Bayesian network

A Bayesian network is a directed graphical model that represents a set of variables and their conditional dependencies as a graph, and updates beliefs by propagating probabilities across it. Judea Pearl proposed the belief-updating algorithm in a 1982 paper ("Reverend Bayes on Inference Engines") and named the approach "Bayesian networks" by 1985, explicitly to give uncertain reasoning "coherence with orthodox probability theory and Bayesian reasoning" — the property he argued the ad hoc combination rules of 1970s expert systems lacked.

In this vault the concept is load-bearing as the destination of a recurring historical pattern: an AI heuristic method, built on hand-tuned uncertainty arithmetic, is later rebuilt on Bayesian-network statistical inference. It is the constructive answer to the certainty-factor scheme of MYCIN — the scheme Heckerman proved incoherent and Pearl replaced — and it recurs, independently, as the formalism two unrelated communities reached for when they finally made their heuristics rigorous: medical diagnosis (the Pathfinder project) and intelligence analysis (Heuer's Analysis of Competing Hypotheses). The pattern is why the vault treats the two senses of "inference" (the classic expert-system engine vs. statistical inference over a distribution) as a genuine connection here rather than a shared word — the bridge is documented, not cosine-deep.

References

• claim-pearl-built-bayesian-networks-as-coherent-alternative-to-certainty-factors — Pearl's own account: built for probabilistic coherence, motivated by the certainty-factor scheme's failure on "explaining away."
• claim-pathfinder-rebuilt-from-rules-to-bayesian-networks — the empirical case in medical diagnosis: a lymph-node pathology system carried from rule-chaining to probabilistic scoring, the expert's blind verdict of improvement landing at the simple-Bayes step.
• claim-valtorta-reformulated-ach-as-bayesian-networks — the second, independent domain: an intelligence-funded CS group recast ACH as Bayesian networks, not as rule-chaining.
• claim-heckerman-1986-certainty-factors-require-independence-assumptions — the critique that motivates the replacement: certainty factors are coherent only under hidden independence assumptions.
• claim-mycin-certainty-factors-non-bayesian-uncertainty — the predecessor scheme the concept is defined against.