HIPODE: Enhancing Offline Reinforcement Learning with High-Quality Synthetic Data from a Policy-Decoupled Approach

TL;DR — We propose a novel data augmentation algorithm HIPODE for offline RL to generate high-return synthetic data, which is beneficial for different downstream policies.

Abstract

Offline reinforcement learning (Offline RL) has gained attention as a means of training reinforcement learning models using pre-collected static data. To address the issue of limited data and improve downstream Offline RL performance, recent efforts have focused on broadening dataset coverage through data augmentation techniques. However, most of these methods are tied to a specific policy (policy-dependent), restricting the generated data to supporting only a specific downstream Offline RL policy. Moreover, the quality of synthetic data is often not well-controlled, which limits the potential for further improving the downstream policy. To tackle these issues, we propose \textbf{HI}gh-return \textbf{PO}licy-\textbf{DE}coupled~(HIPODE), a novel data augmentation method for Offline RL. On the one hand, HIPODE generates high-return synthetic data by selecting states near the dataset distribution with potentially high value among candidate states using the negative sampling technique. On the other hand, HIPODE is policy-decoupled, thus can be used as a common plug-in method to support diverse downstream Offline RL processes. We conduct experiments on the widely studied TD3BC, CQL and IQL algorithms, and the results show that HIPODE outperforms or has competitive results to the state-of-the-art policy-decoupled data augmentation method and most prevalent model-based Offline RL methods on D4RL benchmarks.