One of the proving grounds for artificial intelligence is games. Classic games have a fixed set of rules, and these make it easier for researchers to develop new techniques and algorithms that enable computers to play (and hopefully win) various games. Tic-tac-toe, checkers, and chess are all games where researchers have developed software that is capable of winning or drawing when paired off against the best human players in the world. Last weekend, researchers at the University of Alberta added another classic game to this list: poker. In a series of matches that took place over the Fourth of July weekend in Las Vegas, the researchers' Polaris poker program won against a groupof top-ranked online poker players.
The first three games mentioned above are known as perfect information games. In games of this type, each player has all the available knowledge about the current state of the game. With that information, the player can, in theory, work out every possible outcome from that point. Given a computer's ability to evaluate hundreds of thousands (or more) of scenarios each second, they are an ideal tool for calculating probabilities in such games. Theoretically, with enough computing power, every possible outcome at each point in the game could be calculated, and a computer could never lose.
Poker, on the other hand, is not a perfect information game. Each player has a limited subset of the total information regarding the current state of the game. In Texas Hold 'Em, the poker variation played here, players are aware of the two cards they hold, plus the cards that are open to all on the table. But, critically, they do knot know what their opponents hold in their hands. Each action must be based on this limited amount of information. According to Prof. Bowling, the principal investigator on the Polaris project, "when you look at games where players are asked to make decisions with different amounts of information, missing information, poker is the quintessential game."
This was the second year that the Polaris software went head-to-head, so to speak, with human players. Last year, in a series of four matches, human players Phil Laak and Ali Eslami edged out Polaris 2-1-1. Over the course of the year, the researchers fine-tuned Polaris by improving its learning capability. The new Polaris was capable of identifying opponents' playing styles and could adapt by countering with a strategy that would be expected to give it the edge. To practice, the researchers said Polaris played eight billion games of poker against itself.
This year, Polaris again played four matches and, thanks to these improvements, it came out ahead. Each match consisted of 500 hands of poker. In the first of four matches, Polaris and the human players wound up in a draw. The second match ended with the human players up by $50,000. The third and fourth matches were decisive wins for Polaris—it won the two matches by ending up with nearly $150,000. While Polaris won, it is still not unbeatable, as shown by the first match. According to Bowling, though, given enough computational power, a computer could play perfectly. That would inevitably lead to a win, since people could be counted on to make mistakes.
With the win under their belt, the researchers are headed back to the lab—or cube—to code some new abilities for the Polaris system. Currently, it is limited to playing against two players and to playing a game of poker where the maximum bet is capped. The researchers hope to enable it to start playing more advanced versions of poker, such as the very popular no-limit Texas Hold 'Em. They also hope to extend this line of work into more real-world problems. "In general, problems in the real world are going to be more like poker than chess. You're not always going to have all the information," said Bowling.