David Heckerman

American computer scientist known for foundational work on Bayesian networks and probabilistic reasoning in AI, long associated with Microsoft Research before moving into computational biology and machine learning roles at Insitro and Amazon.

He is the pivot figure of this vault's rule-based-AI-to-Bayesian-networks thread. His 1986 paper on MYCIN's certainty factors is the formal proof that the scheme's combination rules admit a coherent probabilistic reading only under restrictive conditional-independence assumptions and otherwise violate commutativity — the mathematical argument the field's shift to Bayesian networks was built to answer. A decade later he co-authored, with certainty factors' own creator Edward Shortliffe, the paper documenting the field's abandonment of the scheme; separately, with Eric Horvitz and Bharat Nathwani, he rebuilt the Pathfinder diagnostic system from rule-chaining to Bayesian-network inference.

References

• claim-heckerman-1986-certainty-factors-require-independence-assumptions — his formal critique: certainty factors admit a coherent probabilistic reading only under restrictive independence assumptions, and violate commutativity
• claim-pearl-built-bayesian-networks-as-coherent-alternative-to-certainty-factors — Pearl credits Heckerman's "especially effective critique" of certainty factors
• claim-pathfinder-rebuilt-from-rules-to-bayesian-networks — Heckerman, Horvitz, and Nathwani's own account of rebuilding the Pathfinder expert system from rule-chaining to Bayesian-network inference