| This work investigates the specialization of behavior of a full Robocup Soccer Simulation team that learns to play effective and skillful games over time using a machine learning technique known as reinforcement learning. Collaboration among the soccer agents and behavioral diversity depending on the strategy of playing are also addressed in the work. The full team that has been designed is a combination of predefined primary skills and learning of how to apply those skills. Since learning space for these soccer agents is huge and costly in terms of time and storage complexity, a partitioning technique was previously implemented to lessen the computational complexity. We evaluate the benefit of these techniques and tune parameters of the learning method to improve convergence upon an optimal policy. |
