Machine Learning Approaches for Attacking Behaviors in Robot Soccer

Abstract

In this work, Machine Learning approaches were applied to attacking behaviors in RoboCup Small-Size League autonomous robot soccer. Neural networks were used in order to get a binary prediction of an attacking action’s success, while deep reinforcement learning was leveraged to learn low level skills which control the robot’s wheel speeds and kicker. A trained neural network was used to predict whether a shot would be successful, improving the number of goals scored by the attacking behavior by 84 to 186%. The reinforcement learning methodologies used in this work produced behaviors which were efficient in speed, beating manually programmed behaviors in time taken, but can benefit from future refinements to improve accuracy in shooting towards goal.