Balancing Specificity and Generality in a Panmictic-Based Rule-Discovery Learning Classifier System

摘要

A Learning Classifier System has been developed based on industrial experience. Termed iLCS, the methods were designed and selected to function with common data properties found in industry. Interestingly, it considers a different strategy to XCS type systems, with the rule discovery being based pan-mictically. In order to show the worth of the iLCS approach, the benchmark data-mining application of the Wisconsin Breast Cancer dataset was investigated. A competitive level of 95.3% performance was achieved; mainly due to the introduction of a generalisation pressure through a fitness to mate (termed fertility) that was decoupled from a fitness to effect (termed effectiveness). Despite no subsumption deletion being employed the real-valued rule-base was simple to understand, discovering similar patterns in the data to XCS. Much further testing of iLCS is required to confirm robustness and performance. Currently, the iLCS approach represents a flexible alternative to niche-based LCSs, which should further the advancement of the LCS field for industrial application.