Variance-based stochastic projection gradient method for two-stage co-coercive stochastic variational inequalities

Bin Zhou, Jie Jiang, Yongzhong Song, Hailin Sun

Abstract: The existing stochastic approximation (SA)-type algorithms for two-stage stochastic variational inequalities (SVIs) are based on the uniqueness of the second-stage solution, which restricts the use of those algorithms. In this paper, we propose a dynamic sampling stochastic projection gradient method (DS-SPGM) for solving a class of two-stage SVIs satisfying the co-coercive property. With the co-coercive property and the dynamic sampling technique, we can handle the two-stage SVIs when the second-stage problem has multiple solutions and achieve the rate of convergence with \(\varvec{O}(\varvec{1/\sqrt{K}})\). Moreover, numerical experiments show the efficiency of the DS-SPGM.