Saved in:
Bibliographic Details
Main Authors: Zhu, Jiahui, Yu, Kihyun, Lee, Dabeen, Liu, Xin, Wei, Honghao
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2505.21841
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912405208956928
author Zhu, Jiahui
Yu, Kihyun
Lee, Dabeen
Liu, Xin
Wei, Honghao
author_facet Zhu, Jiahui
Yu, Kihyun
Lee, Dabeen
Liu, Xin
Wei, Honghao
contents Online safe reinforcement learning (RL) plays a key role in dynamic environments, with applications in autonomous driving, robotics, and cybersecurity. The objective is to learn optimal policies that maximize rewards while satisfying safety constraints modeled by constrained Markov decision processes (CMDPs). Existing methods achieve sublinear regret under stochastic constraints but often fail in adversarial settings, where constraints are unknown, time-varying, and potentially adversarially designed. In this paper, we propose the Optimistic Mirror Descent Primal-Dual (OMDPD) algorithm, the first to address online CMDPs with anytime adversarial constraints. OMDPD achieves optimal regret O(sqrt(K)) and strong constraint violation O(sqrt(K)) without relying on Slater's condition or the existence of a strictly known safe policy. We further show that access to accurate estimates of rewards and transitions can further improve these bounds. Our results offer practical guarantees for safe decision-making in adversarial environments.
format Preprint
id arxiv_https___arxiv_org_abs_2505_21841
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle An Optimistic Algorithm for online CMDPS with Anytime Adversarial Constraints
Zhu, Jiahui
Yu, Kihyun
Lee, Dabeen
Liu, Xin
Wei, Honghao
Machine Learning
Artificial Intelligence
Online safe reinforcement learning (RL) plays a key role in dynamic environments, with applications in autonomous driving, robotics, and cybersecurity. The objective is to learn optimal policies that maximize rewards while satisfying safety constraints modeled by constrained Markov decision processes (CMDPs). Existing methods achieve sublinear regret under stochastic constraints but often fail in adversarial settings, where constraints are unknown, time-varying, and potentially adversarially designed. In this paper, we propose the Optimistic Mirror Descent Primal-Dual (OMDPD) algorithm, the first to address online CMDPs with anytime adversarial constraints. OMDPD achieves optimal regret O(sqrt(K)) and strong constraint violation O(sqrt(K)) without relying on Slater's condition or the existence of a strictly known safe policy. We further show that access to accurate estimates of rewards and transitions can further improve these bounds. Our results offer practical guarantees for safe decision-making in adversarial environments.
title An Optimistic Algorithm for online CMDPS with Anytime Adversarial Constraints
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2505.21841