Theoretical guarantees for algorithms in multi-agent settings

• Authors: Martin Zinkevich;Avrim Blum
• Venue: Theoretical guarantees for algorithms in multi-agent settings
• Year: 2004

In this thesis we will develop and analyze algorithms in multi-agent settings. We focus on two areas: that of auctions, negotiation, and exchanges, which are of increasing interest to computer scientists with the rise of e-commerce, and that of repeated bimatrix games and multi-stage games, which provide an interesting test bed for agents that will be interacting with other intelligent agents. The main thrust of this thesis is designing algorithms for agents who are missing critical information about the environment or other agents (like what is going to happen in the future, or what the other agents' motivations are) that perform well compared to the optimal behavior which has this information. In the area of auction design, we consider online double auctions (exchanges) for commodities, where there are buyers and sellers trading indistinguishable goods who arrive and depart over time. We consider this from the perspective of the broker, who decides what trades occur and what payments are made. We also consider combinatorial auctions. We show a connection between query learning in machine learning theory and preference elicitation. We show how certain natural hypotheses spaces that can be learned with membership queries correspond to natural classes of preferences that can be learned with value queries. Finally, we consider repeated bimatrix games. One would like two reasonable agents who encounter each other many times in the same setting (e.g. a bimatrix game) to eventually perform well together. We show how a simple gradient ascent technique performs well in a bimatrix game, as well as in an arbitrary, online convex programming domain. One of the hardest parts of working in a domain with other intelligent agents is defining what it means to perform well and understanding what are the right assumptions to be made about the other agent. It is well known that there exists an algorithm that has no regret against an arbitrary algorithm. Furthermore, it is not hard to show that there exists an algorithm that achieves the minimum Nash equilibrium value against a no-external-regret algorithm. However, we show here that no algorithm can achieve both these guarantees. (Abstract shortened by UMI.)