible to analysis in conventional mathematical terms. 2) Characterization of Simple Relations Between Fuzzy Variables by Conditional Statements: In quantitative ap- roaches to system analysis, a dependence between two ‘numerically valued variables x and y is usually charac- terized by a table which, in words, may be expressed as a set of conditional statements, e.g., Fx is 5 THEN y is 10, 1 x is 6 THEN y is 14, etc. ‘The same technique is employed in our approach, except that x and y are allowed to be fuzzy variables. In particular, if'x and y are linguistic variables, the conditional statements describing the dependence of y on x might read (the following italicized words represent the values of fuzzy var ables) x is small THEN y is very large xis not very small THEN y is very very large 1 x is not small and not large THEN y is not very large and so forth. Fuzzy conditional statements of the form uF A THEN B, where A and B are terms with a fuzzy meaning, eg., “IF John is nice to you THEN you should be kind to him,” are used routinely in everyday discourse. However, the meaning of such statements when used in communication between humans is poorly defined. As will be shown in Section V, the conditional statement 1 4 THEN B can be given a precise meaning even when 4 and Bare fuzzy rather than nonfuzzy sets, provided the meanings of A and B are defined precisely as specified subsets of the universe of discourse,0 In the preceding example, the relation between two fuzzy variables and y is simple inthe sense that it can be charac- terized as a set of conditional statements of the form IF A ‘THEN B, where A and B are labels of fuzzy sets representing the values of x and y, respectively. In the case of more complex relations, the characterization of the dependence of y on x may require the use of a fuzzy algorithm. As indicated below, and discussed in greater detail in Section VI, the notion of a fuzzy algorithm plays a basic role in providing a means of approximate characterization of fuzzy ‘concepts and their interrelations. 3) Fuzzy-Algorithmie Characterization of Functions and Relations: The definition of a fuzzy function through the use of fuzzy conditional statements is analogous to the definition of a nonfuzzy function f by a table of pairs (,f@), in which x is a generic value of the argument of f and f(x) is the value of the function. Just as a nonfuzzy function can be defined algorithmically (c.g., by a program) rather than by a table, so a fuzzy function can be defined by a fuzzy algorithm rather than as a collection of fuzzy conditional statements. The same applies to the definition of sets, relations, and other constructs which are fuzzy in nature. Essentially, a fuzzy algorithm [6] is an ordered sequence of instructions (like a computer program) in which some of the instructions may contain labels of fuzzy sets, e. Reduce x slightly it yi large Increase x very slightly i yis not very large and not very small If x is small then stop; otherwise increase x by 2 By allowing an algorithm to contain instructions of this type, it becomes possible t0 give an approximate fuzzy- algorithmic characterization of a wide varity of complex Phenomena. The important feature of such characteriza- tions is that, though imprecise in nature, they may be perfectly adequate for the purposes of a specified task. In this way, fuzzy algorithms can provide an effective means of approximate description of objective functions, con- straints, system performance, strategies, etc. In what follows, we shall elaborate on some of the basic aspects of linguistic variables, fuzzy conditional statements, and fuzzy algorithms. However, we shall not attempt to present a definitive exposition of our approach and its applications. Thus, the present paper should be viewed primarily as an introductory outline of a method which departs from the tradition of precision and rigor in scientific analysis—a method whose approximate nature mirrors the fuzziness of human behavior and thereby offers a promise of providing a more realistic basis for the analysis of humanistic systems. AAs will be seen in the following sections, the theoretical foundation of our approach is actually quite precise and rather mathematical in spirit. Thus, the source of impreci- sion in the approach is not the underlying theory, but the manner in which linguistic variables and fuzzy algorithms are applied to the formulation and solution of real-world problems. In effect, the level of precision in a particular [BEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, JANUARY 1973 application can be adjusted to fit the needs of the task and the accuracy of the available data. This flexibility constitutes cone of the important features of the method that will be described. IL. Fuzzy Sets: A SUMMARY OF RELEVANT PROPERTIES In order to make our exposition self-contained, we shall summarize in this section those properties of fuzzy sets which will be needed in later sections. (More detailed discussions of topics in the theory of fuzzy sets which are relevant to the subject of the present paper may be found in [T-O7)) Notation and Terminology A fuzzy subset 4 of a universe of discourse U is charac- terized by a membership function 4: U + [0,1] which associates with each element y of U a number j.4(9) in the interval [0,1] which represents the grade of membership of yin A, The support of A is the set of points in U at which 44()) is positive. A crossover point in A is an element of 0 ‘whose grade of membership in A is 0.5. A fuzzy singleton is a fuzzy set whose support is a single point in U. If A is fuzzy singleton whose support is the point y, we write A= nly en where jis the grade of membership of y in A. To be con- sistent with this notation, a nonfuzzy singleton will be denoted by I/p. ‘A fuzzy set A may be viewed as the union (see (2.27)) of its constituent singletons. On this basis, A may be repre sented in the form a= f nowy 2) Where the integral sign stands for the union of the fuzzy singletons s.4(y)/y. If A has a finite support {9.,y25°°"sYahs then (2.2) may be replaced by the summation Am daly bo + dale 23) A= ¥ uly Cn) in which jy, = I,--+vn, is the grade of membership of y in A. It should be noted that the + sign in (2.3) denotes the union (See (2.27) rather than the arithmetic sum, In this sense of +, a finite universe of discourse U = (yaya. 1) may be represented simply by the summation Uamnt yt ty es) or u-3% 26 although, strictly, we should write 2.5) and (2.6) as, Ua Wy + Wye to + My en and v= 3S tive 08)Diagrammatic representation of young and old Fig. 1. As an illustration, suppose that Uslt2+--+ 10 29) ‘Then a fuzzy subset? of U labeled several may be expressed as (the symbol A. stands for “equal by definition,” or “is defined to be,” or “denotes") several 8 0.5/3 + 08/4 + YS + 1/6 + 08/7 + 0.5/8 210) Similarly, if U is the interval [0,100], with y & age, then the fuzzy subsets of U labeled young and old may be represented as (here and elsewhere in this paper we do not differentiate between a fuzzy set and its label) yours = [uv + "(0+ PS2)) "oy ean Pos ))", ou =f" (0+ (see Fig. 1). Tne fn of moment na fy st maya be fuzzy set. For example, if U = TOM + JIM + DICK + BOB 2.12) au) and A is the fuzzy subset labeled agile, then we may have medium/TOM + low/3IM + low/DICK + high/BOB. (2.14) agile In this representation, the fuzzy grades of membership Jow, ‘medium, and high are fuzzy subsets of the universe V V=O+01+024° +0941 IS) Which are defined by Jow = 0.50.2 + 0.7/0.3 + 1/0.4 + 0.7/0.5 + 0.5/0.6 2.16 medium = 0.504 + 0.7)0.5 + 110.6 + 0.7/0.7 + 0.5/0.8 Q17 high = 0.5/0.7 + 0.7/0.8 + 0.9/0.9 + 1/1 2.18) 2 Aisa subset of B, written 4 < B,ifand only ifn all yin U. For example, the fuzzy set'A 0.1 + 03/2 Oaji + 032 + O81. 140) 5 wal), for Bs a sabace of 3 Fury Relations A fuzzy relation R trom a set Xto a set Yisa fuzzy subset of the Cartesian product X x Y.(X x Y isthe collection of ordered pairs (x), xX. € ¥), Ris characterized by 2 bivariate membership function (x,y) and is expressed Raf bso) 219) ar More generally, for an nary fuzzy relation R which is a fuzzy subset of Xj x My x Ky we have Ra May MCs Dy — eX, Fa lem 220) As an illustration, if X= {TOM, DICK} and_-¥ = (JOHN, JIM) then a binary fuzzy relation of resemblance between members of X and Y might be expressed as. resemblance = 0.8|(TOM, JOHN) + 0.6/(TOM, JIM) + 0.2(DICK, JOHN) + 0.9/(DICK, JIM), Alternatively, this relation may be represented as a rela- tion matrix: JOHN JIM TOM DICK 08 06) 02 03. 221) in which the (j,/)th element is the value of gig(x,») for the ith value of x and the jth value of y. If Ris a relation from X to ¥ and S is a relation from ¥ to Z, then the composition of R and S is a fuzzy relation denoted by Ro S and defined by ReSA] v Uisboy) A mv2)G52) (2.22) where v and a denote, respectively, max and min.* Thus, for real a,b, a, faz a. ieess. , a, ifa 25. (2.54) 3 ‘Then we can represent the fuzzy subset of U labeled young as (see (2.1) yang = fy + ['"(1+ (S2))'v e509 wih he send monber of 235 pening the Tango at isd msi te nein af oni el Dials ts an tsar i etna acarOPe ago rte Tew ong es III, Lincuisric Hepors As stated in Section Il, the values of a linguistic variable are labels of fuzzy subsets of U which have the form of phrases or sentences in a natural or artificial language. For example, if U is the collection of integers U=04F1424--+4 100 @1) and age isa linguistic variable labeled x, then the values of x might be young, not young, very young, not very young, (old and not old, not very old, not young and not old, et. Tn general, a value ofa linguistic variable is a composite term x = x;%,""* 35 Which is a concatenation of atomic terms x,,"* "ay These atomic terms may be divided into four categories 1) primary terms, which are labels of specified fuzzy subsets of the universe of discourse (e.g, young and cold in the preceding example); 2) the negation not and the connectives and and or; 3) hedges, such as very, much, slightly, more or less (although more or less is comprised of three words, it is regarded as an atomic term), etc 4) markers, such as parentheses. ‘A basic problem P; which arises in connection with the use of linguistic variables isthe following: Given the mean- ing of each atomic term x,, i = 1,*"*7, in @ composite term x = x, "+x, which represents a value of a linguistic variable, compute the meaning of x in the sense of (2.53). This problem is an instance of a central problem in quan- titative fuzzy semantics [4], namely, the computation of the meaning of a composite term. P; is a special case of the latter problem because the composite terms representing the values of a linguistic variable have a relatively simple grammatical structure which is restricted to the four categories of atomic terms 1)-4). ‘As a preliminary to describing a general approach to the solution of P,, it will be helpful to consider a subproblem ‘of P; which involves the computation of the meaning of a composite term of the form x = fu, where iis a hedge and isa term with a specified meaning; ¢.g,,x = very tall man, where h = very and uw = tall man. Taking the point of view described in [15], a hedge f ‘may be regarded as an operator which transforms the fuzzy set M(u), representing the meaning of u, into the fuzzy set ‘M(lu). As stated already, the hedges serve the function ofFig. 2. Elect of hedge very generating a larger set of values for a linguistic variable from a small collection of primary terms. For example, by using the hedge very in conjunction with not, and, and the primary term tall, we can generate the fuzzy sets very tall, tery very tall, not very tall, tall and not very tall, ete. To define a hedge h as an operator, it is convenient to employ some of the basic operations defined in Section II, especially concentration, dilation, and fuzzification. In what follows, ‘we shall indicate the manner in which this can be done for the natural hedge very and the artificial hedges plus and ‘minus. Characterizations of such hedges as more or less, much, slighty, srt of, and essentially may be found in [15] ‘Although in its everyday use the hedge very does not have a well-defined meaning, in essence it acts as an intensifier, generating a subset of the set on which it operates. A simple operation which has this property is that of con- centration (see (2.44)). This suggests that very x, where x is a term, be defined as the square of x, that is very x & x? G2 more explicitly very x & jf wey. G3) For example, if (see Fig. 2) xe otdmen af (14+ (25%) ")'y oo) very old men = ie (+ (4) then yo. 65) ‘Thus, if the grade of membership of JOHN in the class of ‘old men is 0.8, then his grade of membership in the class of | very old men is 0.64. As another simple example, if Uaslt2+34445 66 and small = 1/1 + 08/2 + 0.6/3 + 0.4/4 + 02/5 G7) then very small = 1/1 + 0.64/2 + 0,36/3 + 0.16/4 + 0.04/5. G8) ‘Viewed as an operator, very can be composed with itself, Thus very very x = (very x)? = xt. 6% For example, applying (3.9) to (3.7), we obtain (neglecting ssmall terms) very very small = 1/1 + 0.4/2 + 0.1/3. G.10) TEE TRANSACTIONS ON SYSTEMS, MAN, AND CYRERNETICS, ANUARY 1973 In some instances, to identify the operand of very we have to use parentheses or replace a composite term by an atomic one. For example, it is not grammatical to write X = very not exact ein but if not exact is replaced by the atomic term inexact, then x = very inexact G.12) is grammatically correct and we can write x = (nexact)? @.13) Note that not very exact = —(very exact) = (exact?) (3.14) is not the same as (3.13). The artificial hedges plus and minus serve the purpose of providing milder degrees of concentration and dilation than those associated with the operations CON and DIL (ee (2.44), (2.45), Thus, as operators acting on a fuzzy set labeled x, plus and minus are defined by plus x & xt minus x 2 38-78 G15) G16) In consequence of (3.15) and (3.16), we have the ap- proximate identity plus plus x = minus very x. Gin As an illustration, if the hedge highly is defined as Aighly = minus very very G18) then, equivalently, highly = plus plus very. G19) As was stated earlier, the computation of the meaning of composite terms of the form fu is a preliminary to the problem of computing the meaning of values of a linguistic variable. We are now in a position to turn our attention to this problem, IV. Computarion oF THE MEANING oF VALUES OF & LINGUISTIC VARIABLE Once we know how to compute the meaning of a com= posite term of the form /u, the computation of the meaning ‘of a more complex composite term, which may involve the terms not, or, and and in addition to terms of the form fn, becomes a relatively simple problem which is quite similar to that of the computation of the value of a Boolean expression. As a simple illustration, consider the computa- tion of the meaning of the composite term not very small an where the primary term small is defined as small = 1/1 + 08/2 + 0.6/3 + 0.4/4 + 0.2/5 (4.2) with the universe of discourse being Ual4+2434445, 43){ZADEIL: ANALYSIS OF COMPLEX SYSTEMS AND DETECTION PROCESSES By (2.8), the operation of very on small yields very small = 1/1 + 0.64)2 + 0.36)3 + 0.16/4 + 0.04/5 4) and, by (2.26), not very small = (very small) = 0,36 /2 + 0.64/3 + 0.84/4 + 0.96/5 0.4/2 + 0.6/3 + 0.8/4 + 1/5. (4.5) As a slightly more complicated example, consider the composite term Xx = not very small and not very very large where large is defined by large = 0.2/1 + 0.4/2 + 0.6/3 + 0.8/4 + 1/5. (4.7) In this case, 4.6) ery large = large* 0.04/1 + 0.16/2 + 0.36/3 + 0.64/4 +15 (48) very very large = (large? = 0.1/3 + 0.4/4 + 1/5 49) not very very large = 1/1 + 1/2 + 0.9/3 + 0.6/4 (4.10) and hence not very small and not very very large = (04/2 + 0.6/3 + 0.8/4 + 1/5) AL + 172 + 093 + 0.6/4) = 04/2 + 0.6/3 + 0.6/4. (4.11) ‘An example of a different nature is provided by the values of a linguistic variable labeled likelihood. In this case, we assume that the universe of discourse is given by U=0+01402403 404405 +06 407+ 0840941 4.12) in which the elements of U represent probabilities. Suppose that we wish to compute the meaning of the value x = highly unlikely (413) in which highly is defined as (see (3.18)) highly = minus very very a4) and unlikely = not likely is) ‘with the meaning of the primary term likely given by likely = 1/1 + 1/0.9 + 1/08 + 0.8/0.7 + 0,6)0.6 + 0.5/0.5 + 0.3/0.4 + 0.2/0.3. (4.16) Using (4.15), we obtain 10 + 1/0.1 + 1/0.2 + 0.8/0.3 + 0.7/0.4 + 05)0.5 + 0.4/0.6 + 0.2/0.7. (4.17) unlikely 35 and hence ery very unlikely = (unlikelyy* 2 1/0 + 1/01 + 1/0.2 + 0.4/0.3 + 0.2/0.4. (4.18) Finally, by (4.14) highly unlikely = minus very very unlikely (1/0 + 1/01 + 0.2 + 04/03 + 0.2)0.)°7 = 1/0 + 1/0.1 + 1/0.2 + 0.5/0.3 + 03/04. (4.19) It should be noted that in computing the meaning of ‘composite terms in the preceding examples we have made implicit use of the usual precedence rules governing the evaluation of Boolean expressions. With the addition of hedges, these precedence rules may be expressed as follows. Pret pen Fist not Second and Third on ‘As usual, parentheses may be used to change the precedence order and ambiguities may be resolved by the use of asso- ciation to the right. Thus plus very minus very tall should be interpreted as plus (very (minus (very (tall))). ‘The technique that was employed for the computation of the meaning of a composite term is special case of a ‘more general approach which is described in [4] and [5] ‘The approach in question can be applied to the computa~ tion of the meaning of values of a linguistic variable pro- ‘vided the composite terms representing these values can be generated by a context-free grammar. As an illustration, ‘consider a linguistic variable x whose values are exemplified bby small, not small, large, not large, very small, not very small, small or not very very large, small and (large or not small), not very very small and not very very large, el. ‘The values in question can be generated by a context-free grammar G = (Vy,Vq)5,P) in which the set of terminals Vz comprises the atomic terms small, large, not, and, or, very, etc.; the nonterminals are denoted S, 4, B, C, D, and E; and the production system is given by SoA caD S+Sord C>E AOB D- very D As AadB E-vveryE Bac D— small BntC E> large c+). (4.20)Each production in (4.20) gives rise to a relation between the fuzzy sets labeled by the corresponding terminal and nonterminal symbols. In the case of (4.20), these relations are (We omit the productions which have no effect on the associated fuzzy sets) S+ Sord=S,= Sp + Ap A> Aand B= Ay = Ago Be Bo not C= B, = Cy D = very D> Dy, = Dy? E> very BB, = Eye D~ small > D, = small E> large = By, = large 2 in which the subscripts L and R are used to differentiate between the symbols on the left- and right-hand sides of a production. To compute the meaning of a composite term x, it is, necessary to perform a syntactical analysis of x in terms of the specified grammar G. Then, knowing the syntax tree of % one can employ the relations given in (4.21) to derive a set of equations (in triangular form) which upon solution yield the meaning of x. For example, in the case of the ‘composite term 2 = not very small and not very very large the solution of these equations yields x = (Asmall?) 6 (Alarge*) 422) which agrees with (4.11). Details of this solution may be found in [4] and [5]. ‘The ability to compute the meaning of values of a lin- guistic variable is a prerequisite to the computation of the ‘meaning of fuzzy conditional statements of the form 18 4 {THEN B, eg. 1 xis nof very small THEN y is very very large. ‘This problem is considered in the following section. \V. Fuzzy Conprmiowat, STATEMENTS AND COMPOSITIONAL RULE OF INFERENCE In classical propositional calculus,’ the expression F A tweN B, where A and B are propositional variables, is written as A => B, with the implication = regarded as a ‘connective which is defined by the truth table a rT ror RT FOP ‘Thus, AsB=7AVB 6D * A detailed discussion ofthe significance of implication and its role {in modal loge may be found in [18]. TEBE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, sANUARY 1973 in the sense that the propositional expressions 4 => B (A implies B) and A v B (not A or B) have identical ‘ruth tables. ‘A more general concept, which plays an important role in our approach, is a fuzzy conditional statement: 1F A ‘THEN B or, for short, A => B, in which A (the antecedent) and B (the consequent) are fuzzy sets rather than proposi- tional variables. The following are typical examples of such statements: 1 large THEN small 1 slippery THEN dangerous which are abbreviations of the statements 1 xis large THEN y is small 1 the road is slippery THEN driving is dangerous. In essence, statements of this form describe a relation between two fuzzy variables. This suggests that a fuzzy conditional statement be defined as a fuzzy relation in the sense of (2.19) rather than as a connective in the sense of GD. To this end, it is expedient to define first the Cartesian product of two fuzzy sets. Specifically, let A be a fuzzy subset of a universe of discourse U, and let B be a fuzzy subset of a possibly different universe of discourse V. Then, the Cartesian product of A and B is denoted by A x B and is defined by axBal ud» ufolus) 62) where U x ¥ denotes the Cartesian product of the non fuzzy sets U and V; that is, Ux Vb {us)|weU, ve V). Note that when A and B are nonfuzzy, (5.2) reduces to the conventional definition of the Cartesian product of non- fuzzy sets. In words, (5.2) means that A x Bis a fuzzy set of ordered pairs (2), ue U, ve V, with the grade of membership of (ue) in A x B given by jg(w) » jy(0). In this sense, A x Bis a fuzzy relation from U to V. ‘Asa very simple example, suppose that u=1+2 63) Val+243 4) A= Il + 082 5) B= 06/1 + 09/2 + 1/3. 66) Then A x B= O6/(1,1) + 0.9/(1,2) + 1/(1,3) + 0.6/2.1) + 0.8/2,2) + 08/23). (5.1) The relation defined by (5.7) may be conveniently repre- sented by the relation matrix 1 2G 1706 09 1 alos os os) co)_ZADER: ANALYSIS OF COMPLEX SYSTEMS AND DETECTION PROCESSES The significance of a fuzzy conditional statement of the form A THEN B is made clearer by regarding it asa special case of the conditional expression 1F A THEN B rust. C, where A and (B and C) are fuzzy subsets of possibly different universes U and ¥, respectively. In terms of the Cartesian product, the latter statement is defined as follows: 69) in which + stands for the union of the fuzzy relations A x Band (7A x C). More generally if A,,"**,4, are fuzzy subsets of U, and By" B, are fuzzy subsets of Y, then® Ww A THEN Bese C & A x B+ (74 x C) WF A, THEN B, ELSE IF A, THEN By --- ELSE IF A, THEN By 6.10) Note that (5.10) reduces to (5.9) if wr A THEN B rust C is interpreted as 1F A THEN B ELSE IF 1A THEN C. It should also be noted that by repeated application of (5.9) we obtain DAs x By Ay x By too + Ay X Bye WF A THEN (If BTHEN C sts D) wise E = Ax BX CHAXABKD+ TAKE (ull) If we regard 1F A THEN B as A THEN B rise C with unspecified C, then, depending on the assumption made about C, various interpretations of ir A THEN B will result. In particular, if we assume that C = V, then 1 A THEN B (or A = B) becomes? A> BAwATHNBAAx B+ (0A x V), 6.12) If, in addition, we set A = U in (5.12), we obtain as an alternative definition AeBRUXB+CAXY. — (513) In the sequel, we shall assume that C = V, and hence that A= Bis defined by (5.12). In effect, the assumption that C = V implies that, in the absence of an indication to the contrary, the consequent of 14 => C can be any fuzzy subset of the universe of discourse. As a very simple illustration of (5.12), suppose that 4 and B are defined by (5.5) and (5.6). Then, on substituting (5.8) in (5.12), the relation matrix for A -> Bis found to be 06 09 1 aoa [os Os os It should be observed that when 4, B, and C are non- fuzzy sets, we have the identity We A THEN B Etse C= (tf A THEN B) 7 (IF 1A THEN C) 6.14) It should be noted that, in the sense used in ALGOL, the right- hand side of (10) would beexpressed us y= By + (14) 4) Brrr Cdn ees dyes OA) % By when the Ay and Bi, To '., ae nontuzay sets 7 This debntion should be viewed as tentative in nature. 37 which holds only approximately for fuzzy A, B, and C. ‘This indicates that, in relation to (5.15), the definitions of WF A THEN B ELSE C and iF A THEN B, as expressed by (5.9) and (5.12), are not exactly consistent for fuzzy 4, B, and C. It should also be noted that if 1) U = , 2) x = y, and 3) A= Bholds for all points in U, then, by (5.12), A> B implies and is implied by A < B (5.15) ‘exactly if A and B are nonfuzzy and approximately other- wise. ‘As will be seen in Section VI, fuzzy conditional state- ments play a basic role in fuzzy algorithms. More specif- ically atypical problem which is encountered in the course of execution of such algorithms is the following. We have a fuzzy relation, say, R, from U to V which is defined by a fuzzy conditional statement. Then, we are given a fuzzy subset of U, say, x, and have to determine the fuzzy subset of V, say, », which is induced in V by x. For example, we ‘ay have the following two statements 1) xis very small 2) we x is small THEN yi large ELSE y isnot very large ‘of which the second defines by (5.9) a fuzzy relation R. ‘The question, then, is as follows: What will be the value of 1 if x is very small? The answer to this question is provided by the following rule of inference, which may be regarded as an extension of the familiar rule of modus ponens. Compositional Rule of Inference: If R is a fuzzy relation from U to V, and x is a fuzzy subset of U, then the fuzzy subset y of V which is induced by x is given by the com- position (see (2.22)) of R and x; that is, yoxoR 6.16) in which x plays the role of a unary relation.* As a simple illustration of (5.16), suppose that R and x fare defined by the relation matrices in (5.17). Then y is ‘given by the max-min product of x and R: x R y 08 09 02 [o2 1 o-([s 1 “4]- 06 1 04}. (5:17) 05 08 1 As for the question raised before, suppose that, as in (43), we have Uslt2434445 (5.18) with small and large defined by (4.2) and (4.7), respectively. ‘Then, substituting small for 4, large for Band not very large for C in (5.9), we obtain the relation matrix R for the fuzzy conditional statement IF small THEN large ELSE not very large. The result of the composition of R with x = very "If Ris visualized asa fuzzy graph, then (S.16) may be viewed as the expression for the frzy ornate coresponding Yo fuzrysmall is R 06 06 04 04 04 06 08 08 1 08 08 06 0.6 06 06 04 04 0.64 0.36 0.2 [I 0.64 0.36 0.16 0.04)» ’ = [036 0.4 06 08 1]. (5.19) ‘There are several aspects of (6.16) that are in need of comment, First, it should be noted that when R= A=> B and x = we obtain yaAc(A=B)=B 620) as an exact identity, when 4, B, and C are nonfuzzy, and an approximate one, when 4, B, and C are fuzzy. Ii in this sense that the compositional inference rule (5.16) may be viewed as an approximate extension of modus ponens. (Note that in consequence of the way in which A => B is defined in (5.12), the more different x is from A, the less sharply defined is y.) Second, (5.16) is analogous to the expression for the marginal probability in terms of the conditional probability function; that is = Zapy qe = P(X = x) ry = Priv =») Py = Pr {Y= y/1X =x) (21) where and X and Y are random variables with values x,,x3.°** and y;,y2,°**, respectively. However, this analogy does not imply that (5.16) is a relation between probabilities. Third, it should be noted that because of the use of the ‘max-min matrix product in (5.16), the relation between x and y is not continuous. Thus, in general, a small change in x would produce no change in y until a certain threshold is exceeded. This would not be the case if the composition of x with R were defined as max-product composition, Fourth, in the computation of x R one may take ad- vantage of the distributivity of composition over the union of fuzzy sets. Thus, if x=uoro (5.22) where w and v are labels of fuzzy sets, then (uorv)oR = ue RorvoR (523) For example, if x is small or medium, and R= A> B reads 1F x is not small and not large THEN y is very small, then we can write (small or mediwn) » (not small and not large => very small) = small» (not small and not large => very small) or medium ©(not small and not large = very small). (5.24) [BE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, JANUARY 1973 ‘As a final comment, it is important to realize that in practical applications of fuzzy conditional statements to the description of complex or ill-defined relations, the com- putations involved in (5.9), (5.10), and (5.16) would, in general, be performed in a highly approximate fashion Furthermore, an additional source of imprecision would be the result of representing a fuzzy set asa value of a linguistic variable. For example, suppose that a relation between fuzzy variables x and y is described by the fuzzy condi- tional statement 1F small THEN large ELSE IF medium THEN ‘medium ELSE 1 large THEN very small Typically, we would assign different linguistic values to x and compute the corresponding values of y by the use of (6.16). Then, on approximating to the computed values of by linguistic labels, we would arrive at a table having the form shown below: Given Inferred x y “mall linge "ot small ‘ot very ‘medium — medium Ser mal ay nay iipe Troe” ery mall Ser cory small ery ver a rot cery large _Smaitor Such a table constitutes an approximate linguistic charac- terization of the relation between x and y which is inferred from the given fuzzy conditional statement. As was stated earlier, fuzzy conditional statements play a basic role in the description and execution of fuzzy algorithms. We turn to this subject in the following section. VI. Fuzzy Avcorrrims Roughly speaking, a fuzzy algorithm is an ordered set of fuzzy instructions which upon execution yield an ap- proximate solution to a specified problem. In one form or another, fuzzy algorithms pervade much of what we do. ‘Thus, we employ fuzzy algorithms both consciously and subconsciously when we walk, drive a car, search for an object, tie a knot, park a car, cook a meal, find a number in a telephone directory, etc. Furthermore, there are many instances of uses of what, in effect, are fuzzy algorithms in 1 wide variety of fields, especially in programming, opera- tions research, psychology, management science, and ‘medical diagnosis. The notion of a fuzzy set and, in particular, the concept of a fuzzy conditional statement provide a basis for using fuzzy algorithms in a more systematic and hence more ‘effective ways than was possible in the past. Thus, fuzzy algorithms could become an important tool for an ap- proximate analysis of systems and decision processes which ‘are much too complex for the application of conventional mathematical techniques. ‘A formal characterization of the concept of a fuzzy algorithm can be given in terms of the notion of a fuzzy Turing machine or a fuzzy Markoff algorithm [6]-[8]. In this section, the main aim of our discussion is to relate the concept of a fuzzy algorithm to the notions introduced in ‘the preceding sections and illustrate by simple examples some of the uses of such algorithms.‘The instructions in a fuzzy algorithm fall into the follow- ing three classes. 1) Assignment Statements: 2, xa5 x = small xis large x is not large and not very small 2) Fuzzy Conditional Statements: e.2., wx is small THEN y is large ELSE y is not large 1 x is positive THEN decrease y slightly w x is much greater than 5 THEN stop wx is very small THEN go to 7. Note that in such statements either the antecedent or the consequent or both may be labels of fuzzy sets. 3) Unconditional Action Statements: e., multiply x by x decrease x slightly delete the first few occurrences of 1 0107 print x stop. Note that some of these instructions are fuzzy and some are not. ‘The combination of an assignment statement and a fuzzy ‘conditional statement is executed in accordance with the ‘compositional rule (5.16). For example, if at some point in the execution of a fuzzy algorithm we encounter the instructions 1) x = very small 2) 1 x is small THEN y is large ELSE y is not very large where small and large are defined by (4.2) and (4.7), then the result of the execution of 1) and 2) will be the value of y given by (5.19), that is, y= 36/1 + 0.4/2 + 0.64/3 + 0.8/4 + 1/5. 6.1) An unconditional but fuzzy action statement is executed. similarly. For example, the instruction multiply x by itself @ few times 62) with few defined as few = + 08/2 + 0.63 + 0.4/4 63) would yield upon execution the fuzzy set y= 18 + O8/x? + 06x + 04ix, 64) It is important to observe that, in both (6.1) and (6.4), the result of execution is a fuzzy set rather than a single number. However, when a human subject is presented with ‘a fuzzy instruction such as “take several steps,” with several defined by (See (2.10) several = 0.5/3 + 0.8/4 + 1/5 + 1/6 + 08/7 + 0.5/8 (6.5) the result of execution must be a single number between 3 and 8, On what basis will such a number be chosen? » {As pointed out in [6] it is reasonable to assume that the result of execution will be that element of the fuzzy set which has the highest grade of membership in it. If such an clement is not unique, as is true of (6.5), then a random or arbitrary choice can be made among the clements having the highest grade of membership. Alternatively, an external criterion can be introduced which linearly orders those clements of the fuzzy set which have the highest member- ship, and thus generates a unique greatest element. For example, in the case of (6.5), if the extemal criterion is to ‘minimize the number of steps that have to be taken, then the subject will pick $ from the elements with the highest ‘grade of membership. ‘An analogous question arises in situations in which a fhuman subject has to give a “yes” or “no” answer to a fuzzy question. For example, suppose that a subject is presented with the instruction 1 xis small THEN stop ELSE go 07 (6.6) in which small is defined by (4.2). Now assume that x = 3, ‘which has the grade of membership of 0.6 in small. Should the subject execute “stop” or “go to 7°? We shall assume that in situations of this kind the subject will pick that alternative which is more true than untrue, ¢g,, “x is small” over “xis not small,” since in our example the degree of truth of the statement “3 is small” is 0.6, which is greater than that of the statement “3 is not small.” If both alterna- tives have more or less equal truth values, the choice can be made arbitrarily. For convenience, we shall refer to this rule of deciding between two alternatives as the rule of the preponderant alternative. tis very important to understand that the questions just discussed arise only in those situations in which the result of execution of a fuzzy instruction is required to be a single clement (e.g., a number) rather than a fuzzy set. Thus, if we allowed the result of execution of (6.6) to be fuzzy, then for x = 3 we would obtain the fuzzy set 0.6/stop + 0.4/g0 to 7 ‘which implies that the execution is carried out in parallel ‘The assumption of parallelism is implicit in the composi- tional rule of inference and is basic to the understanding of fuzzy algorithms and their execution by humans and ‘machines. Tn what follows, we shall present several examples of fuzzy algorithms in the light of the concepts discussed in the preceding sections. It should be stressed that these ‘examples are intended primarily to illustrate the basic aspects of fuzzy algorithms rather than demonstrate their effectiveness in the solution of practical problems. It is convenient to classify fuzzy algorithms into several basic categories, each corresponding to @ particular type of application: definitional and identificational algorithms; generational algorithms; relational and behavioral al- gorithms; and decisional algorithms. (It should be noted that an algorithm ofa particular type can include algorithms of other types as subalgorithms. For example, a definitional algorithm may contain relational and decisional sub-ry algorithms.) We begin with an example of a definitional algorithm, Fuzzy Definitional Algorithms One of the basic areas of application for fuzzy algorithms lies in the definition of complex, ill-defined or fuzzy con- cepts in terms of simpler or less fuzzy concepts. The follow- ing are examples of such fuzzy concepts: sparseness of ‘matrices; handwritten characters; measures of complexity; ‘measures of proximity or resemblance; degrees of clustering: criteria of performance; soft constraints; rules of various kinds, e.g., zoning regulations; legal criteria, e., criteria for insanity, obscenity, ete.; and fuzzy diseases such as arthritis, arteriosclerosis, schizophrenia. Since a fuzzy concept may be viewed as a label for a fuzzy set, a fuzzy definitional algorithm is, in effect, a finite set of possibly fuzzy instructions which define a fuzzy set in terms of other fuzzy sets (and possibly itself i.., recursively) or constitute a procedure for computing the grade of membership of any element of the universe of discourse in the set under definition. In the latter case, the definational algorithm plays the role of an identficarional algorithm, that is, an algorithm which identifies whether or not an element belongs to a set or, more generally, determines its grade of membership. An example of such an algorithm is provided by the procedure (see [5]) for computing the grade of membership of a string in a fuzzy language generated by a context-free grammar. ‘Asa very simple example ofa fuzzy definitional algorithm, we shall consider the fuzzy concept oval. It should be em- phasized again that the oversimplified definition that wil be given is intended only for illustrative purposes and has, ro pretense at being an accurate definition of the concept ‘eal. The instructions comprising the algorithm OVAL are listed here. The symbol T in these instructions stands for the object under test, The term Catt. CONVEX represents a call on a subalgorithm labeled CONVEX, which is a definitional algorithm for testing whether or not Tis convex. An instruction of the form 1F A THEN B should be interpreted as tf A THEN B r15e go to next instruction. Algorithm OVAL. 1) iF Tis not closed THEN Tis not oval; stop. 2) WF Tis self-intersecting THEN T is not oval; stop. 3) af Tis not cat. CONVEX THEN Tis not oval; stop. 4) T does not have two more or less orthogonal axes of symmetry THEN Tis not oval; stop. 5) IF the major axis of T is not much longer than the minor axis THEN T is not oval; stop, 6 Tis oval; stop. Subalgorithm CONVEX: Basically, this subalgorithm in- volves a check on whether the curvature of T at each point ‘maintains the same sign as one moves along Tin some initially chosen direction. 1) x = a (some initial point oh 7). 2) Choose a direction of movement along T. 3) 1% direction of tangent to Tat x. [BE TRANSACTIONS OW SYSTEMS, MAN, AND CYBERNETICS, JANUARY 1973 4) x © x + 1 (move from x to a neighboring point). 5) #" & direction of tangent to T'at x’. 6) a ~ angle between 1’ and ¢. Newry. 8) 1 © direction of tangent to Tat x. ext 10) ¢* © direction of tangent to T at x’. 11) B © angle between 1” and 12) Wf docs not have the same sign as « THEN Tis not 13) x’ = a THEN Tis convex; return. 14) Go to 7). ‘Comment: It should be noted that the frst three instruc- tions in OVAL are nonfuzzy. As for instructions 4) and 5), they involve definitions of concepts such as “more or less orthogonal,” and “muuch longer,” which, though fuzzy, are less complex and better understood than the concept of ‘oval. This exemplifies the main function of a fuzzy defini- tional algorithm, namely, to reduce a new or complex fuzzy concept to simpler or better understood fuzzy con- cepts. In a more elaborate version of the algorithm OVAL, the answers to 4) and 5) could be the degrees to which the conditions in these instructions are satisfied. The final result of the algorithm, then, would be the grade of membership of T in the fuzzy set of oval objects In this connection, it should be noted that, in virtue of 6.15), the algorithm OVAL as stated is approximately ‘equivalent to the expression oval = closed 7 non-self-intersecting > convex 19 more or less orthogonal axes of symmetry 19 major axis much larger than minor axis (6.7) which defines the fuzzy set oval as the intersection of the fuzzy and nonfuzzy sets whose labels appear on the right hand side of (6,7). However, one significant difference is that the algorithm not only defines the right-hand side of (6.1), but also specifies the order in which the computations implicit in (6.7) are to be performed Fuzzy Generational Algorithms As its designation implies, @ fuzzy generational algorithm serves to generate rather than define a fuzzy set. Possible applications of generational algorithms include: generation of handwritten characters and patterns of various kinds; cooking recipes; generation of music; generation of sen- tences in a natural language; generation of speech. As a simple illustration of the notion of a generational algorithm, we shall consider an algorithm for generating the letter P, with the height h and the base 6 of P constituting the parameters of the algorithm. For simplicity, P will be generated as a dotted pattern, with eight dots lying on the vertical line. Algorithm P(hb). + (first dot at base.——._£ ‘Add 2 cups gronulted suger to seucepon o ‘Add | cup bottled milk to toucepen Sequence ‘Add ¥ teaspoon aclt te sevcepon of ‘Add 2 ot unsweetened chocelete 'e saucepan stetomonts | | Add 2 tablespoons white arn syrup to educepon Put seucepen over low heat ® Stir the ea Loop for disnleing we (Cts suger distoleed?)— No Ye 1_£ 1 ® Cook gent Removel from heat Loop for sty ‘ooking @ [Drop in 2 voap butter ® 4a si ie oemeere that is) ¥ ‘Brot with spoon Bec with — a lead (_ Becsin held ts shepe?) |beoting Hos mixture lost its gloss? No: ; L | Coot, cut ime squares | Yer Fig. 3. Revie for chocolate fudge (Irom 19D.2 3) X@ + 1) © XG + M6 (put dot approximately h/6 units of distance above ¥(i)). iste. 5) Fi = 7 THEN make right tum and go to 7). 6) Go to 3. 7) Move by h/6 units; put a dot. 8) Tum by 45°; move by /h/6 units; put a dot. 9) Turn by 45°; move by /i/6 units: put a dot. 10) Turn by 45°; move by f/6 units; put a dot. 11) Turn by 45°; move by f/6 units; put a dot; stop. The algorithm as stated is of open-loop type in the sense that it does not incorporate any feedback. To make the algorithm less sensitive to errors in execution, we could introduce fuzzy feedback by conditioning the termination of the algorithm on an approximate satisfaction of a specified test. For example, if the last point in step 11) does not fall on the vertical part of P, we could return to step 8) and either reduce or increase the angle of turn in steps 8)-11) to correct for the terminal error. The flowchart of a cooking recipe for chocolate fudge (Fig. 3), which is reproduced from [19], is a good example of what, in effect, is a fuzzy generational algorithm with feedback. Fuzzy Relational and Behavioral Algorithms AA fuzzy relational algorithm serves to describe a relation or relations between fuzzy variables. A relational algorithm which is used for the specific purpose of approximate description of the behavior of a system will be referred to as a fuzzy behavioral algorithm. ‘A simple example of a relational algorithm labeled R which involves three parameters x, y, and 2 is given. This algorithm defines a fuzzy ternary relation R in the universe of discourse U = 1 +243 +4 + 5 with small and large defined by (4.2) and (4.7). Algorithm R(x,9,2): 1) wx is small and y is large THEN 2 is very small ese is not small, 2) 1 x is large THEN (WF y is small THEN 2 is very large ELSE z is small) ELSE z and y are very very small. If needed, the meaning of these conditional statements can be computed by using (5:9) and (5.11). The relation R, then, will be the intersection of the relations defined by instructions 1) and 2) Another simple example of a relational fuzzy algorithm F(x,y) which illustrates a different aspect of such algorithms is the following Algorithm FCxy): 1) a x is small and x is increased slightly THEN y will increase slighty, 2) WF x is small and x is increased substantially THEN y will increase substantially, 3) 1 2 is large and x is increased slightly THEN y will increase moderately. 4) 1 x is large and x is increased substantially THEN y will increase very substantially. [MEE TRANSACTIONS ON SYSTEMS, MAN, AND CYRERNETICS,SANUARY 1973 As in the case of the previous example, the meaning of the fuzzy conditional statements in this algorithm can be computed by the use of the methods discussed in Sections IV and V if one is given the definitions of the primary terms large and small as well as the hedges slightly, sub- stantially, and moderately As a simple example of a behavioral algorithm, suppose that we have a system S with two nonfuzzy states (ce [3]) labeled q, and qz, two fuzzy input values labeled fow and ‘igh, and two fuzzy output values labeled large and small ‘The universe of discourse for the input and output values, is assumed to be the real line. We assume further that the behavior of Scan be characterized in an approximate fashion by the algorithm that will be given. However, to represent the relations between the inputs, states, and out- puts, we use the conventional state transition tables instead of conditional statements. Algorithm BEHAVIOR. ae % ™ ea a large small ‘malt large where 4, input at time ¢ Je output at time ¢ 1, state at time ¢ (On the surface, ths table appears to define a conventional nonfuzzy finite-state system. What is important to recognize, hhowever, is that in the case of the system under considera- tion the inputs and outputs are fuzzy subsets of the rel line. ‘Thus we could pose the question: What would be the output of S if it isin state g, and the applied input is very low? In the case of 5, this question can be answered by an applica- tion of the compositional inference rule (5.16). On the other hhand, the same question would not be a meaningful one if is assumed to be a nonfuzzy finite-state system charae- terized by the preceding table. Behavioral fuzzy algorithms can also be used to describe the more complex forms of behavior resulting from the presence of random elements in a system. For example, the presence of random elements in S might result in the follow ing fuzzy-probabilistc characterization of its behavior w\ a a a low @alikely—qylikely large aml. oh — qu lkely®— quale? all, Tae “ “ ies bey In this table, the term likely and its modifications by very and not serve to provide an approximate characteriza- tion of probabilities. For example, F the input is Jow and the present state is q,, THEN the next state is likely to be q2. Similarly, the input is high and the present state is q3 THEN the output is very unlikely to be large. If the meaningof likely is defined by (see (4.16) likely = 1/1 + 1]0.9 + 1/08 + 0.8/0.7 + 0.6/0.6 + 0.50.5 + 0.3/0.4 + 0.2/0.3 (6.8) then unlikely = 0.2/0.7 + 0.4/0.6 + 0.5)0.5 + 0.7/0.4 +0.8/03 + 1/0.2 + 1/01 + 1/0 69) very likely = 1/1 + 1/0.9 + 1/08 + 0.6/0.7 + 0.4/0.6 + 0.3/0.5 + 0.1/0.4 (6.10) very unlikely ~ 0.2/0.6 + 0.3/0.5 + 0.5/0.4 + 0.6/0.3 + 10.2 + 1/01 + 1/0. 6.1, Fuzzy Decisional Algorithms A fuzzy decisional algorithm is a fuzzy algorithm which serves to provide an approximate description of a strategy or decision rule. Commonplace examples of such al- ‘gorithms, which we use for the most part on a subconscious level, are the algorithms for parking a car, crossing an intersection, transferring an object, buying a house, ete. To illustrate the notion of @ fuzzy decisional algorithm, wwe shall consider two simple examples drawn from our everyday experiences. ‘Example—Crossing a traffic intersection: It is convenient to break down the algorithm in question into several sub- algorithms, each of which applies to a particular type of intersection. For our purposes, it will be sufficient to describe only one of these subalgorithms, namely, the subalgorithm SIGN, which is used when the intersection has a stop sign. {AS in the case of other examples in this section, we shall make a number of simplifying assumptions in order to shorten the description of the algorithm, Algorithm INTERSECTION: 1) IF signal lights THEN CALL SIGNAL E1se 1F stop sign ‘THEN CALL SIGN ELSE IF blinking light THEN CALL BLINKING E1se cA. UNCONTROLLED. ‘Subalgorithm SIGN: 1) a no stop sign on your side THEN IF no cars in the intersection THEN cross at normal speed ELse wait for cars to leave the intersection and then cross. 2) 1 not clase to intersection THEN continue approach- ing at normal speed for a few seconds; go to 2). 3) Slow down. 4) w in a great hurry and no police cars in sight and no cars in the intersection or its vicinity THEN cross the intersection at slow speed '5) 1 very close to intersection THEN stop; go to 7). 6) Continue approaching at very slow speed; go to 5). 17) i no cars approaching ot in the intersection THEN cross. 8) Wait a few seconds; go to 7). It hardly needs saying that a realistic version of this algorithm would be considerably more complex. The im- 8 oa 14a n we ce stort Fig. 4, Problem of transferring blindfolded subject from start to goal. portant point of the example is that such an algorithm ould be constructed along the same lines as the highly Simplified version just described, Furthermore, it shows that a fuzzy algorithm could serve as an effective means of communicating know-how and experience. ’AS a final example, we consider « decisional algorithm for transferring. a blindfolded subject H from an inital position star to a final postion goa! under the assumption that there may be an obstacle lying between start and goal (ce Fig, 4), (Highly sophisticated nonfuzzy algorithms of this type for use by robots are incorporated in Shakey, the robot built by the Artificial Intelligence Group at Stanford Research Institute. A description of this robot is given in 20} The algorithm, labeled OBSTACLE, is assumed to be used by a human controller C who can observe the way in Which Hf executes his instructions. This fuzzy feedback plays an essential role in making it posible for C to direct. to goal in spite ofthe fuzziness of insrutions as wel as the ors in their execution by H. The algorithm OB- STACLE consists of three subalgorthms: ALIGN, HUG, and STRAIGHT. The function of STRAIGHT i to tran fer H from start to an intermediate goal -goal, and then from Lgoal; to goal, (See Fig. 4) The function of ALIGN is to orient H ina desired direction; the function of HUG is to guide H along the boundary of the obstacle until the goal is no longer obstructed. Tnstead of describing these subalgorithms in terms of fuzzy conditional statements as we have done in previous examples, itis instructive to convey the same inform by floweharts, as shown in Figs. $7. In the flowchart of ALIGN, e denotes the eror in alignment, and we assume for simplicity that e has a constant sign, The flowcharts of HUG and STRAIGHT are selRexplanatory. Expresed in terms of fuzzy conditional statements, the flowchart of STRAIGHT, for example, translates into the following instructions ‘Sualgorithm STRAIGHT: 1) 1 not close THEN take a step; go t0 1). 2) w not very close THEN take a small step; go to 2). 3) IF not very very close THEN take a very small step; 20 103). 4) Stop. ‘VIT.Concuupine Remarks In this and the preceding sections of this paper, we have attempted to develop a conceptual framework for dealingen an by 30 ee 6 ce 0" Bs cb ot Coe wer ome oF Fig. 5. Subalgorithm ALIGN, Fig. 7. Subslgorithm STRAIGHT, [NEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, JANUARY 1973 systems which are too complex or too ill-defined to admit of precise quantitative analysis. What we have done should be viewed, of course, as merely a first tentative step in this direction. Clearly, there are many basic as well as detailed aspects of our approach which we have treated incompletely, if at all. Among these are questions relating to the role of fuzzy feedback in: the execution of fuzzy algorithms; the execution of fuzzy algorithms by humans; the conjunction of fuzzy instructions; the assessment of the goodness of fuzzy algorithms; the implications of the compositional rule of inference and the rule of the pre- onderant alternative; and the interplay between fuzziness and probability in the behavior of humanistic systems. Nevertheless, even at its present stage of development, the method described in this paper can be applied rather effectively to the formulation and approximate solution of ‘ wide variety of practical problems, particularly in such fields as economics, management science, psychology, lin- uistics, taxonomy, artificial intelligence, information re- tieval, medicine, and biology. This is particularly true of those problem areas in these fields in which fuzzy algorithms can be drawn upon to provide a means of description of ill-defined concepts, relations, and decision rules. 10), he Zadeh, “Fuzzy ses," Inform. Contr, vol. 8, pp. 38-353, (2) —~; “Similarity relations and fuzzy orderings" Inform. Sct, vol}, ppt a, 197, ‘pioward a theory of fuzy systems.” in Aypects of Network nd Syavem Theory, RE, Kalan and N. DeClaris, Bas. New York? Holt, Rinehart and Winston, 197 Inform. Sei, vol. 3, ee 2 lest nym tren 7p en rane ee nce ee taal ne Hei a ala bles ae “Fuzzy algorithms,” Inform. Contr, vol. 12, pp. 94-102, i968 71 E, Sins, “Fuzzy algorithms,” Inform. Con, vel. 17, pp. 326- (81 L.A, Zadeh, “On fuzzy algoriths,” Electron, Res. Lab, Univ, aioria, Boktey. eae MES. (9) SK Chg, Gn the exeeston of fy proprams using Bite Rate machi,” IEEE’ Pony Comput, val C3, pp 3H 35, S'S, 1 Chang and L, A. Zadeh, “Fuzzy mapping and conto,” TREE Trans. Syst Man, Cybern, ol. SMC, pp. 305, Tah. (0 KE Mena aL A Zadeh, “Decosaking i try ti Se, Mere Sol ap aah ‘2s bes, = OPE eee 113) Bako “Hedge: sty in meaning criteria and the ope of ” cose Pe Bh Rg Wn Tc Hg me 114] LA: Zadeh, “A system-theoretic view of behavior modification.” Elction. Res. Lab, Univ. Calfrnia, Berkeley, Memo. M30, r Us) —— “A fuzzy set-theoretic interpretation of hedges,” Electeon. Res Lab. Un. California, Berkeley, Memo, M85, 197 116] A. Be Lia and S. Termin, "A deitition of non probabilistic: atropy in the stting of fury sets theory.” Iyform ‘Contr, wot 20, pp. 301-312, 1972 17, RC. 1's, “Fuzzy loge end the resolution principle,” J. Ass Comput, Mach, vol 19, pp. 19-115. 197 U8) G.'E"Hughes"and M.S Creseell: tn inuroduction to. Modal Lexie. "London" Methven, 1968. U9 RS Lede, Forman TY Proginming. New York: MeGiaw- 120) B, Raphael, R, Duda, RE. Fikes, P.E. Hart, N, Nsson,P. W. Thomdyke, and B. M, Wilbur, “Reacarch and appletigns ~ artical iniellgence,” Stanford Res: Inst, Menlo Park Cali, Final Rep, Ot 17Reprinted by permission rom IEEB-TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS Nol SMC, No. January 1973p. 8-4 Copyright 1972 by the latitate of Electral and Electronics Engineer, Inc. PRINTED INTHECS.A.