Learning competition in robot soccer game based on an adapted neuro-fuzzy inference system

Abstract

RoboCup is a worldwide popular research domain. Because of the complexity of the systems, how to describe cooperation and competition between agents is a great challenge in the RoboCup Simulation Game. On one hand, the rich experience of a human soccer player is of great service to the robot players. On the other hand, the difference between the simulation game and the real game make it a must to fit the transcendental knowledge into the new environment. Commonly used reinforcement learning is weak in utilizing transcendental knowledge, thus is limited in complex multi-agent system learning problems. The paper puts forward a supervised learning method on the basis of the adapted neuro-fuzzy inference system (ANFIS) for mapping the competition among the robots. This method can build an ANFIS according to experts' knowledge, and with data obtained in the simulation environment. It can establish a correct map to describe the competition among the robots. We use this method to describe the antagonization between the shooter and goalie, and have successfully applied it in the RoboCup Simulation Game to build the champion team in RoboCup 2000 of China.