An Automated Rule Refinement System

Artificial neural networks (ANNs) are essentially a ‘black box’ technology. The lack of an explanation component prevents the full and complete exploitation of this form of machine learning. This chapter presents an historical perspective on rule extraction from artificial neural networks beginning with 6 ways in which the ANN paradigm may be enriched through the addition of rule extraction / explanation facility. The chapter then describes the general rule extraction task and introduces the Andrews-Diederich-Tickle (ADT) taxonomy for classifying rule extraction techniques. The use of the taxonomy is illustrated by the description of several representative rule extraction algorithms with particular attention being paid to previously described techniques for extracting rules from local function networks. The chapter then describes the concept of rule / theory refinement and includes a discussion of several previously described connectionist approaches to rule / theory refinement. 1.0 Introduction Artificial Neural Networks (ANNs) have qualities that make them attractive to users, viz, i) the ability of the network to ‘learn’ from data through dynamic adaptation of internal numeric weights (as opposed to the tedious ‘knowledge engineering’ phase required to construct a rule based system); ii) the compact (albeit completely numerical) form in which the acquired information/`knowledge' is stored within the trained ANN; iii) the comparative ease and speed with which this `knowledge' can be accessed and used; iv) the robustness of an ANN solution in the presence of `noise' in the input data; and v) the high degree of accuracy reported when an ANN solution is used to generalise over a set of previously unseen examples from the problem domain.