On Zadeh’s Compositional Rule of Inference ∗ Robert Fullér rfuller@abo.fi Hans-Jürgen Zimmermann zi@buggi.or.rwth-aachen.de Abstract This paper deals with Zadeh’s compositional rule of inference [7] under triangular norms: IF X is P AND X and Y have relation R, THEN Y is Q, where P and Q are fuzzy sets of the real line IR, R is a fuzzy relation on IR and the conclusion Q is defined via sup-T composition of P and R: πQ (y) = sup T (πP (x), πR (x, y)), y ∈ IR. x∈IR It is shown that (i) when the triangular norm T and the membership function of the observation P are continuous, then the conclusion Q depends continuously on the observation; (ii) when T and the membership function of the relation R are continuous, then the observation Q has continuous membership function. Furthermore, we present a similar result for the discrete case 1 Preliminaries A fuzzy set A with membership function πA : IR → [0, 1] = I is called fuzzy number if A is normal, continuous,fuzzy convex and compactly supported. The fuzzy numbers will represent the continuous possibility distributions for fuzzy terms. Let A be a fuzzy number, then for any θ ≥ 0 we define ωA (θ), the modulus of continuity of A by ωA (θ) = max |πA (u) − πA (v)|. |u−v|≤θ An α-level set of a fuzzy interval A is a non-fuzzy set denoted by [A]α and is defined by [A]α = {t ∈ IR | πA (t) ≥ α} for α ∈ (0, 1] and [A]α = cl(suppπA ) for α = 0. We metricize F by the metric [4], D(A, B) = sup d([A]α , [B]α ), α∈[0,1] ∗ The final version of this paper appeared in: R.Lowen and M.Roubens eds., Proceedings of the Fourth IFSA Congress, Vol. Artifical intelligence, Brussels, 1991 41-44. 1 where d denotes the classical Hausdorff metric in the family of compact subsets of IR2 , i.e. d([A]α , [B]α ) = max{|a1 (α) − b1 (α)|, |a2 (α) − b2 (α)|}, and [A]α = [a1 (α), a2 (α)], [B]α = [b1 (α), b2 (α)]. When the fuzzy sets A and B have finite support {x1 , . . . , xn }, then their Hamming distance is defined as H(A, B) = n X |πA (xi ) − πB (xi )|. i=1 In the sequel we need the following lemma. Lemma 1 [5] Let δ ≥ 0 be a real number and let A, B be fuzzy intervals. If D(A, B) ≤ δ, then sup |πA (t) − πB (t)| ≤ max{ωA (δ), ωB (δ)}. t∈IR 2 Stability and continuity properties of the compositional rule of inference Consider the compositional rule of inference with different observations P and P’: Observation: Relation: Conclusion: X has property P X and Y have relation R Y has property Q X has property P 0 X and Y have relation R Y has property Q0 According to the compositional rule of inference the membership functions of the conclusions are computed as πQ (y) = sup T (πP (x), πR (x, y)), πQ0 (y) = sup T (πP 0 (x), πR (x, y)). x∈IR (1) x∈IR The following theorem shows that when the observations are closed to each other in the metric D, then there can be only a small deviation in the membership functions of the conclusions. Theorem 1 (Stability theorem) Let δ ≥ 0 and T be a continuous triangular norm, and let P , P 0 be fuzzy intervals. If D(P, P 0 ) ≤ δ then sup |πQ (y) − πQ0 (y)| ≤ ωT (max{ωP (δ), ωP 0 (δ)}). y∈IR 2 Proof. Let y ∈ IR be arbitrarily fixed. From Lemma 2.1. it follows that |πQ (y) − πQ0 (y)| = | sup T (πP (x), πR (x, y)) − sup T (πP (x), πR (x, y))| ≤ x∈IR x∈IR sup |T (πP (x), πR (x, y)) − T (πP 0 (x), πR (x, y))| ≤ sup ωT (|πP (x) − πP 0 (x)|) ≤ x∈IR x∈IR sup ωT (max{ωP (δ), ωP 0 (δ)}) = ωT (max{ωP (δ), ωP 0 (δ)}). x∈IR Which proves the theorem. Theorem 2 (Continuity theorem) Let R be continuous fuzzy relation, and let T be a continuous t-norm. Then Q is continuous and ωQ (δ) ≤ ωT (ωR (δ)) for each δ ≥ 0. Proof. Let δ ≥ 0 be a real number and let u, v ∈ IR such that |u − v| ≤ δ. Then |πQ (u) − πQ (v)| = | sup T (πP (x), πR (x, u)) − sup T (πP (x), πR (x, v))| ≤ x∈IR x∈IR sup |T (πP (x), πR (x, u)) − T (πP (x), πR (x, v))| ≤ sup ωT (|πR (x, u) − πR (x, v)|) ≤ x∈IR x∈IR sup ωT (ωR (|u − v|)) = ωT (ωR (|u − v|)) ≤ ωT (ωR (δ)). x∈IR Which ends the proof. Remark 1 From lim ω(δ) = 0 δ→0 and Theorem 1 it follows that sup |πQ (x) − πQ0 (x)| → 0 as D(Q, Q0 ) → 0 x∈IR which means the stability of the conclusion under small changes of the observation. Other results along this line have appeared in [1, 2, 3, 7, 8]. Consider now the case when P and R are finite. The following theorem shows that the stability property of the conclusion remains valid in the discret case. Theorem 3 Let T be a continuous t-norm. If the observation P and the relation matrix R are finite, then H(Q, Q0 ) ≤ ωT (H(P, P 0 )) (2) where H denotes the Hamming distance and the conclusions Q and Q0 are computed by (1). The proof of this theorem is carried out analogously to the proof of Theorem 1. It should be noted that in the case of T (u, v) = min{u, v} (2) yields H(Q, Q0 ) ≤ H(P, P 0 ). 3 References [1] D.Dubois and H. Prade, On distance between fuzzy points and their use for plausible reasoning, in: Proc. of IEEE Conf. on Cybernetics and Systems, Bombay-New Delhi, December 30, 1983- January 7, 1984 300-303. [2] H.Hellendoorn, Closure properties of the compositional rule of inference, Fuzzy Sets and Systems, 35(1990) 163-183. [3] H.Hellendoorn, The generalized Modus Ponens considered as a fuzzy relation, Fuzzy Sets and Systems, (to appear). [4] O.Kaleva, Fuzzy differential equations, Fuzzy Sets and Systems, 24(1987) 301317. [5] M.Fedrizzi and R.Fullér, Stability in Possibilistic Linear Programming Problems with Continuous Fuzzy Number Parameters, Fuzzy Sets and Systems, 47(1992) 187-191. [6] M. Mizumoto and H.-J. Zimmermann, Comparison of fuzzy reasoning methods, Fuzzy Sets and Systems, 8(1982) 253-283. [7] L.A.Zadeh, Outline of a new approach to the analysis of complex systems and decision processes, IEEE Transactions on Systems, Man and Cybernetics, Vol.SMC-3, No.1, 1973 28-44. [8] H.-J.Zimmermann, Fuzzy set theory - and inference mechanism, in: Mathematical Models for Decision Support Systems, NATO ASI Series, Vol. F48, 1988 727-741. 4