Expert Systems

An expert system is an artificial intelligence program built to emulate the decision-making of a human specialist within a narrow, well-defined domain —diagnosing an infection, interpreting a chemical spectrum or approving a loan—. Rather than learning from examples, it reasons over a body of knowledge handed to it by flesh-and-blood experts and encoded as explicit rules.

Expert systems are the most visible face of so-called symbolic AI, the tradition that for decades treated intelligence as the manipulation of symbols and logical rules. Its promise was crisp: if you can capture in writing what a professional knows, a machine can reproduce that judgement consistently, tirelessly and in a way you can audit.

Architecture: knowledge base and inference engine

Every classic expert system deliberately separates two parts. The knowledge base stores the facts of the domain and, above all, IF-THEN rules («if the fever is high and the culture is gram-negative, then suspect a specific bacterium»), elicited through long interviews with specialists conducted by the so-called knowledge engineer. The inference engine is the mechanism that applies those rules to derive new conclusions. Primary source.

The engine can reason in two directions. Forward chaining starts from the available facts and fires rules until it reaches a conclusion; it is the data-driven approach. Backward chaining starts from a hypothesis —the goal— and works backwards to the evidence that would confirm it, much as a diagnosis does. This separation of knowledge from control is the hallmark of the paradigm. Primary source.

DENDRAL, MYCIN and the golden age

DENDRAL and MYCIN made Stanford one of the key centers of expert-system research from the mid-1960s through the 1970s. DENDRAL began at Stanford in 1965 through Edward Feigenbaum's collaboration with Joshua Lederberg; Bruce Buchanan joined that same year. Feigenbaum's account and archive describe a system that proposed molecular structures and compared their predicted mass spectra with experimental data: domain chemistry embedded in the inference process. MYCIN was Edward Shortliffe's 1972–1979 project, with Bruce Buchanan as his principal mentor and collaborator. Its knowledge base grew to about five hundred rules, applied through goal-directed backward chaining, to propose diagnoses and antimicrobial treatments. In a blinded evaluation of ten meningitis cases, outside experts rated 65% of its recommendations acceptable, versus an average of 55.5% for five Stanford specialists; the authors cautioned that the sample was small. Neither was ever used routinely at the bedside, but they proved the approach was real.

Limits and why machine learning displaced them

Their decline had honest causes. The first was the knowledge acquisition bottleneck, a phrase Feigenbaum himself made famous: extracting an expert's know-how by hand and translating it into rules is slow, costly and artisanal. The second was brittleness: outside the narrow domain they were programmed for, these systems fail without grace, because they have no common sense and no way to say «this is beyond me». The third, and decisive, is that they do not learn from data; every new rule must be written by hand. Primary source. Primary source. Primary source.

That is the contrast with today's machine learning, which flips the premise: instead of being handed the rules, it infers them implicitly from large volumes of examples. This ability to learn on its own is why machine learning displaced expert systems as the frontier of AI. Even so, the paradigm never vanished: it lives on, in good health, inside the business rule engines that govern approvals, pricing and validations across banking, insurance and logistics. Primary source. Primary source. Primary source.