PITTSBURGH _ The machines are taking over _ at least in poker.
Though it had been a point conceded by the humans for the past week, Carnegie Mellon University's poker-playing computer, Libratus, on Monday finally, definitely and soundly defeated four of the world's best Heads-Up, No-Limit, Texas Hold 'em poker players by a resounding $1,766,250 in theoretical dollars in 120,000 total hands played over 20 days at Rivers Casino.
"Yeah, this was a beat down," said a tired human team member, Jimmy Chou, after finishing his 30,000th hand late Monday afternoon.
Tuomas Sandholm, the CMU computer science professor who created Libratus with his doctorate student, Noam Brown, said the win was "awesome" for artificial intelligence research.
"This is a landmark in AI," he said.
His colleagues around the country seemed to agree.
"This is exciting news out of Pittsburgh," said Oren Etzioni, chief executive officer for the Allen Institute for Artificial Intelligence in Seattle. "It is significant because (poker) is an imperfect information game."
Eric Jackson, a private entrepreneur who designed another poker-player bot named Slumbot that finished second to Sandholm's computer in a bot-versus-bot poker tournament last year, agreed with Etzioni.
"It is a surprisingly large jump in improvement after being pretty significantly beaten the last time," Jackson said. "This is the first time anybody has beat the top pros in a No Limit game."
This was the second match between a CMU poker-playing computer and four of the world's best poker players. The last time _ in May 2015 _ the humans won by $732,713 over 80,000 hands, won almost every day, and had three of the four team members beat that computer, nicknamed Claudico, in total winnings.
This time, not one of the four players individually beat Libratus. The best showing came from returning player Dong Kim, who lost by $85,649 over the 20 days. The human team this time only won four of the 20 days.
After the first match in 2015, artificial intelligence experts predicted it could take up to five years for a computer to improve to the point that it could "solve," or win, a comparable tournament against the best human players. Libratus managed to do it just 20 months later _ and win so decisively that it had the two-handed players in awe.
"It is so tough," said Chou. "It not only fixes its leaks and exploits, but I really felt this bot was exploiting me" in certain poker situations.
Jason Les, another returning player, said he was looking forward to sitting down with Sandholm and Brown.
"I just want to know that it's not in our head, because it seems that it changes its strategy not just because it's learning, but against us as individuals," he said.
Libratus is really an algorithm created by Sandholm and Brown that runs on a computer, nicknamed Bridges, in the Pittsburgh Supercomputing Center's stable.
One significant change that clearly helped Libratus quickly adapt to the play of humans and play better this time was the millions of additional "core" hours of research the CMU team was able to perform on Bridges. By the end of the tournament, Libratus used 20 million to 25 million core hours of Bridges' power, compared to just 2 million to 3 million for Claudico.
"And I think that's one of the reasons Libratus is doing better," said Nick Nystrom, senior director of research at the computing center, which is based in Oakland and runs its computers at a site in Monroeville. "Because the actual strategy being run on the casino floor (in Libratus) is being run through our computer."
The computing center approved CMU's request to use its computer because even though it is a poker-playing bot, the same algorithms that allow it to play poker so well have direct applications in other imperfect information games, including negotiations, security and even medicine.
The computer's power was noticeable to the human team early on.
"At day 10, we were still optimistic and thought we might be able to pull it out," said player Daniel McAulay. "But after that, we've known for a few days it was over."
Try as they might, using strategies designed in late-night consultations after each day, nothing they seemed to do helped.
"Yeah, it really gave us a beat down," McAulay said. "What else can you say?"
The human team members even made light of their situation in a video that was posted on Youtube.com on Monday.
Put together by Doug Polk, one of the world's top players who participated in the first contest against Claudico, the video pretended Polk was the team's coach, dismayed at the results he was seeing _ but ultimately decided there was nothing anyone could do.
"You guys are contractually obligated to go out there and get punched in the face repeatedly ...," Polk told the team, looking downcast for laughs.
"It's ain't over till it's over _ but it's really over," he tells them. "I looked at the math; there's just no chance. It's already established you guys are losers. But let's not be big losers."
