Initializing and discretizing for Q-Learning