Lecture 08.2: Fuzzy Expert Systems

In classical Boolean logic, everything is crisp: a statement is either True ($1$) or False ($0$). However, human experts rarely think in binary. An expert might say:

"Though the power transformer is slightly overloaded, I can keep this load for a while."

While another human expert easily understands this statement, a knowledge engineer faces immense difficulty representing such vague and ambiguous terms in traditional binary computer logic.

Fuzzy Logic, introduced by Lotfi Zadeh (Professor and Head of Electrical Engineering at UC Berkeley) in his landmark 1965 paper "Fuzzy sets", resolves this challenge by formalizing fuzziness—mirroring the intrinsic way people think, make decisions, and interpret natural language qualifiers.

---

Crisp Sets vs. Fuzzy Sets

Unlike two-valued Boolean logic, fuzzy logic is multi-valued, dealing in Degrees of Membership where truth values range continuously in the interval $[0.0, 1.0]$.

Logical Value Range: Crisp vs. Fuzzy

• Boolean (Crisp) Logic: Binary step-threshold (Black or White, $\{0, 1\}$).
• Fuzzy Logic: Continuous multi-valued spectrum (Shades of Grey, $[0.0, 1.0]$).

The Boundary Problem: Degree of Membership of 'Tall Men'

Consider evaluating a group of individuals to determine whether they belong to the set of "Tall Men". Under a crisp definition, anyone $\ge 180\text{ cm}$ is considered tall ($1$), and anyone $< 180\text{ cm}$ is not ($0$).

[!IMPORTANT] The Crisp Boundary Paradox: Under crisp logic, Tom (181 cm) is tall ($1$), whereas David (179 cm) is NOT tall ($0$), despite a negligible 2 cm height difference. Fuzzy set theory eliminates this artificial boundary by assigning Tom a membership degree of $0.82$ and David $0.78$.

---

Set Representation & Zadeh's Notation

1. Mathematical Definitions

• Crisp Set: Membership function $\mu_A(x)$ is strictly binary:

$\mu_A(x) = \begin{cases} 1 & \text{if } x \in A \\ 0 & \text{if } x \notin A \end{cases}$

• Fuzzy Set: Membership function $\mu_A(x)$ is a continuous mapping:

$\mu_A(x) \in [0.0, 1.0]$

2. Zadeh's Notation for Discrete Fuzzy Sets

A discrete fuzzy set $A$ is represented using Zadeh's Notation:

$A = \sum_{i=1}^{n} \mu_i / x_i = \mu_1/x_1 + \mu_2/x_2 + \dots + \mu_n/x_n$

[!CAUTION] Zadeh Notation Interpretation:

• The slash / is not division; it links membership degree $\mu_i$ to element $x_i$ ($\text{degree} / \text{element}$).
• The + and $\sum$ symbols represent set union, not arithmetic addition.

• Crisp Example: $A_{\text{crisp}} = \{0/152 + 0/167 + 1/198\} \equiv \{198\}$
• Fuzzy Example: $A_{\text{fuzzy}} = \{0.0/152 + 0.15/167 + 0.98/198\}$

3. Multi-Membership Example: Body Temperature

A patient with a body temperature of $38^\circ\text{C}$ can belong to multiple fuzzy sets concurrently:

$\text{low temp} = \{ (1, 35), (1, 36), (0, 37) \}$ $\text{normal temp} = \{ (0, 36), (1, 36.8), (0, 37.2) \}$ $\text{raised temp} = \{ (0, 37), (1, 37.8), (0.9, 38), (0.39, 39.2) \}$ $\text{strong fever} = \{ (0, 37.5), (0.5, 39.5), (0.9, 41) \}$

At $38^\circ\text{C}$, the patient has raised temperature to degree $0.9$ and strong fever to degree $0.1$.

---

Linguistic Variables and Values

• Linguistic Variable: A variable whose values are words or natural language phrases (e.g., Project Funding, Project Staffing, Risk).
• Linguistic Value: The fuzzy set labels qualifying the linguistic variable.

---

The Four Steps of Fuzzy Inference

Fuzzy inference formulates a mapping from crisp inputs to a crisp output through four sequential steps:

---

Step-by-Step Benchmark: Project Risk Evaluation (2 Inputs, 1 Output)

Rules in the Knowledge Base

---

Step 1: Fuzzification

Given crisp numerical inputs $x_1$ (project_funding) and $y_1$ (project_staffing), we determine their degree of membership in each linguistic fuzzy set:

• Project Funding ($x_1$):
  • $\mu_{A1}(x_1) = 0.5$ (inadequate)
  • $\mu_{A2}(x_1) = 0.2$ (marginal)
  • $\mu_{A3}(x_1) = 0.0$ (adequate)
• Project Staffing ($y_1$):
  • $\mu_{B1}(y_1) = 0.1$ (small)
  • $\mu_{B2}(y_1) = 0.7$ (large)

---

Step 2: Rule Evaluation

We evaluate rule antecedents using standard fuzzy logic operators:

• Fuzzy Disjunction (OR / Optimistic): Evaluated using the maximum operator ($\max$).

$A \vee B \triangleq \max(A, B)$

• Fuzzy Conjunction (AND / Pessimistic): Evaluated using the minimum operator ($\min$).

$A \wedge B \triangleq \min(A, B)$

Evaluating Rule Antecedents:

1. Rule 1 (Disjunction): IF $x$ is $A_3$ ($0.0$) OR $y$ is $B_1$ ($0.1$):

$\mu_{\text{Rule 1}}(z) = \max[\mu_{A3}(x), \mu_{B1}(y)] = \max[0.0, 0.1] = 0.1 \quad \implies \text{Risk is Low } (C_1) \text{ to degree } 0.1$

2. Rule 2 (Conjunction): IF $x$ is $A_2$ ($0.2$) AND $y$ is $B_2$ ($0.7$):

$\mu_{\text{Rule 2}}(z) = \min[\mu_{A2}(x), \mu_{B2}(y)] = \min[0.2, 0.7] = 0.2 \quad \implies \text{Risk is Normal } (C_2) \text{ to degree } 0.2$

3. Rule 3 (Single Antecedent): IF $x$ is $A_1$ ($0.5$):

$\mu_{\text{Rule 3}}(z) = \mu_{A1}(x) = 0.5 \quad \implies \text{Risk is High } (C_3) \text{ to degree } 0.5$

---

Step 3: Aggregation

Aggregation is the process of unifying the outputs of all fuzzy rules into a single aggregate fuzzy set. Applying the evaluated antecedent truths to their respective consequent membership curves yields an overall risk output profile:

• Low Risk ($C_1$) to degree $0.1$
• Normal Risk ($C_2$) to degree $0.2$
• High Risk ($C_3$) to degree $0.5$

---

Step 4: Defuzzification

Defuzzification converts the aggregate output fuzzy set into a single crisp numerical value.

Using the Centroid Method (finding the center of gravity of the combined area under the curve), the aggregate fuzzy risk profile is converted into a definitive risk index:

$\text{Crisp Output Risk} = \mathbf{67.4\%}$

---

Concluding Remarks

Fuzzy Expert Systems bridge the gap between human natural reasoning and precise computer algorithms. While traditional rule-based systems require rigid binary boundaries, fuzzy systems accommodate vagueness and linguistic terms seamlessly. However, designing fuzzy systems requires careful domain expertise to define membership function shapes and rule sets appropriately.