Federated Offline Reinforcement Learning with Proximal Policy Evaluation

Offline reinforcement learning (RL) has gathered increasing attention in recent years, which seeks to learn policies from static datasets without active online exploration. However, the existing offline RL approaches often require a large amount of pre-collected data and hence are hardly implemented by a single agent in practice. Inspired by the advancement of federated learning (FL), this paper studies federated offline reinforcement learning (FORL), whereby multiple agents collaboratively carry out offline policy learning with no need to share their raw trajectories. Clearly, a straightforward solution is to simply retrofit the off-the-shelf offline RL methods for FL, whereas such an approach easily overfits individual datasets during local updating, leading to instability and subpar performance. To overcome this challenge, we propose a new FORL algorithm, named model-free (MF)-FORL, that exploits novel "proximal local policy evaluation" to judiciously push up action values beyond local data support, enabling agents to capture the individual information without forgetting the aggregated knowledge. Further, we introduce a model-based variant, MB-FORL, capable of improving the generalization ability and computational efficiency via utilizing a learned dynamics model. We evaluate the proposed algorithms on a suite of complex and high-dimensional offline RL benchmarks, and the results demonstrate significant performance gains over the baselines.