See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/330957085

Solve Complex Problems using Artificial Neural Network Learned by PSO

Conference Paper · December 2017

CITATION READS
1 2,604

2 authors:

Saeed Shakir Mahmood Gülay Tezel

Selcuk University Selcuk University
1 PUBLICATION 1 CITATION 33 PUBLICATIONS 480 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Saeed Shakir Mahmood on 08 February 2019.

The user has requested enhancement of the downloaded file.

International Conference on Engineering Technologies (ICENTE’17), Dec 07-09, 2017 Konya, Turkey

Solve Complex Problems using Artificial

Neural Network Learned by PSO
SAEED SHAKIR MAHMOOD 1 and GÜLAY.TEZEL2
1
Selcuk University, Konya/Turkey, [email protected]
2
Selcuk University, Konya/Turkey, [email protected]

Abstract - In this work we discuss the problem of classification and intermittent surfaces. Therefore, other kinds of techniques,
of complex data and how to extract preliminary data, which is one such as optimization algorithms, are necessary for training an
of the classification problems faced by many experts in the world ANN.
for decades. Most of the workbooks are designed to learn from the The optimization algorithms have a good admission by the
same data using the training process because no expert in the Artificial Intelligence researchers because they are strong
world can use his or her full experience in determining the optimization tools and can solution complex optimization
parameters of the workbook. This work suggests a mixed problems. For a given problem, NNPSO ability reconnoitering
approach based on an automated learning model. This model large multimodal and non-continuous search extent and can find
integrates a successful detection mechanism that has the ability to the best solving, nearby the optimum value. particle swarm
learn the automatic regulation of the artificial neural network with optimization is based on nature’s behavior describe as swarm
the particle swarm optimization algorithm (PSO). intelligence. This connotation is known in [1] as a possession of
The learning process can also be obtained by the PSO method, systems possessed of unintelligent agents with limited personage
which is one of the hidden layers of the neural network capabilities but with an intelligent mass behavior.
feedforward (NNFF), proving that the record of work done is very There is a lot of research that uses development and optimization
good with quite a few neurons of the hidden layers. In this work, algorithms to train Ann as another basic form of learning [2].
the PSO algorithm is used to determine the best set of parameters Meta-Heuristic methods are based on local research to train
for PSO, and as a result, the number of neurons in the hidden neural networks, population methods, and other collaborative
layers is reduced, thereby improving the overall network examples [3].
performance. The test of the proposed method relies on three sets The researchers have done a good job of showing a broad
of complex data from the machine learning repository and the UCI literature review of developmental algorithms used to develop that
to handle the complex data set. The simulation results show that [2].However, many researchers have relied on the development of
the projected approach can obtain high performance for interlocking weights, parameters [4] Or depends on the evolution
popularization, as compared to the results of other studies. of numbers of hidden layers of neurons, but the number of hidden
layers is created by the designer in advance.
Moreover, researchers have not relied on the development of
Keywords - Artificial neural networks, Hybrid ANN_PSO transfer functions, which are important elements of that which can
,Complex Problem's ,Feedforward neural network , Particle determine the production of all neurons.
Swarm Optimization For example, researchers in [5] suggested a new mechanism
combining the particle squadron to optimize that and improve
I. INTRODUCTION colony ants (ACO) to find a particular structure (s) for that. Using
mixed simulated steel (SA) for ANN’s thresholds and synaptic
Organized neurons in three layers (input layer, output layer and weights to perform modified PSO was done in several researches
hidden layers) represents the Artificial Neural Networks. The such as [6] In [7], the researchers used development software to
artificial neural network is connected to each other by the group obtain architecture and weights to solve classification and
of synaptic weights. Problems such as complex data, Medical prediction problems.
Dataset, pattern recognition, forecasting, and regression has been Another example is [8] where researchers used genetic software to
executed using the ANN. During the learning process, the obtain graphs representing different topologies. In [9], the
interlace values are continuously changed to see enough acquired differential evolution algorithm (DE) was applied to the design
values (until a certain number of repetitions is achieved or until that solved the problem of weather prediction. In [10], the
the value of the target error is achieved). After completion of the researchers used the PSO algorithm to adjust the synaptic weights
training phase, we consider it necessary to assess the for modeling the daily rainfall relationship in Malaysia.
generalization capabilities by using different problem models
that differ from those used in the training phase. The network is In [11], the researchers compared the method of posterior versus
expected to be able to accurately categorize certain types of PSO to adjust only tangled weights to solve classification
problems during the training and testing phase. problems. In [12], the authors worked on the development of a set
of weights using differential evolution and basic PSO.
In recent years several classical algorithms have been proposed In other research such as [13], researchers have worked on the
to train the national network Ann. However, many researchers development of the three main elements of the Artificial Neural
remain trapped in unsolicited solutions; that is, the results network at the same time: architecture, transport functions and
obtained by these researchers may not be good or do not reach synaptic weights.
the level required for the optimal solution or better. Moreover,
these algorithms will be unable to explore multimedia surfaces

1
 the global best solution found so far by any particle in the swarm
2. Materials and methods (population) [16]. Figure 2 shows the updating procedure of a
particle by vectorial representation.
2.1. Neural network learned by backpropagation

In [14] the researchers used Backpropagation to train The algorithm pseudocode is the following:
multilayered neural networks. A multilayered Backpropagation
network consists of an input layer and output layer and at least for each particle do
one hidden layer. The method of the Backpropagation algorithm initialize the particle with random values
is a method of learning under supervision. The aim of learning by end for
Backpropagation is to improve the weights and biases among the
Do
input and output layers based on the error produced by the
network. The input vector is given to the input layer and then for each particle do
passes the hidden layers to the final output layer. In this way, all Calculate fitness value of the particle
the neurons in the network transfer the result to all neurons of the if fitness value of the current particle < fitness value of the
next layer after receiving an arithmetic addition of the weighted pbest particle then
signal from the neurons of the previous layer, Depending on the update the pbest particle
activation function. end if
The ANN's train by backpropagation operates consistently in both end for
forward computing and backward computing, as given in Figure
1, where X1 and X2 are inputs and C1, C2, and C3 are output
gbest = the particle whose fitness value is equal to
vectors of the layers. W1 and W2 are weight matrices; W3 is a min(fitness values of all particles)
weight vector; θ1, θ2, and θ3 are bias vectors; and E1, E2, and E3 for each particle do
bias inputs are chosen as 1. NET1, NET2, and NET3 are net input update velocity and position of the current particle
vectors for the related layer. Sigmoid activation function (φ) is end for
preferred for all neurons. Φ' is the derivative of the activation while stop criterion (maximum generation number or
function. δ1, δ2, and δ3 are local gradient vectors. target fitness value of the gbest particle) is
provided

The 𝑽𝒌𝒊𝒋 and 𝑿𝒌𝒊𝒋 variables in Figure 2 are respectively the j th

velocity component and the jth (j = 1, 2,..., D) position component
of the ith (i = 1, 2, 3, . . . , N) particle at generation k . N is the
population number of particles (10-50) reported as usually
sufficient. D is the dimension size of the search space. Position
figures and tables at the top and bottom of each column if possible.

Figure 1. a) Forward computing schematic structure.

Figure 1.b) backward computing schematic structure (transpose network). Figure 2. The velocity and position updating of a particle .

For the basic PSO [15], calculating the velocity updating and the
2.2. Neural network learned by PSO position updating by Eqs. (1) and (2), are respectively. r1 and r2
In these equations are 2 random numbers, which distributed in the
PSO algorithm, first developed by Kennedy and Earhart in 1995, range between (0,1). c1 personal acceleration coefficient and c2 are
[15] and is considered one of the most exemplary methods based Global acceleration Coefficient, usually c1 + c2 = 4.
on population, and is inspired by fish schools while finding food or
social behavior in flock birds.
The PSO algorithm is initialized with a group of random particles 𝑣𝑖𝑗𝑘+1 = 𝑣𝑖𝑗𝑘 + 𝑐1 × 𝑟1 (𝑝𝑏𝑒𝑠𝑡𝑖𝑗 − 𝑥𝑖𝑗𝑘 ) + 𝑐1 × 𝑟1 × (𝑔𝑏𝑒𝑠𝑡𝑖𝑗 − 𝑥𝑖𝑗𝑘 ) (1)
(candidate solutions for the problem) and then searches for an
optimal solution by updating its individuals. In each generation,
each particle is updated based on 2 special particles: pbest is the
personal best solution of each particle found so far, and gbest is 𝑥𝑖𝑗𝑘+1 = 𝑥𝑖𝑗𝑘 + 𝑣𝑖𝑗𝑘+1 (2)

2
 In an optimized PSO version [17], in the Eq. (3) the inertia
weight (w) parameter has been included the equation of velocity
updating.

𝑘+1 𝑘 𝑘 𝑘
𝑣𝑖𝑗 = 𝑤 𝑘 × 𝑣𝑖𝑗 + 𝑐1 × 𝑟1 (𝑝𝑏𝑒𝑠𝑡𝑖𝑗 − 𝑥𝑖𝑗 ) + 𝑐1 × 𝑟1 × (𝑔𝑏𝑒𝑠𝑡𝑖𝑗 − 𝑥𝑖𝑗 ) (3)

W is used to balance the global and local search [17] and can be
updated using Eq. (4) Or (5) by generations. In Eq. (4), wmax is
maximum of inertia weight and wmin is minimum values of inertia
weight; n is maximum generation number. The 𝜶 & 𝜷 variables in
Eq. (5) are the decrease factors and are used to linearly decrease
inertia weight.

(𝑤𝑚𝑎𝑥 − 𝑤𝑚𝑖𝑛 )
𝑤 𝑘 = 𝑤𝑚𝑎𝑥 − 𝑘 × (4)
𝑛

𝑤 𝑘 = 𝛼 + 𝑟𝑎𝑛𝑑 × 𝛽 (5)

In this work, in order to get an optimized network model and

improve the performance that has been used PSO to train ANN to
ensure good results. During the training phase, the mean squared
error (MSE) is used to calculate the fitness value of a particle (Pi)
by Eq. (6), where e is the error between desired and obtained
outputs after presenting the ith datum to the network, and S is the
number of data in the training dataset. The structure of the Pi
particle is given by Eq. (7). Figure 3. Flowchart for the training and testing of the PSONN.

𝑠
1
𝑀𝑆𝐸 = ∑ 𝑒𝑖2 (6) 3. Experimental Results and Discussion
2𝑆
𝑖=0 The execution of the research ANN_PSO operation has
experimented on three actual benchmark classification problems
(UCI Machine Learning Repository).Out of the data paradigm,
70% is used for training operation and 30% for the testing
operation. The description of these problems is listed in Table 1.
𝑃𝑖 = [𝑤1𝑖11 𝑤1𝑖12 … . 𝜃1𝑖11 … 𝑤2𝑖11 𝑤2𝑖12 … 𝜃2𝑖11 . . 𝑤3𝑖1 … 𝜃3𝑖11 ] (7)
Table 1: Description of datasets.
Dataset Number of Number Number
instances of of classes
features
The flowchart given in Figure 3 [18] shows the training and Pima Indians 768 8 2
testing processes of the PSONN. Diabetes
Artificial neural networks begin the training operation with the Cleveland 296 13 5
initialization random of weights and biases. This indicates the Heart Disease
numerical values associated with the layers of the artificial neural Hepatitis 155 19 2
network. These weights and biases are individuals to each particle,
as given in Eq. (7). The number of connections among layers
indicates to the population size or search space dimension. A stop used in this study. All this data information is reproduced for the
criterion that is selected as the maximum number of generation or benefit of easy reference from [25].
fitness value is the target of the gbest particles as in Figure 3.
Pima Indians Diabetes [25]. The dataset is available at the National
Institute of Diabetes and Digestive and Kidney Diseases. It
contains 768 instances described by eight features used to predict
the presence or absence of diabetes. The features are as follows:
(1) number of pregnancies, (2) plasma glucose concentration, (3)
diastolic blood pressure, (4) tricep skin fold thickness, (5) serum
insulin, (6) body mass index, (7) diabetes pedigree function, and
(8) age in years.

3
Hepatitis [25]. The dataset is obtained from the Carnegie- Mellon Sensitivity (true positive fraction) is the probability that a
University and it contains 155 instances belonging to two classes: diagnostic test is positive, given that the person has the disease:
live or die. There are 19 features (age, sex, steroid, antivirals,
fatigue, malaise, anorexia, big liver, liver film, palpable spleen, 𝑇𝑃
Sensitivity = (8)
𝑇𝑃+𝐹𝑁
spiders, ascites, varices, bilirubin, alk phosphate, SGOT, albumin,
protime, and histology).
Specificity (true negative fraction) is the probability that a
diagnostic test is negative, given that the person does not have the
disease:
Cleveland Heart Disease [25]. The dataset was collected from the
Cleveland Clinic Foundation and contains about 296 instances, 𝑇𝑁
each having 13 features (originally 76 raw features), which are Specificity =
𝑇𝑁+𝐹𝑃
(9)
used to infer the presence (values 1, 2, 3, and 4) or absence (value
0) of heart disease. The features are (1) age, (2) sex, (3) chest pain
type, (4) resting blood pressure, (5) cholesterol, (6) fasting blood Accuracy is the probability that a diagnostic test is correctly
sugar, (7) resting electrocardiographic results, (8) maximum heart performed:
rate, (9) exercise induced angina, (10) depression induced by
exercise relative to segment, (11) slope of peak exercise, (12) (𝑇𝑃+𝑇𝑁)
Accuracy = (10)
(𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁)
number of major vessels, and (13) thal.

The training dataset was used to train the PSONNs. Each network where TP (true positives) is correctly classified positive cases.
consists of an input layer, a hidden layer, and an output layer, as TN (true negative) is correctly classified negative cases.
shown in Figure 4. X1, X2, X3, and X4 are inputs obtained from FP (false positives) is incorrectly classified negative cases.
statistical values as depicted above; Y is the output. The desired FN (false negative) is incorrectly classified positive cases.
output value is 0 for negativity and 1 for positivity. W1 and W2
are connection weight matrices; 𝜽1 and 𝜽𝟐 are bias vectors.
Threshold inputs are used in the layers; their values are chosen as
1. Sigmoid activation function was chosen.

3.1. Pima Indians Diabetes Dataset.

The execution of the various feature chosen propose for the Pima
Indians Diabetes dataset is shown in Table 2. It is noted that when
applying the basic ANN with pso, compared to the accuracy of
other methods and other parameters, the precision value is
88.95%, 99.15 and 100 for sensitivity and privacy with the original
dataset.
The proposed ANN_PSO approach achieved the accuracy rate of
Figure 4. Schematic structure of the neural network..
98.92% with (1,2,6,7,8) selected features, accuracy rate of 99.97%
with (1,2,6,8) selected features and accuracy rate of 99.98% with
(2,6,8) selected features.
In applying the suggest work ANN_PSO algorithm for the
considered medical datasets from the storehouse, the dataset
Table 2: Comparison of Classification results with Pima Indians Diabetes dataset
features are chosen and the parameters, classification accuracy, Methodology adopted Accurac Sensitivi Specificity
sensitivity, and specificity, are notation and scheduling. y ty
Subsequently, in this condition, the execution of the suggested PCA-ANFIS (10x FC) [30] 89.47 70 71.1
method is evaluated in describe of sensitivity, specificity, and LS-ELM (10x FC) [31] 78.21 73.91 80
accuracy. GDA-LS-ELM (10x FC) 79.16 79.1 83.33
[31]
In mostly, sensitivity and specificity are the statistical measures MLNN with LM (10x FC) 79.62 70 70.31
using ultimately to support out the realization of classification [32]
functions. A typical classifier or a predictor will be terminated as PNN (10x FC) [32] 78.05 71 70.5
100% sensitive and 100% specific. Since the proposed application LDA-MWELM [33] 89.74 83.33 93.75
is a set of medical data that is considered complex data, The Mean selection method 76.04 71 78
classification must be exact. Subsequently, sensitivity and [25]
specificity are selected to be the parametric indicator for support
Half selection method [25] 75.91 69 79
of the medical dataset classification.
Neural network for 76.04 71 78
threshold selection [25]
PSO + ELM [34] 91.27 85.26 94.10
SRLPSO + ELM[34] 93.09 91.47 96.29
Proposed ANN_PSO 98,92 100 98

4
3.2. Cleveland Heart Disease Dataset.
Table 4: Comparison of classifcation results with Hepatitis dataset.
Methodology adopted Accuracy( Sensitivity Specificit
The results for the Cleveland Heart Disease dataset can be seen in %) (%) y (%)
Table 3. As a 5-class classification problem is dealt in a state of a Conventional artificial neural 97 92.31 94.5
binary class classification problem, the district of Convergence is network [26]
deleted. It is to be noted that a classification accuracy of 99.61%, Mean selection method [25] 82.58 87 60
sensitivity of 0.98, and a specificity of 0.98 .It should be noted when Half selection method [25] 85.16 90 66
the original features of the dataset were taken into consideration. Neural network for threshold 85.16 90 66
On applying the proposed ANN_PSO approach, the best accuracy selection [25]
of 93.63% is achieved. This accuracy is be realized with only three PSO + ELM [34] 97.43 93.65 95.71
features, Comparison with that of the 13 features of the original SRLPSO + ELM [34] 98.71 94.27 96.04
dataset.
Proposed ANN_PSO 98,44 100 98

The proposed ANN_PSO approach achieved the accuracy rate of

99.33% with (3,9,11,12) selected features, accuracy rate of 99.41%
with (3,8,9,11,12) selected features , accuracy rate of 99.45%
with(3,8,9,10,11,12) selected features and accuracy rate of 99.44%
with(3,8,9,10,11,12,13) selected features. 4. CONCLUSION
In this research a new hybrid algorithm that integrates the
Table 3: Comparison of Classification results with Cleveland Heart Disease dataset. proposed neural network with the particle swarm optimization
Methodology adopted Accuracy Sensitivity Specificity (PSO) for the solution of classification problems is presented. In
(%) (%) (%) order to improve the weight inputs and biases for hidden layer and
C4.5 [27] 81.11 77.23 76.58 the minimum norm least-square scheme for the artificial neural
Naive Bayes [27] 81.48 80.97 81.22 networks, PSO optimization is used to define the analyzed weight
BNND [27] 81.11 82.13 80.42 of output. In order to discover the best algorithm solution, PSO is
BNNF [27] 80.96 76.93 75.81
improved by integrating the mechanism of variation of the particle
AIRS [28] 84.50 75.34 72.96
search method. The effectiveness of the research NN structure by
Hybrid neural network 87.40 93.00 78.50
[30] PSO was bestead than the effectiveness of the other operating
Neural networks 89.01 80.95 95.91 reported in the literature for three benchmark datasets from the
ensemble [29] UCI Machine Learning Repository which are used for estimate.
Mean selection method 81.75 82 82 The results also show that in the proposed framework the number
[25] of neurons in the hidden layer and the relevant input variables
Half selection method 83.44 84 83 does need to be selected by trial-and-error, thus reducing the
[25] network size Improve the ability to generalize.
Neural network for 84.46 82 82
threshold selection [25]
PSO + ELM 89.47 94.49 96.02
SRLPSO + ELM 91.33 95.46 97.29
Proposed ANN_PSO 99,61 98 98

3.3. Hepatitis Dataset.

The results for the Hepatitis dataset are shown in Table 4. From
the results, it is inferred that the proposed ANN_PSO approach
yields a better accuracy of 98.44%, sensitivity 100 %, and
specificity 98 % with that of the original features.

In the proposed ANN_PSO approach achieving, 99.97% was the

accuracy rate with selected features (6, 11, 12, 13, 14, 17, and 19),
98.14% accuracy rate with selected features
(2,5,6,10,11,12,13,14,17,19), 98.18% accuracy rate with selected
features (5,6,11,12,13,14,17,19) and 99.08% accuracy rate with
selected features (11,12,13,14,17,19).

5
REFERENCES [22] M. Karabatak and M. C. Ince, “An expert system for detection
of breast cancer based on association rules and neural network,”
[1] G. Beni and J. Wang, “Swarm intelligence in cellular robotic systems,” in Expert Systems with Applications, vol. 36, no. 2, pp. 3465–3469,
Robots and Biological Systems: Towards a New Bionics? vol. 102 of 2009.
NATO ASI Series, pp. 703–712, Springer, Berlin, Germany, 1993. [23] A. Marcano-Cede˜no, J. Quintanilla-Dom´ınguez, and D. Andina,
[2] X. Yao, “Evolving artificial neural networks,” Proceedings of the IEEE, “WBCD breast cancer database classification applying
vol. 87, no. 9, pp. 1423–1447, 1999. artificial metaplasticity neural network,” Expert Systems with
[3] E. Alba and R. Martí, Metaheuristic Procedures for Training Neural Applications, vol. 38, no. 8, pp. 9573–9579, 2011.
Networks, Operations Research/Computer Science Interfaces Series, [24] P. Jaganathan and R. Kuppuchamy, “A threshold fuzzy entropy
Springer, New York, NY, USA, 2006. based feature selection for medical database classification,” Computers
[4] J. Yu, L. Xi, and S. Wang, “An improved particle swarm optimization for in Biology and Medicine, vol. 43, no. 12, pp. 2222–
evolving feedforward artificial neural networks,” Neural Processing 2229, 2013.
Letters, vol. 26, no. 3, pp. 217–231, 2007. [25] G. Reibnegger, G.Weiss, G.Werner-Felmayer, G. Judmaier, and
[5] M. Conforth and Y. Meng, “Toward evolving neural networks using bio- H.Wachter, “Neural networks as a tool for utilizing laboratory
inspired algorithms,” in IC-AI, H. R. Arabnia and Y. Mun, Eds., pp. 413– information: comparison with linear discriminant analysis and
419, CSREA Press, 2008. with classification and regression trees,” Proceedings of the
[6] Y. Da and G. Xiurun, “An improved PSO-based ANN with simulated National Academy of Sciences of the United States of America,
annealing technique,” Neurocomputing, vol. 63, pp. 527–533, 2005. vol. 88, no. 24, pp. 11426–11430, 1991.
[7] X. Yao and Y. Liu, “A new evolutionary system for evolving artificial [26] K. Polat, S. Sahan, H. Kodaz, and S. G¨unes, “A new classification
neural networks,” IEEE Transactions on Neural Networks, vol. 8, no. 3, method to diagnosis heart disease: supervised artificial immune
pp. 694–713, 1997. system (AIRS),” in Proceedings of the Turkish Symposium on
[8] D. Rivero and D. Periscal, “Evolving graphs for ann development and Artificial Intelligence and Neural Networks (TAINN '05), 2005.
simplification,” in Encyclopedia of Artificial Intelligence, J. R. Rabuñal, [27] R. Das, I. Turkoglu, and A. Sengur, “Effective diagnosis of heart
J. Dorado, and A. Pazos, Eds., pp. 618–624, IGI Global, 2009. disease through neural networks ensembles,” Expert Systems
[9] H. M. Abdul-Kader, “Neural networks training based on differential with Applications, vol. 36, no. 4, pp. 7675–7680, 2009.
evolution algorithm compared with other architectures for weather [28] H. Kahramanli and N. Allahverdi, “Design of a hybrid system
forecasting34,” International Journal of Computer Science and Network for the diabetes and heart diseases,” Expert Systems with
Security, vol. 9, no. 3, pp. 92–99, 2009. Applications, vol. 35, no. 1-2, pp. 82–89, 2008.
[10] K. K. Kuok, S. Harun, and S. M. Shamsuddin, “Particle swarm [29] N. Cheung, Machine learning techniques for medical analysis
optimization feedforward neural network for modeling [M.S. thesis], University of Queensland, 2001.
runoff,” International Journal of Environmental Science and [30] K. Polat and S. G¨unes¸, “An expert system approach based
Technology, vol. 7, no. 1, pp. 67–78, 2010. on principal component analysis and adaptive neuro-fuzzy
[11] B. A. Garro, H. Sossa, and R. A. Vázquez, “Back-propagation vs inference system to diagnosis of diabetes disease,” Digital Signal
particle swarm optimization algorithm: which algorithm is better to Processing, vol. 17, no. 4, pp. 702–710, 2007.
adjust the synaptic weights of a feed-forward ANN?” International
Journal of Artificial Intelligence, vol. 7, no. 11, pp. 208–218, 2011. [31] K. Polat, S.G¨unes¸, andA.Arslan, “Acascade learning systemfor
[12] B. Garro, H. Sossa, and R. Vazquez, “Evolving neural networks: a classification of diabetes disease: generalized discriminant analysis
comparison between differential evolution and particle swarm and least square support vector machine,” Expert Systems
optimization,” in Advances in Swarm Intelligence, Y. Tan, Y. Shi, Y. with Applications, vol. 34, no. 1, pp. 482–487, 2008.
Chai, and G. Wang, Eds., vol. 6728 of Lecture Notes in Computer [32] H. Temurtas, N. Yumusak, and F. Temurtas, “A comparative
Science, pp. 447–454, Springer, Berlin, Germany, 2011. study on diabetes disease diagnosis using neural networks,”
[13] B. A. Garro, H. Sossa, and R. A. Vazquez, “Design of artificial neural Expert Systems with Applications, vol. 36, no. 4, pp. 8610–8615,
networks using a modified particle swarm optimization algorithm,” 2009.
in Proceedings of the International Joint Conference on Neural Networks [33] D. C¸alis¸ir and E. Doˇgantekin, “An automatic diabetes diagnosis
(IJCNN '09), pp. 938–945, IEEE, Atlanta, Ga, USA, June 2009. system based on LDA-wavelet support vector machine classifier,”
[14] J. Kennedy, R.C. Eberhart, \Particle swarm optimization", Proceedings Expert Systems with Applications, vol. 38, no. 7, pp. 8311–
of the IEEE International Conference on Neural Networks, Vol. 4, pp. 8315, 2011.
1942{1948, 1995. [34] S. N. Sivanandamand S. N. Deepa, “ A Machine Learning Paradigm
[15] S. Bağış, \System modelling by using artificial intelligence Integrating Particle Swarm Optimization with Extreme Learning
algorithms", MSc, Erciyes University, Kayseri, Turkey,2009 (in Turkish Machine Classifier,” Department of EEE, Anna University Regional
with English abstract). Centre, Coimbatore, Coimbatore 641 047, 2015.
[16] L. Liu, W. Liu, D.A. Cartes, \Particle swarm optimization based
parameter identi_cation applied to permanent [1] magnet synchronous
motors", Engineering Applications of Arti_cial Intelligence, Vol. 21, pp.
1092{1100, 2008.
[17] Y. Shi, R.C. Eberhart, \Empirical study of particle swarm optimization",
Proceedings of the IEEE Congress on Evolutionary Computation, pp.
1945{1950, 1999.
[18] Y. Shi, R.C. Eberhart, \Empirical study of particle swarm optimization",
Proceedings of the IEEE Congress on Evolutionary Computation, pp.
1945{1950, 1999.
D. E.Goodman Jr., L. C. Boggess, andA. B.Watkins, “Artificial immune
system classification of multiple-class problems,” in Proceedings of the
Artificial Neutral Networks in Engineering Conference (ANNIE ’02),
pp. 179–184, November 2002.
[19] J. Abonyi and F. Szeifert, “Supervised fuzzy clustering for the
identification of fuzzy classifiers,” Pattern Recognition Letters,
vol. 24, no. 14, pp. 2195–2207, 2003.
[20] S. S¸ahan, K. Polat, H. Kodaz, and S. G¨unes¸, “A new hybrid
method based on fuzzy-artificial immune system and k-nn
algorithm for breast cancer diagnosis,” Computers in Biology
and Medicine, vol. 37, no. 3, pp. 415–423, 2007.
[21] M. F. Akay, “Support vector machines combined with feature
selection for breast cancer diagnosis,” Expert Systems with
Applications, vol. 36, no. 2, pp. 3240–3247, 2009.

View publication stats