Correlating internal parameters and external performance: Learning Soccer Agents

Abstract

We have developed a Soccer simulator in Java, which allows us to evaluate agent behavioral strategies for effectively collaborating with other team members while countering threats posed by the opposing team. Whereas action selection mechanisms to decide when to shoot, pass, dribble, guard opponent, tackle, etc. are important from a player's viewpoint, learning the capabilities of individual opponents from repeated encounters can provide critical information for the success of a team. Each agent is described by a set of skill levels for different soccer skills. A soccer agent uses its incomplete perceptions, model of the changing environment, knowledge of the skill levels of its own team players, and estimated skill levels of the opponent team players to select the most prudent action. In this paper, we identify learning opportunities for soccer agents, and investigate some of these possibilities in detail. Initial experimental results demonstrate the advantage of learning agents.