Extracting boolean and probabilistic rules from trained neural networks

Abstract

This paper presents two approaches to extracting rules from a trained neural network consisting of linear threshold functions. The first one leads to an algorithm that extracts rules in the form of Boolean functions. Compared with an existing one, this algorithm outputs much more concise rules if the threshold functions correspond to 1-decision lists, majority functions, or certain combinations of these. The second one extracts probabilistic rules representing relations between some of the input variables and the output using a dynamic programming algorithm. The algorithm runs in pseudo-polynomial time if each hidden layer has a constant number of neurons. We demonstrate the effectiveness of these two approaches by computational experiments.