Enhancing Learning Capabilities by XCS with Best Action Mapping

Masaya Nakata¹, Pier Luca Lanzi², and Keiki Takadama¹

¹Department of Informatics, The university of Electo-Communications, Tokyo, Japan
m.nakata@cas.hc.uec.ac.jp
keiki@inf.uec.ac.jp

²Dipartimento di Elettronica e Informazione, Politecnico di Milano, Milano, Italy
lanzi@elet.polimi.it

Abstract. This paper proposes a novel approach of XCS called XCS with Best Action Mapping (XCSB) to enhance the learning capabilities of XCS. The feature of XCSB is to learn only best actions having the highest predicted payoff with the high accuracy unlike XCS which learns actions having the highest and lowest predicted payoff with the high accuracy. To investigate the effectiveness of XCSB, we applied XCSB to two benchmark problems: multiplexer problem as a single step problem and maze problem as a multi step problem. The experimental results show that (1) XCSB can solve quickly the problem which has a large state space and (2) XCSB can achieve a high performance with a small max population size.

LNCS 7491, p. 256 ff.

Full article in PDF | BibTeX

lncs@springer.com
© Springer-Verlag Berlin Heidelberg 2012