Presenter: Noam Brown
Topic: Learning to Cooperate and Compete in Diplomacy
The EconCS Group holds an Economics and Computer Science research seminar each semester.
AI has made incredible progress in purely adversarial games such as chess, go, and poker. However, the real world involves a complex mixture of cooperation and competition, sometimes with irrational or suboptimal participants, and in these settings past AI techniques fall apart. For this reason, Diplomacy, a popular game focused on negotiation and alliance-building, has recently emerged as a grand challenge for AI that requires radically different techniques compared to prior games and has major implications if AI algorithms eventually succeed. In this talk I will describe Diplomacy and cover recent research results from FAIR, DeepMind, and MILA that make progress on the no-communication version of this game. In particular, I will present results strongly suggesting that self play reinforcement learning alone is incapable of reaching superhuman performance in this game, unlike in purely adversarial and purely cooperative games, but that combining self play with supervised learning on human data achieves a strong human level of play in no-communication Diplomacy. Finally, I will conclude with thoughts on the challenges that await as research shifts from the no-communication version of Diplomacy to versions that involve private communication between the players.