=Paper=
{{Paper
|id=Vol-2608/paper5
|storemode=property
|title=Smart crossover mechanism for parallel neuroevolution method of medical diagnostic models synthesis
|pdfUrl=https://ceur-ws.org/Vol-2608/paper5.pdf
|volume=Vol-2608
|authors=Serhii Leoshchenko,Sergey Subbotin,Andrii Oliinyk,Viktor Lytvyn,Matviy Ilyashenko
|dblpUrl=https://dblp.org/rec/conf/cmis/LeoshchenkoSOLI20
}}
==Smart crossover mechanism for parallel neuroevolution method of medical diagnostic models synthesis==
https://ceur-ws.org/Vol-2608/paper5.pdf
Smart Crossover Mechanism for Parallel Neuroevolution
Method of Medical Diagnostic Models Synthesis
Serhii Leoshchenko1[0000-0001-5099-5518], Sergey Subbotin2[0000-0001-5814-8268], Andrii Oli-
inyk3[0000-0002-6740-6078], Viktor Lytvyn4[0000-0003-4061-4755]
and Matviy Ilyashenko5[0000-0003-4624-4687]
1,2,3,4
Dept. of Software Tools, National University "Zaporizhzhia Polytechnic", 69063 Za-
porizhzhia, Ukraine
5
Dept. of Computer Systems and networks, National University "Zaporizhzhia Polytechnic",
69063 Zaporizhzhia, Ukraine
1
sergleo.zntu@gmail.com 2subbotin@zntu.edu.ua
3
olejnikaa@gmail.com 4lytvynviktor.a@gmail.com
5
matviy.ilyashenko@gmail.com
Abstract. Information technologies significantly expand the capabilities of
modern medicine. Using of artificial neural networks is becoming particularly
promising. They are actively used for diagnostics based on patient data. How-
ever, the problem of network synthesis with a satisfactory topology and accu-
rate diagnosis remains important. Neuroevolution methods allow to solve this
problem without much involvement of an expert. Moreover, these methods
make it possible to effectively use the parallel power of modern computer sys-
tems. On the other hand, parallelization raises a number of new problems. The
paper suggests mechanisms for improving the crossover operator. The proposed
solution allows you to reduce resource consumption and improve the synthesis
process.
Keywords: Medical Diagnosis, Forecasting, Neuroevolution, Synthesis, Adap-
tive Mechanism, Genetic Algorithm, Parallel Genetic Algorithm, Crossover.
1 Introduction
The human factor often causes a number of problems. This also applies, of course, to
medicine. A doctor's mistake can mean the loss of a patient's health or even their life,
and doctors make mistakes not infrequently. Even the highest-level professional can
make mistakes, because the specialist can be tired, irritated, concentrating on the
problem worse than usual [1-3].
In this case, information technologies can come to the rescue. For example, the
IBM Watson cognitive system [4-7] copes with work in the medical field at a fairly
high level (oncology, reading x-rays, etc.) [4], [5]. But there are other solutions pro-
posed by independent researchers. A number of scientists regularly present successful
results of using artificial intelligence systems in medical diagnostics [8-11].
Copyright © 2020 for this paper by its authors. Use permitted under Creative
Commons License Attribution 4.0 International (CC BY 4.0).
Artificial neural networks (ANN) have come into practice wherever it is necessary
to solve problems of forecasting, classification or management. Problems of medical
diagnostics are a sub-division of problems classification and forecasting of the ob-
ject's condition. The following reasons determine the impressive success of the use of
ANNs [10-13]:
a huge number of opportunities [12], [13]. ANNs are a powerful modeling method
that allows reproducing extremely complex dependencies. In particular, neural
networks naturally are nonlinear. In problems where linear application is unsatis-
factory (most medical diagnostics problems), linear models do not work well. In
addition, neural networks cope with modeling linear dependencies in the case of a
large number of variables;
lеraining on examples (data about the object) [14], [15]. Neural networks are trained
using examples. Initially, representative data is selected, and then a training algo-
rithm is run that automatically perceives the data structure.
On the other hand, a user who trains a neural network needs a certain set of knowl-
edge about how to select and prepare data, select the appropriate network architecture,
and interpret the results [14-20]. Therefore, it can be assumed that the probability of
error passes from the expert in knowledge of the problem area to the expert in the
design of the ANN [16], [18].
Neuroevolution methods of ANN synthesis are based on an evolutionary approach
[21]. These methods can simultaneously configure the network structure and weights.
This allows getting ready-made neural network solutions with only data from the
training sample, and in most cases does not require the user to have deep knowledge
of the ANN theory [22], [23].
Advantages of using neuroevolution methods [24-27]:
─ a wide variety of resulting topologies – possible solutions with" non-standard "
ANNs structures;
─ adaptivity;
─ universality;
─ the use does not require a deep knowledge about the ANN;
─ possibility of using a parallel approach.
The most well-established neuroevolution method is the genetic algorithm (GA).
Unlike other optimization technologies, GA contain a population of trial solutions
that are competitively managed using defined operators [28-30]. GA is inherent in
iterative training of a population of individuals. The capacity of GA increases with the
use of distributed computing. Such algorithms are called parallel genetic algorithms.
They are based on splitting a population into several separate subpopulations, each of
which will be processed by GA independently of other subpopulations.
However, when it comes to parallelizing, it should be taken into account that the
main disadvantage of GA is a constant desire for a population containing one local
optimum, which leads to a constant decrease in the genetic diversity of the population,
which reduces the ability of GA to search for a global optimum and/or adapt to chang-
ing parameters of target functions [31], [32]. Therefore, the task of improving the
methods of evolutionary optimization based on increasing the adaptive characteristics
of the method is urgent.
2 Literature review
As noted above, GA is one of the most acceptable ways to synthesize ANN. This is
due to the fact that at the initial stage there is absolutely no information about the
direction of movement in terms of setting the matrix weights. Under conditions of
uncertainty, evolutionary methods, including GA, have the highest chances to achieve
the necessary results.
At the same time, there is a significant disadvantage of using genetic algorithms-the
multi – iterative nature of the methods, serious time costs. To solve such problems,
parallelization can be used to branch out the execution of calculations between the
cores of modern computer systems. However, when designing a parallel approach,
you should consider the most common problems encountered:
selecting or developing a strategy for interaction between the components of the
algorithm;
choosing the frequency of migration between populations;
determination of migrating individuals and their number;
determining the structure of the evolution of individual populations.
Consider the problems in more detail. The structure of a parallel system is an impor-
tant factor in the performance of a parallel algorithm, since it determines how quickly
(or how slowly) a good solution spreads to other populations. If the system is strongly
connected, then good solutions will quickly spread to all streams and can quickly
"saturate" the population [33]. On the other hand, if the network is loosely connected,
solutions will spread more slowly and threads will be more isolated from each other.
Further parallel development and recombination of different solutions can occur to
obtain potentially better solutions.
A common trend in promising parallel genetic algorithms (PGA) [34], [35] is the
using of static system structures that are defined before the algorithm is run and re-
main unchanged.
Another method of constructing a structure is to create a dynamic system [33-35].
In this case, the flow is not limited to links with a certain fixed number of threads;
instead, migrants are directed to threads that meet a certain criterion. As a similar
criterion, the degree of population diversity or the measure of genotypic distance be-
tween two populations (or the distance from a characteristic individual of the popula-
tion, for example, a favorite) is taken. This structure requires mechanisms for tracking
events in neighboring populations, and if an event occurred in one of the neighboring
populations, then an event should be expected in the second population.
The frequency of migrations also has a big impact on the final decision [33], [34].
As it known, too frequent migrations lead to degeneration of populations, and rare
ones, on the contrary, to a decrease in convergence. Various methods are used to
regulate the frequency of migration, which can be divided into two types: adaptive
and event-based. In the first case, adaptation methods are used to adjust the migration
frequency during the algorithm operation. In the second case, methods are used that
determine the need for migration, that is, migration is performed only when an event
occurs.
Selection mechanisms are used to select individuals for migration [36-40]. It is
known that individual chromosomes may contain "good" fragments of the genetic
code, but these parts may be in chromosomes that are poorly adapted. But at the same
time, excluding similar ones can lead to premature convergence, or skipping the glob-
al optimum.
Using different strategies imposes the main restriction: the necessity of forming the
same type of chromosome structure. But the effect that is possible with successful
formation can be much greater than when using a single GA structure in all popula-
tions [37].
It is also worth noting that a large number of migrations and even dynamic ex-
change of intermediate information between threads requires additional overhead,
which sometimes largely negates the achievement of parallel execution of calcula-
tions [36], [38].
Moreover, we should not forget that even the sequential version of GA imposes
significant resource requirements (especially RAM), compared to gradient methods
[39], [40]. This is because in most cases a population of neural networks is used.
Hence, when parallelizing, keep in mind that the requirements will increase according
to the number of threads involved.
Therefore, the use of additional mechanisms and the introduction of hybrid meth-
ods, at different stages of operation, will improve the performance of the PGA.
3 Parallel genetic algorithm with smart crossover
The paper proposes a parallel genetic method. Parallelization will occur in the follow-
ing way. Initially, we will generate a population of individuals on the main core of the
system, where each individual is a separate neural network:
P NN1 , NN2 ,..., NNn , (1)
where, P is population,
NN n is neural network,
n is the size of population.
Also, at the initial stage, the main free parameters of the method will be set: the
stop criteria, population size, and so on.
The next step is to divide the population into subpopulations and distribute sub-
populations between the cores of the multi-core system.
The cores will perform the same sequence of actions with subpopulations: evaluat-
ing the genetic information of individuals, sorting and selecting the best individuals,
crossing and selecting the new best individuals.
Then the selected best individuals are sent to the main core, where they are re-
sorted and selected the best representatives. The new best individual is evaluated for
the stop criterion. If the resulting individual is satisfied, it becomes a solution and the
method ends. If the solution is not satisfactory, the best individuals from the subpopu-
lations are crossed and a new population is obtained, which is re-distributed between
the system cores.
Diversity is maintained in the process of evolution by the fact that each species
(subpopulation) develops without exchanging genetic material with other species.
This is an important aspect of this model. The exchange of genetic material between
two different species usually produces non-viable offspring. In addition, mixing of
genetic material can reduce the diversity of populations. This approach will also sig-
nificantly reduce the share of overhead associated with the transfer of information
between the system cores.
Now let's look at the mechanisms that allow optimizing the use of RAM, while
maintaining the adaptive characteristics of the method. In the proposed method, it is
recommended to use a smart crossover, which is based on uniform crossing and rank
selection enhanced by criteria conditions.
Uniform crossover is one of the most effective recombination operators in standard
GA [41], [42].
Uniform crossover is performed according to a randomly selected standard that
specifies which genes should be inherited from the first parent (other genes are taken
from the second parent) [42]. In other words the general rule of uniform crossover can
be represented as follows:
Ind 3 Crossover( Ind1 , Ind 2 , DataofCros);
g Ind3 {g 1 Rand g1Ind1 , g 1Ind2 ,
g 2 Rand g 2 Ind1 , g 2 Ind2 ,..., (2)
g Rand g
i iInd1
, g iInd2 }.
It has long been known that setting the probability of passing a parent gene to a de-
scendant in a uniform crossover can significantly increase its efficiency [35], [36], and
also allows emulating other crossing operators (single-point, two-point). It is also
known that the use of the uniform crossing operator makes it possible to apply the so-
called multi-parent recombination, when more than two parents are used to generate
one offspring. Despite this, in most studies, only two parents are used and the fixed
probability of transmitting the gene is 0.5 [41].
Even crossing gives greater flexibility when combining rows, which is an impor-
tant advantage when working with GA.
However, it should be noted that even crossing requires additional computing pow-
er. On the other hand, uniform crossover makes it possible to emulate the operation of
simpler types of crossing, such as two-point crossover. Therefore, we will use a two-
point crossover to work on threads. This approach will allow you to implement this
method in the future on computing systems running graphics processors (GPUs).
It is proposed to strengthen the ranking selection by introducing additional criteria
that will help to track various characteristics of neural networks more subtly, namely:
excessive memory usage and approximation properties of the neural network.
The general scheme of the method is shown in Fig. 1.
4 Experiment
4.1 Description of the experiment
A sample of Parkinson's Disease Classification Data Set will be used for the experi-
ment [43].
The data used in this study were gathered from 188 patients with PD (107 men and
81 women) with ages ranging from 33 to 87 (65.1±10.9). The data used in this study
were gathered from 188 patients with PD (107 men and 81 women) with ages ranging
from 33 to 87 (65.1±10.9) at the Department of Neurology in Cerrahpasa Faculty of
Medicine, Istanbul University.
The control group consists of 64 healthy individuals (23 men and 41 women) with
ages varying between 41 and 82 (61.1±8.9). During the data collection process, the
microphone is set to 44.1 KHz and following the physician’s examination, the sus-
tained phonation of the vowel was collected from each subject with three repetitions.
Table 1 shows the main characteristics of the data sample.
Table 1. Main characteristics of the Parkinson's Disease Classification Data Set
Criterion Characteristic Criterion Characteristic
Data Set Characteristics Multivariate Number of Instances 756
Attribute Characteristics Integer, Real Number of Attributes 754
The following hardware and software have been used for experimental verification
of the proposed parallel genetic method for ANN synthesis: the computing system of
the Department of software tools of Zaporizhzhya national technical university (Na-
tional university “Zaporizhzhia politechnic”), Zaporizhzhia: Xeon processor E5-2660
v4 (14 cores), RAM 4x16 GB DDR4, the programming model of Java threads.
The results of the proposed method will be compared with the results of the meth-
od considered in the previous works [44]. Old modification will be called PGM (Par-
allel genetic method) and new variant – PGM SC (Parallel genetic method with smart
crossover). Note that when working, the size of the parent pool for even crossing will
be equal to the number of system cores involved.
4.2 The results of the experiment
In the Fig. 2 is graph of the execution time (in minutes) of the proposed method on
computer systems, which depends on the number of cores involved.
Fig. 1. Parallel genetic method for ANN synthesis with smart crossover
It can be seen from the graphs that the proposed method has an acceptable degree of
parallelism and is effectively performed on MIMD parallel system. In addition, the
processor in multi-core computer supports Turbo Boost technology [45–47], making
the time of the method execution on the single core much less than on the core of
which does not support this technology.
Fig. 3 shows graphs of changes in communication overhead. Since the new method
offers the transfer of only the best individuals, this allows you to significantly reduce
the transfer of excess information. This is explained by the fact that communication
overhead of the proposed method execution on computer systems is relatively small,
and the number of parallel operations significantly exceeds the number of serial op-
erations and synchronizations. In communication overhead, is understood the ratio of
the time spent by the system for transfers and synchronization among cores to the
time of target calculations on a given number of cores.
Fig. 2. Dependence the execution time of the proposed method to the number of involved cores
of the computing system
Fig. 3. Communication overhead performing the proposed method to the number of cores in-
volved of the computing system
Fig. 4 shows graphs of speedup changes. Based on the fact that the communication
overhead has decreased, the speedup of execution increases significantly.
The graph of efficiency of computer systems is presented in Fig. 5. It shows that
the using of even 14 cores of computer systems for the implementation of the pro-
posed method retains the efficiency at a relatively acceptable level and indicates the
potential, if necessary and possibly, to use even more cores.
Fig. 4. The speedup graphics of calculations on computing system
Fig. 5. The efficiency graph of computing systems when executing the proposed method
5 Conclusion
Based on the results of experiments, it can be argued that the proposed method can be
used in the synthesis of neural network diagnostic models. The proposed smart cross-
over mechanism significantly optimizes the synthesis process by using an adjustable
uniform crossover and additional criteria at the selection stage.
However, we are not talking about large-scale implementation of neural networks
in hospitals around the world. The main problem in terms of the spread of these tech-
nologies is that neural networks are a kind of "black box". Specialists enter data and
get a certain result. But the creators of such systems may not fully understand how
this result was obtained, what algorithms and in what sequence are involved. If neural
networks could be made more transparent, and the principle of their operation could
be easily explained to medical practitioners, then the rate of spread of this technology
would be much higher.
Acknowledgment
The work is supported by the state budget scientific research project of National
University "Zaporizhzhia Polytechnic" “Intelligent methods and software for diagnos-
tics and non-destructive quality control of military and civilian applications” (state
registration number 0119U100360).
References
1. Einarson, T.R., Acs, A., Ludwig, C., Panton, U.H.: Prevalence of cardiovascular disease in
type 2 diabetes: a systematic literature review of scientific evidence from across the world
in 2007–2017. Cardiovasc Diabetol 17 (1), 1-19 (2018). https://doi.org/10.1186/s12933-
018-0728-6
2. Petrie, J.R., Guzik, T.J., Touyz, R.M.: Diabetes, Hypertension, and Cardiovascular Dis-
ease: Clinical Insights and Vascular Mechanisms. The Canadian journal of cardiology
34(5), 575-584 (2017). https://doi.org/10.1016/j.cjca.2017.12.005
3. Jungen, C., Scherschel, K., Eickholt, C., Kuklik, P., Klatt, N.: Disruption of cardiac cho-
linergic neurons enhances susceptibility to ventricular arrhythmias. Nature Communica-
tions 8 (2017). https://doi.org/10.1038/ncomms14155
4. Bakkar, N., Kovalik, T., Lorenzini, I. et al.: Artificial intelligence in neurodegenerative
disease research: use of IBM Watson to identify additional RNA-binding proteins altered
in amyotrophic lateral sclerosis. Acta Neuropathol 135, 227–247 (2018).
https://doi.org/10.1007/s00401-017-1785-8
5. Gomes, C., Dietterich, T., Barrett, C. et al.: Computational sustainability: computing for a
better world and a sustainable future. Communications of The ACM 62 (9), 56–65 (2019).
https://doi.org/10.1145/3339399
6. Kolpakova, T., Oliinyk, A., Lovkin, V.: Improved method of group decision making in ex-
pert systems based on competitive agents selection. UKRCON: IEEE First Ukraine Con-
ference on Electrical and Computer Engineering (UKRCON), 939–943. KPI, Kiyv (2017).
https://doi.org/10.1109/UKRCON.2017.8100388
7. IBM Watson Homepage, https://www.ibm.com/watson
8. Pranav, R., Anirudh, J., Anuj, P. et al.: CheXpedition: Investigating Generalization Chal-
lenges for Translation of Chest X-Ray Algorithms to the Clinical Setting (2020),
https://arxiv.org/pdf/2002.11379.pdf
9. Artificial intelligence rivals radiologists in screening X-rays for certain diseases,
https://med.stanford.edu/news/all-news/2018/11/ai-outperformed-radiologists-in-
screening-x-rays-for-certain-diseases.html
10. Stanford researchers develop artificial intelligence tool to help detect brain aneurysms,
https://news.stanford.edu/2019/06/07/ai-tool-helps-radiologists-detect-brain-aneurysms/
11. Leoshchenko S, Oliinyk A., Subbotin, S, Zaiko T: Using modern architectures of recurrent
neural networks for technical diagnosis of complex systems. In: Proceedings of the 2018
international scientific-practical conference problems of infocommunications. science and
technology (PIC S&T), 411–416. IEEE, Kharkov (2018).
https://doi.org/10.1109/infocommst.2018.8632015
12. Schmidt, J., Marques, M.R.G., Botti, S. et al.: Recent advances and applications of ma-
chine learning in solid-state materials science. npj Comput Mater 5, 83 (2019).
https://doi.org/10.1038/s41524-019-0221-0
13. Bradley, J.R., Holan, S.H., Wikle, C.K.: Multivariate spatio-temporal models for high-
dimensional areal data with application to Longitudinal Employer-Household Dynamics.
The Annals of Applied Statistics, 9(4), 1761-1791 (2015).
14. A Tour of Machine Learning Algorithms, https://machinelearningmastery.com/a-tour-of-
machine-learning-algorithms/
15. Sabater-Mir, J., Torra, V., Aguiló, I., González-Hidalgo, M.: Artificial Intelligence Re-
search and Development. IOS Press, Amsterdam (2019).
16. Scher, S., Messori, G.: Generalization properties of feed-forward neural networks trained
on Lorenz systems. Nonlinear Processes in Geophysics 26, 381–399 (2019).
17. How to Scale Data for Long Short-Term Memory Networks in Python,
https://machinelearningmastery.com/how-to-scale-data-for-long-short-term-memory-
networks-in-python/
18. Learning process of a neural network, https://towardsdatascience.com/how-do-artificial-
neural-networks-learn-773e46399fc7
19. Oliinyk, A.O., Zayko, T.A., Subbotin, S.O.: Synthesis of Neuro-Fuzzy Networks on the
Basis of Association Rules. Cybern Syst Anal 50, 348–357 (2014).
https://doi.org/10.1007/s10559-014-9623-7
20. Oliinyk, A.O., Skrupsky, S.Y., Subbotin, S.A.: Using parallel random search to train fuzzy
neural networks. Aut. Control Comp. Sci. 48, 313–323 (2014).
https://doi.org/10.3103/S0146411614060078
21. Stanley, K.O., Clune, J., Lehman, J. et al.: Designing neural networks through neuroevolu-
tion. Nature Machine Intelligence 1, 24–35 (2019). https://doi.org/10.1038/s42256-018-
0006-z.
22. Baldominos, A., Saez, Y. & Isasi, P.: On the automated, evolutionary design of neural
networks: past, present, and future. Neural Comput & Applic 32, 519–545 (2020).
https://doi.org/10.1007/s00521-019-04160-6
23. Gaier, A., Asteroth, A., Mouret, J.-B.: Data-efficient Neuroevolution with Kernel-Based
Surrogate Models. GECCO '18: Proceedings of the Genetic and Evolutionary Computation
Conference, 85–92 (2018). 10.1145/3205455.3205510.
24. Albrigtsen, S.I., Imenes A.: Neuroevolution of Actively Controlled Virtual Characters.
University of Agder, Kristiansand & Grimstad (2017).
25. Bohrer, J., Grisci, B., Dorn, M.: Neuroevolution of Neural Network Architectures Using
CoDeepNEAT and Keras (2020), https://arxiv.org/abs/2002.04634
26. Phillips B.: Deep Neuroevolution: Genetic Algorithms are a Competitive Alternative for
Training Deep Neural Networks for Reinforcement Learning – Paper Summary,
https://towardsdatascience.com/deep-neuroevolution-genetic-algorithms-are-a-
competitive-alternative-for-training-deep-neural-822bfe3291f5
27. Gradient descent vs. neuroevolution, https://towardsdatascience.com/gradient-descent-vs-
neuroevolution-f907dace010f
28. García-Martínez C., Rodriguez F.J., Lozano M. (2018) Genetic Algorithms. In: Martí R.,
Pardalos P., Resende M. (eds) Handbook of Heuristics. Springer, Cham
29. Lee, W.-P., Hsiao, Y.-T., Hwang, W.-C.: Designing a parallel evolutionary algorithm for
inferring gene networks on the cloud computing environment. BMC Systems Biology 8(1),
23-28 (2014). https://doi.org/10.1186/1752-0509-8-5.
30. Gonçalves, I., Gomes, A., Henggeler, C.: Optimizing residential energy resources with an
improved multi-objective genetic algorithm based on greedy mutations. GECCO '18: Pro-
ceedings of the Genetic and Evolutionary Computation Conference, 1246–1253 (2018).
https://doi.org/10.1145/3205455.3205616
31. Zhang, F., Mei, Y., Zhang, M.: A two-stage genetic programming hyper-heuristic ap-
proach with feature selection for dynamic flexible job shop scheduling. GECCO '19: Pro-
ceedings of the Genetic and Evolutionary Computation Conference, 347–355 (2019).
https://doi.org/10.1145/3321707.3321790
32. Stork, J., Eiben, A., Bartz-Beielstein, T.: A new Taxonomy of Continuous Global Optimi-
zation Algorithms (2018), https://arxiv.org/pdf/1808.08818.pdf
33. Potuzak, T.: Optimization of a genetic algorithm for road traffic network division using a
distributed/parallel genetic algorithm. 2016 9th International Conference on Human Sys-
tem Interactions (HSI), 2016, pp. 21-27. https://doi.org/10.1109/HSI.2016.7529603
34. Hoseini Alinodehi, S.P., Moshfe, S., Saber Zaeimian, M., et al.: High-Speed General Pur-
pose Genetic Algorithm Processor. IEEE Transactions on Cybernetics 46(7), 1551–1565,
(2016). https://doi.org/10.1109/TCYB.2015.2451595
35. Hou, N., He, F., Zhou, Y., Chen, Y., Yan, X.: A Parallel Genetic Algorithm With Disper-
sion Correction for HW/SW Partitioning on Multi-Core CPU and Many-Core GPU. IEEE
Access 6, 883–898 (2018). https://doi.org/10.1109/ACCESS.2017.2776295
36. Varun Kumar, S.G., Panneerselvam, Dr.R.: A Study of Crossover Operators for Genetic
Algorithms to Solve VRP and its Variants and New Sinusoidal Motion Crossover Opera-
tor. International Journal of Computational Intelligence Research 13(7), 1717–1733
(2017).
37. Hassanat, A., Alkafaween, E.: On Enhancing Genetic Algorithms Using New Crossovers,
https://arxiv.org/ftp/arxiv/papers/1801/1801.02335.pdf
38. Umbarkar, Dr.A., Sheth, P.: Crossover operators in genetic algorithms: a review. ICTACT
Journal on Soft Computing 6(1), 1083–1092 (2015). https://doi.org/6.
10.21917/ijsc.2015.0150.
39. Oliinyk, A.O., Skrupsky, S.Y., Subbotin, S.A.: Experimental investigation with analyzing
the training method complexity of neuro-fuzzy networks based on parallel random search.
Aut. Control Comp. Sci. 49, 11–20 (2015). https://doi.org/10.3103/S0146411615010071
40. Leoshchenko, S., Oliinyk, A., Skrupsky, S., Subbotin, S., Zaiko T.: Parallel Method of
Neural Network Synthesis Based on a Modified Genetic Algorithm Application. Work-
shop Proceedings of the 8th International Conference on “Mathematics. Information Tech-
nologies. Education”, MoMLeT&DS-2019, 11–23. Lviv Polytechnic National University,
Lviv (2019). https://dblp.org/rec/conf/momlet/LeoshchenkoOSSZ19
41. Baeck, T., Fogel, D.B., Michalewicz, Z.: Evolutionary Computation 1: Basic Algorithms
and Operators. CRC Press, Boca Raton (2000).
42. Das, A.K., Pratihar, D.K. A directional crossover (DX) operator for real parameter optimi-
zation using genetic algorithm. Appl Intell 49, 1841–1865 (2019).
https://doi.org/10.1007/s10489-018-1364-2
43. Parkinson's Disease Classification Data Set, https://archive.ics.uci.edu/ml/datasets/ Parkin-
son%27s+Disease+Classification
44. Leoshchenko, S., Oliinyk, A., Skrupsky, S., Subbotin, S., Lytvyn V.: Parallel Genetic
Method for the Synthesis of Recurrent Neural Networks for Using in Medicine. In: Pro-
ceedings of the Second International Workshop on Computer Modeling and Intelligent
Systems (CMIS-2019), pp. 1-17 (2019).
https://dblp.org/rec/conf/cmis/LeoshchenkoOSSL19
45. Intel Turbo Boost Technology 2.0, https://www.intel.com/content/www/us/en/architecture-
and-technology/turbo-boost/turbo-boost-technology.html
46. Oliinyk, A., Subbotin, S., Lovkin, V., Leoshchenko, S., Zaiko, T. Feature Selection Based
on Parallel Stochastic Computing. 13th International Scientific and Technical Conference
on Computer Sciences and Information Technologies (CSIT), 347-351. IEEE, Lviv (2018).
https://doi.org/10.1109/STC-CSIT.2018.8526729.
47. Shkarupylo, V., Skrupsky, S., Oliinyk, A., Kolpakova, T.: Development of stratified ap-
proach to software defined networks simulation. Eastern-European Journal of Enterprise
Technologies, 5. 67-73 (2017). https://doi.org/10.15587/1729-4061.2017.110142.