Poker AI 'Libratus' Torches Pros in Heads-Up Challenge for $1.76M
Well, that wasn't very close at all.
A new poker AI (aka 'bot') developed by the Carnegie Mellon University Computer Science Department has soundly defeated a team of four professional poker players in a 120,000-hand Challenge.
Dubbed 'Brains vs. AI the Challenge played out at the River Casino in Pittsburgh over the past two weeks with the pros - Dong Kim, Daniel McAulay, Jimmy Chou and Jason Les - getting throttled to the tune of $1.76 million in imaginary money.
This isn't the first time a group of poker pros have taken on a high-powered AI but it is the first time they've come away so soundly defeated.
It's especially significant this time around as the game was Heads-Up No-Limit Hold'em - considered a massively complex of incomplete information previously thought 'unsolvable' by a computer.
The new AI is called 'Libratus' (meaning 'balanced') and is a newer incarnation of the poker bot 'Claudico,' which participated in a similar challenge two years ago.
That time around the pros - Les, Kim, Doug Polk and Bjorn Li - won handily but this time it was a different story. Self-trained by playing trillions of poker hands against itself, Libratus had developed a virtually unbeatable strategy based on 'Counterfactual Regret Minimization.'
In layman's terms it plays a virtually perfect, balanced style of poker that can't be exploited - and it also learns as it goes.
'Way Tougher Than Expected'
The opinions of the pros - all very experienced and talented players in their own right - about the new bot were clear and decisive: It was extremely, extremely tough to play against.
Whenever they would review the hand histories for the day and develop a new counter-strategy, the bot would adjust accordingly and continue to rack up winning days.
All told each player played 30,000 hands again the AI with all of them 'mirrored' - meaning the reverse scenario for each hand would be played out among the other competitors. All ins before the river were also restricted to the equity in the pot when the money went in to minimize variance and luck.
Kim, who played the Challenge last time, fared the best against the bot with a net loss of only about $85,000 in chips. Les, who also played last time, fared the worst by losing over $800,000 in chips.
What Does It All Mean for Poker?
While the success of Libratus shows a remarkable leap forward for Computer Science - and possible applications derived from processing incomplete information better - the affects on poker are likely to be minimal.
The game specifically is heads-up, for starters, which minimizes the number of playing factors the super computer has to process. A game like six-max or Omaha, for example, has an exponentially greater amount of variables to consider.
The technology isn't yet there to do those computations in a timely manner, even with the help of the Bridges Super Computer as Libratus had.
Live poker is of course an entirely different game as well and an impossible one to account for all the variables in one single algorithm - as of now.
Online poker sites also have proper security in place to detect people trying to play with bots - all of which are much less sophisticated than Libratus. Most people also don't have access to a Super Computer to do the necessary trillions of computations needed to play perfectly.
For more on the challenge and the responses from the participants, check the Rivers Casino Brains vs AI page here.