Efficient abstraction selection in reinforcement learning