Improving Convergence Rates in Multiagent Learning Through Experts and Adaptive Consultation

We present a multiagent learning algorithm with guaranteed convergence to Nash equilibria for all games. Our approach is a regret-based learning algorithm which combines a greedy random sampling method with consultation of experts that suggest possible strategy profiles. More importantly, by consulting carefully chosen experts we can greatly improve the convergence rate to Nash equilibria, but in the case where the experts do not return useful advice, we still have guarantees that our algorithm will eventually converge. The goal of our work is to bridge the gap between theoretical and practical learning, and we argue that our approach, FRAME, can serve as a framework for a class of multiagent learning algorithms.