expert system
/ EK-spurt SIS-tum /
An expert system is a program that captures the know-how of a human specialist in a narrow field and then answers questions or gives advice the way that expert would. Imagine bottling a seasoned doctor's diagnostic reasoning, or a master mechanic's troubleshooting instincts, into software that a non-expert can consult. It was the flagship commercial success of symbolic AI in the 1970s and 1980s.
Under the hood, an expert system has two main parts: a knowledge base — a large store of IF–THEN rules and facts elicited from human experts — and an inference engine, which chains those rules together to reach conclusions. Ask it about a sick plant, and it works through rules ('IF leaves are yellow AND soil is wet, THEN suspect root rot') much as the expert would, often explaining each step so you can see its reasoning. Famous examples include MYCIN, which diagnosed bacterial infections, and DENDRAL, which identified chemical structures.
Expert systems shine in stable, well-bounded domains where expertise can be written as rules and the facts don't change much. But they are costly and slow to build — extracting and encoding an expert's knowledge by hand is painstaking — and they are brittle: confront one with a case just outside its rules and it has no common sense to fall back on. They also can't learn from new data on their own. When their limits and high maintenance costs became clear, the boom cooled into one of the AI winters. The idea lives on, though, in modern rule engines and decision-support tools.
MYCIN (Stanford, early 1970s) advised on antibiotic treatment for blood infections. It asked the physician questions, chained through hundreds of expert-supplied rules, and recommended a drug and dose — even rating its own confidence. In tests it sometimes matched specialists, yet it was never deployed clinically, partly over liability and the effort of keeping its rules current.
MYCIN: an early expert system that reasoned like a specialist — and showed both the promise and the brittleness of rule-based AI.
An expert system is only as good and as current as the rules humans put into it; it cannot learn from experience and has no common sense beyond its knowledge base. The 1980s expert-system boom and bust is a textbook example of how an AI winter sets in.