Prediction based segmentation of state space and application to a subgoal finding problem in reinforcement learning

Humans solve problems by segmenting perceived continuous phenomenon and searching for action in compressed problem space. In this paper, we propose a method for segmenting continuous state space based on local prediction and a long-term prediction of continuous phenomenon. Furthermore, we investigate a subgoal finding problem in reinforcement learning as an instance of application of the segmentation result. A state having a subgoal function is found by propagating value from goal in a compressed state space. Reinforcement learning is accelerated by establishing subrewards in the state.