Making Deep Q-learning methods robust to time discretization