Multi-Agent Control: Mechanism Design and Adaptive Learning

University Of California-Berkeley, Berkeley CA

Investigators

Abstract

ECS-0424445 Multi-agent control: mechanism design and adaptive learning PI Pravin Varaiya, University of California, Berkeley, varaiya@eecs.berkeley.edu Abstract We seek to study control strategies for multi-agent systems in which agents have different information and possibly different objectives. We organize these decision-making settings as cooperative or non-cooperative games. The objective is to design the 'rules of the game' in such a way that agents acting selfishly on their own behalf arrive at promoting socially desirable strategies. We will focus on 'combinatorial auctions' as the paradigmatic game. Second, we want to study control strategies with low computational complexity, which rules out dynamic programming or global optimization. The approach is to reduce complexity by resoring to 'learning', based on simulation or experiments, in order to estimate the value of a particular strategy. Examples to test our approach will be drawn from bandwidth trading in communication networks.