New approach for extracting knowledge from the XCS learning classifier system

Abstract

In real-domain problems, having generated a complete map for a given problem, a Learning Classifier System needs further steps to extract minimal and representative rules from the original generated ruleset. In an attempt to understand the generated rules and their complex underlying knowledge, a new rule-driven approach is introduced which utilizes a quality-based clustering technique to generate clusters of rules. Two main outputs are extracted from each cluster: (1) an aggregate average rule which represents the common features of the group of rules, and (2) an aggregate definite rule which presents the common characteristics within the cluster. Initial experimental results show that these extracted patterns are able to classify future domain cases efficiently.