Acquiring Multiagent Cooperative Behavior in the RoboCup Soccer Simulation

Abstract

The RoboCup Soccer Simulation is a research platform for multiagent systems and artificial intelligence. It is based on the RoboCup Soccer 2D Simulator, which enables two teams of 11 autonomous player agents and an autonomous coach agent to play a game of soccer with highly realistic rules and game play. The soccer simulation has devoted more attention to teamwork techniques than to robot control techniques. Therefore, we can avoid the burdens of developing and maintaining mechanical devices and also developing complex robot control tasks such as bipedal walking. These characteristics enable us to concentrate on research efforts related to teamwork. In order to acquire an appropriate teamwork, we need cooperative behavior models for coordinating teammate players, the methods to analyze opponent team behavior, and adaptaion techniques to the opponent strategy. As successful approaches to the first problem, we have proposed two important methods, a positioning model using triangulation and a sequential action planning using a tree-search method. These methods have already been implemented in our released code. In this talk, I introduce the effectiveness of these methods and the overview of our software. The recent research/development topics in the soccer simulation, especially opponent modeling and adaptation technique, are also introduced.