Instruction knowledge acquisition for reinforcement learning scheme by PSO algorithm

In order to realize intelligent agents such as autonomous mobile robots, Reinforcement Learning is one of the necessary techniques in control systems. It is desirable in terms of knowledge or skill acquisition of agents that reinforcement learning is based only upon rewards instead of teaching signals. However, there exist many problems to apply reinforcement learning to real-world tasks. The most severe problem is a huge number of iterations in the learning phase. In order to deal with the problem, the instruction approach for reinforcement learning agents based on sub-rewards and forgetting mechanisms were proposed and shown to be effective. However, the relationship between the instruction and the learning performance of reinforcement learning has not been adequately clarified. In this study, in order to clarify the instruction performance in the reinforcement learning, we propose an instruction knowledge acquisition method for the reinforcement learning scheme by the particle swarm optimization (PSO) algorithm. Through numerical experiments of the mountain car task and the Acrobat task, we show the validness of the proposed approach in terms of learning speed and accuracy.