Bayesian Network Games
Abstract: In many situations agents in a network are affected by the behavior of the whole society while only being able to access information in their local neighborhoods. A rational behavior model inthese situations is to employ a strategy that maximizes expected utility with respect to the information available and a model on the expected strategies of other agents. When the model on the expected strategies of other agents is correct, the strategy profile is defined as the Bayesian Nash equilibrium. We call the repeated play of the equilibrium strategy in an environment with uncertainty and local information exchanges at each stage a Bayesian Network game. In this talk, we first discuss asymptotic properties of these games and for the particular case of quadratic payoffs we introduce the Bayesian Quadratic Network Game filter that agents can run locally to update their beliefs, select corresponding optimal actions, and eventually learn a sufficient statistic of the network's state. In the second part of the talk, we present an application of the class of games considered and the filter to a demand side management model in smart grids with heterogeneous users and renewable energy supply uncertainty. Open research directions are discussed to close the talk.
Bio: Ceyhun Eksin received the B.Sc. degree in Control Engineering from Istanbul Technical University in 2005. He received his M.Sc. degree in industrial engineering from Boğaziçi University, Istanbul, in 2008.During this period, he also spent one semester at Technische Universiteit Eindhoven as an Erasmus exchange student. He joined the department of Electrical and Systems Engineering at the University of Pennsylvania as a Ph.D. student in 2008. His research is on modeling, analysis and design of networked multi-agent systems. Since 2011, he has been focusing on distributed network optimization and learning in games over networks with applications to smart grids.