Global Convergence for Average Reward Constrained MDPs with Primal-Dual Actor Critic Algorithm

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Xu, Yang, Ganesh, Swetha, Mondal, Washim Uddin, Bai, Qinbo, Aggarwal, Vaneet
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866914190384431104
author Xu, Yang
Ganesh, Swetha
Mondal, Washim Uddin
Bai, Qinbo
Aggarwal, Vaneet
author_facet Xu, Yang
Ganesh, Swetha
Mondal, Washim Uddin
Bai, Qinbo
Aggarwal, Vaneet
contents This paper investigates infinite-horizon average reward Constrained Markov Decision Processes (CMDPs) with general parametrization. We propose a Primal-Dual Natural Actor-Critic algorithm that adeptly manages constraints while ensuring a high convergence rate. In particular, our algorithm achieves global convergence and constraint violation rates of $\tilde{\mathcal{O}}(1/\sqrt{T})$ over a horizon of length $T$ when the mixing time, $τ_{\mathrm{mix}}$, is known to the learner. In absence of knowledge of $τ_{\mathrm{mix}}$, the achievable rates change to $\tilde{\mathcal{O}}(1/T^{0.5-ε})$ provided that $T \geq \tilde{\mathcal{O}}\left(τ_{\mathrm{mix}}^{2/ε}\right)$. Our results match the theoretical lower bound for Markov Decision Processes and establish a new benchmark in the theoretical exploration of average reward CMDPs.
format Preprint
id arxiv_https___arxiv_org_abs_2505_15138
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Global Convergence for Average Reward Constrained MDPs with Primal-Dual Actor Critic Algorithm
Xu, Yang
Ganesh, Swetha
Mondal, Washim Uddin
Bai, Qinbo
Aggarwal, Vaneet
Machine Learning
Artificial Intelligence
This paper investigates infinite-horizon average reward Constrained Markov Decision Processes (CMDPs) with general parametrization. We propose a Primal-Dual Natural Actor-Critic algorithm that adeptly manages constraints while ensuring a high convergence rate. In particular, our algorithm achieves global convergence and constraint violation rates of $\tilde{\mathcal{O}}(1/\sqrt{T})$ over a horizon of length $T$ when the mixing time, $τ_{\mathrm{mix}}$, is known to the learner. In absence of knowledge of $τ_{\mathrm{mix}}$, the achievable rates change to $\tilde{\mathcal{O}}(1/T^{0.5-ε})$ provided that $T \geq \tilde{\mathcal{O}}\left(τ_{\mathrm{mix}}^{2/ε}\right)$. Our results match the theoretical lower bound for Markov Decision Processes and establish a new benchmark in the theoretical exploration of average reward CMDPs.
title Global Convergence for Average Reward Constrained MDPs with Primal-Dual Actor Critic Algorithm
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2505.15138