Reduction of symbolic rules from artificial neural networks using sensitivity analysis

This paper shows how sensitivity analysis identifies and eliminates redundant conditions from the rules extracted from trained neural networks, by eliminating irrelevant inputs. This leads to a reduction in the number and size of the rules. The reduced rule set accurately and minimally reflect the classification problems presented. Also, the elimination of redundant input units significantly reduces the combinatorics of the rule extraction algorithm. The resultant rule set compares favorably with traditional symbolic machine learning algorithms