MYCIN's backward chaining is a goal-directed, depth-first recursive search that resolves one target at a time by decomposing rule premises into subgoals

AI-historyexpert-systemsbackward-chainingMYCINinference-enginemechanism

MYCIN — the paradigm 1970s backward-chaining expert system, and the concrete example the vault's inference-engine note cites for "backward chaining" — runs what its own technical team calls a "goal-directed backward chaining of rules." The engine picks a single clinical parameter as a goal, retrieves the pre-indexed rules whose conclusions bear on that goal, and tries each rule's premise. Where a premise clause needs a fact not yet known, the system opens a subgoal for that fact and resolves it the same way — recursively — until the chain bottoms out in a known fact, a question to the user, or a rule that fails. The result, in van Melle's words, is a "highly focused search through the rule base": sequential, depth-first, and single-target, always working on one goal's proof tree at a time. "Questions are asked during the consultation when rules fail to deduce the necessary information" — the user is consulted only where the rule base cannot close a subgoal on its own.

One scope note on the vocabulary, because the two words do different amounts of work: "goal-directed," "subgoal," "subgoal trees" and "highly focused search" are van Melle's own terms, while "depth-first" is this vault's characterisation, not a phrase in the chapter. It is a fair reading of the descent he describes — a premise clause that needs an unknown fact opens a subgoal before the premise completes — but the same chapter adds three qualifications a flat "depth-first" would hide: a preview mechanism that reads rules before invoking them and discards those already provably false (so the system "avoids the useless search of one or more subgoal trees"); a unity path that MYCIN "always seeks… before trying a set of rules or asking a question," so a value derivable with certainty short-circuits the descent; and meta-rules that can "reorder or prune the rule list" ahead of any goal. The control structure is goal-directed and recursive as claimed; it is not an unguided depth-first walk.

This is the Tier-1 primary account of a mechanism the vault had previously held only at a Tier-4 definitional level (claim-inference-classic-ai-engines quotes Wikipedia: backward chaining "starts with goals, and works backward"). The source is Chapter 4 of Rule-Based Expert Systems: The MYCIN Experiments (Addison-Wesley, 1984), authored by William van Melle — who built EMYCIN, the domain-independent shell extracted from MYCIN — and edited by MYCIN's creators Bruce Buchanan and Edward Shortliffe. Per the chapter's own opening footnote, its text first appeared as van Melle's "MYCIN: A knowledge-based consultation program for infectious disease diagnosis," International Journal of Man-Machine Studies 10:313–322 (1978) — so the primary description of this control structure is contemporaneous with MYCIN's own working life, not a retrospective written for the 1984 collection. See entity-mycin for the system hub, and claim-mycin-certainty-factors-non-bayesian-uncertainty for the uncertainty calculus layered on top of this control structure.

The sequential, goal-directed shape of this search is what makes it a mechanism-level false friend for Heuer's ACH, which eliminates hypotheses by a flat matrix scan rather than a recursive proof — see observation-ach-backward-chaining-mechanism-comparison-is-false-friend.