Intelligent Computers? We Still Have A Long Way To "Go," Says Professor

greenspun.com : LUSENET : Human-Machine Assimilation : One Thread

Intelligent Computers? We Still Have A Long Way To "Go," Says Professor ROLLA, Mo. -- If you were impressed with Deep Blue's 1997 triumph over chess master Garry Kasparov, wait until you see the computer that can hold its own with the best players of Go, the ancient Asian game that is astronomically more complex than chess. Some of the world's leading Go players claim it will be a century before such a machine exists, but Dr. Donald C. Wunsch is more optimistic. Wunsch, the Mary K. Finley Missouri Distinguished Professor of computer engineering at the University of Missouri-Rolla, believes the era of truly intelligent machines -- even computer programs that can match wits with a Go master -- is within reach.

The creation of such a machine will involve an "adaptive" computer system that is much more intelligent than today's best computers, Wunsch says. To succeed at Go, he says, such a system would have to be intelligent enough to learn the game, adapt to changing conditions, anticipate several possible moves, and evaluate the "goban," or playing board -- "to help it learn which side is winning and which is losing." And it would have to be "smarter" than Deep Blue, the IBM computer that defeated chess master Garry Kasparov in May 1997.

Intelligent systems of the future also must learn to "think" paradoxically -- suspending judgment of contradictory ideas -- in order to better mimic human intelligence, Wunsch says.

A masterful Go machine "is going to require more human knowledge than chess requires -- and chess requires a lot of human knowledge," says Wunsch, who is an expert in the field of intelligent computer agents.

"The human experts in Go believe it will be a century before we can make a computer as good as a Go master," Wunsch says. "I'm more optimistic, but we still have a long way to go."

Go is a strategy game involving two players. Players take turns placing their markers, called stones, on the goban, a board with grid lines, with the objective of controlling the largest part of the goban. The game is as popular in East Asia as chess is in Russia, Wunsch says. Like chess, Go is beneficial to the intellectual development of school children, he adds.

Today's best Go computer programs play the game "like an amateur -- like someone who has been playing for about six months," says Wunsch, who claimis his own Go skills are "so weak that the stronger computers can beat me."

Go is "easier to learn than chess, but more difficult to master," Wunsch explains. The difficulty in training a computer to play Go is a sheer numbers problem, he adds.

"The game tree in Go has about 10 to the 750th power of possibilities," Wunsch explains. "Chess, on the other hand, is more like 10 to the 150th power. In chess, the possible number of games is greater than the number of atoms in the universe."

So how did Deep Blue become such a chess whiz? According to Wunsch, the answer lies in Deep Blue's parallel processing power, combined with simple "rule-of-thumb" techniques known as heuristics. Thanks to parallel processing -- the use of several processors at once to share the work of evaluating possible moves -- the computer can evaluate an incredible number of possible future game positions. By evaluating current board position, king safety and other variables, the heuristics enable Deep Blue to eliminate a large number of moves.

"Computers aren't good at looking at the static position -- to see who's winning at a given moment -- but they are good at looking at large numbers of possibilities," Wunsch says. "A chess program will look at the board and say, in effect, 'There are 20 possible moves, but 10 of them are lousy.' By doing that, it narrows the possibilities down to a more manageable size."

Still, Deep Blue took a "brute force" approach to processing, Wunsch says. Deep Blue can search through a century of historical chess moves at speeds up to 200 million positions per second. A more truly intelligent approach is needed if computers are to make the leap Wunsch believes is possible in the coming decades.

Future intelligent systems will require more emphasis on what Wunsch calls "massively parallel learning systems." Much of the research in intelligent agents -- from Internet search engines and computer spell-checkers to systems used in high-tech manufacturing -- has focused on developing the "learning systems." Now it's time to emphasize the "massively parallel" part of such systems.

A truly intelligent system is one that can adapt to changes or new information and learn from the changes, Wunsch says.

"'Adaptive' and 'intelligent' should be redundant," says Wunsch, who received a National Science Foundation Faculty Early Career Development (CAREER) Award to support his research into intelligent agents. "To call an agent intelligent that is not adaptive is missing the mark." Many Internet search engines, for example, are not adaptive -- and therefore not truly intelligent -- because they cannot change their databases in a variety of necessary situations, such as when they encounter a defunct World Wide Web site, or when a user ignores the search engine's top hits.

"An adaptive intelligent agent is one that is capable of learning," Wunsch says.

Traditional computer programming, however, limits a computer's ability to think like a human. Wunsch hopes to see future intelligent systems that move beyond the raw logic of most systems.

"We need to create logical systems that are able to embrace contradictions, as humans do," Wunsch says. "The computers need to be able to reason on the data and still be able to prevent coming up with absurdities. We're talking about a system that could constrain contradictions."

"Everything we know about computers is still in the stone ages," Wunsch says.

"They're very clumsy, they require experts -- they have all kinds of problems," he says. "They should be thought of as prototypes. Intelligent systems are in their pre-infancy and there's a long way to go."

But that's not to say intelligent systems aren't going to change the way we view computers -- and perhaps sooner rather than later. "Paradoxically, I think we're closer to those systems than most people realize," Wunsch says.

Editor's Note: The original news release can be found at http://usenet.umr.edu/cgi-bin/news-view?mode=post&group=umr.news.releases&article=2515

-------------------------------------------------------------------------------- Note: This story has been adapted from a news release issued by University Of Missouri-Rolla for journalists and other members of the public. If you wish to quote from any part of this story, please credit University Of Missouri-Rolla as the original source. You may also wish to include the following link in any citation:

http://www.sciencedaily.com/releases/2000/03/000314072236.htm

-- scott (lynx5_5@hotmail.com), March 15, 2000


Moderation questions? read the FAQ