Defining and using ideal teammate and opponent agent models: a case study in robotic soccer

摘要

A common challenge for agents in multiagent systems is trying to predict what other agents are going to do in the future. Such knowledge can help an agent determine which of its current action options are most likely to help it achieve its goals. Ideally, an agent could learn a model of other agents' behavior patterns via direct observation of their past actions. However, that is only possible when agents have many repeated interactions with one another. We explore the use of agent models in an application where extensive interactions with a particular agent are not possible, namely robotic soccer. In robotic soccer tournaments, such as RoboCup (Kitano et al., 1997), a team of agents plays against another team for a single, short (typically 10-minute) period. The opponents' behaviors are usually not observable prior to this game and there are not enough interactions during the game to build a useful model. We introduce "ideal-model-based behavior outcome prediction" (IMBBOP). This technique predicts an agent's future actions in relation to the optimal behavior in its given situation, This optimal behavior is agent-independent and can therefore be computed based solely on a model of the world dynamics. IMBBOP does not assume that the other agent will act according to the theoretical optimum, but rather characterizes its expected behavior in terms of deviation from this optimum.