Earlier this year, more than a dozen professional poker players participated in an unusual competition of Texas Hold'em. The veterans went against a relative newbie: an artificial intelligence-powered bot built by Facebook and Carnegie Mellon University.
AI has crushed professional players of chess and Go, both board games with straightforward rules. Poker, too, has clear rules. But it's considered trickier because you can't see an opponent's hand and it requires manipulating emotions through tactics such as bluffing. The contest added a layer of complexity, as each game featured six players, creating more sets of scenarios for the AI to manage.
None of that stopped the poker-playing bot Pluribus. The bot stomped on its human challengers, which included World Series of Poker and World Poker Tour champions. Researchers called the bot's performance "superhuman."
"This is the first time an AI bot has proven capable of defeating top professionals in any major benchmark game that has more than two players (or two teams)," Facebook said in a blog post.
Pluribus' dominance over the mere mortals represents a breakthrough that might lead to applications of AI in real-world situations. That's because we often deal with multiple people and unknown information when it comes to things like political campaigns, online auctions and cybersecurity threats. AI could help businesses come up with the best strategies to handle those situations, research scientists say.
"We're using poker as a benchmark for a measure in progress in this more complicated challenge of hidden information in a complex multiparticipant environment," said Noam Brown, a research scientist at Facebook AI Research. The research group, which works on advancing AI technology, is also teaching robots to walk on their own.
Brown built Pluribus, which means "more" in Latin, with Tuomas Sandholm, a CMU computer science professor whose team has studied computer poker for more than 16 years. The pair's findings were published in the journal Science on Thursday.
The researchers set up two experiments, one in which a single human played five copies of Pluribus, and another in which five humans played a single copy of the bot. In both cases, Pluribus clearly won.
In the first experiment, Darren Elias and Chris "Jesus" Ferguson, both American poker pros, played 5,000 hands each against five copies of the AI bot. Elias holds the record for most World Poker Tour titles and Ferguson has won six World Series of Poker events. The humans played from their home computers.
Both players were offered $2,000 to participate in the Texas Hold'em game. To encourage them to bring their best game, players could win an extra $2,000 if they performed better against the AI than the other human poker player.
Overall, Pluribus beat the players by an average of 32 milli big blinds (mbb) per game. The big blind is a forced bet in Texas Hold'em, and the milli big blind is a measurement used to compare performance.
In the other experiment 13 players, who've all won more than $1 million each professionally, challenged the AI bot. Pluribus went against five human players at a time over 12 days and played 10,000 hands.
Pluribus won an average of 48 milli big blinds per game. If each chip was worth $1, the bot would've won $1,000 per hour playing against five humans, a Facebook blog post said. The win rate signals that the AI bot is "stronger than the human opponents," the research paper said.
"Sometimes, even if you're a bad player you're going to beat the world's best player just because you have better odds," Sandholm said. "We don't want to measure that luck factor. We want to really measure the skill factor."
Pluribus came up with a strategy for Texas Hold'em from scratch by playing against copies of itself. The bot also used a new algorithm that allowed it to examine its options a few steps ahead rather than at the end of the game.
The bot threw off its human competition by using moves humans typically avoid. For example, the bot placed more "donk bets" than humans. That's a bet at the beginning of a round after the previous one ended in a call.
"Its major strength is its ability to use mixed strategies. That's the same thing that humans try to do," Elias said in a statement. "It's a matter of execution for humans -- to do this inand to do so consistently. Most people just can't."