Computers master the game board

That's why Mr. Schaeffer is happy to put checkers behind him (he led the 18-year effort to crack the game) and focus on poker. His new project, Polaris, poses a much more interesting challenge, he says.

After all, how do you teach a computer to predict an opponent's hand?

Man versus machine

For the July tournament against Mr. Eslami and poker celebrity Phil Laak, the University of Alberta team developed three different computer personalities for Polaris. The first script played it safe, calculating and betting in the hopes of coming out even. The next was aggressive, putting pressure on its opponents. The team programmed the last personality to learn as it went, reading opponents and acting accordingly.

Polaris faced off against the pros at the same time but in separate games. To minimize the luck of the draw, Eslami and Mr. Laak received mirrored hands – so if one player received a great hand it meant that in the other match Polaris was dealt that great hand. At the end of each of the four rounds, the humans' scores were combined.

The pair came out a little behind against the careful, rational program, but the score was close enough that both sides agreed to call it a tie. The aggressive code crushed the humans in the second match. Polaris's learning program failed, handing the humans a solid third-round win. ("It was our fault it didn't work," admits Schaeffer.) For the final round, Eslami and Laak wanted a rematch against the play-it-safe bot. Better prepared, the professionals defeated Polaris.

"It was a really strong, savvy opponent and that has me very excited," says Laak in a phone interview after the tournament. "Life is a myriad of puzzles and this is the first step in some thousands ahead where computers will get better and better."

In six months, Polaris will be much stronger, Schaeffer says. He hopes to fix the learning code and possibly throw in a coaching mechanism, where Polaris can switch between rational and aggressive strategies.

Strengths and weaknesses

Some games are still too complicated for computers to master. The Japanese game of Go stands as the usual example. With a 19 by 19 grid, Go has an astronomical number of possible positions – think 1 followed by 100 zeros. Such a massive scale means computers don't know where to focus.

"They've done eye tracking on Go experts," says Susan Epstein, a computer science professor at Hunter College in New York City. "The studies found that while there are hundred of good moves in front of them, the best [human] players only see three or four."

So how do you teach computers to "see" what humans see? For one, stop relying on programs that simply map out a single game, suggests Michael Genesereth, director of the Logic Group at Stanford University in Stanford, Calif.

Yes, Deep Blue dominates chess. But the supercomputer is a one-trick pony. Without plenty of prep time, it'd be helpless in a game of Othello, Mr. Genesereth says.

Instead, he researches general gaming, where machines learn patterns and principles that work in a variety of puzzles. At the same conference where Polaris battled human opponents, Genesereth held his annual machine-on-machine championship. The competition pitted general-gaming programs against one another in a series of board game mash-ups. May the best code win.

The Air Force Research Laboratory in Rome, N.Y., is even researching time-critical reasoning through asynchronous chess, where two competing computers don't have to wait their turns. They can move any piece at any time they want.

These code-versus-code styles of play are harder to program, but also much easier to translate into real-life situations, he says.

The Logic Group works with firms such as SAP, the world's largest business software designer, to create versatile programs that are ready to shift gears with any new change in interstate law or corporate policy.

"It's impractical to go back to your programmers and say, 'OK, well, here is another change and another one. Start rewriting all the programs,' " Genesereth says. "It's better to change the rules, and let the program figure out how to maximize efficiency under the new conditions."

Page: 1 | 2