Active Deep Q-learning with Demonstration