A Variable Fuzzy-Neural Network to Pattern Classification ... - Read

clangedbivalveAI and Robotics

Oct 19, 2013 (3 years and 5 months ago)

58 views

A Simple Method to Extract Fuzzy Rules by Measure
of Fuzziness



Jieh
-
Ren Chang

Nai
-
Jian Wang

Abstract


Use a
variable fuzzy
-
neural network

structure to
implement the fuzzy rules system.


First, we extract fuzzy rules from different class
region which was named as
activation hyper
-
box.


Second, when the activation hyper
-
boxes

are
overlapped, a recursive process are applied to
additive activation hyper
-
box

in these uncertainty
-
overlap regions.


Third, the stop criterion for the recursive process
--
by measure of fuzziness.

Relation between activation hyper
-
boxes and
overlap regions by 2
-
dimensional example

Contents



Motivation.


Introduction.


Measure of fuzziness for a fuzzy set.


Measure of fuzziness of a fuzzy rule in a fuzzy
system.


Fuzzy
-
neural network.


Learning algorithm.


Compare our method with other methods.


Conclusions.

Motivation


To extract more efficiently fuzzy rules from
numerical information data in classification
problem.


To save computation cost


To get available rules and cancel redundant
rules

Introduction


Human can always collect the knowledge to
discriminate the uncertainty or ambiguous
data by their experience.


But computer still can’t be dealt perfectly in
classification problem.


So, many methods are still proposed to
improve the performance of classification
problem.


The methods of classification problem are
divided into four groups:

1) Statistical method:


It is not practical in solving classification
problem in a real world.

2) Neural network:


It is a system that is constructed to make
use of some organizational principles like
human brain. It is good for many application.


3)Fuzzy inference engine:


By querying experts’ experience or other
techniques directly from training data to build
fuzzy rule database.




4)Hybrid neural
-
fuzzy technique:


It combines the fuzzy inference and neural
network theory to computer
-
based pattern
recognition.


Hong and Lee
, proposed a method based on
the fuzzy clustering technique to setup the
decision tables
. But they need to determine
the scaling it usually takes
more computation
time
.



Hong and Chen
, they propose the other
method to decrease the computation time,
but it still
generates many rules

and
take very
much computation process
, when the training
data increase.



Wu and Chen
have a fuzzy learning algorithm
base on the
α
-
cut, can
induce the fuzzy rule

and reaches a
higher average classification
ratio
. But we don

t know how to
select the
α
-
cut

.



P.K. Simpson

setup the fuzzy rules by an
expansion
-
contraction, it usually generated
too many hyper
-
box that mean
too many
rules to be concerned
.


S. Abe and M
-
S. Lan

extract the fuzzy rules
by resolving overlaps, it can
decrease the
learning process
. But there some
drawback

in following points:


1)It needs more computation time to resolve
overlaps when the data include many classes.


2)It can’t be resolved in some critical
condition.


3)It generate many meaningless fuzzy rules
as the data are chaos.


Our propose is to decrease the computation
time and to extract more efficient fuzzy rule,
the method is described in the following steps:

1)Find the activation hyper
-
box.

2)Find uncertainty overlap.

3)Extracts fuzzy rules .

4)Construct an easy and efficient neural
network by measure of fuzziness.

Measure of Fuzziness of a Fuzzy
Set





To measure uncertainty of vagueness .




Measure of fuzziness is a function
ƒ
, the
function
ƒ

satisfies the following axioms:



Axiom 1

:
ƒ
(A)=0 if only if A is a crisp set.



Axiom 2

: If A B, then
ƒ
(A)
ƒ
(B). Where
A B denotes that A is shaper than B.



Axiom 3

:
ƒ
(A) assumes the maximum
value if and only if A is maximally fuzzy

Degree of fuzziness of fuzzy
set






Normalized measure of
fuzziness


Measure of Fuzziness of a Fuzzy
Rule in a Fuzzy System


In this section, we define a classification system by a
sequence of multi
-
input
-
single
-
output fuzzy rules as
follows





n is the number of attribute of the classification
system


c is the number of class of the system


A
i,k

is the linguistic label, i=1,2,…n,


R
k

can be rewrote by the T
-
norm operator with min
operation in the following:





The membership value of this rule R
k

represented as:





We can define the measure of fuzziness of the
rule
R
k
in the fuzzy rule system as:

According to the formula (3)


We can decide the rule R
k

is worth to exist in
this rule
-
based system or not necessary.



If the rule have high measure of fuzziness of
a rule, it means too much uncertain for this
rule.


A Fuzzy
-
Neural Network Structure

A variable structure


We will leave the rule which is very efficient
and useful, so the number of nodes in the
second layer are variable.



We will reduce the cost, because the
redundant second layer nodes are eliminated.

Second layer includes two Sub
layer


the first sub layer is configured by the hyper
-
box nodes which are created from our
proposed algorithm


the second sub
-
layer is a maximum
-
operation
node, which takes the maximum values of
inputs from the first sub
-
layer.

Learning Process


Step1: set level = 1.



Step2: Set up the hyper
-
boxes and
membership function for each class.



Step3: Find the overlap among the activation
hyper
-
boxes of level l ,then l=l+1.


Step4:Extract activation hyper
-
boxes and set
up feature as in step 1.



Step5:Calculate the measure of fuzziness for
each extracted fuzzy rule. If it is bigger than
threshold, we discard this rule.



Step6:If none of hyper
-
box exist in Step 4,
then stop the process, else go to Step 2.



Step7:Build up the fuzzy
-
neural network
structure by these extracted fuzzy rules



Performance Evaluation


We use Fisher’s iris data, there are three
kinds of flowers, four kinds of attributes.


Three flowers:


Setosa Versicolor Verginica

Four attributes:


Sepal length Sepal width Petal length Petal
width

Original Iris Data

Pseudo
-
Iris data

Sepal
length

Sepal
width

Petal
length

Petal
width

Setosa

4.4~5.8

2.9~4.4

1.0~1.9

0.1~0.6

Versicolor

5.0~7.0

2.0~3.4

3.0~5.1

1.0~1.8

Verginica

4.9~7.7

2.5~3.8

4.8~6.9

1.4~2.5

單位:
cm

Randomly generated area

Conclusions


By this proposed method, we can find more efficient
fuzzy rules.


It generates fewer fuzzy rules than other methods
[9][10][11][14].


It avoids a huge matrix computation [9] so its
computation time decreases.


It provides a simple recursive process and stopping
criteria to extract the fuzzy rules in the uncertainty
-
overlap region. Thus, the network structure is simple
and easy to implement.


The classifier can be generated even for a large
scale of data pattern.