An architecture for co-evolving agents in the ROBOCUP Simulation League

Abstract

Robotics is undoubtedly one of the major application areas of autonomous intelligent agents and multiagent systems research. The RoboCup simulation league provides an environment for the development of multiagent systems, composed of eleven autonomous software agents that are able to play soccer in a simulation system (soccer server) which enables teams to compete against each other. Under this complex, distributed, real-time, noisy and low bandwidth communication setting, agents must be able to reason and cooperate in order to win the game. In this work, a high-level architecture to the development of a successful team is proposed, whose main features are the action selection model based on MultiLayer Perceptrons and the learning process supported by Evolutionary Algorithms. A model based on the co-evolution of team strategies and player skills is proposed. The implementation of this architecture takes as its base available software from the winner of the 2003 competition, the UvA Trilearn team.