Smooth game optimization has recently attracted great interest in machine learning as it generalizes the single-objective optimization paradigm. However, game dynamics is more complex due to the interaction between different players and is therefore fundamentally different from minimization, posing new challenges for algorithm design. Notably, it has been shown that negative momentum is preferred due to its ability to reduce oscillation in game dynamics. Nevertheless, the convergence rate of negative momentum was only established in simple bilinear games. In this paper, we extend the analysis to smooth and strongly-convex strongly-concave minimax games by taking the variational inequality formulation. By connecting momentum method with Chebyshev polynomials, we show that negative momentum accelerates convergence of game dynamics locally, though with a suboptimal rate. To the best of our knowledge, this is the \emph{first work} that provides an explicit convergence rate for negative momentum in this setting.
翻译:平滑的游戏优化最近吸引了人们对机器学习的极大兴趣,因为它概括了单一目标优化模式。 但是,由于不同玩家之间的互动关系,游戏动态更为复杂,因此与最小化截然不同,给算法设计带来了新的挑战。 值得注意的是,人们已经表明,由于它能够减少游戏动态的振动,因此倾向于消极势头。 然而,消极势头的趋同率只在简单的双线游戏中形成。 在本文件中,我们通过采用变式不平等公式,将分析扩大到平滑的和强烈相近的迷你马克思游戏。 通过将动力法与切比谢夫多式游戏联系起来,我们表明,负面势头加速了本地游戏动态的趋同,尽管其速度不理想。 据我们所知,这是为这一环境的负动力提供明确的趋同率的 memph{ 第一次工作} 。