Abstract
In this report, we propose a reinforcement learning that a target agent can learn an unknown environment efficiently by applying a knowledge acquired by another source agent with a different form. Concretely, A Target agent has three tables, Target Q-table, Source Q-table and State-Transfer Table(STT), Target Q-table is unlearned and represented for Target's state-action pair, Source Q-table is already learned and for Source's state-action pair, STT is unlearned and for transferring Target state into Source state. It sets Target Q-table and (STT and Source Q-table in series) in parallel Target's Q-table and Source's Q-table have the almost same size. STT has a smaller size. We expect that the proposed method converges early in detail, since in the learning, STT with Source Q-table converges early but roughly, and Target Q-table converges slowly but in detail.
