Nexus Learn

Lecture 23: Fuzzy Inference: Mamdani and Sugeno

Lecture Slides CS-511 Artificial Intelligence 7 min read Updated 8 Nov 2026

↓ Download slides (PowerPoint)

Recap of Lecture 22 and Homework Check

LAST LECTURE · 5 MIN

Today: the fan has five rules and each gives a different opinion. How do we combine them into ONE fan speed?

Today's Plan (50 minutes)

ROADMAP

0-5

Recap of Lecture 22 and homework check

S02

5-8

Plan and learning outcomes

S03-S04

8-18

Module 1: The fuzzy inference process - five steps on the fan

S05-S08

18-27

Module 2: Mamdani - implication and defuzzification

S09-S11

27-35

Module 3: Sugeno - zero and first order; choosing a method

S12-S15

35-43

Class activity: inference by hand

S16

43-48

Common mistakes and recap quiz

S17-S18

48-50

Homework and next lecture

S19

Learning Outcomes (SLOs)

BY THE END OF THIS LECTURE YOU WILL BE ABLE TO

SLO 23.1

Explain the five steps of fuzzy inference: fuzzify, apply operators, implication, aggregation, defuzzify

→ CLO3 · C2 Understand

SLO 23.2

Apply Mamdani inference: firing strengths, clip or scale, max aggregation and defuzzification methods

→ CLO3 · C3 Apply

SLO 23.3

Apply zero- and first-order Sugeno inference and compare it with Mamdani

→ CLO3 · C3 Apply

  • Course Learning Outcome(s) served
  • CLO3: Apply various techniques and algorithms of AI used in general problem solving, optimization problems, constraint satisfaction problems, and game programming.

Slide 5

MODULE 1

The Fuzzy Inference Process

Five steps from crisp inputs to a crisp output

Five Steps of Fuzzy Inference

MODULE 1 · THE PROCESS

Fuzzify

Operators

Implication

Aggregate

Defuzzify

  • Crisp input -> degrees.
  • 30 °C: comfortable 0.33, hot 0.20

AND = min, OR = max, NOT = 1 - μ -> firing strength of each rule

Shape the THEN set by the strength (clip or scale)

Join all rule outputs into one fuzzy set (max)

One crisp number: centroid of the set

THE FAN RULE BASE (P08-P10)

  • R1 IF temp is cold THEN fan is slow
  • R2 IF temp is comfortable AND hum is dry THEN fan is slow
  • R3 IF temp is comfortable AND hum is NOT dry THEN fan is medium
  • R4 IF temp is hot AND hum is dry THEN fan is medium
  • R5 IF temp is hot AND hum is NOT dry THEN fan is fast

Steps 1 and 2 on the Fan: 30 °C, 40 %

MODULE 1 · WORKED EXAMPLE

REAL OUTPUT: L23/STEPS.PY

  • Step 1 Fuzzify the inputs
  • temp 30 C is cold 0.00
  • temp 30 C is comfortable 0.33
  • temp 30 C is hot 0.20
  • humidity 40 % is dry 0.33
  • Step 2 Apply AND (min) / NOT (1 - mu) -> firing strengths
  • R1 IF cold THEN slow 0.00
  • R2 IF comfortable AND dry THEN slow 0.33
  • R3 IF comfortable AND NOT dry THEN medium 0.33
  • R4 IF hot AND dry THEN medium 0.20
  • R5 IF hot AND NOT dry THEN fast 0.20
  • Steps 3-5 Mamdani (clip, max, centroid) -> 45.04 %
  • Sugeno (weighted average) -> 45.88 %

Firing strength: how strongly a rule applies, from 0 to 1. Several rules usually fire at the same time.

Ask: why is NOT dry = 0.67? Because 1 - 0.33 = 0.67 (complement from Lecture 22).

Steps 3 to 5: Clip, Aggregate, Defuzzify

MODULE 1 · THE INFERENCE DIAGRAM

  • Read across a row: input degrees -> min -> clip the THEN set.
  • R1 adds nothing: 30 °C is not cold (strength 0).
  • Aggregate: max of the four amber shapes.
  • Defuzzify: centroid = balance point = 45.04 %.

Centroid: z* = Σ μ(z)·z / Σ μ(z) over the output axis - the point where the shape would balance on a finger.

Slide 9

MODULE 2

Mamdani Inference

Ebrahim Mamdani, 1975 - fuzzy output sets

Implication: Clip or Scale?

MODULE 2 · STEP 3 CHOICES

REAL OUTPUT

  • input clip scale
  • 30 C, 40 % 45.04 44.51
  • 36 C, 70 % 82.78 83.33
  • 34 C, 65 % 81.43 83.33

Remember: both methods agree on the strongest opinion; they differ a little in how the shapes overlap.

Defuzzification Methods

MODULE 2 · STEP 5 CHOICES

REAL OUTPUT: L23/DEFUZZ_METHODS.PY

  • method 30 C/40 % 36 C/70 %
  • centroid 45.04 82.78
  • bisector 43.66 84.64
  • mom 33.00 95.00
  • som 0.00 90.00
  • lom 66.00 100.00

Why centroid? it uses the whole shape, so a small change in temperature gives a small change in fan speed. MOM can jump from 33 to 95.

Exam tip: for a flat top from 20 to 60: SOM = 20, LOM = 60, MOM = (20 + 60)/2 = 40.

Slide 12

MODULE 3

Sugeno Inference

Takagi-Sugeno-Kang, 1985 - crisp rule outputs

Sugeno: Zero-Order and First-Order

MODULE 3 · WEIGHTED AVERAGE

RULE OUTPUTS AND THE WEIGHTED AVERAGE

REAL OUTPUT: L23/SUGENO_ORDERS.PY

  • zero-order: z = constant e.g. fast: z = 83
  • first-order: z = p·t + q·h + r
  • e.g. fast: z = 50 + 0.8·t + 0.1·h
  • output z* = Σ wi·zi / Σ wi
  • input zero-order first-order
  • 30 C, 40 % 45.88 49.31
  • 36 C, 70 % 83.00 85.80
  • 45 C, 40 % 72.00 85.00
  • 45 C, 90 % 83.00 95.00

Mamdani vs Sugeno

MODULE 3 · COMPARISON

REAL OUTPUT: L23/COMPARE.PY

  • input Mamdani Sugeno
  • 20 C, 50 % 50.00 50.00
  • 30 C, 40 % 45.04 45.88
  • 33 C, 60 % 80.56 83.00
  • 36 C, 70 % 82.78 83.00
  • 42 C, 35 % 65.46 66.50
  • Work per input:
  • Mamdani: 5 clipped sets x 101 points, max, centroid
  • Sugeno: 5 products, 2 sums, 1 division

Choose: Mamdani when people must read the rules; Sugeno for fast control and for learning from data (ANFIS, P11).

Choosing a Method in Practice

MODULE 3 · USES

Mamdani

Sugeno

Sugeno + learning

Rules people can read and check: a crop-disease advisor for Burewala cotton, expert systems (L19).

Fast control in small chips: inverter AC, cruise control. Smooth output, few calculations.

ANFIS learns the formulas and membership functions from data - P11a/P11b this week.

Activity: Inference by Hand

CLASS ACTIVITY · PAIRS · 8 MIN

Pairs, 6 min + 2 min check. Use the fan rules of S06 and your calculator.

Put in order: aggregate, fuzzify, defuzzify, implication, apply operators.

First-order: w = 0.5, z = 30 + t with t = 35; w = 0.5, z = 15. Output?

36 °C, 70 %: comfortable, hot, dry degrees? Which rule fires? Sugeno output?

Flat top from 20 to 60 at height 0.5. SOM, MOM, LOM?

Sugeno: strengths 0.2 (z = 17), 0.6 (z = 50), 0.2 (z = 83). Output?

Mamdani or Sugeno? (a) doctors must check the rules (b) a tiny chip in an AC

Common Mistakes

WATCH OUT

WRONG IDEA

CORRECT IDEA

✗ Only the strongest rule matters.

✓ All firing rules contribute; aggregation joins them.

✗ Fuzzy AND of 0.33 and 0.67 is 0.22.

✓ Standard AND is min: 0.33.

✗ Sugeno divides by the number of rules.

✓ Divide by the sum of strengths Σ wi.

✗ Mamdani clips the input set.

✓ It clips the OUTPUT (THEN) set at the strength.

✗ Defuzzify each rule, then average.

✓ Mamdani aggregates first, then defuzzifies once.

Recap Quiz

CHECK YOUR UNDERSTANDING · 3 MIN

  • Q1. The step that turns a fuzzy set into one number is:
  • (a) Fuzzification (b) Implication (c) Aggregation (d) Defuzzification
  • Q2. Strength of 'comfortable (0.33) AND NOT dry (dry = 0.33)':
  • (a) 0.11 (b) 0.33 (c) 0.67 (d) 1.0
  • Q3. In a zero-order Sugeno system the THEN-part is:
  • (a) A fuzzy set (b) A constant (c) A linear formula (d) A probability
  • Q4. Strengths 0.5 (z = 20) and 0.5 (z = 80). Sugeno output:
  • (a) 40 (b) 50 (c) 60 (d) 100

Homework and Next Lecture

BEFORE THE NEXT CLASS

  • Task 1: Fan rules at 32 °C, 30 % humidity: find the degrees, the five firing strengths and the zero-order Sugeno output (show the weighted average).
  • Task 2: Same input, first-order outputs: slow 15, medium 30 + t, fast 50 + 0.8t + 0.1h. Output?
  • Task 3: Aggregated set {0.2/10, 0.5/30, 0.5/50, 0.8/70, 0.3/90}. Centroid and MOM?

NEXT LECTURE

  • Lecture 24:
  • Machine Learning: Supervised, Unsupervised and Reinforcement

← All CS-511 material