Expert Systems and Rules

AI before machine learning

Today when people say "AI" they usually mean systems that learn from data, like the ones in What Is Machine Learning. But for decades, the most successful AI worked in a completely different way. It used rules — clear if-then instructions written by human experts. These programs are called expert systems, and understanding them shows you that AI is more than one idea.

The core idea: capture an expert's rules

Imagine a doctor who diagnoses skin rashes. Over years they have built up rules in their head:

• If the rash is red and itchy and spreading, then it might be an allergy.
• If there is a fever and the spots are blistered, then consider something more serious.

An expert system tries to bottle that expertise by writing those rules into a program. Then anyone can run the program and get advice that follows the expert's reasoning, even when the expert is not there.

The two main parts

Every classic expert system has two pieces:

1. The knowledge base — a big collection of if-then rules and known facts. This is the "expertise."
2. The inference engine — the part that takes the facts you give it and applies the rules to reach a conclusion.

Here is the flow. You enter facts: "red rash, itchy, no fever." The inference engine searches the rules, finds the ones whose "if" parts match, and fires them. Sometimes one rule's conclusion becomes a fact that triggers another rule. This chaining of rules is how the system reasons its way to an answer like "possible allergy — see a pharmacist."

A worked example

Suppose a simple expert system helps decide what to do about a sick houseplant. Its rules:

• Rule 1: If leaves are yellow AND soil is wet, then overwatering.
• Rule 2: If leaves are crispy AND soil is dry, then needs water.
• Rule 3: If overwatering, then advise "let soil dry out."

You enter: yellow leaves, wet soil. The engine fires Rule 1 → concludes overwatering. That conclusion now matches Rule 3 → advises "let the soil dry out." Two rules chained together gave sensible advice, and you can see exactly how.

The big advantage: it can explain itself

This is a genuine strength. Because every step is an explicit rule, the system can tell you why: "I said overwatering because the leaves are yellow and the soil is wet (Rule 1)." Compare that to a neural network, which often cannot explain its decisions clearly. In fields like medicine, finance or law, being able to audit a decision is hugely valuable. Rule-based systems are transparent by design.

The limits that hold them back

Expert systems sound great, so why did machine learning take over for so many tasks? Because rules have real, stubborn weaknesses:

• No learning. The system never improves on its own. Every rule must be written, tested and updated by hand. This is slow and expensive.
• The knowledge bottleneck. Getting all of an expert's knowledge out of their head and into clean rules is incredibly hard. Experts often cannot fully explain their own intuition.
• Brittleness. Step outside the rules and the system is lost. It has no common sense to fall back on, so an unexpected case gives no answer or a wrong one.
• Explosion of rules. Messy real-world problems can need thousands of rules that start to conflict and become impossible to maintain.

In short, rules shine when a problem is clear, narrow and stable, and break when it is fuzzy, broad or always changing — which is exactly where machine learning, learning patterns from examples, does better.

Two tools, not one winner

Expert systems are not a failed idea — they are a different tool for a different job. Tax software, troubleshooting wizards and safety checklists still run on rules because the logic is fixed and every decision must be explainable. Modern systems often combine both: rules to guarantee safe, clear behaviour, and machine learning to handle the messy patterns. Knowing both helps you see that "AI" was never just one thing. For another rules-based approach you can read, see Decision Trees Explained.