A novel framework of representation policy iteration based on Fuzzy C-means clustering method

摘要

To resolve the subset sampling problem in the representation policy iteration (RPI) method in Reinforcement Learning, a sub-sampling method - Fuzzy C-means clustering method is proposed in this paper. Fuzzy C-means method is used to select the representative points from the collected samples in the proposed method. Based on the proposed method, an integrated RPI algorithm is introduced. Illustrative experiments on Mountain Car problem were accomplished, to show the excellent performance of the introduced RPI algorithm. In addition, the method using the previous trajectory-based sub-sampling method is also presented, to demonstrate the advantages of the proposed method.