Visual Pretraining for Deep Q-Learning