“Reinforcement learning breakthroughs from games like StarCraft and Atari have not been widely deployed in the real world due to challenges in collecting sufficient data and defining clear action spaces and reward functions.”