May 1, 1997
Contact: Robert Irion (408) 459-2495; email@example.com
"DEEP BLUE" INSPIRES DEEP THINKING ABOUT ARTIFICIAL INTELLIGENCE BY UC SANTA CRUZ COMPUTER SCIENTIST
FOR IMMEDIATE RELEASE
SANTA CRUZ, CA--When world chess champ Garry Kasparov and
IBM's "Deep Blue" computer program check in for their much-hyped rematch in New York starting on May 3, Robert Levinson of UC Santa Cruz will watch with more than a passing interest.
Levinson, an associate professor of computer science, is an expert on artificial intelligence (AI), the effort to imbue machines with human qualities such as autonomy, the ability to learn from experience, and even intuition and creativity. Deep Blue is impressive, Levinson feels, but it's little more than a brutally efficient chess-playing automaton. In the field of AI, it doesn't even rate a pawn.
"Deep Blue is a powerful entity, and it represents a wonderful engineering effort," Levinson said this week as he looked forward to following the games live on the Internet. "I do agree that it sits somewhere on the scale of 'intelligence.' But even if it proves the most successful approach toward beating the world champion in chess, it's a long way from artificial intelligence. What it really lacks is autonomy and adaptability."
Levinson and UCSC graduate student Jeff Wilkinson outline their views on Deep Blue and AI in a provocatively titled paper, "Deep Blue is still an infant." They will present the paper on July 28 at the 14th national conference of the American Association for Artificial Intelligence in Providence, RI.
The authors describe several ideals toward which the creators of software must strive in order to make their programs weightier on the intelligence scale. Most critical is the notion of autonomy: Can the program manage its computational resources, assess its errors, generalize from its past experiences, and communicate about its progress? In short, can it succeed--and improve--apart from its human architects? Deep Blue barely touches these regimes, Levinson and Wilkinson note.
Further, it lacks what the authors refer to as "meta-reasoning," the ability to study itself and modify its own decisions as a result.
If Deep Blue had these capacities, Levinson believes, it not only could vie with humans at chess but also could excel at many other complex tasks.
"IBM has written a terrific program to play chess, that's all," Levinson said. "But it doesn't know that it's playing chess. It doesn't have a model of itself as a program. If it wins, it can't tell you why. We need computers that can understand computer science. That's the real AI."
Levinson's own interests in computer chess go beyond monitoring Deep Blue's progress. "When I was 10 years old, someone gave me a book on computer chess, and I've been hooked ever since," he recalled. In recent years, he parlayed that fascination into an ambitious, National Science Foundation-funded project on using computer chess as a means to advance toward certain goals in AI.
His group developed a unique chess program, called "Morph," and continues to refine it.
Morph relies on a far different strategy than Deep Blue's
brute-force calculations: It "learns" chess from the ground up, becoming familiar with the intricate patterns that permeate the intellectual's ultimate board game. Essentially, the system mimics how a child would pick up chess given just two things: lists of legal moves and a playing partner who reveals only whether the child wins or loses.
Early versions of Morph could occasionally defeat novice tournament players, but its performance plateaued. The group's newest model relies heavily on information theory to determine which patterns on the chessboard are worth remembering.
"I can point to the actual equations and statistics that the program uses to make these decisions for itself," Levinson said.
"In the past, much of AI has ignored the importance of such mathematical tools." Instead, it became fashionable for researchers to try to mimic the human brain. According to Levinson, that tactic skirted rigorous scientific method.
Levinson feels confident that Kasparov will prevail in his rematch with Deep Blue, but the new and improved machine should present a stiff challenge. It's speedier than its predecessor, which Kasparov vanquished (after losing Game 1) in February 1996. This Deep Blue analyzes 200 million possible moves per second--nearly 40 billion during each three-minute time limit. Chess grandmaster Joel Benjamin joined IBM's team to improve the program's recognition of strong and weak positions.
Even so, Levinson said, "Kasparov has never lost a world championship match, so you have the feeling that he will do what it takes to win. But he won't underestimate the machine. He'll have to make sure he doesn't overlook anything and that he doesn't create too many complications that make it very difficult to calculate many moves ahead."
What if Deep Blue triumphs? "That would be mind-blowing," Levinson said. "People already have a hard enough time understanding why Kasparov plays chess so much better than anyone else."
Editor's notes: Levinson plans to watch all six games of the Kasparov-Deep Blue match live on the Internet from his office at
UC Santa Cruz, in room 334 of the Applied Sciences Building. He welcomes reporters who seek comment on the match and on the general issue of artificial intelligence. Games are scheduled to begin at 12 noon on each of the following days: Saturday, May 3; Sunday, May 4; Tuesday, May 6; Wednesday, May 7; Saturday, May 10; and Sunday, May 11.
Please contact Levinson in advance to arrange a visit. You may reach him at firstname.lastname@example.org, (408) 459-2087 (office), or
(408) 425-0691 (home).
For a copy of Levinson's paper on Deep Blue, call Robert Irion at
(408) 459-2495 (office) or over the weekend at (408) 335-4185.
The above post is reprinted from materials provided by University of California, Santa Cruz. Note: Materials may be edited for content and length.
Cite This Page: