8 August 2005--It seems the scales are finally tipping in the decades-long struggle of human grand masters and chess machines. The humans are looking increasingly like the palookas that impresarios used to pit against champion boxers, just to attract spectators. And in the one sphere where grand masters always claimed an advantage--playing interesting and fun-to-watch games--the gap also seems to be closing. The machines' play, once marked by bizarre, inscrutable moves, increasingly resembles the play of humans--very smart humans.
The turning of the tide began last October, in Bilbao, Spain, when a special-purpose machine called Hydra and a commercially available program called Fritz, running on a laptop, creamed a bevy of top-ranked grand masters. Hydra and Fritz each scored 3.5 points out of a possible 4. Then, in June, in a match webcast from London, Hydra humiliated Britain's Mickey Adams, the seventh-ranked grand master in the world, with a score of 5.5 out of 6. The same week, a version of Fritz, downloadable for free as part of the Accoona Toolbar (a device that searches a PC's hard drive) drew a one-game exhibition match against Uzbekistan's champion, Rustam Kasimdzhanov.
What explains the rise of the machines? First, the best humans now playing are simply not as good as the former bearer of humanity's torch, Garry Kasparov, the highest-rated player in history. Kasparov recently retired from professional play to devote himself to unseating Vladimir Putin as president of Russia. Second, the machines are getting smarter. Their programmers are learning how to counter anticomputer strategies. These strategies generally involve closing up the computer's position with pawns. That way, the pieces have less mobility, and the machines cannot exploit their fearsome calculating power.
Kasimdzhanov managed to obtain just such a position in his match against the Accoona version of Fritz. But to convert his strategic advantage to a win he had to open things up at some point, potentially allowing the computer more freedom to move. It seems he did so prematurely. At move 21, grand master Lev O. Alburt, who was watching from the audience, said that Kasimdzhanov's winning chances would be around 80 percent if he were playing the same position against another top grand master. Two moves laterAlburt said the win--if there still was one--would be hard to work out in the time the harried human had left. Seventeen moves later and down to less than 3 minutes versus the computer's 35 minutes, Kasimdzhanov offered to repeat the position again and again. The computer--perhaps not programmed to exploit an advantage time on the clock--allowed the draw.
Off in London, Adams had no such luck against Hydra, deemed the best chess computer, in large part because of its intimidating hardware. Based in Abu Dhabi, United Arab Emirates, it yokes together 64 computers, each with an Intel Xeon 3.06-gigahertz processor. The computer cluster boasts a total of 512 gigabytes of memory. You'd expect that so much computing power would lead to high-level play, but what's unusual is the character of the play. As Adams said after the match, in an interview published on the Web site of ChessBase GmbH, in Hamburg, Germany, which produces the Fritz software: "Hydra plays very well indeed. Very often it plays human-style chess, which is strange."
Chessbase is putting new emphasis on making Fritz, its signature program, play more like a human being. This involves playing according to discernible plans or at least seeming to do so, particularly in positions with no immediate threats. In these so-called quiet positions, computers have been wont to shuffle their pieces about without obvious purpose (but with a high tendency to win, nonetheless). According to Frederic Friedel, a principal of Chessbase, the company decided about 18 months ago to start programming in more "chess knowledge."
Formerly, Friedel says, when Fritz entered quiet positions, it evaluated millions of possible end positions only to decide that several continuations were almost exactly equal. It would then often select a strange-looking but serviceable move no master would favor because, Friedel says, "it rated the move a millionth of a pawn better than the alternatives." Now Chessbase has devised a way to structure the data generated in the look-ahead phase to encourage the computer to favor lines that end up moving its pieces to better squares—say, by getting a knight to a central outpost supported by one of its own pawns.
Friedel recalls that Bell Laboratories' Ken Thompson, a pioneer in computer chess and many other aspects of programming, used to joke that no matter how much knowledge you implement, it won't compensate for the loss of one move in search depth. Fritz now looks about 16 moves deep (that's eight moves by white and eight by black). The question was what fraction of a play would have to be sacrificed to make Fritz act like a human.
"We knew implementing more knowledge might weaken Fritz's results against other computers, but that has only one purpose: to beat Shredder [a rival program] by 52 percent," he says. "We once had to do that to survive commercially, but we felt it was no longer so important."
In the end, the programming changes did not make Fritz measurably weaker than rival programs. They have, however, managed to spook some grand masters. After losing to it in Bilbao, Bulgaria's Veselin Topalov, ranked third in the world, asked, "What have you done to Fritz?"