The influential work of Bravo et al. (2018) shows that derivative free play in strongly monotone games converges at the rate $O\left(\frac{d^2}{T^{1/3}}\right)$, where $T$ is the time horizon and $d$ is the ambient dimension. This note shows that the convergence rate is actually $O\left(\frac{d^2}{T^{1/2}}\right)$, thereby matching the classical guarantees for derivative free methods in optimization. The argument we present is elementary, simply interpreting the method as stochastic gradient play on a slightly perturbed strongly monotone game.
翻译:Bravo等人的有影响力的工作(2018年)表明,极强单调游戏中的衍生自由游戏以美元(left)(\\frac{d ⁇ 2 ⁇ T ⁇ 1/3 ⁇ right)的汇率趋同,美元是时间范围,美元是环境维度。本说明表明,趋同率实际上是美元(left)(\frac{d ⁇ 2 ⁇ 2 ⁇ T ⁇ 1/2 ⁇ right),从而与衍生自由方法在优化中的典型保障相匹配。我们提出的论据是基本的,只是将这一方法解释为在略受干扰的极强单调游戏中进行随机梯度游戏。