Optimal last-iterate convergence in matrix games with bandit feedback using the log-barrier

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Fiegel, Come, Menard, Pierre, Kozuno, Tadashi, Valko, Michal, Perchet, Vianney
Natura: Preprint
Pubblicazione: 2026
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866908970693689344
author Fiegel, Come
Menard, Pierre
Kozuno, Tadashi
Valko, Michal
Perchet, Vianney
author_facet Fiegel, Come
Menard, Pierre
Kozuno, Tadashi
Valko, Michal
Perchet, Vianney
contents We study the problem of learning minimax policies in zero-sum matrix games. Fiegel et al. (2025) recently showed that achieving last-iterate convergence in this setting is harder when the players are uncoupled, by proving a lower bound on the exploitability gap of Omega(t^{-1/4}). Some online mirror descent algorithms were proposed in the literature for this problem, but none have truly attained this rate yet. We show that the use of a log-barrier regularization, along with a dual-focused analysis, allows this O-tilde(t^{-1/4}) convergence with high-probability. We additionally extend our idea to the setting of extensive-form games, proving a bound with the same rate.
format Preprint
id arxiv_https___arxiv_org_abs_2604_15242
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Optimal last-iterate convergence in matrix games with bandit feedback using the log-barrier
Fiegel, Come
Menard, Pierre
Kozuno, Tadashi
Valko, Michal
Perchet, Vianney
Machine Learning
We study the problem of learning minimax policies in zero-sum matrix games. Fiegel et al. (2025) recently showed that achieving last-iterate convergence in this setting is harder when the players are uncoupled, by proving a lower bound on the exploitability gap of Omega(t^{-1/4}). Some online mirror descent algorithms were proposed in the literature for this problem, but none have truly attained this rate yet. We show that the use of a log-barrier regularization, along with a dual-focused analysis, allows this O-tilde(t^{-1/4}) convergence with high-probability. We additionally extend our idea to the setting of extensive-form games, proving a bound with the same rate.
title Optimal last-iterate convergence in matrix games with bandit feedback using the log-barrier
topic Machine Learning
url https://arxiv.org/abs/2604.15242