JOURNAL ARTICLE

Corruption-Robust Exploration in Episodic Reinforcement Learning.

  • Published In: Mathematics of Operations Research (INFORMS), 2025, v. 50, n. 2. P. 1277 1 of 3

  • Database: Business Source Ultimate 2 of 3

  • Authored By: Lykouris, Thodoris; Simchowitz, Max; Slivkins, Aleksandrs; Sun, Wen 3 of 3

Abstract

This article focuses on episodic reinforcement learning (RL) under adversarial corruptions affecting both rewards and transition probabilities, extending prior work from multiarmed bandits to more complex RL settings. It introduces a novel algorithmic framework, SuperVIsed.C, which combines optimism under uncertainty with active action elimination while addressing unique RL challenges such as trajectory mismatch and exponential regret in naive elimination. The framework yields near-optimal regret bounds that gracefully degrade with unknown corruption levels and applies to both tabular MDPs (finite states and actions) and linear MDPs (with linear structure in dynamics and rewards). Additionally, the article establishes a lower bound showing that uniformly action-eliminating algorithms suffer exponential regret in episodic RL, highlighting the necessity of the proposed approach.

Additional Information

  • Source:Mathematics of Operations Research (INFORMS). 2025/05, Vol. 50, Issue 2, p1277
  • Document Type:Article
  • Subject Area:Education
  • Publication Date:2025
  • ISSN:0364-765X
  • DOI:10.1287/moor.2021.0202
  • Accession Number:185001418
  • Copyright Statement:Copyright of Mathematics of Operations Research (INFORMS) is the property of INFORMS: Institute for Operations Research & the Management Sciences and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

Looking to go deeper into this topic? Look for more articles on EBSCOhost.