Fuzzy Membership Function Design Using Information Theory Measures and Genetic Algorithms

Outlines
Introduction Problem Statement Proposed Approach Results Conclusion

Masoud Makrehchi

Outlines
Introduction Problem Statement Proposed Approach Results Conclusion

introduction . fuzzy membership function

Fuzzy systems
Perfect operation with fuzzy data Precise data from measurement and interfaces Need to have fuzzy data from precise data Conversion from precise to fuzzy (fuzzification)

Masoud Makrehchi

introduction . fuzzy membership function

introduction . fuzzy membership function

Fuzzification
A gateway to any fuzzy system applications
Masoud Makrehchi

Fuzzification precise world Fuzzy Information Processing Defuzzification fuzzy world

introduction . fuzzy membership function NS NM 1 Z PM PL

introduction . fuzzy membership function NS NM 1 Z PM PL

(NM,0.9)

-1 data/signal from real process

+1
Masoud Makrehchi

-1 data/signal from real process

+1

Normalization

Normalization

introduction . fuzzy membership function NS NM 1 Z PM PL

introduction . fuzzy membership function NS NM 1 Z PM

(NS, 0.2) (NM,0.5)

PL

-1 data/signal from real process

+1
Masoud Makrehchi

-1 data/signal from real process

+1

Normalization

Normalization
10

introduction . fuzzy membership function

introduction . fuzzy membership function

Fuzzy membership function (FMF)

a critical issue in fuzzy information processing, fuzzy control, fuzzy pattern recognition,

Different types of fuzzy membership functions

Masoud Makrehchi

introduction . fuzzy membership function Boundary

introduction . fuzzy membership function

trapezoidal

triangular
1 Prototype

1 Core
Masoud Makrehchi

Support

UD

Support Boundary

UD

13

14

introduction . fuzzy membership function Boundary

introduction . fuzzy membership function

sigmoidal

bell-shaped
1 Prototype

1 Core
Masoud Makrehchi

Support

UD

Support Boundary

UD

introduction . fuzzy membership function

introduction . fuzzy membership function

singleton

Three parts of FMF

support: boundary:
Masoud Makrehchi
xi X : xi > 0

xi X : xi > 0

xi X : 0 < xi < 1

X1

x2

x3

UD

core/prototype:

xi X : xi = 1

17

18

introduction . fuzzy membership function

introduction . fuzzy membership function

Support or fuzzy partition

An essential part of any fuzzy membership function
Masoud Makrehchi Support of partition A

M(x)

information domain-UD (area of interest)

introduction . fuzzy membership function

introduction . fuzzy membership function

MB(x) B A E F D C MA(x) 0
Masoud Makrehchi

M(x)

information domain-UD (area of interest)

X0 X1 X2 X3

X (Support)

Xn-1 Xn

UD

21

22

introduction . design factors

introduction . design factors

FMF design factors

Support: the domain in which the FMF is defined domain of FMF or a partition of desired information and our interest in which fuzzy information is defined

FMF design factors

Shape: determining the boundaries and core/prototype and fuzzy behavior of FMF

Masoud Makrehchi

introduction . design factors

Outlines
Introduction Problem Statement Proposed Approach Results Conclusion

FMF design factors

Number: number of fuzzy partitions assigned to a Linguistic Variable, influencing the size of fuzzy rule base,

25

Masoud Makrehchi

26

problem statement

problem statement

Masoud Makrehchi

How can IT measures help in designing FMF? Which parameters can be optimized by IT measures?

Number
Estimating number of fuzzy partitions is a trade off with fuzzy rules, we can not estimate it independently It can be finalized during optimization of fuzzy rules The number of fuzzy rules is the bottleneck, not the number of fuzzy partitions

problem statement

problem statement

Shape
Shape of FMF is still a heuristic issue There is no proven relation between information domain and degree of fuzziness in that domain, completely related to intuition, expertise, and expert knowledge Learning from examples can be a solution

Support
we can just estimate informational parameters of FMF, not fuzzy issues Support is a part of our information in which an uncertainty is happening IT measures is suitable for estimating support of FMF, or fuzzy partitions
30

29

problem statement

Masoud Makrehchi

Outlines
Introduction Problem Statement Proposed Approach Results Conclusion

Masoud Makrehchi

Finding an optimum set of fuzzy partitions related to a given linguistic variable Optimum fuzzy partitions Optimization problem

proposed approach . requirements

proposed approach . data

Solution requirements
Set of data (simulation or real) for partitioning Fuzzy partitions modeling and Optimization technique FMF design Evaluation procedure

Data
Real data preferred U of Toronto-Mississauga Meteorological Station Temperature information for year 2000 and 2001

33

Masoud Makrehchi

34

proposed approach . modeling & optimization

proposed approach . modeling & optimization

Optimization
To search UD for the best set of support values Genetic Algorithms (GA)
Masoud Makrehchi

Performance indices
Fitness function in GA optimization procedure
Shannon entropy Mutual information

proposed approach . modeling & optimization

proposed approach . modeling & optimization

How we relate FMF to information measure?

Mapping the FMF on the histogram of given data Probability~statistics Maximizing the entropy of partitioned histogram based on given number of partitions (n)
Masoud Makrehchi
37

PDF~histogram

38

proposed approach . modeling & optimization

proposed approach . modeling & optimization

overlaps
In a n-fuzzy-partitioned information, allowed overlaps just between two adjacent partitions, we have n-1 overlaps
Masoud Makrehchi

NS

NM

1 Z PM

PL

How to model the overlaps between partitions?

-1

+1

proposed approach . modeling & optimization

proposed approach . modeling & optimization

Two strategy:
Overlaps as independent partitions: maximize entropy of independent partitions Overlaps as conjunction of two joint partitions: maximize entropy of joint partitions (considering mutual information)

First: Overlaps as independent partitions (2n-1) partitions NS

Masoud Makrehchi

NM

1 Z PM

PL

H1

H2
H3

H4
H5

H6
H7

H8

H9
42

41

proposed approach . modeling & optimization

proposed approach . modeling & optimization

NS

NM

1 Z PM

PL
Masoud Makrehchi

H1

H2
H3

H4
H5

H6
H7

H8

H9

proposed approach . modeling & optimization

proposed approach . modeling & optimization

Algorithm Do optimization for given number of partitions

2 n -1 i =1
Masoud Makrehchi

Increased and enhanced overlaps A conservative strategy

In fuzzy control applications,
Longer rise time Less overshoot Smooth convergence

Change width of partitions Until maximum H

H =

Hi

45

46

proposed approach . modeling & optimization

proposed approach . modeling & optimization

Second: Overlaps as conjunction of two joint partitions NS NM 1 Z PM PL

Masoud Makrehchi

NS

NM

1 Z PM

PL

H2 I1,2 I2,3 I3,4 H3

H4 I4,5

H2 I1,2 I2,3 I3,4 H3

H4 I4,5

H1

H5

H1

H5

proposed approach . modeling & optimization

proposed approach . modeling & optimization

Algorithm Do optimization for given number of partitions

Masoud Makrehchi

Change width of partitions Until maximum H

H= Hi - I (i ,i +1 )
i= 1 i= 1
50

n- 1

49

proposed approach . modeling & optimization

proposed approach . design

Decreased overlaps
In fuzzy control applications
Shorter rise time More overshoot
Masoud Makrehchi

Ready to design FMFs

We have partitions We need values of boundaries to have complete define of FMF
A criteria to choose right value for boundary is necessary

proposed approach . design

proposed approach . design

Importance of boundary
Defining a range instead of an exact value P2 P1 -1 P3 P4
Masoud Makrehchi

In two partitions A and B, if : XA1 < XB1In< XA3 <A X two partitions andB3 B, if :
x <x <x <x Well - defined Wellboundary, defined boundary, W ; W ; B
XA1 < XB1 < XA3 < XB3
A1 B1 A2 B2
B

WB (XA3 - XB1 )
P5 +1 xA1 A xA3 xB1 xB3 B

WB (XA3 - XB1 )

53

54

proposed approach . design xA2

a

proposed approach . design

b < a

b < a

b
xA1 xA2 a < b xB1 xA3 xB3
Masoud Makrehchi

xA1

xB1

a <

xA3

xB3

WB=XA3-XA2 WB>XA3-XB1

WB=XA3-XA2 WB<XA3-XB1

proposed approach . design xA2

b'

proposed approach . evaluation

Evaluation procedure

a a'

b < a
b' > a'
a < xB3
Masoud Makrehchi

xA1

xB1

xA3

WB=XA3-XA2 WB<XA3-XB1

Testing membership function in a complete fuzzy system reacting to a process, compare the output with heuristicdefined membership function Applying the algorithm on other set of data and study the behavior of membership function
58

57

Outlines
Introduction Problem Statement Proposed Approach Results Conclusion Conditions

results

Masoud Makrehchi

Normalized data Five partitions Algorithm test in both two modes

results

results

GA optimization parameters
Search space: 35,184,372,088,832 Population size: 400 Chromosome/string length: 45 Pcross-Over: 0.3 Pmutation: 0.01 Minimum generation: 200

First data set:

Hourly temperature of city of Toronto during year 2000
Masoud Makrehchi

Max: 36.88 Min: -20.67 Mean: 8.90 STD: 10.30

61

62

results

results

Temperature vs. time year 2000

Masoud Makrehchi

Temperature vs. time year 2000 Normalized temperature

results
Mode 1: Overlaps as independent partitions

results

65

Masoud Makrehchi

Temperature vs. time year 2000 Normalized temperature Histogram

Best strings over generations

66

results
Mode 1: Overlaps as independent partitions Mode 1: Overlaps as independent partitions

results

Masoud Makrehchi

Best strings over generations Mean of strings during convergence

Best strings over generations Mean of strings during convergence Resulted fuzzy memberships

results
Mode 2: Overlaps as conjunction of two joint partitions

results
Mode 2: Overlaps as conjunction of two joint partitions

Best strings over generations

69

Masoud Makrehchi

Best strings over generations Mean of strings during convergence

70

results
Mode 2: Overlaps as conjunction of two joint partitions

results

Second data set:

Hourly temperature of city of Toronto during year 2001
Masoud Makrehchi
Best strings over generations Mean of strings during convergence Resulted fuzzy memberships

Max: 32.23 Min: -23.34 Mean: 9.60 STD: 10.20

results

results

Temperature vs. time year 2001

73

Masoud Makrehchi

Temperature vs. time year 2001 Normalized temperature

74

results
Mode 1: Overlaps as independent partitions

results

Masoud Makrehchi

Temperature vs. time year 2001 Normalized temperature Histogram

Best strings over generations

results
Mode 1: Overlaps as independent partitions Mode 1: Overlaps as independent partitions

results

77

Masoud Makrehchi

Best strings over generations Mean of strings during convergence

Best strings over generations Mean of strings during convergence Resulted fuzzy memberships

78

results
Mode 2: Overlaps as conjunction of two joint partitions

results
Mode 2: Overlaps as conjunction of two joint partitions

Best strings over generations

Masoud Makrehchi

Best strings over generations Mean of strings during convergence

results
Mode 2: Overlaps as conjunction of two joint partitions

results
Mode 1: Overlaps as independent partitions

Data set 2000

81

Masoud Makrehchi

Best strings over generations Mean of strings during convergence Resulted fuzzy memberships

Data set 2001

82

results
Mode 2: Overlaps as conjunction of two joint partitions

Outlines
Introduction Problem Statement Proposed Approach Results Conclusion

Data set 2000

Data set 2001

Masoud Makrehchi

conclusion

85

Masoud Makrehchi

A solution for designing fuzzy membership function Besides fuzzy rules generation, a solution for designing fuzzy system by learning from example The idea: having generic membership function for generic data