John M. Talmadge, M.D.

A Blog Covering Many Topics

Part 2: Texas Holdem Poker, Human vs. AI

Texas Holdem is a version of poker that takes five minutes to learn and a lifetime to master. This is the second article about poker professionals taking on the most powerful poker-playing computer yet invented. Following two weeks of battle on the virtual felt in Pennsylvania, the “Brains vs. Artificial Intelligence” challenge concluded after a marathon battle. Although the numbers say there was a winner, in looking deeper into the numbers the contest is being declared a draw.

The competition was set up by the Carnegie Mellon University School of Computer Science, who created a poker-playing program named ‘Claudico’ and were looking for a significant test. The program, with the ability to “learn” as it played and thus is considered an “artificial” intelligence, is the first-of-its-kind in that it was created to play No Limit Texas Hold’em; every other poker playing computer created played the more-statistical Limit version of the game. Once the Carnegie Mellon staff nailed down the players – and the management of the Rivers Casino in Pittsburgh offered an exciting venue to play – the subjects set out on the 14-day competition.

From the start, the representatives of the human race – World Series of Poker bracelet winner and online wunderkindDoug ‘WCGRider’ Polk, Dong Kim, Jason Les and Bjorn Li – moved out to a financial edge that they wouldn’t relinquish. Playing a total of 80,000 hands of $50/$100 Heads Up No Limit Hold’em over the two-week period, the four men built up a $587,231 edge only a week into the play. They would seemingly ride that advantage over the last half of the competition and, once the results were announced on Friday, both sides crowed about their achievement.

When the final tallies were completed, the “Brains” in the competition had vanquished their “Artificial Intelligence” foe by the sizeable figure of $732,713. Leading the way was Li, who accounted for an astounding $528,033 of the total winnings amassed by the humans. Polk didn’t do badly either, racking up $213,671 in winnings and Kim slipped by ‘Claudico’ in taking slightly more than $70,000 in earnings. Only Les would disappoint the human race, dropping $80,482 to be the only one to lose to ‘Claudico’ by the money counts. (The four men divvied up a $100,000 prize provided by the Rivers Casino for their two weeks of work.)

The human players were a bit surprised at the skill that ‘Claudico’ demonstrated. “We know theoretically that artificial intelligence is going to overtake us one day,” Li said during the post-match celebration. “At the end of the day, the most important thing is that the humans remain on top for now.” Les, who had seen a previous version of ‘Claudico’ when it defeated professional players just last year, was stunned by the developers’ skills.

“The advances made in Claudico in just eight months were huge,” Les said, indicating that, at that rate of improvement, an Artificial Intelligence system might need only another year before it clearly plays better than professionals.

Polk seemed to be the only player who critiqued the playing of ‘Claudico’ during the finale. “There are spots where it plays well and others where I just don’t understand it,” Polk noted, stating that some of its bets were highly unusual. Polk saw instances that, where a human might place a bet worth half or three-quarters of the pot, Claudico would sometimes bet a miserly 10% or an over-the-top all-in move. “Betting $19,000 to win a $700 pot just isn’t something that a person would do,” Polk observed.

So who won the event? While the overall numbers would suggest that the “Brains” crushed the “Artificial Intelligence,” a closer look at those figures is necessary. As individuals, the humans once again take a 3-1 winning edge, but the actual analysis of the figures that the players put up indicate that the score might have been closer to 1-0-3, with Li the only outright winner and the remainder of the human team within the statistical range of calling their matches a tie. It wasn’t a point that was missed by the professor who helped to develop ‘Claudico.’

“We knew Claudico was the strongest computer poker program in the world, but we had no idea before this competition how it would fare against four Top 10 poker players,” said Dr. Tuomas Sandholm, the Carnegie Mellon University professor of computer science who helped to create ‘Claudico.’ “It would have been no shame for Claudico to lose to a set of such talented pros, so even pulling off a statistical tie with them is a tremendous achievement.” In replying to Polk’s stab at the unorthodox play during the event, the Carnegie Mellon team admitted they were just as puzzled as to why ‘Claudico’ made the decisions he made.

After the completion of the interesting competition, there have been no indications that there will be another event on the horizon. The Carnegie Mellon team will no doubt head back to the laboratory to tweak on ‘Claudico’ (or potentially a more-potent creation?), while the human race will wait for the next challenge to their ‘superiority’ in this world.

This may turn out to be the latest installment in a grand tradition of computers beating us at our own games. In 1997, IBM's Deep Blue computer famously beat chess great Garry Kasparov. Four years ago, IBM's Watson took part in the TV quiz show Jeopardy! and crushed two contestants with a strong track record. AI has even mastered the popular smartphone game 2048.

Still, poker is a tough nut to crack. In a game like chess, everyone knows where all the pieces are on the board. By contrast, poker is a game of imperfect information: players don't know for sure what cards the others hold or what will come up next in the deck. That makes it a challenge for any player, human or computer, to choose the right play.

(This blog entry was compiled from various sources, and some attribution is lacking. I apologize and will correct this if I can.)

Part 1: Texas Holdem Poker, Human vs AI Computer


Fifteen years ago I, Dr. John Talmadge, became a serious poker player. I have a winning record at the poker table, and although I am (probably) not good enough to quit my day job, I have held my own with professionals in Las Vegas, and I finished in the top third of the field in the World Series of Poker Senior Event in 2009. Although I finished out of the money, I came in ahead of about 2000 other players in that event. In the years since, I have consulted with professional players on "the mental game," and I remain a serious student of Texas Holdem, the world's greatest card game.

Image of Brains vs. AI in Texas Holdem Poker

Now comes Aviva Rutkin, writing in New Scientist that human poker professionals are taking on artificial intelligence (and the strongest poker playing computer built to date), and they are playing for real money. Here's the story.

Computer scientists have already made some progress, at least with simpler forms of the game. But the version being played at the Pittsburgh tournament, called Heads Up No Limit Texas Hold 'em, is "a completely different beast", says pro player Vanessa Selbst. "There's much more human elements and game strategies to employ, so it's a much more complex game." What's more, there are no betting limits, so the computer also has to take into account how much players might stake on each game.

Competing in Pittsburgh is Claudico, a program created at Carnegie Mellon University. Claudico taught itself poker skills by playing trillions of games in search of some kind of optimal strategy. Whatever it has picked is pretty good: last year, Claudico beat all 13 other computer competitors at no-limit poker in the annual contest run by the Association for the Advancement of Artificial Intelligence.

Computers have a few edges over humans, says graduate student Noam Brown, part of the team behind Claudico. For example, a computer can switch randomly between various betting strategies, which may confuse human opponents.

On the other hand, Claudico is slow to pick up on and adapt to people's playing styles – something that many pro players do with ease. "One of our big concerns is that the human will be able to identify weaknesses that Claudico has and exploit them," says Brown.

Because Claudico taught itself to play, even the team that built it don't quite know how it picks its moves. "We're putting our faith in Claudico. It knows much better than we do what it's doing."

Algorithms like those used to play poker could be valuable for other kinds of problems characterised by imperfect information. They could suggest optimal locations for military resources in a war, for example. Rival AIs could also be tasked to negotiate with each other over insurance rates or handle legal squabbles. "In society, sometimes you see one side getting screwed over because someone has more lawyers or more information or more resources at their disposal," says Brown. "Something like this can really level the playing field."

The winner of the poker tournament won't be crowned until the event ends on 7 May. Eric Jackson, a software engineer who creates poker bots as a hobby, is cautiously optimistic that Claudico can win. As we went to press, the pros and Claudico were neck and neck.

Even if AI triumphs, it won't mean programmers have conquered the game. "Beating humans decisively would be a landmark, but it wouldn't mean the end of work on poker," says Jackson. "We still don't know what the perfect strategy is."

The original article by Aviva Rutkin can be found online here.