Contents 1. INTRODUCTION
FNN systems are combinations of the theories of fuzzy logic and neural networks. In the system, the parameters of the fuzzy system are determined by means of the learning algorithms used in neural networks. Usually, FNNs are hybrid systems in which the fuzzy techniques are actually used to create or enhance neural networks. Such integration renders FNN systems more powerful than either one alone. Recently, several FNNs adopt backpropagation (BP) learning algorithm to adjust parameters [1]. Using the steepest descent technique in BP training could minimize the error function, allowing the algorithm to reach the local minimum very fast while never finding a global solution. The reason for local minimum of BP algorithm results from several factors, including error cost function characteristics, initial values, choice of learning parameters, network topology [2]. On the other hand, another central problem in the applications of FNN is how fuzzy rules can be inferred from observations, i.e., from the knowledge of a collection of input-output space [3]. Specially, if the number of fuzzy rules is given by prior expert knowledge and keeps unchanged during learning, it often brings a deficiency or redundancy of fuzzy rules. Therefore, how to build ideal fuzzy rules and rule numbers of FNN systems from observations is a very important issue.
