Rule extraction by successive regularization.

Abstract

Knowledge acquisition is, needless to say, important, because it is a key to the solution to one of the bottlenecks in artificial intelligence. Recently, knowledge acquisition using neural networks, called rule extraction, is attracting wide attention because of its computational simplicity and ability to generalize. Proposed in this paper is a novel approach to rule extraction named successive regularization. It generates a small number of dominant rules at an earlier stage and less dominant rules or exceptions at later stages. It has various advantages such as robustness of computation, better understanding, and similarity to child development. It is applied to the classification of mushrooms, the recognition of promoters in DNA sequences and the classification of irises. Empirical results indicate superior performance of rule extraction in terms of the number and the size of rules for explaining data.