Coordination and Adaptation in Impromptu Teams

Coordinating a team of autonomous agents is one of the major challenges in building effective multiagent systems. Many techniques have been devised for this problem, and coordinated teamwork has been demonstrated even in highly dynamic and adversarial environments. A key assumption of these techniques, though, is that the team members are developed together as a whole. In many multiagent scenarios, this assumption is violated. We study the problem of coordination in impromptu teams, where a team is composed of independent agents each unknown to the others. The team members have their own skills, models, strategies, and coordination mechanisms, and no external organization is imposed upon them. In particular, we propose two techniques, one adaptive and one predictive, for coordinating a single agent that joins an unknown team of existing agents. We experimentally evaluate these mechanisms in the robot soccer domain, while introducing useful baselines for evaluating the performance of impromptu teams. We show some encouraging success while demonstrating this is a very fertile area of research.