Expert systems

1 · What it is

An expert system is a program built for problems that normally take a trained specialist, and it tackles them by applying stored knowledge through a reasoning procedure. It has two main parts, a knowledge base and an inference engine. The knowledge base is where the program keeps what it knows about its subject, such as medicine, mostly written as rules of the form if A and B, then C. The inference engine can chain rules together, so one rule’s conclusion feeds the next. Chaining from known data forward is called forward chaining. MYCIN mostly chained backward instead. It starts from a goal and works back through the rules to find the data that would prove it.

The person who helps an expert write down their knowledge for the program is called a knowledge engineer. Step one was talking with the expert; step two was drafting rules from those talks. Next, the draft rules were tried on sample cases; when the program got one wrong, the engineer traced the rules behind the mistake and asked the expert what to change.

Three systems made the idea famous. DENDRAL, begun at Stanford in 1965, became the early model for the whole approach. DENDRAL was built by Joshua Lederberg, Edward Feigenbaum and Carl Djerassi, and identified the structure of unknown organic compounds. MYCIN advised doctors on antibiotics and computed certainty factors for its conclusions. What it knew about infections fitted into 450 rules plus roughly 1,000 other facts. MYCIN was never tried on the hospital wards. R1, also called XCON, was used at DEC from January 1980. Given a customer’s order, R1 worked out what had to be added or swapped to make it complete and consistent. By November 1983 R1 had about 3,300 rules and 5,500 component descriptions. By 1984 its builders no longer expected it ever to be finished. According to DEC, XCON together with a sister system was worth over 40 million dollars in savings each year.

Expert systems also drew a rush of private money into AI. That boom did not last, since few buyers managed to train their own staff to build and run such systems well. In the mid-1980s enthusiasm cooled and money for AI research got scarce, a slump now known as the AI winter. A Stanford history of AI adds that the field leaned too hard on true-or-false logic and neglected uncertainty. None of these expert systems took over from the chemist or the physician; the most they did was offer a second opinion. The idea lives on in rule engines such as Drools, which can chain rules both forward and backward.