de.mpg.escidoc.pubman.appbase.FacesBean
English
 
Help Guide Disclaimer Contact us Login
  Advanced SearchBrowse

Item

ITEM ACTIONSEXPORT

Released

Journal Article

Derivatives of Logarithmic Stationary Distributions for Policy Gradient Reinforcement Learning

MPS-Authors
http://pubman.mpdl.mpg.de/cone/persons/resource/persons84135

Uchibe E, Yoshimoto J, Peters,  J
Department Empirical Inference, Max Planck Institute for Biological Cybernetics, Max Planck Society;
Dept. Empirical Inference, Max Planck Institute for Intelligent Systems, Max Planck Society;

Locator
There are no locators available
Fulltext (public)
There are no public fulltexts available
Supplementary Material (public)
There is no public supplementary material available
Citation

Morimura, T., Uchibe E, Yoshimoto J, Peters, J., & Doya, K. (2010). Derivatives of Logarithmic Stationary Distributions for Policy Gradient Reinforcement Learning. Neural Computation, 22(2), 342-376. doi:10.1162/neco.2009.12-08-922.


Cite as: http://hdl.handle.net/11858/00-001M-0000-0013-C12A-1
Abstract
Most conventional policy gradient reinforcement learning (PGRL) algorithms neglect (or do not explicitly make use of) a term in the average reward gradient with respect to the policy parameter. That term involves the derivative of the stationary state distribution that corresponds to the sensitivity of its distribution to changes in the policy parameter. Although the bias introduced by this omission can be reduced by setting the forgetting rate 947; for the value functions close to 1, these algorithms do not permit 947; to be set exactly at 947; = 1. In this article, we propose a method for estimating the log stationary state distribution derivative (LSD) as a useful form of the derivative of the stationary state distribution through backward Markov chain formulation and a temporal difference learning framework. A new policy gradient (PG) framework with an LSD is also proposed, in which the average reward gradient can be estimated by setting //!-- MFG_und--//amp;947; = 0, so it becomes unnecessary to learn the value functions. We also test the performance of the proposed algorithms using simple benchmark tasks and show that these can improve the performances of existing PG methods.