Concept · Chapter 1: What Is Artificial Intelligence?
Expert Systems
Expert systems captured a human specialist's knowledge as hundreds or thousands of if–then rules, and were AI's first big commercial success.
The problem
General problem-solvers were weak on real tasks; expertise seemed to come from detailed domain knowledge, not clever general reasoning.
The solution
Interview experts, encode their know-how as rules in a narrow domain, and pair the rules with an inference engine that can explain its conclusions.
The consequence
Systems like MYCIN and XCON worked in their niches, sparking an industry boom in the 1980s — and then revealed how costly it is to acquire and maintain knowledge by hand.
You should understand first
- The Turing Test
- Symbolic AI
- Logic and Rules
- Expert Systems
Two famous examples
MYCIN (Stanford, early 1970s) diagnosed bacterial infections and recommended antibiotics, asking doctors a series of yes/no questions and reasoning backward from possible diagnoses. In a 1979 evaluation, its recommendations were rated acceptable about as often as those of infectious-disease faculty Established. It was never used in practice — it ran standalone on a research computer with manual data entry, and questions of responsibility were unresolved.
XCON (Digital Equipment Corporation, in use from 1980) configured customers' computer orders. It grew to about 2,500 rules and was estimated to save DEC $25 million a year — proof that AI could pay.
The boom — and the cost
Success brought an industry: expert-system shells, specialized LISP machines, corporate AI groups. But every system needed experts' time to extract rules, engineers to encode them, and constant maintenance as the domain changed. Rules interacted in unexpected ways, and systems failed abruptly on cases just outside their scope. When the specialized hardware market collapsed in 1987, a second AI winter followed.
What to remember
- Knowledge from human experts, encoded as rules, in a narrow domain.
- MYCIN (Stanford, 1970s): infection diagnosis — rated as good as specialists, never used clinically.
- XCON (DEC, from 1980): configured orders with about 2,500 rules, reportedly saving about $25M a year.
- Hard to build, hard to maintain, brittle outside their domain.