XCS Revisited: A Novel Discovery Component for the eXtended Classifier System
Published: 2010 Dezember
Herausgeber: Kalyanmoy Deb and others
Buchtitel: Proceedings of the 8th International Conference on Simulated Evolution And Learning (SEAL-2010)
Erscheinungsort: Berlin Heidelberg
The eXtended Classifier System (XCS) is a rule-based evolutionary on-line learning system. Originally proposed by Wilson, XCS combines techniques from reinforcement learning and evolutionary optimization to learn a population of maximally general, but accurate condition-action rules. This paper focuses on the discovery component of XCS that is responsible for the creation and deletion of rules. A novel rule combining mechanism is proposed that infers maximally general rules from the existing population. Rule combining is evaluated for single- and multi-step learning problems using the well-known multiplexer, Woods, and Maze environments. Results indicate that the novel mechanism allows for faster learning rates and a reduced population size compared to the original XCS implementation.
DOI Link: 10.1007/978-3-642-17298-4_30