How Vegas Beat the Scientists Who ‘Solved’ Poker to the Punch

Cepheus, the Canadian artificial intelligence that “solved” Texas hold ‘em should stay off the casino floors.

When the University of Alberta research team behind Cepheus, the poker-playing artificial intelligence, announced a week ago that they had “solved” limit heads up Texas hold ‘em, the news spread fast in the non-poker world. The idea of a bot capable of winning – incapable of losing, really – being made available to the public was worth a screaming headline and a pull quote. Most came from Dr. Michael Bowling, a University of Alberta computer scientist who was braced for the inevitable moment the gambling industry would lash out.

“We don’t play for money,” Dr. Browning told Maxim. “But I was still expecting that I would receive a lot of hate mail.” 

Specifically, Bowling was braced for a reaction from the game makers behind the digital heads up hold ‘em machines common in Las Vegas casinos (heads up, or two-person, hold ‘em is rarely played on floors). He and his team had toyed with the idea of using the AI to play them, a publicity stunt they assumed would make them a bit of cash and get them tossed out of any casino on the strip. Cepheus is, after all, optimized within the margin of error of perfect, fully cognizant of 100,000,000,000,000 possible variations. But they were Canadian – and more interested in the practical application of imperfect information games, like strategic security and insulin monitoring – so they gave their opportunity to bring down the house a miss. Bowling isn’t really the gambling type, but he’s prone to a bit of academic boasting: “I don’t think the machines in Vegas are approaching the strength of our program.”

Pre-flop strategy visualizations from Cepheus

“It would have been funny,” says Gregg Guffria, the owner of G2 Game Design, which makes the machines Bowling had toyed with toppling. “I’d love them to use their theories and come to Vegas. The casinos would have loved it too. Our guy would chew theirs up.”

Guffria’s guy is the AI that governs the brand new hold ‘em machines now on the floor in the Bellagio, the Aria, and the Venetian. Hell, it governs some of the older machines as well and, according to Giuffria, it protects the house from any assault launched by tenured Canadians. The pokerbot is a superhero – complete with origin story. The bot was actually born in a secret government facility in Kjeller, Norway, where a researcher named Fredrik Dahl was using artificial intelligence to build out simulate combat missions. Because gambling games can be, as Bowling says, “a perfect test case,” Dahl had his “neural net” learn to play backgammon, a perfect information game (meaning all the information is available to all players). When that worked, he moved on to card games. 

Years earlier, Giuffria, formerly the lead singer of the rock band Angel, had started scanning patents released into the public sector by the Defense Department with a particular eye to gambling applications. He was turned onto Dahl’s tech by a friend (and Bridge shark) named Bob Hamman who had been arranging poker games between the neural net and serious poker players. Giuffria, who already owned several casinos with Lee Iacocca, found himself deeper in the poker business than ever. The only problem was that the program was perfect.

“Perfect is fine, but there’s no commercial viability for perfect,” Giuffria says. “No one wants to play perfect.”

So his team set about disabling the AI, which actually enabled it. The pokerbot now had layered neural nets and the ability to lie and react so convincingly that it appeared to have moods. A decent UX later, it hit the floors and started making money. It turns out that, when it comes to serious digital gaming – imperfection is the step beyond perfection.

“We wouldn’t have put it out there if we hadn’t solved it,” says Giuffria. “Tell those guys in Alberta someone did this seven years ago.”

That fact – that hold ‘em hard already been perfected – is the reason why Bowling didn’t get any hate mail. Giuffria barely even skimmed the stories. His bot, after all, has been programmed to identify another bot in as few as eight hands. It can confirm its suspicion in 100. That’s a relatively small number when you consider that the only way to ensure success for a robot that is optimized rather than opportunistic (Cepheus doesn’t lie) is to play the long game. Put more succinctly, the Canadian computer, which perfected its game playing against itself, would likely get taken by an electric shark should it swagger onto a casino floor.

Still, having the “solution” to a gambling game in the hands of people not tied to the gambling industry does present an interesting problem for the online gambling industry. Though World Series of Poker spokesman Seth Palansky says that limit heads up hold ‘em probably represent roughly a tenth of a percent of online gambling, he admits that the Cepheus software gets interesting in a non-heads up contest when only two competitors are left standing – when non-heads up becomes heads up by default. 

“Even so,” he says, “you’ve got chip counts at that point so there is an existing situation. Unless everything is even – and it practically never is – the game won’t go on for long enough for an optimized program to guarantee anything.”

That means that if AI researchers want to beat Texas hold ‘em for real, they’ll have to solve it in a non-heads up context. By Bowling’s calculations, that means documenting 100,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000 variations. That’s roughly twice the number of atoms in the observable universe. And if they want to make a buck doing it, they’ll have to beat Giuffria to the punch. That takes a lot of skill and some luck as well.

Photos by Getty Images