An adaptive state space segmentation for reinforcement learning using fuzzy-ART neural network

Abstract

Reinforcement learning has been applied to a variety of physical control tasks. They include many purposive tasks with continuous state variables and discrete-valued actions. The state space segmentation is one of the most important problems for such tasks. However, if they are not given serious damages by "a state-action deviation problem", the conventional methods are unsuitable for them in terms of the cost-performance and the simplicity of the algorithm. To overcome this problem, we propose a new adaptive state space segmentation method based on fuzzy-ART neural network.