=Paper=
{{Paper
|id=Vol-2746/paper8
|storemode=property
|title=Development Strategy Model of the Informational Management Logistic System of a Commercial Enterprise by Neural Network Apparatus
|pdfUrl=https://ceur-ws.org/Vol-2746/paper8.pdf
|volume=Vol-2746
|authors=Valeriy Lakhno,Оlena Kryvoruchko,Alyona Desiatko,Andrii Blozva,Viktoriia Semidotska
|dblpUrl=https://dblp.org/rec/conf/cpits/LakhnoKDBS20
}}
==Development Strategy Model of the Informational Management Logistic System of a Commercial Enterprise by Neural Network Apparatus==
https://ceur-ws.org/Vol-2746/paper8.pdf
Development Strategy Model of the Informational
Management Logistic System of a Commercial
Enterprise by Neural Network Apparatus
Valeriy Lakhno1[0000-0001-9695-4543], Оlena Kryvoruchko2[0000-0002-7661-9227],
Alyona Desiatko2[0000-0002-2284-3418], Andrii Blozva1[0000-0002-4377-0916],
and Viktoriia Semidotska2[0000-0002-1333-1154]
1 National University of Life and Environmental Sciences of Ukraine, Ukraine
2 Kyiv National University of Trade and Economics, Ukraine
valss21@ukr.net
Abstract. Models of the development strategy of the information and manage-
ment system of logistics of the trade enterprise are an integral component that
will increase the efficiency of trade enterprise management. Such a solution to
the problem is possible using the neural network apparatus. The experiment is
performed using the analytical platform Deductor for the construction and visu-
alization of Kohonen maps. The Kohonen network is studied by the method of
successive approximations. During the iterative learning procedure, the network
is organized in such a way that the elements that correspond to the centers located
close to each other in the space of entrances will be located close to each other
and on the topological map. Accordingly, the algorithm of self-learning maps
shows how clustering of multidimensional vectors. The constructed Kochhonen
maps, which show the ratio of assets and liabilities, income and expenses of the
enterprise, provide an opportunity to analyze and determine the structural indica-
tors of the enterprise and manage them. The use of the proposed models and
methods made it possible to identify the main factors and sources of improving
the information management system efficiency through the use of logistics ap-
proaches and system developments.
Keywords: Kohonen Networks, Kohonen Maps, Neural Networks, Logistic
Strategy, Information System, Management System.
1 Introduction
Management of the trade enterprise logistics information systems as part of a single
information space of the organization and the use of various content and form compo-
nents of the information resource management system is associated with the develop-
ment of new approaches and means of combining the interaction of material, financial
and information flows.
Copyright © 2020 for this paper by its authors.
Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0)
88
Coordinated management of logistics information systems and the use of various in
content and form components of the information resource of trade enterprise manage-
ment systems is an important problem that requires the use of scientific approaches and
means to solve it.
Designing an information management system for logistics of a trading enterprise is
a complex and multifaceted process that uses all the achievements of modern infor-
mation technologies, each of which makes it possible to manage the organization and
interaction of structural units of the enterprise successfully.
One of the important ways to increase the efficiency of trading companies and net-
works is creating an effective information management system, which is a set of tools,
techniques, artists, providing the information necessary and sufficient implementation
of all operational management processes, and the use of modern information technol-
ogy.
Thus, the design of a model strategy for the development of information and man-
agement systems of trade enterprise logistics is quite relevant and the introduction of
such an information system will increase the efficiency of the management of trade
enterprises.
To solve this problem, it is proposed to use the means of the neural network appa-
ratus. The increase in information about all production, technological processes and
equipment, raw materials and their quality, range of updated products, staff skills, mar-
ket relations lead to a significant complication of management tasks and especially in
making the right effective decision, implementation of business ideas, creating the strat-
egy of development in the process of enterprise management and the industry as a
whole.
An important role is played by information and management systems, which com-
bine the positive qualities of technical electronic means of collecting information, it is
processing, and the person as a direct participant in the formation of management ac-
tions, effective decision-making on enterprise management.
The article aims to use the method of construction of Kohonen maps to find patterns
in the data, which allows for exploratory analysis of data, which differs from classical
statistical procedures. Based on the forecast, obtain a prediction of the values of the
time series for the number of samples, which corresponds to a given forecasting horizon
(this is a financial structure that will correspond to the required level of commercial
enterprises). Conduct an experiment using the Deductor analytical platform to build
and visualize Kohonen maps.
2 Literature Review
Effective application of various approaches and methods of modeling to the strategy of
development of information and management system of the trade enterprise logistics
acquires a decisive advantage in building business models of trade enterprises. The use
of neural network apparatus is proposed, in particular the use of Kohonen networks and
Kohonen maps.
89
Kohonen mapping methods described in the works of authors Chang, Wui Lee, Pang,
Lie Meng, Tay Kai Meng [1–3], in their study the authors O. Kryvoruchko, A. Desiatko
issues identified information and control of logistics trading enterprises [4–8], and in
the writings scholars V. Lakhno, Y. Matus, V. Malyukov, S. Tsiutsiura, Y. Ryndych,
A. Blozva, A. Desiatko, O. Kryvoruchko, and others [8–11] published on the need for
cognitive components of Smart City, in particular, such as enterprise trading.
But today the issue of using Cohonen networks and Cohonen maps in modeling the
strategy of development of information and management system of logistics of a trad-
ing enterprise is not sufficiently revealed and studied.
3 Models and Methods
One of the methods used to build commercial enterprise information and management
logistic system is the use of artificial neural networks. Artificial neural networks are a
collection of neural elements and the connections between them. The main element of
a neural network is a formal neuron.
The principle of operation of such a neuron is as follows: the input signals (𝑥𝑛 ),
which have the appropriate weights (𝑤𝑛 ), are grouped and pass through the transfer
function, generate the result and at the final stage, the output is obtained.
Various learning algorithms and their modifications are used for network training.
One of the methods used to build commercial enterprise information and management
logistic system is the use of Kohonen neural networks and Kohonen maps [1].
Kohonen networks (layers) [1–3] belong to self-organized neural networks. A self-
organized network allows you to identify clusters (groups) of input vectors which have
some common properties.
Clustering is the division of a large number of objects under study into groups of
identical (similar in properties) objects, called clusters [2]. Synonyms of the term “clus-
ter” are the terms class, taxon, thickening. The task of clustering is fundamentally dif-
ferent from the task of classification.
Solving problems of classification is to assign each object to one of the predefined clas-
ses. The division of objects into clusters is carried out while forming clusters. Kohonen
networks are used to cluster objects that are described by quantitative characteristics
Kohonen network has only two layers: input and output, called self-organizing map.
The elements of the map are located in some space, usually two-dimensional.
Kohonen network operation algorithm:
1. Network initialization. The weights of the network are given small random values.
The initial neighborhood area is shown.
2. Presenting the network with a new input signal.
3. Calculation of the distance to all neurons in the network is the distances dj from the
input signal to each neuron j are determined by the formula:
𝑑𝑗 = ∑𝑁
𝑖=1(𝑥𝑖 (𝑡) ∗ 𝑤𝑖𝑗 (𝑡))
2
(1)
90
where 𝑥𝑖 is the 𝑖 th element of the input signal at time 𝑡, 𝑤𝑖𝑗 (𝑡) is the importance of the
connection from the 𝑖 th element of the input signal to the neuron 𝑗 at time 𝑡.
4. Selection of the neuron with the shortest distance is the winning neuron 𝑗 ∗ is selected,
for which the distance 𝑑𝑗 is the smallest.
5. Adjusting the weights of the neuron 𝑗 ∗ and its neighbors is the scales are adjusted
for the neuron 𝑗 ∗ and all neurons from its vicinity 𝑁𝐸. New weight values:
(𝑡 + 1) = 𝑤𝑖𝑗 (𝑡) + 𝑟(𝑡)(𝑥𝑖 (𝑡) − 𝑤𝑖𝑗 (𝑡)). (2)
6. Return to step 2.
The algorithm uses the coefficient of learning speed, which is gradually reduced for a
more subtle correction at a new level As a result, the position of the center is set in a
certain place, which qualitatively clusters the examples for which the neuron is the win-
ner.
In the learning algorithm, the correction is applied not only to the winning neuron
but also to all neurons in its current environment. As a result of this change of environ-
ment, the initial rather large sections of the network migrate towards the educational
examples [2, 3].
The network forms a rough structure of topological order, in which similar examples
activate groups of neurons that are closed on the topological map. With each new era,
the learning speed and the size of the environment decrease, thus within the areas of the
map are more subtle differences, which ultimately leads to a more precise adjustment
of each neuron. Often training is deliberately divided into two phases: shorter, high-
speed training, and large neighborhoods, and longer training with low learning speed
and zero or near-zero neighborhoods. Once the network is trained to recognize the
structure of data (goods), it can be used as a visualization tool in data analysis.
The algorithm of functioning of self-learning maps (Self Organizing Maps, SOM) is
one of the options for clustering multidimensional vectors. An example of such algo-
rithms is the k-means algorithm [2, 5]. An important difference of the SOM algorithm
is that in it all neurons (nodes, centers of classes) are arranged in some structure (usually
a two-dimensional grid). In this case during training not only the winning neuron is
modified, but also its neighbors (to a lesser extent). Due to this, SOM can be considered
as one of the methods of designing multidimensional space in space with a lower di-
mension. When using this algorithm, vectors similar in the source space are detected
side by side on the resulting map. SOM involves the use of an ordered structure of
neurons. Usually, one and two-dimensional grids are used. Each neuron is an n-dimen-
sional column vector [1–3]
𝑤 = [𝑤1 , 𝑤2 , … 𝑤𝑛 ]𝑇 (3)
where n is determined by the dimension of the output space (dimension of the input
vectors).
The use of one- and two-dimensional grids is since there are problems with the dis-
play of larger dimension spatial structures (problems with reducing the dimension to
two-dimensional appear again).
91
Usually, neurons are located in nodes of a two-dimensional grid with rectangular or
hexagonal cells. In this case, as already mentioned, neurons also interact with each
other. The magnitude of this interaction is determined by the distance between the neu-
rons on the map.
This is easy to see that a hexagonal grid distance between neurons longer coincides
with the Euclidean distance than for a quadrangular grid. The number of neurons in the
grid determines the degree of detail of the result of the algorithm, and as a result, the
accuracy of the generalizing ability of the map depends on it.
Kohonen maps (self-organizing map or SOM) [1–3] are designed for the visual rep-
resentation of multidimensional properties of objects on a plane with two axes. Ko-
honen maps conduct mapping input data of high dimensionality on the elements of a
regular array of small dimensions (usually two-dimensional). The difference between
the cards and Kohonen networks is that the map neurons, which are the centers of clus-
ters are arranged in a certain structure (usually a two-dimensional grid).
As a result, similar in some metric input vector network Kohonen belong to one
neuron (cluster center), and Kohonen map can relate to different closely spaced on a
grid of neurons. Usually, neurons are located in the nodes of a two-dimensional grid of
rectangular or hexagonal cells. Neighbor neurons are determined by the distance be-
tween the neurons on the map.
Each cell corresponds to a neuron of the Kohonen network. The number of cells of
the map depends on the required detail of the image and is selected experimentally. For
each cell, one of the statistical characteristics of the selected component of the input
vectors trapped in the cell is calculated [6, 9]. Kohonen map has the shape of a rectangle
in the corners of which groups of the enterprise of divisions are placed, which are sig-
nificantly different from another, and objects with similar characteristics are grouped
in the center of the map.
Management decisions at different hierarchical levels require increasing their objec-
tivity, which is achieved by using economic and mathematical methods in their calcu-
lation methods including multivariate statistical cluster and factor analysis, and the rel-
atively novel approach based on Kohonen maps.
The process of creating a self-organizing map involves establishing a connection
between the input and output layers of neurons and thus displays larger dimension data
on a smaller dimension map, consisting of “neural lattices” [5]. The construction of the
map takes place as a procedure of establishing a correspondence between input and
output, in the so-called competition of neural networks. It should be noted that this
method has significant advantages over cluster analysis, namely: it takes into account
all the indicators selected for analysis. Thus, the maps can determine the general char-
acteristics of each cluster’s similarities and risks inherent in them. This method can be
used to determine the main risks of the company trade. The first group of units, which
always occupies one of the corners of the map, brings together the worst representa-
tives, indicators of which indicate the negative activity of the logistics system, which
informs about the mandatory changes in the activities of units. The angular position is
always occupied by the least profitable services of the enterprise, that went to cease
their activities and require change and renewal plan.
92
Problem areas occupy the southeast corner of the map. Each small cluster of this angle
has certain features, but they all differ significantly from others by a large level of dam-
age. Clusters located closer to the center have slightly better performance, but also com-
bine the most problematic units of the system. In addition to the group of problem units,
the Kohonen map also identifies other problem units that have unbalanced values of
structural indicators. Among each of these groups, there are more or less stable clusters
that are combined with the same risk profile. In the presence of structural distortions
(dependencies, constraints), even a satisfactory level of individual characteristics of the
logistics system can not indicate sufficient security of the trading company [4]. After
analyzing and determining structural indicators that show the ratio of assets and liabil-
ities, income and expenses of the enterprise depending on the service provided and the
construction of the Kohonen map, a simple, clear, and logical visualization of the dis-
tribution of service units is provided. The service units in different segments of the map
with similar characteristics differ not only in the general level of the logistics system
chosen by the researcher but also in the structural features that affect its achievement.
4 Computational Experiment
Assessment of the financial condition of the trade enterprise is carried out at a qualita-
tively new level, taking into account its structural specifics, receipts, and analysis of
funds from sales, storage, and transportation, belonging to the relevant specialized
groups and the relationship between the objects of each group. The platform for build-
ing and visualizing Kohonen maps is Deductor, an analytical platform that allows you
to create an effective business decision support system in a short time.
With powerful import mechanism, Deductor possible to create a unified analytical
add-on to all existing company systems: “Information system protection;” “Defining
the strategy and technology of resource allocation;” “Optimization of hard wear and
soft wear systems;” “Transportation of goods;” “Sales of goods;” “Material flows of
logistics;” “Information logistics;” etc. [4, 8]. The uniqueness of this solution is that
Deductor, if necessary, automatically combines data from different sources.
Completed in Deductor based technology allows a single architecture to pass all
stages of construction analysis system—from creating data warehouses to automatic
selection of models and visualization of the results. Deductor provides the tools needed
to solve a variety of analytical tasks. Corporate reporting, forecasting, segmentation,
search for patterns—these and other tasks that use such analysis techniques as OLAP,
Knowledge Discovery in Databases, and Data Mining [7]. Consider in practice the use
of data from research and production private enterprise “A.V. New technologies,” tak-
ing into account the period (2018–2019). Using the data of Table 1 in Deductor Studio
Academic the initial data are processed, the table is constructed, statistics have resulted
and the diagram is created, results are shown in Fig. 1–3.
93
Fig. 1. Construction of the table according to the data of research and production private enter-
prise “A.V. New technologies.”
Fig. 2. Construction of statistics according to the data of research and production private enter-
prise “A.V. New technologies.”
94
Fig. 3. The construction of the diagram is based on data from Table 1.
After processing the initial data in Deducto Studio Academic the tool “Sliding Win-
dow” is used in Fig. 4. The Sliding Window tool converts a sequence of row values into
a table where adjacent records are represented as adjacent data fields (a window because
only some continuous section of data is allocated, a sliding one because that window
“moves” throughout the set). The need for quantitative data (tables) often occurs when
building models, analyzing, and forecasting time series when applying for entry models
mentioned several adjacent samples from the original data set [9]. Values in one of the
record fields will refer to the current count, and in others—offset from the current count
“in the future” or “in the past.” Thus, the conversion of the variable window has two
parameters: immersion depth—the number of “past” samples that fall into the window;
forecasting horizon—the number of “future” samples.
Fig. 4. Using the Sliding Window.
Based on the created “Sliding Window” Data Mining tools are used: “Neural Network”
(Fig. 5) and “Self-organizing Kohonen maps” (Fig. 6) based on which the forecast of
the future periods is made Fig. 7.
95
Neural networks are self-learning models that mimic the activity of the human brain.
They are able not only to perform a once programmed sequence of actions on certain
data but also to analyze the coming data [11].
The main advantage of neural networks is the ability to build an effective nonlinear
dependence, a more accurate description of data sets compared to linear methods of
statistics. This data processor allows you to specify the structure of the neural network,
determine its parameters, and teach using one of the algorithms available in the system.
The result will be a neural network emulator, which can be used to solve problems
of forecasting, classification, finding hidden patterns, data compression, and many
other applications [7, 9].
Fig. 5. Neural network graph of trading enterprise with revenue reflection according to research
and production private enterprise “A.V. New technologies.”
Kohonen’s self-organizing maps are a powerful self-learning clustering mechanism that
allows you to display results in the form of compact and easy-to-interpret two-dimen-
sional maps. This handler is used to find patterns in large data sets. This allows for
intelligence analysis of data, which differs from classical statistical procedures, during
which a set of hypotheses is tested. [8, 10]. The main advantages of the algorithm:
resistance to noisy data; fast and unmanaged learning; ability to visualize multidimen-
sional input data.
96
Fig. 6. Self-organizing f Kohonen maps.
Forecasting allows obtaining prediction values of time series on the number of samples
that match the prediction horizon. This algorithm works as follows. Suppose that as a
result of the transformation by the method of “sliding window” a sequence of time
samples was obtained:
𝑋 (−𝑛), . . . , 𝑋 (−2), 𝑋 − (− 1), 𝑋0, 𝑋 (+1) (4)
where 𝑋 + 1 predicted value obtained by the previous processing step (for example,
linear regression) based on 𝑛 previous values.
Then to build a forecast for the value 𝑋 (+2), you need to move the whole sequence,
one count, to the left, so that the previously made prediction 𝑋 (+1) is also included in
the number of initial values. Then the algorithm for calculating the predicted value will
be started again, and 𝑋 (+2) will be calculated taking into account 𝑋 (+1) and so on
according to the forecast given by the horizon [9, 12, 13].
Fig. 7. Self-organizing f Kohonen maps.
97
5 The Result of the Experiment
Prediction built on the results of self-organizing Kohonen maps showed different re-
sults, which is logical.
Deducto Studio Academic allows you to perform various operations using Data Min-
ing and analyzing the history of the periods of each of them to draw conclusions and
use the desired result in the further activities of the trade enterprise.
6 Conclusions
We believe that the proposed apparatus of neural networks and Kohonen maps, which
in contrast to the existing mathematical models reflects the results of the study in the
form of visual two-dimensional maps, which allows increasing the efficiency of visu-
alization and analysis of results and forecasting, allows not only to determine the most
profitable logistics strategy but also to assess the financial condition of each object of
study.
Accordingly, using the above models and methods, the main factors and sources of
improving the management information system efficiency through the use of logistics
approaches and system developments are identified.
The model of the strategy of development of the information-managing system of
logistics of the enterprise of trade using the device of neural networks is presented. It
suggests that this approach encourages companies to apply strategies based on cogni-
tive scientific methods and methodologies.
In the future authors plan to research concepts of modeling of informational man-
agement system strategy to implement this in Smart City models.
References
1. Chang, W. L., Meng, P. L., Meng, T. K.: Application of self-organizing map to failure
modes and effects analysis methodology. Neurocomput. 249: 314–320 (2017).
https://doi.org/10.1016/j.neucom.2016.04.073
2. Shah-Hosseini, H.: Binary tree time adaptive self-organizing map. Neurocomput. 74(11):
1823–1839 (2011). https://doi.org/10.1016/j.neucom.2010.07.037
3. Gorban, A. N., Zinovyev, A.: Principal manifolds and graphs in practice: from molecular
biology to dynamical systems. Int. J. Neural Syst. 3(20): 219–232 (2010). https://doi.org/
10.1142/s0129065710002383
4. Desiatko, A., Kryvoruchko, O.: Integrated flows in logistics. Build-Master-Class Work.
Progr. Proc. Int. Sci.-Pract. Conf. Young Sci. (2017). [Publication in Ukrainian]
5. Doran, G. T.: There’s a S.M.A.R.T. Way to Write Management’s Goals and Objectives.
Manag. Rev. 11(70): 35–36 (1981)
6. Torres, L., Pina, V., Royo, S.: E-government and the transformation of public administra-
tions in EU countries: beyond NPM or just a second wave of reforms? Online Inf. Rev.:
5(29): 531–553 (2005). https://doi.org/10.1108/14684520510628918
98
7. Lakhno, V., Malyukov, V., Kryvoruchko, O., Desiatko, A.: Model of evaluating smart city
projects by group of investors using a multifactorial approach. Appl. Technol. First Int.
Conf. 1193: 13–27 (2020). https://doi.org/10.1007/978-3-030-42517-3_2
8. Lakhno, V., Matus, Y., Malyukov, V., Desiatko, A.: Smart city cybersecurity projects fi-
nancing model in case of description of investors’ resources with fuzzy. IEEE Int. Conf.
Adv. Trends Inf. Theory: 249–252 (2019). https://doi.org/10.1109/ATIT49449.2019.
9030499
9. Lakhno, V., Tsiutsiura, S., Ryndych, Y., Blozva, A., Desiatko, A., Usov, Y., Kaznadiy, S.:
Optimization of information and communication transport systems protection tasks. Int. J.
Civ. Eng. Technol. 1(10): 1–9 (2019)
10. Hollnagel, E., Woods, D. D.: Joint Cognitive Systems: Foundations of Cognitive Systems
Engineering (2005). https://doi.org/10.1201/9781420038194
11. Lakhno, V., Malyukov, V., Kasatkin, D., Blozva, A., Litovchenko, T.: Model for supporting
decisions of a group of a smart city project investors, taking into consideration multifactori-
ality. Cybersecur. Edu. Sci. Tech. 6(2): 57–70 (2019). https://doi.org/10.28925/2663-
4023.2019.6.5770
12. Lakhno, V., Malyukov, V., Plyska, L. D.: Model of investment strategies of transport situa-
tional centers in cyber security systems. Cybersecur. Edu. Sci. Tech. 2(2): 68–79 (2018).
https://doi.org/10.15588/1607-3274-2018-2-9
13. Rzаiеvа, S., Rzаiеv, D., Kraskevich, V., Roskladka, A., Gamaliy, V.: Automated logistic
flow system for trading enterprise. Cybersecur. Edu. Sci. Tech. 7(3): 72–84 (2020).
https://doi.org/10.28925/2663-4023.2020.7.7284