Karlsruhe brainstormers - A Reinforcement Learning approach to robotic soccer

Abstract

Our long-term goal is to build teams of agents where the decision making is based completely on Reinforcement Learning (RL) methods. It requires an appropriate modelling of the learning task and the paper describes how robotic soccer can be seen as a multi-agent Markov Decision Process (MMDP). It discusses how optimality of behaviours of agents can be defined and what difficulties one encounters in developing concrete algorithms which are supposed to reach such optimal agent/team policies. We also give an overview of already incorporated algorithms in our 'Karlsruhe Brainstormers' simulator league team and report some results on learning of offensive team behaviour. © 2002 Springer-Verlag Berlin Heidelberg.