8. History — Expert Systems (1969–1986)

Source: AIMA 4th Ed, §1.3.4

The Shift: From General Methods to Domain Knowledge

The lesson from the first AI winter: general search over large spaces doesn’t scale.

The insight: to solve hard problems, you have to almost know the answer already — i.e., you need domain-specific knowledge that allows large reasoning steps and can handle real-world complexity.

This led to expert systems: AI programs that encoded the knowledge of human experts in a narrow domain.

Weak methods (general-purpose search) → Strong methods (domain-specific knowledge + rules)

Key Expert Systems

DENDRAL (1969) — Stanford

• Task: Infer molecular structure from mass spectrometry data.

• Team: Feigenbaum, Buchanan (philosopher), Lederberg (Nobel laureate geneticist).

• Approach: Encode chemist expertise as production rules:

  IF molecule mass M and peaks at x1, x2 such that
     x1 + x2 = M + 28 and x1-28 is high and x2-28 is high
  THEN there is a ketone (C=O) subgroup

• Significance: First successful knowledge-intensive system — expertise came from large numbers of special-purpose rules, not first principles.

• This inspired the Heuristic Programming Project (HPP) at Stanford to investigate how far expert systems could go.

MYCIN (1974) — Stanford

• Task: Diagnose bacterial blood infections and recommend antibiotic treatments.
• Knowledge base: ~450 production rules.
• Performance: Matched some experts; significantly outperformed junior doctors.
• Key innovations:
  1. Rules had to be acquired from interviews with experts (unlike DENDRAL which had theoretical foundations).
  2. Medical knowledge is uncertain — MYCIN introduced certainty factors to handle this (a pre-Bayesian heuristic for uncertainty).
• Limitation: Certainty factors were ad-hoc, not grounded in probability theory. But they worked well enough in practice.

R1 / XCON (1980) — Digital Equipment Corporation

• Task: Configure orders for new DEC computer systems.
• Deployed in 1980; by 1986 saving DEC ~$40 million/year.
• Impact: The first large-scale commercial success of AI — proved that expert systems had real economic value.
• By 1988: DEC had 40 deployed expert systems; DuPont had 100 in use, 500 in development.

The Expert Systems Boom (1980–1987)

• AI industry grew from a few million dollars (1980) to billions of dollars (1988).
• Hundreds of companies built expert systems, vision systems, robots, and specialized hardware.
• Japan’s “Fifth Generation” project (1981): $1.3 billion (today’s terms) to build massively parallel intelligent computers running Prolog.
• U.S. responded with the Microelectronics and Computer Technology Corporation (MCC).

The Second AI Winter (1987–1993)

The boom collapsed because:

1. Expert systems were brittle. They broke down when encountering situations outside their narrow training. They couldn’t handle uncertainty well, couldn’t learn from experience, and couldn’t handle common-sense knowledge.
2. Maintenance nightmare. As domains grew, keeping the rule base consistent and up to date became impossibly expensive.
3. Hardware disappointment. Specialized LISP machines became obsolete when standard workstations became more powerful and cheaper.
4. Unmet promises. The Fifth Generation project never delivered.

Pattern repeats: hype → investment → disappointment → funding cuts = second AI winter.

Legacy

Expert systems taught AI an important lesson: knowledge matters. You cannot have intelligence without knowledge of the domain. This lesson survived the winter and drives modern AI:

• Knowledge representation remains an active field.
• Modern LLMs can be seen as a new form of knowledge encoding — but learned from data rather than hand-coded by experts.
• The fundamental tension between symbolic (logic-based) and subsymbolic (learned) AI that emerged in this era continues today.