Floor 7 — Expert Systems: The Civilisation Built Out of If-Else

The floor

From roughly 1970 to the late 1980s, the ambitious research programme in AI was the expert system. Take the knowledge of a specialist — a doctor, a chemist, a geologist — and encode it as if-then rules. Feed the rules into a machine with an inference engine. The machine answers questions in the domain.

MYCIN diagnosed bacterial infections. DENDRAL inferred molecular structures. PROSPECTOR found mineral deposits. XCON configured VAX computers for DEC and saved the company an audited forty million dollars a year.

It genuinely worked, for a while, within a narrow room.

In 1982, Japan launched the Fifth Generation Computer Systems project. Four hundred million dollars. Ten years. A national mission to build computers whose native language was Prolog — pure rule-based logic. The Americans, panicked, founded the Microelectronics and Computer Technology Corporation to compete. The British responded with the Alvey Programme. Expert systems were the future.

What was picked

Intelligence as a pile of rules, assembled by hand. If X and Y and not Z, then conclude A. Ten thousand of these per domain, tuned by "knowledge engineers" who spent months interviewing specialists. The system was deterministic, inspectable, auditable. You could point at a rule and say "this is why it said that."

The assumption beneath the whole programme was that expertise decomposes — that what a physician knows can be written down as a finite rule set, if you only interview them long enough. The assumption beneath that assumption was that decision-making is fundamentally a classification problem. You present a case, the system selects a category, you read off the advice for that category.

Floor 1, still. All the way up.

What could have been picked

Cyc, started by Doug Lenat in 1984, is the reductio ad absurdum. Lenat believed — and believes — that intelligence is common sense, and common sense is several million rules, each hand-written by a knowledge engineer, covering every quiet assumption about how the world works. A person is usually alive. Water is usually wet. Things fall down unless held up. He has been adding to Cyc for forty years.

It is one of the most remarkable and heroic research programmes in computer science, and it is also a confession. If you need several million rules to express common sense, the thing you are modelling is not rules.

Meanwhile, in the same decade, Hopfield networks were running. Boltzmann machines were running. Backpropagation was getting rediscovered. Small groups — Sejnowski, McClelland, Rumelhart, Hinton — were building systems that learned their own representations from data. Nobody was funding them seriously. The journals of the day politely declined their papers on the grounds that neural networks had already been shown not to work.

(See Floor 6.)

What we missed

A twenty-year investment in continuous, learned, context-sensitive reasoning, made during the period when compute was growing tenfold every few years. We chose instead to invest in hand-crafted, brittle, context-blind rule systems, which collapsed the moment they encountered a situation their knowledge engineer had not anticipated.

The Fifth Generation Project ended quietly in 1994. The Japanese did not get intelligent machines. Expert systems did not spread into every workplace. XCON became a maintenance nightmare at DEC because every new VAX configuration required a new rule and the rules had started contradicting each other. The knowledge engineers burned out.

In the alternate timeline, the four hundred million dollars — or even a tenth of it — went into continuous systems that learned their own representations. By the time the field came out of the second winter, it would not have been starting from zero. It would have been a generation deep.

What the next floor will ask

In 1986, the door opened again. But the thing that walked back through it was carrying a very specific kind of objective function in its hands.

That's Floor 8.