We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

math.OC

Change to browse by:

References & Citations

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Mathematics > Optimization and Control

Title: Convergence of Optimistic Gradient Descent Ascent in Bilinear Games

Abstract: We study the convergence of Optimistic Gradient Descent Ascent in unconstrained bilinear games. In a first part, we consider the general zero-sum case and extend previous results by Daskalakis et al. in 2018, Liang and Stokes in 2019 and others: we prove for the first time the exponential convergence of OGDA to a saddle point, for any payoff matrix. We also provide a new, optimal, geometric ratio for the convergence, and in case of multiple Nash equilibria we characterize the limit equilibrium as a function of the initialization. In a second part, we introduce OGDA for general-sum bilinear games: we provide sufficient conditions for the convergence of OGDA to a Nash equilibrium, and show that in an interesting class of games, including common payoffs games and zero-sum games, either OGDA converges exponentially fast to a Nash equilibrium, or the payoffs of both players converge exponentially fast to $+\infty$ (which might be interpreted as the endogenous emergence of coordination, or cooperation, among players). We also present a very simple trick, the Double OGDA, ensuring convergence to a Nash equilibrium in any general-sum bilinear game. We finally illustrate our results on an example of Generative Adversarial Networks.
Subjects: Optimization and Control (math.OC)
Cite as: arXiv:2208.03085 [math.OC]
  (or arXiv:2208.03085v1 [math.OC] for this version)

Submission history

From: Etienne De Montbrun [view email]
[v1] Fri, 5 Aug 2022 10:42:26 GMT (981kb,D)

Link back to: arXiv, form interface, contact.