IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 14, No. 1, February 2025, pp. 547~560

ISSN: 2252-8938, DOI: 10.11591/ijai.v14.i1.pp547-560  547

Hyperparameter optimization of convolutional neural network

using particle swarm optimization for emotion recognition

Dian Palupi Rini1, Tri Kurniasari2, Winda Kurniasari3, Novi Yusliani1

1
Department of Informatics Engineering, Faculty of Computer Science, Universitas Sriwijaya, Palembang, Indonesia
2
Department of Information Technology, Padang Selasa Public Heatlh Center, Palembang, Indonesia
3
Department of Information Systems, Faculty of Computer Science, Universitas Sriwijaya, Palembang, Indonesia

Article Info ABSTRACT

Article history: Emotion identification has been widely researched based on facial
expressions, voice, and body movements. Several studies on emotion
Received Feb 3, 2024 recognition have also been performed using electroencephalography (EEG)
Revised Jul 9, 2024 signals and the results also show that the technique has a high level of
Accepted Jul 26, 2024 accuracy. EEG signals that detected by standart method using exclusive
representations of time and frequency domains presented unefficient results.
Some researchers using the convolutional neural network (CNN) method
Keywords: performed EEG signal for emotional recognition and obtained the best results
in almost all benchmarks. Although CNN has shown fairly high accuracy,
Convolutional neural network there is still a lot of room for improvement. CNN is sensitive to its
Particle swarm optimization hyperparameter value because it has considerable effect on the behavior and
Electroencephalography signal efficiency of the CNN architecture. So that the use of optimization algorithms
Emotion is expected to provide an alternative selection of appropriate hyper parameter
Hybrid values on CNN. Particle swarm optimization (PSO) algorithm is a
metaheuristic-based optimization algorithm with many advantages. This PSO
algorithm was chosen to optimize the hyperparameter values on CNN. Based
on the evaluation results in each model, hybrid CNN-PSO showed better
results and achieved the best value in 80:20 split data which is 99.30%
accuracy.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Dian Palupi Rini
Department of Informatics Engineering, Faculty of Computer Science, Universitas Sriwijaya
St. Palembang-Prabumulih KM.32, Ogan Ilir, Palembang, Indonesia
Email: [email protected]

1. INTRODUCTION
Research on emotion recognition has also been performed using electroencephalography (EEG)
signals and the results show that the technique has a high level of accuracy [1]. Through EEG, information
about mental activity and human emotional states can be known and the information produces different values
when emotional states change. Therefore, the use of EEG is considered effective for emotional identification
or prediction [2].
Traditional methods of EEG signal classification, which use separate representations for time and
frequency domains, often yield unsatisfactory results. In this approach, EEG signals are decomposed into
time-frequency representations using discrete wavelet transform (DWT). Statistical features are then calculated
to represent the distribution of the signals [3], [4]. Research on EEG signal recognition using end-to-end models
based on convolutional neural networks (CNNs) has been conducted and obtained good results in almost all
classification benchmarks [5], [6]. The EEG signal representation problem in the database for emotion analysis

Journal homepage: http://ijai.iaescore.com

548  ISSN: 2252-8938

using physiological (DEAP) dataset shows that the proposed method achieves an accuracy of 77.98% on
valence recognition and 72.98% on arousal recognition [6].
Although CNN standards have shown considerable accuracy, there is still a lot of room for
improvement [7]. CNN performance cannot be expected by applying the same architecture to different types
of tasks, it needs to be adjusted the architecture for specific tasks so as to produce better performance [8]. CNN
also requires a large number of samples for the training phase. In addition, CNN has many hyperparameters
[9] and a wide variety of architectures that are considered a challenge and it is difficult to determine the best
value of hyperparameters manually. CNN is sensitive to its hyperparameter value because it has a considerable
effect on the behavior and efficiency of the CNN architecture [10]. To attain hyperparameters with better
performance, experts must manually configure a set of hyperparameter options. However, different datasets
require different models or hyperparameter combinations, making it complicated [11].
Research on efficient optimization of hyperparameter values in CNNs using particle swarm
optimization (PSO) resulted in an increase in image recognition accuracy by 0.7%-5.7% using the
Alexnet-CNN standard [9]. Research on the use of PSO on CNN, which is one of the basic methods in deep
learning has been conducted by several researchers [7], [12], [13]. The use of PSO in the training process aims
to optimize the results of the solution vector on CNN in order to improve recognition accuracy. Experiments
show that the accuracy that can be achieved in 4 epochs is 95.08%. These results are better than conventional
CNNs and deep belief networks (DBNs). Its execution time is also similar to conventional CNN. This study
proposes a CNN and PSO deep learning approach to optimize multiple hyperparameters, resulting in higher
results [14].
Therefore, based on previous research, researchers will test the CNN method for EEG signal
classification to detect types of emotions. Then optimization will be carried out with the PSO method on the
initialization of learning rate, epoch, and batch size to improve the accuracy of signal recognition in this study.
Next will be made a comparison with the CNN algorithm itself without the optimization process from the PSO.

2. METHOD
2.1. Hybrid modeling of convolutional neural network–particle swarm optimization
The hybrid CNN-PSO algorithm is built to provide many alternatives in determining the value of
CNN hyperparameters using the PSO algorithm, so that it is expected to avoid local optimal solutions. CNNs
take their name from mathematical linear operations between matrices called convolutions. The architecture of
CNN that shown in Figure 1 has many layers, including a convolutional layer, a non-linear layer, a pooling
layer and a fully-connected layer. Convolutional and fully-connected layers have parameters but pooling and
non-linear layers have no parameters.

Figure 1. CNN architecture

PSO is a method inspired by the nature of a group of animals such as birds, termites, bees, or ants.
Three crucial aspects of PSO include particles, cognitive and social components, and particle velocity. Each
particle represents the solution to a given problem. Particle learning includes two elements, which are particle
experience (cognitive learning) and the combination of whole swarm learning (social learning) [15], [16].
Cognitive learning in the form of pBest is the best position a particle can ever reach, while social learning in
the form of gBest is the best position of all particles in a swarm. The parameters pBest and gBest are used to
calculate the speed of particle (1) as well as the speed of calculating the position (2) of the next particle.
Speed update:

𝑛𝑒𝑤 𝑜𝑙𝑑 𝑙𝑜𝑐𝑎𝑙 𝑏𝑒𝑠𝑡 𝑜𝑙𝑑 𝑔𝑙𝑜𝑏𝑎𝑙 𝑏𝑒𝑠𝑡 𝑜𝑙𝑑

𝑉𝑖,𝑚 = 𝑊. 𝑉𝑖,𝑚 + 𝐶1 × 𝑅 × (𝑃𝑖,𝑚 − 𝑋𝑖,𝑚 ) + 𝐶2 × 𝑅 × (𝑃𝑚 − 𝑋𝑖,𝑚 ) (1)

Position update:
𝑛𝑒𝑤 𝑜𝑙𝑑 𝑛𝑒𝑤
𝑋𝑖,𝑚 = 𝑋𝑖,𝑚 + 𝑉𝑖,𝑚 (2)

Int J Artif Intell, Vol. 14, No. 1, February 2025: 547-560

Int J Artif Intell ISSN: 2252-8938  549

Where 𝑉𝑖,𝑚 is speed of the ith particle at the ith iteration, W is the inertial weight coefficient, 𝐶1 , 𝐶2 are the
acceleration constants (learning rate), 𝑅 is a random number (0-1), 𝑋𝑖,𝑚 is the current position of the i-th particle
in the i-th iteration, 𝑃𝑏𝑒𝑠𝑡𝑖 is the previous best position of the i-th particle, and 𝑃𝑔𝑏𝑒𝑠𝑡 is the best particle
among all particles in a group or population. The optimization process is carried out by providing alternative
hyperparameter values on CNN a number of particles in the PSO algorithm. It then makes the output value of
CNN an objective function of the PSO, by finding the maximum value of several alternative processes in each
article [17].

2.2. Dataset
For experimental purposes, the dataset used was EEG data that focused on capturing brainwave
patterns associated with various emotions and feelings. This dataset consists of 2,549 variables, of which 2,548
variables contain decimal data, while 1 other variable contains data in the form of strings used as labels. For
three minutes under each state (positive, neutral, and negative), data was recorded from two participants,
consisting of one male and one female. This data collection process was carried out using an EEG muse
headband equipped with dry electrodes to record EEG activity at TP9, AF7, AF8, and TP10 location points
[18], [19]. The dataset is shown in Table 1, with the following content specifications:

Table 1. Dataset specifications

Data label Number of data
Positive 708
Negative 708
Neutral 716
Total 2,132

In the context of 2,549 variables, this study applies an even division to form two datasets, each
consisting of 1,275 variables. Of these, 1,274 variables will contain decimal data, while 1 variable will serve
as a label. The next step is to train both datasets that have been formed. The first dataset will be named
"eeg-emotion-1," while the second dataset will be named "eeg-emotion-2." This process is designed to create
two similar but independent subsets of data for training purposes.

2.3. Pre-processing
The data pre-processing that has been done in this study includes data transformation and
normalization, preparing it for data analysis and deep learning implementation. In the context of using machine
learning algorithms, categorical data cannot be directly processed by those algorithms. Therefore, categorical
data needs to be converted into numerical form before it can be used in the analysis process [20], [21]. This
research involves using label encoding to convert class variables from string data to numeric values. This
transformation process maps three classes of strings, namely neutral to class 0, positive to class 1, and negative
to class 2. Next, StandardScaler is applied as a normalization method to the dataset. This process aims to change
each feature (column) in the dataset so that it has a mean of zero and a standard deviation of one.

2.4. Convolutional neural network hybrid modeling–particle swarm optimization

The process of modeling data in deep learning will be depends on selecting the right model
architecture, building a model structure with appropriate layers, initializing parameters, selecting loss functions
and optimizers, as well as training the model using training data by updating parameters through optimization
algorithms. After training, the model is evaluated using validation data to check its performance, and
hyperparameters can be reset to improve results [19], [22]. Final testing is done using unseen test data.
Figure 2 is the complete process of CNN's PSO.
CNN is very effective in understanding visual data, recognizing patterns, as well as extracting features
hierarchically [23]–[25]. CNN architecture is a combination of three components; Convolutional layer, which
contains a number of filters to apply to the input. Each filter scans the input using point products and delivery
methods to generate a number of feature maps in a single convolutional layer [4].
In this study, a one-dimensional CNN architecture was used that has a specific configuration to
process data. The architecture consists of a convolution layer with 16 filters, each measuring 10, followed by
a maxpooling layer with size 2. The hidden layer of the CNN has 100 neurons and uses the activation function
of hyperbolic tangent (tanh). The importance of selecting this activation function is proven by experiments,
where the use of the rectified linear unit (ReLU) activation function results in unsatisfactory performance. In
this study, PSO was formed with a configuration of 3 particles and 5 dimensions. In this scenario, the PSO

Hyperparameter optimization of convolutional neural network using particle swarm … (Dian Palupi Rini)
550  ISSN: 2252-8938

algorithm will use 3 particles as search agents to explore a search space that has 5 dimensions, which includes
hyperparameter variables to be optimized.

Figure 2. Detail process of CNN–PSO

2.5. Hyperparameter optimizer validation

In this study, the Optunity library was used to perform hyperparameter optimization on a model.
Libraries such as optunity and optunity. metrics are imported, as well as the particle swarm solver module from
the optunitysolvers library. The search dictionary is used to define the range of hyperparameter values to be
explored in the optimization process. Then, the performance function is created with three hyperparameters:
lrate, epochs, and batchsize.
The validation process of hyperparameter optimization is carried out by considering the range of
values on several key parameters, including learning rates ranging from 0.0001 to 0.01, epoch numbers ranging
from 1 to 100, and batch sizes ranging from 0 to 1. For batch sizes less than 1, a value of 32 is used, a batch
size less than 2 uses a value of 64, and batch sizes not included in both use a value of 128. In addition, there
are other parameters such as the number of particles is 3, PSO dimensions are 5, and the number of evaluations
is 10.

3. RESULTS AND DISCUSSION

This research involved training ten models, namely five models (models 1-5) on the eeg-emotion-1
dataset and five models (models 6-10) on the eeg-emotion-2 dataset. Each model is trained and evaluated
according to the configuration described earlier. Table 2 shows the overall experimental results of the confusion
matrix for the eeg-emotion-1 dataset using different data sharing. Meanwhile, Table 3 shows the results of

Int J Artif Intell, Vol. 14, No. 1, February 2025: 547-560

Int J Artif Intell ISSN: 2252-8938  551

confusion matrix experiments for the eeg-emotion-2 dataset using different data divisions. It can be seen that
90:10 data division gives the lowest false positive and false negative values.

Table 2. Comparison of confusion matrix eeg-emotion-1 classification results

Ratio data TP FP FN TN
90:10 212 2 2 424
80:20 415 12 12 830
70:30 621 21 21 1,242
60:40 830 23 23 1,660
50:50 1,033 33 33 2,066

Table 3. Comparison of confusion matrix eeg-emotion-2 classification results

Ratio data TP FP FN TN
90:10 208 6 6 416
80:20 416 11 11 832
70:30 624 16 16 1,248
60:40 837 16 16 1,674
50:50 1,037 29 29 2,074

3.1. Model 1
The ratio of datasets used in model 1 is 90:10. The training accuracy obtained in model 1 is 99.07.
After the data was normalized, the data was trained by adding a PSO algorithm and obtained an accuracy of
98.60. In the case of model 1, CNN coupled with PSO does not affect the accuracy increase.To see the
comparison of training and testing per epoch in the CNN accuracy curve, it can be seen in Figure 3, while the
loss curve is in Figure 4. Then, the accuracy and loss curves of both training and testing for CNN+PSO can be
observed in Figures 5 and 6. Although the accuracy of the CNN model reached 99%, the performance curve
showed a less significant improvement, as well as CNN+PSO which showed indications of overfitting on the
loss curve, despite obtaining high accuracy [12].

Figure 3. CNN accuracy curve Figure 4. CNN loss curve

Figure 5. CNN+PSO accuracy curve Figure 6. CNN+PSO loss curve

Hyperparameter optimization of convolutional neural network using particle swarm … (Dian Palupi Rini)
552  ISSN: 2252-8938

3.2. Model 2
In model 2, training and testing are performed by dividing the data in an 80:20 ratio. The results of
the performance evaluation of the CNN in this model showed an accuracy of 97.19%. After going through the
data normalization stage, we decided to involve a PSO algorithm to improve performance. The increase
resulted in an accuracy of 98.36%. Figures 7 and 8 show the accuracy and loss curves on the CNN, while
Figures 9 and 10 represent the accuracy and loss curves on the optimized model using the PSO algorithm. The
curve produced by the second model shows less satisfactory quality compared to the first model, with
indications of overfitting.

Figure 7. CNN accuracy curve Figure 8. CNN loss curve

Figure 9. CNN+PSO accuracy curve Figure 10. CNN+PSO loss curve

3.3. Model 3
In the third model, we repeated similar steps to the previous model, with a 70:30 variation in data
comparison. This model achieved an accuracy of 97.03%, which did not surpass the performance of the two
previous models. The accuracy and loss curves can be seen in Figures 11 and 12. After going through the
normalization stage, we continued CNN training with the addition of a PSO algorithm for comparison of
results. Finally, the accuracy achieved by the CNN+PSO model is 97.19%. In addition, the accuracy and loss
curves are also formed which are seen in Figures 13 and 14. The performance of the third model still shows
unsatisfactory quality due to overfitting detected, both on the CNN architecture and models that have been
improved using the CNN+PSO algorithm.

3.4. Model 4
Furthermore, the fourth model involved training the eeg-emotion-1 dataset with a data comparison of
60:40. Confusion matrix is used to calculate precision, recall, and F1-score with an accuracy of 97.30%.
Information on the accuracy and loss curves of the CNN model is presented in Figures 15 and 16. Furthermore,
this model gets additional training with the application of PSO algorithms to improve accuracy. This results in
an accuracy of 97.66%. The accuracy and loss curves for the CNN+PSO model can be observed in
Figures 17 and 18. On the loss chart of the fourth model, train loss shows a positive increase by continuing to
decrease near 0. However, this does not apply to loss validation which actually shows an upward trend, even
though it remains below the value of 1, resulting in overfitting.

Int J Artif Intell, Vol. 14, No. 1, February 2025: 547-560

Int J Artif Intell ISSN: 2252-8938  553

Figure 11. CNN accuracy curve Figure 12. CNN loss curve

Figure 13. CNN+PSO accuracy curve Figure 14. CNN+PSO loss curve

Figure 15. CNN accuracy curve Figure 16. CNN loss curve

Figure 17. CNN+PSO accuracy curve Figure 18. CNN+PSO loss curve

Hyperparameter optimization of convolutional neural network using particle swarm … (Dian Palupi Rini)
554  ISSN: 2252-8938

3.5. Model 5
In the fifth model, the data was trained and tested at a ratio of 50:50. By still using the same
hyperparameters, the accuracy obtained reaches 96.90%. However, the results from the fifth model showed the
lowest accuracy when compared to the four previous models using the eeg-emotion-1 dataset. In addition, an
overview of the accuracy and loss curves of the CNN model can be found in Figures 19 and 20. Furthermore,
this model gets improved performance through retraining by integrating the PSO algorithm. As a result,
accuracy increased by 0.66% to reach 97.56%. Information on the accuracy and loss curves of the CNN+PSO
model can be found in Figures 21 and 22. The performance of this model is far behind compared to the other
four models, this can be seen from the CNN+PSO loss curve which has a significant difference. Although both
train loss and validation loss remain below 0, the results obtained are still unsatisfactory.

Figure 19. CNN accuracy curve Figure 20. CNN loss curve

Figure 21. CNN+PSO accuracy curve Figure 22. CNN+PSO loss curve

3.6. Model 6
In the sixth model, we trained the eeg-emotion-2 dataset by dividing the data in a 90:10 ratio. By
adopting the same parameters as the previous model, the accuracy of CNN on this model reaches 97.20%. The
next step involves training the data with the application of the PSO algorithm, which results in an accuracy of
97.66%. This suggests that the use of PSOs can slightly improve accuracy, although not significantly.
Comparison of accuracy and loss curves in the CNN model is shown in Figures 23 and 24. Meanwhile, the
accuracy and loss curves in the CNN+PSO model are seen in Figures 25 and 26. When compared with the
training results of the eeg-emotion-1 dataset using the same data comparison and using both CNN and
CNN+PSO architectures, the sixth model resulted in lower accuracy and variable loss values, although still
below the value of 0.

3.7. Model 7
In the seventh model, data training was carried out by dividing the dataset in an 80:20 ratio. As a result,
the accuracy achieved by this model reaches 97.42%. By involving the PSO algorithm in the next stage of
training, accuracy was successfully increased to 99.30%. Comparison with the second model, which involves
sharing the same data and identical methods, shows improved performance on the seventh model. However, the
resulting performance curve is still not optimal. Details on the CNN accuracy and loss curves are shown in
Figures 27 and 28, while the CNN+PSO accuracy and loss curves are shown in Figures 29 and 30.

Int J Artif Intell, Vol. 14, No. 1, February 2025: 547-560

Int J Artif Intell ISSN: 2252-8938  555

Figure 23. CNN accuracy curve Figure 24. CNN loss curve

Figure 25. CNN+PSO accuracy curve Figure 26. CNN+PSO loss curve

Figure 27. CNN accuracy curve Figure 28. CNN loss curve

Figure 29. CNN+PSO accuracy curve Figure 30. CNN+PSO loss curve

3.8. Model 8
Performance results from the confusion matrix that will be used to calculate precision, recall, and
F1-score with an accuracy of up to 97.50%. After the implementation of the PSO algorithm, the accuracy was

Hyperparameter optimization of convolutional neural network using particle swarm … (Dian Palupi Rini)
556  ISSN: 2252-8938

successfully increased to 98.12%. Information on the comparison of training and testing per epoch in the
accuracy curve of the CNN model is presented in Figure 31, while the loss curve is shown in Figure 32.
Furthermore, comparison of accuracy curves and losses of training and testing of CNN+PSO models can be
observed in Figures 33 and 34. It can be seen that the loss curve in the CNN+PSO model shows suboptimal
results, which may be caused by a decrease in accuracy at the validation stage that causes overfitting.

Figure 31. CNN accuracy curve Figure 32. CNN loss curve

Figure 33. CNN+PSO accuracy curve Figure 34. CNN+PSO loss curve

3.9. Model 9
In the ninth model, the ratio of datasets used is 60:40. The training accuracy of the CNN model reached
98.12%. After normalizing the data, the model was trained with the PSO optimization algorithm involved, but
this did not result in a significant increase in accuracy, only about 98.01%. Details about the accuracy and loss
curves of the CNN model are seen in Figures 35 and 36. The results of the CNN+PSO model also illustrate a
similar situation, where the accuracy curve at the validation stage decreases compared to the training accuracy,
as a result of the validation loss value continues to increase. This information is shown in Figures 37 and 38.

Figure 35. CNN accuracy curve Figure 36. CNN loss curve

Int J Artif Intell, Vol. 14, No. 1, February 2025: 547-560

Int J Artif Intell ISSN: 2252-8938  557

Figure 37. CNN+PSO accuracy curve Figure 38. CNN+PSO loss curve

3.10. Model 10
The tenth model, which was the last model to be trained and tested using 50:50 data sharing from the
eeg-emotion-2 dataset, achieved an accuracy of 97.28%. As with previous models, it also faces challenges in
the form of suboptimal performance curves. A decrease in validation accuracy and an increase in validation
loss value also affect the performance of this model.
Information about the accuracy and loss curves of the CNN model is shown in Figures 39 and 40.
Similarly, the accuracy curve shown in Figure 41 and loss in the CNN+PSO model reflect a similar situation,
almost identical to the fifth model tested with the same data sharing. In both cases, there was an increase in
validation loss which resulted in an overfitting tendency as shown in Figure 42.

Figure 39. CNN accuracy curve Figure 40. CNN loss curve

Figure 41. CNN+PSO accuracy curve Figure 42. CNN+PSO loss curve

Hyperparameter optimization of convolutional neural network using particle swarm … (Dian Palupi Rini)
558  ISSN: 2252-8938

Results from all ten CNN and CNN+PSO models that have been trained and tested, documented and
compared in Table 4. In the case of the eeg-emotion-1 dataset, the first model with 90:10 data sharing using
CNN obtained the highest accuracy of 99.07%. Meanwhile, training and testing using the eeg-emotion-2
dataset, the seventh model involving the PSO algorithm, achieved a peak accuracy of 99.30%. However, when
evaluating the quality of the performance curve, models trained and tested using CNN produced better curves.

Table 4. Comparison of accuracy results

Model Metode Accuracy (%)
1 CNN 99.07
CNN+PSO 98.60
2 CNN 97.19
CNN+PSO 98.36
3 CNN 97.03
CNN+PSO 97.19
4 CNN 97.30
CNN+PSO 97.66
5 CNN 96.90
CNN+PSO 97.56
6 CNN 97.20
CNN+PSO 97.66
7 CNN 97.42
CNN+PSO 99.30
8 CNN 97.50
CNN+PSO 98.12
9 CNN 98.12
CNN+PSO 98.01
10 CNN 97.28
CNN+PSO 97.37

We compared our model with two other models which also use eeg based brainwave dataset. The first
model explores single and ensemble methods for classifying emotions by reducing them to smaller datasets
through feature selection using OneR scores, Bayes network, information gain, and symmetrical uncertainty.
They achieved an overall accuracy of about 97.89% [19] and the second model used an applied evolutionary
algorithm to select the most informative features from the initial feature set. Multilayer perceptron (MLP)
optimization is performed using an evolutionary approach before classification to estimate the best
hyperparameters of the network. Deep learning and adaptation using long short-term memory (LSTM) are also
explored, and the adaptive boosting of both types of models is examined for each problem. The obtained results
show that an adaptive boosted LSTM can achieve an accuracy of 84.44%, 97.06%, and 9.94% attention,
emotion, and numerical datasets, respectively [26]. Compared with the existing techniques, the proposed
techniques achieved a higher level of accuracy of 98% by applying the hybrid CNN with PSO.

4. CONCLUSION
The main focus in this study was to classify EEG signal data for emotion detection using CNN
algorithms. To enhanced the CNN performance, the hyper parameters in the CNN algorithm are optimized
with the PSO algorithm. Based on the evaluation results in each model, hybrid CNN-PSO showed better results
an accuracy rate of 99.30% compared to CNN. The comparison between the existing algorithms that used the
same data (i.e OneR, Bayes network, information gain, and an adaptive boosted LSTM) shows that hybrid
CNN-PSO has better performance. However, although hybrid CNN PSO provides better results, the quality of
the performance curve CNN produces a better result. So, for further research is still needed to improve the
performance of machine learning by tuning the hyper parameter or using another optimization algorithm.

REFERENCES
[1] J. Yang, X. Huang, H. Wu, and X. Yang, “EEG-based emotion classification based on bidirectional long short-term memory
network,” Procedia Computer Science, vol. 174, pp. 491–504, 2020, doi: 10.1016/j.procs.2020.06.117.
[2] S. M. Alarcao and M. J. Fonseca, “Emotions recognition using EEG signals: a survey,” IEEE Transactions on Affective Computing,
vol. 10, no. 3, pp. 374-393, 2019, doi: 10.1109/TAFFC.2017.2714671.
[3] A. Subasi and M. I. Gursoy, “EEG signal classification using PCA, ICA, LDA and support vector machines,” Expert Systems with
Applications, vol. 37, no. 12, pp. 8659–8666, 2010, doi: 10.1016/j.eswa.2010.06.065.
[4] Mustaqeem and S. Kwon, “A CNN-assisted enhanced audio signal processing for speech emotion recognition,” Sensors, vol. 20,
no. 1, 2020, doi: 10.3390/s20010183.

Int J Artif Intell, Vol. 14, No. 1, February 2025: 547-560

Int J Artif Intell ISSN: 2252-8938  559

[5] F. E. F. Junior and G. G. Yen, “Particle swarm optimization of deep neural networks architectures for image classification,” Swarm
and Evolutionary Computation, vol. 49, pp. 62–74, 2019, doi: 10.1016/j.swevo.2019.05.010.
[6] Z. Wen, R. Xu, and J. Du, “A novel convolutional neural networks for emotion recognition based on EEG signal,” 2017
International Conference on Security, Pattern Analysis, and Cybernetics, SPAC 2017, vol. 2018, pp. 672–677, 2017, doi:
10.1109/SPAC.2017.8304360.
[7] S. B. Sulistyo, W. L. Woo, and S. S. Dlay, “Regularized neural networks fusion and genetic algorithm based on-field nitrogen status
estimation of wheat plants,” IEEE Transactions on Industrial Informatics, vol. 13, no. 1, pp. 103–114, 2017, doi:
10.1109/TII.2016.2628439.
[8] M. Zhou et al., “Epileptic seizure detection based on EEG signals and CNN,” Frontiers in Neuroinformatics, vol. 12, 2018, doi:
10.3389/fninf.2018.00095.
[9] T. Yamasaki, T. Honma, and K. Aizawa, “Efficient optimization of convolutional neural networks using particle swarm
optimization,” Proceedings - 2017 IEEE 3rd International Conference on Multimedia Big Data, BigMM 2017, pp. 70–73, 2017,
doi: 10.1109/BigMM.2017.69.
[10] A. Darwish, D. Ezzat, and A. E. Hassanien, “An optimized model based on convolutional neural networks and orthogonal learning
particle swarm optimization algorithm for plant diseases diagnosis,” Swarm and Evolutionary Computation, vol. 52, 2020, doi:
10.1016/j.swevo.2019.100616.
[11] N. M. Aszemi and P. D. D. Dominic, “Hyperparameter optimization in convolutional neural network using genetic algorithms,”
International Journal of Advanced Computer Science and Applications, vol. 10, no. 6, pp. 269–278, 2019, doi:
10.14569/ijacsa.2019.0100638.
[12] V. Passricha and R. K. Aggarwal, “PSO-based optimized CNN for Hindi ASR,” International Journal of Speech Technology, vol.
22, no. 4, pp. 1123–1133, 2019, doi: 10.1007/s10772-019-09652-3.
[13] Y. Wang, H. Zhang, and G. Zhang, “cPSO-CNN: an efficient PSO-based algorithm for fine-tuning hyper-parameters of
convolutional neural networks,” Swarm and Evolutionary Computation, vol. 49, pp. 114–123, 2019, doi:
10.1016/j.swevo.2019.06.002.
[14] G. Silva, T. Valente, A. Silva, A. Paiva, and M. Gattass, “Convolutional neural network-based PSO for lung nodule false positive
reduction on CT images,” Computer Methods and Programs in Biomedicine, vol. 162, pp. 109–118, 2018, doi:
10.1016/j.cmpb.2018.05.006.
[15] M. Marhatang and R. D. Muhammad, “Optimal economic dispatch using particle swarm optimization in Sulselrabar system,” IAES
International Journal of Artificial Intelligence, vol. 11, no. 1, pp. 221–228, 2022, doi: 10.11591/ijai.v11.i1.pp221-228.
[16] D. Wang, D. Tan, and L. Liu, “Particle swarm optimization algorithm: an overview,” Soft Computing, vol. 22, pp. 387–408, 2018,
doi: 10.1007/s00500-016-2474-6.
[17] P. Singh, S. Chaudhury, and B. K. Panigrahi, “Hybrid MPSO-CNN: multi-level particle swarm optimized hyperparameters of
convolutional neural network,” Swarm and Evolutionary Computation, vol. 63, pp. 1-13, Jun. 2021, doi:
10.1016/j.swevo.2021.100863.
[18] J. J. Bird, L. J. Manso, E. P. Ribeiro, A. Ekart, and D. R. Faria, “A study on mental state classification using EEG-based brain-
machine interface,” 9th International Conference on Intelligent Systems 2018: Theory, Research and Innovation in Applications,
IS 2018 - Proceedings, pp. 795–800, 2018, doi: 10.1109/IS.2018.8710576.
[19] J. J. Bird, A. Ekart, C. Buckingham, and D. R. Faria, “Mental emotional sentiment classification with an EEG-based brain-machine
interface,” Proceedings of the International Conference on Digital Image and Signal Processing (DISP’19), Oxford University,
UK, pp. 1-7, 2019.
[20] J. T. Hancock and T. M. Khoshgoftaar, “Survey on categorical data for neural networks,” Journal of Big Data, vol. 7, no. 1, 2020,
doi: 10.1186/s40537-020-00305-w.
[21] J. Lanchantin, T. Wang, V. Ordonez, and Y. Qi, “General multi-label image classification with transformers,” Proceedings of the
IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pp. 16473–16483, 2021, doi:
10.1109/CVPR46437.2021.01621.
[22] V. A. Yulianto, N. Effendy, and A. Arif, “Finger vein identification system using capsule networks with hyperparameter tuning,”
IAES International Journal of Artificial Intelligence, vol. 12, no. 4, pp. 1636–1643, Dec. 2023, doi: 10.11591/ijai.v12.i4.pp1636-
1643.
[23] S. Cebollada, L. Payá, M. Flores, A. Peidró, and O. Reinoso, “A state-of-the-art review on mobile robotics tasks using artificial
intelligence and visual data,” Expert Systems with Applications, vol. 167, 2021, doi: 10.1016/j.eswa.2020.114195.
[24] L. Xie, J. Tao, Q. Zhang, and H. Zhou, “CNN and KPCA-based automated feature extraction for real time driving pattern
recognition,” IEEE Access, vol. 7, pp. 123765–123775, 2019, doi: 10.1109/ACCESS.2019.2938768.
[25] A. A. Barbhuiya, R. K. Karsh, and R. Jain, “CNN based feature extraction and classification for sign language,” Multimedia Tools
and Applications, vol. 80, no. 2, pp. 3051–3069, 2021, doi: 10.1007/s11042-020-09829-y.
[26] J. Bird, D. Faria, L. Manso, A. Ekárt, and C. Buckingham, “A deep evolutionary approach to bioinspired classifier optimisation for
brain-machine interaction,” Complexity, vol. 2019, 2019, doi: 10.1155/2019/4316548.

BIOGRAPHIES OF AUTHORS

Dian Palupi Rini holds a Doctor of Computer Science degree from University
Technology Malaysia, in 2017. She also received her B.Sc. (Mathematics) from Sriwijaya
University, Indonesia. and M.Sc. (Computer Science) from Gadjahmada University, Indonesia
in 2000 and 2003, respectively. She is currently a lecturer at Computer Science Faculty,
Sriwijaya University, Indonesia. Her research includes meta-heuristics, global optimization,
machine learning, deep learning, data mining, fuzzy systems, and artificial neural network. She
can be contacted at email: [email protected].

Hyperparameter optimization of convolutional neural network using particle swarm … (Dian Palupi Rini)
560  ISSN: 2252-8938

Tri Kurnia Sari holds the Bachelor of Informatic degree (S.Kom.) from Sriwijaya
University in 2019 and also Master of Computer Science (M.Kom.) from Sriwijaya University
in 2023. She is currently a civil servant as an information technology staff in Padang Selasa
Public Heatlh Center at Palembang City. Her research interest includes data mining. She can be
contacted at email: [email protected].

Winda Kurnia Sari earned a Bachelor's degree in Information Systems (S.SI.)

with a focus on knowledge management, and a Master's degree in Computer Science (M.Kom.)
specializing in Business Intelligence and Natural Language Processing. Currently, she teaches
in the Department of Information Systems at the Faculty of Computer Science, Universitas
Sriwijaya. Her research interests include natural language processing, machine learning, and
deep learning. She can be contacted at email: [email protected].

Novi Yusliani is currently a lecturer and researcher in Faculty of Computer Science,

Universitas Sriwijaya, Indonesia. Her research interest includes natural language processing,
deep learning, and machine learning. She can be contacted at email: [email protected].

Int J Artif Intell, Vol. 14, No. 1, February 2025: 547-560