Q&A with 2019 Innovator Under 35 Noam Brown


The MIT Technology Review has named Facebook AI Research Scientist Noam Brown one of this year’s Innovators Under 35 for his research on AI and games. The award recognizes talented young innovators by region for their potential to transform the world through their contributions to science and technology. Previous winners have included Mark Zuckerberg in 2007 and Google cofounders Larry Page and Sergey Brin in 2002.

Noam Brown is best known for his work on the poker-playing AI system Libratus, which he developed at Carnegie Mellon University with his PhD adviser Tuomas Sandholm in 2017. Libratus was the first AI to defeat top poker players in two-player no-limit Texas Hold’em.

Brown took a moment to share thoughts about his past, current, and future research on game-playing bots, as well as the possible applications of his research. He also shared advice to anyone looking to pursue research in AI and game theory.

Q: Describe some key highlights from your research work. What projects are you proudest of?

Noam Brown: The project I am proudest of is definitely Libratus, which beat top human poker professionals in two-player no-limit poker. This was a challenging problem that had existed for decades. Poker involves hidden information, which makes it resistant to other AI techniques that were successful in games like chess and Go. The first four years of my PhD were focused on figuring out how to crack this problem by building off of decades of research from previous researchers in the field. Eventually it became clear to me what the path to victory was, but it still took another year of implementation to get it all working and to actually beat top humans in the game. It was very rewarding to see all the pieces come together into an actual agent that could beat top humans.

Q: What led you to focus on AI and game theory?

NB: My original goal was to get a PhD in economics, but after spending a couple of years working in the economics research field, I realized I also wanted to build things. Economics doesn’t provide that opportunity as much as computer science, and especially AI, does. I had always been interested in both AI and game theory, and the economics angle of game theory seemed like a natural fit for me given my background.

Q: What does research in this area look like five years from now? What problems still need to be solved?

NB: There has been tremendous progress in recent years on AI for purely adversarial zero-sum games like checkers, chess, Go, poker, Starcraft, and Dota 2. But the real world isn’t zero-sum. Researchers still don’t know how to tackle AI for partly cooperative and partly adversarial settings, like negotiations. The state of the art in this area is way behind human performance. I think this will be a major area of AI research in the next five years, and it is an area that can have tremendous real-world impact.

Q: How does research on game-playing bots tie to real-world applications? How does it tie to other fields of AI research?

NB: While Libratus plays poker, the techniques are not limited to poker. Poker is just a benchmark that allows us to compare the performance of these techniques with the peak of human ability. That’s true for other AI milestones in games as well. The research I’m doing is really about developing AI techniques that can handle strategic reasoning and hidden information in multi-agent settings. This is very important because most real-world strategic interactions involve some amount of hidden information. If an AI agent is to act and to help people in the real world, it must be able to cope with hidden information.

Q: What surprised you most about how the research in this field has evolved? What’s been harder or easier than you might have expected?

NB: As with most research, it was very hard to predict what the “magic ingredient” would be that would lead to superhuman performance. My early PhD research was focused on techniques that seemed like good ideas at the time but ultimately didn’t make a huge difference in performance. But if you keep making good shots, eventually you’ll score a goal.

Q: What would you say to other AI researchers or students who are considering focusing on game theory and AI?

NB: This is a very exciting time to be in this research area. Research on imperfect-information games was historically a bit outside of the mainstream of AI, but recent results have shown convincingly that it holds answers to questions that have vexed AI researchers for decades. This is an underexplored field with a lot left to be done. But most important, I think the key to doing good research is loving what you do.