Edward Feigenbaum

Who he is

Edward Albert Feigenbaum was born on January 20, 1936, in Weehawken, New Jersey, and is known as "the father of expert systems".^[1] The phrase describes a research direction rather than a single program: instead of building a system that tries to reason about everything, he built systems that carried the concentrated knowledge of a specialist — the working rules a chemist or a physician actually uses — and applied it to a narrow class of problems.

Education and early research

He received a B.S. in Electrical Engineering from Carnegie Institute of Technology in 1956 and a Ph.D. from its Graduate School of Industrial Administration in September 1960.^[1] His dissertation work was supervised by Herbert Simon and produced EPAM (Elementary Perceiver and Memorizer), a program that modelled how a person takes in and remembers items it is shown.^[1] The method behind that kind of work is to compare a program's behaviour with human behaviour observed in experiments, and to treat the places where they diverge as information about the mind.

Berkeley and Stanford

From 1960 to 1965 he held research and faculty appointments at the University of California, Berkeley, including Assistant Professor at the School of Business Administration from 1960 to 1963.^[1] He then moved to Stanford University, where he was Associate Professor of Computer Science from 1965 to 1968, Professor of Computer Science from 1969 to 1995, and Kumagai Professor of Computer Science from 1995 to 2000.^[1] Administrative work ran alongside the teaching and research: he directed the Stanford Computation Center from 1965 to 1968, chaired the Computer Science Department from 1976 to 1981, and was Director and then Co-Director of the Knowledge Systems Laboratory from 1965 to 2000.^[1] The laboratory's name is a fair summary of the programme he pursued there — the study of systems whose main asset is stored knowledge.

Reading molecules by machine

His Heuristic DENDRAL program inferred the structure of complex compounds from chemical formulas and mass spectral data, and META-DENDRAL derived new rules to improve it.^[1] Mass spectra record the fragments a molecule breaks into and their weights; getting back from those fragments to a whole structure means choosing among a very large number of possible arrangements. What made the program work was not a general-purpose reasoning trick but the chemistry it had been given, and META-DENDRAL addressed the obvious follow-up question: where the rules come from. A program that can propose new rules of its own shifts part of the burden of building a knowledge base from the person who writes the rules to the machine that tests them.

Medical systems and the idea of a shell

Medical expert systems from his projects include MYCIN and ONCOCIN, and EMYCIN was the first portable general expert system "shell".^[1] A shell is what is left when the medical knowledge is taken out of a consultation program: an inference mechanism plus a way of writing rules, with no domain built in. Another field could then encode its own knowledge on top of the same machinery instead of rebuilding the engine, which is what makes a program reusable rather than a one-off. The term "shell" is still used in this sense in software that separates a general engine from the rules supplied for a particular job.

Naming the field

He pioneered and used the phrases "Knowledge is Power", "Expert Systems", and "Knowledge Engineering".^[1] Naming is not decoration in a young research area. Once "knowledge engineering" existed as a phrase, the labour of extracting rules from experts and organising them for a machine could be described, taught and argued about as a distinct kind of work rather than as an unexplained preliminary to programming.

A wider record

He co-edited the book "Computers and Thought" with Julian Feldman, published in 1963.^[1] Public service included a term as Chief Scientist of the United States Air Force from 1994 to 1997.^[1] He was elected to the National Academy of Engineering in 1986, served as past president of the American Association for Artificial Intelligence, and received the ACM Turing Award jointly with Raj Reddy in 1994.^[1]

What the approach left behind

The systems Feigenbaum built and named put knowledge, rather than a universal reasoning method, at the centre of artificial intelligence. That choice also defines the limits of the style: such a system is only as good as the rules it has been given, collecting those rules from busy specialists is slow, and a domain that shifts under the rules forces them to be revised. The shell idea and the practice of knowledge engineering were answers to those costs, and they are the part of the work that later software carried forward, even where the machinery underneath changed.