Simultaneous Abstract and Concrete Reinforcement Learning