IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 12, No. 3, September 2023, pp. 1250~1259

ISSN: 2252-8938, DOI: 10.11591/ijai.v12.i3.pp1250-1259  1250

The evaluation of convolutional neural network and genetic

algorithm performance based on the number of
hyperparameters for English handwritten recognition

Muhammad Munsarif1,2, Edi Noersasongko3, Pulung Nurtantio Andono3, Moch Arief Soeleman3
1
Graduated program of computer science, Dian Nuswantoro University Semarang, Semarang, Indonesia
2
Department of computer science, University of Muhammadiyah Semarang, Semarang, Indonesia
3
Department of computer science, Dian Nuswantoro University Semarang, Semarang, Indonesia

Article Info ABSTRACT

Article history: Convolutional neural network (CNN) has been widely applied to image
recognition, especially handwritten English recognition. CNN's performance
Received Oct 14, 2022 is good if the hyperparameter values are correct. However, the determination
Revised Jan 1, 2023 of precise hyperparameters is not a trivial task. This task is made more
Accepted Jan 10, 2023 difficult when combined with a larger number of hyperparameters resulting in
a high dimensionality of the search space. Usually, hyperparameter
optimization uses a finite number. Previous studies have shown that a large
Keywords: number of hyperparameters can result in optimal CNN performance.
However, the studies only apply to text mining datasets. This study offers two
Convolutional neural networks novelties. First, it applied 20 hyperparameters and their ranges to handwritten
Genetic algorithms English. Second, this paper conducted seven experiments based on different
Handwritten digit recognition hyperparameters and the number of hyperparameters. This paper also
Hyperparameter optimization compares the existing methods, namely random and grid search. The
experiment resulted in the proposed model being superior to the existing
methods. EX3 is better than other experiments and a larger number of
hyperparameters and layer-specific hyperparameter values are unimportant.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Edi Noersasongko
Department Computer science, Dian Nuswantoro University Semarang
Semarang, Indonesia
Email: [email protected]

1. INTRODUCTION
The challenge of finding suitable hyperparameters for convolution neural network (CNN) is still a
concern for researchers. CNN has been widely used in computer vision, especially image recognition. In the
recognition process, CNN has several advantages, such as automatically extracting important features from
each image as well as, saving memory and complexity. The number of hyperparameters strongly influences
the CNN performance process. However, hyperparameter optimization is not easy because the accuracy
depends on the hyperparameters' quality. Furthermore, there is no standard rule in hyperparameter optimization
because each hyperparameter has certain characteristics or is still local optimum. Therefore, the task of
hyperparameter optimization is challenging. It is difficult to improve performance with a large number of
hyperparameters, the weakness of deep neural networks [1]. So, hyperparameter optimization is a solution [2].
Hyperparameter optimization techniques have been proposed. Grid search (GS) [3] and random search
(RS) [4] are popular optimization algorithms. However, the two algorithms do not have a precise learning
mechanism to produce an optimal solution, and each solution produced is independent. This mechanism is

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

Int J Artif Intell ISSN: 2252-8938  1251

fundamental as part of the search activity, which ensures that the optimal solution is global or not trapped in
the local optimal in the high dimension of the search space.
Evolutionary algorithm (EA) provides this ability. EA or Neuroevolution as an optimization tool is,
popular and widely used to direct the learning process [5], [6]. Based on previous research, neuroevolution is
played for optimizing global hyperparameters or the small number of hyperparameters. Most of the studies
related to neuroevolution use genetic algorithm (GA) [7]–[11]. GA has two operators which are crossover and
mutation. These two operators are used for exploration and exploitation of search space. This paper proposes
optimization of hyperparameter and architecture on CNN by using GA or CNN-GA. CNN is a deep learning
model widely applied in various objects because of its high performance such as in computer vision, including
gender classification [12], image classification [13], [14], vehicle tracking system [15], and e-detection [16].
Despite its capabilities, CNN still has some challenges. The selection of hyperparameters strongly
influences its performance. Empirical studies by [17] with text mining data show the strong influence of the
larger number of hyperparameters and values of hyperparameter on CNN performance. However, this claim is
not necessarily proven in different data. Meanwhile, many papers on CNN hyperparameter using
neuroevolution focus on image classification [9], [10], [18]–[23]. Therefore, we focus on CNN optimization
using GA on image recognition.
The paper provides some contributions. First, the CNN-GA optimizes 20 hyperparameters for English
handwritten recognition. To our knowledge, no research has optimized all hyperparameters (Global,
architecture, and layer). Previous research only focused on global parameters or layers. Max-norm weight
constraints in convolutional and dense layers also need to be optimized because values significantly impact the
search process, and different values affect model performance. Second, this paper designed seven experiments
based on the different number of hyperparameters and the different number of optimized hyperparameters.
Based on the study of [17], the larger the number of hyperparameters tends to produce the best model
performance.

2. METHOD
The flow chart for the proposed method is shown in Figure 1. The first stage is pre-processing data.
This research used English handwritten recognition (HR) obtained from the National Institute of Standards and
Technology (NIST) [24]. The number of datasets was 372.450 records. The details are shown in Figure 2. This
figure showa that the distribution of each alphabet digit number the is unbalance, so it becomes an imbalance
dataset. Therefore, this research useds the under-sampling approach to balance the datasets. The balance dataset
is shown in Figure 3. Next, the image size is reshaped and converted into grayscale, then divided or split into
training data and testing data with a proportion of 80:20.

Figure 1. Workflow of the proposed GA-CNN algorithm

We operated standard GA for hyperparameter optimization [25], in which there were five processes:
population initialization, evaluation, selection, crossover, and mutation. The initialization of the population used
is similar to the study of [17]. However, we eliminated embedding in this process. The initialization of the
population is shown in Figure 4. Apart from the same dense layers number (ND) and convolutional layers
number (NC), each individual has the same chromosome length based on the number of hyperparameters
optimized for each experiment. The list, range, and default of hyperparameter values are presented in Table 1.
We refer to the study of [17] that the number of optimized hyperparameters affects the model's performance.
Therefore, we conducted seven experiments as done by the study of [17]. This study generated genes for all
hyperparameter layers for each layer up to a maximum of NC and ND. This generation process was carried out

The evaluation of convolutional neural network and genetic algorithm … (Muhammad Munsarif)
1252  ISSN: 2252-8938

to facilitate crossover and mutation. As a result, the number of genes in populastion size (PS) is a maximum of
NC1. This number is based on the max-pooling layer following every convolution layer except the last layer.

Figure 2. The imbalance of A-Z English HR datasets

Figure 3. The balance of A-Z English HR datasets

Figure 4. Population initialization of the CNN-GA

After the initialization process, the evaluation process was based on the fitness function, namely the
Confusion matrix (accuracy, precision, F1-score, and recall). Figure 1 shows the evaluation process of
transferring the architecture and the resulting hyperparameter values to CNN. Next, CNN returneds the fitness
value. The higher the fitness value, the greater the chance to survive and be selected for the next generation.
Furthermore, the proposed model useds three operators: crossover, mutation, and selection. This
research used a uniform mutation type and randomly selected three types of crossover (one-point, two-point,
and uniform). NC and ND produced a mated architecture using a one-point crossover operator with a global
max-pooling layer. These layers serve as intersection points, which mutate dense layers, or add or remove
convolution layers. The selected individual is the best from each generation, which was then selected using
elitism.

Int J Artif Intell, Vol. 12, No. 3, September 2023: 1250-1259

Int J Artif Intell ISSN: 2252-8938  1253

Table 1. Hyperparameter range and value

Hyperparameter Description Range/Values
Global NE The Number of 𝛼:1,𝛽:100, ℮: 10
epochs
BS Batch Size 𝛼:32, 𝛽::256, ℮: 32
OP Optimizer [‘'Nadam”, 'Adagrad', 'Adadelta', 'Sgd', 'Rmsprop', ‘adam', 'Adamax'] ℮: 'Adam'
LR Learning Rate 𝛼:1e-4, 𝛽 : 1e-2, ℮: 1e-4
MO Momentum 𝛼: 0.0, 𝛽: 1.0, ℮: 0.9
Layer NF Number of filters 𝛼:32, 𝛽:512, ℮: 64
(Convolution)
KS Kernel Size 𝛼: 1, 𝛽: 5, ℮: 3
(Convolution)
AFC Activation function [Sf', 'El', 'SL', 'SF', 'SG', 'Tn', Sid', ' HsG ', 'Linear'], ℮: ℮: ‘RL’
(Convolutional)
KIC Kernel initializer ['Zeros', 'Glorot_Uniform', 'Ones', 'Uniform', 'Normal', 'Glorot_Normal',
(Convolution) 'Lecun_Normal', Lecun_Uniform', 'He_Normal','He_Uniform'] ℮: ℮: ‘Glorot Uniform’
WC Max-norm weight 𝛼: 1, 𝛽: 5, ℮: 3
C constraint
(Convolutional)
NN Number of neuron 𝛼: 1, 𝛽: 5, ℮: 1
(Dense)
AFD Activation function ['RL', 'Sf', 'Elu', 'SL', 'SF', 'SG', 'Tn', Sid', 'HsG', 'Linear'], ℮: 'RL'
(Dense)
KID Kernel initializer [Zeros','Ones', 'Uniform', 'Lecun_Normal', Lecun_Uniform' 'Normal', 'Glorot_Normal',
(Dense) 'Glorot_Uniform', 'He_Normal','He_Uniform'], ℮: 'Glorot_Uniform'
WC Max-norm weight 𝛼: 1, 𝛽: 5, ℮: 3
D constraint (Dense)
DR Drop rate 𝛼: 0.0, 𝛽 1.0, ℮: 0.2
(Dropout)
PS Pool size (Max- 𝛼: 2, 𝛽: 6, ℮: 5
pooling)
KIO Kernel initializer ['Zeros,’Glorot_Normal’,'Ones', 'Uniform', 'Normal', Glorot_Uniform',
(Output) 'He_Normal','He_Uniform',],
℮: 'Glorot_Uniform'
Architect NC Convolutional 𝛼: 1, 𝛽: 15, ℮: 1
ure layers number
ND Dense layers 𝛼: 0, 𝛽: 15, ℮: 1
number
Note: 𝛼: Min; 𝛽:Max, ℮: Default, Sf:softmax,El:Elu,SL: Selu,SF: Softplus, SG: Softsign, Tn:Tanh, Sid: Sigmoid, HsG:
Hard_Sigmoid,RL; Relu

3. RESULT AND DISCUSSION

This section discusses the performance of CNN+GA at obtaining the near-optimum combination of
the hyperparameters and architecture of CNN. English HR dataset was used to assess the performance of the
model [24]. In this study, splits the dataset was split by 80% for training and 20% for testing. The training
dataset is also split training dataset and the validation dataset with 80:20 ration [26]. The recognition of English
HR is an image recognition that recognises the digits of the alphabet into 28 classes, namely A-Z or a-z. In this
experiment, the optimisation method used the training and validation data to produce the near-optimum
architecture and combination of hyperparameters. Python was used to implement all hyperparameter
optimizations. SciKit-Learn [3] calculated the confusion matrix and visualization tools using Matplotlib [27]
and Seaborn. Pandas [28] are used to process datasets, and NumPy [29] handle all scientific computing.
Tensorflow [30] and Keras libraries were used to build the CNN model. Distributed evolutionary algorithms
in Python (DEAP) was used to create GA. Finally, Google colabPro Plus with GPU high ram was used to
perform all the experiments in Table 2.
The use of different hyperparameter values in each layer is marked with an asterisk (*). We refer to
[17] to perform seven experiments based on the different number of hyperparameters and the different number
of optimized hyperparameters as presented in Table 2. In [17], the optimization of the number of
hyperparameters depends on whether each layer will have the same or different values. This is realized by
putting one star (*) on the hyperparameter layer, which means that the hyperparameter is selected once and is
the same for all layers. Meanwhile, the hyperparameter layer with two stars (**) means that each layer will
have a different hyperparameter value. Experiment pairs 2-3, 4-5, and 6-7 have the same number of
hyperparameters, but the number of optimized hyperparameters differs.
Determination of GA parameter values, namely crossover rate (CR) and mutation rate (MR), is
essential in the hyperparameter optimization process. This paper determines CR=0.8 and MR = 0.2. This value
is determined based on papers [17], which refer to several papers [8], [10]. The number of generations (Ngen)
and population (Npop) selected is 25 in this experiment. Generally, a high population size and a large

The evaluation of convolutional neural network and genetic algorithm … (Muhammad Munsarif)
1254  ISSN: 2252-8938

generation size will result in better performance. However, this selection will take longer. Based on previous
research, the selection of small sizes has also been widely used and proven to produce a good performance
[8]–[10], [17], [31].

Table 1. Hyperparameter design for each experiment

Description Experiment
1 2 3 4 5 6
Global NE V V V V V V
BS V V V V V V
OP V V V V V V
LR V V V V V V
MO V
Layer NF V* V** V* V** V* V**
KS V* V** V* V** V* V**
A F C V* V** V* V** V* V**
KIC V* V** V* V**
WCC V* V** V* V**
NN V* V** V* V** V* V**
AFD V* V** V* V**
KID V* V** V* V**
WCD V* V** V* V**
DR V V
PS V* V**
KIO V V
Architecture NC V V V V V V
ND V V V V V V
Hyperparameter numbers 6 10 10 15 15 20 20
Optimized hyperparameters numbers 6 10 66 15 141 20 159
Notes: *) All layers with the same Value, **) All layers with different values

This study used an evaluation matrix, which is accuracy, in all experiments. Figure 5 shows the
minimum and maximum accuracy results in all experiments. This Figure shows that the higher the size of the
population (Npop), the accuracy increases. Experiment 3 (EX 3) has superior accuracy than the other
experiments. Meanwhile, Figure 6 shows Mean and Standart Deviasion of Accuracy on CNN-GA. Figure 6 is
also in line with Figure 5; a higher number of population sizes results in better average performance.
Meanwhile, each population in each generation produces almost the same accuracy value. This means that the
distribution of accuracy performance results is uniform so that each population produces a small standard
deviation value. EX3 excels with other experiments. Figure 7 shows the average distance (AvgDis) and time
execution of accuracy in all experiments.
The smaller AvgDis of each population, the better the accuracy performance produced. This figure
shows that EX1 and EX3 have a stable AvgDis, but EX3 has better performance than EX1. Then the experiment
resulted in better accuracy performance requiring a longer time. This is in accordance with previous studies,
showing that to produce optimal hyperparameters using GA requires more time. Therefore EX3 is an
experiment with a longer time than other experiments.
As Figures 5-7 show EX3 is the best experiment among other experiments, with an accuracy of
93.77% and a total execution time of 10 hours 22 minutes. These results indicate that the GA's hyperparameter
optimization process can produce the best accuracy if the execution time is long. The comparison of the best
accuracy and total execution time is shown in Figure 8.
Furthermore, this paper also presents a comparison of CNN+RS as a comparison of the proposed
model shown in Figure 9. Similarly, CNN+RS was built in seven experiments with the same list and range of
hyperparameters on CNN+GA. This study set iterations = 150 to run CNN+RS. The best accuracy results show
that CNN+GA is superior in all experiments from CNN+RS. Meanwhile, the total execution time resulted in
CNN+GA being longer in all experiments than CNN+RS. The best comparison of the best accuracy and total
execution time of the two models is shown in Figure 8. Besides RS, this study executed GS by optimizing only
six hyperparameters from EX1. This choice was made because only EX1 could cover all 525 CNN's evaluated.
If all hyperparameters were optimized, it would require at least 1,48,576 CNNs. Meanwhile, GS could not
optimize a number of these CNNs, and the characteristics of CNNs are known to be time-consuming [32].
Therefore, this study used the list and range hyperparameters in GS, similar to the paper [17] presented in
Table 3.

Int J Artif Intell, Vol. 12, No. 3, September 2023: 1250-1259

Int J Artif Intell ISSN: 2252-8938  1255

Figure 5. Minimum and maximum of accuracy on CNN-GA

Figure 6. Mean and standart deviasion of accuracy on CNN-GA

The evaluation of convolutional neural network and genetic algorithm … (Muhammad Munsarif)
1256  ISSN: 2252-8938

Figure 7. The average distance and time execution of accuracy on CNN-GA

Figure 8. The summarize of best accuracy and total time excecution on CNN-GA

Figure 9. The comparison of CNN+GA and CNN+RS

Int J Artif Intell, Vol. 12, No. 3, September 2023: 1250-1259

Int J Artif Intell ISSN: 2252-8938  1257

Table 3. Hyperparameter value sets of GS [17]

Hyperparameter Set of Values
NE [20,40,60,80,100]
BS [64,256]
OP [‘Adam’, ‘Adamax’, ‘Nadam’, ‘Adagrad’, ‘ RmsProp’ ]
LR [1e-3, 1e-2]
NCL [5, 10, 15]]
NDL [5, 10, 15]]

Based on the comparison of the three models, it is found that CNN+GA is superior to CNN+GS and
CNN+RS in terms of accuracy performance. The comparison results can be seen in Figure 10. This result is in
line with the research obtained by Fatyanosa suggesting that. GA is still superior to GS and RS in optimizing
CNN hyperparameters with image and text datasets.

Figure 10. The comparison of three models

The resulting architecture in the best model is shown in Figure 11. The architecture consists of four
convolution layers and four dense layers. Deep architecture can study a wide variety of handwritten character
datasets and generate architecture for learning handwritten datasets.

Figure 11. The best individual GA-CNN in EX3

4. CONCLUSION
This paper describes an automated approach to optimize 20 CNN hyperparameters using GA
(CNN+GA) on an English handwritten dataset. The results of the seven experiments show that the larger
number of hyperparameters and layer-specific hyperparameters are sometimes important as the use of max-
norm weight constraint hyperparameters does not significantly impact the search process. It is the standard
CNN architecture that produces the best performance compared to CNN with complete hyperparameters. The
best result in this study is that EX3 achieves an accuracy of 97.77%.

The evaluation of convolutional neural network and genetic algorithm … (Muhammad Munsarif)
1258  ISSN: 2252-8938

REFERENCES
[1] A. Ghatak, “Deep learning with R,” Deep Learning with R, pp. 1–245, 2019, doi: 10.1007/978-981-13-5850-0.
[2] A. Mishra, “Evaluating Machine Learning Models,” Machine Learning in the AWS Cloud, pp. 115–132, 2019, doi:
10.1002/9781119556749.ch5.
[3] F. Pedregosa et al., “Scikit-learn: Machine learning in Python,” Journal of Machine Learning Research, vol. 12, pp. 2825–2830,
2011.
[4] J. Bergstra and Y. Bengio, “Random search for hyper-parameter optimization,” Journal of Machine Learning Research, vol. 13,
pp. 281–305, 2012.
[5] K. O. Stanley, J. Clune, J. Lehman, and R. Miikkulainen, “Designing neural networks through neuroevolution,” Nature Machine
Intelligence, vol. 1, no. 1, pp. 24–35, 2019, doi: 10.1038/s42256-018-0006-z.
[6] Z. Fouad, M. Alfonse, M. Roushdy, and A. B. M. Salem, “Hyper-parameter optimization of convolutional neural network based on
particle swarm optimization algorithm,” Bulletin of Electrical Engineering and Informatics, vol. 10, no. 6, pp. 3377–3384, 2021,
doi: 10.11591/eei.v10i6.3257.
[7] N. Bansal, A. Sharma, and R. K. Singh, “An evolving hybrid deep learning framework for legal document classification,” Ingenierie
des Systemes d’Information, vol. 24, no. 4, pp. 425–431, 2019, doi: 10.18280/isi.240410.
[8] A. Dahou, M. A. Elaziz, J. Zhou, and S. Xiong, “Arabic Sentiment Classification Using Convolutional Neural Network and
Differential Evolution Algorithm,” Computational Intelligence and Neuroscience, vol. 2019, 2019, doi: 10.1155/2019/2537689.
[9] T. Hinz, N. Navarro-Guerrero, S. Magg, and S. Wermter, “Speeding up the Hyperparameter Optimization of Deep Convolutional
Neural Networks,” International Journal of Computational Intelligence and Applications, vol. 17, no. 2, 2018, doi:
10.1142/S1469026818500086.
[10] Y. Sun, B. Xue, M. Zhang, G. G. Yen, and J. Lv, “Automatically Designing CNN Architectures Using the Genetic Algorithm for
Image Classification,” IEEE Transactions on Cybernetics, vol. 50, no. 9, pp. 3840–3854, 2020, doi: 10.1109/TCYB.2020.2983860.
[11] S. Sanders and C. Giraud-Carrier, “Informing the use of hyperparameter optimization through metalearning,” Proceedings - IEEE
International Conference on Data Mining, ICDM, vol. 2017-November, pp. 1051–1056, 2017, doi: 10.1109/ICDM.2017.137.
[12] Y. Nie, W. Ren, B. Liang, J. Dai, and P. Huang, “A study on image based gender classification using convolutional neural network,”
ACM International Conference Proceeding Series, pp. 81–84, 2019, doi: 10.1145/3342999.3343011.
[13] F. Sultana, A. Sufian, and P. Dutta, “Advancements in image classification using convolutional neural network,” Proceedings -
2018 4th IEEE International Conference on Research in Computational Intelligence and Communication Networks, ICRCICN
2018, pp. 122–129, 2018, doi: 10.1109/ICRCICN.2018.8718718.
[14] L. Niharmine, B. Outtaj, and A. Azouaoui, “Tifinagh handwritten character recognition using optimized convolutional neural
network,” International Journal of Electrical and Computer Engineering, vol. 12, no. 4, pp. 4164–4171, 2022, doi:
10.11591/ijece.v12i4.pp4164-4171.
[15] H. J. Kim, “Multiple vehicle tracking and classification system with a convolutional neural network,” Journal of Ambient
Intelligence and Humanized Computing, vol. 13, no. 3, pp. 1603–1614, 2022, doi: 10.1007/s12652-019-01429-5.
[16] R. Ranjan et al., “A Fast and Accurate System for Face Detection, Identification, and Verification,” IEEE Transactions on
Biometrics, Behavior, and Identity Science, vol. 1, no. 2, pp. 82–96, 2019, doi: 10.1109/tbiom.2019.2908436.
[17] T. N. Fatyanosa and M. Aritsugi, “Effects of the Number of Hyperparameters on the Performance of GA-CNN,” Proceedings -
2020 IEEE/ACM International Conference on Big Data Computing, Applications and Technologies, BDCAT 2020, pp. 144–153,
2020, doi: 10.1109/BDCAT50828.2020.00016.
[18] E. Real et al., “Large-scale evolution of image classifiers,” 34th International Conference on Machine Learning, ICML 2017, vol.
6, pp. 4429–4446, 2017.
[19] T. Desell, “Developing a volunteer computing project to evolve convolutional neural networks and their hyperparameters,”
Proceedings - 13th IEEE International Conference on eScience, eScience 2017, pp. 19–28, 2017, doi: 10.1109/eScience.2017.14.
[20] I. Loshchilov and F. Hutter, “CMA-ES for Hyperparameter Optimization of Deep Neural Networks,” 2016, [Online]. Available:
http://arxiv.org/abs/1604.07269.
[21] A. Martín, R. Lara-Cabrera, F. Fuentes-Hurtado, V. Naranjo, and D. Camacho, “EvoDeep: A new evolutionary approach for
automatic Deep Neural Networks parametrisation,” Journal of Parallel and Distributed Computing, vol. 117, pp. 180–191, 2018,
doi: 10.1016/j.jpdc.2017.09.006.
[22] H. Xie, L. Zhang, and C. P. Lim, “Evolving CNN-LSTM Models for Time Series Prediction Using Enhanced Grey Wolf Optimizer,”
IEEE Access, vol. 8, pp. 161519–161541, 2020, doi: 10.1109/ACCESS.2020.3021527.
[23] B. Van Stein, H. Wang, and T. Back, “Automatic Configuration of Deep Neural Networks with Parallel Efficient Global
Optimization,” Proceedings of the International Joint Conference on Neural Networks, vol. 2019-July, 2019, doi:
10.1109/IJCNN.2019.8851720.
[24] P. J. Grother and K. K. Hanaoka, “NIST Special Database 19,” pp. 1–30, 2016, [Online]. Available: https://s3.amazonaws.com/nist-
srd/SD19/1stEditionUserGuide.pdf%0Ahttps://www.nist.gov/srd/nist-special-database-19.
[25] R. L. Haupt, “An Introduction to Genetic Algorithms for Electromagnetics,” IEEE Antennas and Propagation Magazine, vol. 37,
no. 2, pp. 7–15, 1995, doi: 10.1109/74.382334.
[26] T. S. Gunawan, A. F. R. M. Noor, and M. Kartiwi, “Development of english handwritten recognition using deep neural network,”
Indonesian Journal of Electrical Engineering and Computer Science, vol. 10, no. 2, pp. 562–568, 2018, doi:
10.11591/ijeecs.v10.i2.pp562-568.
[27] J. D. Hunter, “Matplotlib: A 2D graphics environment,” Computing in Science and Engineering, vol. 9, no. 3, pp. 90–95, 2007, doi:
10.1109/MCSE.2007.55.
[28] W. McKinney, “Data Structures for Statistical Computing in Python,” Proceedings of the 9th Python in Science Conference, pp.
56–61, 2010, doi: 10.25080/majora-92bf1922-00a.
[29] S. Van Der Walt, S. C. Colbert, and G. Varoquaux, “The NumPy array: A structure for efficient numerical computation,” Computing
in Science and Engineering, vol. 13, no. 2, pp. 22–30, 2011, doi: 10.1109/MCSE.2011.37.
[30] M. Abadi et al., “TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems,” 2016, [Online]. Available:
http://arxiv.org/abs/1603.04467.
[31] G. C. Felbinger, “Optimal CNN Hyperparameters for Object Detection on NAO Robots,” 2018.
[32] C. Fox, “Python for Data Science Primer,” pp. 15–25, 2018, doi: 10.1007/978-3-319-72953-4_2.

Int J Artif Intell, Vol. 12, No. 3, September 2023: 1250-1259

Int J Artif Intell ISSN: 2252-8938  1259

BIOGRAPHIES OF AUTHORS

Muhammad Munsarif received the Master Degree a n d Graduated program of

computer science, Dian Nuswantoro University Semarang, Indonesia.He is a lecturer in
informatics Engineering at Muhammadiyah University, Semarang (UNIMUS). His research
interests include computer vision, data science and technopreneuership. He can be contacted
at email: [email protected]

Edi Noersasongko was born in Semarang, Central Java, Indonesia, in June 1955.
He is currently a Senior Professor of artificial intelligence and technopreneurship. He also has a
position as the President or Rector of Universitas Dian Nuswantoro, Semarang, Indonesia. His
area of interests includes artificial intelligence and technopreneurship. He can be contacted at
email: [email protected]

Pulung Nurtantio Andono was born in Central Java, Indonesia, in September 1982.
He received the Bachelor of Engineering from Universitas Trisakti, Jakarta, in 2006, and the
Master of Computer Science from Universitas Dian Nuswantoro (UDINUS), in 2009, and the
Ph.D. degree from the Institut Teknologi Sepuluh Nopember (ITS), Surabaya, in 2014. He is
currently an Associate Professor at Dian Nuswantoro University. His area of interest includes
3D image reconstruction and computer vision. He can be contacted at email:
[email protected]

Moch Arief Soeleman received bachelor and master degree in 1999 and 2004
respectively from Computer Science Department Dian Nuswantoro University, Semarang,
Indonesia. Since 2016, he have received Ph.D degree at the Electrical Engineering of Institut
Teknologi Sepuluh Nopember (ITS), Surabaya, Indonesia. He works as a lecturer of Computer
Science Faculty, Dian Nuswantoro University, Semarang, Indonesia. His research interest
includes image processing, computer vision and video processing. He can be contacted at email:
[email protected]

The evaluation of convolutional neural network and genetic algorithm … (Muhammad Munsarif)