Google Trains Reinforcement Learning Agents to Ask the Right Questions