Groping Toward Humanity

News of Deep Blue's victory over reigning world chess champion Garry Kasparov has been splattered across newspapers everywhere, including our very own Crimson. But what does it all mean? Does it deserve the status of an historic event? Has IBM built artificial intelligence, or does a devoted chess-playing super-computer deserve the same status as an electronic pencil sharpener?

Richard Reti, the great Czech grandmaster of the first half of the 20th century might be though to have proposed an answer in his response to a question as to how many moves he prepared ahead in a chess game. "Only one," he replied, "but it is always the best move."

Reti was joking, of course. Grandmasters typically look about five moves ahead in a given position, and sometimes many more. But Reti's quip captures something at the heart of human chess-playing ability: intuition. And in so far as humans rely on intuition rather than calculation for the decisions of everyday life, his statement captures what is at the heart of human intelligence more generally.

I believe that Deep Blue's victory over Kasparov is, and should be considered, an historic event. But why? Some may claim to have known all along that a fast enough computer would be able to defeat any human, even the very best. If a human is stronger, then the computer only need be that much faster. So if none of this is news, why should May 11, 1997--the first time a computer beat a reigning world champion in a match--be considered the date to immortalize?

All historic events of this type are symbolic. Scientists knew long before they put a man on the moon that it would be possible. Why not scrap the project after the blueprints for the spacecraft and the mission plan had been written and all the key calculations had been completed? In part, it is because one cannot know with certainty whether something is possible until it has been done. Theory stands upon tenuous soil where there are not facts to back it up.

Kasparov still maintains that he will easily defeat Deep Blue in a rematch and that the best humans will always be able to beat computers, "barring human error." Some may balk at the claim and consider Kasparov's excuses of tiredness and lack of spirit to be mere poor sportsmanship, but a part of me (albeit a small part) wonders if maybe Kasparov is right.

It is conceivable that human understanding may be so penetrating, human intuition so deep, that computers cannot physically become fast enough (able to see far enough ahead) to beat the best humans consistently. Even if computers can become fast enough, we may be unable to program their evaluation functions well enough for them to judge (as well as humans can "feel") what moves are best.

The question inevitably follows: Has IBM created intelligence? Unfortunately, this question is impossible to answer without a common understanding of what intelligence is. As a friend of mine put it, "the only thing that everybody agrees on is that the IQ test does not measure it."

What we can agree on, however, is that IBM has created a computer that plays chess very well. And insofar as playing chess well is something that we think only intelligent beings can do, it can thus be said that a computer was constructed which displays one aspect of intelligent behavior. But one might object that Deep Blue just does a whole lot of calculations very quickly, and that's not intelligence.

Indeed, most modern computer scientists would not consider Deep Blue an intelligent machine. And yet, ironically, this very fact underscores how much has been learned in the field of artificial intelligence. This is because the earliest efforts to create intelligent machines involved trying to program computers to play chess. In the first experiments, computers were modeled to play chess like humans--in a given position, the computers examined a small set of moves according to the heuristics that good human players seemed to use. The scientists were shocked to learn that the computers played miserably.

Computer scientists very quickly gave up trying to program computers to play like humans. Experiments that took advantage of the brute calculating force of the machines--where the computer analyzes every possible legal move in every position--produced much stronger moves. As computers got faster and the evaluation functions were made better and better, computer chess players just kept getting stronger.

Recently, however, progress has slowed. In 1987, chess computers may have been twice as skilled as computers in 1977, but computers in 1997 are not twice as good as they were a decade ago. The reason is that chess is an enormously complicated game. Even if computers continue to get faster at a faster rate, they will be better players than their predecessors by smaller and smaller amounts. This is because looking one move further is exponentially more time consuming the more moves a computer looks ahead.

Interestingly, programmers have reacted to this law of diminishing returns to faster processing by reintroducing heuristics into the computer's algorithm. Grandmaster Joel Benjamin spent a year "training" Deep Blue for its match with Kasparov. What did Benjamin do? He helped the programmers refine Deep Blue's evaluation function to incorporate human "feeling" about how good a position is and he helped Deep Blue zone in on the kinds of moves a human grandmaster would intuitively leap to analyze. Of course, Deep Blue relies primarily on brute force, and it is still very much an open question what the optimal mix of brute force and heuristics is.

But its programmers were remarkably successful in making Deep Blue's play resemble the best aspects of human chess players. Grandmaster commentators repeatedly referred to its moves as "creative" and as being indicative of "a deep understanding of chess." In fact, Kasparov, who trained for this match by preparing against "computer-ish" moves, thought Deep Blue's moves were so unlike the typical computer that he has accused the IBM team of tinkering with Deep Blue during the game.

This accusation raises an important point. If computers become so good at an intellectual task that their actions are indistinguishable from those of humans, then it becomes difficult to see why it matters whether neurons and synapses or circuits and C++ are behind the moves. At least in that task, it seems, the computer has become, for all intents and purposes, intelligent.

It must be remembered, however, that although chess is extraordinarily complex, it is inevitably finite--there are always a concrete and limited number of possible moves. By contrast, humans face an infinite number of choices in everyday life, and here, brute force is not an option. Computers, then, will need to find ways to grasp the infinite if they are ever to earn the stamp of true intelligence.

Daniel J. Benjamin '99 is president of the Harvard Chess Club.