A Multi-agent Algorithm for Robot Soccer Games

In this paper, we demonstrate the design and implementation of a multi-agent algorithm including a set of strategies for robot soccer games in Fira [1] simulation league. Soccer games are pseudo matches performed on a simulator for testing real-timedecisionmaking schemes for cooperative multi-agent systems. One of most important issues in the soccer game is how to pass a ball among teammates. Many proposed strategies [2][3], based on reinforcement learning methods [21] [22][23][24] or heuristic schemes to infer the possible risks and costs, do not consider ball passing as a multi-criteria optimum problem. To determine whether passing the ball to a teammate in the goal area, a multi-criteria decision-making strategy is proposed by considering the multiple criteria, the angle of the ball to the two goalposts, the distance between the ball and the goal, and the position of the enemies. As to the formation of a team, we introduce an idea, called virtual forces [18], to suggest the optimum positions of team players. The optimum positions will depend on the real-time conditions such as the ball position and the positions of all enemy players. To evaluate the proposed strategies, we take the team UvA-Trilearn[4], the world champion in 2002, as an opponent. The results show that the proposed strategies attain better performance than UvA-Trilearn.