OE-11.5 Fuzzy Logic System Components

What this is and why it exists

This unit assembles the previous one into something that works. Crisp measurement in, fuzzy reasoning through a rule base, crisp action out.

The rule base is where the system's knowledge lives, and it has an unusual property worth appreciating. It is written in near-plain language: *if the temperature is warm and rising slowly, reduce the heat a little*. And it is usually written by a domain expert rather than learned from data.

That is the practical appeal of fuzzy control. You can read the system's reasoning, and whoever understands the process can write it without becoming a programmer.

The vocabulary

• Crisp value — an ordinary number, such as a sensor reading.
• Fuzzification — turning a crisp value into memberships of fuzzy sets.
• Rule base — the collection of if-then rules.
• Antecedent — the if part of a rule.
• Consequent — the then part.
• Inference — combining the rules that fired into one fuzzy output.
• Defuzzification — turning that fuzzy output back into a crisp value.
• Centroid method — taking the centre of area of the output shape.

The mental model

The pipeline has four stages and runs in order.

Fuzzification takes each crisp input and finds its membership in the relevant sets. A temperature of thirty-two degrees might be 0.3 warm and 0.7 hot. Note that several sets can be true at once, to different degrees — that overlap is intended, and it is what produces smooth behaviour.

The rule base is consulted. Typically several rules fire, each to the degree its conditions are met. The rules are the knowledge, and they are readable, which means a domain expert can check them and argue with them. That is a genuine advantage over a network whose knowledge is spread across weights and cannot be read at all.

Inference combines the rules that fired into one fuzzy output. A rule that fired weakly contributes weakly, and the results are aggregated into a single output shape.

Defuzzification turns that shape back into one number, because an actuator needs a number. Several methods exist. The centre-of-area approach is the common one and gives smooth output, at the cost of more computation. Taking the largest membership is cheap and can jump abruptly as which rule dominates changes. Choose by whether smoothness or cost matters more. And notice that the choice affects behaviour, so it is a design decision rather than an implementation detail.

Here is why all this machinery is worth it. The overlap between sets means the output moves gradually as inputs change, instead of jumping at a threshold. That is the smoothness a classical rule-based controller cannot give, and it is why these turn up in appliances, cameras and process control.

The honest limitation is worth stating too. The rule base has to come from somebody who understands the process. When no such person exists, or the process is not understood, this approach has nothing to offer. A method that learns from data does.

What you should now be able to explain or do

Run the four stages in order and say what each converts. Say why several sets being true at once is intended. Say what makes a rule base readable and why that is an advantage. Explain inference as weighted aggregation of the rules that fired. Compare defuzzification methods on smoothness against cost, and treat the choice as a design decision. State when this approach has nothing to offer.

Check yourself