EXPERT SYSTEMS
Computers as sages
by Howard Rheingold


Howard Rheingold is the author of Software Odyssey and co-author of Higher Creativity.

Should you ever want to drill for oil, diagnose a disease or synthesize a new molecule, you can ask Prospector, MYCIN or Dendral for some sage advice. They are certified experts in their respective fields. They are also computer programs.
    We all depend on expert assistance-from doctors, attorneys, automobile mechanics, computer repairmen. Wouldn't it be nice to have our own experts? Right now, only large institutions like medical schools or oil companies can afford to purchase cybernetic expertise, but we may soon see the day when "intelligent assistants" are sold off the shelf for personal computers along with word processing, spreadsheet and games programs.

Higher IQ
"Knowledge engineering" or expert systems started out as pure research, and were in fact developed and studied by artificial intelligence laboratories for more than a decade before knowledge engineering began to be offered commercially. As commercial products, expert systems are already serving as "intelligent assistants" for experts in fields as diverse as medical diagnosis and higher mathematics, chemistry and geology, computer design and genetic engineering.
    Chemists can ask the computer program Dendral to "Display the chemical structure of this compound we just invented" and expect to get an answer. Oil companies are paying significant fees to query Prospector: "Here's a whole bunch of data about geological site X. Where's the best place to drill for oil?"
    Cyrus, a program developed at Yale (purely for research purposes), can answer any question you might want to put to former Secretary of State Cyrus Vance.
If you ask, "When did your wife meet Begin's wife?" it will reply "At a state dinner in Israel in January 1980," even though that fact is not specified in its data base and the answer is inferred from other biographical information.
    Using an expert system is relatively simple. MYCIN, for example, can be fed a configuration of symptoms exhibited by a patient. If these symptoms can lead to diagnosis of a disease the program "knows about," it will ask the attending physician for more input about specific symptoms to narrow down the possibilities until it can venture a diagnosis. When the physician questions the computer about the specific line of reasoning the program uses to make a decision, the expert system is able to explain its decision through a simple human-language dialogue:

expert systems



MYCIN: HAS THE PATIENT RECENTLY HAD SYMPTOMS OF PERSISTENT HEADACHE OR OTHER ABNORMAL NEUROLOGIC SYMPTOMS (DIZZINESS, LETHARGY, ETC.)?
HUMAN: YES.
MYCIN: HAS THE PATIENT RECENTLY HAD OBJECTIVE EVIDENCE ABNORMAL NEUROLOGIC SIGNS (MUSCLE RIGIDITY, COMA, SEIZURES, ETC.) DOCUMENTED BY PHYSICIAN OBSERVATION OR EXAMINATION?
MYCIN: THE CSF CULTURES WILL BE CONSIDERED TO BE ASSOCIATED WITH MENINGITIS. PLEASE GIVE THE DATE ON WHICH CLINICAL EVIDENCE (SYMPTOMS, SIGNS OR LABORATORY TESTS) OF MENINGITIS FIRST APPEARED. PLEASE ENTER CSF FINDING. IS MENINGITIS A HOSPITAL-ACQUIRED INFECTION?
HUMAN: WHY?
MYCIN: THIS WILL AID IN DETERMINING THE ORGANISMS (OTHER THAN THOSE SEEN ON CULTURES OR SMEARS) WHICH MIGHT BE CAUSING THE INFECTION. IT HAS ALREADY BEEN ESTABLISHED THAT
(2.1) The infection which requires therapy is meningitis, and
(2.2) organisms were not seen on the stain of the pending csf culture, and
(2.3) the type of infection is bacterial, and
(2.4) neurosurgery is not one of the relevant items from the history of this patient, and
(2.5) the age of this patient is greater than ten days, therefore, if
(2.6) the infection was acquired while the patient was hospitalized then there is evidence that the organisms (other than those seen on cultures or smears) which might be causing the infection are e. coli (75), staphylococcus-coag-pos (3), pseudomonas-aeruginosa (5).

    With software like Meta-Dendral you can build systems such as Prospector, Cyrus or MYCIN by guiding the acquisition of knowledge from a human expert with appropriate queries and embodying that knowledge in a problem-solving program that can answer questions about its line of "reasoning."
    Bacon, a program developed at Carnegie-Mellon University, is named after the man who brought the inductive method to science. (Francis Bacon was also the first to proclaim, in 1620, that "knowledge is power.") Given a set of facts about a consistent system of knowledge, this program can make inferences about the system in a way that mimics human-style creative problem solving. It looks at a collection of facts and tries to guess at general laws, tests the laws against the facts, refines the guess and produces a theory.
    If Bacon is fed everything the astronomer Johannes Kepler was likely to have known about the cosmos, the program can independently produce Kepler's Third Law. Bacon, however, takes only about one minute to arrive at the conclusion Kepler spent a lifetime formulating. The program also "discovered" Ohm's law. While rediscovering the great insights of history is probably of little use outside a classroom, nobody will worry about its uselessness if and when one of Bacon's software successors discovers something new.
    Is the future successor to Newton and Einstein just a newborn microprocessor awaiting assembly and programming? The question of whether an artificial inference engine can ever discover anything new and significant goes directly to the heart of the artificial intelligence controversy.

Can Machines Think?
This question has been vigorously debated ever since Alan Turing first formalized it a quarter-century ago. It is also a big question-the kind that requires big science, mountains of high-tech hardware, armies of progammers, and institutes full of theorists.
    Lately, more restricted, more manageable and, to some, more profitable questions are being raised:. Can a program formulate a theory? Can a computer use rules to extract answers from an information base and tell us how it did this, the way a human expert can? Finding an answer to this last variant was the original goal of the first expert systems researchers in the mid-1960s.
    Expert systems as they exist today are composed of three parts: a base of task-specific knowledge, a set of rules to make decisions about that knowledge and a means of answering questions about the reasons for its decisions. The "expert" program knows what it knows, not through the raw volume of facts fed to the computer's memory, but by virtue of a "reasoning" process of applying the rule system to the knowledge base; it chooses between alternatives not through brute-force calculation, but through some of the same rules of thumb that human experts use.
    Before you can construct a program capable of making expert decisions, you have to be able to define what distinguishes an expert opinion. How do human experts gain expertise? First, they learn the rules for the kind of reasoning needed. The rules of English grammar constitute the basic unit of reasoning if American literature is the field of expertise, and mathematics and physics furnish a different set of rules if electrical engineering is the field.
    After learning how to learn by practicing the basic method of gaining knowledge within the chosen field, the human expert spends a long time learning specific facts. Part of what we need to become experts is a large amount of task-specific information. (Cognitive scientists who study human expertise now say that experts also need a lot of knowledge outside the specific field of expertise in order to have a large pool of available analogies.) Then there is a phase of "hands-on" direct experience in the field-an internship, in which the learning comes directly from the necessity of making decisions in the real world.
    Finally, after years of study and practice, the human expert knows how to look at a problem and see solutions that nonexperts are unable to see. This ability to sift through the possibilities (is it meningitis or the flu? an oil deposit or worthless shale?) and quickly reach one decision out of myriad alternatives is ultimately what distinguishes an expert.
    The unadorned statistics on how often experts turn out to be right is the ultimate criterion for their expertise-whether the expert is a person who studied for years or a computer program that was literally born yesterday. Research conducted at the Stanford Medical School found MYCIN to be more effective than most physicians who are not specialists in diagnosing bacterial infections, and 80 percent as effective as those physicians who are themselves experts in the field.
    It turns out that you can't just feed all the known facts into a computer and expect to get a coherent answer. That isn't the way human experts make decisions, and apparently it isn't the way you coax a computer into making a decision. What you need is an "inference engine" to fit together the rules of the game, the body of previously known facts and the mass of new data, then venture a guess about what it means.

Future Computer Experts
Expert systems are a key element in the so-called fifth generation of computers planned for the 1990s. These are devices that will tell you not only what you want to know, but also how to find out without learning a computer language. Instead of requiring the computer user to try to think like a computer, fifth-generation devices will be expert at helping users figure out what they want to know.
    It's hard to argue with a molybdenum deposit or a statistically significant high rate of successful diagnoses. As the debate over whether software is capable of acting intelligent dies down in the face of what mathematicians call "existence proof," the question of how much computer technology ought to be applied to such areas as medicine, air traffic control, nuclear power plant operations or nuclear weapons delivery systems is just beginning.

expert systems



COMPUTER WORMS
Late in the 1970s, at Xerox's Palo Alto Research Center, scientists raised the specter of artificial life by creating software "worms" that jump through networks and replicate themselves on idle machines.
    Initially intended to test security in the Xerox network, these worm programs are composed of many individual segments-smaller programs that operate on different machines. From the computers where they begin life, individual programs can migrate to any accessible network and take over the resources of "cooperating" computers.
    Xerox researchers John Schoch and Jon Hupp actually designed and tested different types of worms. The simplest species, an "existential" worm whose purpose was to stay alive in a network even when some of the on-line computers were turned off, was later modified to display the message I'M A WORM, CATCH ME IF YOU CAN on the screen of whatever computer it was inhabiting. The "town crier" worm carries a message with it; the "alarm clock" worm replicates, then performs a specific task on cue; the "diagnostics" worm goes troubleshooting.
The name "worm" was taken from Shockwave Rider, John Brunner's story about an authoritarian government whose existence depends on an omnipotent network of computers. The government falls when a rebel programmer lets loose an unstoppable "tapeworm" in the network. From here, it's not a far jump to the notion of hostile worms raiding unsuspecting computers to "liberate" information or generally wreak havoc.
    Indeed, Xerox has already experienced a renegade worm. Arriving at work one day, researchers found that more than a hundred personal computers in one of the experimental networks had crashed mysteriously during the night. Reconstructing the accident, they discovered to their embarrassment that a defective worm had gotten loose in the network, jumping quickly from computer to computer and rendering each inoperable as it went. While a frustrating morning was spent tracking down worm segments in various corners of the large research center, some researchers wondered briefly whether the worm had been able to jump through a special "gateway" to crash other computers at remote locations around the country-but this turned out not to be the case.
    Fortunately, Xerox had already developed a worm killer. Researchers were able to inject a special antibody packet into the network to tell every running worm to stop, no matter what it was doing, and for a time all worm behavior ceased.
    Is it possible that in the future such worms may disable computer networks? Probably not. But some designers have speculated about an even more bizarre worm program known as the "vampire." Such a program would hide itself during the day in an individual computer, then emerge at night to run long, time-consuming computations by taking advantage of idle computer power. The next morning, when users reclaimed their terminals, the vampire would shrink back into its original host to wait for sundown.

JOHN MARKOFF, West Coast editor of Byte

EXPERT SYSTEMS HISTORY
Feigenbaum, Lederberg, Buchanan, Shortliffe
Expert expertise: (left to right) Edward Feigenbaum, Joshua Lederberg,
Bruce Buchanan, Edward Shartliffe.

Edward A. Feigenbaum was one of the people in artificial intelligence research who decided, in the mid-1960s, that it was important to know how much a computer program can know and that the best way to find out would be to try to construct an artificial expert. While looking for an appropriate field of expertise, Feigenbaum encountered Joshua Lederberg, the Nobel laureate biochemist, who suggested that organic chemists sorely needed assistance in determining the molecular structure of chemical compounds.
    Together with Bruce Buchanan, Lederberg and Feigenbaum began work on Dendral, the first expert system, in 1965 at Stanford University. Conventional computer-based systems had failed to provide organic chemists with a tool for forecasting molecular structure. Human chemists know that the possible structure of any chemical compound depends on a number of basic rules about how different atoms can bond to one another. They also know a lot of facts about different atoms in known compounds. When they make or discover a previously unknown compound, they can gather evidence about the compound by analyzing the substance with a mass spectroscope-which provides a lot of data, but no clues to what it all means.
    Building the right kind of "if-then" program, with enough flexibility to use rules of thumb employed by human experts, was only the first major problem to be solved. When you think you've created a program structure capable of manipulating expert knowledge, you have to get some knowledge into the system. After feeding the computer program lots of data, the creators of Dendral interviewed as many expert chemists as they could to find out how they made their decisions. This "knowledge acquisition" phase has problems of its own. When asked how they know what they know, they're unable to articulate the answer. You just have to show them a program that makes decisions and ask them where the program is wrong, and why.
    "Knowledge engineering" is the art, craft and science of observing human experts, building models of their expertise and refining the model until the human experts agree that it works. One of the first spinoffs from Dendral was Meta-Dendral, an expert system for those people whose expertise lies in building expert systems. By separating the inference engine from the body of factual knowledge, Buchanan was able to produce a tool for expert-systems builders.
    In the mid-1970s MYCIN was developed by Edward H. Shortliffe, a physician and computer scientist at Stanford Medical School. The problems associated with diagnosing a certain class of brain infections was an appropriate area for expert system research and an area of particularly pressing human need because the first twenty-four to forty-eight hours are critical if the treatment of these illnesses is to succeed. With all its promise, and all its frightening ethical implications, medicine appears to be one of the most active areas of application for commercial knowledge engineering.
    MYCIN's inference engine, known as E-MYCIN, was used by researchers at Stanford and Pacific Medical Center to produce Puff, an expert system that assists in diagnosing certain lung disorders. An even newer system, Caduceus, now has a knowledge base-larger than any doctor's-of raw data comprising about 80 percent of the world's medical literature.
    Prospector, developed by SRI International, looks at geological data instead of molecules or symptoms. Recently this program accurately predicted the location of a molybdenum deposit that may be worth tens of millions of dollars.
    About two dozen corporations are currently selling expert systems and services. Teknowledge, founded by Feigenbaum and associates in 1981, was the first. IntelliGenetics is perhaps the most exotic, specializing in expert systems for the genetic engineering industry. Start-ups in this field tend toward science-fiction names-Machine Intelligence Corporation, Computer Thought Corp., Symbolics, etc. Other companies already established in non-AI areas have entered the field-among them, Xerox, DEC, IBM, Texas Instruments and Schlumberger.
    Expert systems are now in commercial and research use in a number of fields:
• KAS (Knowledge Acquisition System) and Teiresias help knowledge engineers build expert systems.
• ONCOCIN assists physicians in managing complex drug regimens for treating cancer patients.
• Molgen helps molecular biologists in planning DNA experiments.
• Guidon is an education expert that teaches students by correcting answers to technical questions.
• Genesis assists scientists in planning cloning experiments.
• TATR is used by the Air Force in planning attacks on enemy air bases.

HOWARD RHEINGOLD


Return to Table of Contents | Previous Article | Next Article