Identification algorithms for fuzzy relational matrices - Part 2. Optimizing algorithms

This paper, Part 2 of a two part series, reviews and evaluates four (4) alg orithms that identify fuzzy relational matrices by optimizing a user-specif ied performance index [6,8,29,34]. The performance of the Recursive Paramet er method [34] was unsatisfactory but the Probabilistic Descent [6], Neural Learning [29] and Quasi-Newton [8] methods all gave comparable results tha t, in general, were better than the non-optimizing algorithms [3,9,23,32] r eviewed in Part 1 [1]. However, the tuning and iteration required for these optimizing algorithms makes them less desirable for most on-line applicati ons than the non-optimizing techniques of Shaw et al. [23] and Pedrycz [9]. It was also noted;that results expressed in terms of fuzzy indices, Q(q), were very poorly correlated (in fact tended to an inverse correlation) with results based on the non-fuzzy discrete indices. J(q) preferred in many pr actical applications. (C) 2000 Elsevier Science B.V. All rights reserved.