论文信息 - A Michigan style architecture for learning finite state controllers: a first step

A Michigan style architecture for learning finite state controllers: a first step

In this paper, we describe FACS, a new Michigan style architecture able to build Finite-State Automata controllers for agents learning to solve non-Markov problems. FACS relies on a population of partial automata and implements a Reinforcement Learning algorithm to compute the strength of each automaton and a Genetic Algorithm to select and discover efficient automata. We detail our approach and present very preliminary results.

Olivier Sigaud | Samuel Landau

[1] Larry Bull,et al. An accuracy based corporate classifier system , 2002, Soft Comput..

[2] Richard S. Sutton,et al. Reinforcement Learning: An Introduction , 1998, IEEE Trans. Neural Networks.

[3] Ron Sun,et al. Multi-Agent Reinforcement Learning with Bidding for Segmenting Action Sequences , 2000 .

[4] M. Pelikán,et al. Analyzing the evolutionary pressures in XCS , 2001 .

[5] Stewart W. Wilson,et al. Toward Optimal Classifier System Performance in Non-Markov Environments , 2000, Evolutionary Computation.