Learning Tree Interpretation from Object Representation for Deep Reinforcement Learning