Chapter 1 · Lesson 1.2
Expert systems
First major stage: a system able to reason from facts and logic.
Architecture
Domain expert
Holds the business knowledge.
Knowledge engineer
Extracts the expertise and translates it into rules.
Knowledge base
Base of faits (raw knowledge) + base of rules (“If condition Then action”).
Inference engine
Chooses which rules to apply to the facts to solve the problem.
User interface
Handles communication between user ↔ system.
Example of a knowledge base (company diagnosis)
Facts: Legal status = SARL, Revenue = €100 M.
Rule: If Revenue < €10 M AND Subcontractor = qualified, then risk = low.
The inference engine
It goes through the tree of rules until a stopping rule is reached. If several rules are applicable, a selection process arbitrates. Applying a rule makes it possible to deduce new facts by modus ponens : “A is true” and “if A then B” ⇒ “B is true”.
| Mode | Principle |
|---|---|
| Forward chaining | Starts from known facts, applies the rules, reaches a conclusion. |
| Backward chaining | Starts from a hypothesis, checks the conditions, validates or invalidates the hypothesis. |
| Mixed chaining | A mix of both. |
Your turn
Rules: r1 flies ∧ lays eggs ⇒ bird · r2 has feathers ⇒ bird · r3 bird ∧ long neck ∧ long legs ⇒ ostrich.
Famous expert systems
| System | Scope | Distinguishing feature |
|---|---|---|
| DENDRAL (1965) | Analysis of mass spectra, molecular structures | First successful expert system |
| MYCIN (1972) | Diagnosis of bacterial infections, antibiotics | Accuracy ≈ 65% |
| XCON/R1 (1980) | Configuration of VAX computers (Digital Equipment) | ≈ USD 40 M/year in savings |
Building the rule base: it requires a lot of expert work and does not update itself.
An expert system applies rules written by experts (fact base + rule base + inference engine). It is transparent and explainable, but costly to build and maintain: this is what machine learning replaces by learning from data.