Corruption-Tolerant Asynchronous Q-Learning with Near-Optimal Rates

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Maity, Sreejeet, Mitra, Aritra
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866911705328517120
author Maity, Sreejeet
Mitra, Aritra
author_facet Maity, Sreejeet
Mitra, Aritra
contents We study the problem of learning the optimal policy in a discounted, infinite-horizon reinforcement learning (RL) setting in the presence of adversarially corrupted rewards. To address this problem, we develop a novel robust variant of the \(Q\)-learning algorithm and analyze it under the challenging asynchronous sampling model with time-correlated data. Despite corruption, we prove that the finite-time guarantees of our approach match existing bounds, up to an additive term that scales with the fraction of corrupted samples. We also establish an information-theoretic lower bound, revealing that our guarantees are near-optimal. Notably, our algorithm is agnostic to the underlying reward distribution and provides the first finite-time robustness guarantees for asynchronous \(Q\)-learning. A key element of our analysis is a refined Azuma-Hoeffding inequality for almost-martingales, which may have broader applicability in the study of RL algorithms.
format Preprint
id arxiv_https___arxiv_org_abs_2509_08933
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Corruption-Tolerant Asynchronous Q-Learning with Near-Optimal Rates
Maity, Sreejeet
Mitra, Aritra
Machine Learning
Systems and Control
Optimization and Control
We study the problem of learning the optimal policy in a discounted, infinite-horizon reinforcement learning (RL) setting in the presence of adversarially corrupted rewards. To address this problem, we develop a novel robust variant of the \(Q\)-learning algorithm and analyze it under the challenging asynchronous sampling model with time-correlated data. Despite corruption, we prove that the finite-time guarantees of our approach match existing bounds, up to an additive term that scales with the fraction of corrupted samples. We also establish an information-theoretic lower bound, revealing that our guarantees are near-optimal. Notably, our algorithm is agnostic to the underlying reward distribution and provides the first finite-time robustness guarantees for asynchronous \(Q\)-learning. A key element of our analysis is a refined Azuma-Hoeffding inequality for almost-martingales, which may have broader applicability in the study of RL algorithms.
title Corruption-Tolerant Asynchronous Q-Learning with Near-Optimal Rates
topic Machine Learning
Systems and Control
Optimization and Control
url https://arxiv.org/abs/2509.08933