#1220 1980 · SRI International (Richard Duda, Peter Hart) · Mining / mineral exploration
An expert system re-read a passed-over survey and found a $100 million ore deposit
the problem
Only a few senior geologists could judge whether a site was worth drilling — too few to recheck deprioritized sites
background
Evaluating whether a site was worth drilling required a scarce kind of judgment: weighing rock types, mineral signals and structural clues against each other the way a senior porphyry-deposit specialist would. Busy specialists necessarily judged quickly and moved to the next site, and once a site had been looked at and passed over, there was rarely time or headcount to have anyone look again.
SRI International researchers Richard Duda and Peter Hart set out to test whether that judgment was really inexpressible field intuition or a set of describable rules, by interviewing a porphyry-molybdenum exploration specialist about exactly how he weighed evidence, and encoding those rules with Bayesian reasoning into a program called PROSPECTOR.
what everyone would do
The standard answer to a shortage of senior exploration geologists was to hire and train more of them, or to reprioritize which sites got a scarce expert's personal attention and leave lower-priority sites unexamined.
what they saw
Duda and Hart saw a geologist's judgment as weighing describable evidence, not field intuition, and used Bayesian rules so PROSPECTOR could reconsider Mount Tolman more exhaustively than a specialist who looked once.
the move
Fed nothing but published geological survey data on Mount Tolman, Washington, a site human exploration teams had already surveyed and passed over, PROSPECTOR re-weighed the same evidence using the specialist's own encoded rules and flagged a specific zone as a promising, previously unrecognized molybdenum target.
why it works
A geologist's site judgment is really weighing dozens of rock, mineral and structural signals against each other, and busy human specialists necessarily judge quickly before moving to the next site. PROSPECTOR held the same expert's weighting rules but never got tired or rushed, so it could reconsider the full weight of Mount Tolman's existing survey data more exhaustively than the humans who had already looked at it and passed. Because the answer was sitting in data already collected, the discovery cost nothing but computer time, no new drilling or fieldwork, until PROSPECTOR's flag justified it.
the payoff
Drilling confirmed PROSPECTOR's target held molybdenum worth an estimated $100 million — a discovery credited to the program.
where it breaks
It only works when the expert's reasoning really is reducible to explicit rules over available data; PROSPECTOR could reconsider evidence already gathered, but could not go collect evidence a human hadn't thought to record, and it depended entirely on the specialist's rules being sound for that class of deposit. Applied outside the geological setting it was built from, it would fail silently rather than flag its own limits.
what came after
PROSPECTOR became one of the most cited early proofs that expert-system technology could outperform, not just imitate, trained specialists, credited with legitimizing serious industrial investment in knowledge-based systems through the 1980s.
references
- [1]Recognition of a Hidden Mineral Deposit by an Artificial Intelligence ProgramScience (AAAS), 1982science.org