Learning State Representation for Deep Actor-Critic Control