Q-learning with Nearest Neighbors