In plain words
In the 1970s and 1980s, AI researchers tried a practical idea: interview experts, write down how they decide, and put it in a computer.
A doctor might say: if the infection is in the blood and the bacteria are rod-shaped, it is probably this type. Each such judgement becomes an if-then rule. Collect hundreds of them and you have a knowledge base.
A second part, the inference engine, asks the user questions, checks which rules apply and chains them together towards a conclusion. It can even explain its advice by listing the rules it used.
Expert systems were one of AI's first commercial successes, then hit limits. Rules were costly to write and update, and the systems broke on cases nobody had foreseen. They are a classic form of symbolic AI.
Everyday example
The scooter mechanic's checklist
Your scooter will not start. The mechanic runs through it in his head: fuel in the tank? Yes. Spark at the plug? No. Then it is the plug or the wiring, so change the plug first. Write every such question and rule into a program and anyone can get his advice when he is not around. That program is an expert system.
Where the example breaks: The mechanic can spot something odd he has never seen before and use common sense. An expert system only knows its written rules and cannot tell when a problem falls outside them.
How it actually works
A classic expert system has three parts: a knowledge base, an inference engine and a way to talk to the user.
Knowledge base
Facts and if-then rules gathered by interviewing specialists, a slow job called knowledge engineering. MYCIN, a Stanford system for blood infections, had about 600 rules.
Inference engine
The engine matches rules against known facts. Forward chaining starts from facts and fires rules to reach conclusions. Backward chaining starts from a possible conclusion and asks questions to confirm it.
Handling uncertainty
Experts are rarely certain, so systems attached confidence numbers to rules. MYCIN used certainty factors running from minus 1 to 1.
Explanation
Because every step is a written rule, the system can answer "why?" by showing the chain. That transparency is something modern neural networks struggle with, as explainability research shows.
- First
- DENDRAL, 1965, by Edward Feigenbaum and Joshua Lederberg at Stanford, for identifying chemical compounds
- Famous
- MYCIN, Stanford, 1970s: suggested causes of and antibiotics for blood infections
- How good
- Experts rated MYCIN's advice acceptable about 75% of the time, comparable to Stanford specialists
- In practice
- MYCIN was never used routinely on patients, partly for ethical and legal reasons
Questions people ask
What is an example of an expert system?
MYCIN, built at Stanford in the 1970s, suggested which bacteria were causing a blood infection and which antibiotics to give. DENDRAL, begun in 1965, helped chemists work out the structure of molecules. Both were built from rules gathered from human experts.
What is the difference between an expert system and machine learning?
An expert system's rules are written by people from interviews with experts. A machine learning model learns its own patterns from data. Expert systems are easy to explain but hard to scale; learned models scale but are harder to explain.
Are expert systems still used?
The name has faded but the approach has not. Business rules engines, built on the same if-then idea, still decide eligibility, pricing and compliance checks in banking, insurance and tax software, sometimes alongside machine learning models.
Why did expert systems decline?
Capturing knowledge was slow and expensive, rules multiplied and clashed, and systems failed on cases outside their rules. Updating them meant interviewing experts again. Learning from data proved cheaper and more flexible for many tasks.