The modifiability of rule-based expert systems

Abstract

This thesis examines the claim that rule representations of knowledge are conveniently modified. The thesis falls into two parts and in the first, a precise notion of a convenient modification, called an extension, is developed and it is shown that extensible knowledge-bases are very convenient to modify and develop. We show that- rule representations of knowledge are extensible only if they incorporate a suitable organisation of knowledge. Furthermore, we show that non-rule representations with a suitable organisation of knowledge are also extensible. We therefore conclude that rule representations of knowledge are no more or less extensible than non-rule representations. In the second part, we consider the more pragmatic aspects of knowledge-base modifiability. In each of two detailed case studies, we compare the modifiability of a rule-based expert system with its "second generation” counterpart which incorporates non-rule representations of knowledge. We conclude that in practice the modifiability of extensible knowledge-bases can be compromised if the organisation of knowledge is represented obscurely. Above all, the thesis emphasizes the importance of the organisation of knowledge in an expert system.