Inducing and refining rule-based knowledge from inexact examples

The knowledge acquisition bottle-neck (Feigenbaum, 1981) greatly obstructs the development of expert systems. This paper describes AKARS-1, a domain independent Automatic Knowledge Acquisition and Refinement System which can automatically induce and refine knowledge in rule form (exact and approximate) from exact and inexact examples. Its major components, AKA-2 (Automatic Knowledge Acquisition system) and HERES (HEuristic REfinement System), are detailed. AKA-2 employs a new discriminatory coefficient to evaluate discriminatory ability of each attribute/value pair and a novel method for calculating certainty factors of rules. HERES collects performance statistics of every rule and calculates the overall adequacy of the initial knowledge base, and then employs heuristics and the information collected to determine how to refine the knowledge base. The functionality and effectiveness of AKARS-1 are verified through various case studies. It has been verified that AKARS-1 can successfully induce and refine knowledge bases. The learning and refinement methods can handle imprecise and uncertain examples and generate approximate rules. In this aspect, they are better than other famous learning algorithms like ID3 (Quinlan, 1983), AQ11 and INDUCE (Michalski, 1973, Michalski, 1980, Michalski, 1983). AKARS-1's methods are currently unique in processing inexact examples and creating approximate rules.