Policy transfer using Value Function as Prior Information

This work proposes an approach based on reward shaping techniques in a reinforcement learning setting to approximate the opti- mal decision-making process (also called the optimal policy) in a desired task with a limited amount of data. We extract prior information from an existing family of policies have been used as a heuristic to help the construction of the new one under this challenging condition. We use this approach to study the relationship between the similarity of two tasks and the minimal amount of data needed to compute a near-optimal pol- icy for the second one using the prior information of the existing policy. Preliminary results show that for the least similar existing task consid- ered compared to the desired one, only 10% of the dataset was needed to compute the corresponding near-optimal policy.