Pedagogical method for extraction of symbolic knowledge

Abstract

This paper addresses the extraction of symbolic knowledge from trained artificial neural networks. Specifically, for that purpose the so-called pedagogical approach is incorporated, where the trained network is used as an oracle when inducing the symbolic description. We present an essential extension of the TREPAN algorithm proposed originally by Craven and Shavlik [4][5]. The crucial modification concerns the way of generating artificial training instances. The paper ends with an empirical verification of the proposed method on popular machine learning benchmarks and comparison with the original TREPAN.