论文信息 - Using Expectation-Maximization for Reinforcement Learning

Using Expectation-Maximization for Reinforcement Learning

We discuss Hinton's (1989) relative payoff procedure (RPP), a static reinforcement learning algorithm whose foundation is not stochastic gradient ascent. We show circumstances under which applying the RPP is guaranteed to increase the mean return, even though it can make large changes in the values of the parameters. The proof is based on a mapping between the RPP and a form of the expectation-maximization procedure of Dempster, Laird, and Rubin (1977).

Geoffrey E. Hinton | Peter Dayan | P. Dayan

[1] Ronald A. Howard,et al. Dynamic Programming and Markov Processes , 1960 .

[2] F. Downton. Stochastic Approximation , 1969, Nature.

[3] L. Baum,et al. A Maximization Technique Occurring in the Statistical Analysis of Probabilistic Functions of Markov Chains , 1970 .

[4] D. Rubin,et al. Maximum likelihood from incomplete data via the EM - algorithm plus discussions on the paper , 1977 .

[5] Richard S. Sutton,et al. Temporal credit assignment in reinforcement learning , 1984 .

[6] P. Anandan,et al. Pattern-recognizing stochastic learning automata , 1985, IEEE Transactions on Systems, Man, and Cybernetics.

[7] Kumpati S. Narendra,et al. Learning automata - an introduction , 1989 .

[8] Geoffrey E. Hinton. Connectionist Learning Procedures , 1989, Artif. Intell..

[9] 永福智志. The Organization of Learning , 2005, Journal of Cognitive Neuroscience.

[10] Michael I. Jordan,et al. Reinforcement Learning by Probability Matching , 1995, NIPS 1995.