ABSTRACT
This paper describes an algorithm for extracting knowledge, in the form of fuzzy rules, from a self-organizing supervised learning neural network called fuzzy ARTMAP. Rule extraction proceeds in two stages: pruning removes those recognition nodes whose confidence index falls below a selected threshold; and quantization of continuous learned weights allows the final system state to be translated into a usable set of rules. Using a molecular biology problem of recognizing DNA subsequences, comparisons are drawn between fuzzy matching rules of ARTMAP networks and Craven and Shavlik’s NofM rules extracted from backpropagation networks. Preliminary results indicate that while the predictive performance of both systems is comparable, there are tradeoffs to be made between learning speed and system size.
