Textbook of AI

An expert system is a knowledge-based AI system that solves problems within a narrow domain by applying rules elicited from human experts. The classical architecture has three components: a knowledge base of domain rules and facts, an inference engine that applies the rules (typically by forward chaining from data to conclusions, or backward chaining from goals to required premises), and a user interface for entering data and explaining conclusions.

The first major expert systems, DENDRAL (Feigenbaum, Buchanan, Lederberg, from 1965) for inferring molecular structure from mass spectra, and MYCIN (Shortliffe, Buchanan, Feigenbaum, from 1972) for diagnosing bacterial infections, established the methodology and demonstrated that, in narrow domains, rule-based systems could match human expert performance. The 1980s saw an enormous commercial expert-systems boom: XCON (configuring DEC VAX computers, hundreds of millions of dollars in savings), MYCIN-derived systems in medicine, and a flotilla of expert-system shells (KEE, ART, EMYCIN, OPS5).

The boom collapsed around 1987–1992. Expert systems proved expensive to maintain (rule bases needed continuous updating), brittle outside their narrow domains, and unable to integrate with the growing volumes of data that statistical machine learning could exploit. The collapse was the central episode of the second AI winter.

Expert-system techniques have not vanished. Modern business rules engines (Drools, IBM ODM), decision tables, and medical decision-support systems descend from the same line, and contemporary neuro-symbolic approaches that combine LLM reasoning with explicit rule bases are a current research front.