Transfer with Model Features in Reinforcement Learning