RD 2 : Reward Decomposition with Representation Decomposition