On learning coordination among soccer agents

摘要

The paper applies machine learning to learn coordination between two soccer agents. The prime focus is on designing the role of a support player whose job is to support the attack player as the attacker dribbles the ball towards the opponent goal. The traditional way of designing coordination among players is via manual scripting. This, however, requires a detailed specification of routines related to path planning, team formation, collision avoidance, etc. In this paper, we learn the coordination skill by observing log files of the matches played by one of the better teams in the RoboCup Soccer 3D Simulation league. For effective learning, we have extracted knowledge from log files by defining events that relates to a team's strategy. The coordination skill is learned as classification and regression models using neural networks. The goal is to predict the next position of the support robot based on the game state and other relevant variables. Experiments have shown very promising results.