Using Spatial Hints to Improve Policy Reuse in a Reinforcement Learning Agent 

By Bruno da Silva

We study the problem of knowledge reuse by a reinforcement learning agent. We are interested in how an agent can exploit policies that were learned in the past to learn a new task more efficiently in the present. Our approach is to elicit spatial hints from an expert suggesting the world states in which each existing policy should be more relevant to the new task. By confronting these hints with domain exploration, the agent is able to distinguish those portions of existing policies that are beneficial to the new task, therefore learning a new policy more efficiently. We call our approach Spatial Hints Policy Reuse (SHPR). Experiments demonstrate the effectiveness of our method, and its robustness against bad inputs.

Visit the LCI Forum page