Hybrid Fuzzy Classification Algorithm with Modifed Initialization and
Crossover
Tatiana Pleshkova
a
and Vladimir Stanovov
b
Siberian Federal University, Krasnoyarsk, Russian Federation
Keywords:
Classification, Fuzzy Logic, Genetic Algorithm.
Abstract:
The article proposes two modifications of initialization and crossover operations for the design of a genetic
fuzzy system. A fuzzy logic system is used to solve data classification problems and is automatically generated
by a genetic algorithm. The paper uses a genetic algorithm to encode of several fuzzy granulations into a single
rule, while each individual encodes a rule base. The proposed algorithm uses several training objects of the
same class to create a single rule during initialization. The modified crossover creates a new rule base from
most efficient rules selected from parents. To evaluate the effectiveness of the modification, the computational
experiments were carried out on several datasets, followed by verification using Mann-Whitney U test. The
proposed initialization modification allows reducing the number of rules in a fuzzy rule base and increasing
the accuracy and F-score on some datasets. The crossover modification shows higher efficiency only on one
dataset.
1 INTRODUCTION
In the modern world, data analysis is a necessary
process in all spheres. One of the popular methods
of knowledge systematization is classification, which
implies the organization of data into categories for ef-
fective use. Various tasks are solved with classifica-
tion methods, for example, the classification of data
from satellite images or medical diagnosis and prog-
nosis of diseases (S¸tefan, 2012). There is no regular
method for solving data classification problems. De-
pending on the formulated conditions, it is necessary
to select a method that would classify the data into
the appropriate categories. There are many methods
for solving data classification problems that can be
divided into two groups, based on “black box” and
“white box” principles (Komargodski et al., 2017).
Depending on the problem, a suitable method is se-
lected before starting the classification. First of all,
due to the fact that the user of the system that classi-
fies data does not always have the opportunity to un-
derstand the decision-making process, there is a need
to create a system that allows the classification pro-
cess to be performed while getting information about
which characteristics influence the classifications. In
a
https://orcid.org/0000-0001-7761-7808
b
https://orcid.org/0000-0002-1695-5798
other words, the users’ involvement is aimed at study-
ing the decision-making process, which gives them
the opportunity to use their knowledge of the subject
area to increase the reliability of data classification.
To fulfill these conditions, it is necessary to use the
“white box” model when implementing the data clas-
sification method. If the solver type (neural network
and so on) is not important, then it makes sense to
use the data classification method without providing
information about which characteristics influence the
classification results.
In this study it was decided to use Fuzzy Rule-
Based System due to its flexibility and interpretabil-
ity. Many researchers use Fuzzy Rule-Based System
to implement the “white box” model (Bodenhofer and
Herrera, 1997). L. Zadeh proposed a fuzzy logic in
1965 that was able to cope with the task of describing
vague definitions of human language (Zadeh, 1965).
And in 1974, I. Mamdani designed the first function-
ing controller for a steam turbine (Mamdani, 1974),
which operated based on the L. Zadeh algebra. Mod-
ern researchers also continue to develop this area and
have achieved great success. For example, Y. Nojima,
H. Ishibuchi and F. Herrera are doing research with
the classification of objects based on fuzzy rules. In
the current study, it was decided to use the method
discussed in the article (Ishibuchi et al., 2013) as base-
line, where H. Ishibuchi proposed a machine learn-
Pleshkova, T. and Stanovov, V.
Hybrid Fuzzy Classification Algorithm with Modifed Initialization and Crossover.
DOI: 10.5220/0011587500003332
In Proceedings of the 14th International Joint Conference on Computational Intelligence (IJCCI 2022), pages 225-231
ISBN: 978-989-758-611-8; ISSN: 2184-3236
Copyright © 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
225
ing algorithm based on fuzzy logic. Our goal is to
improve the quality of fuzzy classifiers by proposing
new genetic operations, to make the algorithm work
more efficiently, that is, to find better solutions faster,
but with the same level of complexity. In particu-
lar, the contributions of this study are: initializing
rules from several objects is more efficient and select-
ing better rules during crossover may improve perfor-
mance.
This paper is organized as follows: Section II con-
tains related works. In Section III we briefly ex-
plain the basic method and its features. Next we
explain proposed modifications for initialization and
crossover in Section IV. In Section V we present the
results. Finally we conclude this paper in Section VI.
2 RELATED WORKS
Genetic fuzzy systems have attracted considerable at-
tention in the artificial intelligence community in the
last decades. Since 2000 there was a growing need
to find a compromise between interpretability and
accuracy in the tasks of linguistic fuzzy modeling.
First of all in (Ishibuchi and Nakashima, 2000) au-
thors introduce the effective use of rule weights in
fuzzy rule-based classification systems. In (Alcal
´
a
et al., 2007) a new post-processing method was in-
troduced. This method was based on the well-known
SPEA2 algorithm. In (Fern
´
andez et al., 2010) evo-
lutionary approaches that help to search for a set of
rules were considered. The hybrid fuzzy genetics-
based machine learning (GBML) algorithm was pro-
posed in (Ishibuchi et al., 2013), where algorithm had
a Pittsburg-style framework in which a rule set is han-
dled as an individual. The operation of the algorithm
will be described in more detail in the next section.
Each of the works uses the representation of the
decision-making process in the language understand-
able for the expert. It is possible thanks to the use
of fuzzy rules in the classification, which determine
whether an object with known characteristics belongs
to a particular class. A fuzzy rule consists of a con-
dition of the type “if... then... with fuzzy terms in
the“if... part and the corresponding class number in
the “then...” part (Herrera and Magdalena, 1997).
Rule R
n
: if x
1
is L
q1
and...and x
v
is L
qv
then Class C
q
with CF
q
, (1)
where n number of rules in the rule base, v number
of variables in the data sample, L – this is a linguistic
term, C class label, CF rule weight (which is a
real number in the unit interval [0, 1]).
3 GENETICS-BASED MACHINE
LEARNING
The method in (Ishibuchi et al., 2013) is based on the
search for the optimal rule base. This search is carried
out using evolutionary algorithms, where the main
idea is based on Charles Darwin’s theory of natural
selection (Bleckmann, 2006). The process of finding
the best solution begins with a set of individuals, i.e.
a population. The process of natural selection begins
with the choice of the individuals with better fitness
from the population. The selected individuals pro-
duce an offspring that inherits the characteristics of
the parents. If the parents are better i.e. have a higher
fitness than the others, there is an opportunity that
their offspring will be better than the parents and will
have a higher chance of survival. This process contin-
ues to be repeated for a certain number of generations,
and at the end of the generation the fittest individual
is found (Mitchell, 1996). One individual consists of
n fuzzy rules, where upper limit is n 50. Each rule
is designed using linguistic terms L
1
, L
2
, ..., L
14
. One
of the features of study (Ishibuchi et al., 2013) is the
use of several fuzzy granulations for each linguistic
variable. Figure 1 shows this concept. There are 14
linguistic variables and a “don’t care” condition (DC),
which means that for this variable in this rule there is
no difference what value the variable has. This is de-
scribed in (Ishibuchi et al., 2013).
Figure 1: Fuzzy granulations (L
1
, L
2
,...,L
14
).
After that, using the fitness function, the fitness
of the individual is determined, i.e. how well the
fuzzy rules base performs classification. After receiv-
ing fitness function value of each individual, we can
determine the probability of choosing a particular in-
dividual for following reproduction. The genetic al-
gorithm, which is one of the variations of the evolu-
tionary algorithm and is used in the study, consists
of several stages: initialization, selection, crossover,
mutation and formation of a new generation (Banzhaf
FCTA 2022 - 14th International Conference on Fuzzy Computation Theory and Applications
226
et al., 1998). The feature of the Ishibuchi algorithm
is that after the mutation, the Michigan-style part is
applied.
The first stage is initialization, in which individu-
als are formed. When creating a rule, a random ob-
ject is first selected from the training sample, then a
rule is formed based on the parameters of the selected
object, that is, the corresponding linguistic variable
is selected for each parameter. Then the confidence
is calculated and applied in evaluating the rule class
number, which is the most corresponding to a partic-
ular rule. To specify the consequent class C
q
and the
rule weight CF
q
, we first calculate the confidence of
the association from the antecedent fuzzy vector L
q
to
each class k (k = 1, 2, ..., M) (Ishibuchi et al., 2013) as
follows:
Con f (L
q
Class k) =
x
p
Class k
µ
L
q
(x
p
)
m
p=1
µ
L
q
(x
p
)
(2)
where µ
L
q
is the product of all memberships values of
the antecedent fuzzy set L
qi
at the input value x
pi
.
If the confidence value for a particular class C
q
is
larger than 0.5 in eq. 2 in (Ishibuchi et al., 2013) com-
putational experiments, we generate a fuzzy rule with
the antecedent fuzzy rule and the consequent class C
q
.
Then the rule weight CF
q
in (Ishibuchi et al., 2013) is
specified as follows:
CF
q
= 2 Con f (L
q
Class C
q
) 1. (3)
It is established that if CF
q
> 0, then the rule
is added to the fuzzy rule base. Fuzzy rules with
negative weights are not applicable for classification.
One of the advantages of this work is the use of a
Michigan-style algorithm probabilistically applied to
each set of rules. The Michigan-style part is used as
local search after main genetic algorithm.
The fitness of the main genetic algorithm was cal-
culated as shown in eq. 4, and the fitness of the
Michigan-style part is the fitness of the individual
rules at the selection stage, which was calculated
based on the number of correctly classified objects.
f itness = w
1
f
1
+ w
2
f
2
+ w
3
f
3
(4)
where w
1
= 100, w
2
= 1, w
3
= 1 and f
1
the error rate
on training patterns in percentage, f
2
– the number of
active fuzzy rules (rules that classified at least 1 object
correctly), f
3
the length of the rule (the number of
variables other than 0 - don’t care).
The second stage is selection, in which from the
entire population, individuals are selected with the
same probability in an amount equal to the size of the
tournament, that is, the number pre-set by the user,
the number must necessarily be greater than 2 and less
than the population size. After the tournament is per-
formed, the individual with the highest fitness in this
group is selected (Banzhaf et al., 1998).
The selected individuals enter the next stage
crossover. Offspring is the random combination of
the genes (fuzzy rules) of two individuals. Two indi-
viduals are selected with tournament in the algorithm,
and the number of active rules for each of them is cal-
culated. Then an offspring is formed from these indi-
viduals (Poli et al., 2008). The offspring should not
have the number of active rules less than the number
of classes.
The last stage is mutation, where certain parts of
the rule base change in accordance with a given for-
mula:
min(
3
v f
2
, 1), (5)
where v is the number of variables in the dataset.
In the next section the proposed modifications is
described.
4 MODIFICATION
Due to the fact that the rule describes several objects,
the following modification was proposed. Instead of
forming a fuzzy rule on one random object, the use of
several objects of the same class to create a rule was
implemented. To form a rule, a random object a is
first selected, then several more objects b
k
of the same
class are selected, where k = 1...r and r is a number
of neighbors. The search is carried out for the nearest
several objects to the selected one using the Euclidean
distance, as shown in Figure 2. For example, based
on three objects a, b
k1
, b
k2
a fuzzy rule is generated.
The analysis of the best combination of the number of
objects to form the rule was also performed.
Figure 2: Visualization of fuzzy rules generation modifica-
tion.
Hybrid Fuzzy Classification Algorithm with Modifed Initialization and Crossover
227
As for crossover modification, we proposed to
change the choice of rules for the formation of an
offspring. The new method selects the most suitable
rules, then forms an offspring from these rules that
have the largest number of correctly classified objects.
An example is shown in Figure 3.
Figure 3: Visualization of crossover modification.
We also made additional changes in the algorithm.
For instance, we used F-score in fitness of the main
genetic algorithm. As we said before, one of the fea-
tures in (Ishibuchi et al., 2013) is the Michigan-style
part. Depending on the number of active rules of an
individual, the number of rules that will be formed
heuristically or genetically was calculated using for-
mula with one addition: if the value is 1, then the
number of rules generated by the genetic algorithm is
2, heuristic is 2. Heuristically, a rule was formed for
an object that was not correctly classified by any rule.
In addition, we set the minimum threshold parame-
ter for confidence equal to 0.6. In our experiment
we checked the results of weak, average and strong
mutation and we came to the conclusion that average
mutation has better effect on accuracy.
5 RESULTS
Testing of the basic and the modified algorithms was
carried on several tasks taken from the UCI repository
(Asuncion, 2007). The algorithm was run with the
number of individuals set to 100.
In the first experiment with selecting the best num-
ber of objects we used three tasks and 500 genera-
tions. In the second experiment we used four tasks
and we decided to increase the number of generations
to 1000 in order to check whether improvements are
noticeable when the number of generations is dou-
bled.
The 10-fold cross-validation procedure was iter-
ated three times using different data partitions into
ten subsets. Average results over 30 runs are sum-
marized in Table 1, Table 3 and Table 5, NR is the
number of rules. Since the average values are pre-
sented, in order to evaluate the efficiency of the mod-
ification, it is necessary to use the Mann-Whitney U
test (Mann and Whitney, 1947), popular nonparamet-
ric test to compare outcomes between two indepen-
dent groups. Table 2, Table 4 and Table 6 also show
the results of a statistical test accordingly, where the
symbol “=” shows that differences are insignificant,
the symbol “+” shows that differences are significant
and the modification is more efficient and “–” means
that differences are significant and the modification is
worse than the original algorithm.
Table 1: Results on phoneme dataset, first experiment.
Number of Values
objects NR F-score Accuracy
1 10.267 0.722 0.792
2 12.060 0.705 0.783
2 of 5 6.200 0.745 0.801
3 10.367 0.721 0.782
3 of 6 4.367 0.774 0.807
4 of 6 4.900 0.775 0.811
5 of 7 7.300 0.776 0.808
6 10.833 0.709 0.789
Table 2: Results Mann-Whitney U-test on phoneme dataset.
Number of Values
objects NR F-score Accuracy
2 = =
2 of 5 + = =
3 = = =
3 of 6 + + +
4 of 6 + + +
5 of 7 + + +
6 = = =
Table 3: Results on ring dataset, first experiment.
Number of Values
objects NR F-score Accuracy
1 20.107 0.831 0.831
2 20.300 0.825 0.825
2 of 5 20.100 0.832 0.834
3 20.000 0.829 0.829
3 of 6 18.000 0.832 0.833
4 of 6 19.900 0.828 0.829
5 of 7 19.333 0.829 0.829
6 19.533 0.830 0.831
FCTA 2022 - 14th International Conference on Fuzzy Computation Theory and Applications
228
Table 4: Results Mann-Whitney U-test on ring dataset.
Number of Values
objects NR F-score Accuracy
2 = = =
2 of 5 = = =
3 = = =
3 of 6 + = =
4 of 6 + = =
5 of 7 = = =
6 + = =
Table 5: Results on segment dataset, first experiment.
Number of Values
objects NR F-score Accuracy
1 22 0.903 0.903
2 22.567 0.898 0.899
2 of 5 22.600 0.904 0.905
3 21.767 0.903 0.904
3 of 6 17.933 0.905 0.905
4 of 6 22.700 0.907 0.908
5 of 7 21.100 0.903 0.903
6 21.567 0.907 0.907
Based on Tables 1, 3, 5, it can be seen that the
modification with the design of a fuzzy rule on nearby
objects works more efficiently than on randomly se-
lected ones from the same class, regardless of their
number. For the second experiment, it was decided
to use 3 nearest objects selected from 6 objects of the
same class to form the rule as shown in Figure 3. It
can be seen from the results of the modification that
the number of rules decreases, which has a positive
effect on interpretation, because the fewer rules, the
easier the rule base. The results are presented in Ta-
ble 7 and results from Mann-Whitney U-test is in Ta-
ble 8. Comparing the results of the first and second
experiments, it can be seen that the changes are in-
significant, which means there is no need to increase
the number of generations for the data used in the ex-
periment. The algorithm finds a high-quality rule base
in 500 generations, and therefore in less time.
Table 6: Results Mann-Whitney U-test on segment dataset.
Number of Values
objects NR F-score Accuracy
2 = =
2 of 5 = =
3 = = =
3 of 6 + = =
4 of 6 = = =
5 of 7 = = =
6 = = =
Table 7: Results of second experiment.
Number of Values
objects NR F-score Accuracy
Phoneme
1 10.833 0.721 0.789
3 of 6 4.933 0.766 0.825
Ring
1 20.567 0.832 0.832
3 of 6 18.533 0.828 0.828
Satimage
1 22.700 0.806 0.863
3 of 6 16.833 0.837 0.885
Page-blocks
1 9.900 0.435 0.909
3 of 6 9.000 0.440 0.910
Table 8: Results Mann-Whitney U-test of second experi-
ment.
Number of Values
objects NR F-score Accuracy
Phoneme
3 of 6 + + +
Ring
3 of 6 + = =
Satimage
3 of 6 + + +
Page-blocks
3 of 6 + = +
Based on Table 7, it can be noted that the proposed
method of generating fuzzy rules allows the modified
algorithm to significantly reduce the number of rules
and at the same time improve the accuracy for some
data classification problems.
The results of crossover modifications and Mann-
Whitney U test are presented in Table 9 and Table 10.
Table 9: Crossover modification.
Data Values
set NR F-score Accuracy
Phoneme 11.400 0.731 0.786
Ring 23.133 0.833 0.834
Satimage 17.100 0.833 0.858
Based on the results in Tables 9-10, we can con-
clude that crossover modification shows higher effi-
ciency only on one dataset “Satimage”. That means
that there is a possibility that the modification may
work in some cases. The difference between “Satim-
age” and other datasets is the number of classes and
range of numbers. Dataset “Satimage” has posi-
tive numbers, others have both (positive and nega-
tive numbers). The main contributions of this paper
Hybrid Fuzzy Classification Algorithm with Modifed Initialization and Crossover
229
Table 10: Results Mann-Whitney U-test for crossover mod-
ification.
Data Values
set NR F-score Accuracy
Phoneme = = =
Ring = =
Satimage + + +
are: 1) it is shown that initialization on several objects
works better than on one; 2) it is shown that modified
crossing can work better on some data. It is necessary
to conduct additional experiments on other datasets
to find out what conditions influence the efficiency of
modified crossover.
We decided to test both modifications on the same
data in order to check if there will be improvements.
The results are presented in Table 11.
Table 11: Results of both modifications.
Data Values
set NR F-score Accuracy
Phoneme 5.833 0.747 0.797
Ring 23.200 0.833 0.835
Satimage 14.667 0.832 0.861
Segment 21.852 0.899 0.896
As we can see from the results of both modifica-
tions, the differences are insignificant. In future ex-
periments, it will be necessary to try changing the pa-
rameters to test the efficiency.
The modified method was also compared with al-
ternative approaches: Decision Tree (DT), Support
Vector Machine (SVM), Logistic Regression (LR),
Neural Networks (NN). These methods were taken
from sklearn library, the standard parameters were
used. The accuracy are presented at Table 12.
Table 12: Results of alternative approaches.
Data Methods
set DT SVM LR NN
Phoneme 0.770 0.774 0.774 0.738
Ring 0.737 0.726 0.721 0.762
Satimage 0.772 0.869 0.867 0.789
Page-blocks 0.960 0.967 0.967 0.944
Segment 0.541 0.961 0.962 0.913
As can be seen from the table, the proposed modi-
fication with initialization is comparable classification
quality to other known methods. For example, on the
“Satimage” and “Ring” tasks, it proved to be better
than neural networks and on the “Phoneme” task the
proposed modification showed better results compar-
ing to others.
6 CONCLUSIONS
In this paper, the search for better way to find accu-
rate and compact rule bases was investigated. In this
regard, the modification with initialization was pro-
posed in this paper, and showed efficiency on several
datasets, decreasing number of rules and increasing
accuracy and F-score. The design of a rule on several
nearby objects from the same class showed higher ef-
ficiency compared to using a single object. As for the
modification with crossover, the improved efficiency
of this approach was observed only on one dataset.
The proposed method shows similar accuracy com-
pared to alternative methods. In future studies, it is
possible to consider other modifications to improve
the quality of the algorithm. For instance, we want to
implement multi-objective optimization.
ACKNOWLEDGEMENTS
The work was carried out within the framework of the
state support program for leading scientific schools
(grant of the President of the Russian Federation NSh-
421.2022.4).
REFERENCES
Alcal
´
a, R., Gacto, M. J., Herrera, F., and Alcal
´
a-Fdez, J.
(2007). A multi-objective genetic algorithm for tun-
ing and rule selection to obtain accurate and compact
linguistic fuzzy rule-based systems. Int. J. Uncertain.
Fuzziness Knowl. Based Syst., 15:539–557.
Asuncion, A. U. (2007). Uci machine learning repository,
university of california, irvine, school of information
and computer sciences.
Banzhaf, W., Francone, F. D., Keller, R. E., and Nordin, P.
(1998). Genetic programming - an introduction: On
the automatic evolution of computer programs and its
applications.
Bleckmann, C. A. (2006). Evolution and creationism in sci-
ence: 1880–2000.
Bodenhofer, U. and Herrera, F. (1997). Ten lectures on ge-
netic fuzzy systems.
Fern
´
andez, A., Garc
´
ıa, S., Luengo, J., Bernad
´
o-Mansilla,
E., and Herrera, F. (2010). Genetics-based machine
learning for rule induction: State of the art, taxonomy,
and comparative study. IEEE Transactions on Evolu-
tionary Computation, 14:913–941.
Herrera, A. and Magdalena, L. (1997). Genetic fuzzy sys-
tems : A tutorial.
Ishibuchi, H., Mihara, S., and Nojima, Y. (2013). Parallel
distributed hybrid fuzzy gbml models with rule set mi-
gration and training data rotation. IEEE Transactions
on Fuzzy Systems, 21:355–368.
FCTA 2022 - 14th International Conference on Fuzzy Computation Theory and Applications
230
Ishibuchi, H. and Nakashima, T. (2000). Effect of rule
weights in fuzzy rule-based classification systems.
Ninth IEEE International Conference on Fuzzy Sys-
tems. FUZZ- IEEE 2000 (Cat. No.00CH37063), 1:59–
64 vol.1.
Komargodski, I., Naor, M., and Yogev, E. (2017). White-
box vs. black-box complexity of search problems:
Ramsey and graph property testing. 2017 IEEE 58th
Annual Symposium on Foundations of Computer Sci-
ence (FOCS), pages 622–632.
Mamdani, E. H. (1974). Applications of fuzzy algorithms
for control of a simple dynamic plant. Proceedings of
the IEEE.
Mann, H. B. and Whitney, D. R. (1947). On a test of
whether one of two random variables is stochastically
larger than the other. Annals of Mathematical Statis-
tics, 18:50–60.
Mitchell, M. (1996). An introduction to genetic algorithms.
Poli, R., Langdon, W. B., and McPhee, N. F. (2008). A field
guide to genetic programming.
S¸ tefan, R.-M. (2012). A comparison of data classification
methods. Procedia. Economics and finance, 3:420–
425.
Zadeh, L. A. (1965). Fuzzy sets. Inf. Control., 8:338–353.
Hybrid Fuzzy Classification Algorithm with Modifed Initialization and Crossover
231