In the name of God

GENERAL FUZZY MIN-MAX

NEURAL NETWORK
Presented by: Habib Alizadeh
Adviser: Dr. Farokhi

December 14

FUZZY SETS & NEUROFUZZY SYSTEMS

Fuzzy sets have been proposed by Zadeh. Compared to the classic
sets, fuzzy sets and their operations are more compatible with realworld systems and are highly efficient in pattern recognition and
machine learning problems.

Fuzzy logic usually is combined with a learning instrument.

Neurofuzzy systems are created by combining fuzzy logic and artificial
neural networks. Computational efficiency of neural networks and
capability of fuzzy logic to present complex class boundaries make
these networks a perfect tool for pattern recognition.
2

FUZZY MIN-MAX NEURAL NETWORK

GFMM

FMM & GFMM

The fuzzy min-max (FMM) clustering and classification neural
networks, with their representation of classes as hyper boxes in ndimensional pattern space and their conceptually simple but powerful
learning process, provided a natural basis for our development.

The proposed generalized fuzzy min-max (GFMM) neural network

incorporates significant modifications that improve the effectiveness of
the original algorithms.
5

FMM & GFMM

An

important development of the GFMM algorithm relates to the

interpretation of the membership values, both during the training and
the operation of the GFMM neural network.

THE ORIGINAL FMM ALGORITHMS

The FMM neural networks are built using hyperbox fuzzy sets.
A hyperbox defines a region of the n-dimensional pattern space, and
all patterns contained within the hyperbox have full cluster/class
membership.

A hyperbox is completely defined by its min point and its max point.

THE ORIGINAL FMM ALGORITHMS

Each hyperbox belongs to one of the classes and covers a part of
pattern space, defined as

Bj = {A, Vj ,Wj , f (A, Vj ,Wj )} A I n

FMMNN: 1-CLASSIFICATION (HYPERBOX

MEMBERSHIP FUNCTION)

The example of membership function bj presented in FMM

Classification NN for the hyperbox defined by min point V=
[0.2 0.2] and max point W= [0.3 0.4]: Sensitivity parameter =
4

FMMNN: 2-CLUSTRING (HYPERBOX

MEMBERSHIP FUNCTION)

The example of membership function bj

presented in FMM clustering NN for the
hyperbox defined by min point V = [0.2 0.2]
and max point W = [0.3 0.4]: Sensitivity
parameter = 4

10

FMMNN LEARNING
The fuzzy min-max neural network learning algorithm is a four-step
process consisting of:

1.
2.
3.
4.

Initialization
Expansion
Overlap Test

Contraction

The last three steps repeated for each training input pattern.
11

GFMM ALGORITHM
A- Basic Definitions
1)

Input: The input is specified as the ordered pair: ,

Where = is the hth input pattern in a form of lower, , and

upper, , limits vectors contained within the n-dimensional unit cube
.
And 0,1, , is the index of one of the + 1 classes, where = 0
means that the input vector is unlabeled.
12

GFMM ALGORITHM (MEMBERSHIP

FUNCTION)
2) Fuzzy Hyperbox Membership Function:
Where

is the min point for the jth hyperbox,

is the max point for the jth hyperbox, and the

membership function for the jth hyperbox is

13

GFMM ALGORITHM (MEMBERSHIP

FUNCTION)

where

f is a two parameter ramp threshold function and is the

sensitivity parameters regulating how fast the membership
values decrease.

14

GFMM ALGORITHM (MEMBERSHIP

FUNCTION)
One-dimensional (1-D) membership function
where

15

GFMM ALGORITHM (MEMBERSHIP

FUNCTION)

The 1-D illustration of membership value finding for an input in form

of lower and upper bounds.
16

GFMM ALGORITHM (MEMBERSHIP

FUNCTION)
Two-dimensional (2-D) membership function

The hyperbox is defined by min point V= [0.2 0.2] and max point W= [0.3 0.4]:
Sensitivity parameter = 4

17

GFMM ALGORITHM (LEARNING)

B - GFMM Learning Algorithm:
1) Initialization: The hyperbox is adjusted for the first time
using the input pattern = the min and max points
of this hyperbox would be Vj= and Wj= .

2) Hyperbox Expansion: When the hth input pattern Xh is

presented, the hyperbox Bj with the highest degree of
membership and allowing expansion (if needed) is found.

18

GFMM ALGORITHM(EXPANSION)
The expansion criterion, consists of the following two parts:
and

19

GFMM ALGORITHM (EXPANSION)

with the adjust Bj operation defined as:

The parameter

is a user-defined value that imposes a

bound on the maximum size of a hyperbox and its value significantly
affects the effectiveness of the training algorithm.
20

GFMM ALGORITHM
3) Hyperbox Overlap Test: Assuming that hyperbox Bj was expanded
in the previous step, test for overlapping with Bk if

21

GFMM ALGORITHM(OVERLAP TEST)

The four cases are being considered
(where initially = 1).

If overlap for the th dimension has

been detected (one of the above four
cases is valid) and
,
then
,
.

If overlap for the ith dimension has not

been detected, set
signifying
that the contraction step is not

22

GFMM ALGORITHM
(CONTRACTION)
4) Hyperbox Contraction: If
then only the th dimensions of
the two hyperboxes are
adjusted.

23

AN EXAMPLE ILLUSTRATING THE

LEARNING ALGORITHM

24

GFMM ALGORITHM ()
5) An Adaptive Maximum Size of the Hyperbox:

In the original FMM NNs the user defined parameter .

To find the best value of this parameter the network has to be
trained for several different s and verified by checking the number of
misclassifications.
A large value of can cause too many misclassifications.
When is small, many unnecessary hyperboxes may be created.
25

GFMM ALGORITHM ()
The training is completed when:
a) after presentation of all training patterns there have been no
misclassifications for

the training data;

b) or the minimum user-specified value of the parameter has been reached.

where is the coefficient responsible for the speed of decrease of .

26

GFMM ALGORITHM ()

The result of NN training for the 42 input pattern data set (three
classes).

27

was constant during training.

TOPOLOGY OF THE NETWORK

28

THE EXAMPLE OF
CLUSTERING/CLASSIFICATION OF
LABELED AND UNLABELED FUZZY INPUT
PATTERNS
The data set consists of 26

patterns from which 15 are

labeled as belonging to one of
four classes and the remaining
11 are unlabeled.

29

FMM & GFMM(3 REAL DATA SET)

30

COMPARISON OF THE PERFORMANCE OF THE

GFMM
WITH SEVERAL OTHER TRADITIONAL
CLASSIFIERS

31

Thanks from your attention.

32