Fuzzification

By considering quantities as uncertain:

Imprecision
Ambiguity
Vagueness

Membership Assignment
Many ways to do it!

Intuition
Using our intelligence and understanding.
Intuition involves contextual and semantic knowledge
about an issue. It can also involve linguistic truth-values
about the knowledge.

Note: they are overlapping.

Inference
Using knowledge to perform deductive reasoning.
Example:
Let U be a universe of triangles.
U = { (A B C) | A > B > C > 0 A + B + C =
180 }
We can define the following 5 types of triangles:
I: Approximate isosceles triangle
R: Approximate right triangle
IR: Approximate isosceles and right triangle
E: Approximate equilateral triangle
T: Other triangles

Inference
I(A B C) = 1 1/60 min(A B,B C)
R(A B C) = 1 1/90 |A - 90|
IR = I R
IR(A B C) = min [ I(A B C), R(A B C) ]
= 1 max[1/60 min(A B,B C),1/90 |A 90|]
E(A B C) = 1 1/180 (A C)
T = (I R E) = I R E
= min{1 - I,1 -1 R,1 - E}
= 1/180 min{3(A B),3(B C),2| A - 90|,A C}

Rank Ordering
Assessing preference by a single individual, a pole, a
committee, and other opinion methods can be used to
assign membership values to a fuzzy variable.
Preference is determined by pair wise comparisons
which determine the order of memberships.

Angular Fuzzy Sets

Angular Fuzzy sets are defined on a universe of angles
with 2 as cycle.
The linguistic values vary with and their memberships
are
t() = t tan()

Angular Fuzzy sets are useful for situations:

Having a natural basis in polar coordinates, or the
variable is cyclic.

Neural Networks
We have the data sets for inputs and outputs, the
relationship between I/O may be highly nonlinear or not
known.
We can classify them into different fuzzy classes.

Training
R1 Testing
R2
R3

Then, the output may not only be 0 or 1!

Neural Networks
R1

0.2

R2

0.7

R3

0.1

memberships

Once the neural network is trained and tested, it can be

used to find the membership of any other data points in
the fuzzy classes (# of outputs)

Genetic Algorithms
Crossover
Mutation

random selection

Reproduction
Chromosomes
Fitness Function
Stop (terminate conditions)
Converge
Reach the #limit

Inductive Reasoning
Deriving a general consensus from the particular (from
specific to generic)
The induction is performed by the entropy minimization
principle, which clusters most optimally the parameters
corresponding to the output classes.
The method can be useful for complete systems where
the data are abundant and static.
The intent of induction is to discover a law having
objective validity and universal application.

Inductive Reasoning
Particular General
Maximize entropy
Computing mean probability
Minimize entropy
The entropy is the expected value of information.
Many entropy definitions!
A survey paper

Inductive Reasoning
One example:
S(x) = p(x) Sp(x) + q(x) Sq(x)
Sp(x) = -[P1(x) ln(P1(x)) + P2(x) ln(P2(x))]
Sq(x) = -[q1(x) ln(q1(x)) + q2(x) ln(q2(x))]
Pk ( x )

nk ( x ) 1
n( x ) 1

qk ( x )

Nk ( x ) 1
N( x ) 1

P( x )

n( x )
n

Inductive Reasoning
Where:
nk(x): # of class k samples in [x1,x1+x]
n(x): Total # of samples in [x1,x1+x]
Nk(x): # of class k samples in [x1+x,x2]
N(x): Total # of classes in [x1+x,x2]
n = Total # of samples in [x1,x2]
Move x in [x1,x2], and compute the entropy for each x to
find the maximum / minimum entropy.
Note: there are many approaches to compute entropy.

Defuzzification
(Fuzzy-To-Crisp conversions)
Using fuzzy to reason, to model
Using crisp to act
Like analog digital analog
Defuzzification is the process: round it off to the nearest
vertex.

Fuzzy set (collection of membership values).

Defuzzification
(Fuzzy-To-Crisp conversions)
A vector of values reduce to a single scalar quantity:
most typical or representative value.
Fuzzification Analysis Defuzzification Action
-cuts for fuzzy sets
(-cuts, some books)
A , 0 < < 1
A = {x | A(x) > }
Note: A is a crisp set derived from the original fuzzy set.
[0,1] can have an infinite number of values. Therefore,
there can be infinite number of -cut sets.

Defuzzification
(Fuzzy-To-Crisp conversions)
Example:
A = {1/a + 0.9/b + 0.6/c + 0.3/d + 0.01/e + 0/f}
A1 = {a} or A1 == {1/a + 0/b + 0/c + 0/d + 0/e + 0/f}
A0.9 = {a,b}
A0.3 = {a,b,c,d}
A0.6 = {a,b,c}
A0.01 = {a,b,c,d,e}
A0 = x = {a,b,c,d,e,f}

Defuzzification
(Fuzzy-To-Crisp conversions)
-cut re-scales the memberships to 1 or 0
The properties of -cut:
1. (A B) = A B
2. (A B) = A B
3. (A) (A) except for x = 0.5
4. A A < and 0 < < 1
A0 = X
Core = A1
Support = A0+
Boundaries = [A0 + A1]

Defuzzification
(Fuzzy-To-Crisp conversions)

0.6
0.3

-cuts for fuzzy relations

R=

0.8

0.1

0.2

0.8

0.4

0.9

0.4

0.1

0.5

0.2

0.9

0.5

Defuzzification
(Fuzzy-To-Crisp conversions)
We can define -cut for relations similar to the one for sets
R = {(x y) | R(x y) > }
1

R1
=

R0.9
=

1
1

R0 = E

Defuzzification
(Fuzzy-To-Crisp conversions)
-cuts on relations have the following properties:
(R S) = R S
(R S) = R S
(R) (R)
R < R and 0 1

Defuzzification Methods
fuzzy set a single scalar quantity
fuzzy quantity precise quantity

O1

O2

O = O1 O2

Defuzzification Methods
A fuzzy output can have many output parts
k

C Ci
i1

Many methods can be used for defuzzification.

They are listed in the following slides

Defuzzification Methods
Max-membership principle

c(Z*) c(z) z Z
z*

Centroid principle

z
*

c ( z ) zdz

c ( z )dz
z*

Note: It relates to moments.

Defuzzification Methods
Weighted average method

(Only valid for symmetrical output membership functions)

z
*

c ( z )

c ( z )
~

.9

.5

Mean-max membership

(middle-of-maxima method)

z* a b 2
0

z* b

Defuzzification Methods
Example:
A railroad company intends to lay a new rail line in a particular part of a
county. The whole area through which the new line is passing must be
purchased for right-of-way considerations. It is surveyed in three stretches,
and the data are collected for analysis. The surveyed data for the road are
given by the sets B~ 1,B~ 2 and B~ 3 , where the sets are defined on the universe of
right-of-way widths, in meters. For the railroad to purchase the land, it must
have an assessment of the amount of land to be bought. The three surveys
on the right-of-way width are ambiguous , however, because some of the land
along the proposed railway route is already public domain and will not need
to be purchased. Additionally, the original surveys are so old (circa 1860) that
some ambiguity exists on the boundaries and public right-of-way for old utility
B, ,shown
B and Bin the figures
lines and old roads. The three fuzzy sets
~1 ~ 2
~3
below, represent the uncertainty in each survey as to the membership of the
right-of-way width, in meters, in privately owned land.
We now want to aggregate these three survey results to find the single most
nearly representative right-of-way width (z) to allow the railroad to make its
initial estimate

Defuzzification Methods

Defuzzification Methods
Centroid method:
z

( z ) zdz

( z )dz
B
~

B
~

3.6

5 .5
6
7
8

z 3
zdz 4 (.5) zdz 5.5 z 5 zdz 6 zdz 7 8 z zdz
2

0 (.3z ) zdz 1 (.3 z )dz 3.6

.3z dz

4.9meters

3 .6

5 .5
6
7
8

z 3
(.3)dz
dz (.5)dz ( z 5)dz dz (8 z )dz

3 .6
4
5 .5
6
7
2

Defuzzification Methods
Weighted-Average Method:
z

.3 2.5 .5 5 1 6.5

.3 .5 1

5.41meters

Mean-Max Method: (6 7) / 2 6.5meters

Defuzzification Methods

Defuzzification Methods

Defuzzification Methods
According to the centroid method,
z

( z ) zdz

( z )dz
B

B
~

1.5

(.67 z ) zdz 2 .67 z zdz z 1 zdz

1.8

1.5

1.8

2.33

3 z zdz

2.33 .5 z .5 zdz 3 2.5 .5 z zdz

.
67
z
dz

.
67
z
dz

(
z

1
)
dz

z
dz

0
1.5
1.8
2

5
3

2.33 .5 z .5 dz 3 2.5 .5 z dz

1.5

2. 5

1.8

2.33

Defuzzification Methods
The centroid value obtained, z*, is shown in the figure
below:

Defuzzification Methods
According to the weighted average method:

11.5 1 2 1 3

2.25
111

Defuzzification Methods
Center of sums Method
Faster than any defuzzification method
Involves algebraic sum of individual output fuzzy sets,
instead of their union
Drawback: intersecting areas are added twice.

n
k 1

k 1

C z dz
~k

C z dz
~k

It is similar to the weighted average method, but the

weights are the areas, instead of individual membership
values.

Defuzzification Methods
z1 = 4

z2 = 8

z1s1 z 2 s2 4 3 8 2 28

5.6
s1 s2
3 2
5
or

z 0.25 z dz 0.5 zdz 0.25 z 2 zdz

2

0.5 z 3 zdz 0.5 z 5 zdz

8

10

z3
0.25
3
z2

z2
0.5
2
0
2

z3
0.25
3
2

5 2 10
z3
z 0.5
2 8
3
6

10
8

28

8
6

Defuzzification Methods

0.25 z dz 0.5dz 0.25 z 2 dz

2

0.5 z 3 dz 0.5 z 5 dz
8

10

z2
0.25
2
5

z2
0.5 z 0.25
2
0
2
2

z 28 5 5.6
*

z2
z 0.5
2
6
6

10
8

5z

10
8

Defuzzification Methods
Center of Sums Method

Defuzzification Methods
Using Center of sums:
S1 = 0.5 * 0.5(8+4) = 3
S2 = 0.5 * 1 * 4 = 2
Center of the largest area: if output has at least two
convex sub-regions
z

z zdz
z dz
Cm
~

Cm
~

Where Cm is the convex sub-region that has the largest

area making up Ck. (see figure)

Defuzzification Methods
Center of sums method

Defuzzification Methods
First (or Last) of Maxima method
This method uses the overall output or union of all
individual output fuzzy sets to determine the smallest
value of the domain with maximized membership
degree in each output set. The equations for z* are as
follows:
First, the largest height in the union is determined:

hgt

Ck
~

sup Ck z
zZ

Then the first of the maxima is found:

z * inf z Z | Ck z hgt Ck
~
~
zZ

Defuzzification Methods
First (or last) of Maxima method
An alternative to this method is called the last of maxima,
and it is given by:

z sup z Z | Ck z hgt
~
zZ

Supremum (Sup): the least upper bound

Infimum (Inf): the greatest lower bound

Ck
~

Defuzzification Methods
Continuation of the railroad example, the results of
the different methods can be shown graphically as
follows:

Defuzzification Methods

Fuzzy Arithmetic, Numbers, Vectors

The Extension Principle

f(x)

How to find y if x is fuzzy, f is fuzzy or both are fuzzy

Crisp function, Mapping and Relation

For a set A defined on universe X, its image, set B
on the universe Y is found from the mapping
B = f(A) = { y | x A, y = f(x)
B is defined by its characteristic value
XB(y) = Xf(A)(y) = XA(x)
y = f(x)

Note: means max

Crisp function, Mapping and Relation

Example:
A = {0/-2 +0/-1 +1/0 +1/1 +0/2}
X = {-2,-1,0,0,1,2}
If y = |4x| + 2
Y = {2,6,10}
XB(2) = {XA(0)} = 1
XB(6) = {XA(-1),XA(1)} = {0,1} = 1
XB(10) = {XA(-2),XA(2)} = {0,0} = 0
B = {1/2 + 1/6 + 0/10} or B = {2,6}

Crisp function, Mapping and Relation

We may consider the universe
X = {-2,-1,0,1,2} and universe Y = {0,1,2,,9,10}
The relation describing this mapping
0

-2
-1

R=

0
1
2

1
0
1

1
1

Crisp function, Mapping and Relation

If A = {0/-2 + 0/-1 + 1/0 + 1/1 + 0/2}
Then, B = A R
XB(y) = (XA(A) XR(x y))
xX

1 for y = 2,6
=
0 otherwise
or
B = {0/0 + 0/1 + 1/2 + 0/3 + 0/4 + 0/5 + 1/6 + 0/7 +
0/8 + 0/9 + 0/10}

Function of Fuzzy Sets Extension Principle

B = f(A)
If A is fuzzy, B is also fuzzy.
B(y) = A(x)
f(x) = y

Fuzzy Vectors:

a A x1 ,... A xn A xi
i 1,2,3,..., n

b B y1 ,... B y n B y j
j 1,2,3,..., m
B A R
or

b aR

Function of Fuzzy Sets Extension Principle

General case

f : P x1 x2 ... xn P Y
Let A1,A2,An be defined on X1,X2,,Xn
Then B = f(A1,A2,,An)
B y max min A1 y1 , A2 y2 ,..., An yn
y f x1 , x2 ,..., xn

This is called Zadehs extension principle.

Fuzzy Transform (Mapping)

Extending fuzziness in an input set to an output set.
I: fuzzy

f:A B

O: fuzzy f: crisp

If x X thenB f x is called fuzzy mapping, ~ indicates

~
~
fuzzy.
It can be described as a fuzzy relation.

R
~

y1

y2

ym

x1

r11

r12

r1m

x2

r21

r22

r2m

xn

rn1

rn2

rnm

Fuzzy Transform (Mapping)

Bi f xi

Bi yi rij
or

bi ri1 , ri 2 ,..., rim

More general:

B f A
~

B y
~

x xy
A
~

x X

R
~

or

b aR
b j max min ai , rij
~

b j is the jth element of the fuzzy image B

~

Fuzzy Transform (Mapping)

1
0.8

R=

0.8 0.2 0.1

1

0.2 0.8

0.8 0.2 0.1

1

0.1 0.2 0.8

0

0.8 0.2
1

40
50
60

0.8

70

80

0.1 0.2 0.8

A 0.8 / 40 1 / 50 0.6 / 60 0.2 / 70 0 / 80

~

or
a 0.8,1,0.6,0.2,0
~

b a R 0.8,1,0.8,0.6,0.2
~

Practical Considerations
F:u v, u U and v V
A U
~

A 1 / u1 2 / u2 ... n / un
~

Then the extension principle is

f A f 1 / u1 2 / u 2 ... n / un
~

1 / f u1 2 / f u 2 ... n / f u n

It is a mapping called one-to-one.

Practical Considerations
Example:
u = {1,2,3}
v= f(u) = 2u 1
A = {0.6/1 + 1/2 + 0.8/3}
Then f(A) = {0.6/1 + 1/3 + 0.8/5}
If A U 1 U 2
~

Then

min 1 i , 2 j
f A
i U 1, j U 2
~
f i, j

Where 1(i) and 2(j) are the separable membership

projections of (I,j) from U1 U2, when (I,j) cannot be
determined.

Practical Considerations
Example:
U1 = U2 = {1,2,,10}
A = 2 = Approximately 2 = {0.6/1 + 1/2 + 0.8/3}
B = 6 = Approximately 6 = {0.8/5 + 1/6 + 0.7/7}
min 0.6,0.8 min 0.6,1 min 0.6,0.7

5
6
7

min 1,0.8 min 1,1 min 1,7 min 0.8,0.8

10
12
14
15

min 0.8,1 min 0.8,0.7

18
21

0.6 / 5 0.6 / 6 0.6 / 7 0.8 / 10 1 / 12

0.7 / 14 0.8 / 15 0.8 / 18 0.7 / 2

This mapping is unique. If not, we have to perform

maximum operation!

Practical Considerations
A u1 , u2 max min 1 u1 , 2 u2
u f u1 ,u 2

Example:
A = {0.2/1 + 1/2 + 0.7/4}
B = {0.5/1 + 1/2}

f A, B A B

min 0.2,0.5 max min 0.2,1 , min 0.5,1

1
2

max min 0.7,0.5 , min 1,1 min 0.7,1

4
8
0.2 / 1 0.5 / 2 1 / 4 0.7 / 8

Practical Considerations
Example:
We want to map ordered pairs from the input universe X1
= {a,b} and X2 = {1,2,3} to an output universe Y={x,y,z},
for instance
1

Crisp

if A 0.6 / a 1 / b
~

B 0.2 / 1 0.8 / 2 0.4 / 3

~

C f A, B
~

f is described by R.

Practical Considerations
C x max min 0.2,0.6 , min 0.4,0.6 , min 0.2,1 0.4
~

C y min 1,0.8 0.8

~

C z max min 0.6,0.8 , min 1,0.4 0.6

~

C 0.4 / x 0.8 / y 0.6 / z

~

Note :
0.6
min 0.6,0.2 , min 0.6,0.8 , min 0.6,0.4

0.2,0.8,0.4

1
min 1,0.2 , min 1,0.8 , min 1,0.4

Practical Considerations
Consider:

x cos w t
~

x x
~

w
w

x cos wt

For t = 0, all values of w~ map into a single point.

wt 0 x 1
~

x x
~

1 if x = 1
0 otherwise

If t 0, but small, the supp w (support of w)

~
~
The membership value of in this interval
determined in a one-to-one mapping

is