When world chess champion Garry Kasparov abruptly resigned the sixth and final game of his match in May against Deep Blue-a.k.a. the IBM RS/6000 SP supercomputer-a machine finally fulfilled one of the oldest challenges in artificial intelligence. Chess has tantalized computer researchers since the 1830s, when the eccentric English inventor Charles Babbage thought of luring investors to his idea of a programmable “analytical engine” by holding out the possibility of a chess-playing machine. After all, the rules of chess are precisely defined and easy to program, yet they give rise to strategic complexities that challenge the finest human minds. But despite researchers’ best efforts, no machine proved able to beat the finest human players. Until Deep Blue.
Ironically, the victory comes when the computer-chess community has long abandoned any pretense of mimicking human thought. Chess masters, like the rest of us, are now known to reason by recognizing patterns, forming concepts, and creating plans-processes that computers do poorly, if at all. Deep Blue, like all the top chess-playing machines since the 1960s, relies instead on brute force-it looks as far ahead as it can at all possible moves and evaluates the strength of each position according to preprogrammed rules. Because of the rule that the faster the computer, the more positions it can search and the better it can play, Deep Blue relies on 32 high-speed processors operating simultaneously, each coordinating the work of 16 special-purpose “chess chips” that run in parallel. This computing firepower enables Deep Blue to evaluate a total of 200 million positions each second.
M. Mitchell Waldrop, author of the bestseller Complexity and of a forthcoming book on the history of computing, recently spoke with Deep Blue’s principal designer at IBM, Feng-Hsiung Hsu, about the implications of the machine’s victory and its value for other uses.
TR: In February 1996, when Deep Blue was brand new, it went up against Garry Kasparov and lost. Many people felt vindicated-as if that proved the human mind’s innate superiority over a mere machine. But now that Deep Blue has won, many feel as if the computer has humbled humanity. Should they feel threatened?
HSU: No. Remember, Deep Blue didn’t play chess by itself. Before the match even started, humans programmed the machine to rise to Garry’s level. And then during the match we actually went in between games, looked at Deep Blue’s mistakes, and adjusted its criteria for evaluating the situation accordingly, so it wouldn’t make the same mistake twice. Without that, Deep Blue could not have competed with Garry. So you could say that last year, Garry won one for humanity’s past. This year, Deep Blue won one for humanity’s future.
TR: How so?
HSU: When Garry plays chess, he is relying on the intellect he is born with, his knowledge of the game, and the experience he has gained from playing both people and computers. This is the old-fashioned way of playing chess; Garry, despite his brilliance, is limited by what is biologically possible. Deep Blue represents any technology that allows us to exceed the limits nature normally imposes on us. Right now we’re talking over the telephone: just by shouting I cannot reach you. The principle is the same with chess. Garry may be the top player ever in chess, but while the chess players on Deep Blue’s team can’t claim to reach anywhere near Garry’s ability, with Deep Blue we exceeded our limits and won.
TR: When you put it that way, the match sounds a little unfair. Garry wasn’t playing against one machine or even one person but a whole team.
HSU: But Garry was also part of a team. Between games he would consult with his coach, and even his own chess computer, to find out more about what Deep Blue would do. That is actually a normal part of any master’s-level chess match. So you could say that Garry was playing against a computer relying on human power-but Deep Blue was playing against a human relying partly on computer power.
TR: Fair enough. But you could have said that last year when Garry won. Yet this year he lost. What made the difference?
HSU: The most obvious differences are that Deep Blue was twice as fast this year because it had new central-processing-unit chips, as well as twice as many chips designed only for the purpose of playing chess.
But for the match those hardware advances weren’t as critical as two other considerations. First, we addressed the knowledge gap. Garry is a remarkable human being, with vast stores of knowledge and intuition about chess gained over 30 years of playing. Last year Deep Blue went into the match as a newborn baby: it had just been built and didn’t know much about chess. So afterward we asked International Grand Master Joel Benjamin to come in with us and essentially take the machine to chess school. Actually, we went to chess school and used what we learned to completely reprogram the machine’s basic software code and redesign the chess chips to incorporate much more chess knowledge. By this year’s match, in Joel’s words, Deep Blue had started to play human-level chess.
Second, we addressed the question of continued learning on Garry’s part. For a computer scientist, the idea of building a machine to compete with the world chess champion is like climbing Mt. Everest. Unfortunately for us last year, the human Mt. Everest grew 100 feet a day while the match was proceeding: Garry has a human being’s ability to adapt to what Deep Blue is doing. We knew that Deep Blue would never be as adaptive as a human, since that’s not the way a computer is constructed. But we built software tools that allowed us to go in between the games and adjust Deep Blue’s programming much faster than we could before. That turned out to be critical. The situation was like competing in the Indy 500, where you go to the pit stop and use your own high-speed tool to change the wheel.
TR: As you note, Deep Blue isn’t as adaptive as a person. You and your colleagues have emphasized again and again that the computer operates by numerical brute force. Why not try to simulate human cognition and adaptability?
HSU: While people are very good at pattern recognition, concept formation, and so on, those tasks are very difficult for computers. Computers can complement humans, however, because they’re good at calculations. So from an engineering point of view, if you want to attack chess problems by computer, you figure out how to use the ability of the machine to calculate fast.
The ability to compute quickly is quite useful in many other fields. One application is called data mining. Big organizations use this technique to extract select information from a vast number of details-for instance, businesses employ it to analyze financial markets. Data mining could also help solve a myriad of problems for individuals, such as the information overload people are now experiencing in the wake of increased access to, among other entities, the Internet. Just as we used our special-purpose chess chips to speed up Deep Blue-employing many of them in parallel-we can create computer systems good for data mining the World Wide Web. Such technology could find and present you with information in a nutshell so that you don’t have to spend your whole life surfing the Web.
TR: Wouldn’t such a tool reinforce what one might call the “quantification fallacy”-the notion that all judgments and decisions can be reduced to calculations?
HSU: That danger exists. But data mining eventually leads to the discovery of empirical findings and rules, after which people stop to figure out why those exist. In other words, we can use computers to extract knowledge from data, but human beings still have to turn that knowledge into wisdom. That’s how humanity progresses.
TR: What’s next, now that Deep Blue has beaten the foremost human chess master?
HSU: Deep Blue’s basic search blueprint is actually not specific to chess. So we’ve started looking at other areas such as pharmaceutical research, where Deep Blue could help design new drugs faster. That’s important, since if a disease is very deadly and also very contagious, we need to be able to fight it with the best tools we have. Toward that end we are designing a molecular-modeling chip-one that can help predict how a candidate drug molecule would interact with, say, the protein envelope of a virus. We plan to install a number of such chips in a computer next year.
TR: Having come this far with Deep Blue, what would you say would actually constitute artificial intelligence?
HSU: Deep Blue would exhibit real AI if it would not allow me to unplug it.