A NEW APPROACH TO RULE LEARNING BASED ON FUSION OF FUZZY-LOGIC AND NEURAL NETWORKS

A new method is developed to generate fuzzy rules from numerical data. This method consists of two algorithms: Algorithm 1 is used to identi fy structures of the given data set, that is, the optimal number of ru les of system; Algorithm 2 is used to identify parameter of the used m odel. The former is belonged to unsupervised learning, and the latter is belonged to supervised learning. To identify parameters of fuzzy mo del, we developed a neural network which is referred to as Unsymmetric al Gaussian Function Network (UGFN). Unlike traditional fuzzy modellin g methods, in the present method, a) the optimal number of rules (clus ters) is determined by input-output data pairs rather than by only out put data as in sugeno's method, b) parameter identification of the pre sent model is based on a like-RBF network rather than backpropagation algorithm. Our method is simple and effective because it integrates fu zzy logic with neural networks from basic network principles to neural architecture, thereby establishing an unifying framework for differen t fuzzy modelling methods such as one with cluster analysis or neural networks and so on.