Epistemic Robust Offline Reinforcement Learning

TL;DR

This paper introduces a new framework for offline reinforcement learning that uses uncertainty sets instead of ensembles to improve robustness and generalization, especially under risk-sensitive policies.

Contribution

It proposes a unified approach replacing ensembles with uncertainty sets and introduces an Epinet-based model for better uncertainty estimation in offline RL.

Findings

Our method outperforms ensemble baselines in robustness and generalization.

The approach is effective in both tabular and continuous state domains.

A new benchmark for evaluating offline RL under risk-sensitive policies is introduced.

Abstract

Offline reinforcement learning learns policies from fixed datasets without further environment interaction. A key challenge in this setting is epistemic uncertainty, arising from limited or biased data coverage, particularly when the behavior policy systematically avoids certain actions. This can lead to inaccurate value estimates and unreliable generalization. Ensemble-based methods like SAC-N mitigate this by conservatively estimating Q-values using the ensemble minimum, but they require large ensembles and often conflate epistemic with aleatoric uncertainty. To address these limitations, we propose a unified and generalizable framework that replaces discrete ensembles with compact uncertainty sets over Q-values. %We further introduce an Epinet based model that directly shapes the uncertainty sets to optimize the cumulative reward under the robust Bellman objective without relying on ensembles. We also introduce a benchmark for evaluating offline RL algorithms under risk-sensitive behavior policies, and demonstrate that our method achieves improved robustness and generalization over ensemble-based baselines across both tabular and continuous state domains.