The Future of Small-Scale Multi-agent Competitions

DIMACs, Rutgers University (Directions)

Jul 27th, 2012, 9 AM-5 PM

Room 433

Schedule

Summary

Competitions are a growing part of the research in multi-agent systems. Robocup, the Trading Agent Competitions, and the Annual Computer Poker Competition have stirred a great deal of interest in multiagent systems. They have pushed the envelope, in the computation of Nash equilibria, robotics, and have taught countless competitors experience in building complex systems on a deadline. However, there are fundamental questions in perscriptive, multi-agent systems that remain unanswered. Moreover, the complexity of these tasks usually means that it is difficult to get to the multi-agent systems aspects of the problems.

An alternative to this is to design simple competitions. For example, there have been competitions on the Prisoner's Dilemma, Rock-Paper-Scissors, and the Lemonade Stand Game. The questions that these competitions ask are often more philosophical than computational, but they get at a heart of a problem that increasingly faces researchers: I have a wealth of computational resources, so I can sort, or calculate posteriors, et cetera. But what do I want to achieve and what do I believe? A fundamental problem is that the answer to these two questions are impossible to answer in a vacuum.

Competitions can help develop priors and objectives. There are questions as how specific these are to the competition at hand. Good competitions will help us develop methods that extend to new domains. If these methods extend, they will likely result in publications.

This meeting is to discuss the future of small scale competitions. In particular, we will discuss some of the results from the Lemonade Stand Game Competition, and talk about directions for this competition. We will also have invited speakers on the theory of mind and communication, to understand what problems exist out there for working in multi-agent systems. The three big questions that have been presented so far are questions of human players in the competition, knowledge of other's utilities, and explicit communication.