Summary
International Symposium on Nonlinear Theory and its Applications
2005
Session Number:1-3-3
Session:
Number:1-3-3-6
An Adaptive State Space Segmentation Method Based on ART Neural Network with Two Learning Phases for Reinforcement Learning
Taisuke Nakamura, Takeshi Kamio, Kunihiko Mitsubori, Hisato Fujisaka,
pp.266-269
Publication Date:2005/10/18
Online ISSN:2188-5079
DOI:10.34385/proc.40.1-3-3-6
PDF download (137.3KB)
Summary:
The trade-off between exploration and exploitation has often been discussed in studies on reinforcement learning (RL). This is because exploration and exploitation influence the quality of solutions and the learning efficiency respectively. Previously, we have proposed an adaptive state space segmentation method based on ART neural network (ART) to execute RL effectively. However, if the exploration strength is too large, the learning efficiency degreases rapidly. Since the appropriate strength is generally unknown, this problem must be solved. In this paper, we propose a new segmentation method based on ART with two learning phases to improve our conventional method in the tolerance of exploration strength.