Operations Management and Econometrics
Distributed Optimisation by Learning in Games with Pure Nash Equilibria
Dr Archie Chapman, The Discipline of Operations Management and Econometrics
15th Apr 2011 11:00 am - Room 498, Merewether Building
An emerging framework for optimisation in large distributed problems is that of multi-agent systems, in which control of a system is partitioned among several autonomous decision makers. Within this context, potential games are used as a design template for constructing agents' utility functions, resulting in games with pure strategy Nash equilibria that can be solved for using iterative learning algorithms. This presentation investigates the convergence properties of one such iterative learning procedure, called fictitious play, in repeated normal form games. Specifically, using methods from the theory of differential inclusions and stochastic approximations, we analyse the rest points of fictitious play. We also discuss how to extend fictitious play's use to solve games with unknown rewards or perturbations to action observations, and also to multi-agent sequential decision-making problems (such as Decentralised-POMDPs).