RORL: Robust Offline Reinforcement Learning via Conservative Smoothing
This addresses robustness issues in offline RL for decision-making tasks under realistic conditions like sensor errors, though it appears incremental as it builds on existing conservative methods.
The paper tackles the trade-off between robustness and conservatism in offline reinforcement learning by proposing RORL with a conservative smoothing technique, achieving state-of-the-art performance on benchmarks and showing robustness to adversarial perturbations.
Offline reinforcement learning (RL) provides a promising direction to exploit massive amount of offline data for complex decision-making tasks. Due to the distribution shift issue, current offline RL algorithms are generally designed to be conservative in value estimation and action selection. However, such conservatism can impair the robustness of learned policies when encountering observation deviation under realistic conditions, such as sensor errors and adversarial attacks. To trade off robustness and conservatism, we propose Robust Offline Reinforcement Learning (RORL) with a novel conservative smoothing technique. In RORL, we explicitly introduce regularization on the policy and the value function for states near the dataset, as well as additional conservative value estimation on these states. Theoretically, we show RORL enjoys a tighter suboptimality bound than recent theoretical results in linear MDPs. We demonstrate that RORL can achieve state-of-the-art performance on the general offline RL benchmark and is considerably robust to adversarial observation perturbations.