Reproduce “On Finding Local Nash Equilibria (and Only Local Nash Equilibria) in Zero-Sum Games” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Eric V. Mazumdar, Michael I. Jordan, S. Shankar Sastry
A research team from the University of California, Berkeley, introduced Local Symplectic Surgery (LSS), a gradient-based algorithm for two-player zero-sum continuous games, engineered to converge exclusively to local Nash equilibria. This method eliminates problematic non-Nash locally asymptotically stable equilibria and suppresses oscillatory dynamics, demonstrating improved stability and quality in Generative Adversarial Network training compared to standard approaches.
Key insights
- The anti-symmetric component of the game's Jacobian matrix is the primary cause of non-Nash locally asymptotically stable equilibria and oscillatory behavior in gradient dynamics.
- By specifically modifying the dynamics to eliminate this anti-symmetric part *at equilibrium points*, only true local Nash equilibria can become attracting fixed points.
- Complex dynamical systems that conceptually use higher-order information can be approximated effectively and efficiently via a two-timescale stochastic procedure using only Jacobian-vector products.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.