The Second Man-Machine
Poker Competition
U of A programmers joining the big table
Poker-playing computer challenges Vegas
Keith Gerein, The Edmonton Journal
Published: Sunday, June 29
EDMONTON - A poker-playing computer created at the University of Alberta grabbed worldwide attention a year ago when it challenged two of the game's top players in head-to-head Texas Hold 'Em.
Polaris suffered a narrow defeat that day in Vancouver, but its designers felt they learned enough to give their creation the edge in the next man-versus-machine tussle.
This week, a new and improved Polaris II will again try to strike a blow for artificial intelligence, only this time the scientists are so confident they are pulling up a chair in Las Vegas.
"I think we have it," said Michael Bowling, the U of A researcher who heads the Polaris team.
"What was proven last year is that we can play competitively with professional-level players," he says.
"This year we're aiming higher, so that the statement afterwards is that we're 'better than human players.' " Taking on the computer will be a group of world-class players who serve as coaches on stoxpoker.com, an online poker training site.
Some are specialists at the head-to-head limit Hold 'Em game that Polaris plays.
"Right now this program is better than all but maybe a handful of players in the world," said Stoxpoker's lead coach Bryce Paradis, who has contributed to the Polaris project. "But we are bringing in some very good people, so it's going to be interesting." The Las Vegas tournament, which runs Thursday to Sunday at the Rio All-Suite Hotel & Casino, will be conducted in a similar format to last year's event.
Each day, Polaris will play two games simultaneously.
In one room, it will play 500 hands against one human, and in separate room it will play a different human. The exact cards dealt to Polaris in Room 1 will be given to Polaris's opponent in Room 2. The format is designed is to reduce luck as a factor, so that the match is truer test of artificial intelligence versus human ingenuity.
Polaris employs two different abilities to play poker.
First, it uses a database of strategies to try to determine the ideal move based on the cards it is dealt. However, there's an astronomical number of outcomes for a poker game, and Polaris does not have the computer memory necessary to figure them all out. So instead, it makes the best guess it can.
Polaris's second ability is learning. The program studies how opponents are playing and makes adjustments to its style in response.
At last year's event in Vancouver, the machine played the first two matches purely on its memory, earning a draw and a victory. For the third and fourth matches, the scientists activated its learning mechanism, programming Polaris based on what they had seen from opponents Phil Laak and Ali Eslami.
The humans beat the machine twice in row to win the tournament.
"We maybe got a little cocky after winning the second match, and decided to put in some experimental ideas we had on learning," Bowling said. "It was very crude and it had a bug in it, which meant it wasn't doing any learning." In the 12 months since, the U of A scholars have been refining the ideas they tried in Vancouver, and now feel they have a system that works. Of course, the trick with any learning mechanism is to ensure the machine doesn't learn the wrong things.
"The opponents we are playing against are trying to set traps," Bowling said. "You might think the opponent is folding too much, so you start raising your bets, only to find out that's exactly what the human wanted you to think." Another gap was that Polaris failed to recognize that some cards on the table were shared, and instead simplified the game by thinking that both players had completely different hands. That created a weakness that sophisticated players could exploit. For example, in a situation where an opponent was close to completing a straight, Polaris might have failed to realize it was a good time to pressure the human to fold with a high bet, Bowling said.
Modifications were made for this year's event and Bowling believes Polaris has more "balanced" programming that will make it harder for the humans to identify weaknesses.
Besides the tournament's four official matches, the Polaris team also plans to stage two or three exhibition games that will not be seen by the public.
Paradis, who says he's won $2.5 million playing limit Hold 'Em, is taking part in one of the exhibition matches.
The first part of that match was already played -- Paradis and his partner lost a close one to the machine -- but the duo are hoping for a better result in Vegas when the second part is held.
"Right now it does have a slight edge on me," Paradis said.
The point of the Polaris project is not all fun and games, Bowling said.
The science under development could one day lead to sophisticated computer systems that can help humans make good decisions in situations where some information is unknown, he said.
kgerein@thejournal.canwest.com