Textbook of AI

MYCIN was an expert system developed by Edward Shortliffe at Stanford from 1972 onwards as the basis of his 1974 PhD thesis, in collaboration with Bruce Buchanan, Stanley Cohen and the Heuristic Programming Project. The system diagnosed bacterial infections of the blood (bacteraemia) and meningitis and recommended appropriate antibiotic therapy. It is one of the most-studied expert systems in the history of AI and the system that, more than any other, defined what the second wave of symbolic AI looked like.

Architecture

MYCIN encoded approximately 600 rules elicited from infectious-disease specialists at Stanford Medical Center, of the canonical IF–THEN form:

IF the infection is primary bacteraemia, AND the site of the culture is one of the sterile sites, AND the suspected portal of entry is the gastrointestinal tract, THEN there is suggestive evidence (0.7) that the identity of the organism is Bacteroides.

The inference engine used backward chaining: starting from a goal (e.g. "what is the identity of the organism?"), it worked backwards through rules, asking the physician for data only when the system itself could not derive an answer. A simple certainty factor calculus, devised by Shortliffe and Buchanan, propagated degrees of belief through chains of rules, values between $-1$ and $+1$ representing strength of evidence for or against a hypothesis.

MYCIN included an explanation facility that could answer questions of the form "why are you asking this?" and "how did you conclude that?" by reconstructing the chain of rule firings. This was a substantial advance over earlier black-box systems and an important precursor to modern interest in explainable AI.

Performance

MYCIN's clinical performance was strong. In the formal Yu et al. (1979) evaluation, MYCIN's recommendations for ten meningitis cases were judged blind by a panel of eight infectious-disease experts; MYCIN's prescriptions were rated acceptable in 65% of cases, slightly above the average for the human experts (42.5–62.5%) and substantially above general-practice physicians.

Why it was never deployed

Despite its performance, MYCIN was never used clinically. The reasons were partly technical (it required typing diagnostic data into a terminal in the pre-network era), partly economic (no obvious deployment path for a stand-alone PDP-10 program), and partly legal, there was genuine concern in the 1970s about the liability implications of a computer system prescribing drugs.

Legacy

MYCIN's architecture became the template for an entire generation of expert systems. Decoupling the inference engine from the rule base produced EMYCIN (Empty MYCIN, 1981), the first general-purpose expert-system shell, which spawned commercial products such as KEE and ART. The MYCIN rule format and certainty-factor calculus were copied widely throughout the 1980s expert-systems boom.

The certainty factors themselves were later (Heckerman, 1986) shown to be a particular and somewhat ill-founded special case of probabilistic reasoning, foreshadowing the rise of Bayesian networks under Judea Pearl in the late 1980s and the field's gradual shift from symbolic rule systems to probabilistic graphical models.