Basic Research on Speed-Up of Reinforcement Learning Using Parallel Processing for Combination Value Function

In this paper we use parallel processing to combine value functions in order to speedup reinforcement learning. We propose an asynchronous method of periodically composing Q table of local learning clusters to form global Q table. In this research, two approaches are implemented. First is discontinuance learning. Second is combination of value function by asynchronous communication. The asynchronous combination method is compared with a synchronous combination method in order of learning times. A cluster of 40 PCs were used in the experiments are presented. The convergence time and learning times are evaluated and discussed.