Even when it comes to Texas Hold’em, card sharks can’t out-bluff the AI
Researchers from Facebook and Carnegie Mellon University have built artificial intelligence that consistently defeats expert human players in six-player, no-limit Texas Hold’em poker. The research is being jointly published in the publication Science today.
Winning in multiplayer poker has been seen as an especially tough challenge for AI. It involves contemplating hidden information like an opponent’s hand, as well as bluffing based on one’s own hand. AI has defeated single opponents in poker in recent years, but Facebook and Carnegie Mellon say their success marks the first time an AI has defeated human poker pros in full-scale poker with multiple opponents.
Facebook and Carnegie Mellon researchers created an AI bot called Pluribus to confront elite human poker players. Pluribus defeated the humans in games where five AIs and one human played and in games involving Pluribus and five human players. Facebook says that if each chip had been worth a dollar, Pluribus would have won roughly a thousand dollars an hour, or roughly $5 a hand, playing against five human players.
The Pluribus AI is the big brother to the Libratus bot developed at Carnegie Mellon that defeated humans at two-player Texas Hold’em poker back in 2017. Pluribus uses a new algorithm that lets it evaluate its options a few moves ahead in the game, as opposed to working through options for the entire game. It also uses a new self-play strategy, in which it plays against copies of itself without interference from humans.
“The AI starts from scratch by playing randomly and gradually improves as it determines which actions, and which probability distribution over those actions, lead to better outcomes against earlier versions of its strategy,” the Facebook researchers say in a blog post.
The researchers say these advancements allow for an AI that can be trained using minimal resources, including less than $150 worth of cloud computing time.
Of course, Facebook’s and Carnegie Mellon’s interest in the research goes beyond poker. More serious applications also involve multiplayer scenarios with hidden information. Facebook says these may include things like “taking action on harmful content and dealing with cybersecurity challenges, as well as managing an online auction or navigating traffic.”
This success for Facebook and Carnegie Mellon is the latest in a growing line of AI victories over humans in gaming scenarios.
IBM’s Deep Blue supercomputer famously defeated Garry Kasparov, then the world chess champion, in 1997. Google sibling DeepMind created an AI called AlphaGo that defeated Korean Go champion Lee Sedol in 2016. Deepmind’s AlphaZero AI defeated the world’s best chess-playing program in 2017 after teaching itself how to play the game.