Introduction

• What is a fuzzy inference system (FIS)?

A nonlinear mapping that derives its output based
on fuzzy reasoning and a set of fuzzy if-then rules.
The domain and range of the mapping could be
fuzzy sets or points in a multidimensional spaces.
• Also known as
▫ Fuzzy models
▫ Fuzzy associate memory
▫ Fuzzy-rule-based systems
▫ Fuzzy expert systems
▫ F
Fuzzy Logic
L i C Controller
t ll
 Structure of FIS
• The basic structure of a fuzzy inference system consists
of three conceptual components:
▫ A rule base, which contains a selection of fuzzy rules
▫ A database (or dictionary),
dictionary) which defines the
membership functions used in the fuzzy rules
▫ And a reasoning mechanism, which performs the
inference procedure upon the rules and given facts to
derive a reasonable output or conclusion.

Rule base Database

(Fuzzy rules) (MFs)
input
p output
Fuzzy reasoning
Structure of FIS
 Example: Fuzzy Control Systems

Input
Inference
Fuzzifier Defuzzifier Plant Output
Engine

Fuzzy
Knowledge base
 Inputs and Outputs of FIS

• The basic fuzzy y inference system

y can take
either fuzzy inputs or crisp inputs, but the
outputs it produces are almost always fuzzy
sets.
• Sometimes it is necessary to have a crisp
output especially in a situation where a fuzzy
output,
inference system is used as a controller.
• Therefore,
Therefore we need a method of defuzzification
to extract a crisp value to represent a fuzzy set.
 Block Diagram for a FIS with Crisp
Output
• FIS implements
p pp g from its
a nonlinear mapping
input space to output space. This mapping is
accomplished by a number of fuzzy if-then rules.
 Nonlinearity

In the case of crisp inputs & outputs, a fuzzy

inference system implements a nonlinear
mapping from its input space to output space.
 Three Popular Fuzzy Models (FIS)

• Ebrahim Mamdani Fuzzy

y Models
• Sugeno Fuzzy Models
• Tsukamoto Fuzzy Models

• The differences between these three FISs lie in

the consequents of their fuzzy rules, and thus
their aggregation and defuzzification procedures
differ accordingly
accordingly.
 Mamdani Fuzzyy Models
• The most commonly used fuzzy inference technique is the so-
called M d i method.
ll d Mamdani th d
• In 1975, Professor Ebrahim Mamdani of London University
built one of the first fuzzy systems to control a steam engine
and boiler combination
combination. He applied a set of fuzzy rules
supplied by experienced human operators.
• The Mamdani-style fuzzy inference process is performed in
four steps:

1. Fuzzification of the input variables

2. Rule evaluation (inference)

3. Aggregation of the rule outputs (composition)

4. Defuzzification
 Example:
p fuzzy
y control

Crisp
Crisp Inference Defuzzilier Values
Values Fuzzifier
Engine

Membership Functions

Fuzzy (IF THEN) Rules

Fuzzy Variables
Linguistic Variables
 Mamdani Fuzzyy Models

Mamdani composition of three SISO fuzzy outputs

http://en.wikipedia.org/wiki/Fuzzy_control_system
 Two-rule Mamdani with min and max
operators
The mamdani FIS using min and max for T-norm and S-norm
, and subject to two crisp inputs x and y
 Two-rule Mamdani FIS with max and
product operators
The mamdani FIS using product and max for T-norm and S-norm
, and subject to two crisp inputs x and y
 Example

Rule 1: If p
pressure is low and temperature
p is high
g then p
power is low
Rule 2: If pressure is average and temperature is warm
then power is moderate
 Mamdani Fuzzyy Models

Two input, one

Two-input, ouput example:
one-ouput
If x is Ai and y is Bk then z is Cm(i,k)

y Y

L M H VH
B1 B2 B3 B4
\ C2 C3 C4 L Z L M H
A1
M L M H VH
x A2 C5 C6 C7 C8 x
H M H VH VH
A3 C9 C10 C11 C12

z Playing time
 Whyy Defuzzification Is Needed?

• In many applications we have to use crisp

values as inputs for controlling of machines and
systems.

•S
So, we have
h tto use a defuzzifier
d f ifi tot convertt a
fuzzy set to a crisp value.
 Defuzzification

• Defuzzification refers to the way

y a crisp
p value is
extracted from a fuzzy set as a representative
value.
• Defuzzification Methods:
▫ Centroid of Area
▫ Bi
Bisector
t off Area
A
▫ Mean of Max
▫ Smallest of Max
▫ Largest of Max
 Centroid of Area ((zCOA )

• where µA is aggregated output MF.

• This is the most widely adopted defuzzification
strategy, which is reminiscent of the calculation
of expected values of probability distributions.
 Bisector of Area ((zBOA )

• zBOA satisfies

| ∈   | ∈  

• That is, the vertical line z = zBOA partitions the

region between z = α, z = β, y = 0 and y = µA(z)
i t two
into t regions
i with
ith th
the same area.
 Mean of Maximum ((zMOM )
• zMOM is the mean of maximizing z at which the MF
reaches maximum µ*. In Symbols,

| ∗
′ ∈    
• In particular, if µA(z) has a single maximum at
z = z*, then the zMOM = z*.
• Moreover,
M if µA(z)
( ) reaches
h itits maximum
i whenever
h
 

then
zMOM = (zleft + zright)/2
 Smallest of Maximum (zSOM ) and
Largest of Maximum (zLOM )

• zSOM is the minimum ((in terms of magnitude)

g ) of
the maximizing z.
• zLOM is the maximum (in terms of magnitude) of
the maximizing z.

•B
Because off their
th i obvious
b i bi
bias, zSOM and
d zLOM are
not used as often as the other three
defuzzification methods.
methods
Defuzzification
Example
 Example: Single-input Single-output
Mamdani Fuzzy Model
• An example
p of a single-input
g p single-output
g p
Mamdani fuzzy model with three rules can be
expressed as
▫ If X is small then Y is small.
▫ If X is medium then Y is medium.
▫ If X is large then Y is large
large.
• Figure (next slide) plots the membership
functions of input X and output Y Y.
• With max-min composition and centroid
defuzzification,, we can find the overall input-
p
output curve, as shown in the Figure.
 Single-input single-output Mamdani
Fuzzy Model
Membership Grades

small medium large

1

0.5 10

8
0
-10 -5 0 5 10
6
X

Y
es
ship Grade

4
small medium large
1
2

05
0.5 0
Members

-10 -5 0 5 10
X

0
0 2 4 6 8 10
Y
a) MFs of the input and output
b) Overall input-output curve
 Example: Two-input Single-output
Mamdani Fuzzy Model
• An example of a two-input single-output
Mamdani fuzzy model with four rules can be
expressed as
▫ If X is small and Y is small then Z is negative large
large.
▫ If X is small and Y is large then Z is negative small.
▫ If X is large and Y is small then Z is positive small.
▫ If X is
i llarge and
d Y iis llarge th
then Z iis positive
iti llarge.
• Figure (next slide) plots the membership
functions of input X and Y and output Z.
• With max-min composition and centroid
defuzzification, we can find the overall input-
output
t t surface,
f as shown
h in
i the
th Figure.
Fi
 Two-input Single-output Mamdani
Fuzzy Model
1 small large
0.5
0
-5 0 5
Grades

X
mbership G

1 smallll l
large 2

Z
0.5 -2

0
-5 0 5 5
Mem

Y 0
0

large small small large Y -5 -5 X

1 negative negative positive positive

05
0.5
0
-5 0 5
Z
a) MFs of the inp
inputs
ts and output
o tp t
b) Overall input-output curve
 Operations in Mamdani Fuzzy Model

• To completely specify the operation of a Mamdani

fuzzy inference system,
system we need to assign a
function for each of the following operators:
▫ AND operator (usually T-norm) for calculating the
firing strength of a rule with AND
AND'ed ed antecedents.
antecedents
▫ OR operator (usually T-conorm) for calculating the
firing strength of a rule with OR'ed antecedents.
▫ Implication operator (usually T T-norm)
norm) for calculating
qualified consequent MFs based on given firing
strength.
▫ Aggregate operator (usually T-conorm)
T conorm) for
aggregating qualified consequent MFs to generate an
overall output MF.
▫ Defuzzification operator
p for transformingg an output
p
MF to a crisp single output value.
 Yet more on Mamdani model

• One such examplep is to use p

product for the
implication operator and Point-wise summation
(sum) for the aggregation operator (Note that
sum is not even a T-conorm operator !).
)
• Advantage:
▫ Th
The Final
Fi l crisp
i output
t t via
i centroid
t id defuzzification
d f ifi ti
is equal to the weighted average of the centroids
of co
o consequent
seque MFs. s
 Theorem:
Computation shortcut for Mamdani FIS
Defuzzification can be
computationally
t ti ll expensive.
i What
Wh t is
i
the center of area of this fuzzy set?

Theorem 4.1: If we use the product T‐norm, and the 
summation S‐norm (which is not really an S‐norm), and 
( h h ll )
centroid defuzzification, then the crisp output z is: 

z
 wa zi i i wi = firing strength (input MF value)
wa i i
ai = consequent MF area
zi = consequent MF centroid
ai and zi  can be calculated ahead of time!
 Example
We examine a simple two-input one-output problem that
includes three rules:

Rule: 1 IF x is A3 OR y is B1 THEN z is C1
Rule: 2 IF x is A2 AND y is B2 THEN z is C2
Rule: 3 IF x is A1 THEN z is C3

Real-life example for these kinds of rules:

Rule: 1 IF project_funding is adequate OR project_staffing is small THEN

risk is low
Rule: 2 IF project_funding is marginal AND project_staffing is large THEN
risk
i k iis normall
Rule: 3 IF project_funding is inadequate THEN
risk is high
 Step 1: Fuzzification
• The first step is to take the crisp inputs, x1 and y1 (project funding
and project staffing), and determine the degree to which these
inputs belong to each of the appropriate fuzzy sets.

34
 Step 2: Rule Evaluation
• The second step is to take the fuzzified inputs,
(x=A1) = 0.5,, (x=A2) = 0.2,, (y=B1) = 0.1 and (y=B2) = 0.7,,
and apply them to the antecedents of the fuzzy rules.
• If a given fuzzy rule has multiple antecedents, the fuzzy operator
((AND or OR)) is used to obtain a single g number that represents
p the
result of the antecedent evaluation.

j
RECALL: To evaluate the disjunction of the rule antecedents,, we use
the OR fuzzy operation. Typically, fuzzy expert systems make use of
the classical fuzzy operation union:
AB(x) = max [A(x), B(x)]
Similarly, in order to evaluate the conjunction of the rule antecedents,
we apply the AND fuzzy operation intersection:
AB(x) = min [A(x), B(x)]
 Step 2: Rule Evaluation
1 1 1
A3 B1 C1 C2 C3
0.1 OR 0.1
0.0
(max)
0 x1 X 0 y1 Y 0 Z
Rule 1: IF x is A3 (0
(0.0)
0) OR y is B1 (0
(0.1)
1) THEN z is C1 (0
(0.1)
1)
1 1 1
0.7
C1 C2 C3
A2 0.2 B2 AND 0.2
(min)
0 x1 X 0 y1 Y 0 Z
Rule 2: IF x is A2 (0.2) AND y is B2 (0.7) THEN z is C2 (0.2)
1 1
A1 0.5 0.5 C1 C2 C3

0 x11 X 0 Z
Rule 3: IF x is A1 (0.5) THEN z is C3 (0.5)
 Step 2: Rule Evaluation
• Now the result of the antecedent evaluation Degree of
can be applied to the membership function of Membership
1.0
10
the
th consequent. t
• The most common method is to cut the consequent C2
membership function at the level of the antecedent
truth. This method is called clipping (alpha-cut).
▫ Since the top of the membership function is sliced, 0.2
the clipped fuzzy set loses some information.
▫ However, clipping is still often preferred because it 0.0
Z
involves less complex and faster mathematics, and clipping
pp g
generates an aggregated d output surface
f that
h iis easier
i
to defuzzify. Degree of
Membership
• While clipping is a frequently used method, scaling 1.0
offers a better approach for preserving the original
shape of the fuzzy set. C2
▫ The original membership function of the rule
consequent is adjusted by multiplying all its
membership degrees by the truth value of the rule 0.2
antecedent.
▫ This method, which generally loses less information, 0.0
can be very useful in fuzzy expert systems. scaling Z
 Step 3: Aggregation of the Rule
Outputs
• Aggregation is the process of unification of the outputs of all
rules.
• We take the membership functions of all rule consequents
previously clipped or scaled and combine them into a single
fuzzy set.
• The input of the aggregation process is the list of clipped or
scaled consequent membership functions, and the output is one
fuzzy set for each output variable.

1 1 1
C1 C2 C3
0.5 0.5
0.2 0.2
0.1 0.1
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) 
 Step 4: Defuzzification
• The last step in the fuzzy inference process is defuzzification.
• Fuzziness helps us to evaluate the rules, but the final output of a
fuzzy system has to be a crisp number.
• The input for the defuzzification process is the aggregate output
fuzzy set and the output is a single number.
• There are several defuzzification methods, but probably the most
popular one is the centroid technique. It finds the point where a
vertical line would slice the aggregate set into two equal masses.
Mathematically this centre of gravity (COG) can be expressed as:
b
  A x  x dx
COG  a
b
  A x  dx
a
 Step 4: Defuzzification
• Centroid defuzzification method finds a point representing the centre
of gravity of the aggregated fuzzy set A, on the interval [a, b ].
• A reasonable estimate can be obtained by calculating it over a
sample of points.
Degree of
Membership
1.0
0.8
0.6

0.4
0.2
0.0
0 10 20 30 40 50 60 70 80 90 100
67.4 Z

(0  10  20)  0.1  (30  40  50 0  60)  0.2  (70 0  80  90  100)  0.5

COG   67.4
0 .1  0 .1  0 .1  0 .2  0 .2  0 .2  0 .2  0 .5  0 .5  0 .5  0 .5
 Sugeno
g Fuzzy
y Inference
• Mamdani-style inference, as we have just seen, requires
us to find the centroid of a two-dimensional shape by
integrating across a continuously varying function. In
general,, this process
g p is not computationally
p y efficient.
• Michio Sugeno suggested to use a single spike, a
singleton, as the membership function of the rule
consequent.
consequent
• A singleton, or more precisely a fuzzy singleton, is a
fuzzyy set with a membership p function that is unity
y at a
single particular point on the universe of discourse and
zero everywhere else.
 Sugeno
g Fuzzyy Model

• Also known as the TSK fuzzy y model (p p

(proposed
by Takagi, Sugeno, and Kang)
• For developing a systematic approach to
generating fuzzy rules from a given input-output
data set
• A typical
t i l ffuzzy rule l in
i aSSugeno ffuzzy model:d l
if x is A and y is B then z = f (x, y)
• A and B: fuzzy sets
• z = f(x, y): a crisp function (usually polynomial in
the input variables x and y)
 Sugeno
g Fuzzyy Model
• Sugeno-style fuzzy inference is very similar to the Mamdani method.
• Sugeno changed only a rule consequent: instead of a fuzzy set, he used a
mathematical function of the input variable.
• The format of the Sugeno-style fuzzy rule is
IF x is A AND y is B THEN z is f(x,
f(x y)
where:
▫ x, y and z are linguistic variables;
▫ A and B are fuzzy sets on universe of discourses X and Y, respectively;
▫ f (x, y) is a mathematical function.

• The most commonly used zero-order Sugeno fuzzy model applies fuzzy
rules in the following form:
IF x is A AND y is B THEN z is k
• where k is a constant.
• In this case,, the output
p of each fuzzyy rule is constant and all consequent
q
membership functions are represented by singleton spikes.
 So …

• First-order Sugeno
g ( , y) is a
y model: f(x,
fuzzy
first-order polynomial
• zero-order Sugeno fuzzy model: f is a
constant
▫ a special case of the Mamdani fuzzy inference
system in which each rule's consequent is
system,
specified by a fuzzy singleton;
▫ or a special case of the Tsukamoto fuzzy model
(to be introduced next) in which each rule's
consequent is specified by an MF of a step
f
function
ti center
t att the
th constant
t t
 Sugeno
g Fuzzy
y Models

The output is a weighted average:

z
  Ai , Bk ( x, y ) f m ( i , k ) ( x, y )
Double summation 
over all i (x MFs) and 
over all i (x MFs) and
 Ai , Bk ( x, y ) all k (y MFs)


 w f ( x, y )
i i SSummation over all i
ti ll i
w i
(fuzzy rules)

where wi is the firing strength of the i-th output

Fuzzy Reasoning Procedure For a First-
Order Sugeno Fuzzy Model
 Example: Fuzzy and nonfuzzy rule
sets
• An example
p of a single-input
g p Sugeno
g y
fuzzy
model:
▫ If X is small then Y = 0.1X + 6.4.
▫ If X is medium then Y = -0.5X + 4.
▫ If X is large then Y = X - 2.
 Example: Fuzzy and nonfuzzy rule
set a comparison
set-a
• If "small,"
, "medium,", and "large"
g are nonfuzzy
y
sets with membership functions shown in figure
(a), then the overall input-output curve is
piecewise linear, as shown in figure
f (b):
( )

(a) Crisp Antecedent MFs (b) Crisp I/O Curve

p Grades

small medium large 8

1
6
embership

4
Y

0.5
2
0 0
Me

-10 0 10 -10 0 10
X X
 Example: Fuzzy and nonfuzzy rule
sets
• If we have smooth membership p functions
[figure (c)] instead, the overall input-output curve
[figure (d)] becomes a smoother one:

(c) Fuzzy Antecedent MFs (d) Fuzzy I/O Curve

bership Grrades

small medium large

g 8
1
6
4

Y
0.5
2
Memb

0 0
-10 0 10 -10 0 10
X X
 Example: Two-input single-output
Sugeno fuzzy model
• An example
p of a two-input
p single-output
g p Sugeno
g
fuzzy model with four rules:
▫ If X is small and Y is small then z = -x + y + 1.
▫ If X is small and Y is large then z = -y + 3.
▫ If X is large and Y is small then z = -x + 3.
▫ If X is large and Y is large then z = x + y + 2
2.
 Example: Two-input single-output
Sugeno fuzzy model
Grades

Small Large
1
Membership

0.5
10

0
-5 0 5
M

5
X

Z
p Grades

Small Large 0
1
5
Membership

5
0.5 0
0
Y -5 -5
0 X
-5 0 5
M

Y
a) MFs of the inputs and output
b) Overall input-output curve
• Th
The surface
f i composed
is d off ffour planes,
l each
h off which
hi h iis
specified by the output equation of a fuzzy rule.
 Another Example
1 1 1
A3 B1
0.1 OR 0.1
0.0
(max)
0 x1x1 X 0 y1y1 Y 0 k1k1 Z
Rule
Rule1:
1: IF x is A3 (0.0)
(0(0
(0.0)
0)0) OR
OR yy is
is B1
B1 (0.1)
(0 1) THENTHEN z is
z is
k1k1
(0.1)
(0(0.1)
(0
1)1)
1 1 1
0.7
A2 0.2 B2 AND 0.2
(min)
0 x1x1 X 0 y1
y1 Y 0 k2k2 Z
Rule
Rule 2:
2: IF
IF x is A2 (0.2)
(0.2) AND yyisisB2
B2(0.7)
(0.7) THENTHENz isz k2
is k2 (0.2)
(0.2)
1 1
A1 0.5 0.5

0 x1x1
11 X 0 k3
k3 Z
Rule 3: IF x is A1 (0.5)
Rule (0.5) THEN THENz isz k3
is k3 (0.5)
(0.5)
 Another Example

COG becomes Weighted Average (WA)

(k1)  k1  (k 2)  k 2  (k 3)  k 3 0.1 20  0.2  50  0.5  80

WA    65
(k1)  (k 2)  (k 3) 0.1  0.2  0.5
 More on Sugeno
g model

• Unlike the Mamdani fuzzy y model,, the Sugeno

g
fuzzy model cannot follow the compositional rule
of inference strictly in its fuzzy reasoning
mechanism
• Without the time-consuming and mathematically
intractable defuzzification operation,
operation the Sugeno
fuzzy model is by far the most popular candidate
for sample data-based
data based fuzzy modeling (we will
see an application in ANFIS)
 How to make a decision on which
method to apply  Mamdani or Sugeno?
 Mamdani method is widely accepted for capturing
expert knowledge. It allows us to describe the
expertise in more intuitive, more human-like
manner However,
manner. However Mamdani-type
Mamdani type fuzzy inference
entails a substantial computational burden.
 On the other hand
hand, Sugeno method is
computationally effective and works well with
optimisation and adaptive techniques, which
makes it very attractive in adaptive problems,
particularly for dynamic nonlinear systems.
 Tsukamoto Fuzzyy Model
• In the Tsukamoto fuzzy model, the consequent
of each fuzzy if-then
if then rule is represented by a
fuzzy set with a monotonical MF, as shown in
the following figure:
 Tsukamoto Fuzzyy Model

• The inferred output

p of each rule is defined as a
crisp value induced by the rule's firing strength
• The overall output is taken as the weighted
average of each rule's output
▫ Thus avoids the time-consuming process of
defuzzification
• This model is not used often since it is not as
transparent as either the Mamdani or Sugeno
fuzzy models
 Example: Single-input Tsukamoto
fuzzy model
• An example
p of a single-input
g p Tsukamoto fuzzy
y
model:
▫ IT X is small then Y is C1
▫ IT X is medium then Y is C 2

▫ IT X is large then Y is C
3,
Example: Single-input Tsukamoto
fuzzy model
 Example: Single-input Tsukamoto
fuzzy model
• The overall input-output
p p curve is equal
q to:

▫ Where fi is the output of each rule induced by the

firing strength wi and MF for Ci
• If we plot each rule's output fi as a function of x,
we e obtain fig
figure
re (c)
(c), which
hich is not q
quite
ite ob
obvious
io s
from the original rule-base and MF plots.
• The output is always crisp even when the inputs
are fuzzy.
 FIS Optimisation

• Why
y to optimise?
p
▫ trial and error tuning is laborious
▫ it can be impossibly complicated if number of
input parameters are large (100’s or more)
▫ far too many parameters to tune: number of rules,
membership functions
functions, rule consequents
▫ arbitrariness of FIS is eliminated
▫ if not optimised,
p , the FIS performance
p is not
optimum
 FIS Optimisation
• Optimisation methods:
▫ Adaptive Neuro-Fuzzy
Neuro Fuzzy systems
multilayer perceptron neural networks (ANFIS, FuNeI)
RBFN  we will see these systems later on
▫ evolutionary techniques (genetic algorithms)
▫ clustering methods (cluster analysis, Hard C-means, Fuzzy C-
means)
▫ ...
 FIS Optimisation

• Optimisation
p can be viewed as knowledge g
discovery
• Optimisation = Learning
• Resulting model describes input-output
relationships qualitatively and quantitatively.
• Easy
E validation
lid ti and d verification
ifi ti off FIS
 Tuning
g fuzzyy systems
y
1. Review model input and output variables, and if
required redefine their ranges.
2. Review the fuzzy sets, and if required define
additional sets on the universe of discourse.
3. Provide sufficient overlap between neighbouring
sets. It is suggested that triangle-to-triangle and
trapezoid to triangle fuzzy sets should overlap
trapezoid-to-triangle
between 25% to 50% of their bases.
 Tuning
g fuzzyy systems
y

4. Review the existing g rules,, and if required

q add new
rules to the rule base.
5. Examine the rule-base for opportunities to write
hedge rules to capture the pathological behaviour
of the system.
6. Adjust the rule execution weights. Most fuzzy
l
logic tools
l allow
ll controll off the
h importance off rules
l
by changing a weight multiplier (we did not see this
one in this course !). )
7. Revise shapes of the fuzzy sets. In most cases,
fuzzy systems are highly tolerant of a shape
approximation.
approximation
 Other Considerations

• certain common issues concerning

g all these
three fuzzy inference systems
▫ how to partition an input space
▫ how to construct a fuzzy inference system for a
particular application
 Input Space Partitioning
g
• The antecedent of a fuzzy rule defines a local fuzzy
region,
region
• The consequent describes the behavior within the
region via various constituents
▫ constituent
tit t can b
be a consequentt MF (Mamdani
(M d i andd
Tsukamoto fuzzy models),
▫ a constant value (zero-order Sugeno model),
▫ or a linear
li equation
ti (fi
(first-order
t d Sugeno
S model).
d l)
• Different consequent constituents result in different
fuzzy inference systems,
• but
b t their
th i antecedents
t d t are alwaysl th
the same.
Therefore, methods of partitioning input spaces to
form the antecedents of fuzzy rules is applicable to
allll th
three ttypes off ffuzzy inference
i f systems
t
 Grid partition

p
input 2
Grid Partition
The curse of
dimensionality:
If we have n inputs and m
ut 1

FMs per input,

input then we
inpu

have mn if-then rules.

p 6 inputs
For example, p and
5 memberships per input
 56 = 15,625 rules!
 Tree partition

p
input 2
Tree Partition
• Reduces the number of
rules
• Requires more MFs per
ut 1
inpu

i
inputt
• MFs do not have clear
linguistic meanings
 Scatter partition

p
input 2
Scatter Partition
• In many systems,
extremes occur rarely
• Number of active rules
ut 1
inpu

d
depends
d on iinputt
values
 Input Space Partitioning
g

• A more flexible p y , when MFs are

partition style,
defined on certain transformations of the input
Variables:
 Fuzzy
y Modeling
g

• In g
general,, we design
g a fuzzy
y inference system
y
based on the past known behavior of a target
system.
• now consider how we might construct a fuzzy
inference system for a specific application

• Fuzzy Modeling: The standard method for

constructing a fuzzy inference system
 Fuzzy
y Modeling
g

• fuzzy g features:
g has the following
y modeling
▫ Takes advantage of domain knowledge
 Making it easy to incorporate human expertise about
th target
the t t system
t directly
di tl into
i t the
th modeling
d li process
▫ The use of numerical data also plays an
important role in fuzzy modeling
 When the input-output data of a target system is
available, conventional system identification
t h i
techniques can b
be used
d ffor ffuzzy modeling
d li
 Fuzzy
y Modeling
g

• fuzzy
y modelingg can be p g ,
pursued in two stages,
which are not totally disjoint:
1. the identification of the surface structure
2. the identification of deep structure
 Fuzzy
y Modeling
g

• The identification of the surface structure

includes the following tasks:
1. Select relevant input and output variables
2. Choose a specific type of fuzzy inference
system.
3 Determine the number of linguistic terms
3.
associated with each input and output variables.
((For a Sugeno
g model, determine the order of
consequent equations.)
4. Design a collection of fuzzy if-then rules.
 Fuzzy
y Modeling
g

• After the first stage

g of fuzzyy modeling,
g, we obtain
a rule-base that can describe the behavior of the
target system by means of linguistic terms.
• The meaning of these linguistic terms is
determined in the second stage, the
identification of deep structure,
structure which
determines the MFs of each linguistic term (and
the coefficients of each rule's
rule s output polynomial
if a Sugeno fuzzy model is used).
 Fuzzy
y Modeling
g

• The identification of deep

p structure includes the
following tasks:
1. Choose an appropriate family of parameterized
MFs
2. Interview human experts familiar with the target
systems to determine the parameters of the MFs
used in the rule-base.
3. Refine the p
parameters of the MFs using g
regression and optimization techniques.
 Reading
g

• JJ-S
S R Jang and C T Sun,
C-T Sun Neuro Fuzzy
Neuro-Fuzzy
and Soft Computing, Prentice Hall, 1997
(Chapter 4)
4).