International Journal of Engineering Research (ISSN : 2319-6890)

Volume No.2, Issue No.2, pp : 29-32 01 April 2013

IJER@2013 Page 29

Literature Survey on Genetic Algorithm Approach for Fuzzy Rule-Based
System

Dinesh P.Pitambare, Pravin M.Kamde
M.E. Computer (Networks), Sinhgad College of Engineering, Pune-41, Maharashtra, INDIA
[email protected]


Abstract- Fuzzy-Rule Based Clustering (FRBC) is
proposed, for automatically exploring potential clusters in
dataset. This uses supervised classification approach to
achieve the unsupervised cluster analysis. Fusion of clustering
and fuzzy set theory is nothing but fuzzy clustering, which is
appropriate to handle problems with imprecise boundaries of
clusters. A fuzzy rule-based classification system is a special
case of fuzzy modeling, in which the output of system is crisp
and discrete. Fuzzy modeling provides high interpretability
and allows working with imprecise data. To explore the
clusters in the data patterns, FRBC appends some randomly
generated auxiliary patterns to the problem space. It then uses
the main data as one class and the auxiliary data as another
class to enumerate the unsupervised clustering problem as a
supervised classification one.

I. INTRODUCTION

Clustering is unsupervised approach where we do not need
to provide training data. The task of separating different object
in to groups is nothing but cluster, in which the objects in the
same group are much more identical to each other than objects
in other group. Ref.[2], one of the major problems in cluster
analysis is the determination of the number of clusters in
unlabeled data, which is a basic input for most clustering
algorithms, which can be find using cluster validity
function[7]. Ref. [10], synthesis between clustering and fuzzy
set theory is fuzzy clustering, in which one object may belongs
to number of clusters. Fuzzy clustering is useful in handling
unclear boundaries of clusters. Ref. [9], fuzzy c-means
clustering (FCM) is the most popular fuzzy clustering
algorithm whose time requirement is high. It also suffers from
the presence of noise and outliers and the difficulty to identify
the initial partitions. A special case of fuzzy modeling is the
fuzzy rule-based classification system, whose output is crisp
and discrete. Ref.[3], then fuzzy rules are generated using
SGERD A steady state genetic algorithm for extracting fuzzy
classification rules from data, whose product is best rule on
the basic of that best rule process of clustering takes place.
Genetic algorithms (GAs) are search algorithms that provide
a robust search capability in complex spaces hence fulfills an
efficient and effective searching feature.


Ref. [3], E. G. Mansoori et al in SGERD, proposed method for
extraction of fuzzy rule from given data set. This algorithm is
generational and population-based, where its generations are
finite and bounded to the problem dimension. Individual
selection of this algorithm is serial, and only the best ones can
survive. Each parent produces a finite number of offspring
through reproduction. SGERD uses fitness function which is
, based on a rule evaluation criterion, provides the best rules
among all candidates.
Ref. [4] H. Ishibuchi and T. Yamamoto stated in
comparison of heuristic criteria for fuzzy rule selection in
classification problems, fuzzy rules are used for knowledge
representation with high interpretability. Some heuristic
criterias are used for extraction of prespecified number of
fuzzy rule .Genetic algorithm based rule selection criteria
improves classification ability of extracted fuzzy rule.
Ref. [4], there are seven different rule selection criteria viz.
confidence, confidence with minimum support level, support
with minimum confidence level, SLAVE (Structural Learning
Algorithm in a Vague Environment) criteria and Castro
criteria. Heuristic procedure is used here for obtaining good
rule set.
In this paper, we present SGERD to generate fuzzy rules
from a set of training data. The rest of this paper is organized
as follows. In Section II, we briefly describe blueprint of
membership function. In Section III, we present fuzzy rule
generation methods. In Section IV, we present SGERD
algorithm. The conclusion is discussed in Section V.

II. BLUEPRINT OF MEMBERSHIP
FUNCTION

Fuzzy rules are simply IF-THEN rules, used for
knowledge representation with high interpretability. For a
pattern classification problem, Fuzzy IFTHEN rules include
two clauses viz. antecedent and consequent. Antecedent clause
includes conditions for the occurrence of the event; while
consequent contain consequence of antecedent clause. Fig. 1
shows membership functions for four different values of K,
where L3, L4 and L5 are the linguistic labels, which interprets
linguistic values small, medium, and large, respectively.
International Journal of Engineering Research (ISSN : 2319-6890)
Volume No.2, Issue No.2, pp : 29-32 01 April 2013
IJER@2013 Page 30


Fig. 1.Different partitioning of each input attributes [1].

For generating fuzzy rules we need to draw membership
function for corresponding input data. The length of
membership function is obtained using difference between
maximum and minimum value of the attribute. Membership
function rescaled each input attribute to unit interval [0, 1] by
using linear transformation that preserves the distribution of
training patterns. Then, partitioning the pattern into fuzzy
subspaces took place where each subspace is identified by a
fuzzy rule. By assigning linguistic values to each input
attribute we can do partitioning. Generally, triangular
membership functions are used for this purpose, as they are
simpler and more human understandable with high
interpretability [1].
When an input partitioning of pattern is given, then we can
generate fuzzy rules for all combinations by considering all
possible combinations of antecedent linguistic values.
However, this is not easily handled as far as high dimensional
problems are concern, as it generates numerous rules. For
example, for a data set having n input attributes, K
n
fuzzy rules
might be generated.
This problem can be solved using some rule evaluation
criteria to select a small subset of rules among all candidates
[4].This approach simultaneously considers all membership
functions in Fig. 1 for each attribute. That is, one of the 14
fuzzy sets can be used for each attribute when generating a
candidate rule. In this case, for an n-dimensional problem, 14
n

antecedent combinations should be considered.
However, practically it is not possible to consider such a huge
number of antecedent combinations when dealing with high-
dimensional problems. Solution for this problem is presented
in [6] by adding the fuzzy set dont care (with linguistic label
L0) to each attribute.
Example1: Let us consider two data patterns x1 and x2.
If x1 is small and x2 is small then Class 1 . (i)
If x1 is small and x2 is large then Class 2 ..... (ii)
If x1 is large and x2 is small then Class 3 (iii)
If x1 is large and x2 is large then Class 4 .... (iv)

When fuzzy rule-based systems are used for two-dimensional
problems, generated fuzzy rules [4] from above equations are
shown in fig. 2.

Fig 2. Four fuzzy rules in two-dimensional pattern [0, 1]
[0, 1].

III. GENERAL DESIGN OF FUZZY RULE-
BASED CLASSIFIERS

Fuzzy IFTHEN rules for a pattern classification problem for n
attributes can be written as follows [1]:
Rule Rj: IF x1 is Aj1 and . . . and xn is Ajn
THEN class Cj, for j = 1, N (1)
Where X = [x1, x2, xn]
n-dimensional pattern vector,
Aji Antecedent linguistic value of Rj,
Cj Consequent class,
N Number of fuzzy rules.
Generally, for an M-class problem with m labeled patterns
Xp = [xp 1 , xp 2 , . . . , xpn ],where p = 1, . . . ,m, our task is to
design the classifier to generate N fuzzy rules as in (1). Using
training pattern in the corresponding fuzzy subspace
consequent class Cj of the fuzzy rule Rj in (1) can be
determined. The compatibility grade of the training pattern Xp
is defined with the antecedent part of the rule, by the usage of
the product operator as

[
=
=
n
i
xpi ji Xp j
1
) ( ) (

Where ji (.) is the membership function of the antecedent
fuzzy set Aji .Using heuristic method which is based on
confidence we select the consequent class of a rule [4]. This
confidence is given as

=
e
=
m
p
classT Xp
Xp j
Xp j
ClassT Aj conf
1
) (
) (
) (

Hence from (3), the consequent class Cj is determined using
the maximum confidence, shown using (4).
} ,...., 1 ) ( max{ arg M T classT Aj conf Cj = =
Input partitioning of the pattern space is given, there are many
approaches to generate fuzzy classification rules from data.
Using heuristic methods candidate rules can be evaluated
easily [3]. Their basic criterion, which is a fuzzy version of the
difference between the number of true positives and false
positives, is specified as in (5).


e e
=
classCj Xp classCj Xp
Xp j XP j classCj Aj f ) ( ) ( ) (

(5)
(2)

(3)
(4)
International Journal of Engineering Research (ISSN : 2319-6890)
Volume No.2, Issue No.2, pp : 29-32 01 April 2013
IJER@2013 Page 31

Single winner rule is the most popular and simple reasoning
method in fuzzy rule-based classifiers. Using this method, new
pattern Xt = [xt 1, xt 2, xtn] is classified according to the
consequent class of the winner rule Rw. In fact, the winner rule
has the maximum compatibility grade with Xt among the fired
rules. This can be stated as
} ,...., 2 , 1 ), ( max{ ) ( N j Xt j Xt w = = (6)

IV. FUZZY RULE-BASED CLUSTERING
ALGORITHM

Our anticipated approach uses supervised classification
approach to do the unsupervised cluster analysis by the
addition of some auxiliary data patterns to the main data and
usage of a fuzzy classifier to solve this new problem.
For extracting each cluster FRBC considers all unlabeled
data pattern as main data pattern and label them as class 1.Then
generation of random data pattern takes place which are
uniformly distributed data patterns. This random data pattern
can be treated as auxiliary data pattern.
For high dimensional problem, it is impossible to generate
specific number of uniform pattern, so they are produced
randomly and added to pattern space as a class 2 to form two-
class problem. Auxiliary data generation is controlled by
number of main data and their distribution.
For appropriately estimation of the number of auxiliary data
pattern should be added, the summation of a within-cluster
point to-point scatter matrix is used. This value for main data is
defined

<
=
j i
Xi Xj d q ) , (
(7)
Where Xi is one of the mM main data patterns, and d (.,.) is a
distance metric, which is usually the Euclidean norm.
Similarly, for auxiliary data {X

i, i
=1, mA}

<
=
j i
j X i X d q ) ' , ' ( '
(8)
Auxiliary random patterns are added incrementally using q and
q until q exceeds q. In this regard, the distribution of main
data patterns influences the size of auxiliary instances.
After the preparation of the two-class problem, the FRBC
make use of SGERD for a fuzzy rule generation to solve two-
class problem. Then rule evaluation measures can be applied
for the rule R
j
.
j
p X j Xp j
class Aj f
class Xp class Xp
'
) ' ( ) (
) 1 (
2 1
q


e e

=
(9)
Where
= j ' q
Number of auxiliary patterns that are covered by
the rule Rj
After generation of the fuzzy rules, the FRBC sets best rule
as the initial members of the first cluster and removes them
from the problem space to not be reconsidered for other
clusters.
In this manner FRBC explore all potential clusters with
assignment of distinct class labels to the consequent of fuzzy
rules that represent the explored clusters [8]. It then uses the
fuzzy rules simultaneously to classify the main data patterns
and, therefore, identify the clusters boundaries. Since the
FRBC uses the single winner-rule reasoning method, it selects
fuzzy rule with highest compatibility grade as in (6).To
increase the clustering accuracy [5], centroids can be calculated
as in (10)

e
e

=
Gj Xp
Gj Xp
XP j
Xp Xp j
Cj
) (
) (

(10)

Algorithm: Fuzzy rule-based clustering algorithm.
Inputs: mI unlabeled data patterns and threshold .
Outputs: The number of clusters J and their members that
are identified by J fuzzy classification rules.

1. Name the unlabeled patterns {Xp , p = 1, . . . , mI } of
the problem as the main data.
2. Let j = 1 (j: the cluster number).
3. Let mM = mI (mM : number of main data
(non clustered) patterns).
4. Generate some uniformly random patterns and append
them incrementally to the main data patterns until q_ in
(8) exceeds q in (7) (call these added patterns {Xp, p
=1, mA} as auxiliary data).
5. Label mM main data patterns as Class 1 and mA
auxiliary data as Class 2.
6. By the usage of SGERD with the criterion in (9),
generate fuzzy rules to classify this two-class problem
(with mM + mA data patterns).
7. Among the rules that are generated for Class 1, choose
the best one and name it Rj.
8. If e (Rj) < goto Step 12.
9. Set aside the main data patterns in the fuzzy subspace
of Rj as the initial members of cluster Gj.
10. Remove the members of Gj from the problem space,
and let mM = mM |Gj |.
11. Let j = j + 1, then goto Step 4.
12. Let J = j 1 as the number of explored clusters and
R1, RJ as their representing fuzzy rules, where Rj
identifies cluster Gj.
13. Replace the consequent of the rule Rj with j (for
j =1,. , J) so that the fuzzy rule base, which represents
clusters {G1, GJ}, would be {R1, . . . , RJ } with the
consequents {1, . . . , J}.
14. By the usage of the fuzzy rule base, classify the mI
unlabeled patterns to identify the actual members of
J explored clusters.
15. Compute the centroid of J clusters from (10).
International Journal of Engineering Research (ISSN : 2319-6890)
Volume No.2, Issue No.2, pp : 29-32 01 April 2013
IJER@2013 Page 32

16. Regroup the mI unlabeled patterns according to the
nearest centroid of J clusters.
17. Stop.

V. FUZZY RULE GENERATION METHOD

Method used in generating fuzzy rules is SGERD, which is
discussed bellow.

A. Steady state genetic algorithm for extracting fuzzy
classification rules from data:-
SGERD generates a prespecified number of Q rules per class
(R = M Q rules in total at most). Input for this algorithm is
labeled pattern and projected output is fuzzy rules.

Algorithm: SGERD.
I nputs: m labeled patterns of an n-dimensional M-class
problem and Q.
Outputs: Possibly R = MQ fuzzy classification rules.
1. i = 1 (i: generation number).
2. Generate all fuzzy rules having only one active
antecedent variable (at most C = 14 n
candidate rules would be generated).
3. Determine the consequent class of each candidate rule
using (4).
4. Divide the candidate rules into M groups according to
their consequent class.
5. Rank, in descending order of their fitness values, the
Candidate rules in each group.
6. Choose the best Q rules from each class (i.e., possibly
R = M Q rules in total) as the population in the i
th

generation. In the first generation only, choose the second best
R rules as the auxiliary population and put away for mutation.
7. Increment i, if i > n, goto step 11.
8. Use all individuals in the previous generation (i.e., R
rules) as parents and do reproduction (i.e., crossover, mutation,
or elitism) on them. That is, for each parent rule, generate as
offspring all fuzzy rules having one more active antecedent
variable than its parent, provided each new offspring is fitter
than its parent. In
this case, the number of offspring will totally be R 14 at
most.
9. If no offspring fitter than the parents is produced in
step (8), goto step 11.
10. Consider both parents and offspring in step 8 as
Candidate rules (at most C = R + R 14 rules in total) and goto
step 3.
11. Use R = M Q rules (obtained in step 6 for the i
th

generation) as the final population and stop. The actual length
of these rules is i or less.

As mentioned before, the rule selection scheme in SGERD
only considers the evaluation measure of each rule to select the
best ones through competition.

V.CONCLUSSION

In this paper, we proposed FRBC is a novel fuzzy rule-
based clustering algorithm to automatically explore the
potential clusters in the datasets. It looks at the clustering
issue as a classification problem by the addition of some
auxiliary data patterns to the main data and then generation
of some fuzzy rules to classify the new pattern space.
The generated fuzzy rules, which represent the clusters, are
human understandable with acceptable accuracy. SGERD is a
steady-state genetic algorithm to extract fuzzy classification
rule from data. Best rule is generated in SGRED.


REFERENCES

I. Eghbal G. Mansoori,FRBC: A Fuzzy Rule-Based
Clustering Algorithm, IEEE Trans. Fuzzy Syst., vol. 19, NO.
5, October 2011
II. L. Wang, C. Leckie, K. Ramamohanarao, and J.
Bezdek, Automatically determining the number of clusters in
unlabeled data sets, IEEE Trans. Knowl. Data Eng., vol. 21,
no. 3, pp. 335350, Mar. 2009.
III. E. G. Mansoori, M. J. Zolghadri, and S. D. Katebi,
SGERD: A steady state genetic algorithm for extracting
fuzzy classification rules from data, IEEE Trans. Fuzzy Syst.,
vol. 16, no. 4, pp. 10611071, Aug. 2008.
IV. H. Ishibuchi and T. Yamamoto, Comparison of
heuristic criteria for fuzzy rule selection in classification
problems, Fuzzy Optim. DecisionMaking, vol. 3, no. 2, pp.
119139, 2004.
V. R.Xu and D.Wunsch,Survey of clustering
algorithms, IEEE Trans.Neural Netw., vol.16, no.3, pp.645-
678, May 2005.
VI. H. Ishibuchi and T. Yamamoto, Fuzzy rule selection
by multi-objective genetic local search algorithms and rule
evaluation measures in data mining, Fuzzy Sets Syst., vol.
141, no. 1, pp. 5988, 2004.
VII. G. J. Klir and B. Yuan, Fuzzy sets and fuzzy logic:
Theory and Applications, in Pattern Recognition. Englewood
Cliffs, NJ: Prentice-Hall, 1995.
VIII. R. Dubes and A. K. Jain, Validity studies in
clustering methodologies, Pattern Recognit., vol. 11, pp.
235254, 1979.
IX. J. C. Bezdek, Cluster validity with fuzzy sets, J.
Cybern., vol. 3, pp. 58 71, 1974.
X. L. Zadeh, Fuzzy sets, Inf. Control, vol. 8, pp. 338
353, 1965.