Layered decision-making and planning in ShaoLing team

Abstract

Multi-agent systems in complex, real-time domains require agents to act effectively both autonomously and as a part of a team. Real-time, noisy, collaborative and adversative domains are four characteristics in multi-agent systems. As such, simulated robotic soccer-the RoboCup soccer server-has served as our research test bed; it is a domain which fits all the above characteristics. The paper defines a team member agent architecture which fits the above characteristics. Especially, flexible roles have been used to fit the characteristic of collaborative in robotic soccer. The decision-making is classified in three layers: strategy, tactics, and individual execution. Machine learning and planning are used in the different layers of the decision-making. Using the techniques described, the ShaoLing team was built, which has won a championship in the first competition on RoboCup of China.