Fuzzy Logic Outline Mamdani fuzzy inference

Part 2 „ The Mamdani fuzzy inference involves four steps:

Andrew Kusiak „ Mamdani fuzzy inference
Intelligent Systems Laboratory „ Sugeno fuzzy inference z Fuzzification of the input variables
2139 Seamans Center
The University of Iowa „ Case study z Rule evaluation
Iowa City, Iowa 52242 - 1527
„ Summary z Aggregation of the rule outputs
andrew-kusiak@uiowa.edu z Defuzzification
http://www.icaen.uiowa.edu/~ankusiak
Tel: 319 - 335 5934 Fax: 319 - 335 5669

Based on material provided by Professor Michael Negnevitsky

Consider a simple two-

two-input one-
one-output application that Step 1: Fuzzification Step 2: Rule Evaluation
includes three rules: The first step is to take the crisp inputs, x1 and y1 The second step is to take the fuzzified inputs,
Rule: 1 Rule: 1 (project funding and project staffing),
staffing), and determine μ(x=A1) = 0.5, μ(x=A2) = 0.2, μ(y=B1) = 0.1 and μ(y=B2) =
IF x is A3 IF project_funding is adequate the degree to which these inputs belong to each of the
OR y is B1 OR project_staffing is small 0.7, and apply them to the antecedents of the fuzzy
appropriate fuzzy sets. rules. If a given fuzzy rule has multiple antecedents,
THEN z is C1 THEN risk is low
Rule: 2 Rule: 2
the fuzzy operator (AND or OR) is used to obtain a
IF x is A2 IF project_funding is marginal Project Crisp Input Project Crisp Input single number that represents the result of the
funding x1 staffing y1
AND y is B2 AND project_staffing is large Inadequate Marginal Adequate Small Large
antecedent evaluation. This number (the truth value)
THEN z is C2 THEN risk is normal 1
A1 A2 A3
1 B1 B2 is then applied to the consequent membership
0.7
Rule: 3 Rule: 3
0.5
0.2 0.1
function.
0 0
IF x is A1 IF project_funding is inadequate x1 X y1 Y
μ (x = A1) = 0.5 μ (y = B1) = 0.1
THEN z is C3 THEN risk is high μ (x = A2) = 0.2 μ (y = B2) = 0.7
 R1: IF project_funding is adequate

To evaluate the disjunction of the rule antecedents,

OR project_staffing is small
THEN risk is low
Mamdani’
Mamdani’s rule evaluation Now the result of the antecedent evaluation can be
we use the OR fuzzy operation.
operation. Typically, fuzzy 1 1 1
applied to the membership function of the
expert systems make use of the classical fuzzy A3 B1
0.1 0.1
C1 C2 C3 consequent.
0.0 OR

operation union:
union: 0 x1 X 0 y1 Y
(max)
0 Z „ The most common method of correlating the rule
Rule 1: IF x is A3 (0.0) OR y is B1 (0.1) THEN z is C1 (0.1) consequent with the truth value of the rule
μA∪B(x) = max [μA(x), μB(x)] 1 1
0.7
1
antecedent is to cut the consequent membership
C1 C2 C3
A2 0.2 B2 0.2
AND function at the level of the antecedent truth. This
Similarly, in order to evaluate the conjunction of the 0 0
(min)
0
x1 X y1 Y Z
method is called clipping.
clipping. Since the top of the
rule antecedents, we apply the AND fuzzy operation Rule 2: IF x is A2 (0.2) AND y is B2 (0.7) THEN z is C2 (0.2)
membership function is sliced, the clipped fuzzy set
intersection:
intersection: 1 1
A1 0.5 0.5 C1 C2 C3 loses some information. However, clipping is still
μA∩B(x) = min [μA(x), μB(x)] often preferred because it involves less complex and
0 x1 X 0 Z
Rule 3: IF x is A1 (0.5) THEN z is C3 (0.5) faster mathematics, and generates an aggregated
R2: IF project_funding is marginal R3: IF project_funding is inadequate output surface that is easier to defuzzify.
defuzzify.
AND project_staffing is large THEN risk is high
THEN risk is normal

„ While clipping is a frequently used method, scaling Clipped and scaled membership functions Step 3: Aggregation of the rule outputs
offers a better approach for preserving the original Aggregation is the process of unification of the
shape of the fuzzy set. The original membership Clipped Scaled
outputs of all rules. We take the membership
function of the rule consequent is adjusted by Degree of Degree of functions of all rule consequents previously clipped or
multiplying all its membership degrees by the truth Membership Membership
scaled and combine them into a single fuzzy set.
1.0 1.0
value of the rule antecedent.
antecedent. This method, which
The input of the aggregation process is the list of
generally loses less information, can be useful in C2 C2
clipped or scaled consequent membership functions,
functions,
fuzzy expert systems.
0.2 0.2
and the output is one fuzzy set for each output
variable.
0.0 0.0
Z Z
 „ There are several defuzzification methods, but
Aggregation of the rule outputs Step 4: Defuzzification probably the most popular one is the centroid
The last step in the fuzzy inference process is technique.
technique. It finds the point where a vertical line
Clipped membership functions
defuzzification.
defuzzification. Fuzziness helps us to evaluate the would slice the aggregate set into two equal masses.
rules, but the final output of a fuzzy system has to be Mathematically this center of gravity (COG) can be
1 1 1
a crisp number. The input for the defuzzification expressed as:
C1 C2 C3
0.5 0.5
0.2
process is the aggregate output fuzzy set and the
0.1 0.2 b
0.1
output is a single number.
∫ μ A ( x ) x dx
0 Z 0 Z 0 Z 0 Z
z is C 1 (0.1) z is C 2 (0.2) z is C 3 (0.5) ∑
COG = a
b
∫ μ A ( x ) dx
a

„ Centroid defuzzification method finds a point Center of gravity (COG): Sugeno fuzzy inference
representing the centre of gravity of the fuzzy set, A, (0 + 10 + 20 ) × 0.1 + (30 + 40 + 50 + 60 ) × 0.2 + (70 + 80 + 90 + 100 ) × 0.5
COG = = 67 .4
on the interval, ab.
ab. 0 . 1 + 0 . 1 + 0 . 1 + 0 .2 + 0 . 2 + 0 . 2 + 0 .2 + 0 . 5 + 0 . 5 + 0 .5 + 0 .5 „ Mamdani-
Mamdani-style inference, requires finding the
„ A reasonable estimate can be obtained by calculating centroid of a two-
two-dimensional shape by integrating
it over a sample of points.
Degree of
Membership
across a continuously varying function. In general,
this process is not computationally efficient.
μ( x ) 1.0
1.0 0.8
0.8 0.6 „ Michio Sugeno suggested to use a single spike, a
0.6 A 0.4 singleton,
singleton, as the membership function of the rule
0.4 0.2 consequent. A singleton,
singleton, or more precisely a fuzzy
0.0 singleton,
singleton, is a fuzzy set with a membership function
0.2 0 10 20 30 40 50 60 70 80 90 100
a b
that is unity at a single particular point on the
67.4 Z
0.0 X universe of discourse and zero everywhere else.
150 160 170 180 190 200 210
 Sugeno-
Sugeno-style fuzzy inference is similar to the The most commonly used zero-
Sugeno-
Sugeno-style rule evaluation
zero-order Sugeno fuzzy
Mamdani method. Sugeno has modified a rule model applies fuzzy rules in the following form: 1 1 1
A3 B1
consequent. Instead of a fuzzy set, he used a 0.1 0.1
IF x is A 0.0 OR
(max)
mathematical function of the input variable. The 0 x1 X 0 y1 Y 0 k1 Z
AND y is B
format of the Sugeno-
Sugeno-style fuzzy rule is Rule 1: IF x is A3 (0.0) OR y is B1 (0.1) THEN z is k1 (0.1)
THEN z is k 1 1 1
0.7
IF x is A
where k is a constant. A2 0.2 B2 AND 0.2
AND y is B 0 x1 X 0 y1 Y
(min)
0 k2 Z
THEN z is f (x, y)
y) Rule 2: IF x is A2 (0.2) AND y is B2 (0.7) THEN z is k2 (0.2)
In this case, the output of each fuzzy rule is constant.
where x, y and z are linguistic variables; A and B are All consequent membership functions are represented
1
A1 0.5
1
0.5

fuzzy sets on universe of discourses X and Y, by singleton spikes. 0 x1 X 0 k3 Z

respectively; and f (x, y)
y) is a mathematical function. Rule 3: IF x is A1 (0.5) THEN z is k3 (0.5)

Weighted average (WA): How to make a decision on which method

Sugeno-
Sugeno-style aggregation of the rule outputs
to apply − Mamdani or Sugeno?
Sugeno?
μ(k1) × k1 + μ(k 2) × k 2 + μ(k 3) × k 3 0.1× 20 + 0.2 × 50 + 0.5 × 80
WA = = = 65 „ Mamdani method is widely accepted for capturing
μ(k1) + μ(k 2) + μ(k 3) 0.1 + 0.2 + 0.5
expert knowledge. It allows us to describe the
1 1 1 1
Sugeno-
Sugeno-style defuzzification expertise in more intuitive, more human-
human-like
0.5 0.5
0.1 0.2 0.1
0.2 manner. However, Mamdani-
Mamdani-type fuzzy inference
0 k1 Z 0 k2 Z 0 k3 Z 0 k1 k2 k3 Z entails a substantial computational effort.
z is k1 (0.1) z is k2 (0.2) z is k3 (0.5) ∑ „ Sugeno method is computationally effective and
0 z1 Z works well with optimization and adaptive
Crisp Output techniques, which makes it very attractive in control
z1 problems, particularly for dynamic nonlinear
systems.
Building a fuzzy expert system: A case study
„ A service center keeps spare parts and repairs failed
ones.
„ A customer brings a failed item and receives a spare
of the same type.
„ Failed parts are repaired, placed on the shelf, and
thus become spares.
„ The objective here is to advise a manager of the
service center on certain decision policies to keep
the customers satisfied.
Process of developing a fuzzy expert system
1. Specify the problem and define linguistic variables.
2. Determine fuzzy sets.
3. Elicit and construct fuzzy rules.
4. Encode the fuzzy sets, fuzzy rules and procedures
to perform fuzzy inference into the expert system.
5. Evaluate and tune the system.
Step 1: Specify the problem and define
linguistic variables
There are four main linguistic variables: average
waiting time (mean delay) m, repair utilisation
factor of the service centre ρ, number of servers s,
and initial number of spare parts n.

Linguistic variables and their ranges

Linguistic Value
Linguistic Variable: Mean Delay, m
Notation Numerical Range (normalised) Step 2: Determine fuzzy sets Fuzzy sets of Mean Delay m
Very Short VS [0, 0.3]
Short S [0.1, 0.5]
Degree of
Medium M [0.4, 0.7]
Linguistic Variable: Number of Servers, s
Fuzzy sets can have a variety of shapes. However, Membership
Linguistic Value
Small
Notation
S
Numerical Range (normalised)
[0, 0.35]
a triangle or a trapezoid can often provide an 1.0

Medium
Large
M
L
[0.30, 0.70]
[0.60, 1]
adequate representation of the expert knowledge, 0.8 VS S M
Linguistic Variable: Repair Utilisation Factor, ρ
Utilisation Factor, and at the same time, significantly simplifies the 0.6
Linguistic Value Notation Numerical Range
Low L [0, 0.6] process of computation. 0.4
Medium M [0.4, 0.8]
High H [0.6, 1] 0.2
Linguistic Variable: Number of Spares, n
Linguistic Value Notation Numerical Range (normalised) 0.0
Very Small VS [0, 0.30] 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Small S [0, 0.40] Mean Delay (normalized)
Rather Small RS [0.25, 0.45]
Medium M [0.30, 0.70]
Rather Large RL [0.55, 0.75]
Large L [0.60, 1]
Very Large VL [0.70, 1]
 Fuzzy sets of Number of Servers s Fuzzy sets of Repair Utilisation Factor ρ Fuzzy sets of Number of Spares n
Degree of Degree of Degree of
Membership Membership Membership
1.0 1.0 1.0
0.8 S M L 0.8 L M H 0.8 VS S RS M RL L VL
0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

0.0 0.0 0.0

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Number of Servers (normalized) Repair Utilization Factor Number of Spares (normalized)

The Fuzzy Associative Memory Rule Base 1

Step 3: Elicit and construct fuzzy rules
To accomplish this task, we might ask the expert to
describe how the problem can be solved using the
fuzzy linguistic variables defined previously.
Required knowledge also can be collected from
other sources such as books, computer databases,
flow diagrams and observed human behaviour.
(FAM) square
1. If (utilization_factor is L) then (number_of_spares is S)
2. If (utilization_factor is M) then (number_of_spares is M)
No of No of 3. If (utilization_factor is H) then (number_of_spares is L)
s spares
servers 4. If (mean_delay is VS) and (number_of_servers is S) then (number_of_spares is VL)
5. If (mean_delay is S) and (number_of_servers is S) then (number_of_spares is L)
6. If (mean_delay is M) and (number_of_servers is S) then (number_of_spares is M)
7. If (mean_delay is VS) and (number_of_servers is M) then (number_of_spares is RL)
L M S VS 8. If (mean_delay is S) and (number_of_servers is M) then (number_of_spares is RS)
9. If (mean_delay is M) and (number_of_servers is M) then (number_of_spares is S)
10. If (mean_delay is VS) and (number_of_servers is L) then (number_of_spares is M)
M RL RS S 11. If (mean_delay is S) and (number_of_servers is L) then (number_of_spares is S)
12. If (mean_delay is M) and (number_of_servers is L) then (number_of_spares is VS)
S VL L M

VS S M m Delay
3 + 3x3 rules

k dimensional cube for k variables

 Cube FAM of Rule Base 2
s
Rule table 2 Step 4: Encode the fuzzy sets, fuzzy rules
s

L S
VS VS VS
VS VS VS
S VS
L RL M RS
Rule m s ρ n Rule m s ρ n Rule m s ρ n
and procedures to perform fuzzy
VS VS VS M M M S
M
VS VS VS
VS VS VS
ρ VL L M ρ
1 VS S L VS 10 VS S M S 19 VS S H VL inference into the expert system
VS VS VS S
VS VS VS 2 S S L VS 11 S S M VS 20 S S H L
S VS VS VS H H
M VS S M
L s 3 M S L VS 12 M S M VS 21 M S H M
VS S M m
m L M RS S
4 VS M L VS 13 VS M M RS 22 VS M H M
One may choose one of two options:
M RS S VS
5 S M L VS 14 S M M S 23 S M H M • Build the system using a programming language
S S VS VS 6 M M L VS 15 M M M VS 24 M M H S such as C/C++ or Pascal, or
M
VS S M
s
m
7 VS L L S 16 VS L M M 25 VS L H RL • Apply a fuzzy logic development tool such as
L S S VS 8 S L L S 17 S L M RS 26 S L H M MATLAB Fuzzy Logic Toolbox or Fuzzy
9 M L L VS 18 M L M S 27 M L H RS
M VS VS VS
Knowledge Builder.
S VS VS VS
VS S M
L
3x3 = 27 rules The rules constructed from a 3-dimensional cube
m (3x3x3 = 27 rules)

Three-
Three-dimensional plot of Rule Base 1 Three-
Three-dimensional plot of Rule Base 1
Step 5: Evaluate and tune the system

0.6
The last, and the most laborious, task is to evaluate 0.6

and tune the system. We want to see whether the

number_of_spares

0.5

number_of_spares
0.5
fuzzy system meets the requirements. 0.4

0.4

Several test situations depend on the mean delay, 0.3

0.3
number of servers and repair utilisation factor. 0.2

0.2
0
The Fuzzy Logic Toolbox can generate surfaces 1
0
0.2 0.2 1
useful in analysis of the system’
system’s performance. 0.4 0.6
0.8
0.4 0.6
0.8

0.4 0.4
0.6 0.2 0.6 0.2
mean_delay 0 number_of_servers mean_delay 0 utilization_factor
 Three-
Three-dimensional plot of Rule Base 2 Three-
Three-dimensional plot of Rule Base 2

0.35 0.5 However, even now, the expert might not be

satisfied with the system performance.
number_of_spares

0.3
0.4

number_of_spares
0.25
0.3
To improve the system performance, we may use
0.2 additional sets − Rather Small and Rather Large −
0.2
on the universe of discourse Number of Servers,
Servers,
0.15
and then extend the rule base.
0 0

0.2 1 0.2 1
0.8 0.8
0.4 0.6 0.4 0.6
0.4 0.4
0.6 0.2 0.6 0.2
mean_delay 0 number_of_servers mean_delay 0 utilization_factor

Cube FAM of Rule Base 3 Three-

Three-dimensional plots for Rule Base 3
Modified fuzzy sets of Number of Servers s
s

VS VS VS L L M RS
VS VS VS
SL S VS
VS VS VS RL RL M RS
Degree of VS VS VS
RL S S VS 0.35
VS VS VS M M M S
Membership VS VS VS
1.0 M VS VS VS
VS VS VS s RS VL RL RS
VS VS VS 0.3
RS VS VS VS

number_of_spares
0.8 VS VS VS ρ L M RS S S VL L M ρ
S RS M RL L VS VS VS
S VS VS VS H H
M VS S M 0.25
RL M RS S
0.6 VS S M
L m
m M RS S VS
0.4 0.2
s RS S VS VS
0.2 L S S VS S S VS VS 0.15
M
0.0 VS S M
RL S S VS
m 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
M VS VS VS
Number of Servers (normalized) 0.2 1
RS VS VS VS 0.8
0.4 0.6
S VS VS VS 0.4
0.6 0.2
L
VS S M 0
mean_delay number_of_servers
m
Three-
Three-dimensional plots for Rule Base 3 Tuning fuzzy systems 4. Review the existing rules, and if required add new
1. Review model input and output variables, and if rules to the rule base.
0.5
required redefine their ranges. 5. Examine the rule base for opportunities to write
0.4 2. Review the fuzzy sets, and if required define hedge rules to capture the pathological behaviour
number_of_spares

additional sets on the universe of discourse. of the system.

0.3
The use of wide fuzzy sets may cause the fuzzy 6. Adjust the rule execution weights. Most fuzzy
0.2 system to perform roughly. logic tools allow control of the importance of rules
3. Provide sufficient overlap between neighbouring by changing a weight multiplier.
0

0.2 1
sets. It is suggested that triangle-
triangle-to-
to-triangle and 7. Revise shapes of the fuzzy sets. In most cases,
0.8
0.4 0.6 trapezoid-
trapezoid-to-to-triangle fuzzy sets should overlap fuzzy systems are highly tolerant of a shape
0.4
0.6
0
0.2 between 25% to 50% of their bases. approximation.
mean_delay utilization_factor