Monday, Sep. 02, 1985
How to Clone an Expert
By Philip Elmer-DeWitt.
It had all the high-tech razzle-dazzle of a consumer-electronics trade show. But most of the computer systems on display started at $50,000 and did a good deal more than play video games. At the booth of a company called Intellicorp, engineers from Ford Aerospace were showing off a program for troubleshooting balky satellites. At the Apollo Computer display, a firm called Visual Intelligence had a system to help nuclear-plant operators quickly interpret the kind of instrument readings that confused technicians at Three Mile Island. On a Digital Equipment computer, newspaper specialists from Composition Systems exhibited a program that lets editors accommodate late- breaking news by reducing from hours to minutes the time it takes to lay out and print a new edition.
The event, held last week at the University of California, Los Angeles, was the Ninth International Joint Conference on Artificial Intelligence, the premier showcase of the most esoteric of computer sciences. Five years ago, a similar meeting at Stanford University was attended by 900 earnest academics. The UCLA gathering drew a crowd of 10,000 that included military brass, investment bankers and scores of corporate headhunters.
Why this sudden interest in artificial intelligence? In two words: expert systems. An expert system is an AI program that captures in computer code the knowledge and informal rules of thumb used by a particular human expert to solve a particular problem. "It's the closest thing to cloning a human mind," says Randall Davis, a professor at M.I.T.'s Artificial Intelligence Lab. Expert systems for use in nuclear power plants, for example, are programmed with the relevant knowledge of the handful of top engineers who know just what to do when a dozen different alarms and signals go off all at once. Once transferred into the computer through a painstaking debriefing of the expert by a team of programmers--a process known as knowledge engineering --this expertise can be made available for consultation round the clock. Similar systems have cloned insurance underwriters, geologists, military field commanders and scores of other human specialists.
The technique behind all these systems can be traced to MYCIN, a computer program written in the mid-1970s by a Stanford physician and computer scientist named Edward Shortliffe. Using tools developed for AI research, Shortliffe boiled down everything he knew about diagnosing infectious blood diseases and meningitis into about 500 "if-then"rules. Rule 27, for example, said that if an organism found in a patient's blood is rod shaped, gram- negative and able to survive in the absence of oxygen, then there is a strong likelihood that the organism is a type of bacteria called Bacteroides. In tests that applied these rules to cases reported in the medical literature, MYCIN was eventually able to diagnose as well as or better than most medical experts.
But MYCIN had a couple of fatal flaws. For one thing, it was designed to replace doctors in their roles as diagnosticians, an approach that was not likely to endear it to its potential users; in fact, it was never actually put into regular hospital use. Moreover, it lacked a quality possessed by most human experts: awareness of its limitations. Confronted with a case outside its narrow field of expertise, MYCIN failed miserably. Says M.I.T.'s Davis: "If you brought MYCIN a bicycle with a flat tire, it would do its best to find you an antibiotic."
Today's expert systems are far less ambitious and much more practical. Shortliffe's current project, for example, is a program called ONCOCIN, which helps cancer specialists plan their patients' drug therapies. As the doctor fills out a series of forms displayed on the screen, the program interprets his data, recommending dosages, watching out for cumulative toxicity effects, searching its store of medical lore for similar cases and offering bits of diplomatic advice. "This is a fundamental shift," says Beau Sheil, manager of product development for Xerox Artificial Intelligence Systems. "More and more expert systems are being configured to act not as experts, but as knowledgeable assistants."
Last week in Los Angeles, these computerized assistants were everywhere. Scientists at Pittsburgh's Carnegie Group had developed a system that helps Navy firemen track the spread of fires on aircraft carriers. Another firm, Teknowledge, was touting a prototype system for the National Weather Service that might someday help predict flash floods and hailstorms. Inference Corp., working with NASA, is developing a battery of expert assistants for the space shuttle that would monitor orbital trajectories, maintain life-support systems and help manage the hundreds of routine glitches that slow up the turnaround time between missions. Says Michael Taylor of Tektronix: "The purpose is not to replace human beings, but to make them more productive."
For every system demonstrated at UCLA, there are many more being kept under wraps. The Department of Defense, for example, discreetly supports artificial intelligence projects in battlefield management and satellite surveillance. The details of many commercial systems are also being closely guarded. One of these is a financial adviser's "workstation," developed by Applied Expert Systems of Cambridge, Mass. Due for release late in September, it is designed to help top brokers sort through the fast-changing array of new investment vehicles and determine which mix of 120 different possibilities will yield the greatest return.
Although expert systems grew out of AI research, there is little danger that these narrowly focused systems will supplant many of man's cerebral, decision- making roles. "This is a nuts-and-bolts business," says Lee Hecht, president of Teknowledge. "The only thing we tell our clients is that we can save them time and money, improve their operations and make them more effective and efficient." Not bad for nuts and bolts.
With reporting by Cristina Garcia/Los Angeles