DRU Learning and Social Efficiency in Large Games

Georgia Tech Research Corporation, Atlanta GA

Investigators

Abstract

Game theorists, economists, and computer scientists approach decision making from very different perspectives. While computer scientists focus on complexity aspects, in a worst-case framework, game theorists and economists focus on coordination and strategic aspects, typically employing Bayesian models. This project attempts to integrate these two approaches by formulating and analyzing practical models that weaken or eliminate prior distribution assumptions, taking into account strategic behavior and yielding efficient outcomes. The focus is on large games, that is games with many players, both because of the intrinsic interest in them and because many of the central issues are more easily addressed in such games. Many important social and economic problems with many agents, such as economic trade, auctions, transportation, and communication, can be modeled as large games. The research included in the project aims at extending our understanding of when equilibria of large games are robust, in the sense that optimal play does not depend on the fine details of the game such as sequencing of moves, communication, etc. The investigators propose to study when players who learn over time in such games might converge to Nash equilibria. Through a series of examples we also illustrate how one might design mechanisms that avoid the sometimes very strong informational and communicational demands implicit in existing mechanism design.