Between MDPs and Semi-MDPs: A Framework for Temporal Abstraction in Reinforcement Learning