Sparse Offline RL Robust to Data Corruption

Published:Dec 31, 2025 10:28
1 min read
ArXiv

Analysis

This paper addresses the challenge of robust offline reinforcement learning in high-dimensional, sparse Markov Decision Processes (MDPs) where data is subject to corruption. It highlights the limitations of existing methods like LSVI when incorporating sparsity and proposes actor-critic methods with sparse robust estimators. The key contribution is providing the first non-vacuous guarantees in this challenging setting, demonstrating that learning near-optimal policies is still possible even with data corruption and specific coverage assumptions.

Reference

The paper provides the first non-vacuous guarantees in high-dimensional sparse MDPs with single-policy concentrability coverage and corruption, showing that learning a near-optimal policy remains possible in regimes where traditional robust offline RL techniques may fail.