Convergent gradient ascent in general-sum games

Abstract

In this work we look at the recent results in policy gradient learning in a general-sum game scenario, in the form of two algorithms, IGA and WoLF-IGA. We address the drawbacks in convergence properties of these algorithms, and propose a more accurate version of WoLF-IGA that is guaranteed to converge to Nash Equilibrium policies in self-play (or against an IGA learner). We also present a control theoretic interpretation of variable learning rate which not only justifies WoLF-IGA, but also shows it to achieve fastest convergence under some constraints. Finally we derive optimal learning rates for fastest convergence in practical simulations.