Libratus: What we learned
• It wasn't ready to go until the night before the tournament started. Sandholm and Brown played a few test hands against it — and lost badly — before turning it loose against the pros.
• It didn't adapt to individual players. Libratus adapted to the play of the team as a whole, but not to individual players, despite what the pros thought. But as Brown said, “Humans are really good at finding patterns in data, and they're also really good at finding patterns that aren't there.”
Sign up for one of our email newsletters.
Updated 57 minutes ago
Libratus wasn't concerned with beating the humans.
No. The pokerbot that just trounced four of the best poker players in the world during a 20-day Heads-Up, No-Limit Texas Hold'em poker tournament just wanted to be more perfect.
Instead of focusing on exploiting weaknesses in its challengers, Libratus would look for holes in its own play and fix them, said Carnegie Mellon University professor Tuomas Sandholm, who created Libratus with graduate student Noam Brown.
Meanwhile, the pros — Jason Les, Jimmy Chou, Daniel McAuley and Dong Kim — Casino siteleri spent hours each night pouring over the day's hands looking for weaknesses in Libratus to exploit. But when they showed up the next day, armed with new strategies to target holes they thought they had found, Libratus was ready.
“We came in guns blazing, and it was just a very sad sight to see,” Les said Tuesday, as the four pros sat around a poker table inside the Rivers Casino poker room talking about the tournament with Sandholm and Brown. “We ran into a stone wall.”
“The best way to put it is we were literally one step behind Libratus every day,” Kim said. “Every day that we tried a new strategy, that we found a hole, it either A) wasn't a hole, it was just us being wrong the whole time or B) it patched it up.” Avrupabet
Libratus ended the tournament with about 1.77 million more chips than the humans.
Sandholm said Libratus was programmed to fix the three most important holes each night. Big holes sometimes took more than one day. It also studied and learned how to adapt each night to different bet sizes used by its human opponents.
Focusing on improving and perfecting its own play rather than exploiting weaknesses in its opponent was a departure from the work Sandholm, his students and others in the field have done on artificial intelligence and game play in the past. But achieving near perfect game play will eventually allow a program like Libratus to handle scenarios outside of poker, Sandholm said.
Sandholm has secured more federal grant money to continue working on the core algorithms in Libratus and will use a new grant from the National Science Foundation to look for applications for his AI in biology. His work will ultimately focus on developing an AI to be used in negotiations, he said.
He also wants to take a nap.
The supercomputer powering Libratus, on the other hand, hardly broke a sweat. Libratus ran on the equivalent of 3,300 laptops, but used less than half of the supercomputer's computing capacity. While it trounced four of the best poker pros in the world, it also looked for root causes of cancer, identified diseases in crops and worked on the next generation of nuclear power.
The poker pros split $200,000 in prize money for their efforts — first place took home about $75,000, last place $20,000. They doubted whether they could incorporate some of seemingly random game-play of Libratus into their future tournaments. The human brain might not be able to handle it, they said.
Aaron Aupperlee is a Tribune-Review staff writer. Reach Aupperlee at aaupperlee@tribweb.com or 412-336-8448.
Our editors found this article on this site using Google and regenerated it for our readers.
