Diffusion optimistic learning for min-max optimization

H. Cai, Sulaiman A Alghunaim, Ali H. Sayed

Abstract:

This work introduces and studies the convergence of a
stochastic diffusion-optimistic learning (DOL) strategy for
solving distributed nonconvex (NC) and Polyak–Lojasiewicz
(PL) min-max optimization problems. Problems of this type
are of interest due to a wide range of applications, including
in generative adversarial networks (GANs), adversarial machine learning, and reinforcement learning. We prove that
the DOL algorithm approaches an $ε$-stationary point through
cooperation among agents following a left-stochastic communication protocol. The good performance of the proposed
algorithm is illustrated by means of computer simulations.