On Stability of Formal l~"~ness Systems ALEKSANDER A. KANIA
Institute of Applied Mechanica, Mathematica Dioiaion, Technical Unit~raiOpof laelce, AL Ty~'~lecia Pa~twa Polakiego 7, 25-314 Kielce~ Poland and JERZY B. KISZKA, MARIAN B. G and MARIAN S. STACHOWICZ
O
~
,
JAROSLAW It. MAJ,
Institute of Automatic Contro~ Technical Unioer$iO~of lOeic~ AL Tys~lecia P~mstwa Polakiego 7, 25-314 Kielc~ Poland Communicated by L. A. 7adeh
ABSTRACT The structure of formal fslT'~inesssystems as some abstraction on real f-~,iness systems is debated. Then the global definition of a-stability is given. First some necessary and sufficient conditions for a-stability are given for the case of R .. A ~ B (Sec. 2) and later for the case of R defined by an implication chain of any finite length (Sec. 4). In Secs, 5 and 6 notions of a-stability and a-fl strong decision stability are introduced, and some theorems on necessary or sufficient (or both) conditions for these kinds of stability, imposed on the structure of an impfication chain, are proved. In Sec. 3 the good-mapping property of a fuzzy relation matrix is a n M y ' / ~ and because of its heuristic importance it is assumed in further ,.,ections. In many cases examples are given to illustrate definitions and theorems or conclusions.
At . . . . . A, defined on X a n d a family of fuzzy sets Bi . . . . . B, defined on Y, one can construct a chain of implications (AliBi) or ... or (A,~B,). By definition, the length of this implication chain is s > 1. Then in the preceding case, R=(A~B), R was defined by an implication chain of length 1. In the latter, the chain of length s can also form a relation R. In what follows, when we say "relation," we mean a chain of implications of length s, s • 1. A fuzzy relation R is a form of mapping with its d o m a i n in ~ , a family of all fuzzy sets defined on X, and a space of values contained in ~ , a family of all fuzzy sets defined on Y. This mapping can be denoted in terms of the compositional rule of inference (see Z a d e h [1]):
UoR=V, where U is any fuzzy set defined on X, and V any fuzzy set defined on Y. This mapping can be treated as a model of a real transformation system which consists of a fuzzy input, black box and fuzzy output:
Fuzzy input--. ~
--,fuzzy output.
W e assume that this real transformation system can be m a p p e d on to the formal model:
u--,®--)v
(,)
or equivalently,
UoR=V, where U and V are also fuzzy input and output, respectively, but they are formally well defined. Similarly R is not a mere idea of fuzzy transformation, but a real matrix, suitably defined. U a n d V are abstractions on the real fuzzy input a n d output, a n d R is an abstraction on a black box, a natural transforming system. Then the formula U o R = V, until you hear further, will be called a "formal fuzziness system." In the present p a p e r we are not interested in studying the adequacy of relation R with respect to a real transformation system. But when such adequacy occurs, some problems of great importance in real transformation systems are converted to corresponding problems in their formal models. One of them, the problem of stability, is a subject of study in the present paper.
ON STABILITY OF FORMAL FUZZINESS SYSTEMS
53
As far as we know, stability has been considered only in [2,3,4,5]. As Negoita [4] maintains, a system has the stability property if the state trajectory can be held inside some given domain for some nonempty set of inputs. This is consistent with the classical definition of stability, but his insight into the problem merits for further attention, especially in view of its possible applications. Mamdanl [3] strongly stresses the importance of stability problem for fuzzy system both in practice and theoretical debate. But in his opinion, this problem has not been satisfactorily stated, and therefore has not been solved yet. The main reason, in his view, lies in the absence of a developed theory for the design of fuzzy systems--in particular, in the absence of good theoretical tools for studying fuzzy-system stability. Tong [5] defines a fuzzy system by means of fuzzy sets of finite dimension. Stability of fuzzy system is defined in terms of an equivalence relation in the space of fuzzy sets. This relation is generated by the peak-pattern function. This attempt is applied to closed-loop fuzzy systems and arises in connection with the controLlability of a system. It is of great practical importance, but the application of the peak-pattern criterion restricts studies on other aspects of stability of fuzzy systems. Kickert and Mandani [2], recalling the classical approach, treat the fuzzy system as multilevel relay. Applying the described function method, the relay's stability can be investigated; thus investigation of the stability of a fuzzy system, under some assumptions, is made possible. These assumptions, however, seem to be very restrictive for practical purposes. This short review suggests that research on stability is caLled for, but up to now there have been but a few commonly accepted definitions and criteria for those fuzzy-system properties which can be compared with stability in the nonfuzzy case. The present paper is an attempt to jump over the gap dividing fuzzy and nonfuzzy analysis of transformation systems. 2.
a-STABILITY, T H E CASE OF SIMPLE IMPLICATION
At the be#nning let us consider a formal fuTziness system in a special form. That is, let U, V be nonfuzzy sets, and R a nonfuzzy mapping R : ~ ( X ) - - , (Y). Let Y D F(Y) be the set of feasible points in Y. Then we consider the system (*) as having the stability property with respect to some subfamily P ( X ) C ~ (X) if for each U ~ P ( X ) , U o R N [ Y - F( Y)] = 0 . In other words, the system has the stability property with respect to some class P ( X ) if for each input X E P(X), X o R does not contain any non.feasible point. This definition of stability in nonfuzzy systems can be generalized to include also the fu77iness case,
DEFINITION 1. The formal fuzziness system has the a-stability (a-S) property with respect to some family t~ c ~(~ and some feasible subset F(Y) C Y if
54
A L E K S A N D E R K A N I A ET AL.
~.0.
or4
I% '
~tY)
4,
~r Y- F(Y~
Fig. I for every C D U fuzzy set, U o R = V is such that t t v ( y ) < a for every y E Y F(Y). E X A M P L E 1. F o r some input C ~ U, let the output set V have the m e m b e r ship function shown in Fig. 1. W e see that in this case the system has not the a-S property with respect to the family ~ a n d given F(Y), because for some U E C, U o R = V is such that there is a point, say Y0 E Y - F ( Y ) , such that #v(Y0)>a. Of course, in this example a~-stability (a < a t ) is not excluded; however, it can h a p p e n that there is another set W E A such that m a x y _ ~ r ) l z w . R ( y ) > a I. This would imply that the system ( * ) is not a i-S with respect to the given family A a n d feasible subset F ( Y ) . In the case of the nonfuzzy system ( * ) it is obvious that its stability implies a-stability for any a E [0, 1]. 2.1. SUFFICIENT AND NECESSARY CONDITIONS FOR a-STABILITY OF A SYSTEM (*) Let the fuzzy relation matrix R, R = A ~ B = A × B, be given. Elements of this matrix can be written as follows: r#=l~A(xi)AuB(yj), i = 1 . . . . . n a n d j = 1 . . . . . m. Then any output V, associated with an input t) can be characterized as follows: #v(Yj) -- VT-I tZu(Xi)Aro" -~ V~ l I ~u(Xi) AU~(Xl)Au~(Yj) ~ ~B(Yj) A k/7.11~v(xi)Aua(xj), j = l . . . . . m. It can readily be seen that uv(Yj)
#B(Yj) A V # u ( x i ) A u A ( x i ) < a . i--I
(1)
ON STABILITY O F F O R M A L F U Z Z I N E S S SYSTEMS