Low Dimensional State Representation Learning with Reward-shaped Priors