Putting Knowledge to Work

By EUGENE LINDEN

Although supercomputers are dazzling in their power and engineering virtuosity, hardware alone will only partly achieve the eventual goal of computer scientists: the creation of systems that can mimic the decision- making powers of human beings. This goal is called AI, for artificial intelligence, and it has eluded computer programmers for decades. Now, however, even as supercomputers open up new worlds of possibility, researchers are taking major strides toward making their machines both smarter and more versatile.

Their work has spawned a new phase of the great computer revolution that has been going on for the past 40 years or so. Whereas the early use of computers revolutionized information handling, late developments promise to better manage raw computer power and the increasing complexity of modern information technology. For the first time in history, these systems allow computers to deal with ambiguity and questions of judgment that are too subtle for conventional data processing, however powerful. After years of false starts and overblown promises, the new systems, called expert or knowledge-processing systems, have exploded onto the commercial scene in the U.S., Western Europe and Japan, which is also trying to develop AI technologies. "We have spent hundreds of billions of dollars developing computer power that has set us adrift in a sea of data," says Thomas P. Kehler, CEO of IntelliCorp, a California software company. The new systems promise to put that information to work.

Eighteen months ago scarcely a handful of these systems existed in business and government. Now there are an estimated 1,000 to 3,000 in daily use, and the number is increasing by 50% annually. They grew out of much touted artificial-intelligence research into human decision making in the 1960s and '70s. AI thus far has failed to reduce human intelligence to hardware and software. But in the quest to build machines that see, move, communicate and think like humans, AI has produced offshoots with evident commercial potential. Says Herbert Schorr, who spearheads IBM's efforts to commercialize AI: "Knowledge processing allows you to handle new, tough problems that are too costly or too painful to do with conventional programming techniques."

Commercial systems derived from artificial intelligence suddenly seem to be everywhere. Some examples:

-- At American Express, a new computer system contains the cloned expertise of platoons of specialists who approve unusual credit requests for the company's estimated 20 million U.S. cardholders. For the first time, the computer will decide whether to okay the purchase of, say, a $5,000 Oriental rug by a usually prudent spender -- or nix the transaction on the suspicion that the cardholder is on a buying spree.

-- In their supersecret war on terrorism, U.S. intelligence agents routinely consult a specially developed computer system, programmed with the arcane knowledge of a handful of terrorism experts, to anticipate and avert terrorist actions. The year-old system has reportedly helped predict terrorist attacks in Western Europe.

-- By the end of next year, Ford dealers across the country may no longer have to call Dearborn, Mich., to talk with Company Expert Gordy Kujawski every time they run into a hard-to-diagnose engine problem. Instead, they will simply plug into a new nationwide computer system developed by Ford to duplicate the reasoning Kujawski uses to untangle the knottiest problems.

Two years ago IBM's Schorr proclaimed the "second wave" of the information revolution. "While the first wave automated data processing," he said, "the second wave will automate decision making." IBM now considers itself the world leader in second-wave technology and is either using or developing expert systems throughout the company. Big Blue's claims to leadership, however, get spirited argument from companies like Digital Equipment Corp. and E.I. du Pont de Nemours. They and others are using second-wave technology not only to bring computers to bear on problems that until now have been bypassed by the information revolution but also to extend the range and availability of human expertise. Says Edward Feigenbaum, an AI pioneer and co-author of a | forthcoming book on second-wave success stories: "Every system we have looked at improved productivity by more than an order of magnitude -- that's like the difference between a car and a jet plane."

Despite these gains, current systems operate within strict limits and too often behave more like idiots savants than experts. Second-wave systems as yet have no common sense or awareness of the world outside their narrow slice of expertise. At high-tech redoubts like Xerox's Palo Alto Research Center in California, scientists are planning decision-making systems that will behave more like real experts. Example: an all-purpose electronic repairman that uses knowledge and common sense about electricity to diagnose any problem put before it. At Xerox and elsewhere, other scientists are examining the very foundations of artificial intelligence. Their aim: a theory that will enable them to build computers that can step outside the limits of a specific expertise and understand the nature and context of the problems they are confronting.

Still, the impetus behind second-wave technology is not its potential but what it can deliver now in financial returns and improved productivity. In April 1986, IBM brought on line its first expert system, called DEFT (for Diagnostic Expert-Final Test). Its task: to perform the mundane but critical job of diagnosing problems during the final testing of the giant disk drives that store information for IBM's mainframe computers. Since then the testing system has been adapted as a diagnostic tool for IBM service experts and to perform a variety of different tests on IBM equipment. IBM's initial cost: roughly $100,000. The payoff: $12 million in annual savings.

At Xerox, a leading U.S. manufacturer of copying machines, expert systems like RIC (for Remote Interactive Communications) are giving the first practical hints about what the second-wave revolution will mean. Employing the reasoning of a special Xerox team of diagnosticians, RIC reads data from a copier's internal instruments, senses when something is about to go wrong, and sends a report to a repairman, who can warn the customer that an imminent breakdown can be avoided by taking appropriate steps. Theoretically, Xerox copiers hooked up to RIC systems should never break down.

The technological lineage of RIC and almost every other second-wave system can be traced back to Mycin, an expert system written at Stanford in the mid- 1970s. Named for a group of antibiotics, Mycin was the brainchild of a Ph.D. candidate named Edward Shortliffe, who designed it to help physicians diagnose certain infectious diseases and choose appropriate remedies. After painstakingly interviewing doctors about the process of diagnosis and treatment, Shortliffe and company programmed Mycin with some 500 rules to guide its decisions.

Unlike the basic unit of conventional computer programming -- the algorithm, which details a precise series of steps that will yield a precise result -- those rules (referred to in computerese as heuristics) state a relationship that is likely, but not guaranteed, to yield an outcome. Heuristics allow computers to deal with situations that cannot be reduced to mathematical formulas and may involve many exceptions. It is the kind of reasoning that governs countless everyday decisions, ranging from the mundane, such as choosing the appropriate clothes for a job interview, to the apocalyptic, such as deciding whether a Soviet missile launch is a routine test or an all-out attack.

Mycin took some 20 man-years to complete. It turned out to be more accurate than the humans against whom it was tested: in one trial, the system prescribed the correct treatment 65% of the time, in contrast to human specialists, who were right in 42.5% to 62.5% of cases. Still, Mycin did not have a clue that it was diagnosing a human being, nor did it have any idea what a human is. In fact, it was perfectly capable of trying to prescribe penicillin to fix a broken window. All it could do was rigidly test the applicability of various rules to pieces of data. This led critics like Joseph Weizenbaum, a professor of computer science at M.I.T., to dismiss expert systems like Mycin as "Potemkin villages. You move a little to the left, and you see it's all a facade."

While Weizenbaum and other critics insisted on measuring Mycin against human intelligence and knowledge, others looked at the system and saw a computer- handling expertise that had previously resisted automation. No one, however, was going to build expert systems if they took several years to construct. Solution: create a Mycin without medical knowledge -- in effect, construct an empty shell into which programmers could pour all kinds of different expertise. In 1977 a team of Stanford researchers under Feigenbaum dubbed the new shell Emycin (for Empty Mycin) and used it to build several more expert systems. Emycin spurred a number of start-up companies, led by AI entrepreneurs like Feigenbaum, to build knowledge shells for the commercial market.

The second wave had a rocky start. Too often, enthusiastic young computer nerds babbling in technospeak would sell flashy systems to computer-dazzled counterparts in the research divisions of Fortune 500 companies. In turn, the corporate techies built glitzy prototypes that ran on exotic hardware. By the mid-1980s it became clear that both groups had missed the point: big companies did not want sexy technology for its own sake; they wanted solutions to business problems. Consequently, a number of once gung-ho companies began to sour on artificial-intelligence technology as expensive and impractical.

What saved the fledgling industry was the discovery that applied artificial intelligence could produce concrete results when properly used. In 1978 the Massachusetts-based Digital Equipment Corp. joined forces with AI Theoretician John McDermott of Carnegie Mellon University to develop XCon (for Expert Configurator), a system to assist salesmen in choosing parts for DEC computer systems from among tens of thousands of alternatives. XCon went on line in 1981, and for several years it was the only expert system in commercial use that companies could employ to gauge the worth of their technology. Today XCon configures almost every Digital computer system and saves the company $25 million annually.

IBM was a latecomer to the second wave. It was not until 1984 that Schorr, a respected computer designer within the company, was assigned to take corporate responsibility for artificial-intelligence projects. "Three years ago it became apparent that this technology had gone past the research phase and had become commercial," he recalls. "IBM decided we could make money in it, and that we should be the world leader." Cautiously at first, IBM began to search for opportunities to apply expert systems internally -- for "the low-hanging fruit," as Schorr puts it today.

After the company's first shot -- the DEFT system to diagnose troubles in IBM's giant disk drives -- proved a bull's-eye, IBM Chairman John Akers became an enthusiast. He gave Schorr the green light to promote expert systems throughout the company. IBM now has 50 knowledge systems up and running, and Schorr expects that number to double each year for the next few years.

If IBM, as Schorr claims, is the world leader in applied artificial intelligence, Du Pont is running close behind. Ed Mahler, the Delaware multinational's program director for artificial intelligence, says the company currently has 200 knowledge systems in use and expects to have 2,000 systems running by 1990. The reason for the explosion, according to Mahler: knowledge- processing technology is now affordable, and even the most sophisticated systems are available on personal computers.

Rule-based systems such as XCon and DEFT, however, still have drawbacks. When asked a question, the expert system blindly searches through its data base to see which rules apply, then searches through the data base again to find the data for the rule. More sophisticated knowledge systems store information in frames, which organize it along with its relevant attributes. AI Pioneer Marvin Minsky of M.I.T. noticed that when people enter a room, they have a set of expectations about what they will find -- a desk or chair, perhaps, but certainly not, for example, an ocean. His idea was to package information in a way that accommodates those expectations: a room might also contain a bed, window and lamp. Minsky's frame concept allowed for more efficient use of the computer by enabling it to find what it needed directly, minimizing blind searches.

Growing numbers of U.S. companies are no longer arguing about whether second-wave technology is worth adopting; instead, they are concerned about how best to use it. They are finding all sorts of ingenious applications. United Airlines has developed a simple frame-based system called GADS (Gate- Assignment and Display System) to help prevent the infuriating delays that occur when weather and scheduling problems scramble gate assignments for incoming planes. The system encodes the reasoning that gate controllers use when scheduling gate assignments (for example, two adjacent gates cannot accommodate two DC-10s at once). Before GADS, United's gate controllers would physically move magnetic pieces around on a big metal board. Now they use GADS for playing what-if games to head off problems long before they develop.

The military has tapped second-wave technology in its efforts to come to grips with the complexities of modern warfare. The Navy monitors the strategic status of the Pacific fleet with a system that tracks 600 ships, submarines and aircraft and alerts the fleet commander to changes in readiness and the probable impact of those changes. The system analyzes everything that affects readiness, from firepower and fuel consumption to morale (which it estimates by keeping track of the time that has elapsed since a ship's last shore leave). Complex fleet-deployment problems that used to require several days can now be resolved in a matter of hours.

More ambitious is ALBM (AirLand Battle Management), a system designed to address every aspect of planning and fighting an air and land battle. ALBM is intended to supply computerized intelligence to the "electronic battlefield" that the military has been developing as part of its evolving command-and- control strategy. When completed, this system will enable commanders to explore war games and battle scenarios, test tactical hypotheses and plan weapons and troop deployment. But the information-processing requirements of a major-theater war would be enormous. Managing a battle is not a case of dealing with one source of data rapidly but, rather, simultaneously processing data about air threats, supply lines, weather and the positioning of hundreds of thousands of soldiers. Whether ALBM will be a device for war games or a battlefield tool depends on the military's ability to harness the power of massive parallel processors -- computers with thousands of processors that can work simultaneously on a problem.

Processing power is an even more daunting problem for Pilot's Associate, a knowledge system the military hopes to field in the 1990s. The device is designed to advise electronically a fighter pilot in combat about everything from weather to ground and air threats. It will include several expert systems with sophisticated three-dimensional data bases. But if it is to deliver its advice effectively to pilots who have only seconds to respond and act, this system too will require putting into fighter aircraft the type of computing power that today fills entire rooms.

Moreover, even the developers wonder whether pilots in a crunch will trust their lives to silicon advisers. Chris Spiegl of Texas Instruments, which is developing the system with McDonnell Douglas, notes that to better their concentration, many pilots begin turning off automatic systems the closer they get to combat.

Pilots are not the only ones worrying about the reliability of sophisticated military expert systems. Terry Winograd, an AI pioneer turned critic who is now at Stanford, has formed a Palo Alto-based group called Computer Professionals for Social Responsibility to oppose the use of second-wave systems in military applications. Winograd believes that isolating experts from the unforeseen consequences of their decisions is "perhaps the most ! subtle and dangerous consequence of the patchwork rationality of present expert systems." He is specifically concerned about the use of expert systems in President Reagan's Strategic Defense Initiative, or Star Wars system. In the 1960s, Winograd notes, a computer system announced a Soviet attack when radar signals bounced off the moon, an occurrence that had not been anticipated by the programmer. He contends that the potential for similar errors is greatly magnified with expert systems.

In the near term, the future of the second wave will involve novel applications built with existing software technology such as frames and rules. It has already produced some unanticipated benefits. Companies have discovered, for example, that their engineers use the technology as a reasoning tool. While in the past they would tell a programmer what they wanted in the way of a computer application and hope for the best, now they are creating prototypes for their own systems, then fiddling with them until they are right.

Others, however, are already thinking beyond existing technologies. Johan de Kleer, a respected knowledge-system designer at Xerox, envisions an all- purpose electrical diagnostician that would have specific knowledge, such as the various laws that govern electrical flow and conductivity. But it would also have the common sense to decide whether it was faced with a broken VCR or a broken computer. To build this system, de Kleer has spent ten years codifying what he calls "qualitative" calculus that will provide the language to build "common-sense physics." The problem with common sense is that it requires the computer to skip nimbly among many different perspectives in order to find the approach that best fits a problem. The computer must be able to simultaneously maintain the assumptions underlying these different perspectives, and de Kleer says that this, again, will require massive processing power. He looks to parallel processing for the power to run his systems. "Running my applications on a serial supercomputer would require all the computer time in history," he says.

De Kleer's diagnostic systems are at least five years away. Even further out are general-knowledge systems that would not be limited to a specific function or even a preset agenda but would instead be able to respond appropriately to unexpected tasks and problems. To develop such systems, a rebel generation of AI scientists believes that it is necessary to rebuild their field from the ground up. Their emphasis, says Philosopher Daniel Dennett of Tufts University, is on figuring out how people manage to accomplish the plain, everyday things that account for most human behavior, rather than on creating a mathematical model of the intellect, as an older generation of AI researchers have tried and failed to do.

Some breakaway AI scientists suspect that the answer to this problem lies in the way humans respond to the context of an activity or conversation. Brian Smith, a Xerox theorist studying the foundations of knowledge, believes that people derive a tremendous amount of information from the physical setting of a conversation, and that meaning that is not evident usually emerges during the dialogue. If he can reduce this process to theory, Smith believes, it will be possible to build a machine that would know what is meant in ordinary human conversation. Getting a machine to act on what it "understands" is yet another problem. Stan Rosenschein, former head of artificial-intelligence research at SRI International in Menlo Park, Calif., is testing a robot called Flakey, which he hopes will have the ability to carry out a physical chore like delivering a package. Right now Flakey can follow only simple, specific instructions, then exclaim, "I did it! I did it!"

How long will it take before machines are developed that are truly intelligent and able to make their way in the world? It is, of course, foolish to predict where any new technology is going or when it will get there. But in his 1986 book on artificial intelligence, Machinery of the Mind, Science Writer George Johnson offers a guidepost. He recounts the story of a Chinese student who became disillusioned with the study of artificial intelligence. It was as if, said the student, a modern American had asked an ancient Greek to build a television, then offered only the information that TV is a system that projects images across long distances; logically, the ancient might proceed to place a long sequence of reflecting mirrors across the landscape and claim to have built such a machine.

That, the student concluded, was an apt analogy to AI. Like their Greek counterparts, AI scientists can build crude models and they have a rough idea of the principles and properties involved in achieving their goal. But it may be centuries, if ever, before all those elements are sufficiently understood to enable mere mortals to fulfill the dream of AI: to create electronic replicas of themselves.

With reporting by Scott Brown/Dallas and J. Madeleine Nash/Chicago