Reward Design Method Adapting to Agents' Learning Ability based on Self-Organizing Map with Evaluation Value

Abstract

In education for children and guidance of sports, it is important to give appropriate instruction to learners. It is necessary to grasp the ability and characteristic of the learner by observing the learning process and to change the teaching method as needed. In this paper, we consider the learning parameter and appropriate giving rewards method, using simulation data which makes agent learn maze. For learning of the maze, we used Q-learning well known in the field of reinforcement learning. And we conducted experiments using multiple agents with different learning parameters. Agent behavior data at the middle stage of learning is classified by SOM and learning parameters are estimated. After that, we change the giving rewards method, and consider it according to the learning parameters from learning result.