Classic Computer Magazine Archive COMPUTE! ISSUE 65 / OCTOBER 1985 / PAGE 37



Expert Systems:
Shortcut To Artificial Intelligence?


Kathy Yakal, Feature Writer


The term expert system is rapidly becoming a new catch-phrase, like user-friendly. Some people point to "smart" computers now being used for diagnosis and trouble-shooting in medicine and industry as proof that expert systems are possible and practical. Even some personal computer software publishers claim that their products possess artificial intelligence or expert system capabilities. But others maintain that few, if any, true expert systems really exist. Here's a look at what's happening.


If an "expert" is defined as someone who knows more than most people about a given subject, then you probably seek advice from several experts every week. If you or someone in your family is ill, you probably go to a physician. After asking several questions and running some tests, the doctor arrives at a diagnosis and recommends treatment. If your car keeps stalling at intersections, you probably take it to a mechanic, who checks the car and recommends a repair. If you find yourself owing too much federal income tax on April 15, a tax consultant can offer ways to help. And if you think you've been wronged by someone, a lawyer can usually decide if it's worthwhile to bring a lawsuit.
    All of these people you consult-these experts-are trusted to have a sufficient database of knowledge in certain areas so that their advice is worth following (and worth paying for).
    You can also buy programs for your personal computer that have been designed to act as consultants in such areas as personal finance and health care. Are they replacements for real experts? Not according to their publishers, who stress that the programs are consultants only, and that you should almost always seek additional help from professionals.
    But the day may not be too distant when a new type of computer program will replace experts-or at least, take over part of what experts do. These sophisticated programs, called expert systems, contain a database of knowledge that human experts can spend years acquiring. More significantly, the most advanced expert systems now under development also incorporate some of the rules of logic and analysis that experts combine with their storehouse of facts to solve real-life problems. Already, there are programs in everyday use that analyze geological data to find likely spots for new reservoirs of oil-a job which was formerly the exclusive domain of geologists and engineers.
    Some people even believe that expert systems will become commonplace on the next generation of home computers, bringing the advice of family doctors and other professionals into the home at the touch of a key. But others warn that the premature application of expert systems could result in serious trouble, especially if they're based on an incomplete understanding of the decision-making process.
    Though still in their infancy, expert systems are opening another chapter in the debate over artificial intelligence.

Several years ago, Joseph Weizenbaum, professor of computer science at the Massachusetts Institute of Technology (MIT), wrote a computer program called Eliza. His intention was to show how a computer could act like a psychologist. Eliza would ask the user questions about how he or she was feeling, then pick up on key words or phrases in the answer to guide its "therapy."
    Some people are now calling Eliza an early expert system.
    "I hadn't even heard that phrase used when I wrote it," says Weizenbaum today.
    Part of the challenge of designing an expert system is deciding on the definition of what it's supposed to be and how it's supposed to work: Even the experts can't agree. For example, Weizenbaum thinks Eliza is being characterized as an early expert system because he consulted experts before writing it. Although Eliza may seem like it's really listening to you and responding, the program just follows a set of rules given it by Weizenbaum. If you say you're having a bad day, the program may ask you to talk about it. Then it may ask how certain events made you feel, or what you think you should do about it. Eliza is really more of an interactive diary than an expert.
    Now the term expert system appears to be changing to apply to systems that perform expertly.
    That's still too vague, says Weizenbaum. "If one were to characterize systems that perform expertly as expert systems, then huge libraries of scientific and business programs that have accumulated over the years-many of which are doing a perfectly expert job at whatever they do-would all be expert systems. So it's not a very precise term.
    "Here is an example of something that nobody considers to be an expert system: Today, almost all landings of wide-bodied airplanes are done automatically by onboard computers. I often wonder what the world would be like if that particular work had been done at the AI (artificial intelligence) lab at MIT or Stanford. I don't think we'd ever hear the end of it. But as a matter of fact, it was done, one might say, anonymously. I have no idea who did it, and certainly it does a job that it takes a lot of years to train a human being to do, but it's not considered an expert system. That's odd."

Yet, defining an expert system isn't as simple as pointing to a computer which replaces the performance of a human. Computers have been doing that for years. For instance, though they may not be labeled by some academics as expert systems, process control computers perform functions previously carried out by people with extensive training. "Today, for example, one can see a very large-I mean acres and acres-petroleum processing factory, and if you look very, very hard, you might find two people in these hundreds of acres," says Weizenbaum. "The whole thing is done under computer control.
    "So there's this whole world of computerized process control which has been doing this for a long, long time, and it doesn't think of itself, or hasn't, as expert control."
    Instead, true expert systems seem to be defined according to their evolution and architecture - such as a database of rules and inference mechanisms. Process control computers were developed by other means. "There are lots of process control applications that have been done very well that today might have been tackled differently in the light of expert systems," says Weizenbaum.
    The point at which expert systems cross the border of artificial intelligence is hazier still. To some, there is a definite difference; to others, a perfectly functioning expert system  implies  artificial intelligence.
    Part of the problem is that AI researchers diverge over how to approach the development of expert systems and artificial intelligence. A long time ago, says Weizenbaum, those in the field recognized two fundamentally different ways of doing business.
    The first is to look at AI basically as a branch of psychology; that is, to use a computer to understand the operations of the human mind by programming it do high-level tasks as we think a human mind might do them. The other approach is to program a computer to do very clever things that ordinarily would require human intelligence, but to perform the tasks in ways that might not be considered by (or even possible for) a human being.

These two schools of thought are referred to as theory mode and performance mode. Weizenbaum gives an example of theory mode:
    "Very early on, people got interested in the idea of computers playing chess. It was thought that if we could find out somehow what goes on in a chess player's mind and somehow program that into the computer, not only would we have a good chess-playing machine, but we'd also learn a lot about psychology, about human thought processes. People started trying to do that, but if nothing else, people got tempted to take shortcuts, to take advantage of some features that were built into the computer that no one thought were built into the human mind.
    "So from the very beginning, the temptation couldn't be resisted, and people started designing chess playing programs which took enormous advantage of all the peculiarities of computers but left behind any consideration of how the mind does it. And today we have powerful chess-playing computers, without the slightest claim that they teach us anything at all about human thinking.
    "We've sort of drifted from theory into performance mode."
    And due to a number of circumstances, including the military's interest in and funding of performance mode AI research, says Weizenbaum, there's very little theory work going on today.
    One place where theory work is being pursued is at the University of California at San Diego, in a research center called the Institute for Cognitive Science. Paul Smolensky, one of the researchers there, has been primarily involved in research on neurally inspired mathematical models of learning, memory processes, and problem solving. Using what are currently believed to be some very general characterizations of the brain, Smolensky's work is focused on one primary area: to understand people, and how to educate them and advance knowledge in scientific fields.
    An outgrowth of this research is that it suggests various kinds of novel computers that could be built-such as connecting lots of processors together and letting them work in parallel the same way neurons work in the brain. Only a few prototypes of such machines exist today.
    "There's the platonic idea of what an expert system is, and then there's a whole bunch of actual systems that people have developed that they use the label for," says Smolensky. "I'm not aware of any that are actually in practice except the one that everyone in computer science is aware of, and that's the DEC [Digital Equipment Corporation] expert system for designing installations of their VAX computer systems."
    This expert system, called R1/XCON, was developed by Dr. John McDermott, principal scientist and associate head of the computer science department at CarnegieMellon University. It configures a VAX minicomputer system to the customer's specifications, saving DEC more than $2.5 million annually in field costs. R1/XCON takes roughly a minute to execute the work it took its human predecessors an hour to complete.
    McDermott and a number of other scientists, engineers, and programmers at Carnegie-Mellon have formed a corporation called the Carnegie Group to design and market AI-based systems for commercial applications. The Carnegie Group is looking into many areas that could benefit from expert systems, including engineering design, project management, production management, and sensor-based machine diagnosis and control.

One of the first steps in creating an expert system is to interview the experts the program is supposed to emulate. By asking a series of highly detailed questions, the designers try to figure out the decision-making process they'll attempt to reconstruct in the program. When this thinking process is coupled with a database of facts, the ideal expert system should have a similar capacity for analyzing information and arriving at the right decision.
    A potential flaw has been cited in this approach, however: the difficulty of taking into account the role of human intuition, and even emotion, in decision-making.
    This is a vital point for some critics of expert systems and artificial intelligence. For instance, if you ask someone what the movie War Games was about, they'll probably say something like, "Oh, this kid broke into the national defense system with his home computer and almost started a nuclear war." But the defense system wasn't exposed to this vulnerability until after the government decided that human beings could not be trusted to enter the codes and push the buttons that would launch our nuclear weapons. So the weapons were placed under computer control, because computers would not falter for emotional reasons at the crucial moment.
    "There's a tremendous amount of human judgment that has to go into a decision about whether to give a computer a certain role in a decision-making system," says Smolensky.
    Computers may be able to take over jobs previously done by human beings, but that does not make them intelligent, let alone experts, he says. "Expertise derives in a very significant way from intuition and intuitive processes. Experts do not have any access to that when they introspect about how they do what they do, and no amount of asking an expert questions is going to get at the information and the knowledge that allows the expert to do what he or she does. And if we're going to understand expertise, we have to understand intuition."
    Smolensky warns of the dangers of employing too much technology too fast, especially in areas that have a direct effect on human life. He points out that even when a relatively simple computer system is first installed in a business, there are inevitable last-minute bugs and problems that must be solved before it functions smoothly. "Am it's only because these systems car make a lot of bad mistakes and people can go in and fix them afterward-basically putting Band-Aids on top of Band-Aids on top of Band-Aids-that we don't have a lot of permanent disaster stories.
    "If you look at the problem of making decisions intelligently as something that we can only under stand when we understand intuition, and if you realize that intuition is something that we're not going to understand for a long time, then you realize that we shouldn't be giving computers the power to make decisions that are important."