Hybrid Learning Methods

Abstract

In Chapters 4 and 5 the connectionist, multi-layer architectures of fuzzy systems, called fuzzy inference neural networks, were presented. These architectures are similar to neural networks (see Section 3.1), so learning algorithms can be proposed to tune the parameters of the networks, analo gously to tuning weights in neural networks. The parameters of the neuro fuzzy architectures define the shape of membership functions of the fuzzy sets in the IF-THEN rules. Tuning these parameters thus optimizes the form of the rules. Moreover, the number of rules in the rule base of the fuzzy systems can be determined using a learning method. The number of elements in the first layers of the neuro-fuzzy architectures depends on the number of the rules, so this kind of algorithms determines the architectures. Hybrid learning, which is the subject of this chapter, consists of a combina tion of different learning methods, such as gradient, genetic, and clustering algorithms. These methods are first described, and then the hybrid algo rithms for rule generation and parameter tuning are presented, including the algorithms proposed in [479], [438], [439], [440], [480], [481].