Fuzzy Sets, Linguistic Variables & Approximate Reasoning

A crisp set answers whether an element belongs: $x \in A$ or $x \notin A$. A fuzzy set answers to what degree an element matches a concept by a membership function $\mu_A(x) \in [0,1]$. This is useful when the category boundary is conventional or gradual—“high temperature,” “near the target,” “large error”—and a hard threshold would invent a discontinuity that the concept does not possess.

Membership is not probability #

For a measured temperature of $27\,^{\circ}\mathrm{C}$, $\mu_{\text{warm}}(27)=0.8$ means that 27 degrees is highly compatible with the designed concept warm. It does not mean there is an 80% chance that the temperature is warm, nor that the sensor is 80% reliable.

Probability distributes belief or frequency across uncertain alternatives. Fuzzy membership grades compatibility with a set whose boundary is vague. The two can coexist: a noisy sensor can give a probability distribution over temperature, and each possible temperature can have a fuzzy membership in warm.

Common membership functions #

A triangular membership function with feet $a,c$ and peak $b$ is

$$\mu(x)=\max\!\left(\min\!\left(\frac{x-a}{b-a},\frac{c-x}{c-b}\right),0\right), \qquad a<b<c.$$

Trapezoids add a plateau; Gaussian functions provide smooth tails; sigmoids express one-sided concepts. Shape choice changes interpolation and sensitivity, so it is part of the model rather than decoration.

For a temperature universe from 0 to $40\,^{\circ}\mathrm{C}$, overlapping sets might define cool, warm, and hot. At $27\,^{\circ}\mathrm{C}$ both warm and hot may be active. That overlap lets rules blend rather than switch abruptly.

flowchart LR
  X[Crisp value x] --> C[Cool membership]
  X --> W[Warm membership]
  X --> H[Hot membership]
  C --> V[Membership vector]
  W --> V
  H --> V

Alpha-cuts recover ordinary sets #

For $\alpha \in (0,1]$, the $\alpha$-cut is

$$A_\alpha=\{x\mid\mu_A(x)\ge\alpha\}.$$

Alpha-cuts turn a fuzzy set into nested crisp regions. They are useful for interval reasoning and for checking whether membership functions behave monotonically as the required compatibility level rises. If $\alpha_2>\alpha_1$, then $A_{\alpha_2}\subseteq A_{\alpha_1}$.

Set operations require an operator choice #

Zadeh's original operators use

$$\mu_{A\cap B}(x)=\min(\mu_A(x),\mu_B(x)),\qquad \mu_{A\cup B}(x)=\max(\mu_A(x),\mu_B(x)),$$

and $\mu_{\neg A}(x)=1-\mu_A(x)$. Other valid fuzzy logics use product for AND, probabilistic sum for OR, or different implication operators. These choices are not generally interchangeable: they change rule firing strength and the final surface.

A t-norm generalizes conjunction and a t-conorm generalizes disjunction. At minimum, the selected operators should satisfy boundary, monotonicity, commutativity, and associativity properties appropriate to their role.

Linguistic variables bind words to a domain #

A linguistic variable is not a string label. It consists of:

• a name such as temperature error;
• a numerical universe and units, such as $[-20,20]\,^{\circ}\mathrm{C}$;
• terms such as negative large, near zero, and positive large;
• one membership function for each term;
• modifiers and rules that define how the terms are used.

Scaling is critical. Reusing the same normalized partition after a sensor-range change silently changes the physical meaning of every rule.

From propositions to approximate rules #

A typical rule is:

IF error is positive-large AND error-rate is positive-small, THEN cooling-command is high.

For input values $(e,\dot e)$, the antecedent firing strength using minimum conjunction is

$$w_i=\min\bigl(\mu_{\text{positive-large}}(e), \mu_{\text{positive-small}}(\dot e)\bigr).$$

The inference system must then define how $w_i$ modifies the consequent, how multiple rules aggregate, and how a numerical output is recovered. Those steps distinguish families such as Mamdani and Takagi–Sugeno inference.

Partition design is model design #

Useful partitions usually satisfy four practical properties:

1. Coverage: every admissible input activates at least one rule.
2. Overlap: neighboring terms blend through transition regions.
3. Distinguishability: terms are not so similar that rules become redundant.
4. Boundary behavior: saturation and out-of-range inputs have explicit semantics.

Adding more terms increases resolution but also multiplies the rule base. With $m$ terms on each of $n$ inputs, a full grid may require $m^n$ antecedent combinations. Sparse rule bases avoid that explosion but create uncovered regions unless coverage is checked directly.

What fuzzy reasoning can and cannot claim #

Fuzzy logic can make vague operational categories explicit, interpolate among local rules, and preserve readable linguistic structure. It does not by itself:

• learn correct membership functions;
• prove a rule base complete or consistent;
• quantify random uncertainty;
• guarantee closed-loop stability;
• make a subjective label objective.

The system remains an engineered function from inputs to outputs. Its operators, partitions, and rules require the same validation envelope, sensitivity analysis, and failure testing as any other model.