Solving Zero-Sum Convex Markov Games

Fuente: arXiv
Enregistré dans:
Détails bibliographiques
Auteurs principaux: Kalogiannis, Fivos, Vlatakis-Gkaragkounis, Emmanouil-Vasileios, Gemp, Ian, Piliouras, Georgios
Format: Preprint
Publié: 2025
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866915350807838720
author Kalogiannis, Fivos
Vlatakis-Gkaragkounis, Emmanouil-Vasileios
Gemp, Ian
Piliouras, Georgios
author_facet Kalogiannis, Fivos
Vlatakis-Gkaragkounis, Emmanouil-Vasileios
Gemp, Ian
Piliouras, Georgios
contents We contribute the first provable guarantees of global convergence to Nash equilibria (NE) in two-player zero-sum convex Markov games (cMGs) by using independent policy gradient methods. Convex Markov games, recently defined by Gemp et al. (2024), extend Markov decision processes to multi-agent settings with preferences that are convex over occupancy measures, offering a broad framework for modeling generic strategic interactions. However, even the fundamental min-max case of cMGs presents significant challenges, including inherent nonconvexity, the absence of Bellman consistency, and the complexity of the infinite horizon. We follow a two-step approach. First, leveraging properties of hidden-convex--hidden-concave functions, we show that a simple nonconvex regularization transforms the min-max optimization problem into a nonconvex-proximal Polyak-Lojasiewicz (NC-pPL) objective. Crucially, this regularization can stabilize the iterates of independent policy gradient methods and ultimately lead them to converge to equilibria. Second, building on this reduction, we address the general constrained min-max problems under NC-pPL and two-sided pPL conditions, providing the first global convergence guarantees for stochastic nested and alternating gradient descent-ascent methods, which we believe may be of independent interest.
format Preprint
id arxiv_https___arxiv_org_abs_2506_16120
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Solving Zero-Sum Convex Markov Games
Kalogiannis, Fivos
Vlatakis-Gkaragkounis, Emmanouil-Vasileios
Gemp, Ian
Piliouras, Georgios
Computer Science and Game Theory
Machine Learning
Multiagent Systems
Optimization and Control
We contribute the first provable guarantees of global convergence to Nash equilibria (NE) in two-player zero-sum convex Markov games (cMGs) by using independent policy gradient methods. Convex Markov games, recently defined by Gemp et al. (2024), extend Markov decision processes to multi-agent settings with preferences that are convex over occupancy measures, offering a broad framework for modeling generic strategic interactions. However, even the fundamental min-max case of cMGs presents significant challenges, including inherent nonconvexity, the absence of Bellman consistency, and the complexity of the infinite horizon. We follow a two-step approach. First, leveraging properties of hidden-convex--hidden-concave functions, we show that a simple nonconvex regularization transforms the min-max optimization problem into a nonconvex-proximal Polyak-Lojasiewicz (NC-pPL) objective. Crucially, this regularization can stabilize the iterates of independent policy gradient methods and ultimately lead them to converge to equilibria. Second, building on this reduction, we address the general constrained min-max problems under NC-pPL and two-sided pPL conditions, providing the first global convergence guarantees for stochastic nested and alternating gradient descent-ascent methods, which we believe may be of independent interest.
title Solving Zero-Sum Convex Markov Games
topic Computer Science and Game Theory
Machine Learning
Multiagent Systems
Optimization and Control
url https://arxiv.org/abs/2506.16120