Prof. Maryam Kamgaprour, EPFL

Title: Learning equilibria in dynamic games with zeroth-order information

Abstract: A rising challenge in control of large-scale systems such as the energy and the transportation networks is to address autonomous decision making of interacting agents. Game theory provides a framework to model and analyse this class of problems. In several realistic applications, such as power markets, each player has partial information about the cost functions and actions of other players. Thus, a learning approach is needed to design optimal decisions for each player.

My talk will summarize the work with my research group on addressing decentralized learning in games under zeroth-order online information. Leveraging the understanding we have gained from design of algorithms with provable convergence for static games, I will discuss additional challenges that arise in the dynamic game setting. For the latter class of games, I will discuss our single time-scale approach to efficiently learn Nash equilibria in zero-sum Markov games. Furthermore, I will sketch our approach for learning correlated equilibria in general sum Markov games. Bringing in simulation results of our algorithms on energy and transportation systems, based on state-of-the-art simulation models of the systems, I will discuss their potential applicability in the real world.