An expert system is a program that tries to capture a specialist's know-how in a collection of explicit "if… then…" rules (production rules) plus an engine that chains them together. MYCIN, developed at Stanford in the first half of the 1970s, is the laboratory specimen: it advised on which antibiotics to give patients with serious bacterial infections (first of the blood, later meningitis). According to M. Mitchell (2019) it had about six hundred rules (its authors' final assessment speaks of 450, plus about a thousand facts in tables), and it was "a paradigmatic example" of symbolic AI (2019, ch. 2). Its name is the suffix of many antibiotics.
How it works
It has three parts: a knowledge base, an inference engine, and an explanation module (Shortliffe, 1976). A typical rule, translated from its internal Lisp representation, reads: if the identity of the organism is not known with certainty, the stain is gram-negative, the morphology is rod and the organism is aerobic, then there is strongly suggestive evidence (0.8) that it belongs to the enterobacteriaceae (Buchanan & Shortliffe, 1984). That 0.8 is a certainty factor: a number between −1 and +1 that subtracts degrees of disbelief from degrees of belief, and that its authors presented as a computational device, not as a probability.
MYCIN reasons backwards. It starts from the goal (which organism must be covered, which therapy is advisable), looks for rules whose conclusion would establish it, and the premises of those rules become new subgoals; when no rule can establish a datum, it asks the physician. The result is a dialogue. At any point the user can type WHY ("why are you asking this?") or HOW ("how did you conclude that?"), and the system displays the chain of rules: the explanation is the very trace of the inference. Choosing Lisp was not a technical detail, since it allowed rules to be treated as data the program could read, display and edit. Emptied of medicine, the engine became EMYCIN, a shell for building expert systems in other domains.
What changed
Logic Theorist searched for proofs using general heuristics; expert systems bet on the opposite: little general method and a great deal of specific knowledge. Their direct ancestor was DENDRAL, in organic chemistry, and the craft of translating an expert's know-how into rules was called knowledge engineering. Buckner (2024) places this approach at the point of the pendulum where AI hand-programmed as much structure as possible, what Haugeland (1985) dubbed GOFAI (2024, p. 26).
And it worked. In a blinded evaluation published in 1979, eight specialists rated prescriptions for ten meningitis cases without knowing who had written them: MYCIN's were judged acceptable in 65% of the ratings, against 42.5 to 62.5% for those of five Stanford faculty members (Yu et al., 1979). Even so, MYCIN was never tried on the hospital wards, and its infectious-disease knowledge base was shelved in 1978 (Buchanan & Shortliffe, 1984). Its commercial successors did prosper: R1/XCON configured VAX computer orders at Digital Equipment (McDermott, 1982), and by around 1985 companies were spending over a billion dollars a year on expert systems (Chollet, 2021, sec. 1.1.7).
Limits
By the early 1990s these systems had proven expensive to maintain, hard to scale and limited in scope, and the second "AI winter" set in (Chollet, 2021, sec. 1.1.7).
- The bottleneck. MYCIN's authors concluded that knowledge acquisition was "the critical bottleneck in building expert systems" (Buchanan & Shortliffe, 1984). Witten et al. (2017) show the alternative: in diagnosing soybean diseases, rules induced from about 300 examples got 97.5% of the test cases right, against 72% for the rules elicited from the expert himself (2017, pp. 19–21). That is the niche the decision tree will fill.
- Brittleness. By the mid-1980s it was clear that these systems failed in new situations, because the experts relied on common sense they never put into the rules (M. Mitchell, 2019, ch. 2).
- Apparent precision. Certainty factors worked well in practice, but what exactly a 0.8 means is far less clear than the decimal suggests.