On-the-Fly Knowledge Base Refinement by a Classifier System

Abstract

A Classifier System (CS) is a machine learning system composed of a production system, a reinforcement learning mechanism, and a rule generation function by genetic algorithms (GAs). This paper presents a new method for knowledge base refinement by CS technique and describes its application to rule-based simulation for an automated transportation system in a steel manufacturing process.

The key idea of the proposed method is that the condition part of a rule should be divided into two parts: indispensable conditions and discriminate conditions. The former are generated by a diagnosing type knowledge-based system. The latter and action parts are generated and refined by genetic algorithms. Using this method, we can easily input initial rule sets, refine the rules without generating unapplicable ones, and reduce the computation time for learning. The method enables us to develop an on-the-fly knowledge refinement mechanism for rulebased simulation systems.

Intensive experiments on the transportation system have shown that 1) the generated rules prevent blocking of indispensable events from occurring, and 2) the rules also generate useful sequences of events by means of the minimization of loss time of the shops in the process. The prerequisites of the proposed method are so general that the method can be widely applied to the rule refinement tasks in various kinds of rule-based systems.