Robust Asynchronous Q-Learning under Reward and State Corruption via Batching

A new robust Q-learning algorithm, BR-Async-Q, achieves the first robustness guarantee for asynchronous Q-learning under both reward and state corruption, matching vanilla Q-learning error bounds up to a small additive term proportional to the corruption fraction. The algorithm, introduced in arXiv:2607.20822v1, uses epoch-based batching and robust Bellman operator estimates to defend against adversarial perturbations in harsh environments, with minimax optimal corruption dependence when only rewards are corrupted.

arXiv:2607.20822v1 Announce Type: new Abstract: Motivated by reinforcement learning in harsh environments, we consider the problem of learning an optimal policy subject to adversarially corrupted feedback. Specifically, at each time-step, an adversary can perturb both the reward and state observations of the learner following the Huber contamination model. To defend against such data corruption, we propose {{\texttt{BR-Async-Q}}}: a novel, epoch-based, robust \ Q\ -learning algorithm built upon two key ideas: i partitioning the online data stream into batches to reduce variance, and ii constructing robust estimates of the Bellman optimality operator using such batched data. We prove a high-probability $\ell \infty$ error bound for {{\texttt{BR-Async-Q}}} that matches that for vanilla \ Q\ -learning, up to a small additive term that scales with the fraction of corrupted samples. To our knowledge, this provides the first robustness guarantee for asynchronous \ Q\ -learning subject to both reward and state corruption. Furthermore, when only rewards are corrupted, the dependence of our algorithm's bound on the corruption fraction is minimax optimal.