Pattern Recognition with Fuzzy Competitive Learning

摘要

A Fuzzy Competitive Learning (FCL) model is proposed in this paper. Unlike the crisp competitive learning algorithms in which one and only one neurode who has minimum distance will win and adjust its weights at each competition, Fuzzy Competitive Learning (FCL) allows every neurode in the network to win to a certain degree and adjust its weights to some fuzzy extent, depending on its distance to the input pattern. In other words, every neurode learns depending on how well it wins. We adaptively update weights according to sample data. For each unknown input pattern, the system applies a membership function to measure a set of feature similarities to the existing exemplars. Then the system maps this array of membership function values to a character similarity measure and so classifies the input pattern. If the input pattern is not found similar to any existing exemplars, the system creates a new class. Fuzzy Competitive Learning is very useful in pattern classification. It performs much better and converges more often especially for overlapping classes.