Neural architecture search with reinforcement learning. We've used an actor critic method internally with good results as well.