Summary
International Symposium on Nonlinear Theory and its Applications
2005
Session Number:1-3-3
Session:
Number:1-3-3-2
An efficient reinforcement learning method for dynamic environments using short term adjustment
Hidehiro Nakano, Satoko Takada, Shuichi Arai, Arata Miyauchi,
pp.250-253
Publication Date:2005/10/18
Online ISSN:2188-5079
Summary:
This paper proposes a novel reinforcement learning method for dynamic environments. A learning agent estimates changing environments by comparing rule sequence with each action selection probability. If the change is estimated, action selection probabilities are temporarily adjusted. We derive the condition for the amount of adjustment to be flexibly adaptive for dynamic environments. Our method provides better learning performances in various dynamic environments than conventional methods. We present some numerical results for our method applied to dynamic maze problems.