We propose a reinforcement learning solution to the \emph{soccer dribblingtask}, a scenario in which a soccer agent has to go from the beginning to theend of a region keeping possession of the ball, as an adversary attempts togain possession. While the adversary uses a stationary policy, the dribblerlearns the best action to take at each decision point. After definingmeaningful variables to represent the state space, and high-level macro-actionsto incorporate domain knowledge, we describe our application of thereinforcement learning algorithm \emph{Sarsa} with CMAC for functionapproximation. Our experiments show that, after the training period, thedribbler is able to accomplish its task against a strong adversary around 58%of the time.
展开▼
