A new perspective to the keepaway soccer: the takers

Keepaway is a sub-problem of RoboCup Soccer Simulator in which 'the keepers' try to maintain the possession of the ball, while 'the takers' try to steal the ball or force it out of bounds. By using Reinforcement Learning as a learning method, a lot of research has been done in this domain. In these works, there has been a remarkable success for the intelligent keepers part, however most of these keepers are trained and tested against simple hand-coded takers. We tried to address this part of the problem by using Sarsa(λ) as a Reinforcement Learning method with linear tile-coding as function approximation and used two different state spaces that we specially designed for the takers. As the results of the experiments confirm, we outperformed the hand-coded taker which results in creating a better trainer and tester for the keepers. Also when designing the new state space, we noticed that smaller state spaces can also be successful for this part of the problem.