The First Man-Machine
Poker Championship

Man Versus Machine: Computer Giving Players Real Test

Phil Laak and Ali Eslami Face polaris, the Deep Blue of Poker Bots

Tuesday, July 24, 2007

Cardplayer.com

Phil Laak and Ali Eslami are representing the human race at the Association for the Advancement of Artificial Intelligence (AAAI), and, well, they aren't doing so good. The two are in the middle of a two-day match with a poker playing computer program named Polaris that was developed by a team of students and professors at the University of Alberta who are obsessed with solving the "problem" of heads-up limit poker.

After playing two sessions yesterday, Laak, a self described Geekaphile, believes the scientists are extremely close to solving it.

"(Polaris) is phenomenal. They just spent all their (freaking) time perfecting heads-up limit poker," Laak said. "That thing just beat us."

Laak and Eslami didn't go to Vancouver just to help out the big brains in Canada. Like most scientific competitions, there's money on the line. Here's how the competition, which is taking place at the AAAI conference in Vancouver, works:

The players are playing a total of four 500-hand sessions of $10-$20 limit hold'em in two days. If the humans win a total of 25 small bets or more from Polaris in each session, they will split $5,000. If neither the computer nor the humans manages a 25 small-bet profit, the session is considered a tie and the humans split $2,500. If the humans sweep all four sessions, they split $50,000. If Polaris comes out 25 small bets or more ahead, the humans get nothing.

The first session ended with Eslami and Laak squeaking out a profit of seven small bets (Eslami ended up $395, but Laak lost $465) for the tie. Yesterday's second session was won outright by Polaris. Laak fared much better in this session, coming out $1,560 ahead, but Eslami ended up down $2,515. Laak said he was getting great cards, though, the same cards that the Polaris playing Eslami was getting.

Both players play simultaneously -- one in front of an audience and one in a private room -- and they are playing a form of duplicate poker. This means the cards dealt in the first match to the human are dealt to the computer in the second match, and vice versa. According to the computing team, this significantly reduces the role of luck in determining who the better player is.

So far, it's been Polaris. Unlike the computer that Laak beat two years ago as the human representative at the World Poker Robot Championship, Polaris is designed to bob and weave like a real player, adjusting playing styles and recognizing weakness. But Polaris also plays the way many pros play -- it tries to play basic solid poker and wait until its opponent makes a mistake.

"That (strategy) basically defines my entire net worth," Laak said. "It's always thinking like this: I will break even."

The University of Alberta has one of the biggest departments dedicated to artificial intelligence anywhere in the world. The faculty and students there dedicate many, many hours to developing programs that could beat humans at games.

The school received a ton of publicity this week when it announced that it had "solved" checkers. Basically, researchers there created a computer program, called Chinook, that will never lose at checkers. The best a player can do against it is play to a draw.

But checkers, as well as chess, is different from poker in one drastic way. In both those games, all the information a player needs to win is right there in front of them. The positions of the pieces aren't hidden, nor are the spots where they could be moved. Very, very smart players can look at the board and know exactly how to counter the other player in order to win.

In fact, Schaeffer said computer scientists knew the solution to chess since the 1960s, but technology wasn't ready to handle the amount of calculations needed to beat a grand master like Gary Kasparov, who lost to IBM's Deep Blue in 1997. The scientists just had to wait until enough computing power was available.

Poker is different. Poker players make decisions by analyzing only the information that is available to them. There's a whole lot of educated guessing going on in poker, and this makes the whole pokerbot problem much more complex than other games.

Like so many decisions in life, there's no one right way to win at poker. Since 1991, when the researchers at the University of Alberta started to try to crack poker, dozens and dozens of different poker playing machines have been built and tested. Schaeffer says all the programs have different strengths and weaknesses and fare differently against different players. The trick is building a program that can recognize which adjustments to make on the fly that will give it the best chance to win against anyone.

One of the major strengths of humans is we're able to adjust our thoughts very easily, at least compared to a computer program. A person just has to make his mind up and do it. Changing the way a computer program "thinks" is much more difficult. Computers are not very flexible when it comes to "thought."

That worries Schaeffer, especially with players like Laak and Eslami, who are both severely smart, curious, and competitive people who want to badly beat the machine. They want to know how the program works. They want to figure out the weaknesses and learn how to exploit them. They've read papers written by the group of developers and have exchanged emails discussing theories. They understand their roles in the research group as guinea pigs, and they love it.

Laak said he just loves to hang around people who are smarter than he, and Schaffer appreciates the help.

"These guys are superb. They're friendly, they're outgoing, they're inquisitive. We consider these players to be very dangerous for us because they do understand the mathematics that are going on," Schaeffer said. "It's very impressive. I'm in awe with the depth of their insights."

Schaeffer said it was obvious listening to the players talk about today's matches that they have spent time thinking how they could beat this machine. But last night, Laak said he doesn't think it's really possible to consistently beat it, no matter what caliber of player takes the mouse in hand.

"Let's put it this way. The bot is so strong, I would bet on it against any human in the world," Laak said.

He thinks the best players would always play to a tie because that's what the computer is programmed to do. Players of lesser skill will make mistakes, and the mistakes will translate into wins for Polaris.

That said, Laak believes that the University of Alberta team only has heads-up limit poker about 10 percent "solved." The other 90 percent will come in the next few years, when the researchers figure out all the small details that Polaris needs to know to always win, the same details that many poker players are trying to figure out, too.

Schaeffer feels the same way. No matter what happens during the last two sessions today, he believes that this specific form of poker -- and heads-up limit poker is about as specific as you can get -- will be solved. This exhibition is an important step for the team.

Although the data obtained from this experiment will be taken with a grain of salt because of many uncontrollable factors, Schaeffer is pleased going into the second day of matches.

"Boy, based on what happened yesterday, it looks really, really good."