Improved Regret Bound and Experience Replay in Regularized Policy Iteration