A bot has beaten the world’s top poker players, a milestone showing that AI can work with incomplete information.
Libratus, a program created by researchers at Carnegie Mellon University, has demonstrated new capabilities for artificial intelligence after beating four poker champions in a 20-day tournament for the first time. For marketers, whose work involves situations that are much more similar to poker than to board games like chess and Go, the milestone could signal new possibilities for the use of AI.
Developed by Tuomas Sandholm, professor of computer science, and his PhD student Noam Brown, Libratus took on poker professionals Dong Kim, Jason Les, Jimmy Chou and Daniel McAulay and ended up winning more than $1.7 million in chips at Rivers Casino in Pittsburgh, Pennsylvania.
AI bots have beaten human experts at many tasks, from IBM Watson’s Jeopardy triumph in 2011 to DeepMind’s AlphaGo win in 2016. What makes this victory different is that the AI was able to use imperfect information to win. Poker is a complex game that requires intuition, reasoning and an ability to bluff. It’s different from other recreational games that AI has won in the past because an opponent’s hand is hidden and it is impossible to know with certainty what a player has.
In order to win, Libratus trained over several months using reinforcement learning, a type of trial and error learning in which the system plays against itself to help it learn. “We didn’t tell Libratus how to play poker. We gave it the rules of poker and said ‘learn on your own,’” Brown told The Guardian. “The bot started playing randomly but over the course of playing trillions of hands was able to refine its approach and arrive at a winning strategy.”
The real-world applications go far beyond poker. Tuomas Sandholm, the co-creator of Libratus, told the Telegraph that his invention “can be used in any situation where information is incomplete including business negotiation, military strategy, cyber security and medical treatment.”
Professor Nick Jennings, a computer science and engineering expert at Imperial College who holds a chair in artificial intelligence, confirmed this to the Innovation Group. Libratus could be used in game theory research, in which researchers analyze and model outcomes in a wide variety of real-world situations, he said.
“One could treat advertising as marketing as game, against an opponent with unknown preferences and resources, where the game is to make them make a purchase or view an ad for a period of time,” Jennings said. “The algorithms would be able to explore such strategies and learn from what others do … It could then learn which strategies are successful and deploy them.”
Jennings said that the Libratus system was mostly not specific to poker, which would make it fairly easy to adapt to other contexts. “It is a significant step forward for AI and shows it how it can perform effectively in real world situations in which there are inherent uncertainties and many possible courses of action,” he added.
For more on advances in artificial intelligence and evolving consumer attitudes toward them, read our Control Shift trend report.