Offline Reinforcement Learning via Conservative Smoothing and Dynamics Controlling

Abstract

Offline Reinforcement Learning (RL) optimizes policy using pre-collected data instead of direct environment interaction, offering a safe and cost-effective solution for sequential decision-making in the real world. However, it faces challenges such as distribution shift issues and vulnerability under perturbations. Researchers have developed various conservative methods to improve the robustness of offline RL. Nevertheless, model-based methods can result in transition distribution shift issues, while model-free value-based uncertainty penalty methods may not be sufficiently robust. To address these problems, we propose a new method called Robust Offline RL via Conservative Smoothing and Dynamics Controlling (RCSD). To achieve reliable value estimation of out-of-distribution (OOD) actions, RCSD uses both model-free uncertainty penalty and model-based simulation methods. It introduces a new one-step simulation method with conservative dynamics controlling to avoid value overestimation caused by transition distribution shifts. Moreover, RCSD considers both current and next states when generating OOD states to ensure cautious value estimation and efficient data utilization. RCSD uses conservative Q-smoothing and policy smoothing to strengthen the policy against sudden changes under perturbations. Experiments on D4RL benchmark demonstrate that RCSD can achieve state-of-the-art performance compared to baselines in either benchmark or adversarial attack tests.