Abstract
On an automatic rule extraction from set of input-output data examples, decision tree generating methods such as ID3 proposed by Quinlan (1986) are major. ID3 is, however, applicable only to the case that both input and output data are discrete or symbolic. As a method extended to case that input data are numerical and extracting fuzzy rules, Fuzzy ID3 has been proposed by Umano (1993) and Sakurai (1993). These crisp or fuzzy decision tree generating methods, however, need to recreate the trees from the beginning as often as a tendency of learning examples and are hard to be applied in such case that a tendency of examples changes changes dynamically during the inference process is in progress. As a method to overcome above two shortcomings and extract fuzzy rules adaptively, clustering input-output data space with neural networks, especially, Kohonen's Self-Organizing Map (SOM) is considerable.In this paper, we propose a neural network that has the architecture of SOM and the function of fuzzy clustering, called "Fuzzy Self-Organizing Map (FSOM)", and the learning methods based on Competitive Learning.And we propose a neural network that learns tendency of examplese, represents results of learning as fuzzy rules, and does fuzzy inference with the architecture of FSOM, called "Fuzzy Inference Network (FIN)", and a method of an automatic and adaptive rule extraction with the neural network. Furthermore, we present results of simulations for the comparison with other methods of automatic and adaptive rule extraction.
