Similarity Evaluation of Multiple Policies Applied to Transfer Learning Agents

In this study, we have been developing a policy transfer method based on a spreading activation model in the cognitive psychology for a reinforcement learning agent for suitable learning in unknown or dynamic environments. This policy transfer method categorizes policies and constructs a network among policies. Previously, the similarity between policies was compared between Q-tables for policy categorization in transfer learning using Q-learning. This paper proposes a method for evaluating the similarity between policies using the variance of Q values as an indicator for transferring and reusing policies with different rewards. We verify the effectiveness of the proposed method by the shortest path problem in a grid world.