Learning to predict with the Delayed Action Classifier System

This paper describes an extended version of the Delayed Action Classifier System (DACS), a rule-based system which employs the genetic algorithm to discover temporal rules for learning in environments with temporal structure. The extended version of DACS (called DACS2) is described and experimentally evaluated. A mathematical analysis is provided which offers a theoretical justification of the advantages of DACS compared to non-temporal classifier systems. Areas for further development are briefly discussed and possible applications of the system in the field of intelligent control are suggested.