MULTI-SCALE REINFORCEMENT LEARNING WITH FUZZY STATE

Zhuang, Xiaodong
MULTI-SCALE REINFORCEMENT LEARNING WITH FUZZY STATE
conference proceedings (Compressed PDF - 207KB)

Abstract: In this paper, multi-scale reinforcement learning is presented based on fuzzy state. The concept of fuzzy state is proposed to enable multi-scale representation of the state space. The performance of different learning scales is investigated, based on which a multi-scale learning approach is proposed to increase the learning speed while keeping the learning accuracy. The multi-scale learning approach is applied to the robot navigation problem in the computer simulation experiment. In a multi-obstacle environment, the multi-scale reinforcement learning approach shows better performance than the traditional reinforcement learning method.