Layered Learning and Flexible Teamwork in RoboCup Simulation Agents

RoboCup was introduced as a challenge area at IJCAI-97. We have been actively pursuing research in this area and have participated in the RoboCup competitions, winning the RoboCup-98 and RoboCup-99 simulator competitions. In this paper, we report on the main technical issues that we encountered and addressed in direct response to the learning and teamwork challenges stated in the IJCAI-97 challenge paper. We describe “layered learning” in which off-line and online, individual and collaborative, learned robotic soccer behaviors are combined hierarchically. We achieve effective teamwork through a team member agent architecture that encompasses a “flexible teamwork structure.” Agents are capable of decomposing the task space into flexible roles and can switch roles while acting. We report detailed empirical results verifying the effectiveness of the learned behaviors and the components of the team member agent architecture.