With more losses than wins, strategies that work one day but not the next, and 12-hour days sapping their energy, four of the of the world’s best Heads-Up, No-Limit Texas Hold ’em poker players know they are facing a tough battle over the coming days.
In the second competition pitting humans against the latest version of Carnegie Mellon University’s poker-playing computer, nicknamed Libratus, the human players say they are more than impressed with their opponent, which has beaten them seven out of the first 10 days and held a cumulative lead of $676,584 (theoretical dollars) as of Friday night.
“It is very impressive, actually,” Dong Kim, the only human beating Libratus as of Friday morning, said of the computer. “It does a little of everything [in terms of strategy] very well.”
Dubbed “Brains Vs. Artificial Intelligence: Upping the Ante,” the 20-day competition being held in the poker room at Rivers Casino on the North Shore was CMU’s second attempt to get a computer to be able to beat four top-level human poker players.
The first competition, in 2015, resulted in a win by the humans — though not by “statistical significance.”
This time to try to reach that magical “statistical significance’ point, the humans agreed to play 50 percent more hands of poker, increasing the overall number from 80,000 hands to 120,000 hands. Each player competes against the computer one-on-one, but the final result will be determined by the combined winnings of all four players vs. the computer’s winnings.
The players knew that CMU professor Tuomas Sandholm and his Ph.D. student Noam Brown had built Libratus from scratch to be more powerful and less easy to take advantage of than the 2015 edition, known as Claudico.
But they did not realize how much better it would be.

PG graphic: Beat a computer?
(Click image for larger version)
“The way we beat it last time was to try some quirky stuff and exploit it when it worked,” Mr. Kim said. “That’s not working this time.”
Jason Les — who, like Mr. Kim, played in the 2015 competition — said there is an obvious big difference between the two computers.
“It’s very apparent to me that [Libratus] is learning very quickly as we play,” he said. “The way it played on Day 3 to how it is playing now is an insane difference. Things we identified as a weakness on Day 3 now are gone.”
Mr. Brown said that is exactly how Libratus was designed to perform, and it is benefitting from additional time on the computers at the Pittsburgh Supercomputing Center at night to learn from each day’s hands.
“We’re not going to comment on how the bot is changing except to say that when the humans are sleeping, the bot is working at night trying to improve its game,” he said.
Mr. Sandholm, obviously feeling confident from the success, said a win for Libratus “looks good.”
“And if we can maintain the current win rate, it would be a very statistically significant win,” he said Friday. “But these humans are very crafty. They’re awesome at poker. They’re also very serious and try to improve every day, like professional athletes.”
Mr. Kim, Mr. Les and their two new teammates — Daniel McAulay and Jimmy Chou — don’t just end their day with the last hand. They gather at night after they finish to strategize and go over data, often until early the next morning.
So far, it has only resulted in two small wins — less than $50,000 on two days — and one better day — $108,775 won this past Monday, while Libratus had two days this past week when it won by more than $180,000.
“I guess if you wanted to make this into a drama, it’d be pretty good for that so far,” said Mr. Les. “We lost right out of the gate, then we clawed our way back, and then we went on a [two-day] $400,000 drop.”
Is there a chance for the humans to win?
Given the way the computer now plays, Mr. Les said, “It will make it really hard for us to have a big win day.”
“We still have a fighting chance. We can try to get small wins every day. But if we win, it won’t be by much,” he said.
“The optimist in me wants to think we can claw our way back in the end,” he said. “But I can’t ignore the data.”
Sean D. Hamill: shamill@post-gazette.com or 412-263-2579 or Twitter: @SeanDHamill




Comments
Welcome to Post-Gazette subscriber-only commenting community. Please review our community guidelines.