Outspoken, controversial and a fugitive from academe, Doug Lenat is one of the few computer scientists who appreciates the link between artificial intelligence and natural stupidity.
"Even computer programs should be able to learn from their mistakes," insists Lenat, who argues that the current crop of artificial-intelligence programs do "things that any four-year-old child would know is stupid; they lack common sense."
Lenat views knowledge as something that the truly intelligent program should be able to acquire, not something that must be built in.
Top computer scientists from MIT to Stanford assert that Lenat's work at the Microelectronics and Computer Technology Consortium here is defining the future of machine intelligence.
"He's clearly one of the most creative people in the field," said MIT's Marvin Minsky, a pioneer of artificial-intelligence research. "What Lenat has done is started up a whole new field of knowledge."
What sets Lenat's software apart is that it isn't programmed just to solve existing problems; it's programmed to learn the rules about how to solve problems. In essence, Lenat is programming computers to learn how to become experts.
One Lenat program, EURISKO, literally programmed itself to become the world's first designer of three-dimensional computer chips. Another version of the program kicked the stuffing out of its human opponents for two years running in the national championship of a science fiction spacewar game called Traveller -- even after a last-minute rules change.
EURISKO offers just a glimmer of what could prove to be the most powerful implementation of artificial-intelligence techniques yet. Computer programs with the power to test new ideas, make analogies and learn would have applications in everything from weapons design to the genetic engineering of life.
EURISKO signals that computers are prepared to move to a new level of complexity and application, Lenat said: the capability to program themselves.
"His work is highly innovative and often brilliant," said Edward Feigenbaum, chairman of Stanford University's computer science department. "And . . . what he's done is probably the most important work that's been done along the road of modelling creativity and discovery by machine."
Ironically, Lenat was denied tenure at Stanford, where he won his Ph.D -- Lenat believes it was because his research wasn't deemed scientific enough. Now, the 35-year-old computer scientist is at the Microelectronics and Computer Technology Consortium -- a research-and-development group run by Bobby Ray Inman, which made a special effort to recruit him.
"It is the next serious step in artificial intelligence," asserted Al Clarkson, director of man/machine technology at the advanced military science division of ESL, a subsidiary of defense giant TRW Inc.
"It's a radical departure from expert systems," Clarkson said. "If you were a smart venture capitalist, this is what you'd invest in."
The Pentagon and U.S. intelligence officials are exploring the national-security implications of the EURISKO program, Clarkson said.
What sets Lenat's work apart from the rest of the artificial-intelligence intelligentsia is his radically different approach to representing knowledge and learning in computer code. Where most researchers today are intent on transferring human knowledge directly into machine form, Lenat's programs embody such anthropomorphic concepts as curiousity, discovery and, yes, stupidity -- some of the tools through which knowledge is achieved.
"I honestly think that the only way we will learn things is if the program surprises people," Lenat said. "People should be using computers the way astronomers use telescopes: as a new way of looking at things."
Lenat's research represents a vital enhancement of "expert systems," the hottest research area in artificial intelligence. Dozens of companies -- including International Business Machines Corp., Texas Instruments Inc., Digital Equipment Corp., Ford Motor Co. and General Motors Corp. -- are spending tens of millions of dollars to develop expert-systems software for everything from writing software to designing automobile engines.
Essentially, expert systems seek to transfer the expertise of human specialists onto a computer program in the form of rules. For example, certain medical diagnoses could be expressed as chains of IF/THEN rules: IF a certain symptom is present, THEN check the blood-sugar level. The thrust of the expert-systems approach is that all kinds of expertise can be represented in properly ordered sequences of IF/THEN rules.
The problem with expert systems is that they generally address only a tiny base of knowledge, and the rules are fixed, Lenat pointed out. That means expert systems are frozen slices of information that can't adapt to change.
Crudely put, an expert system knows something by rote; a Eurisko system can learn it from experience. An expert system that gives advice on designing a certain kind of computer chip is virtually useless when it comes to designing other kinds of chips. By contrast, a Eurisko system could learn to add to its chip-design knowledge and become more productive.
In working on the design of integrated circuits, for example, EURISKO discovered that symmetry is a desirable property for such chips -- even though it didn't understand why.
Later, when EURISKO designed its conquering Traveller game fleets, the program decided to make them symmetrical -- and justified the decision by referring to its earlier experience in designing circuits.
The business implications of expert systems that could learn to reason by analogy are staggering. These "learning expert systems" could, over time, reprogram themselves to be more productive. Theoretically, they could discover new ways of performing tasks more effectively.
The intellectual theme of Lenat's machine learning research is "heuristics." Heuristics are the intellectual tools -- the rules of thumb -- that people use in the course of attempting to solve problems and make decisions.
The trick is figuring out how to represent these rules in computer code in such a way that a program can succesfully transform them into learning behavior.
Lenat likens intelligent problem-solving to using a map to find one's way. Without the map, a traveller has no idea which way to go. The map makes searching for the right way to go much easier.
The key is constraints, argued Lenat, who asserted that "intelligence is the ability to zero in effectively on a solution despite the apparent size of the search space."
To understand concepts such as "intelligence" and "knowledge," one has to understand the rules, Lenat said.
The importance of the rules of rules burned into Lenat's research efforts when he developed a computer program called AM (Automated Mathematician) over a decade ago for his doctoral thesis at Stanford. AM was a novel approach to expert systems because it was designed to discover new concepts rather than solve old problems.
"AM was probably the only program written that didn't know what it was going to do," Lenat said. "I was ready to build a program that surprised me."
Lenat gave AM a couple of hundred basic mathematical concepts such as equivalence and addition. He then programmed AM to explore "interesting" relationships between concepts to see if new mathematical concepts could be uncovered. Thus, one AM heuristic was "If X is an interesting operation, THEN look at the inverse of X."
AM thus would spew out hundreds of mathematical conjectures and rules based on permutations of these core concepts and heuristics. In a Darwinian form of conceptual evolution, AM derived hypotheses that either proved interesting and true, and thus were kept in the program's repertoire, or were intellectual dead ends and thus were discarded.
This survival-of-the-fittest approach to generating and testing hypotheses according to AM's built-in rules was the heart of Lenat's approach to machine learning.
AM rediscovered such basic mathematical concepts as numbers, set theory, prime numbers, Goldbach's conjecture (that every even number greater than 2 is the sum of two prime numbers) and "several conjectures that don't have names, for good reasons," Lenat said.
Ultimately, however, the AM approach proved sterile. The further away the program explored from its core set of rules, the wilder and stupider its hypotheses became. The "hit rate" for useful hypotheses dropped from more than 60 percent to less than 10 percent.
"AM ground to a halt because, unlike humans," it never learned any new rules based on its mistakes, Lenat said. "It put together concepts that were awful and that people instantly recognized were awful" but that the machine did not.
What Lenat recognized was that, to be truly effective, AM had to do more than just generate new concepts: It had to discover and test rules about discovering concepts. In effect, AM had to learn how to learn. That was the genesis of EURISKO.
"The initial experience was disastrous," Lenat said. "For the first five years, nothing good came out of it."
Then, Lenat figured out what he was doing wrong. Rather than express rules in complicated forms of IF/THEN statements, he decided to break them down into their fundamental attributes. In effect, he based EURISKO rules in terms of key words -- adjectives and nouns -- rather than complete sentences.
Think of a Eurisko concept expressed as a name modified by descriptions. In Traveller, for example, a weapon might be described as a BEAM WEAPON, DEFENSIVE and WORTH 500 POINTS.
Instead of testing something broad, such as an IF/THEN hypothesis, EURISKO also is capable of testing all the components, all the names and descriptions, within that hypothesis and determining what about a hypothesis makes it work or not work.
Mutations create new concepts that are tested against the environment. Some concepts "succeed" and EURISKO adopts them; others "fail" and become the programming equivalent of dinosaurs. It may take thousands of EURISKO generations to spawn useful concepts.
For Traveller, Lenat wrote, "At first, mutations were random. Soon, patterns were perceived: More ships were better; smaller ships were better; etc. Gradually, as each fleet beat the previous one . . . its 'lessons' were abstracted into new specific heuristics."
These improved rules, evolved over hundreds of hours of computer time, enabled EURISKO to design the winning Traveller space fleet.
At that level, EURISKO took off.
But Lenat soon ran into another brick wall. The program needed a broader base of knowledge -- common sense.
"Common sense is everything they assume you already know when you read an encyclopedia article," said Lenat, whose reasearch group at MCC is busy trying to create a common-sense database from an encyclopedia. But that will take years. Consequently, "EURISKO is frozen now," Lenat said. "It will be thawed out once we have this body of common sense."
Armed with common sense to go along with its programmed expertise in a specialty subject, this next-generation EURISKO will be a powerful intellectual tool for design, analysis and discovery, according to Lenat.
An artificial-intelligence system that can learn to become an expert could be the most significant step yet to rivaling human intelligence.