Sample-efficient Reinforcement Learning via Difference Models