A Fuzzy-Logic-Based Approach Qualitative Modeling: Abstract

lEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. I, NO. I .
FEBRUARY 1993
A Fuzzy-Logic-Based Approach to Qualitative Modeling

Michio Sugeno and Takahiro Yasukawa
Abstract- This paper discusses a general approach to qualitative modeling based on fuzzy logic. The method of qualitative modeling is divided into two parts: fuzzy modeling and linguistic approximation. It proposes to use a fuzzy clustering method (fuzzy c-means method) to identify the structure of a fuzzy model. To clarify the advantages of the proposed method, it also shows some examples of modeling, among them a model of a dynamical process and a model of a human operators control action.
I. INTRODUCTION
N this paper we discuss a method of qualitative modeling based on fuzzy logic. Related terminologies for qualitative modeling in fuzzy theory are known as fuzzy modeling and linguistic modeling. Qualitative modeling is not so popular in general but this concept is indirectly implied by fuzzy modeling or linguistic modeling. Here we deal with a qualitative model as a system model based on linguistic descriptions just as we use them in sociology or psychology. Though the terminology of linguistic modeling may be straightforward and more appropriate, we use qualitative modeling partly because we like to discuss certain problems in comparison with the qualitative reasoning approach in artificial intelligence and also because qualitative modeling has been one of the most important issues since the very beginning of fuzzy theory. Here artificial intelligence will be abbreviated to AI as usual and the terminology fuzzy modeling will be used in a narrow sense. Before entering the main subject we give an overview of fuzzy modeling in fuzzy theory and qualitative reasoning in AI; then we state the problems concerning qualitative modeling based on fuzzy logic from a systems theory point of view. After these discussions, we propose a method of qualitative modeling in a general framework known as the black box approach in systems theory. That is, we build a qualitative model of a system without LI priori knowledge about a system provided that we are given numerical input-output data.
A. F i i p Modeling
Though the term fuzzy modeling has not been used so often, fuzzy modeling is the most important issue in fuzzy logic or more widely in fuzzy theory. In fact we can find
Manuscript received August 2, 1991; revised April 12, 1992. M. Sugeno is with the Department of Systems Science, Tokyo lnatitute of Technology, 4259 Nagatauta. Midoriku, Yokohama 227, Japan. T. Yasukawa is with the Department of Control Engineering, Tokyo lnstitutc of Technology, 4259 Nagatauta, Midoriku, Yokohama 227, Japan. IEEE Log Number 9205336.
the seminal ideas of fuzzy modeling in the early papers of Zadeh. Thus the research concerned with fuzzy niodeling has a history of more than 20 years. There are many interpretations of fuzzy modeling. For instance, we can consider a fuzzy set as a fuzzy model of a human concept. In this study, we simply understand the fuzzy modeling to be an approach to form a system model using a description language based on fuzzy logic with fuzzy predicates. In a broader sense we can interpret the fuzzy modeling as a qualitative modeling scheme by which we qualitatively describe system behavior using a natural language. The fuzzy modeling in a narrow sense is a system description with fuzzy quantities. Fuzzy quantities are expressed in terms of fuzzy numbers or fuzzy sets associated with linguistic labels where a fuzzy set usually does not have a tight relation with a linguistic label; either a fuzzy number does not need to have a linguistic interpretation or a linguistic label is expressed as if it were a frill of a membership function. On the other hand, what we mean by a qualitative model is a generalized fuzzy model consisting of linguistic explanations about system behavior. Linguistic terms in linguistic explanations are found such that they linguistically approximate the fuzzy sets in an underlying fuzzy model. Furthermore, when a relation of state variables in a system is to be expressed linguistically in a qualitative model, we can use a method of linguistic approximation to accomplish this aim. In this paper, we focus our attention on qualitative modeling since we want to stress that one of the final targets of fuzzy modeling is qualitative modeling. We could even say that fuzzy sets and/or fuzzy logic were suggested with qualitative modeling in mind. Let us go back to Zadehs early ideas. In his paper in 1968 [ I ] following the first paper, Fuzzy Sets. in 1965, he suggested using an idea of fuzzy algorithm such as a) set y approximately equal to I O if .I is approximately equal to 5 ; b) if .I is large. increase l/J by several units. Those fuzzy algorithms are nothing but qualitative descriptions of a human action, or decision making. As examples, he shows cooking recipes, directions for repairing a TV set, instructions on how to treat a disease, and instructions for parking a car. As for the necessity of fuzzy algorithms, he notes that most realistic problems tend to be complex, and many complex problems are either algorithmically unsolvable or, if solvable in principle, are computationally infeasible. The most remarkable paper related to qualitative modeling is his paper of 1973 [21 on linguistic analysis, where he states the principle of incompatibility, according to which as the
1063-6706/93$03 00 G 1993 IEEE
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. I , NO. I , FEBRUARY 1993
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 groups. Based on the above considerations Zadeh suggested linguistic analysis in place of quantitative analysis. If we look at his idea of linguistic analysis from the viewpoint of modeling, it is seen to be precisely qualitative modeling. As the main characteristics of this approach, he suggests 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. It is well-known that, motivated by these ideas of fuzzy algorithm and linguistic analysis, Mamdani first applied fuzzy logic to control [3]. This topic has come to be known as fuzzy algorithmic control or linguistic control. Fuzzy control can be viewed in a certain sense as the result of the qualitative modeling of a human operator working at plant control. For example, fuzzy control rules may be as follows: 1) if the error is positive big and the change of error is positive medium, then the change of control is positive big; 2 ) if the error is positive small and the change of error is positive small, then the change of control is zero, where the error is the difference between the reference input and the plant output, the change of error is the derivative of the error, and the change of control is the derivative of the control input. As seen in the above example, we use, to describe control rules, linguistic variables which take linguistic values such as positive big, positive medium, zero, and negative small. These are not merely symbols but rather the linguistic labels with quantitative semantics given by underlying fuzzy sets which are associated with membership functions. So we may call a label positive big a fuzzy quantity. In fact we use sometimes the term fuzzy number or fuzzy interval. The main problem of fuzzy control is to design a fuzzy controller where we usually take an expert-system-like approach. That is, we derive fuzzy control rules from a human operators experience and/or engineers knowledge, which are mostly based on their qualitative knowledge of an objective system. The design procedure is thus something like the following: first, we build linguistic control rules; second, we adjust the parameters of fuzzy sets by which the linguistic terms in the control rules are quantitatively interpreted. For example, let us look at the linguistic rules for controlling a cement kiln [4], as shown in Fig. 1. From these, we can easily derive fuzzy control rules. In this sense, we may say that a set of fuzzy control rules is a linguistic model of human control actions which is not based on a mathematical description of
Case
1
Condition BZ low Ox low

BE low
Action to be taken
When BZ is drasticall low: (a) reduce klln s p e J (b) reduce fuel When Br is sli htly low: (c) increase I.%. speed (d) decrease fuel rate
BZ IOW Ox low BE O.K. BZ low Ox low BE high
(a) reduce klln speed (b) reduce fuel rate (c) reduce 1. D. fan speed a reduce kiln speed [b] reduce fuel rate (c) reduce 1. D. fan speed
BE = back end temperature, BZ = burning zone temperature, OX= percentage of oxygen gas In kiln exit gas. Total of 27 rules.
Fig. 1. Operators manual of cement kiln control.
human control actions but is directly based on a human way of thinking about plant operation. Apart from fuzzy control, we have many studies on fuzzy modeling. Those are divided in two groups. The studies of the first group deal with fuzzy modeling of a system itself or a fuzzy modeling for simulation [5]-[8]. Some of those are considered as examples of qualitative modeling. The studies of the second group deal with fuzzy modeling of a plant for control [9]-[13]. Just as with the modem control theory, we can design a fuzzy controller based on a fuzzy model of a plant if a fuzzy model can be identified. Fuzzy modeling in the latter sense is not necessarily viewed as qualitative modeling unless the derivation of a qualitative model from the identified fuzzy model is discussed. It is, of course, quite interesting to derive qualitative control rules based on the qualitative model of a plant. For example [14], assume that we have a dynamical plant model such as 1) if u , ~ positive and yn-l is positive, then y is positive is , big; 2 ) if U,, is negative and yn-l is positive, then yn is positive small; 3) if U , is positive and yn-l is negative, then yn is negative small; 4) if U , is negative and yn-l is negative, then yn is negative big. Then given a reference input 7 = 0, we can derive the following control rules based on the above model: 1 ) if y,,-1 is positive and y is positive big, then U , , is , negative big; 2 ) if yla-1 is positive and y is negative, then U,, is positive; , 3) if yTa-l is negative and yn is positive, then uTL is negative; 4) if y,,-1 is negative and yT2is negative big, then U,, is positive big. Here the error e,, is equal to -yn since the reference input (the set point) is z x o . In the fuzzy modeling, the most important problem is the identification method of a system. The identification for the
SUGENO AND YASUKAWA: FUZZY-LOGIC-BASED APPROACH TO QUALITATIVE MODELING
fuzzy modeling has two aspects as usual: structure identification and parameter identification. This problem will be discussed in subsection D.
E. Qualitative Reasoning
Outside the research area of fuzzy logic, the concept of qualitative reasoning has been suggested in AI. Earlier studies of qualitative reasoning were found on mechanical systems in 1975 [15], and also on naive physics in 1979 [16]. We can identify as typical studies of qualitative reasoning the following: 1) qualitative physics or naive physics [ 171, 2 ) qualitative process theory [ 181, 3) qualitative simulation [ 191. Here we call these ideas qualitative reasoning. The aims and characteristics of qualitative reasoning may be summarized as follows. According to de Kleer [17], the behavior of a physical system can be discussed by the exact values of its variables (forces, velocities, positions, pressures, etc.) at each time instant. Such a description, although complete, fails to provide much insight into how the system functions. Our longtemi goal is to develop an alternate physics in which these same concepts are derived from a far simpler, but never the less formal qualitative basis . . . our proposal is to reduce the quantitative precision of the behavioral descriptions but retain the crucial distinctions. Also if we refer to Kuipers [ 191, we find an expert system is often a shallow model of its application domain, in the sense that conclusions are drawn directly from observable features of the presented situation. One major line of research toward the representation of deep models is the study of qualitative causal models. Research on qualitative causal models differs from more general work on deep models in focusing on qualitative descriptions of the deep mechanism, capable of representing incomplete knowledge of the structure and behavior of the mechanism. As we find in the above statements, there are small distinctions and big similarities between fuzzy modeling and qualitative reasoning. One distinction is that fuzzy modeling starts from the fact that a precise mathematical model of a complex system cannot be obtained, whereas qualitative reasoning starts from the fact that, although a complete model may be available, it cannot provide insight into the system; a description based on deep knowledge is needed. On the other hand, similarities are found in the confidence of the advantage of qualitative expressions, in the goals, and in some parts of description languages for modeling. The common idea of qualitative system theory should be emphasized. Qualitative reasoning makes use of a quantity space on which landmarks are defined. Usually one landmark 0 is set, and then three values {+. 0 . - } are used which are qualitative in nature, similar to fuzzy values {positive, zero, negative} in fuzzy modeling. The difference between them is that qualitative values are crisp, acting as symbols, though they call {+. 0 . - } semantics, while fuzzy values are not symbols but quantitative semantics represented by their membership functions. In qualitative reasoning such phenomena as
1) temperature c Q+ pressure ( 2 ) pressure x Q- volume are interpreted as 1) if the temperature increases (decreases), then pressure increases (decreases); 2) if pressure increases (decreases), then volume decreases (increases). As we have seen, this sort of description was found in Zadehs paper in 1968, which shows our motivation to use fuzzy concepts. As far as reasoning is concemed, in qualitative reasoning the following arithmetic is defined:
[:J; X
y] =
[:I;] X
where [ x ] means a qualitative value of a physical quantity .r. n is a discrete time, and i3.r is the qualitative derivative of .I with the values of {+. 0. -}. Equations ( I ) and ( 2 ) describe binary operations on qualitative values; (3) describes the dynamical relation of a system. We must point out that, in fuzzy theory, the arithmetic of fuzzy numbers is in a much more general setting. For example we have
where n, is a fuzzy number, approximately U, associated with a membership function, e.g.,
The arithmetic in (4) and ( 5 ) is a generalization of the ordinary arithmetic, which of course precisely includes ( 1 ) and ( 2 ) . In fuzzy modeling, the fuzzy arithmetic is used for computation and if-then rules are used for inference together with fuzzy reasoning based on fuzzy logic. Fuzzy reasoning is interpreted as approximate reasoning based on vague or incomplete knowledge. As far as the tools for model description are concerned, those in qualitative reasoning seem to be very simple compared with those in fuzzy modeling. We can recall a dentists way of -. description: a dentist uses the symbols {++. i. - -}, and a control engineer uses in the bang-bang control the symbols {+l.-l}, where the symbol zero can be added as an ideal case in the middle of + I and - I from a theoretical point of view. The dentists symbols are fuzzy in nature, much more similar- to fuzzy values, while the control symbols are crisp and look similar to qualitative reasoning. Now let us look at the basic approach of qualitative modeling. As far as the authors can determine, this approach suggests I ) to make a model by observations of phenomena based on,
+.
in
for instance, empirical and/or physical knowledge or 2) to derive a qualitative model based on a differential equation of a system. This approach is effective if a system is simple, for example a mechanical system with a small number of state variables and simple interactions among state variables. But, of course, even if a system is composed of many state variables, this approach is useful provided that the system is a lumpedparameter system such as an electric circuit. As a conclusion it seems that qualitative modeling is concerned with wellstructured systems; in most studies on qualitative modeling, they deal with toy systems, if we look at those from a systems theory point of view. In modeling we do not take this approach to First of all, if a mathematical model is available, we can use it for the purpose of engineering (analysis, control and prediction). This fact of course does not prevent us for making a qualitative model of a system, for instance, in order to explain its general structure and behavior to students. Second, even if we can find the local mechanisms of a system, it is often the case that we cannot build the whole system mode] by aggregating the local mechanisms. This is why we take a so-called black box approach to Systems identification in systems engineering. We have to note that what we mean by a complex system is a system for which we cannot have a relevant mathematical model; in other words, we cannot build a global system model by integrating local models. There is another reason for a black box approach. It is often the case that we cannot measure the states of a dynamical system. We can observe only the inputs and the outputs of a system in many cases. Summing up the above the discussions on qualitative reasoning and fuzzy modeling, we may conclude: l ) motivations are different, 2) goals look similar, 3) tools are much more powerful in fuzzy modeling, and 4) methods of approach to modeling are different and those in fuzzy modeling are more varied and applicable. In any case, however, we see similarities between qualitative reasoning and fuzzy modeling rather than differences. The research domain of fuzzy logic started adopting a qualitative approach to systems analysis in the late 1960s. We are quite sure that fuzzy modeling can be directly applied to problems considered by a qualitative reasoning method. Despite these facts, it is a great pity that we can hardly find any reference to fuzzy logic in the papers related to qualitative reasoning.
2) If z is small and z 2 is more or less small or medium 1 small, then y is big, dy/dzl is very negative, and dy/dxz is very negative. We can regard most of fuzzy models and fuzzy control rules [24], [25] as qualitative models. We distinguish, however, in this paper, a qualitative model from a fuzzy model. The terminology a fuzzy model is used in a narrow sense. A qualitative model is considered a fuzzy model with something more, i.e., with more linguistic expressions. For example, a fuzzy model of the type 1) if z is approximately 3, then y is approximately 5 ; 2) if z increases by approximately 5, then y decreases by approximately 4 should not be called a qualitative model. We can call the above model a fuzzy number model to distinguish it from a qualitative model. Further, in an ordinary fuzzy model that is used in fuzzy control such as
l ) if is positive then is negative 2, if is positive medium, then ! is positive / the terms positive small, negative small, etc., are the labels conventionally attached to fuzzy sets, where the fuzzy sets play an important role, not the labels. We go beyond this stage by utilizing of the concept of linguistic approximation. That is, given a conventional fuzzy model with fuzzy sets, we improve its qualitative nature by using linguistic approximation techniques in fuzzy logic. In other words, we deal with a model in which we focus our attention on how to linguistically or qualitatively explain a system behavior as we shall see in what follows. Now let us discuss available sources, i.e., information or data, for qualitative modeling. We find the following classification of the sources: 1) conventional mathematical models 2) observation based on knowledge and/or experience 3) numerical data 4) image data 5 ) linguistic data. In qualitative reasoning as we have seen, a model is built based on 1) and 2). A conventional mathematical model is usually identified based on 2) and 3). A fuzzy model is also based on 2) and 3), as in the case of a mathematical model. Sources 4) and 5 ) are also useful to build a qualitative model. A qualitative model based on 4), image data, can be interpreted as image understanding, that is, a linguistic explanation of an image. As for the examples concerned with 5 ) , we can consider translation of one language into another or making a summary of a story. In this paper we deal with qualitative modeling based partly on 2) and mainly on 3), as usual; a black box approach. The reason for this is that system identification is a key issue in the black box approach and all the basic matters in model building appear in the context of system identification. Indeed we think that any method of modeling should be able to solve the problems discussed from now on.
__
C. Qualitative Modeling
What we imply here by a qualitative model [20]-[23] is a linguistic model. A linguistic model is a model that is described or expressed using linguistic terms in the framework of fuzzy logic instead of mathematical equations with numerical values or conventional logical formula with logical symbols. For example as we shall see later, a linguistic model of a two input-single output system is something like the following:
1 ) if .r1 is more than medium and .12 is more than medium. then is small, ag/13.r1 is sort of negative, and tl,y/tl,rz
is sort of negative.
TABLE I CLASSIFICATION OF IDENTII-ICATIOV Stucture Identification I a: Input candidates

b: Input variables
a: Number o rules f
Stucture Identification I1 Parameter identification
b: Partition of input space
Let us first consider the problems in modeling. Modeling is classified from the viewpoint of a description language. However, there are some common problems to be solved in modeling independently of both the description language and the data type. Thus we refer to the systems theory. The most complicated problems arise when we take a black box approach to modeling. In the black box approach, we have to build a dynamical model using only input-output data. This stage of modeling is usually referred as identification. It is worth noting that the concept of system identification and its formal definition were introduced by Zadeh in the early 1960's. We may suppose that Zadeh suggested the idea of fuzzy sets since he found difficulties in the identification of a complex system based on differential equations with numerical parameters. According to Zadeh's definition [26], given a class of models, system identification involves finding a model which may be regarded as equivalent to the objective system with respect to input-output data.
D. Identification in Fuzzy Modeling
this problem in multivariate analysis. In a conventional black box approach in systems theory, this type of identification is, however, not explicitly discussed, and models are based on preassigned input-output variables. Generally speaking, we need some criterion for identification to evaluate the performance of a model. For instance, we can use the output error, i.e., the difference between the model output and the real output. However as far as the structure identification is concerned, it is well known that we cannot use the output error. We need a special criterion for the type Ib. The structure identification of type 11, which is concerned with input-output relations, is further divided into the subtypes IIa and IIb. Let us first consider IIa. In this type we have to determine the number of fuzzy rules in a fuzzy model. By structure identification in ordinary systems theory, what we mean is to find the relations between the inputs and the outputs. Given a description language for modeling, it simply means the determination of the order of a model. Let us consider an input-output model of a static system:
,y = a[)
fL1.l'
+f L * 2 +. +
' '
U I 1 'I.'L
(7)
and also consider an input-output system:
model of a dynamical
Now let us look at the problems in the identification of a fuzzy model. The identification is divided into two kinds: structure identifications and parameter identification. Structure identification can be divided into two types, called, in this paper, type I and type 11, where each type is also divided into two subtypes. Before going into details, we show a classification of identification in Table I. Later we shall consider a multi-input and single-output system. I ) Structure Identijkation: Generally speaking, the structure identification of a system has to solve two problems: in type I, one is to find input variables and in type 11, one is to find input-output relations. Type I consists of Ia and Ib. In Ia, we find possible input candidates for the inputs to a system. There are of course an infinite number of possible candidates, which should be restricted to a certain number. This type of identification, like induction, i.e., selection of a finite number out of an infinite number, cannot be solved in general. Let us just recall Newton's law of kinetic motion. There is no systematic way to find the exact causes of an unknown phenomenon. We have to take a heuristic method based on experience andlor common sense knowledge; we do not discuss the identification of type Ia in this study. In the structure identification of type Ib, given the possible input candidates, we find a set of input variables to a system which affect the output. In this case we select a finite number out of a finite number and so there are some systematic ways to solve this problem. For example, we are familiar with
where f is a discrete time. In both cases, the order, it,, must be identified. We know that AIC [27], Akaike's information criterion, can be used to this aim in particular when a system is linear. The idea of the identification of the order, 71, can be explained as follows. We consider the model in (7). Given a set of the inFig, put-output data indicated by 0, 2(a) shows an identified model with a high order which has a good fitting for the data. This model minimizes the output error. However since the data are contaminated by noise, we might also observe another set of data indicated by A in the figure for which the model has no fitting. What is a model which fits both sets of data, independently of noise? Fig. 2(b) shows a model with a lower order which has a better fitting on the average than the model in Fig. 2(a): this model does not minimize the output error. So the identification of the order, i i , is very important in modeling: usually 'rt is determined so as to minimize the value of AIC, which, roughly speaking, means minimizing both the output error and the order. In a fuzzy model, the structure identification of this kind is stated in a different way. A fuzzy model consists of a number of if-then rules. The number, 71, of rules in a fuzzy model corresponds to the order, n, in a conventional model. This identification of 71 is called type IIa. There are two parts of an if-then rule: the premise part and the consequent part. So the rules have two structures in principle: the premise structure and the consequent structure. If we use a fuzzy model (see Appendix II), the consequent of which is of a functional type [ 121 as in (8), we have the same structure problem as in a conventional model. We will not use
12
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. I , NO. I. FEBRUARY 1993
i.-l.I'xl
(b)
xl
Fig. 3.
Partition of the input space.
(b)
Fig. 2 . Input+utput data and model
this sort of a fuzzy model in this paper. Below we shall discuss only the premise structure, which is concerned with IIb. Supposz we obtain the following fuzzy model with three rules by the identification of IIa: 1) if 5 1 is small, then y is small; 2 ) if 5 1 is big and z 2 is small, then y is medium; 3) if x:1 is big and 2 2 is big, then y is big. The premise space of the above model, which is twodimensional space of the inputs 2 1 and 5 2 , is partitioned into three fuzzy subspaces as we see; the number of rules corresponds to the number of subspaces. This partition is called the premise structure in a fuzzy model. So the identification IIb implies determining how the input space should be partitioned. Fig. 3(a) shows the partition of the input space in the above model. As we find in Fig. 3(b) and (c), there are a number of different partitions. The problem is combinatorial: we need a heuristic method to find an optimal partition together with a criterion [ 121. However in our approach, this problem is automatically solved in the process of the structure identification of IIa, where we use a fuzzy clustering method. As we discussed in the above, we deal with the problems of selecting the input variables (Ib) and finding the number of rules (Ha). 2) Parameter Identijication: In ordinary system identification, parameters are the coefficients in a functional system model. In a fuzzy model, the parameters are those in the membership functions of the fuzzy sets. There is not a big difference between the two except in the number of the parameters, there being many more in a fuzzy model. We can use a conventional criterion, i.e., the output error, for parameter identification also in a fuzzy model. We have to notice that the structure identification and the parameter identification cannot
be separately performed in principle. This fact makes the identification very complicated. In our approach, the parameter identification can, however, be separately performed after the structure identification. By summing up the above-mentioned tasks to be done for modeling, we could say that the ratio of importance of the structure identification of type I to that of I1 and the parameter identification would be, moderately speaking, 100 : 10 : 1. In any event, if we know the input candidates to a system, our problem is almost solved. After this, we can certainly find an algorithm of the identification depending on a description language. Unfortunately we do not have a systematic approach to structure identification of type Ia: it is a problem of induction. Structure identification of type Ib is a combinatorial problem. Type IIa is similar to the identification of the model order in ordinary identification. Type IIb is also a combinatorial problem which appears only in the case of if-then rules. The parameter identification is merely an optimization problem with an objective function. In this paper we will mainly deal with structure identification since parameter identification is neither valuable for discussions nor theoretically important. 11. FUZZYMODEL In this paper we use the following type of a fuzzy model for a multi-input and single-output system:
R' : if :c1 is Ai and :1:2 is A i . . . and xTtis Ai, then y is B i . (9)

where R' is the %thrule (1 5 i 5 vi), 1 5 j 5 71) are input x:j( variables, :I/ is the output, and Ai and B' are fuzzy variables. We can simply rewrite the form (9) as
R'
: if z is Ai then .
T J
is B i
(9')
where z = (XI.. . . . : r T l and A = ( A I , . . . A 7 , ) . ) For simplicity, the membership function of AS is denoted Ai ( e ) . In this section we regard the fuzzy variables A and B as those taking as values fuzzy numbers which are not necessarily
13
associated with linguistic labels. Our method of qualitative modeling consists of two steps. In the first step we deal with a fuzzy model in terms of fuzzy numbers. In the second step we give linguistic interpretations to this fuzzy model. That is, we deal with a fuzzy model with linguistic terms which we call a qualitative model. Occasionally we let B take singletons, i.e., Z real numbers, instead of fuzzy numbers. As far as fuzzy control is concerned, it is well known that we can use fuzzy rules with singletons in the consequents without losing the performance of the control. However in order to derive a qualitative model from the fuzzy model in (9), it is better to use fuzzy numbers in the consequents. As for reasoning, we modify partly the ordinary method as seen in the following steps 2) and 3): 1) Given the inputs xy. x ! , . . . , and x:, calculate the degree of match, wi, the premises for the ith rule, 1 5 i 5 m in as
W'
Ri:ifxlis Aiand
x2
isA2then i
.....
, ,
m
0
Fig. 4. Distance from the input S to the rule.
= A",z:) x
A ; ( : E ~ ). . . x Ak(z:). x
(10)
2) Then defuzzify B z in the consequents by taking the center of gravity:
b2
B()/ 'Y&
J' B"Y)4!.
I
111
(1 1)
3) Calculate the inferred value, ij, by taking the weighted f average of ba with repsect to u
m ...
2) Calculate the distance di between the input and the core region of the ith rule as shown in Fig. 4. As we see, the core region, in a two-dimensional case, of the ith rule: 2 Z" "if 5 1 is Ai and z is Ai then y is B , is the crisp region determined by the fuzzy sets Ai and A i , where membership grades Ai and Ai are 1. 3) Calcualte the inferred output, y, by
ij =
m
i=l
i=l
l l I ( d i ) * (hi+@
c#~711(di),
where m is the number of rules. As we find in the process of reasoning, the rule Ra translate to the form
i=l
i=l
R'
: if
z IS A' then y is b'.
(13)
In what follows we shall call a fuzzy model in the form (9) a position type model. Next we propose occasionally to use a position-gradient type model. It is often the case that we cannot build a fuzzy model over the whole input space because we lack data. In this case we need to take an extrapolation for estimating the output using local fuzzy rules. Note that in conventional fuzzy reasoning we take an interpolation using some number of rules. A. Structure Identification A position-gradient model is of the following form: Let us now discuss the method of structure identification. R' : if z is A' then y is B'and d!g/i)z is C'. (14) As stated in the Introduction we have to deal with structures C' where ay/dz = (ay/a.rl... . . t l ! ~ / t l . r , ~ ) . = (Ci. . . . . C;L). of type Ib: to find a set of input variables among the possible and a y / a ~ ~ partial derivative of y with respect to ,rJ. candidates, and with structures of type 11: to find the number is the Using the rules shown in (14), we can infer the output for a of rules and a fuzzy partition of the input space (see Table given input for which no rule of the position type is available; I). We shall use an ordinary method for Ib but propose a new if w' = 0 in all the rules for the inputs, 4 cannot be inferred method based on fuzzy clustering for 11. In general, it seems in the position type model. The reasoning algorithm of the that we can neither separate the identification of type Ib from that of type I1 nor separate the structure identification from the position-gradient rules is the following: parameter identification; these are mutually related. However 1 ) Defuzzify B' and C', and let those values be b' and c', we can separate these by using our method. This is a great respectively, where c' = ( r i . . . . . (.it). We can rewrite advantage of the new method. (13) as In principle, the algorithm for the identification is of the R' : if z is A' then y is b' and a*y/azis c' . ( 15) iterative type. We present the algorithm through numerical
(16) where 71 is the number of inputs, m the number of rules, d i the component of di on the x j coordinate axis, and rw(d2) = e x p ( - d i ) is the weight of the ith rule dependeing on the distance d'. As is seen in (16), the term bz x,"=,(dje:) is the x extrapolated value of the output using the value of its partial derivatives a y / d z j . We shall see below that the membership functions in this study are almost always of a trapezoidal type as shown in Fig. 4.
14
TABLE I1 INPUT-OUTPUT DATA NONLINEAR OF SYSTEM Group A No.

1
Group B
.rl 1.40 4.28 1.18 1.96 1.85 3.66 3.64 4.5 1 3.77 4.84 1.05 4.5 1 1.84 1.67 2.03 3.62 1.67 3.38 2.83 I .48 3.37 2.84 1.19 4.10 1.65
.r2 1.80 4.96 4.29 1.90 1.43 1.60 2.14 1.52 1.45 4.32 2.55 1.37 4.43 2.8 1 I .88
.r2 3.00 3.02
.r3 3.80 4.39 3.80 I .59 3.30 3.33 2.64 2.54 1.86 1.70 2.02
No.
.r 1
.1.2 4.13 1.11 2.27 3.74 3.18 4.66 3.88 2.55 2.12 4.42 4.71 1.06 3.66 1.76 1.41 1.35 4.25 2.68 4.97 3.80 1.97 4.76 3.90 1.35
.r3 4.38 3.13 2.27 4.25 3.31 4.1 I 1.27 2.07 1.1 1 1.71 1.53 2.91 1.23 3.00 1.92 4.96 3.24 4.89 2.77 4.73 4.16 2.63 2.55 2.75
.r4 3.21 1.80 3.61 3.26 2.07 3.74 2.21 4.42 4.44 1.23 2.08 2.80 3.62 3.82 2.25 4.04 3.89 1.03 2.63 2.69 2.95 3.88 1.65 1.01
2 3 4 5 6 7 8 9
1.31 3.35 2.10 3.52 2.46 1.95 2.51 2.70 1.33 4.63 2.80 1.97 2.47 2.66 2.08 2.75 1.51 2.40 2.44 1.99 3.42 4.99 2.27 3.94
27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50
IO
11
12 13 14 15 16 17 18 19 20 21 22 23 24 25
I .95 2.23
3.70 1.77 4.44 2.13 1.24 1.53 1.71 1.38
1.60 1.71 4.15 3.44 1.64 4.53 2.50 2.15 3.03 3.97 4.20 2.23 1.41 4.93 3.93 4.65 2.61 1.33 2.42 4.42
2.54 2.54 4.57
2.71 1.78 3.61 2.24 1.81 4.85 3.41 1.38 2.46 2.66 4.44 3.11 4.47 1.35 1.24 2.81 1.92 4.61 3.04 4.82 2.58 4.14 4.35 2.22
1.58 4.71 1.87 1.79 2.20 1.30 1.48 3.14 2.22 1.56 1.32 4.08 1.42 3.91 5.05 1.97 1.92 1.63 1.44 1.39 2.29 1.33 1.40 3.39
I .70 I .38 4.5 1 1.10 1.58 1.06 1.28 4.50 3.25 3.95 1.21 3.22 1.76 4.03
examples. Let us consider the following nonlinear static system with two inputs, x1 and 2 2 , and a single output, y:
We show a three-dimensional input-output graph of this system in Fig. 5. From this system equation, 50 input-output data are obtained (Table 11). The data of 2 3 and x4 are put as dummy inputs to check the appropriateness of the identification method. 1 ) Structure IdentGcation of Type I: In the proposed algorithm, the identification of type Ib is done between the identification of type I1 and the parameter identification. Let us, however, first consider the identification Ib. , We have four candidates, 2 1 - ~ 4 for the inputs to the system and have to find among them the actual inputs affecting the output, y. This is a combinatorial problem. For instance we can count 15 cases in this example: four cases if the system has only one input, six cases if it has two inputs, and so on. In general, let X be a set of possible input candidates 5 1 , 2 2 . . . . .x T L ; then the total number of cases is the number of subsets except an empty subset of X , i.e., 2" - 1. Here we take a heuristic method to select some inputs from among the candidates; we increase the number of inputs one by one, watching a criterion.
Fig. 5. Inputboutput relation of nonlinear system.
First we divide the data into two groups, A and B, as in Table 11. As a criterion to this purpose, we use the so-called regularity criterion, RC [28], in GMDH (group method of data handling), which is defined as follows:
-
Y?B)2/h
+ E : - Ip"!'/kB] L Y
1=l
kB
/2.
(18)
15
TABLE Ill STRUCTURE IDENTIFICATION NONLINEAR Ib OF SYSTEM Input Variables
level
RC
level
.r 1
Step 1
s2
0
0.863 0.830 0.937
.r3 .r1
level
.rl - .r2
Step 2
.rl - 1.3 .rl - . x i

Step 3 .r1 - s 2 - .r3
. r l - .rL - .rf
0.57 I 0.583
level 4 - xl x2 x3 x
Fig. 6. Search tree in identification of type Ib
0.483 0.493
X
X
see, the tree in this case consists of four levels. Each node of the tree corresponds to a subset of the set of input candidates: the node at the level 0 corresponds to an empty set. Thus the where total number of meaningful nodes in 2" - 1 for a case of n, input candidates. The nodes 2 1 and ~ 1 . 2 2 with double circles the number of data of the groups A k 1 and k ~ : are selected in searching the tree as shown in Table 111. Only and B; one node at each level is selected and so those nodes connected the output data of the groups A and B; y,' and yp: by dotted lines are not evaluated. Since the values of RC at the model output for the group A input y IB: level 3 become bigger than that of :c1 0 2 2 at level 2, the search estimated by the model identified is terminated at level 2 . If not, the search is continued until using the group B data; the final level. As a result of this algorithm, we evaluate at , B I. the model output for the group B input most n ( n l ) / 2 nodes in this algorithm out of 21L- 1. estimated by the model identified We can further reduce the number of evaluations. For using the group A data. example, suppose the RC of the node z1 0 2 4 at level 2 is not improved compared with that of the best node, 1c1, at level 1. Then we consider that the input xq should be eliminated As we can guess from the form of RC, we build two models and thus we stop evaluating the node 5 1 0 x2 0 5 4 at level 3. for two data groups at each stage of the identification. Notice This case will be shown later in the example of modeling a that we have to make the structure identification of type I1 and gas furnace (see Table VII). the parameter identification in order to calculate RC. We can 2 ) Structure ldenrijcation of Dpe II: Usually in the design easily find the meaning of RC in Fig. 2. of a fuzzy controller we first pay attention to rule premises and Now we show the outline of a heuristic algorithm for Ib. find an optimal partition based on a certain criterion. Here we First, we begin with a fuzzy model with one input. We make propose a different method; that is, we first pay attention to the four models: one model for one particular input. After the consequents of the rules and then find a partition concerning identification of the structure I1 and parameter identification, the premises. Also in our method, we do not take an ordinary which will be described later, we calculate RC of each model fuzzy partition of the input space, as is shown in Fig. 7, for and select one model to minimize RC from among the oneif we take this kind of partition, the number of rules increases input models. Next we fix the one input selected above and add exponentially with the number of inputs. For this reason, we another input to our fuzzy model from among the remaining introduce the fuzzy c-means method, abbreviated FCM, for three candidates. Our fuzzy model has two inputs at this stage. the structure identification of type IIb. The algorithm of FCM We select the second input as we do at the first step, according will be shown in Appendix I. Using FCM, we make fuzzy to the value of RC. clustering of the output data; we use all the data. As a result, We continue the above process until the value of RC every output y is associated with the grade of membership increases. The result is shown in Table 111. As is shown, .c1 belonging to a fuzzy cluster B. Notice that we now have the is selected at the first step, .r2 at the second step. At the third following data associated with the grade of the membership step, however, both the values of RC for the third inputs, . r ~ of y i ,in BJ(l 5 :j 5 c ) : and .rA are bigger than the minimal RC at the second step. So the search is terminated at this stage. As a result we have evaluated nine of the 15 case5 and succeeded in finding the true inputs, . r 1 and .r2. We can induce a fuzzy cluster A in the input space as in Let us look at Fig. 6 to understand the algorithm. This figure shown in Fig. 8. By making the projection of the cluster A shows a tree structure of this combinatorial problem. As we onto the axes of the coordinates .rl and .r2, we obtain the
- -
--I
16
x2
Input space
XI
Fig. 7. Ordinary fuzzy partition of input space
Output space
fi,/
X
7 p a c e
(c) Fig. IO. Construction of a membership function: (a) input cluster; (b) approximated convex fuzzy set; (c) approximated trapezoid fuzzy set.
Fig. 8. Fuzzy cluster in the input space.
Output space
Fig. 9. Projection of a fuzzy cluster. Fig. 1 I .

Two fuzzy clusters induced from one output cluster.
fuzzy sets A I and A2 as shown in Fig. 9. As is easily seen, we have at this stage the following relation:
where B is the output cluster. Now this cluster gives a fuzzy rule: if : c 1 is A I and 1c2 is A2, then :y is B. Remark I : Although the output cluster B is convex, the input cluster A corresponding to B might not be convex: for instance we might obtain A l , as is shown in Fig. 10(a). In this case we approximate the input cluster in Fig. 10(a) with a convex fuzzy set as in Fig. 10(b). Finally we approximate this convex fuzzy set and B as well, with a fuzzy set of trapezoidal type as shown in Fig. 10(c), which is used in the fuzzy model.
Remark 2: the next problem is that we might have more than two fuzzy clusters, A' and A2 in the input space which corresponds to the same fuzzy cluster B in the output space. In this case we carefully form two convex fuzzy clusters as illustrated in Fig. 11. We obtain the following two rules with the same consequent:
R1 : if : c 1 is A: and z 2 is Ai then ;y is B R 2 : if :rl is A: and L I : ~is Ai then y is B .

As we can easily understand, a fuzzy partition of the input space is obtained as a direct result of fuzzy clustering; we do not have to consider the structure identification of type IIb.
17
-20
-30
1 4\
-21.58
3.66
R ; if x l i
s
1.70 1.80
m
4.95 3.60
and x2 i s f o then y is
1.73 1.95 4.97 4.43
-32.47
R2; if XI s [ B i
1.50 4.80 4.50
and x2 s i m then y is
1.42 1.52 4.55 4.43
+
-41.69
-42.70
rb
1.51 2.60 2.51 2.66 1.53 3.31 3.91 3.94
1.30
20 .0
R3; if x l s i m and x2 i
0.50
1.18 4.60 0.91
s
1.24
g
7.97 4.29
then y is
2.84
Fig. 12. Behavior of S ( c ) in nonlinear system model.
1.35 1.65
R5; if XI w s i
and x2 i
1.06
s
2.40
c then y is
Let us discuss a method to determine the number of rules which is related to the number of fuzzy clusters in fuzzy clustering. Note that, because of Remark 2 the number of fuzzy rules is not exactly the same as that of fuzzy clusters in the output space, as we have just discussed above. The determination of the number of clusters is the most important issue in clustering. There are many studies on this issue [29]-[31]. Here we use the following criterion [31] for this purpose:
n c
1.02
4.98
R;
if x l
isfilla
1.05 1.81
and x2 i
s
1.06
K
4.62
then y is
a
4.05
3.22
4.94
5.06
Fig. 13. Model of the nonlinear sustem.
k l i=l
where
U:
c:
,rh.:
number of data to be clustered; number of clusters, c 2 2; kth data, usually vector; average of data: s1. .r2 . . ..rrz ; vector expressing the center of ith cluster; nom; grade of kth data belonging to ith cluster; adjustable weight (usually
171
After identifying the structure of type Ib (input variables), we combine the two groups of the data and induce fuzzy clusters in the input space from the output clusters to find the parameters A. We use those parameters obtained in the process of fuzzy clustering as an initial guess in the parameter identification. Fig. 13 shows the fuzzy model of (17). Below we shall use the mean square error of output as a performance index of a fuzzy model:
.r: :
11 . 11:
pLk:
111:
= 1.5
3).
The number of clusters, c, is determined so that S ( e ) reaches a minimum as c increases: it is supposed to be a local minimum as usual. As in seen in (21), the first term of the right-hand side is the variance of the data in a cluster and the second term is that of the clusters themselves. Therefore the optimal clustering is considered to minimize the variance in each cluster and to maximize the variance between the clusters. Fig. 12 shows the change of S ( c ) in fuzzy modeling of the system (17). In this case, the optimal number of fuzzy clusters is found to be 6. From this, the number of rules is identified by taking account of the case where a fuzzy cluster in the input space is divided into two fuzzy clusters. As it is easily seen, in the process of fuzzy clustering, we refer neither to the input nor to the parameters in premises. Therefore we can separate the identification of type I1 from that of Ib and the parameter identification. Futher as we find in the above discussion, a fuzzy partition of the input space IIb as well as the parameters Ai in the premises and Bi in the consequents are obtained as by-products of IIa. We use these A and Bi to calculate RC in the identification of type Ib.
where m is the number of data, yz is the ith actual output, and is the ith model ouput. In this example, we have P I = 0.318. 3) Identication of a Position-Gradient Model: Finally let us discuss the identification of a position-gradient type model of the form (14): the determination of the term dyldz, is Cj. Since the partial derivative dy/dzi is not given as data, we have to estimate it from the given input-output data. We assume an equation around the input-output data (z?.. . x;L,yo):
y = yo
+ dy/dzl(zl - z:) + . . . + dy/dz,(z,,
- x) :.
(23)
We estimate the coefficients dy/dzj in (23) by using the method of weighted least squares using the other input-output data. As weights, we use
where .c;(l 5 ,j 5 n ) is the kth data. Then we can obtain a fuzzy set Ci in the form of (14) from the output cluster B z just as we do to obtain the fuzzy sets A i .
18
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. I, NO. I , FEBRUARY 1993
Find the Darameters 6
t
Divide data into two groups
1
Assume single variable among input candidates
Pl
P2
P3
P4
I
t
i
Find the parameters A with respect to each group of data
Fig. 15. Trapezoidal fuzzy set.
Calculate RC for each input variable I
Combine two groups of data to reform the parameters A
I
I
Estimate b y / b x and find the parameters C for a positionoradient tvDe model
Make parameters identificationI
I
Fig. 16. Constraint for adjusting parameter
Fig. 14. Algorithm for identification.
Summing up the above method of identification, we arrive nonlinear optimization method for parameter identification, at the algorithm given in Fig. 14. We should notice as pointed out in the preceding section e.g., the complex method [12]. In the method we propose, we have a great advantage in that the parameter identification can be done separately from the structure identification in our method since the parameters that we can find, as we did in our example, the approximate obtained as by-products in fuzzy clustering are available to values of the parameters; we can use these values as an initial calculate RC in the process of the structure identification guess in the parameter identification. Now we show an algorithm for the identification process. Ib. In ordinary algorithms of identification, the parameter identification must be performed in the process of the structure Set the value f of adjustment. identification, which makes an algorithm complicated. Our Assume that the lcth parameter of the jth fuzzy set is pj". method also enables us to omit the identification IIb of the Calculate pj" f and p: - f . If lc = 2 , 3 , 4 , and pj" - f partition of the input space. is smaller than p$-', then 6; = else 6 = pj" - f . : Also if lc = 1 , 2 , 3 and p: f is bigger than pS+l, then
$';
B. Parameter Ident$cation
As discussed in the preceding sections, we determine at the stage of parameter identification the values of parameters in a system model. In the case of a fuzzy model, the parameters are those concerned with membership functions. In this paper, we approximate a convex fuzzy set with a trapezoidal fuzzy set. A trapezoidal fuzzy set has four parameters, as shown in Fig. 15, where p l 5 p 2 5 p3 5 p4. As mentioned, we have already found the parameters by making the projection of the clusters onto the axes of coordinates. We can derive a qualitative model from a fuzzy model with these parameters. However it is better to improve the parameters in order to use a fuzzy model for simulation. So we adjust the parameters as we do in the ordinary parameter identification. Since parameter identification is a problem of nonlinear optimization, there are studies on fuzzy modeling using the
p j = p;+l; else fij = pj" f . f is a constraint for adjusting parameters, as shown in Fig. 16. Choose the parameter which shows the best performance -IC P I in (21) among {fi:,p;,fij} and replace pj" with it. Go to step 2 while unadjusted parameters exist. Repeat step 2 until we are satisfied with the result. We use 5% of the width of the universe of discourse as the value of f and we repeat 20 times steps 1 to 6. Note that we do not adjust the parameters in the consequents of the rules.
-IC
MODEL 111. QUALITATIVE
A. Linguistic Approximation
In this section we discuss a way to derive a qualitative model from a fuzzy model. To this aim we use a method of linguistic approximation [32], [33] of fuzzy sets. We can
Fig. 17. Fuzzy set
and linguistic approximations P and
&.
state our problem as follows: Given a proposition with fuzzy predicates, find a word or a phrase out of a given set of words to linguistically approximate it, with hedges and connectives. After this procedure, we can obtain a qualitative model with linguistic rules from the identified fuzzy model. We use the following two indices to measure the matching degree of two fuzzy sets [34]: Degree of similarity:
l)f 2
(2) more or less A
.............
I-\*
(6) more than A
(3) slightly A
(7) less than A
S ( N ,L ) =II N n L
II / II N
UL
II,
(25)
.............
............. :
where (1 . 11 is the cardinality of the fuzzy set, and N and L are fuzzy sets. Degree of inclusion:
-: modified A
(4) sort of A
Fig. 18. Linguistic hedges.
I ( N , L ) =II N n L
II / II L II .
(26)
the above setting, N is a fuzzy set considered as a fuzzy number and L is a fuzzy set associated with a linguistic label. In order to choose the closest L to a given N , we may refer only S ( N ,L ) . In some cases, however, it is better to refer also I ( N , L ) .In Fig. 17, for two fuzzy sets P and Q, we have
S ( N ,P ) = S ( N ,Q) W , ) < I ( N ,Q ) . p As Q is included in N , we can say that Q approximates more appropriately N than P. We suggest leaving the exact
use of these to the users preference. In this paper we consider linguistic approximation on three levels.
Level 1: Level 2: Level 3:
approximation with linguistic terms. approximation with linguistic terms and hedges. approximation with linguistic terms, hedges and connectives.
0 (b)
10 (mk)
We shall use the following hedges [35]:very, more or less, slightly, sort o j not, more than, less than, as shown in Fig. 18. We apply certain constraints in the use of hedges. For instance we may say very big, or more than middle, but we do not say more than big. As an example, let us approximate two fuzzy numbers, A and B , concerning the strength of wind in Fig. 19(a) with the linguistic terms in Fig. 19(b). The top and bottom parts of Table IV show the linguistic approximations of A and B according to three levels where the values of two indices are shown.
Fig. 19. Fuzzy numbers and words for linguistic approximation: (a) fuzzy numbers and (b) wind scale.
B. Control of Number of Linguistic Rules
It is often the case that we are asked to explain phenomena more precisely or more simply. We have a limitation in responding to this sort of request if we merely try to refine or simplify the linguistic approximation. We can manage this problem by increasing or decreasing the number of linguistic rules. Two features contribute to this flexibility: hedges and connectives.
20
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 1 , NO. 1, FEBRUARY 1993
TABLE IV
2.10
RESULTS LINGUISTIC OF APPROXIMATION

Level
~~
R'; if x i s im and x2 i
I
0.96 0.97 0.94
1.06 1.84 4.85 3.45
s
1.15
m then y is
4.97
Linguistic Expression calm
S
0.62 0.87 0.94
1.93 3.20
less than soft

more or less calm or more or less soft
R'; if x i i s t n
1.46
I 4 n
and x2 i m then y is s 1.16 2.02 3.79
4.06
Del
r1.30 1.51 2.60 2.51 2.65 1.W 3.31 3.42
.. .-
R3; if XI s i
1.10 2.76
and x2 i
1.63
s
1.24 2.92
h then y is
5.21
Level I
Linguistic Expression moderate more or less moderate not strong but more or less moderate
,
S
0.76 0.81 0.86
I
0.92 0.83 0.98
2 0 227 .0
3.35
R 4 ; if XI s i
2.08 2.32
and x2 is&,
124
then y is
3.60
2.88
3.98
R5; if XI is&
1.40 5.95
and x2 i
1.03
s
2.88
then y is
t/l
3.22 4.94
4.05
3.91 3.94
As we adopt a fuzzy clustering method for the identification of IIa i.e., the number of rules, we can simply control the number of linguistic rules by adjusting the number of fuzzy clusters. This is also a great advantage of the proposed method. IV. ILLUSTRATIVE EXAMPLES We deal with six illustrative examples in this section. The first example involves qualitative modeling of a nonlinear system, which has been partly discussed in the previous section. The second is also a numerical example in which we discuss the performance of our fuzzy modeling method. In the third example we make a qualitative model of a dynamical process. The fourth and fifth examples are concerned with qualitative modeling of human operation at process control. In the final example we make a qualitative model to explain the trend in the time series data of the price of a stock. In order to understand the examples in connection with the preceding discussions, let us list the key points: a) fuzzy model of position type, b) fuzzy model of position-gradient type, c) qualitative model, d) structure of type Ib: input variables, e) control of the number of linguistic rules, f) structure of type IIa: number of rules, g) parameter identification. In what follows, we shall attach symbols (a)-(e) to each example. Throughout the examples, (f), the structure identification of type IIa, and (g), the parameter identification, are performed. A . Nonlinear System ( a , b, c, d ) This example deals with the explanation of the proposed method of identification. Let us recall a static and nonlinear two input-single output system of the type shown in (17):
Re; if x i i s f i ,
1.m 2.57
and x2 is&
1.06 4.93
then y is
5.m
Fig. 20. Model of nonlinear system after parameter identification.
This system shows the nonlinear characteristic illustrated in Fig. 5 and we use the data in Table 11. The process of the identification of type IIa has been shown in Table 111 and we have found the true inputs .c1 and .r2. The optimal number of
rules has been found to be 6. In this example, the number of rules is found to be equal to the number of clusters. After the parameter identification, we obtain a fuzzy model as in Fig. 20, where the performance index of the model is P I = 0.079. Before the parameter identification, we have obtained the fuzzy model with unadjusted parameters as we see in Fig. 13, where P I = 0.318. We can certainly improve the model by adjusting its parameters. In this example, we also obtain a fuzzy model of positiongradient type as illustrated in Fig. 21. In the case of the latter model, the performance is greatly improved. We have P I = 0.010. Finally we show a qualitative model of the position-gradient type. I ) If z is more than MEDIUM and 2 2 is more than 1 MEDIUM, then y is SMALL, .3y/dzl is sort of NEGATIVE, and dy/da2 is sort of NEGATIVE. 2 ) If 2 1 is not SMALL but less than MEDIUM BIG and 2 2 is not SMALL but less than MEDIUM BIG, then y is MEDIUM SMALL, dy/dzl is sort of NEGATIVE, and dy/dz2 is sort of NEGATIVE. 3) If 2 1 is not SMALL but less than MEDIUM BIG and :cp is more or less MEDIUM BIG, then y is more or less MEDIUM SMALL and MEDIUM, dy/d:1;1 is sort of NEGATIVE, and i)y/d:~2 is sort of NEGATIVE. 4) If 21 is less than MEDIUM and 1c2 is less than MEDIUM BIG, then y is MEDIUM and more or less MEDIUM BIG, i)y/d:cl is sort of NEGATIVE, and i)y/dzp is sort of NEGATIVE. 5) If : ~ ' 1 more or less MEDIUM SMALL or MEDIUM is and 2 2 is more or less SMALL, then y is MEDIUM BIG, i)y/i)zl is sort of NEGATIVE, and dy/i):c2 NEGATIVE. 6) If :c1 is SMALL and x 2 is more or less SMALL or MEDIUM SMALL, then y is BIG, dy/i)zl is very NEGATIVE, and dy/d:1;2 is very NEGATIVE.
21
TABLE V SOME INPUT-OUTPUT DATAOF FUZZY SYSTEM

.1'
12 '
13 '
11 '
Y
~~ ~
8.85 3.89 4.43 1.98 1.69 6.23

Consequents are singletons
9.32 9.19 3.67 6.74 7.20 7.26
0.69 2.49 2.95 7.08 5.41 6.50
5.84 9.00 5.17 6.62 9.23 6.25
6.00 1.56 -1.32 -3.04 -2.00 2.92
*
Fig. 21.
Position-gradient model of nonlinear system. TABLE VI STRUCTURE IDENTIFICATION OF Fuzzy SYSTEM Ib
x2
Input Variables
.' 1 1
RC
0
1 1.335
Step 1
.r2
.2.3
.rl
19.820 18.249
.rl
.r2
Step 2
.r1 - .r3 .2.1- . ' 11
10.904 10.476
1 xl
0
3
Fig. 22.
10
Fuzzy system.
We can understand the performance of the model by referring to the characteristics of the system shown in Fig. 5.
B. Fuzzy System Model (a, d )

In this example we investigate the reproductivity of the proposed method of system identification. That is, we obtain input4utput data from a fuzzy system consisting of fuzzy rules and make its fuzzy model using the data. Fig. 22 shows a fuzzy system with nine rules in the form of (13) which has singletons in the consequents. We use 100 data points of X I . 2 2 . : X S , :CA, and y taken from the system where 2 3 and :CA are dummy input variables; some data are shown in Table V. After the fuzzy clustering, we find the optimal number of clusters to be 8. From this we derive nine rules; one of eight clusters is divided into two rules because of the nonconvexity of an input cluster. The structure identification of type Ib proceeds as is shown in Table VI. In step 2 of this table, the two inputs, :1:1 and 2 2 , with the minimal RC are correctly selected. After parameter identification, we obtain a fuzzy model of the type shown in Fig. 23. We can see a good reproductivity by comparing the identified model with the original system.
.xl
Fig. 23. Fuzzy system model.
C. Box and Jenkins's Gas Furnace (b, c, d )

We discuss the qualitative modeling of a dynamical process using a famous example of the system identification given by Box and Jenkins [36]. The process is a gas furnace with single input u ( t )and single output y(t): gas flow rate and CO2 concentration, re5pectively. We omit to show data which are found in [36]. Since the process is dynamical, we consider as candidates ) the ten variables y ( t - 1).. . . , g ( t - 4 .u(t - 1).. . . . u ( t - 6) to affect the present output y(t). Using 296 data points, we make fuzzy clustering and find six clusters. Table VI1 shows
22
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. I , NO. 1, FEBRUARY 1993
TABLE VI1 Ib OF STRUCTURE IDENTIFICATION GAS FURNACE

Input Variables Y ( t - 1)
Y(t - 2 )
Y(t - 3 )
I y(t-1)
R'
u(t-4)
55.16
0.01
u(t-3)
1.4 -3.51
3
13.59
y(t) ayby(t-1) ayhu(t-4) ay/w(t-3)

47.66 0.79 -2.36 -1.39
RC
m
2.276 4.151 5.720 6.724 4.816 3.257 1.869 1.465 2.040 0.973 1.067
R( rr;11.3:lJ1.:
38.51
47.50 55.10 -2.61 2.1
50.22
0.79 -2.27 -1.02
-292 2.79
Step 1
Y(t - 4) u ( t - 1)
u(t- 2)
u(t - 3)
u ( t - 4)
u(t - 5 )
u ( t - 6)
Step 2
Y(t - 1) - Y ( t - 2 ) ?At - 1 ) - Y(t - 3 ) Y(t - 1 ) - Y ( t - 4) g ( t - 1) - u ( t - 1 ) y(t - 1) - u ( t - 2 ) Y ( t - 1) - u ( t - 3 ) g ( t - 1 ) - u ( t - 4)

y(t - 1 ) - u ( t - 5 )
Y(t
X X
1.020
0.745 0.598 0.439
58.74
51.11 63,13 4.62
1.m
4.41
0.79 -1.77 -0.72
286
Fig. 24.
Fuzzy model of gas furnace (position-gradient type).
0.565 0.846 0.439 0.429

X X
- 1)- u ( t - 5 )
g ( t -1)- u ( t -4)- u ( t -1)
Y ( t-1)-
u ( t- 4 ) -
u ( t- 2 )
step 3
y(t-l)--(t-4)-u(t-3) y(t - 1 ) u(f- 4 ) - u ( t
-5)
0.454 0.499
g ( t-1)-
u ( t- 4 ) - ~ ( f - 6 )
the structure identification process of type Ib. As we find, we can select y(t - l),u(t - 4), and u(t - 3) by referring to the values of RC. In Table VII, some sets of input candidates, e.g., {y(t 11,Y ( t - 211, {Y(t - I), Y ( t - 3)) and { d t - 11,Y(t - 4)) in step 2 are associated with the mark x . The mark x shows that the values of RC for those input sets are bigger than the minimal RC at the previous step, i.e., RC for y(t - 1). As discussed in subsection 11-A-1, we eliminate the input candidates y(t - 2), y(t - 3 ) , and y(t - 4) after step 2 since we may not have good prospects by considering those as candidates at proceeding steps. So we evaluate only five sets of input candidates at step 3 as seen in Table VII. After the parameter identification, we obtain a fuzzy model of the position-gradient type as shown in Fig. 24. Fig. 25 shows the model behavior in comparison with the actual process. In Table VIII, we compare our fuzzy model with other fuzzy models where we see a high performance of our model; the number of rules are remarkably reduced, and P I also. Now let us derive a qualitative model from a fuzzy model of position type using linguistic approximation at level 3. We obtain the following model.
TIE
- PROCESS '.... MODEL

Fig. 25. Output of fuzzy model of gas furnace.
1) If y ( t
- 1) is sort of SMALL or MEDIUM SMALL u(t - 4) is MEDIUM BIG or sort of BIG and u(t - 3)
is more or less MEDIUM BIG or sort of BIG, then y(t) is SMALL. 2) If y(t - 1) is not SMALL but more or less MEDIUM SMALL, u(t - 4) is not BIG but more or less MEDIUM BIG, and u(t - 3) is more or less MEDIUM, then y(t) is MEDIUM SMALL. 3) If y(t - 1) is not MEDIUM BIG but more or less MEDIUM, u(t - 4) is not MEDIUM SMALL but more or less MEDIUM, and u(t-3) is more or less MEDIUM, then y(t) is MEDIUM SMALL and MEDIUM. 4) If y(t - 1) is not MEDIUM SMALL but more or less MEDIUM, u(t - 4) is MEDIUM SMALL or more or less MEDIUM, and u(t - 3 ) is not small but less than MEDIUM BIG, then y(t) is MEDIUM or MEDIUM BIG.
23
TABLE VI11 COMPARISONFUZZYMODEL OF WITH OTHER MODELS Model Name Inputs Number of Rules
19
Y ( t - 1). tr(t - 4 )
(Candidates of input variables ) Monomer concentration Change of monomer concentration Monomer flow rate
( 6 rules)
( output )
Model Error 0.469 0.320 0.328
Tongs model [51 Pedrycs model [6] Xus model ~71
81 25
Set point for monomer flow rate
0.193
Temperature 2
r/(t- 1 , 1
Y ( t - 21,
8, :Selected 2 0.068 Fig. 26.
input variables
Takagi Sugeno model [35]
- 31, u(t - l ) , n(t - 2 ) , l / ( t- 3 )

Y(t Y(t
Structure of plant operation
Positiongradient model
u ( t - 4), tt(t - 3)
0.190
t
-50
5) If y(t
U(
1) is not BIG but more or less MEDIUM BIG,
t -4) is MEDIUM SMALL or MEDIUM, and U ( t - 3 )
is not SMALL but more or less MEDIUM SMALL, then y ( t ) is MEDIUM BIG and sort of BIG. 6) If y(t-1) is MEDIUM BIG or more or less BIG, u(t-4) is more or less MEDIUM SMALL or MEDIUM, and u(t - 3 ) is less than MEDIUM, then y(t) is very BIG. As is well known, we can make a linear model of this process. The performance index of the identified linear model is P I = 0.193. The P170.190,of our fuzzy model is almost the same as that of the linear model. We should remark that we do not aim at a good numerical model by fuzzy modeling in this study. If we want to do it by fuzzy modeling, we can use Takagi-Sugenos fuzzy model and we can improve the performance quite a bit more such that P I = 0.068
K
II
2
-62.52
-70
3 I
I 4
b b 7
Y C
Fig. 27. Behavior of S ( c ) in plant operation model.
ctl:
tt2:
monomer concentration, change of monomer concentration, monomer flow rate, local temperatures inside the plant, set point for monomer flow rate.
113:
114.tt.5:
Y:
Wl.
We will show this fuzzy model as well as a linear model in Appendix 11.
D.Human Operation ut a Chemical Plant ( U , c, d, e )

We deal with a model of an operators control of a chemical plant. The plant is for producing a polymer by the polymerization of some monomers. Since the start-up of the plant is very complicated, a man has to make the manual operation at the plant. The structure of the human operation is shown in Fig. 26. There are five input candidates which a human operator might refer to for his control, and one output, i.e., his control. These are the following:
where an operator determines the set point for the monomer flow rate and the actual value of the monomer flow rate to be put into the plant is controlled by a PID controller. We obtain 70 data points of the above six variables from the actual plant operation as shown in Appendix 111. First we find six clusters by fuzzy clustering; we obtain six rules in this case. Fig. 27 shows the change of S ( c ) in the process of IIa. The identification process of type Ib is as shown in Table IX. As a result, we conclude that the operator must refer to the three informations u l (monomer concentration), u2 (change of u l ) and 213 (monomer flow rate) to decide his control action. We show the obtained fuzzy model in Fig. 28. Fig. 29 shows the excellent performance of the fuzzy model.
24
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. I , NO. 1, FEBRUARY 1993
TABLE 1X STRUCTURE IDENTIFICATION CHEMICAL Ib OF PLANT OPERATION Input Variables

U1
RC
6027 15 6077539
U2
Step I
U3
U4
U5
lGOTjGI
6663660 5570 199
U3
- U1
46178
20
40
MODEL
60
Step 2
U3
U3
-U2 - U4 U3 - U 5
-U2 -U1
U5
00
64124 60277
X
TIRE - OPERATOR
Fig. 29.
Ouput of plant operation model.
Step 3
U3
138950
41846
000
X
U3-U2 -
5 ) If ul is more or less SMALL, and u2 is very INCREASED, and u3 is MEDIUM BIG, then y is BIG. 6) If ul is more or less SMALL, and u2 is sort of INCREASED, and u3 is BIG, then y is very BIG. We have shown this qualitative model to some operators and obtained their agreements with the model. It is worth noticing that we can make a fuzzy controller using the obtained fuzzy model to automate the operators control. In general it is difficult to evaluate this sort of a qualitative model since a model error in an ordinary sense is not available. One method is taken here is to ask an expert about the performance of a model. Also in such a case as subsection IV-A, we can qualitatively evaluate a model by comparing a model with the original shape in Fig. 5. If we are asked to explain the operators action in a simple form, we can present a qualitative model, for instance, with three linguistic rules by adjusting the number of clusters. Those are the following.
MI)
I
R
U1
U2
U3 = y )
765
aybul
-1604
yU aylau3 a2
67 0.98
47 .1
9.32 4.28 0.22 -721
6047
157;92
5.03
6.37 4.38 0.M
1629
769
5224 4.01 0.10
2072 3730
-1558 2594 0.99
aurn
4.97 5.21 3.74
6.09 4 3 0.15 .0
2545
-989 1166 0.99
16m 5977
d!lntn FKED
5.25 6.14 4.19 0.15 4261 4924 4.16 6.57 4 3 0.15 .3 1398 6636
3917
-1535
63 0.99
4.47
6.66 4 4 .4
,42h 0 6 2703 011 .1
5599
-1410 2858 0.84
R6
6806
-1440 373 0.86
4.47
726 4.30 0.20
2358
7032
M2)
Fig. 28.
Fuzzy model of plant operation (position-gradient type). M3)
From the fuzzy model, we can derive the following qualitative model.
If u l is MEDIUM SMALL, and rr2 is less than NO CHANGE, and (13 is sort of SMALL, then y is sort of SMALL or more or less MEDIUM SMALL. If ul is more or less MEDIUM, and u2 is ANY, and i t 3 is more or less MEDIUM, then y is more or less MEDIUM. If u l is sort of SMALL or more or less MEDIUM SMALL, and tr2 is more or less NO CHANGE, and I13 is more than MEDIUM BIG, then y is more then MEDIUM BIG.
1) If ul is more or less BIG, and u2 is not INCREASED, and u3 is SMALL, then y is SMALL or MEDIUM SMALL. 2 ) If u1 is more or less MEDIUM, and u2 is DECREASED, and u3 is SMALL or MEDIUM SMALL, then y is MEDIUM SMALL. 3 ) If ul is MEDIUM, and u2 shows NO CHANGE, and u3 is MEDIUM SMALL or MEDIUM, then y is MEDIUM. 4) If u1 is more or less MEDIUM, and u2 is ANY VALUE, and u3 is MEDIUM, then y is MEDIUM or MEDIUM BIG.
By comparing those MlLM3 with the above 1-6, we can find that M 1 is a summary of 1, 2, and 3, M2 is that of 3 and 4, and M3 is that of 4, 5, and 6. That is, the original six rules are fuzzily summarized in three rough rules.
E. Human Operation ut Water Purification Process (a, e, d )

A water purification process is a process to produce clean water for civil use. Turbid water, called raw water, flows into the process from rivers or lakes and then a chemical product,
XI
x2
x3
x4
x5
R'
2 R
R3
Ft
Fl
0.W
30.14
R
15.91 171.68 256.11 7.11 7.21 28.99 14.76
7 R
Fig. 30. Fuzzy model of water purification process operation.
called a floccul ting agent, is put into the process so that the turbid part of the water form flocks by cohesion which settle out in the settling basin. After sterilization by chlorine, we get clean water. The chemical reactions in this process are very complicated and so a human operator has to decide the amount of the flocculating agent by referring to some state variables. In principle, his control is of a feedforward type; it is generally difficult to automate such a human operator behavior. The referred variables are:
XI
:
2 :
change of turbidity of raw water, turbidity of raw water, turbidity of water in the settling basin, alkalinity, Fig. 3 1.
TIRE
- OPERATOR
MODEL
.I,:$:
.r.~: .r::
Output of operation model at water purification process.
PH,
amount of flocculating agent.
CJ:
F. Trend of Stock Prices
(U,
c)
We use 624 data points. Making the identification of the operator's control action, we obtain a fuzzy model with seven rules as is shown in Fig. 30. Fig. 31 shows the performance of the model. From this fuzzy model, we derive a qualitative model as is shown in Table X.
Finally we deal with the trend data of stock prices. We use the daily data of two different stocks: 100 data points of the first stock, A, and 150 data points of the second stock, B. The data of stock A are shown in Appendix IV. Both data consists of ten inputs and one output. These are
26
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 1, NO. I , FEBRUARY 1993
TABLE X QUALITATIVE MODEL WATER PURIFICATION PRWESS OPERATION OF Rule Change of Turbidity of Raw Water very small slightly small slightly small medium medium more or less big medium Turbidity of Raw Water slightly small slightly small slightly big medium medium sort of small slightly small less than medium slightly small more than medium Turbidity of Water in the Settling Basin Alkalinity P Amount of Flocculating Agent small medium-small sI i ghtI y medium-small medium slightly medium-big medium-big big
1
2
more than medium less than medium
more than medium sort of small
sort of small
sort of small
medium more or less big medium slightly big
5
6
7
more or less small small slightly small
-30
-20
1EZl
20
40
TIME
60
80
40
80
TIUE
120
160
- STDCK PRICE ..... EODEL

Fig. 32. Output of stock price model A.
- STOCK PRICE ..... MODEL

Fig. 33. Output of stock price model B.
past change of moving average ( I ) over a middle period; present change of moving average ( I ) over a middle period; past separation ratio ( I ) with respect to moving average over a middle period; present separation ratio ( I ) with respect to moving average over a middle period; present change of moving average (2) over a short period; past change of price ( I ) , for instance, change on one day before; present change of price ( I ) : past separation ratio (2) with respect to moving average over a short period; present change of moving average (3) over a long period; present separation ratio (3) with respect to moving average over a short period; prediction of stock price.
The qualitative models of the trend data of the two stock prices are shown in Tables XI and XII. This sort of study is quite interesting. It implies that we induce qualitative laws, by observing data, to which economic phenomena are subject.
V. CONCLUSIONS
We have discussed an approach to qualitative modeling based on fuzzy logic. A qualitative model is derived from a fuzzy model using the linguistic approximation method. We have proposed the use of a fuzzy clustering method for the structure identification of a fuzzy model. The proposed method has been examined in several case studies. What we have to do further in qualitative modeling is to improve the linguistic approximation method.
Here the separation ratio is a value concerning the difference between a moving average of a stock price and price of a stock. Making fuzzy clustering, we obtain five rules in both cases. Here we omit fuzzy models. Figs. 32 and 33 show a comparison of the price with the actual price; we can see the good performance of the models.
APPENDIXI ALGORITHM FUZZY C-MEANS OF METHODS [30]

Given a set of vector data Z k , 1 5 k 5 ri, we classify those into a certain number of fuzzy clusters. A fuzzy cluster is characterized by p,k. which shows the grade of membership
27
TABLE XI QUALITATIVE MODEL TREND OF OF Change of Moving Average 1 Rule Past

1
STOCK PRICE
A Separation Ratio 2 Change of Moving Average

7
Separation Ratio I
Change of Moving Average
Change of Price 1
Separation Ratio 3 Prediction
Now
level or rise level or decline level or decline
Past
sort of plus
Now
less than zero minus
Now
level more or less level more or less level more or less level
sort of rise
Past level
Now
rise
Past more or less zero more or less zero more or less zero more or less zero
Now
level level more or less level more or less level level or rise
Now
more less zero more or less zero more or less zero more or less zero more than zero
rise
more or less rise
2
3
level level or
rise
sort of rise level sort of decline decline
more or less zero plus more than zero
sort of minus
sort of decline
level level
E: r
5
level or rise
level rise
more or less zero plus
sort of rise sort of rise
sort of plus
TABLE XI1 QUALITATIVE MODEL TREND STOCK OF OF Change of Moving Average 1 Rule Past
1
PRICE
B Separation Ratio 2 Change of Moving Average Separation Ratio 3 Prediction
Separation Ratio 1
Change of Moving Average
Change of Price 1
3
Past sort of decline
sort of decline
sort of rise
Now
level or decline level or decline level more or less rise
Past less than zero
Now
sort of minus
more or less zero more or less zero slightly plus more than zero
Now
level level or rise
Now
more or less level
Past
sort of minus
Now
sort of decline
sort of rise
Now
more or less zero more than zero rise
sort of rise
sort of rise
level or rise
less than zero

more or less zero more or less zero plus
level or rise
level or rise level or rise more or less zero

sort of minus
sort of rise
level less than zero more or less zero

sort of decline
4
5
more or less rise
more or less level level

sort of decline
level or decline
more or less rise
level
decline
of the kth data, xk,belonging to the ith cluster. We assume that
3) Calculate the element of a new U ( ' ) as follows for 1=1+1:
where c is the number of clusters, and we define a matrix U consisting of p i k . Our problem now is to find c and to determine U . The algorithm is as follows 1) Set c = 2, an initial value U(1) of U and 1 = 1. 2) Calculate the center wi of the fuzzy cluster by
n
We define the distance from the kth data to the center of the ,ith cluster by
11 U ( l - l ) - U ( l ) I(<- E , then stop; otherwise go to step 2. In this algorithm, r n is an adjustable parameter, which is set as 2 in this paper.
4) If
28
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. I . NO. I . FEBRUARY 1993
APPENDIXI1 TAKAGI-SUGENOS FUZZY MODELAND A LINEAR MODELOF GAS FURNACE Takagi-Sugenos fuzzy model is of the following form;
APPENDIX I11
DATA HUMAN OF OPERATION AT A CHEMICAL PLANT

U1
Ri: if x1 is Ai and
yi = c6
. . . and x, is AA then
+ c;x1 + . . . + cix,,
(31)
where yz is the outut from the ith rule, A; is a fuzzy set, and c; is a consequent parameter. :, Given an input (xy, x:, . . . , x) the final output of this model is inferred by taking the weighted average of the yzs:
n
n
where yio is calculated for the input by the consequent equation of the ith implication and the weight w 2 implies the degree of match in the premise as of the ith rule calculated as in (10). This fuzzy model can express a highly nonlinear functional relation in spite of small number of rules. The simple algorithm of the identification of this model is: 1 ) Choose the premise structure and the consequent structure. 2) Estimate the parameters of the structure determined in 1). 3) Evaluate the model. 4) Repeat 1 ) until we are satisfied with the result. For further details of the identification of Takagi-Sugenos model, the reader is referred to [12]. Using the above type of modeling method, we obtain a fuzzy model of Box and Jenkinss gas furnace with two rules as follows [37]:
6.80 6.59 6.59 6.50 6.48 6.54 6.45 6.45 6.20 6.02 5.80 5.51 5.43 5.44 5.51 5.62 5.77 5.94 5.97 6.02 5.99 5.82 5.79 5.65 5.48 5.24 5.04 4.81 4.62 4.61 4.54 4.71 4.72 4.58
U2 -0.05 -0.21 0.00 -0.09 -0.02 0.06 -0.09 0.00 -0.25 -0.18 -0.22 -0.29 -0.08 0.01 0.07 0.11 0.15 0.17 0.03 0.05 -0.03 -0.17 -0.03 -0.14 -0.17 -0.24 -0.20 -0.23 -0.19 -0.01 -0.07 0.17 0.01 -0.14
U3
401.00 464.00 703.00 797.00 717.00 706.00 784.00 794.00 792.00 1211.00 1557.00 1782.00 2206.00 2404.00 2685.00 3562.00 3629.00 3701.00 3775.00 3829.00 3896.00 3920.00 3895.00 3887.00 3930.00 4048.00 4448.00 4462.00 5078.00 5284.00 5225.00 5391.00 5668.00 5844.00
U4 -0.20 -0.10 -0.10 0.10 -0.10 -0.20
U5 -0.10 0.10 0.10 0.10 0.10 0.10 0.00 0.10 -0.20 0.10 0.00 0.00 0.00 0.10 -0.20 0.00 -0.10 0.00 -0.10 0.10 -0.10 -0.10 0.10 0.00 -0.40 0.10 -0.10 0.00 -0.20 0.10 -0.10 0.00 -0.10 -0.10 0.20 -0.10 0.20 -0.10 0.20 -0.10 -0.10 0.00 0.20 0.00 0.10 0.00 0.00 0.00 0.00 0.10 -0.30 0.30 -0.10 0.20 -0.30 0.10 -0.10 0.00 0.00 -0.10 -0.20 0.10
Y 500.00 700.00 900.00 700.00 700.00 800.00 800.00 800.00 1000.00 1400.00 1600.00 1900.00 2300.00 2500.00 2800.00 3700.00 3800.00 3800.00 3800.00 3900.00 3900.00 3900.00 3900.00 3900.00 4000.00 4400.00 4700.00 4900.00 5200.00 5400.00 5600.00 6000.00 6000.00 6100.00
U1
then y(t) = 1.45y(t - 1) - 0.87y(t - 2)
+0.19~(t- 1) - 0.214t - 2)
+ 0.23y(t - 3) + 10.3
Rz : i f ~ (-t 1) is
4. 88
59.0
then y(t) = 1.44y(t - 1) - 0.70y(t - 2)
+O.lOu(t - 1) - 0 . 2 6 ~ ( - 2) - 0.27u(t - 3) t
+ O.lOy(t - 3) + S.32
On the other hand, we can also obtain a linear model of this system using the conventional system modeling method. The linear model is:
y(t) = 0.57y(t - 1)
+ 0.02y(t - 2)
-
0.53,/L(t- 3 )
- 0 . 3 3 * ~-t 4 ) (
0 . 5 1 ~ (- 5) t
+ 21.87.
4.55 4.59 4.65 4.70 4.81 4.84 4.83 4.76 4.77 4.77 4.77 4.73 4.73 4.74 4.77 4.71 4.66 4.70 4.63 4.61 4.57 4.56 4.54 4.51 4.47 4.47 4.48 4.48 4.50 4.50 4.48 4.54 4.57 4.56 4.56 4.57
U2 -0.03 0.04 0.06 0.05 0.11 0.03 -0.01 -0.07 0.01 0.00 0.00 -0.04 0.00 0.01 0.03 -0.06 -0.05 0.04 -0.07 -0.02 -0.04 -0.01 -0.02 -0.03 -0.04 0.00 0.01 0.00 0.02
0.00
-0.02 0.06 0.03 -0.01
0.00 0.01
U4 U3 - U5 Y 6068.00 -0.20 0.00 6400.00 6250.00 -0.20 -0.10 6400.00 6358.00 -0.10 -0.10 6400.00 6368.00 -0.10 0.00 6400.00 6379.00 -0.30 0.00 6400.00 6412.00 -0.10 -0.10 6400.00 6416.00 0.10 -0.10 6500.00 6514.00 0.00 0.00 6600.00 6587.00 -0.10 0.10 6600.00 6569.00 0.00 -0.10 6600.00 6559.00 0.00 0.00 6700.00 6672.00 0.00 0.00 6700.00 6844.00 -0.10 0.00 6800.00 6775.00 -0.20 0.00 6800.00 6779.00 0.00 -0.10 6800.00 6783.00 0.00 0.00 6800.00 6816.00 0.00 0.00 6800.00 6812.00 0.00 0.00 6800.00 6849.00 0.00 0.00 6800.00 6803.00 0.00 0.00 6800.00 6832.00 0.00 0.10 6800.00 6832.00 -0.10 0.10 6900.00 6862.00 -0.10 -0.10 7000.00 6958.00 0.10 -0.10 7000.00 6998.00 0.00 0.10 7000.00 6986.00 -0.10 0.10 7000.00 6975.00 0.00 0.00 7000.00 6973.00 0.00 0.00 7000.00 7006.00 0.00 0.10 7000.00 7027.00 0.00 0.00 7000.00 7032.00 0.00 0.00 7000.00 6995.00 0.00 0.00 7000.00 6986.00 0.10 -0.10 7000.00 7009.00 -0.10 0.10 7000.00 7022.00 0.00 0.00 7000.00 6998.00 - 0.00 7000.00 -0.10
29
APPENDIX IV
DAILYDATA OF STOCK A
xl 0.239 0.279 0.319 0.346 0.351 0.269 0.250 0.256 0.296 0.326 0.354 0.395 x2 0.237 0.246 0.272 0.270 0.303 0.268 0.267 0.336 0.379 0.400 0.350 0.315 0.295 0.294 0.350 0.340 0.404 0.478 0.401 0.427 0.361 0.378 0.413 0.457 0.455 0.408 0.415 0.404 0.394 0.410 0.364 0.318 0.238 0.176 -0.026 0.070 0.053 -0.035 -0.044 -0.097 0.0 -0.070 -0.114 -0.026 0.026 0.053 0.035 0.018 0.105 x3 8.307 16.771 21.253 19.867 19.447 14.320 11.157 12.787 12.453 12.088 13.587 15.971 19.283 22.512 28.465 26.016 19.058 13.125 11.815 10.504 11.021 10.674 13.067 11.830 9.733 5.862 7.374 6.214 5.066 2.165 2.820 5.222 5.819 4.654 7.848 5.809 5.527 11.263 14.309 16.444 14.317 14.815 12.767 16.698 19.689 21.823 26.391 30.825 29.467 x4 2.820 6.222 6.819 4.654 7.848 6.809 5.527 11.263 14.309 16.444 14.317 14.815 12.767 16.698 19.689 21.sa3 26.391 30.825 29.467 28.084 20.990 22.186 25.798 26.044 28.734 25.776 28.491 25.557 24.262 24.554 24.102 17.360 17.873 10.554 -5.226 2.607 -2.184 -10.837 -9.218 -14.666 -5.967 -13.023 -16.885 -9.731 -6.585 -9,802 -12.209 -10.641 -5.201 x5 -0.498 -0.250 -0.418 -0.420
x6 0.810
2.588 0.841
-0.834
3.364 -1.627 0.0 5.790 3.127 2.274 -1.483 0.752 -1.494 3.791 2.922 2.ia9 4.170 4.003 -0.641 -0.646 -5.198 1.371 3.381 0.654 2.599 -1.900 2.582 -1.888
-0.641
0.084
0.253 0.084 0.756 1.168 1.649 1.379 1.200 0.949 1.410 1.390 1.752 2.171 2.051 1.651 1.342 0.906 0.967 1.436 1.146 1.133 0.725 0.589 0.0
0.0
0.388
0.433 0.466 0.466 0.391 0.346 0.348 0.357 0.357 0.314 0.292 0.297 0.266 0.255 0.264 0.249 0.243 0.223 0.227 0.246 0.272 0.270 0.303 0.268 0.267 0.336 0.379 0.400 0.350 0.315 0.295 0.294 0.350
0.340
0.404 0.478 0.401
0.130 0.650 0.0 -0.258 -0.906 -2.482 -1.674 -2.384 -3.001 -2.878 -3.556 -2.842 -3.004 -3.586 -2.198 -0.173 -1 385 -1.141 0.0 0.444
0.646 0.0 -5.131 0.676 -6.044 -14.296 8.340 -4.619 -8.878 1.771 -6.092 10.195 -7.569 -4.550 8.580 3.512 -3.393 -2.634 1.803 6.200
x7 -3.281 1.696 -0.834 -0.841 -2.545 0.870 2.588 0.841 -0.834 3.364 -1.627 0.0 5.790 3.127 2.974 -1.483 0.752 -1.494 3.791 2.922 2.129 4.170 4.003 -0.641 -0.646 -5.198 1.371 3.381 0.654 2.599 -1.900 2.582 -1.888 -0.641 0.646 0.0 -5.131 0.676 -6.044 -14.296 8.340 -4.619 -8.878 1.771 -6.092 10.195 -7.569 -4.550 8.580
x8 -2.569 -4.568 -3.255 -0.502 0.756

0.338
3.626 1.682 1.597 6.672 8.739 9.408 6.320 5.850 3.289 5.714 7.311 7.710 9.817 11.917 9.393 7.247 0.760 11 3 .6 3.102 2.600 4.087 1.374 3.383 1.431 0.781 1.299 0.646 -4.519 -3.625 -8.622 -19.692 -11.512 -13.538 -18.777 -14.889 -17.127 -6.008 -10.432 -11.327 -1.556 2.078 0.0 -1.510
x9 -2.523 0.848 4.352 2.688 3.364 0.834 1.682 4.068 9.098 11.580 3.909 1.516 -2.224 3.010 5.228 9.098 9.496 10.646 7.644 2.668 -6.414 -4.519 -0.650 5.483 6.761 1.308 3.249 -1.267 00 . -1.888 0.0 -4.519 -4.490 -10.263 -18.932 -12.760 -11.437 -5.838 -1 1.547 -12.914 5.314 -4.352 -2.780 -4.205 7.279 8.580 -2.634 -4.241 5.268
x10 -3.355 -0.502 0.756
Y
34.612 24.390 24.187 17.662 -2.441 7.444 2.481 -11.728 -12.889 -20.015 -10.534 -17.924 -20.470 -16.801 -16.324 -20.848 -26.017 -27.582 -22.595 -23.392 -19.191 -24.341 -26.161 -22.092 -21.533 -20.013 -23.914 -18.602 -20.658 -19.243 -19.243 -14.875 -16.790 -13.581 0.0 -10.778 -4.036 3.543 0.0 5.561 -3.364 5.459 16.206 5.268 2.544 0.878 2.705 4.429 -3.336
038 .3
3.626 1.682 1.597 6.672 8.739 9.408 6.320 5.850 3.289 5.714 7.31 1 7.710 9.817 11.917 9.393 7.247 0.760 1.163 3.102 2.600 4.087 1.374 3.383 1.431 0.781 1.299 0.646 -4.519 -3.625 -8.622 -19,692 -11.512 -13.538 -18.777 -14.889 -17.127 -6.008 -10.432 -11.327 -1.556 2.078 0.0 -1.510 0.266 6.012
(Table continues on next p a g e )
30
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 1, NO. I , FEBRUARY 1993
xl 0.427 0.361 0.378 0.413 0.457 0.455 0.408 0.415 0.404 0.394 0.410 0.364 0.318 0.238 0.1 76 -0.026 0.070 0.053 -0.035 -0.044 -0.097 0.0 -0.070 -0.114 -0.026 0.026 0.053 0.035 0.018 0.105 0.123 0.088 0.061 0.017 0.061 0.105 0.149 0.113 0.174 0.148 0.174 0.130 0.052 -0.009 -0.026 -0.035 -0.026 0.026 -0.009 -0.026 -0.035
x3 28.084 20.990 001 .6 22.186 0.017 25.798 001 . 6 . 26.044 0.106 28.734 0.149 25.776 0.113 28.491 0.174 25.557 0.148 24.262 0.174 24.554 0.130 24.102 17.360 0.052 -0.009 17.873 -0.026 10.654 -0.035 -5.226 -0.026 2.607 -2.184 0.026 -0.009 -10.837 -0.026 -9.218 -0.035 -14.666 -0.043 -5.967 -0.061 -13.023 -16.885 -0.043 -9.731 -0.096 -6.585 -0.148 -9.802 -0.183 -0.131 -12.209 -0.044 -10.641 -0.052 -5.201 -6.897 -0.009 -0.017 -6.978 0.017 -11.766 -10.993 -0.044 -5.533 -0.035 0.009 -2.484 -2.628 -0.017 -0.061 -5.094 -0.035 -0.557 -0.026 -3.834 0.053 -1.658 -1.786 0.149 0.096 -1.837 -3.388 0.087 0.052 -5.702 -6.450 0.0 -9.547 0.009 -7.230 0.044 -8.783 0.026 -10.320 -0.044 -11.070 -0.079
x2 0.123 0.088
x4 -6.897 -6.978 -11.766 -10.993 -5.533 -1.484 -2.628 -5.094 -0.557 -3.834 -1.658 -1.786 -1.837 -3.388 -5.702 -6.450 -9.547 -7.230 -8.783 -10.320 -11.070 -10.250 -9.414 -4.683 -6.157 -5.234 -9.772 -10.440 -7.255 -8.780 -5.624 -5.607 -2.475 -5.582 -5.549 -3.194 -8.692 -11.001 -9.393 -9.369 -4.686 4.618 2.158 2.069 -1.911 -3.482 -4.275 -1.177 2.720 -0.375 -1.082
x5 0.884 -0.088 0.175 0.700 0.522 0.519
x6 -1.668 0.0 -5.089 0.894 6.200 3.336 0.0 081 .6 -2.421 0.853 4.963 1.184 -3.152 0.251 2.441 0.667 0.663 0.0 0.0 1.152 -1.589 0.895 0.081 -2.421 -0.827 -0.322 -3.336 -0.647 -0.163 2.588 -0.815 -1.682 -0.657 -1.711 -0.870 -0.993 0.878 -0.919 0.870 -0.843 5.177 -0.170 -1.641 -0.085 0.834 0.085 -0.085 -4.963 -0.870 -0.426 3.512 0.086 -1.696 0.085 3.451 0.513 0.0 0.425 3.336 0.677 -0.168 -3.228 0.0 0.0 0.168 2.502 -5.696 0.084 -0.084 -2.588 -0.252 1.771 -0.084 0.0 5.222 0.084 9.926 1.096 0.500 -2.257 0.o 0.830 -3.849 0.412 -1.601 0.0 -0.814 0.492 1.060 3.281 1.291 3.971 -3.056 0.956 -0.788 0.395
x7 3.512 -3.393 -2.634
xa 0.266 6.012 3.330 3.411 1.803 -2.014 6.200 -1.826 -1.668 3.720 0.0 6.626 -5.089 5.717 0.894 2.284 6.200 6.104 3.336 2.502 0.0 4.308 -2.421 3.621 4.963 2.441 -3.152 2.441 -0.081 -2.579 0.0 -3.072 0.0 -5.696 -1.589 -3.097 -2.431 -3.944 -0.827 -4.963 -3.336 -4.845 2.588 -3.120 -1.682 -1.446 -1.711 3.833 -0.870 2.217 0.878 2.981 0.870 -2.046 5.177 -2.483 -1.641 0.855 0.834 -0.940 -4.963 1.956 -0.870 1.524 3.512 4.205 -1.696 3.451 1.011 0.0 1.011 3.364 3.336 -2.605 -3.228 -5.046 0.0 -3.120 2.502 -3.038 -5.696 -2.588 1.939 1.771 1 0.842 7.801 0.0 5.222 6.914 2.377 9.926 -2.257 0.738 -0.571 0.0 -3.849 1.614 -1.601 4.303
x9 6.312 4.429 -6.672 -4.241 1.696 10.724 9.743 0.834 2.421 -0.807 4.136 -0.788 2.441 -1.589 -3.971 -4.766 -6.457 -1.654 -2.502 -0.863 -4.205 -1.711 0.870 7.024 4.352 4.314 -5.742 -5.004 -2.481 0.870 5.268 1.696 6.902 0.0 0.0 -0.807 -3.336 -5.838 -6.509 -0.863 7.086 15.666 13.055 7.444 -6.020 -5.389 -6.159 0.801 6.509 4.102 0.0
x10 3.330 3.421 -2.014 -1.826 3.720 6.626 5.717 2.284 6.104 2.502 4.308 3.621 2.441 -0.081 -2.579 -3.072 -5.696 -3.097 -3.944 -4.963 -4.845 -3.120 -1.446 3.833 2.217 2.981 -2.046 -2.483 0.855 -0.940 1.956 1.524 4.205 1.011 1.011 3.364 -2.605 -5.046 -3.120 -3.038 1.939 10.842 7.801 6.914 2.377 0.738 -0.571 1.614 4.303 0.158 -1.022
Y
1.696 1.696 10.724 6.200 0.0 -0.807 -6.457 -6.617 -9.456 -6.509 -3.971 5.560 3.177 4.843 3.308 2.502 5.177 5.887 11.976 10.444 10.536 6.962 5.177 0.820 1.668 1.654 6.962 7.024 5.089 8.628 5.004 2.502 -0.807 1.668 1.668 1.627 6.040 10.629 7.833 10.444 4.963 -5.267 -3.849 -3.849 3.203 4.882 6.563 0.794 -1.528 4.728 4.766
REFERENCES
[ I ] L. A. Zadeh, Fuzzy algorithm, Information and Control, vol. 12, pp. 94-102, 1968. [2] L. A. Zadeh, Outline of a new approach to the analysis of complex systems and decision processes, IEEE Trans. Syst.. Man, Cybem.. vol. 3 , pp. 2 8 4 4 , 1973. [3] E. H. Mamdani and S. Assilian, Applications of fuzzy algorithms for control of simple dynamic plant, Proc. Inst. Elec. Eng., vol. 121, pp. 1585-1588, 1974. [4] L. P. Holmblad and J. J. Ostergaard, Control of a cement kiln by fuzzy logic, in Fuzzy Information and Decision Processes, M. M. Gupta and E. Sanchez, Eds. North-Holland, 1982. [SI R. M. Tong, The evaluation of fuzzy models derived from experimental data, Fuzzy Sets and Systems, vol. 4, pp.1-12, 1980. [6] W. Pedrycz, An identification algorithm in fuzzy relational systems, Fuzzy Sets and Systems, vol. 13, pp.153-167, 1984. [7] C. W. Xu and 2. Yong, Fuzzy model identification and self-learning for dynamic systems, IEEE Truns. Syst., Man. Cybern., vol. 17, no. 4, pp. 683-689, 1987.
[SI D. Filev, Fuzzy modeling of complex systems, Int. J. Approximate Reasoning, vol. 5, pp. 281-290, 1991. [9] D. Willaeys, N. Malvache, and P. Hammond, Utilization of fuzzy sets for systems modeling and control, in Proc. 16th IEEE Con$ Decision and Control (New Orleans, LA), Dec. 1975. [ I O ] E. Czogala and W. Pedrycz, On identification in fuzzy systems and its applications in control problem, Fuzzy Sets and Systems, vol. 6, pp. 73-83, 1981. [I I] T. Takagi and M. Sugeno, Fuzzy identitication of systems and its application to modeling and control, IEEE Trans. Syst., Man, Cybern., vol. 15, pp. 116132, 1985. [I21 M. Sugeno and G. T. Kang, Fuzzy modeling and control of multilayer incinerator, Fuzzy Sets and Systems, vol. 18, pp. 329-346, 1986. [ 131 K. Tanaka and M. Sugeno, Stability of analysis of fuzzy systems using Lyapunovs direct method and construction procedure for Lyapunov functions, in Proc. 6th Fuzzy Systems Symposium (Tokyo), 1990, pp. 353-356. [ 141 M. Sugeno and T. Takagi, A new approach to design of fuzzy controller, in Advcinces in Fuzzy Sets. Possibility Theory, and Applications,
31
P. P. Wang and S. K. Chang, Eds. New York Plenum Press, 1983, pp. 325-334. J. De kleer, Qualitative and quantitative knowledge in classical mechanics, Artificial Intelligence Laboratory, TR-352, MIT, Cambridge, MA, 1975. Hayes, P. J., The naive physics manifesto, in Expert in the MicroEdinburgh: Edinburgh University electronic Age, D. Michie, Ed. Press, 1979. J. De Kleer and J. S. Brown, A qualitative physics based on confluences, Artijicial Intelligence, vol. 24, pp. 7-83, 1984. K. D. Forbus, Qualitative process theory, Artijicial Intelligence, vol. 24 pp. 85-168, 1984. b. Kuipers, Qualitative simulation, Artificial Intelligence, vol. 29, pp. 289-338, 1986. F. Wentop, Deductive verbal models of organizations, Int. J. ManMachine Srudies, vol. 8, pp. 293-31 I , 1976. W. J. M. Kickert, An example of linguistic modeling: the case of Mulders theory of power, in Advances in Fuzzy Ser Theory and Applications, M. M. Gupta, R. K. Ragade, and R. R, Yager, Eds. North-Holland, 1979, pp. 5 19-540. W. J. M. Kickert, Towards an analysis of linguistic modeling, Fuzzy Sers and Systems, vol. 2, pp. 293-307, 1979. B. DAmbrosio, Qualitative Process Theory Using Linguistic Variables. New York: Springer-Verlag, 1989. E. H. Mamdani, Application of fuzzy logic to approximate reasoning using linguistic synthesis, IEEE Trans. Comput. vol. 26, pp. 1182-1 191, 1977. T. J. Procyk and E. H. Mamdani, A linguistic self-organizing process controller, Auromatica, vol. 15, pp. 15-30, 1979. L. A. Zadeh, From circuit theory to system theory, in Facers of Systems Science, G. J. Klir, Ed. New York: Plenum Press, 1991. H. Akaike, A new look at the statical model identification, IEEE Trans. Auromar. Contr., vol. 19 pp. 716-723, 1974. J. Ihara, Group method of data handling towards a modeling of complex systems-IV, Sustems and Conrrol (in Japanese), vol. 24, pp. 158-168, 1980. J. C. Dunn, Well-separated cluster and optimal fuzzy partition, J. Cybern., vol. 4, pp. 95-104, 1974. J. C. Bezdek, Pattern Recognition wirh Fuzzy Objective Function Algorithm. New York: Plenum Press, 1981. Y. Fukuyama and M. Sugeno, A new method of choosing the number of clusters for fuzzy c-means method, in Proc. 5th Fuzzy System Symposium (in Japanese), 1989, pp. 247-250. F. Eshragh and E. H. Mamdani, A general approach to linguistic approximation, Inr. J. Man-Machine Studies, vol. 11, pp. 501-519, 1979. A. Ralescu and H. Narazaki, Integrating artifical intelligence tech-
[34] [35] [36] [37]
niques in linguistic modeling from numerical data, in Proc. Int. Fuzzy Engineering Symp. (Yokohama), 1991. D. Dubois and H. Prade, Fuzzy Sets and Systems-Theory and ApplicaNew York: Academic Press, 1980. tions. L. A. Zadeh, A fuzzy-set-theoretic interpretation of linguistic hedges, J. Cybern., vol. 2, pp. 4-34, 1972. G. E. P. Box and G. M. Jenkins, Time Series Analysis, Forecasring and Control. San Francisco: Holden Day, 1970. K. Tanaka, An approach to modeling, analysis and design of fuzzy control systems, Ph.D. thesis, Tokyo Institute of Technology, 1990.
Michio Sugeno was born in Yokohama, Japan. He received the B.S. degree in physics from the University of Tokyo in 1962. From 1962 to 1965 he worked in nuclear engineering for Mitsubishi Atomic Power Industry. From 1965 to 1976 he was a Research Associate in control engineering at the Tokyo Institute of Technology, Tokyo. He received the doctor of engineering degree from the Tokyo Institute of Technology in 1975. In 1976 he joined the Graduate School of the Tokyo Institute of Technology, Yokohama, as an Associate Professor. He is currently a Professor in systems science. Since 1989 he has been a Leading Advisor at the Laboratory for International Fuzzy Engineering Research, Yokohama. Since 1991 he has been the President of the Japan Society for Fuzzy Theory and Systems. His research interests include fuzzy computing based on systemic functional grammar and fuzzy logic, fuzzy control of helicopters, and fuzzy measure analysis.
Takahim Yasukawa was born in Iizuka, Japan, in 1960. He graduated in aeronautical engineering from the National Defense Academy, Yokosuka, Japan, in 1983, and received the M.S. degree in control engineering from the Tokyo Institute of Technology, Tokyo, Japan in 1991. Since 1983 he has been an officer of the Japan Air Self-Defense Force and since 1989 a graduate student in control engineering at the Tokyo Institute of Technology. His research interests include fuzzy control, modeling, and linguistic modeling.

A Fuzzy-Logic-Based Approach Qualitative Modeling: Abstract

Uploaded by

Document Informationclick to expand document information

Copyright:

Available Formats

A Fuzzy-Logic-Based Approach Qualitative Modeling: Abstract

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Fuzzy-Logic-Based Approach Qualitative Modeling: Abstract

Uploaded by

Copyright:

Available Formats

lEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. I, NO. I .

A Fuzzy-Logic-Based Approach to Qualitative Modeling

1063-6706/93$03 00 G 1993 IEEE

IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. I , NO. I , FEBRUARY 1993

Condition BZ low Ox low

BZ IOW Ox low BE O.K. BZ low Ox low BE high

Fig. 1. Operators manual of cement kiln control.

SUGENO AND YASUKAWA: FUZZY-LOGIC-BASED APPROACH TO QUALITATIVE MODELING

where n, is a fuzzy number, approximately U, associated with a membership function, e.g.,

IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. I , NO. I , FEBRUARY 1993

SUGENO AND YASUKAWA: FUZZY-LOGIC-BASED APPROACH TO QUALITATIVE MODELING

TABLE I CLASSIFICATION OF IDENTII-ICATIOV Stucture Identification I a: Input candidates

Stucture Identification I1 Parameter identification

b: Partition of input space

and also consider an input-output system:

IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. I , NO. I. FEBRUARY 1993

Partition of the input space.

Fig. 2 . Input+utput data and model

R' : if :c1 is Ai and :1:2 is A i . . . and xTtis Ai, then y is B i . (9)

SUGENO AND YASUKAWA: FUZZY-LOGIC-BASED APPROACH TO QUALITATIVE MODELING

2) Then defuzzify B z in the consequents by taking the center of gravity:

z IS A' then y is b'.

IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. I , NO. I , FEBRUARY 1993

TABLE I1 INPUT-OUTPUT DATA NONLINEAR OF SYSTEM Group A No.

.r2 3.00 3.02

Fig. 5. Inputboutput relation of nonlinear system.

SUGENO AND YASUKAWA: FUZZY-LOGIC-BASED APPROACH TO QUALITATIVE MODELING

TABLE Ill STRUCTURE IDENTIFICATION NONLINEAR Ib OF SYSTEM Input Variables

.rl - 1.3 .rl - . x i

IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. I , NO. I , FEBRUARY 1993

Fig. 7. Ordinary fuzzy partition of input space

Fig. 8. Fuzzy cluster in the input space.

Fig. 9. Projection of a fuzzy cluster. Fig. 1 I .

R1 : if : c 1 is A: and z 2 is Ai then ;y is B R 2 : if :rl is A: and L I : ~is Ai then y is B .

SUGENO AND YASUKAWA: FUZZY-LOGIC-BASED APPROACH TO QUALITATIVE MODELING

Fig. 12. Behavior of S ( c ) in nonlinear system model.

Fig. 13. Model of the nonlinear sustem.

+ dy/dzl(zl - z:) + . . . + dy/dz,(z,,

IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. I, NO. I , FEBRUARY 1993

Find the Darameters 6

Assume single variable among input candidates

Fig. 15. Trapezoidal fuzzy set.

Calculate RC for each input variable I

Combine two groups of data to reform the parameters A

Estimate b y / b x and find the parameters C for a positionoradient tvDe model

Make parameters identificationI

Fig. 14. Algorithm for identification.

MODEL 111. QUALITATIVE

SUGENO AND YASUKAWA: FUZZY-LOGIC-BASED APPROACH TO QUALITATIVE MODELING

Fig. 17. Fuzzy set

and linguistic approximations P and

(2) more or less A

(6) more than A

(7) less than A

B. Control of Number of Linguistic Rules

IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 1 , NO. 1, FEBRUARY 1993

RESULTS LINGUISTIC OF APPROXIMATION

Linguistic Expression calm