Improving MACS Thanks to a Comparison with 2TBNs

Olivier Sigaud, Thierry Gourdin, and Pierre-Henri Wuillemin

Abstract. Factored Markov Decision Processes is the theoretical framework underlying multi-step Learning Classifier Systems research. This framework is mostly used in the context of Two-stage Bayes Networks, a subset of Bayes Networks. In this paper, we compare the Learning Classifier Systems approach and the Bayes Networks approach to factored Markov Decision Problems. More specifically, we focus on a comparison between MACS, an Anticipatory Learning Classifier System, and Structured Policy Iteration, a general planning algorithm used in the context of Two-stage Bayes Networks. From that comparison, we define a new algorithm resulting from the adaptation of Structured Policy Iteration to the context of MACS. We conclude by calling for a closer communication between both research communities.

LNCS 3103, p. 810 ff.

Full article in PDF

lncs@springer.de
© Springer-Verlag Berlin Heidelberg 2004