Robust Offline Actor-Critic with On-Policy Regularized Policy Evaluation

To alleviate the extrapolation error and instability inherent in Q-function directly learned by off-policy Q-learning (QL-style) on static datasets, this article utilizes the on-policy state-action-reward-state-action (SARSA-style) to develop an offline reinforcement learning (RL) method termed robust offline Actor-Critic with on-policy regularized policy evaluation (OPRAC). With the help of SARSA-style bootstrap actions, a conservative on-policy Q-function and a penalty term for matching the on-policy and off-policy actions are jointly constructed to regularize the optimal Q-function of off-policy QL-style. This naturally equips the off-policy QL-style policy evaluation with the intrinsic pessimistic conservatism of on-policy SARSA-style, thus facilitating the acquisition of stable estimated Q-function. Even with limited data sampling errors, the convergence of Q-function learned by OPRAC and the controllability of bias upper bound between the learned Q-function and its true Q-value can be theoretically guaranteed. In addition, the sub-optimality of learned optimal policy merely stems from sampling errors. Experiments on the well-known D4RL Gym-MuJoCo benchmark demonstrate that OPRAC can rapidly learn robust and effective task-solving policies owing to the stable estimate of Q-value, outperforming state-of-the-art offline RLs by at least 15%.