強化学習における方策の性能を向上するサンプリング方策

抄録

When applying a reinforcement learning method, the estimation accuracy of the state transition probabilities affects the performance of the policy obtained from the estimated plant. Therefore, we find a sampling condition guaranteeing that the optimal policy from the estimated plant is also optimal for the real plant with the desired degree of reliability, and a sampling methods based on it is proposed. Not by the sampling for the reliability in which the policy is optimal for the real plant, but by the sampling for the policy to be effective irrespective of estimation errors, we can further reduce the number of samples. We show the problem setting for finding the policy which is guaranteed to be effective for estimation errors from the real transition probabilities with the desired degree of reliability, and we propose a sampling method as a solution of this problem. The effectiveness of the proposed method is verified by numerical simulations.