Karlsruhe Brainstormers 2000 Team Description

The main motivation behind the Karlsruhe Brainstormer’s effort in the robotic soccer simulation league of the RoboCup project is to develop and to apply Reinforcement Learning (RL) techniques in complex domains. Our long term goal is to have a learning system which is able to learn by itself the best winning behaviour. The soccer simulation domain allows more than (108x50)23 different positionings of the 22 players and the ball - the complete state space considering object velocities and player’s stamina is magnitudes larger. In every cycle, an agent can choose between more than 300 basic commands (parametrized turns and dashes), which makes a choice of 30011 joint actions for the team per cycle! A problem of such complexity is a big challenge for today’s RL methods; in the Brainstormers project we are investigating methods to practically handle learning problems of such size.