Knowledge Selection based on Value in Transfer Learning

摘要

This paper is aimed at reducing the amount of knowledge to avoid lower learning performance of an agent in transfer learning. In transfer or multitask reinforcement learning problems, the agent reuses policies which were learned in past tasks in order to efficiently solve unknown tasks. Therefore,the agent has a large number of state-action pairs as knowledge. But, at the same time, it causes both explosively increasing the amount of knowledge and decreasing the learning speed. This paper proposes a method for reducing the amount of knowledge on the basis of value. The effectiveness of the proposed method was verified with the simulation of the reaching problem for a multi-link robot arm. The proposed method achieves a reduction of the amount of knowledge and learning time. It also improves learning performance of the agent.