Co-evolutionary Design of Team Level Play in Soccer Server

Recently, RoboCup soccer simulation has been regarded as a good benchmark problem for multiagent researches. Soccer agents have to make decision based on visual and auditory information, which are sent from the soccer server. In order to develop a strong team, we have to design decision-making process of each player agent. However, it is very difficult for us to design the decision-making processes in detail, because we don't know what actions of each player are effective for the team. In this paper, we attempt to apply co-evolutionary method, which is one type of analogies of evolution, to improve the team play. Agents have hand coded basic skills, which include dribble, shoot, pass etc. Agents already can play autonomously and independently. Individual agent skills are characterized by some parameters. By coevolving teams with these parameters, we obtained relatively interesting teams, in which players behave cooperatively in order to win the soccer game. From some experiments, we discuss what teams are evolved.