In this paper we study a nonconvex-strongly-concave constrained minimax problem. Specifically, we propose a first-order augmented Lagrangian method for solving it, whose subproblems are nonconvex-strongly-concave unconstrained minimax problems and suitably solved by a first-order method developed in this paper that leverages the strong concavity structure. Under suitable assumptions, the proposed method achieves an \emph{operation complexity} of $O(\varepsilon^{-3.5}\log\varepsilon^{-1})$, measured in terms of its fundamental operations, for finding an $\varepsilon$-KKT solution of the constrained minimax problem, which improves the previous best-known operation complexity by a factor of $\varepsilon^{-0.5}$.
翻译:本文研究一个非凸-强凹约束极小极大问题。具体而言,我们提出一种求解该问题的一阶增广拉格朗日方法,其子问题为非凸-强凹无约束极小极大问题,并适合采用本文开发的一种利用强凹结构的一阶方法进行求解。在适当假设下,所提方法在基本操作次数上达到 $O(\varepsilon^{-3.5}\log\varepsilon^{-1})$ 的**操作复杂度**,以找到该约束极小极大问题的一个 $\varepsilon$-KKT 解,这比先前已知的最佳操作复杂度改进了 $\varepsilon^{-0.5}$ 倍。