Discovering state-of-the-art reinforcement learning algorithms