Computing Equilibria in Stochastic Nonconvex and Non-monotone Games via Gradient-Response Schemes

Abstract: We consider a class of smooth $N$-player noncooperative games, where player objectives are expectation-valued and potentially nonconvex. In such a setting, we consider the largely open question of efficiently computing a suitably defined {\em quasi}-Nash equilibrium (QNE) via a single-step gradient-response framework. First, under a suitably defined quadratic growth property, we prove that the stochastic synchronous gradient-response ({\bf SSGR}) scheme and its asynchronous counterpart ({\bf SAGR}) are characterized by almost sure convergence to a QNE and a sublinear rate guarantee. Notably, when a potentiality requirement is overlaid under a somewhat stronger pseudomonotonicity condition, this claim can be made for NE, rather than QNE. Second, under a weak sharpness property, we show that the deterministic synchronous variant displays a {\em linear} rate of convergence sufficiently close to a QNE by leveraging a geometric decay in steplengths. This suggests the development of a two-stage scheme with global non-asymptotic sublinear rates and a local linear rate. Third, when player problems are convex but the associated concatenated gradient map is potentially non-monotone, we prove that a zeroth-order asynchronous modified gradient-response ({\bf ZAMGR}) scheme can efficiently compute NE under a suitable copositivity requirement. Collectively, our findings represent amongst the first inroads into efficient computation of QNE/NE in nonconvex settings, leading to a set of single-step schemes that are characterized by broader reach while often providing last-iterate rate guarantees. We present applications satisfying the prescribed requirements where preliminary numerics appear promising.