強化学習の方策再利用時におけるステップ単位の方策忘却手法

抄録

This paper presents the policy forgetting method in the step units when the reinforcement learning agent reuses policy which is obtained from reinforcement learning. Policy reuse methods such as transfer learning has been discussed for the real robotics application. However, transfer learning has possibility that the agent is encountered negative transfer which is caused by mismatch between task settings and reusing policies. In this paper, we propose the policy forgetting method when the agent is encountered negative transfer, further evaluate the effectiveness of proposed method in computer simulation.