Fitted Q-Learning for Relational Domains