| Programming |
“The winnah — and new champeen — Dr. Hans Berliner’s computer backgammon program.”
Well, not exactly. But the Carnegie-Mellon University research computer scientist’s brainchild did trounce the world champion of that game 7–1 in a seven-point exhibition match recently.
The program was born when Berliner, world champion of correspondence chess 1968–72 and for 15 years one of the top 10 players in the country, returned to school for his doctorate.
Dr. Hans Berliner sits at screen of his backgammon computer.
“One of the reasons I came back to school was to try to build a chess-playing program,” says Berliner, who received his doctorate in computer science from CMU four years ago.
“There’s a particular interest around this university in trying to emulate human behavior with a computer. The field is called artificial intelligence, getting machines to do things which would be considered human if a human did them. We try to build programs so computers can do intelligent things.
“I was trying to devise a chess program. There are lots of avid chess players here, and there are standardized chess tournaments where all the players have ratings, so you can bring a computer program in, let it play a dozen games and find out how good a program it is and how ‘intelligent’ it is.”
But chess was too difficult to program then, having 10 to 20 times more aspects to weigh than backgammon, says Berliner, who himself has been playing backgammon for six years and is an intermediate-level player.
“Backgammon filled the bill. The program understands many different aspects of backgammon and understands them very well. It understands the critical things; it has a very good idea of when to try to apply these things.”
In the match against champ Luigi Villa of Italy in Monte Carlo last July, the program showed its “intelligence” by “doubling” Villa (one of the critical things it understands) and backing him down.
Doubling, Berliner explains, is a 20th-century twist to the ancient game in which a player who thinks victory is assured challenges his opponent to continue the game, with double points for the winner, or concede defeat at the regular point value by rejecting the challenge.
“At one point, Villa did reject a double,” Berliner says. “He decided it was too risky and did not accept. Other top players there said he should have accepted. I ran it out, and they were right — he should have continued. He had enough of a chance to make it mathematically worthwhile.”
Microprocessor backgammon games, ranging from $200 to $400, are available commercially, but they have “nowhere near the amount of information my program has,” he says. “I played my program against the best of these and it really kills it.
“It’s not perfect — I see a number of things that need to be fixed — but it gets harder and harder to improve it. The next version will be able to hold conversations with the user, to point out errors if the user is interested in an assessment of his play. We also hope to help the program by having it keep records of its own play” for later analysis.
OK, it’s better than the home versions — is it better than its creator?
“I’m just a little bit better at the moment, but it’s getting close,” Berliner says. “Ultimately it will know almost as much as I know but it will know it better and will not make mistakes.”
But if it’s not quite as good as Berliner, an intermediate-level player, how did it clobber the world champion so badly?
“They both made mistakes, but the dice favored the program (Villa rolled his own, Berliner ran the terminal connected by trans-Atlantic cable to CMU computers and a third person rolled dice for the program).
“There’s no denying there is an element of luck. When you beat a champion like that, you have to have luck on your side.”